$(INSTALL) -m0755 -D initramfs/hook $(DESTDIR)$(INITRAMFS_HOOK)
$(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mkfs.bcachefs
$(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/fsck.bcachefs
- $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mount.bcachefs
+ # See: #1057295
+ # $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mount.bcachefs
$(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mkfs.fuse.bcachefs
$(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/fsck.fuse.bcachefs
$(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mount.fuse.bcachefs
+bcachefs-tools (24+really1.3.4-2~exp1) experimental; urgency=medium
+
+ [ Chris Hofstaedtler ]
+ * Non-maintainer upload.
+ * Install files into /usr instead of /. (Closes: #1059373)
+
+ [ Jonathan Carter ]
+ * Do not install bcachefs.mount symlink (Closes: #1057295)
+
+ -- Jonathan Carter <jcc@debian.org> Wed, 27 Dec 2023 19:22:06 +0200
+
bcachefs-tools (24+really1.3.4-1) unstable; urgency=medium
* New upstream release
--- /dev/null
+Description: do not install mount.bcachefs symlink
+Bug-Debian: https://bugs.debian.org/1059373
+Last-Update: 2023-12-27
+
+--- bcachefs-tools-24+really1.3.4.orig/Makefile
++++ bcachefs-tools-24+really1.3.4/Makefile
+@@ -159,7 +159,8 @@ install: bcachefs
+ $(INSTALL) -m0755 -D initramfs/hook $(DESTDIR)$(INITRAMFS_HOOK)
+ $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mkfs.bcachefs
+ $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/fsck.bcachefs
+- $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mount.bcachefs
++ # See: #1057295
++ # $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mount.bcachefs
+ $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mkfs.fuse.bcachefs
+ $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/fsck.fuse.bcachefs
+ $(LN) -sfr $(DESTDIR)$(ROOT_SBINDIR)/bcachefs $(DESTDIR)$(ROOT_SBINDIR)/mount.fuse.bcachefs
--- /dev/null
+do-not-install-mount-symlink
export NO_RUST=-true
PREFIX := /usr
+ROOT_SBINDIR := /usr/sbin
DEB_BUILD_ARCH ?= $(shell dpkg-architecture -qDEB_BUILD_ARCH)
dh $@
override_dh_auto_install:
- dh_auto_install -- "PREFIX=$(PREFIX)"
+ dh_auto_install -- "PREFIX=$(PREFIX)" "ROOT_SBINDIR=$(ROOT_SBINDIR)"
override_dh_auto_test:
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BBPOS_TYPES_H
+#define _BCACHEFS_BBPOS_TYPES_H
+
+struct bbpos {
+ enum btree_id btree;
+ struct bpos pos;
+};
+
+static inline struct bbpos BBPOS(enum btree_id btree, struct bpos pos)
+{
+ return (struct bbpos) { btree, pos };
+}
+
+#define BBPOS_MIN BBPOS(0, POS_MIN)
+#define BBPOS_MAX BBPOS(BTREE_ID_NR - 1, POS_MAX)
+
+#endif /* _BCACHEFS_BBPOS_TYPES_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "bset.h"
+#include "btree_journal_iter.h"
+#include "journal_io.h"
+
+#include <linux/sort.h>
+
+/*
+ * For managing keys we read from the journal: until journal replay works normal
+ * btree lookups need to be able to find and return keys from the journal where
+ * they overwrite what's in the btree, so we have a special iterator and
+ * operations for the regular btree iter code to use:
+ */
+
+static int __journal_key_cmp(enum btree_id l_btree_id,
+ unsigned l_level,
+ struct bpos l_pos,
+ const struct journal_key *r)
+{
+ return (cmp_int(l_btree_id, r->btree_id) ?:
+ cmp_int(l_level, r->level) ?:
+ bpos_cmp(l_pos, r->k->k.p));
+}
+
+static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
+{
+ return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
+}
+
+static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
+{
+ size_t gap_size = keys->size - keys->nr;
+
+ if (idx >= keys->gap)
+ idx += gap_size;
+ return idx;
+}
+
+static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
+{
+ return keys->d + idx_to_pos(keys, idx);
+}
+
+static size_t __bch2_journal_key_search(struct journal_keys *keys,
+ enum btree_id id, unsigned level,
+ struct bpos pos)
+{
+ size_t l = 0, r = keys->nr, m;
+
+ while (l < r) {
+ m = l + ((r - l) >> 1);
+ if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
+ l = m + 1;
+ else
+ r = m;
+ }
+
+ BUG_ON(l < keys->nr &&
+ __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
+
+ BUG_ON(l &&
+ __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
+
+ return l;
+}
+
+static size_t bch2_journal_key_search(struct journal_keys *keys,
+ enum btree_id id, unsigned level,
+ struct bpos pos)
+{
+ return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
+}
+
+struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
+ unsigned level, struct bpos pos,
+ struct bpos end_pos, size_t *idx)
+{
+ struct journal_keys *keys = &c->journal_keys;
+ unsigned iters = 0;
+ struct journal_key *k;
+search:
+ if (!*idx)
+ *idx = __bch2_journal_key_search(keys, btree_id, level, pos);
+
+ while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
+ if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
+ return NULL;
+
+ if (__journal_key_cmp(btree_id, level, pos, k) <= 0 &&
+ !k->overwritten)
+ return k->k;
+
+ (*idx)++;
+ iters++;
+ if (iters == 10) {
+ *idx = 0;
+ goto search;
+ }
+ }
+
+ return NULL;
+}
+
+struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
+ unsigned level, struct bpos pos)
+{
+ size_t idx = 0;
+
+ return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
+}
+
+static void journal_iters_fix(struct bch_fs *c)
+{
+ struct journal_keys *keys = &c->journal_keys;
+ /* The key we just inserted is immediately before the gap: */
+ size_t gap_end = keys->gap + (keys->size - keys->nr);
+ struct btree_and_journal_iter *iter;
+
+ /*
+ * If an iterator points one after the key we just inserted, decrement
+ * the iterator so it points at the key we just inserted - if the
+ * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
+ * handle that:
+ */
+ list_for_each_entry(iter, &c->journal_iters, journal.list)
+ if (iter->journal.idx == gap_end)
+ iter->journal.idx = keys->gap - 1;
+}
+
+static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
+{
+ struct journal_keys *keys = &c->journal_keys;
+ struct journal_iter *iter;
+ size_t gap_size = keys->size - keys->nr;
+
+ list_for_each_entry(iter, &c->journal_iters, list) {
+ if (iter->idx > old_gap)
+ iter->idx -= gap_size;
+ if (iter->idx >= new_gap)
+ iter->idx += gap_size;
+ }
+}
+
+int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bkey_i *k)
+{
+ struct journal_key n = {
+ .btree_id = id,
+ .level = level,
+ .k = k,
+ .allocated = true,
+ /*
+ * Ensure these keys are done last by journal replay, to unblock
+ * journal reclaim:
+ */
+ .journal_seq = U32_MAX,
+ };
+ struct journal_keys *keys = &c->journal_keys;
+ size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
+
+ BUG_ON(test_bit(BCH_FS_RW, &c->flags));
+
+ if (idx < keys->size &&
+ journal_key_cmp(&n, &keys->d[idx]) == 0) {
+ if (keys->d[idx].allocated)
+ kfree(keys->d[idx].k);
+ keys->d[idx] = n;
+ return 0;
+ }
+
+ if (idx > keys->gap)
+ idx -= keys->size - keys->nr;
+
+ if (keys->nr == keys->size) {
+ struct journal_keys new_keys = {
+ .nr = keys->nr,
+ .size = max_t(size_t, keys->size, 8) * 2,
+ };
+
+ new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
+ if (!new_keys.d) {
+ bch_err(c, "%s: error allocating new key array (size %zu)",
+ __func__, new_keys.size);
+ return -BCH_ERR_ENOMEM_journal_key_insert;
+ }
+
+ /* Since @keys was full, there was no gap: */
+ memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
+ kvfree(keys->d);
+ *keys = new_keys;
+
+ /* And now the gap is at the end: */
+ keys->gap = keys->nr;
+ }
+
+ journal_iters_move_gap(c, keys->gap, idx);
+
+ move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
+ keys->gap = idx;
+
+ keys->nr++;
+ keys->d[keys->gap++] = n;
+
+ journal_iters_fix(c);
+
+ return 0;
+}
+
+/*
+ * Can only be used from the recovery thread while we're still RO - can't be
+ * used once we've got RW, as journal_keys is at that point used by multiple
+ * threads:
+ */
+int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bkey_i *k)
+{
+ struct bkey_i *n;
+ int ret;
+
+ n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
+ if (!n)
+ return -BCH_ERR_ENOMEM_journal_key_insert;
+
+ bkey_copy(n, k);
+ ret = bch2_journal_key_insert_take(c, id, level, n);
+ if (ret)
+ kfree(n);
+ return ret;
+}
+
+int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bpos pos)
+{
+ struct bkey_i whiteout;
+
+ bkey_init(&whiteout.k);
+ whiteout.k.p = pos;
+
+ return bch2_journal_key_insert(c, id, level, &whiteout);
+}
+
+void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
+ unsigned level, struct bpos pos)
+{
+ struct journal_keys *keys = &c->journal_keys;
+ size_t idx = bch2_journal_key_search(keys, btree, level, pos);
+
+ if (idx < keys->size &&
+ keys->d[idx].btree_id == btree &&
+ keys->d[idx].level == level &&
+ bpos_eq(keys->d[idx].k->k.p, pos))
+ keys->d[idx].overwritten = true;
+}
+
+static void bch2_journal_iter_advance(struct journal_iter *iter)
+{
+ if (iter->idx < iter->keys->size) {
+ iter->idx++;
+ if (iter->idx == iter->keys->gap)
+ iter->idx += iter->keys->size - iter->keys->nr;
+ }
+}
+
+static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
+{
+ struct journal_key *k = iter->keys->d + iter->idx;
+
+ while (k < iter->keys->d + iter->keys->size &&
+ k->btree_id == iter->btree_id &&
+ k->level == iter->level) {
+ if (!k->overwritten)
+ return bkey_i_to_s_c(k->k);
+
+ bch2_journal_iter_advance(iter);
+ k = iter->keys->d + iter->idx;
+ }
+
+ return bkey_s_c_null;
+}
+
+static void bch2_journal_iter_exit(struct journal_iter *iter)
+{
+ list_del(&iter->list);
+}
+
+static void bch2_journal_iter_init(struct bch_fs *c,
+ struct journal_iter *iter,
+ enum btree_id id, unsigned level,
+ struct bpos pos)
+{
+ iter->btree_id = id;
+ iter->level = level;
+ iter->keys = &c->journal_keys;
+ iter->idx = bch2_journal_key_search(&c->journal_keys, id, level, pos);
+}
+
+static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
+{
+ return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
+ iter->b, &iter->unpacked);
+}
+
+static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
+{
+ bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
+}
+
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
+{
+ if (bpos_eq(iter->pos, SPOS_MAX))
+ iter->at_end = true;
+ else
+ iter->pos = bpos_successor(iter->pos);
+}
+
+struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
+{
+ struct bkey_s_c btree_k, journal_k, ret;
+again:
+ if (iter->at_end)
+ return bkey_s_c_null;
+
+ while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
+ bpos_lt(btree_k.k->p, iter->pos))
+ bch2_journal_iter_advance_btree(iter);
+
+ while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
+ bpos_lt(journal_k.k->p, iter->pos))
+ bch2_journal_iter_advance(&iter->journal);
+
+ ret = journal_k.k &&
+ (!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
+ ? journal_k
+ : btree_k;
+
+ if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
+ ret = bkey_s_c_null;
+
+ if (ret.k) {
+ iter->pos = ret.k->p;
+ if (bkey_deleted(ret.k)) {
+ bch2_btree_and_journal_iter_advance(iter);
+ goto again;
+ }
+ } else {
+ iter->pos = SPOS_MAX;
+ iter->at_end = true;
+ }
+
+ return ret;
+}
+
+void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
+{
+ bch2_journal_iter_exit(&iter->journal);
+}
+
+void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+ struct bch_fs *c,
+ struct btree *b,
+ struct btree_node_iter node_iter,
+ struct bpos pos)
+{
+ memset(iter, 0, sizeof(*iter));
+
+ iter->b = b;
+ iter->node_iter = node_iter;
+ bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
+ INIT_LIST_HEAD(&iter->journal.list);
+ iter->pos = b->data->min_key;
+ iter->at_end = false;
+}
+
+/*
+ * this version is used by btree_gc before filesystem has gone RW and
+ * multithreaded, so uses the journal_iters list:
+ */
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+ struct bch_fs *c,
+ struct btree *b)
+{
+ struct btree_node_iter node_iter;
+
+ bch2_btree_node_iter_init_from_start(&node_iter, b);
+ __bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
+ list_add(&iter->journal.list, &c->journal_iters);
+}
+
+/* sort and dedup all keys in the journal: */
+
+void bch2_journal_entries_free(struct bch_fs *c)
+{
+ struct journal_replay **i;
+ struct genradix_iter iter;
+
+ genradix_for_each(&c->journal_entries, iter, i)
+ if (*i)
+ kvpfree(*i, offsetof(struct journal_replay, j) +
+ vstruct_bytes(&(*i)->j));
+ genradix_free(&c->journal_entries);
+}
+
+/*
+ * When keys compare equal, oldest compares first:
+ */
+static int journal_sort_key_cmp(const void *_l, const void *_r)
+{
+ const struct journal_key *l = _l;
+ const struct journal_key *r = _r;
+
+ return journal_key_cmp(l, r) ?:
+ cmp_int(l->journal_seq, r->journal_seq) ?:
+ cmp_int(l->journal_offset, r->journal_offset);
+}
+
+void bch2_journal_keys_free(struct journal_keys *keys)
+{
+ struct journal_key *i;
+
+ move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
+ keys->gap = keys->nr;
+
+ for (i = keys->d; i < keys->d + keys->nr; i++)
+ if (i->allocated)
+ kfree(i->k);
+
+ kvfree(keys->d);
+ keys->d = NULL;
+ keys->nr = keys->gap = keys->size = 0;
+}
+
+static void __journal_keys_sort(struct journal_keys *keys)
+{
+ struct journal_key *src, *dst;
+
+ sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
+
+ src = dst = keys->d;
+ while (src < keys->d + keys->nr) {
+ while (src + 1 < keys->d + keys->nr &&
+ src[0].btree_id == src[1].btree_id &&
+ src[0].level == src[1].level &&
+ bpos_eq(src[0].k->k.p, src[1].k->k.p))
+ src++;
+
+ *dst++ = *src++;
+ }
+
+ keys->nr = dst - keys->d;
+}
+
+int bch2_journal_keys_sort(struct bch_fs *c)
+{
+ struct genradix_iter iter;
+ struct journal_replay *i, **_i;
+ struct jset_entry *entry;
+ struct bkey_i *k;
+ struct journal_keys *keys = &c->journal_keys;
+ size_t nr_keys = 0, nr_read = 0;
+
+ genradix_for_each(&c->journal_entries, iter, _i) {
+ i = *_i;
+
+ if (!i || i->ignore)
+ continue;
+
+ for_each_jset_key(k, entry, &i->j)
+ nr_keys++;
+ }
+
+ if (!nr_keys)
+ return 0;
+
+ keys->size = roundup_pow_of_two(nr_keys);
+
+ keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
+ if (!keys->d) {
+ bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
+ nr_keys);
+
+ do {
+ keys->size >>= 1;
+ keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
+ } while (!keys->d && keys->size > nr_keys / 8);
+
+ if (!keys->d) {
+ bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
+ keys->size);
+ return -BCH_ERR_ENOMEM_journal_keys_sort;
+ }
+ }
+
+ genradix_for_each(&c->journal_entries, iter, _i) {
+ i = *_i;
+
+ if (!i || i->ignore)
+ continue;
+
+ cond_resched();
+
+ for_each_jset_key(k, entry, &i->j) {
+ if (keys->nr == keys->size) {
+ __journal_keys_sort(keys);
+
+ if (keys->nr > keys->size * 7 / 8) {
+ bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
+ keys->nr, keys->size, nr_read, nr_keys);
+ return -BCH_ERR_ENOMEM_journal_keys_sort;
+ }
+ }
+
+ keys->d[keys->nr++] = (struct journal_key) {
+ .btree_id = entry->btree_id,
+ .level = entry->level,
+ .k = k,
+ .journal_seq = le64_to_cpu(i->j.seq),
+ .journal_offset = k->_data - i->j._data,
+ };
+
+ nr_read++;
+ }
+ }
+
+ __journal_keys_sort(keys);
+ keys->gap = keys->nr;
+
+ bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);
+ return 0;
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_JOURNAL_ITER_H
+#define _BCACHEFS_BTREE_JOURNAL_ITER_H
+
+struct journal_iter {
+ struct list_head list;
+ enum btree_id btree_id;
+ unsigned level;
+ size_t idx;
+ struct journal_keys *keys;
+};
+
+/*
+ * Iterate over keys in the btree, with keys from the journal overlaid on top:
+ */
+
+struct btree_and_journal_iter {
+ struct btree *b;
+ struct btree_node_iter node_iter;
+ struct bkey unpacked;
+
+ struct journal_iter journal;
+ struct bpos pos;
+ bool at_end;
+};
+
+struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *, enum btree_id,
+ unsigned, struct bpos, struct bpos, size_t *);
+struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *, enum btree_id,
+ unsigned, struct bpos);
+
+int bch2_journal_key_insert_take(struct bch_fs *, enum btree_id,
+ unsigned, struct bkey_i *);
+int bch2_journal_key_insert(struct bch_fs *, enum btree_id,
+ unsigned, struct bkey_i *);
+int bch2_journal_key_delete(struct bch_fs *, enum btree_id,
+ unsigned, struct bpos);
+void bch2_journal_key_overwritten(struct bch_fs *, enum btree_id,
+ unsigned, struct bpos);
+
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *);
+struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *);
+
+void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *);
+void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
+ struct bch_fs *, struct btree *,
+ struct btree_node_iter, struct bpos);
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
+ struct bch_fs *,
+ struct btree *);
+
+void bch2_journal_keys_free(struct journal_keys *);
+void bch2_journal_entries_free(struct bch_fs *);
+
+int bch2_journal_keys_sort(struct bch_fs *);
+
+#endif /* _BCACHEFS_BTREE_JOURNAL_ITER_H */
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_KEY_CACHE_TYPES_H
+#define _BCACHEFS_BTREE_KEY_CACHE_TYPES_H
+
+struct btree_key_cache_freelist {
+ struct bkey_cached *objs[16];
+ unsigned nr;
+};
+
+struct btree_key_cache {
+ struct mutex lock;
+ struct rhashtable table;
+ bool table_init_done;
+
+ struct list_head freed_pcpu;
+ size_t nr_freed_pcpu;
+ struct list_head freed_nonpcpu;
+ size_t nr_freed_nonpcpu;
+
+ struct shrinker shrink;
+ unsigned shrink_iter;
+ struct btree_key_cache_freelist __percpu *pcpu_freed;
+
+ atomic_long_t nr_freed;
+ atomic_long_t nr_keys;
+ atomic_long_t nr_dirty;
+};
+
+struct bkey_cached_key {
+ u32 btree_id;
+ struct bpos pos;
+} __packed __aligned(4);
+
+#endif /* _BCACHEFS_BTREE_KEY_CACHE_TYPES_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_gc.h"
+#include "btree_io.h"
+#include "btree_iter.h"
+#include "btree_journal_iter.h"
+#include "btree_key_cache.h"
+#include "btree_update_interior.h"
+#include "btree_write_buffer.h"
+#include "buckets.h"
+#include "errcode.h"
+#include "error.h"
+#include "journal.h"
+#include "journal_reclaim.h"
+#include "replicas.h"
+#include "snapshot.h"
+
+#include <linux/prefetch.h>
+
+static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
+{
+#ifdef CONFIG_BCACHEFS_DEBUG
+ struct bch_fs *c = trans->c;
+ struct bkey u;
+ struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
+
+ if (unlikely(trans->journal_replay_not_finished)) {
+ struct bkey_i *j_k =
+ bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
+
+ if (j_k)
+ k = bkey_i_to_s_c(j_k);
+ }
+
+ u = *k.k;
+ u.needs_whiteout = i->old_k.needs_whiteout;
+
+ BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
+ BUG_ON(i->old_v != k.v);
+#endif
+}
+
+static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
+{
+ return i->path->l + i->level;
+}
+
+static inline bool same_leaf_as_prev(struct btree_trans *trans,
+ struct btree_insert_entry *i)
+{
+ return i != trans->updates &&
+ insert_l(&i[0])->b == insert_l(&i[-1])->b;
+}
+
+static inline bool same_leaf_as_next(struct btree_trans *trans,
+ struct btree_insert_entry *i)
+{
+ return i + 1 < trans->updates + trans->nr_updates &&
+ insert_l(&i[0])->b == insert_l(&i[1])->b;
+}
+
+inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b)
+{
+ struct bch_fs *c = trans->c;
+
+ if (unlikely(btree_node_just_written(b)) &&
+ bch2_btree_post_write_cleanup(c, b))
+ bch2_trans_node_reinit_iter(trans, b);
+
+ /*
+ * If the last bset has been written, or if it's gotten too big - start
+ * a new bset to insert into:
+ */
+ if (want_new_bset(c, b))
+ bch2_btree_init_next(trans, b);
+}
+
+static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
+{
+ while (--i >= trans->updates) {
+ if (same_leaf_as_prev(trans, i))
+ continue;
+
+ bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
+ }
+
+ trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
+ return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
+}
+
+static inline int bch2_trans_lock_write(struct btree_trans *trans)
+{
+ struct btree_insert_entry *i;
+
+ EBUG_ON(trans->write_locked);
+
+ trans_for_each_update(trans, i) {
+ if (same_leaf_as_prev(trans, i))
+ continue;
+
+ if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
+ return trans_lock_write_fail(trans, i);
+
+ if (!i->cached)
+ bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
+ }
+
+ trans->write_locked = true;
+ return 0;
+}
+
+static inline void bch2_trans_unlock_write(struct btree_trans *trans)
+{
+ if (likely(trans->write_locked)) {
+ struct btree_insert_entry *i;
+
+ trans_for_each_update(trans, i)
+ if (!same_leaf_as_prev(trans, i))
+ bch2_btree_node_unlock_write_inlined(trans, i->path,
+ insert_l(i)->b);
+ trans->write_locked = false;
+ }
+}
+
+/* Inserting into a given leaf node (last stage of insert): */
+
+/* Handle overwrites and do insert, for non extents: */
+bool bch2_btree_bset_insert_key(struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b,
+ struct btree_node_iter *node_iter,
+ struct bkey_i *insert)
+{
+ struct bkey_packed *k;
+ unsigned clobber_u64s = 0, new_u64s = 0;
+
+ EBUG_ON(btree_node_just_written(b));
+ EBUG_ON(bset_written(b, btree_bset_last(b)));
+ EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
+ EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
+ EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
+ EBUG_ON(insert->k.u64s >
+ bch_btree_keys_u64s_remaining(trans->c, b));
+ EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
+
+ k = bch2_btree_node_iter_peek_all(node_iter, b);
+ if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
+ k = NULL;
+
+ /* @k is the key being overwritten/deleted, if any: */
+ EBUG_ON(k && bkey_deleted(k));
+
+ /* Deleting, but not found? nothing to do: */
+ if (bkey_deleted(&insert->k) && !k)
+ return false;
+
+ if (bkey_deleted(&insert->k)) {
+ /* Deleting: */
+ btree_account_key_drop(b, k);
+ k->type = KEY_TYPE_deleted;
+
+ if (k->needs_whiteout)
+ push_whiteout(trans->c, b, insert->k.p);
+ k->needs_whiteout = false;
+
+ if (k >= btree_bset_last(b)->start) {
+ clobber_u64s = k->u64s;
+ bch2_bset_delete(b, k, clobber_u64s);
+ goto fix_iter;
+ } else {
+ bch2_btree_path_fix_key_modified(trans, b, k);
+ }
+
+ return true;
+ }
+
+ if (k) {
+ /* Overwriting: */
+ btree_account_key_drop(b, k);
+ k->type = KEY_TYPE_deleted;
+
+ insert->k.needs_whiteout = k->needs_whiteout;
+ k->needs_whiteout = false;
+
+ if (k >= btree_bset_last(b)->start) {
+ clobber_u64s = k->u64s;
+ goto overwrite;
+ } else {
+ bch2_btree_path_fix_key_modified(trans, b, k);
+ }
+ }
+
+ k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
+overwrite:
+ bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
+ new_u64s = k->u64s;
+fix_iter:
+ if (clobber_u64s != new_u64s)
+ bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
+ clobber_u64s, new_u64s);
+ return true;
+}
+
+static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
+ unsigned i, u64 seq)
+{
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
+ struct btree_write *w = container_of(pin, struct btree_write, journal);
+ struct btree *b = container_of(w, struct btree, writes[i]);
+ struct btree_trans *trans = bch2_trans_get(c);
+ unsigned long old, new, v;
+ unsigned idx = w - b->writes;
+
+ btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
+ v = READ_ONCE(b->flags);
+
+ do {
+ old = new = v;
+
+ if (!(old & (1 << BTREE_NODE_dirty)) ||
+ !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
+ w->journal.seq != seq)
+ break;
+
+ new &= ~BTREE_WRITE_TYPE_MASK;
+ new |= BTREE_WRITE_journal_reclaim;
+ new |= 1 << BTREE_NODE_need_write;
+ } while ((v = cmpxchg(&b->flags, old, new)) != old);
+
+ btree_node_write_if_need(c, b, SIX_LOCK_read);
+ six_unlock_read(&b->c.lock);
+
+ bch2_trans_put(trans);
+ return 0;
+}
+
+int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
+{
+ return __btree_node_flush(j, pin, 0, seq);
+}
+
+int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
+{
+ return __btree_node_flush(j, pin, 1, seq);
+}
+
+inline void bch2_btree_add_journal_pin(struct bch_fs *c,
+ struct btree *b, u64 seq)
+{
+ struct btree_write *w = btree_current_write(b);
+
+ bch2_journal_pin_add(&c->journal, seq, &w->journal,
+ btree_node_write_idx(b) == 0
+ ? bch2_btree_node_flush0
+ : bch2_btree_node_flush1);
+}
+
+/**
+ * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
+ * @trans: btree transaction object
+ * @path: path pointing to @insert's pos
+ * @insert: key to insert
+ * @journal_seq: sequence number of journal reservation
+ */
+inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
+ struct btree_path *path,
+ struct bkey_i *insert,
+ u64 journal_seq)
+{
+ struct bch_fs *c = trans->c;
+ struct btree *b = path_l(path)->b;
+ struct bset_tree *t = bset_tree_last(b);
+ struct bset *i = bset(b, t);
+ int old_u64s = bset_u64s(t);
+ int old_live_u64s = b->nr.live_u64s;
+ int live_u64s_added, u64s_added;
+
+ if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
+ &path_l(path)->iter, insert)))
+ return;
+
+ i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
+
+ bch2_btree_add_journal_pin(c, b, journal_seq);
+
+ if (unlikely(!btree_node_dirty(b))) {
+ EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
+ set_btree_node_dirty_acct(c, b);
+ }
+
+ live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
+ u64s_added = (int) bset_u64s(t) - old_u64s;
+
+ if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
+ b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
+ if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
+ b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
+
+ if (u64s_added > live_u64s_added &&
+ bch2_maybe_compact_whiteouts(c, b))
+ bch2_trans_node_reinit_iter(trans, b);
+}
+
+/* Cached btree updates: */
+
+/* Normal update interface: */
+
+static inline void btree_insert_entry_checks(struct btree_trans *trans,
+ struct btree_insert_entry *i)
+{
+ BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
+ BUG_ON(i->cached != i->path->cached);
+ BUG_ON(i->level != i->path->level);
+ BUG_ON(i->btree_id != i->path->btree_id);
+ EBUG_ON(!i->level &&
+ btree_type_has_snapshots(i->btree_id) &&
+ !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
+ test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
+ i->k->k.p.snapshot &&
+ bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
+}
+
+static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
+ unsigned flags)
+{
+ return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
+ trans->journal_u64s, flags);
+}
+
+#define JSET_ENTRY_LOG_U64s 4
+
+static noinline void journal_transaction_name(struct btree_trans *trans)
+{
+ struct bch_fs *c = trans->c;
+ struct journal *j = &c->journal;
+ struct jset_entry *entry =
+ bch2_journal_add_entry(j, &trans->journal_res,
+ BCH_JSET_ENTRY_log, 0, 0,
+ JSET_ENTRY_LOG_U64s);
+ struct jset_entry_log *l =
+ container_of(entry, struct jset_entry_log, entry);
+
+ strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
+}
+
+static inline int btree_key_can_insert(struct btree_trans *trans,
+ struct btree *b, unsigned u64s)
+{
+ struct bch_fs *c = trans->c;
+
+ if (!bch2_btree_node_insert_fits(c, b, u64s))
+ return -BCH_ERR_btree_insert_btree_node_full;
+
+ return 0;
+}
+
+noinline static int
+btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
+ struct btree_path *path, unsigned new_u64s)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ struct bkey_cached *ck = (void *) path->l[0].b;
+ struct bkey_i *new_k;
+ int ret;
+
+ bch2_trans_unlock_write(trans);
+ bch2_trans_unlock(trans);
+
+ new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
+ if (!new_k) {
+ bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
+ bch2_btree_id_str(path->btree_id), new_u64s);
+ return -BCH_ERR_ENOMEM_btree_key_cache_insert;
+ }
+
+ ret = bch2_trans_relock(trans) ?:
+ bch2_trans_lock_write(trans);
+ if (unlikely(ret)) {
+ kfree(new_k);
+ return ret;
+ }
+
+ memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
+
+ trans_for_each_update(trans, i)
+ if (i->old_v == &ck->k->v)
+ i->old_v = &new_k->v;
+
+ kfree(ck->k);
+ ck->u64s = new_u64s;
+ ck->k = new_k;
+ return 0;
+}
+
+static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
+ struct btree_path *path, unsigned u64s)
+{
+ struct bch_fs *c = trans->c;
+ struct bkey_cached *ck = (void *) path->l[0].b;
+ struct btree_insert_entry *i;
+ unsigned new_u64s;
+ struct bkey_i *new_k;
+
+ EBUG_ON(path->level);
+
+ if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+ bch2_btree_key_cache_must_wait(c) &&
+ !(flags & BCH_TRANS_COMMIT_journal_reclaim))
+ return -BCH_ERR_btree_insert_need_journal_reclaim;
+
+ /*
+ * bch2_varint_decode can read past the end of the buffer by at most 7
+ * bytes (it won't be used):
+ */
+ u64s += 1;
+
+ if (u64s <= ck->u64s)
+ return 0;
+
+ new_u64s = roundup_pow_of_two(u64s);
+ new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT);
+ if (unlikely(!new_k))
+ return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
+
+ trans_for_each_update(trans, i)
+ if (i->old_v == &ck->k->v)
+ i->old_v = &new_k->v;
+
+ ck->u64s = new_u64s;
+ ck->k = new_k;
+ return 0;
+}
+
+/* Triggers: */
+
+static int run_one_mem_trigger(struct btree_trans *trans,
+ struct btree_insert_entry *i,
+ unsigned flags)
+{
+ struct bkey_s_c old = { &i->old_k, i->old_v };
+ struct bkey_i *new = i->k;
+ const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
+ const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
+ int ret;
+
+ verify_update_old_key(trans, i);
+
+ if (unlikely(flags & BTREE_TRIGGER_NORUN))
+ return 0;
+
+ if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
+ return 0;
+
+ if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
+ ret = bch2_mark_key(trans, i->btree_id, i->level,
+ old, bkey_i_to_s_c(new),
+ BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
+ } else {
+ struct bkey _deleted = KEY(0, 0, 0);
+ struct bkey_s_c deleted = (struct bkey_s_c) { &_deleted, NULL };
+
+ _deleted.p = i->path->pos;
+
+ ret = bch2_mark_key(trans, i->btree_id, i->level,
+ deleted, bkey_i_to_s_c(new),
+ BTREE_TRIGGER_INSERT|flags) ?:
+ bch2_mark_key(trans, i->btree_id, i->level,
+ old, deleted,
+ BTREE_TRIGGER_OVERWRITE|flags);
+ }
+
+ return ret;
+}
+
+static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
+ bool overwrite)
+{
+ /*
+ * Transactional triggers create new btree_insert_entries, so we can't
+ * pass them a pointer to a btree_insert_entry, that memory is going to
+ * move:
+ */
+ struct bkey old_k = i->old_k;
+ struct bkey_s_c old = { &old_k, i->old_v };
+ const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
+ const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
+
+ verify_update_old_key(trans, i);
+
+ if ((i->flags & BTREE_TRIGGER_NORUN) ||
+ !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
+ return 0;
+
+ if (!i->insert_trigger_run &&
+ !i->overwrite_trigger_run &&
+ old_ops->trans_trigger == new_ops->trans_trigger) {
+ i->overwrite_trigger_run = true;
+ i->insert_trigger_run = true;
+ return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
+ BTREE_TRIGGER_INSERT|
+ BTREE_TRIGGER_OVERWRITE|
+ i->flags) ?: 1;
+ } else if (overwrite && !i->overwrite_trigger_run) {
+ i->overwrite_trigger_run = true;
+ return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
+ } else if (!overwrite && !i->insert_trigger_run) {
+ i->insert_trigger_run = true;
+ return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
+ } else {
+ return 0;
+ }
+}
+
+static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
+ struct btree_insert_entry *btree_id_start)
+{
+ struct btree_insert_entry *i;
+ bool trans_trigger_run;
+ int ret, overwrite;
+
+ for (overwrite = 1; overwrite >= 0; --overwrite) {
+
+ /*
+ * Running triggers will append more updates to the list of updates as
+ * we're walking it:
+ */
+ do {
+ trans_trigger_run = false;
+
+ for (i = btree_id_start;
+ i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
+ i++) {
+ if (i->btree_id != btree_id)
+ continue;
+
+ ret = run_one_trans_trigger(trans, i, overwrite);
+ if (ret < 0)
+ return ret;
+ if (ret)
+ trans_trigger_run = true;
+ }
+ } while (trans_trigger_run);
+ }
+
+ return 0;
+}
+
+static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
+{
+ struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
+ unsigned btree_id = 0;
+ int ret = 0;
+
+ /*
+ *
+ * For a given btree, this algorithm runs insert triggers before
+ * overwrite triggers: this is so that when extents are being moved
+ * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
+ * they are re-added.
+ */
+ for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
+ if (btree_id == BTREE_ID_alloc)
+ continue;
+
+ while (btree_id_start < trans->updates + trans->nr_updates &&
+ btree_id_start->btree_id < btree_id)
+ btree_id_start++;
+
+ ret = run_btree_triggers(trans, btree_id, btree_id_start);
+ if (ret)
+ return ret;
+ }
+
+ trans_for_each_update(trans, i) {
+ if (i->btree_id > BTREE_ID_alloc)
+ break;
+ if (i->btree_id == BTREE_ID_alloc) {
+ ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
+ if (ret)
+ return ret;
+ break;
+ }
+ }
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trans_for_each_update(trans, i)
+ BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
+ (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
+ (!i->insert_trigger_run || !i->overwrite_trigger_run));
+#endif
+ return 0;
+}
+
+static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ int ret = 0;
+
+ trans_for_each_update(trans, i) {
+ /*
+ * XXX: synchronization of cached update triggers with gc
+ * XXX: synchronization of interior node updates with gc
+ */
+ BUG_ON(i->cached || i->level);
+
+ if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
+ ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
+ if (ret)
+ break;
+ }
+ }
+
+ return ret;
+}
+
+static inline int
+bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
+ struct btree_insert_entry **stopped_at,
+ unsigned long trace_ip)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ struct btree_write_buffered_key *wb;
+ struct btree_trans_commit_hook *h;
+ unsigned u64s = 0;
+ int ret;
+
+ if (race_fault()) {
+ trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
+ return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
+ }
+
+ /*
+ * Check if the insert will fit in the leaf node with the write lock
+ * held, otherwise another thread could write the node changing the
+ * amount of space available:
+ */
+
+ prefetch(&trans->c->journal.flags);
+
+ trans_for_each_update(trans, i) {
+ /* Multiple inserts might go to same leaf: */
+ if (!same_leaf_as_prev(trans, i))
+ u64s = 0;
+
+ u64s += i->k->k.u64s;
+ ret = !i->cached
+ ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
+ : btree_key_can_insert_cached(trans, flags, i->path, u64s);
+ if (ret) {
+ *stopped_at = i;
+ return ret;
+ }
+
+ i->k->k.needs_whiteout = false;
+ }
+
+ if (trans->nr_wb_updates &&
+ trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
+ return -BCH_ERR_btree_insert_need_flush_buffer;
+
+ /*
+ * Don't get journal reservation until after we know insert will
+ * succeed:
+ */
+ if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
+ ret = bch2_trans_journal_res_get(trans,
+ (flags & BCH_WATERMARK_MASK)|
+ JOURNAL_RES_GET_NONBLOCK);
+ if (ret)
+ return ret;
+
+ if (unlikely(trans->journal_transaction_names))
+ journal_transaction_name(trans);
+ }
+
+ /*
+ * Not allowed to fail after we've gotten our journal reservation - we
+ * have to use it:
+ */
+
+ if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
+ !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
+ if (bch2_journal_seq_verify)
+ trans_for_each_update(trans, i)
+ i->k->k.version.lo = trans->journal_res.seq;
+ else if (bch2_inject_invalid_keys)
+ trans_for_each_update(trans, i)
+ i->k->k.version = MAX_VERSION;
+ }
+
+ if (trans->fs_usage_deltas &&
+ bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
+ return -BCH_ERR_btree_insert_need_mark_replicas;
+
+ if (trans->nr_wb_updates) {
+ EBUG_ON(flags & BCH_TRANS_COMMIT_no_journal_res);
+
+ ret = bch2_btree_insert_keys_write_buffer(trans);
+ if (ret)
+ goto revert_fs_usage;
+ }
+
+ h = trans->hooks;
+ while (h) {
+ ret = h->fn(trans, h);
+ if (ret)
+ goto revert_fs_usage;
+ h = h->next;
+ }
+
+ trans_for_each_update(trans, i)
+ if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
+ ret = run_one_mem_trigger(trans, i, i->flags);
+ if (ret)
+ goto fatal_err;
+ }
+
+ if (unlikely(c->gc_pos.phase)) {
+ ret = bch2_trans_commit_run_gc_triggers(trans);
+ if (ret)
+ goto fatal_err;
+ }
+
+ if (unlikely(trans->extra_journal_entries.nr)) {
+ memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
+ trans->extra_journal_entries.data,
+ trans->extra_journal_entries.nr);
+
+ trans->journal_res.offset += trans->extra_journal_entries.nr;
+ trans->journal_res.u64s -= trans->extra_journal_entries.nr;
+ }
+
+ if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
+ struct journal *j = &c->journal;
+ struct jset_entry *entry;
+
+ trans_for_each_update(trans, i) {
+ if (i->key_cache_already_flushed)
+ continue;
+
+ if (i->flags & BTREE_UPDATE_NOJOURNAL)
+ continue;
+
+ verify_update_old_key(trans, i);
+
+ if (trans->journal_transaction_names) {
+ entry = bch2_journal_add_entry(j, &trans->journal_res,
+ BCH_JSET_ENTRY_overwrite,
+ i->btree_id, i->level,
+ i->old_k.u64s);
+ bkey_reassemble((struct bkey_i *) entry->start,
+ (struct bkey_s_c) { &i->old_k, i->old_v });
+ }
+
+ entry = bch2_journal_add_entry(j, &trans->journal_res,
+ BCH_JSET_ENTRY_btree_keys,
+ i->btree_id, i->level,
+ i->k->k.u64s);
+ bkey_copy((struct bkey_i *) entry->start, i->k);
+ }
+
+ trans_for_each_wb_update(trans, wb) {
+ entry = bch2_journal_add_entry(j, &trans->journal_res,
+ BCH_JSET_ENTRY_btree_keys,
+ wb->btree, 0,
+ wb->k.k.u64s);
+ bkey_copy((struct bkey_i *) entry->start, &wb->k);
+ }
+
+ if (trans->journal_seq)
+ *trans->journal_seq = trans->journal_res.seq;
+ }
+
+ trans_for_each_update(trans, i) {
+ if (!i->cached) {
+ bch2_btree_insert_key_leaf(trans, i->path, i->k, trans->journal_res.seq);
+ } else if (!i->key_cache_already_flushed)
+ bch2_btree_insert_key_cached(trans, flags, i);
+ else {
+ bch2_btree_key_cache_drop(trans, i->path);
+ btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
+ }
+ }
+
+ return 0;
+fatal_err:
+ bch2_fatal_error(c);
+revert_fs_usage:
+ if (trans->fs_usage_deltas)
+ bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
+ return ret;
+}
+
+static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
+{
+ struct btree_insert_entry *i;
+ struct btree_write_buffered_key *wb;
+
+ trans_for_each_update(trans, i)
+ bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
+
+ trans_for_each_wb_update(trans, wb)
+ bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
+}
+
+static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
+ enum bkey_invalid_flags flags,
+ struct btree_insert_entry *i,
+ struct printbuf *err)
+{
+ struct bch_fs *c = trans->c;
+
+ printbuf_reset(err);
+ prt_printf(err, "invalid bkey on insert from %s -> %ps",
+ trans->fn, (void *) i->ip_allocated);
+ prt_newline(err);
+ printbuf_indent_add(err, 2);
+
+ bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
+ prt_newline(err);
+
+ bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
+ bch2_print_string_as_lines(KERN_ERR, err->buf);
+
+ bch2_inconsistent_error(c);
+ bch2_dump_trans_updates(trans);
+
+ return -EINVAL;
+}
+
+static int bch2_trans_commit_journal_pin_flush(struct journal *j,
+ struct journal_entry_pin *_pin, u64 seq)
+{
+ return 0;
+}
+
+/*
+ * Get journal reservation, take write locks, and attempt to do btree update(s):
+ */
+static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
+ struct btree_insert_entry **stopped_at,
+ unsigned long trace_ip)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ int ret = 0, u64s_delta = 0;
+
+ trans_for_each_update(trans, i) {
+ if (i->cached)
+ continue;
+
+ u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
+ u64s_delta -= i->old_btree_u64s;
+
+ if (!same_leaf_as_next(trans, i)) {
+ if (u64s_delta <= 0) {
+ ret = bch2_foreground_maybe_merge(trans, i->path,
+ i->level, flags);
+ if (unlikely(ret))
+ return ret;
+ }
+
+ u64s_delta = 0;
+ }
+ }
+
+ ret = bch2_trans_lock_write(trans);
+ if (unlikely(ret))
+ return ret;
+
+ ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
+
+ if (!ret && unlikely(trans->journal_replay_not_finished))
+ bch2_drop_overwrites_from_journal(trans);
+
+ bch2_trans_unlock_write(trans);
+
+ if (!ret && trans->journal_pin)
+ bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
+ trans->journal_pin,
+ bch2_trans_commit_journal_pin_flush);
+
+ /*
+ * Drop journal reservation after dropping write locks, since dropping
+ * the journal reservation may kick off a journal write:
+ */
+ if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
+ bch2_journal_res_put(&c->journal, &trans->journal_res);
+
+ return ret;
+}
+
+static int journal_reclaim_wait_done(struct bch_fs *c)
+{
+ int ret = bch2_journal_error(&c->journal) ?:
+ !bch2_btree_key_cache_must_wait(c);
+
+ if (!ret)
+ journal_reclaim_kick(&c->journal);
+ return ret;
+}
+
+static noinline
+int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
+ struct btree_insert_entry *i,
+ int ret, unsigned long trace_ip)
+{
+ struct bch_fs *c = trans->c;
+
+ switch (ret) {
+ case -BCH_ERR_btree_insert_btree_node_full:
+ ret = bch2_btree_split_leaf(trans, i->path, flags);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
+ break;
+ case -BCH_ERR_btree_insert_need_mark_replicas:
+ ret = drop_locks_do(trans,
+ bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
+ break;
+ case -BCH_ERR_journal_res_get_blocked:
+ /*
+ * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
+ * flag
+ */
+ if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
+ (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
+ ret = -BCH_ERR_journal_reclaim_would_deadlock;
+ break;
+ }
+
+ ret = drop_locks_do(trans,
+ bch2_trans_journal_res_get(trans,
+ (flags & BCH_WATERMARK_MASK)|
+ JOURNAL_RES_GET_CHECK));
+ break;
+ case -BCH_ERR_btree_insert_need_journal_reclaim:
+ bch2_trans_unlock(trans);
+
+ trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
+
+ wait_event_freezable(c->journal.reclaim_wait,
+ (ret = journal_reclaim_wait_done(c)));
+ if (ret < 0)
+ break;
+
+ ret = bch2_trans_relock(trans);
+ break;
+ case -BCH_ERR_btree_insert_need_flush_buffer: {
+ struct btree_write_buffer *wb = &c->btree_write_buffer;
+
+ ret = 0;
+
+ if (wb->state.nr > wb->size * 3 / 4) {
+ bch2_trans_unlock(trans);
+ mutex_lock(&wb->flush_lock);
+
+ if (wb->state.nr > wb->size * 3 / 4) {
+ bch2_trans_begin(trans);
+ ret = __bch2_btree_write_buffer_flush(trans,
+ flags|BCH_TRANS_COMMIT_no_check_rw, true);
+ if (!ret) {
+ trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
+ ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
+ }
+ } else {
+ mutex_unlock(&wb->flush_lock);
+ ret = bch2_trans_relock(trans);
+ }
+ }
+ break;
+ }
+ default:
+ BUG_ON(ret >= 0);
+ break;
+ }
+
+ BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
+
+ bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
+ (flags & BCH_TRANS_COMMIT_no_enospc), c,
+ "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
+
+ return ret;
+}
+
+static noinline int
+bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
+{
+ struct bch_fs *c = trans->c;
+ int ret;
+
+ if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
+ test_bit(BCH_FS_STARTED, &c->flags))
+ return -BCH_ERR_erofs_trans_commit;
+
+ ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
+ if (ret)
+ return ret;
+
+ bch2_write_ref_get(c, BCH_WRITE_REF_trans);
+ return 0;
+}
+
+/*
+ * This is for updates done in the early part of fsck - btree_gc - before we've
+ * gone RW. we only add the new key to the list of keys for journal replay to
+ * do.
+ */
+static noinline int
+do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ int ret = 0;
+
+ trans_for_each_update(trans, i) {
+ ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
+ if (ret)
+ break;
+ }
+
+ return ret;
+}
+
+int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i = NULL;
+ struct btree_write_buffered_key *wb;
+ int ret = 0;
+
+ if (!trans->nr_updates &&
+ !trans->nr_wb_updates &&
+ !trans->extra_journal_entries.nr)
+ goto out_reset;
+
+ ret = bch2_trans_commit_run_triggers(trans);
+ if (ret)
+ goto out_reset;
+
+ trans_for_each_update(trans, i) {
+ struct printbuf buf = PRINTBUF;
+ enum bkey_invalid_flags invalid_flags = 0;
+
+ if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
+ invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
+
+ if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
+ i->bkey_type, invalid_flags, &buf)))
+ ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
+ btree_insert_entry_checks(trans, i);
+ printbuf_exit(&buf);
+
+ if (ret)
+ return ret;
+ }
+
+ if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
+ ret = do_bch2_trans_commit_to_journal_replay(trans);
+ goto out_reset;
+ }
+
+ if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
+ unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
+ ret = bch2_trans_commit_get_rw_cold(trans, flags);
+ if (ret)
+ goto out_reset;
+ }
+
+ if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
+ mutex_trylock(&c->btree_write_buffer.flush_lock)) {
+ bch2_trans_begin(trans);
+ bch2_trans_unlock(trans);
+
+ ret = __bch2_btree_write_buffer_flush(trans,
+ flags|BCH_TRANS_COMMIT_no_check_rw, true);
+ if (!ret) {
+ trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
+ ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
+ }
+ goto out;
+ }
+
+ EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
+
+ trans->journal_u64s = trans->extra_journal_entries.nr;
+ trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
+ if (trans->journal_transaction_names)
+ trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
+
+ trans_for_each_update(trans, i) {
+ EBUG_ON(!i->path->should_be_locked);
+
+ ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
+ if (unlikely(ret))
+ goto out;
+
+ EBUG_ON(!btree_node_intent_locked(i->path, i->level));
+
+ if (i->key_cache_already_flushed)
+ continue;
+
+ if (i->flags & BTREE_UPDATE_NOJOURNAL)
+ continue;
+
+ /* we're going to journal the key being updated: */
+ trans->journal_u64s += jset_u64s(i->k->k.u64s);
+
+ /* and we're also going to log the overwrite: */
+ if (trans->journal_transaction_names)
+ trans->journal_u64s += jset_u64s(i->old_k.u64s);
+ }
+
+ trans_for_each_wb_update(trans, wb)
+ trans->journal_u64s += jset_u64s(wb->k.k.u64s);
+
+ if (trans->extra_journal_res) {
+ ret = bch2_disk_reservation_add(c, trans->disk_res,
+ trans->extra_journal_res,
+ (flags & BCH_TRANS_COMMIT_no_enospc)
+ ? BCH_DISK_RESERVATION_NOFAIL : 0);
+ if (ret)
+ goto err;
+ }
+retry:
+ bch2_trans_verify_not_in_restart(trans);
+ if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
+ memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+
+ ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
+
+ /* make sure we didn't drop or screw up locks: */
+ bch2_trans_verify_locks(trans);
+
+ if (ret)
+ goto err;
+
+ trace_and_count(c, transaction_commit, trans, _RET_IP_);
+out:
+ if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
+ bch2_write_ref_put(c, BCH_WRITE_REF_trans);
+out_reset:
+ if (!ret)
+ bch2_trans_downgrade(trans);
+ bch2_trans_reset_updates(trans);
+
+ return ret;
+err:
+ ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
+ if (ret)
+ goto out;
+
+ /*
+ * We might have done another transaction commit in the error path -
+ * i.e. btree write buffer flush - which will have made use of
+ * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
+ * how the journal sequence number to pin is passed in - so we must
+ * restart:
+ */
+ if (flags & BCH_TRANS_COMMIT_no_journal_res) {
+ ret = -BCH_ERR_transaction_restart_nested;
+ goto out;
+ }
+
+ goto retry;
+}
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_update.h"
+#include "btree_iter.h"
+#include "btree_journal_iter.h"
+#include "btree_locking.h"
+#include "buckets.h"
+#include "debug.h"
+#include "errcode.h"
+#include "error.h"
+#include "extents.h"
+#include "keylist.h"
+#include "snapshot.h"
+#include "trace.h"
+
+static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
+ const struct btree_insert_entry *r)
+{
+ return cmp_int(l->btree_id, r->btree_id) ?:
+ cmp_int(l->cached, r->cached) ?:
+ -cmp_int(l->level, r->level) ?:
+ bpos_cmp(l->k->k.p, r->k->k.p);
+}
+
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
+ struct bkey_i *, enum btree_update_flags,
+ unsigned long ip);
+
+static noinline int extent_front_merge(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k,
+ struct bkey_i **insert,
+ enum btree_update_flags flags)
+{
+ struct bch_fs *c = trans->c;
+ struct bkey_i *update;
+ int ret;
+
+ update = bch2_bkey_make_mut_noupdate(trans, k);
+ ret = PTR_ERR_OR_ZERO(update);
+ if (ret)
+ return ret;
+
+ if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
+ return 0;
+
+ ret = bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p) ?:
+ bch2_key_has_snapshot_overwrites(trans, iter->btree_id, (*insert)->k.p);
+ if (ret < 0)
+ return ret;
+ if (ret)
+ return 0;
+
+ ret = bch2_btree_delete_at(trans, iter, flags);
+ if (ret)
+ return ret;
+
+ *insert = update;
+ return 0;
+}
+
+static noinline int extent_back_merge(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_i *insert,
+ struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ int ret;
+
+ ret = bch2_key_has_snapshot_overwrites(trans, iter->btree_id, insert->k.p) ?:
+ bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p);
+ if (ret < 0)
+ return ret;
+ if (ret)
+ return 0;
+
+ bch2_bkey_merge(c, bkey_i_to_s(insert), k);
+ return 0;
+}
+
+/*
+ * When deleting, check if we need to emit a whiteout (because we're overwriting
+ * something in an ancestor snapshot)
+ */
+static int need_whiteout_for_snapshot(struct btree_trans *trans,
+ enum btree_id btree_id, struct bpos pos)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ u32 snapshot = pos.snapshot;
+ int ret;
+
+ if (!bch2_snapshot_parent(trans->c, pos.snapshot))
+ return 0;
+
+ pos.snapshot++;
+
+ for_each_btree_key_norestart(trans, iter, btree_id, pos,
+ BTREE_ITER_ALL_SNAPSHOTS|
+ BTREE_ITER_NOPRESERVE, k, ret) {
+ if (!bkey_eq(k.k->p, pos))
+ break;
+
+ if (bch2_snapshot_is_ancestor(trans->c, snapshot,
+ k.k->p.snapshot)) {
+ ret = !bkey_whiteout(k.k);
+ break;
+ }
+ }
+ bch2_trans_iter_exit(trans, &iter);
+
+ return ret;
+}
+
+int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
+ enum btree_id id,
+ struct bpos old_pos,
+ struct bpos new_pos)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter old_iter, new_iter = { NULL };
+ struct bkey_s_c old_k, new_k;
+ snapshot_id_list s;
+ struct bkey_i *update;
+ int ret = 0;
+
+ if (!bch2_snapshot_has_children(c, old_pos.snapshot))
+ return 0;
+
+ darray_init(&s);
+
+ bch2_trans_iter_init(trans, &old_iter, id, old_pos,
+ BTREE_ITER_NOT_EXTENTS|
+ BTREE_ITER_ALL_SNAPSHOTS);
+ while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
+ !(ret = bkey_err(old_k)) &&
+ bkey_eq(old_pos, old_k.k->p)) {
+ struct bpos whiteout_pos =
+ SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
+
+ if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
+ snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
+ continue;
+
+ new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
+ BTREE_ITER_NOT_EXTENTS|
+ BTREE_ITER_INTENT);
+ ret = bkey_err(new_k);
+ if (ret)
+ break;
+
+ if (new_k.k->type == KEY_TYPE_deleted) {
+ update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
+ ret = PTR_ERR_OR_ZERO(update);
+ if (ret)
+ break;
+
+ bkey_init(&update->k);
+ update->k.p = whiteout_pos;
+ update->k.type = KEY_TYPE_whiteout;
+
+ ret = bch2_trans_update(trans, &new_iter, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+ }
+ bch2_trans_iter_exit(trans, &new_iter);
+
+ ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
+ if (ret)
+ break;
+ }
+ bch2_trans_iter_exit(trans, &new_iter);
+ bch2_trans_iter_exit(trans, &old_iter);
+ darray_exit(&s);
+
+ return ret;
+}
+
+int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
+ struct btree_iter *iter,
+ enum btree_update_flags flags,
+ struct bkey_s_c old,
+ struct bkey_s_c new)
+{
+ enum btree_id btree_id = iter->btree_id;
+ struct bkey_i *update;
+ struct bpos new_start = bkey_start_pos(new.k);
+ bool front_split = bkey_lt(bkey_start_pos(old.k), new_start);
+ bool back_split = bkey_gt(old.k->p, new.k->p);
+ int ret = 0, compressed_sectors;
+
+ /*
+ * If we're going to be splitting a compressed extent, note it
+ * so that __bch2_trans_commit() can increase our disk
+ * reservation:
+ */
+ if (((front_split && back_split) ||
+ ((front_split || back_split) && old.k->p.snapshot != new.k->p.snapshot)) &&
+ (compressed_sectors = bch2_bkey_sectors_compressed(old)))
+ trans->extra_journal_res += compressed_sectors;
+
+ if (front_split) {
+ update = bch2_bkey_make_mut_noupdate(trans, old);
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ return ret;
+
+ bch2_cut_back(new_start, update);
+
+ ret = bch2_insert_snapshot_whiteouts(trans, btree_id,
+ old.k->p, update->k.p) ?:
+ bch2_btree_insert_nonextent(trans, btree_id, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+ if (ret)
+ return ret;
+ }
+
+ /* If we're overwriting in a different snapshot - middle split: */
+ if (old.k->p.snapshot != new.k->p.snapshot &&
+ (front_split || back_split)) {
+ update = bch2_bkey_make_mut_noupdate(trans, old);
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ return ret;
+
+ bch2_cut_front(new_start, update);
+ bch2_cut_back(new.k->p, update);
+
+ ret = bch2_insert_snapshot_whiteouts(trans, btree_id,
+ old.k->p, update->k.p) ?:
+ bch2_btree_insert_nonextent(trans, btree_id, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+ if (ret)
+ return ret;
+ }
+
+ if (bkey_le(old.k->p, new.k->p)) {
+ update = bch2_trans_kmalloc(trans, sizeof(*update));
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ return ret;
+
+ bkey_init(&update->k);
+ update->k.p = old.k->p;
+ update->k.p.snapshot = new.k->p.snapshot;
+
+ if (new.k->p.snapshot != old.k->p.snapshot) {
+ update->k.type = KEY_TYPE_whiteout;
+ } else if (btree_type_has_snapshots(btree_id)) {
+ ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
+ if (ret < 0)
+ return ret;
+ if (ret)
+ update->k.type = KEY_TYPE_whiteout;
+ }
+
+ ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+ if (ret)
+ return ret;
+ }
+
+ if (back_split) {
+ update = bch2_bkey_make_mut_noupdate(trans, old);
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ return ret;
+
+ bch2_cut_front(new.k->p, update);
+
+ ret = bch2_trans_update_by_path(trans, iter->path, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+ flags, _RET_IP_);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
+static int bch2_trans_update_extent(struct btree_trans *trans,
+ struct btree_iter *orig_iter,
+ struct bkey_i *insert,
+ enum btree_update_flags flags)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ enum btree_id btree_id = orig_iter->btree_id;
+ int ret = 0;
+
+ bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
+ BTREE_ITER_INTENT|
+ BTREE_ITER_WITH_UPDATES|
+ BTREE_ITER_NOT_EXTENTS);
+ k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+ if ((ret = bkey_err(k)))
+ goto err;
+ if (!k.k)
+ goto out;
+
+ if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
+ if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
+ ret = extent_front_merge(trans, &iter, k, &insert, flags);
+ if (ret)
+ goto err;
+ }
+
+ goto next;
+ }
+
+ while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
+ bool done = bkey_lt(insert->k.p, k.k->p);
+
+ ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
+ if (ret)
+ goto err;
+
+ if (done)
+ goto out;
+next:
+ bch2_btree_iter_advance(&iter);
+ k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+ if ((ret = bkey_err(k)))
+ goto err;
+ if (!k.k)
+ goto out;
+ }
+
+ if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
+ ret = extent_back_merge(trans, &iter, insert, k);
+ if (ret)
+ goto err;
+ }
+out:
+ if (!bkey_deleted(&insert->k))
+ ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+
+ return ret;
+}
+
+static noinline int flush_new_cached_update(struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree_insert_entry *i,
+ enum btree_update_flags flags,
+ unsigned long ip)
+{
+ struct btree_path *btree_path;
+ struct bkey k;
+ int ret;
+
+ btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+ BTREE_ITER_INTENT, _THIS_IP_);
+ ret = bch2_btree_path_traverse(trans, btree_path, 0);
+ if (ret)
+ goto out;
+
+ /*
+ * The old key in the insert entry might actually refer to an existing
+ * key in the btree that has been deleted from cache and not yet
+ * flushed. Check for this and skip the flush so we don't run triggers
+ * against a stale key.
+ */
+ bch2_btree_path_peek_slot_exact(btree_path, &k);
+ if (!bkey_deleted(&k))
+ goto out;
+
+ i->key_cache_already_flushed = true;
+ i->flags |= BTREE_TRIGGER_NORUN;
+
+ btree_path_set_should_be_locked(btree_path);
+ ret = bch2_trans_update_by_path(trans, btree_path, i->k, flags, ip);
+out:
+ bch2_path_put(trans, btree_path, true);
+ return ret;
+}
+
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
+ struct bkey_i *k, enum btree_update_flags flags,
+ unsigned long ip)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i, n;
+ int cmp;
+
+ EBUG_ON(!path->should_be_locked);
+ EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
+ EBUG_ON(!bpos_eq(k->k.p, path->pos));
+
+ n = (struct btree_insert_entry) {
+ .flags = flags,
+ .bkey_type = __btree_node_type(path->level, path->btree_id),
+ .btree_id = path->btree_id,
+ .level = path->level,
+ .cached = path->cached,
+ .path = path,
+ .k = k,
+ .ip_allocated = ip,
+ };
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trans_for_each_update(trans, i)
+ BUG_ON(i != trans->updates &&
+ btree_insert_entry_cmp(i - 1, i) >= 0);
+#endif
+
+ /*
+ * Pending updates are kept sorted: first, find position of new update,
+ * then delete/trim any updates the new update overwrites:
+ */
+ trans_for_each_update(trans, i) {
+ cmp = btree_insert_entry_cmp(&n, i);
+ if (cmp <= 0)
+ break;
+ }
+
+ if (!cmp && i < trans->updates + trans->nr_updates) {
+ EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
+
+ bch2_path_put(trans, i->path, true);
+ i->flags = n.flags;
+ i->cached = n.cached;
+ i->k = n.k;
+ i->path = n.path;
+ i->ip_allocated = n.ip_allocated;
+ } else {
+ array_insert_item(trans->updates, trans->nr_updates,
+ i - trans->updates, n);
+
+ i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
+ i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
+
+ if (unlikely(trans->journal_replay_not_finished)) {
+ struct bkey_i *j_k =
+ bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
+
+ if (j_k) {
+ i->old_k = j_k->k;
+ i->old_v = &j_k->v;
+ }
+ }
+ }
+
+ __btree_path_get(i->path, true);
+
+ /*
+ * If a key is present in the key cache, it must also exist in the
+ * btree - this is necessary for cache coherency. When iterating over
+ * a btree that's cached in the key cache, the btree iter code checks
+ * the key cache - but the key has to exist in the btree for that to
+ * work:
+ */
+ if (path->cached && bkey_deleted(&i->old_k))
+ return flush_new_cached_update(trans, path, i, flags, ip);
+
+ return 0;
+}
+
+static noinline int bch2_trans_update_get_key_cache(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct btree_path *path)
+{
+ if (!iter->key_cache_path ||
+ !iter->key_cache_path->should_be_locked ||
+ !bpos_eq(iter->key_cache_path->pos, iter->pos)) {
+ struct bkey_cached *ck;
+ int ret;
+
+ if (!iter->key_cache_path)
+ iter->key_cache_path =
+ bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+ BTREE_ITER_INTENT|
+ BTREE_ITER_CACHED, _THIS_IP_);
+
+ iter->key_cache_path =
+ bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
+ iter->flags & BTREE_ITER_INTENT,
+ _THIS_IP_);
+
+ ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
+ BTREE_ITER_CACHED);
+ if (unlikely(ret))
+ return ret;
+
+ ck = (void *) iter->key_cache_path->l[0].b;
+
+ if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+ trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
+ return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
+ }
+
+ btree_path_set_should_be_locked(iter->key_cache_path);
+ }
+
+ return 0;
+}
+
+int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
+ struct bkey_i *k, enum btree_update_flags flags)
+{
+ struct btree_path *path = iter->update_path ?: iter->path;
+ int ret;
+
+ if (iter->flags & BTREE_ITER_IS_EXTENTS)
+ return bch2_trans_update_extent(trans, iter, k, flags);
+
+ if (bkey_deleted(&k->k) &&
+ !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+ (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
+ ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
+ if (unlikely(ret < 0))
+ return ret;
+
+ if (ret)
+ k->k.type = KEY_TYPE_whiteout;
+ }
+
+ /*
+ * Ensure that updates to cached btrees go to the key cache:
+ */
+ if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+ !path->cached &&
+ !path->level &&
+ btree_id_cached(trans->c, path->btree_id)) {
+ ret = bch2_trans_update_get_key_cache(trans, iter, path);
+ if (ret)
+ return ret;
+
+ path = iter->key_cache_path;
+ }
+
+ return bch2_trans_update_by_path(trans, path, k, flags, _RET_IP_);
+}
+
+int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
+ enum btree_id btree,
+ struct bkey_i *k)
+{
+ struct btree_write_buffered_key *i;
+ int ret;
+
+ EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
+ EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
+
+ trans_for_each_wb_update(trans, i) {
+ if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
+ bkey_copy(&i->k, k);
+ return 0;
+ }
+ }
+
+ if (!trans->wb_updates ||
+ trans->nr_wb_updates == trans->wb_updates_size) {
+ struct btree_write_buffered_key *u;
+
+ if (trans->nr_wb_updates == trans->wb_updates_size) {
+ struct btree_transaction_stats *s = btree_trans_stats(trans);
+
+ BUG_ON(trans->wb_updates_size > U8_MAX / 2);
+ trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
+ if (s)
+ s->wb_updates_size = trans->wb_updates_size;
+ }
+
+ u = bch2_trans_kmalloc_nomemzero(trans,
+ trans->wb_updates_size *
+ sizeof(struct btree_write_buffered_key));
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ return ret;
+
+ if (trans->nr_wb_updates)
+ memcpy(u, trans->wb_updates, trans->nr_wb_updates *
+ sizeof(struct btree_write_buffered_key));
+ trans->wb_updates = u;
+ }
+
+ trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
+ .btree = btree,
+ };
+
+ bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
+ trans->nr_wb_updates++;
+
+ return 0;
+}
+
+int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
+ enum btree_id btree, struct bpos end)
+{
+ struct bkey_s_c k;
+ int ret = 0;
+
+ bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
+ k = bch2_btree_iter_prev(iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ bch2_btree_iter_advance(iter);
+ k = bch2_btree_iter_peek_slot(iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ BUG_ON(k.k->type != KEY_TYPE_deleted);
+
+ if (bkey_gt(k.k->p, end)) {
+ ret = -BCH_ERR_ENOSPC_btree_slot;
+ goto err;
+ }
+
+ return 0;
+err:
+ bch2_trans_iter_exit(trans, iter);
+ return ret;
+}
+
+void bch2_trans_commit_hook(struct btree_trans *trans,
+ struct btree_trans_commit_hook *h)
+{
+ h->next = trans->hooks;
+ trans->hooks = h;
+}
+
+int bch2_btree_insert_nonextent(struct btree_trans *trans,
+ enum btree_id btree, struct bkey_i *k,
+ enum btree_update_flags flags)
+{
+ struct btree_iter iter;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, btree, k->k.p,
+ BTREE_ITER_CACHED|
+ BTREE_ITER_NOT_EXTENTS|
+ BTREE_ITER_INTENT);
+ ret = bch2_btree_iter_traverse(&iter) ?:
+ bch2_trans_update(trans, &iter, k, flags);
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+int bch2_btree_insert_trans(struct btree_trans *trans, enum btree_id id,
+ struct bkey_i *k, enum btree_update_flags flags)
+{
+ struct btree_iter iter;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
+ BTREE_ITER_CACHED|
+ BTREE_ITER_INTENT);
+ ret = bch2_btree_iter_traverse(&iter) ?:
+ bch2_trans_update(trans, &iter, k, flags);
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+/**
+ * bch2_btree_insert - insert keys into the extent btree
+ * @c: pointer to struct bch_fs
+ * @id: btree to insert into
+ * @k: key to insert
+ * @disk_res: must be non-NULL whenever inserting or potentially
+ * splitting data extents
+ * @flags: transaction commit flags
+ *
+ * Returns: 0 on success, error code on failure
+ */
+int bch2_btree_insert(struct bch_fs *c, enum btree_id id, struct bkey_i *k,
+ struct disk_reservation *disk_res, int flags)
+{
+ return bch2_trans_do(c, disk_res, NULL, flags,
+ bch2_btree_insert_trans(trans, id, k, 0));
+}
+
+int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
+ unsigned len, unsigned update_flags)
+{
+ struct bkey_i *k;
+
+ k = bch2_trans_kmalloc(trans, sizeof(*k));
+ if (IS_ERR(k))
+ return PTR_ERR(k);
+
+ bkey_init(&k->k);
+ k->k.p = iter->pos;
+ bch2_key_resize(&k->k, len);
+ return bch2_trans_update(trans, iter, k, update_flags);
+}
+
+int bch2_btree_delete_at(struct btree_trans *trans,
+ struct btree_iter *iter, unsigned update_flags)
+{
+ return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
+}
+
+int bch2_btree_delete_at_buffered(struct btree_trans *trans,
+ enum btree_id btree, struct bpos pos)
+{
+ struct bkey_i *k;
+
+ k = bch2_trans_kmalloc(trans, sizeof(*k));
+ if (IS_ERR(k))
+ return PTR_ERR(k);
+
+ bkey_init(&k->k);
+ k->k.p = pos;
+ return bch2_trans_update_buffered(trans, btree, k);
+}
+
+int bch2_btree_delete(struct btree_trans *trans,
+ enum btree_id btree, struct bpos pos,
+ unsigned update_flags)
+{
+ struct btree_iter iter;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, btree, pos,
+ BTREE_ITER_CACHED|
+ BTREE_ITER_INTENT);
+ ret = bch2_btree_iter_traverse(&iter) ?:
+ bch2_btree_delete_at(trans, &iter, update_flags);
+ bch2_trans_iter_exit(trans, &iter);
+
+ return ret;
+}
+
+int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
+ struct bpos start, struct bpos end,
+ unsigned update_flags,
+ u64 *journal_seq)
+{
+ u32 restart_count = trans->restart_count;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret = 0;
+
+ bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
+ while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
+ struct disk_reservation disk_res =
+ bch2_disk_reservation_init(trans->c, 0);
+ struct bkey_i delete;
+
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ bkey_init(&delete.k);
+
+ /*
+ * This could probably be more efficient for extents:
+ */
+
+ /*
+ * For extents, iter.pos won't necessarily be the same as
+ * bkey_start_pos(k.k) (for non extents they always will be the
+ * same). It's important that we delete starting from iter.pos
+ * because the range we want to delete could start in the middle
+ * of k.
+ *
+ * (bch2_btree_iter_peek() does guarantee that iter.pos >=
+ * bkey_start_pos(k.k)).
+ */
+ delete.k.p = iter.pos;
+
+ if (iter.flags & BTREE_ITER_IS_EXTENTS)
+ bch2_key_resize(&delete.k,
+ bpos_min(end, k.k->p).offset -
+ iter.pos.offset);
+
+ ret = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
+ bch2_trans_commit(trans, &disk_res, journal_seq,
+ BCH_TRANS_COMMIT_no_enospc);
+ bch2_disk_reservation_put(trans->c, &disk_res);
+err:
+ /*
+ * the bch2_trans_begin() call is in a weird place because we
+ * need to call it after every transaction commit, to avoid path
+ * overflow, but don't want to call it if the delete operation
+ * is a no-op and we have no work to do:
+ */
+ bch2_trans_begin(trans);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ ret = 0;
+ if (ret)
+ break;
+ }
+ bch2_trans_iter_exit(trans, &iter);
+
+ return ret ?: trans_was_restarted(trans, restart_count);
+}
+
+/*
+ * bch_btree_delete_range - delete everything within a given range
+ *
+ * Range is a half open interval - [start, end)
+ */
+int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
+ struct bpos start, struct bpos end,
+ unsigned update_flags,
+ u64 *journal_seq)
+{
+ int ret = bch2_trans_run(c,
+ bch2_btree_delete_range_trans(trans, id, start, end,
+ update_flags, journal_seq));
+ if (ret == -BCH_ERR_transaction_restart_nested)
+ ret = 0;
+ return ret;
+}
+
+int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
+ struct bpos pos, bool set)
+{
+ struct bkey_i *k;
+ int ret = 0;
+
+ k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
+ ret = PTR_ERR_OR_ZERO(k);
+ if (unlikely(ret))
+ return ret;
+
+ bkey_init(&k->k);
+ k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
+ k->k.p = pos;
+
+ return bch2_trans_update_buffered(trans, btree, k);
+}
+
+__printf(2, 0)
+static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
+{
+ struct printbuf buf = PRINTBUF;
+ struct jset_entry_log *l;
+ unsigned u64s;
+ int ret;
+
+ prt_vprintf(&buf, fmt, args);
+ ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
+ if (ret)
+ goto err;
+
+ u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
+
+ ret = darray_make_room(entries, jset_u64s(u64s));
+ if (ret)
+ goto err;
+
+ l = (void *) &darray_top(*entries);
+ l->entry.u64s = cpu_to_le16(u64s);
+ l->entry.btree_id = 0;
+ l->entry.level = 1;
+ l->entry.type = BCH_JSET_ENTRY_log;
+ l->entry.pad[0] = 0;
+ l->entry.pad[1] = 0;
+ l->entry.pad[2] = 0;
+ memcpy(l->d, buf.buf, buf.pos);
+ while (buf.pos & 7)
+ l->d[buf.pos++] = '\0';
+
+ entries->nr += jset_u64s(u64s);
+err:
+ printbuf_exit(&buf);
+ return ret;
+}
+
+__printf(3, 0)
+static int
+__bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
+ va_list args)
+{
+ int ret;
+
+ if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
+ ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
+ } else {
+ ret = bch2_trans_do(c, NULL, NULL,
+ BCH_TRANS_COMMIT_lazy_rw|commit_flags,
+ __bch2_trans_log_msg(&trans->extra_journal_entries, fmt, args));
+ }
+
+ return ret;
+}
+
+__printf(2, 3)
+int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
+{
+ va_list args;
+ int ret;
+
+ va_start(args, fmt);
+ ret = __bch2_fs_log_msg(c, 0, fmt, args);
+ va_end(args);
+ return ret;
+}
+
+/*
+ * Use for logging messages during recovery to enable reserved space and avoid
+ * blocking.
+ */
+__printf(2, 3)
+int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
+{
+ va_list args;
+ int ret;
+
+ va_start(args, fmt);
+ ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
+ va_end(args);
+ return ret;
+}
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/log2.h>
+#include <linux/slab.h>
+#include "darray.h"
+
+int __bch2_darray_resize(darray_void *d, size_t element_size, size_t new_size, gfp_t gfp)
+{
+ if (new_size > d->size) {
+ new_size = roundup_pow_of_two(new_size);
+
+ void *data = krealloc_array(d->data, new_size, element_size, gfp);
+ if (!data)
+ return -ENOMEM;
+
+ d->data = data;
+ d->size = new_size;
+ }
+
+ return 0;
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_DISK_GROUPS_TYPES_H
+#define _BCACHEFS_DISK_GROUPS_TYPES_H
+
+struct bch_disk_group_cpu {
+ bool deleted;
+ u16 parent;
+ u8 label[BCH_SB_LABEL_SIZE];
+ struct bch_devs_mask devs;
+};
+
+struct bch_disk_groups_cpu {
+ struct rcu_head rcu;
+ unsigned nr;
+ struct bch_disk_group_cpu entries[] __counted_by(nr);
+};
+
+#endif /* _BCACHEFS_DISK_GROUPS_TYPES_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+#ifndef NO_BCACHEFS_FS
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "bkey_buf.h"
+#include "fs-io.h"
+#include "fs-io-buffered.h"
+#include "fs-io-direct.h"
+#include "fs-io-pagecache.h"
+#include "io_read.h"
+#include "io_write.h"
+
+#include <linux/backing-dev.h>
+#include <linux/pagemap.h>
+#include <linux/writeback.h>
+
+static inline bool bio_full(struct bio *bio, unsigned len)
+{
+ if (bio->bi_vcnt >= bio->bi_max_vecs)
+ return true;
+ if (bio->bi_iter.bi_size > UINT_MAX - len)
+ return true;
+ return false;
+}
+
+/* readpage(s): */
+
+static void bch2_readpages_end_io(struct bio *bio)
+{
+ struct folio_iter fi;
+
+ bio_for_each_folio_all(fi, bio) {
+ if (!bio->bi_status) {
+ folio_mark_uptodate(fi.folio);
+ } else {
+ folio_clear_uptodate(fi.folio);
+ folio_set_error(fi.folio);
+ }
+ folio_unlock(fi.folio);
+ }
+
+ bio_put(bio);
+}
+
+struct readpages_iter {
+ struct address_space *mapping;
+ unsigned idx;
+ folios folios;
+};
+
+static int readpages_iter_init(struct readpages_iter *iter,
+ struct readahead_control *ractl)
+{
+ struct folio **fi;
+ int ret;
+
+ memset(iter, 0, sizeof(*iter));
+
+ iter->mapping = ractl->mapping;
+
+ ret = bch2_filemap_get_contig_folios_d(iter->mapping,
+ ractl->_index << PAGE_SHIFT,
+ (ractl->_index + ractl->_nr_pages) << PAGE_SHIFT,
+ 0, mapping_gfp_mask(iter->mapping),
+ &iter->folios);
+ if (ret)
+ return ret;
+
+ darray_for_each(iter->folios, fi) {
+ ractl->_nr_pages -= 1U << folio_order(*fi);
+ __bch2_folio_create(*fi, __GFP_NOFAIL|GFP_KERNEL);
+ folio_put(*fi);
+ folio_put(*fi);
+ }
+
+ return 0;
+}
+
+static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
+{
+ if (iter->idx >= iter->folios.nr)
+ return NULL;
+ return iter->folios.data[iter->idx];
+}
+
+static inline void readpage_iter_advance(struct readpages_iter *iter)
+{
+ iter->idx++;
+}
+
+static bool extent_partial_reads_expensive(struct bkey_s_c k)
+{
+ struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+ struct bch_extent_crc_unpacked crc;
+ const union bch_extent_entry *i;
+
+ bkey_for_each_crc(k.k, ptrs, crc, i)
+ if (crc.csum_type || crc.compression_type)
+ return true;
+ return false;
+}
+
+static int readpage_bio_extend(struct btree_trans *trans,
+ struct readpages_iter *iter,
+ struct bio *bio,
+ unsigned sectors_this_extent,
+ bool get_more)
+{
+ /* Don't hold btree locks while allocating memory: */
+ bch2_trans_unlock(trans);
+
+ while (bio_sectors(bio) < sectors_this_extent &&
+ bio->bi_vcnt < bio->bi_max_vecs) {
+ struct folio *folio = readpage_iter_peek(iter);
+ int ret;
+
+ if (folio) {
+ readpage_iter_advance(iter);
+ } else {
+ pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
+
+ if (!get_more)
+ break;
+
+ folio = xa_load(&iter->mapping->i_pages, folio_offset);
+ if (folio && !xa_is_value(folio))
+ break;
+
+ folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
+ if (!folio)
+ break;
+
+ if (!__bch2_folio_create(folio, GFP_KERNEL)) {
+ folio_put(folio);
+ break;
+ }
+
+ ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
+ if (ret) {
+ __bch2_folio_release(folio);
+ folio_put(folio);
+ break;
+ }
+
+ folio_put(folio);
+ }
+
+ BUG_ON(folio_sector(folio) != bio_end_sector(bio));
+
+ BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
+ }
+
+ return bch2_trans_relock(trans);
+}
+
+static void bchfs_read(struct btree_trans *trans,
+ struct bch_read_bio *rbio,
+ subvol_inum inum,
+ struct readpages_iter *readpages_iter)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_buf sk;
+ int flags = BCH_READ_RETRY_IF_STALE|
+ BCH_READ_MAY_PROMOTE;
+ u32 snapshot;
+ int ret = 0;
+
+ rbio->c = c;
+ rbio->start_time = local_clock();
+ rbio->subvol = inum.subvol;
+
+ bch2_bkey_buf_init(&sk);
+retry:
+ bch2_trans_begin(trans);
+ iter = (struct btree_iter) { NULL };
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (ret)
+ goto err;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+ SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
+ BTREE_ITER_SLOTS);
+ while (1) {
+ struct bkey_s_c k;
+ unsigned bytes, sectors, offset_into_extent;
+ enum btree_id data_btree = BTREE_ID_extents;
+
+ /*
+ * read_extent -> io_time_reset may cause a transaction restart
+ * without returning an error, we need to check for that here:
+ */
+ ret = bch2_trans_relock(trans);
+ if (ret)
+ break;
+
+ bch2_btree_iter_set_pos(&iter,
+ POS(inum.inum, rbio->bio.bi_iter.bi_sector));
+
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ break;
+
+ offset_into_extent = iter.pos.offset -
+ bkey_start_offset(k.k);
+ sectors = k.k->size - offset_into_extent;
+
+ bch2_bkey_buf_reassemble(&sk, c, k);
+
+ ret = bch2_read_indirect_extent(trans, &data_btree,
+ &offset_into_extent, &sk);
+ if (ret)
+ break;
+
+ k = bkey_i_to_s_c(sk.k);
+
+ sectors = min(sectors, k.k->size - offset_into_extent);
+
+ if (readpages_iter) {
+ ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
+ extent_partial_reads_expensive(k));
+ if (ret)
+ break;
+ }
+
+ bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
+ swap(rbio->bio.bi_iter.bi_size, bytes);
+
+ if (rbio->bio.bi_iter.bi_size == bytes)
+ flags |= BCH_READ_LAST_FRAGMENT;
+
+ bch2_bio_page_state_set(&rbio->bio, k);
+
+ bch2_read_extent(trans, rbio, iter.pos,
+ data_btree, k, offset_into_extent, flags);
+
+ if (flags & BCH_READ_LAST_FRAGMENT)
+ break;
+
+ swap(rbio->bio.bi_iter.bi_size, bytes);
+ bio_advance(&rbio->bio, bytes);
+
+ ret = btree_trans_too_many_iters(trans);
+ if (ret)
+ break;
+ }
+err:
+ bch2_trans_iter_exit(trans, &iter);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ goto retry;
+
+ if (ret) {
+ bch_err_inum_offset_ratelimited(c,
+ iter.pos.inode,
+ iter.pos.offset << 9,
+ "read error %i from btree lookup", ret);
+ rbio->bio.bi_status = BLK_STS_IOERR;
+ bio_endio(&rbio->bio);
+ }
+
+ bch2_bkey_buf_exit(&sk, c);
+}
+
+void bch2_readahead(struct readahead_control *ractl)
+{
+ struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch_io_opts opts;
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct folio *folio;
+ struct readpages_iter readpages_iter;
+ int ret;
+
+ bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+ ret = readpages_iter_init(&readpages_iter, ractl);
+ BUG_ON(ret);
+
+ bch2_pagecache_add_get(inode);
+
+ while ((folio = readpage_iter_peek(&readpages_iter))) {
+ unsigned n = min_t(unsigned,
+ readpages_iter.folios.nr -
+ readpages_iter.idx,
+ BIO_MAX_VECS);
+ struct bch_read_bio *rbio =
+ rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
+ GFP_KERNEL, &c->bio_read),
+ opts);
+
+ readpage_iter_advance(&readpages_iter);
+
+ rbio->bio.bi_iter.bi_sector = folio_sector(folio);
+ rbio->bio.bi_end_io = bch2_readpages_end_io;
+ BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
+
+ bchfs_read(trans, rbio, inode_inum(inode),
+ &readpages_iter);
+ bch2_trans_unlock(trans);
+ }
+
+ bch2_pagecache_add_put(inode);
+
+ bch2_trans_put(trans);
+ darray_exit(&readpages_iter.folios);
+}
+
+static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio,
+ subvol_inum inum, struct folio *folio)
+{
+ bch2_folio_create(folio, __GFP_NOFAIL);
+
+ rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
+ rbio->bio.bi_iter.bi_sector = folio_sector(folio);
+ BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
+
+ bch2_trans_run(c, (bchfs_read(trans, rbio, inum, NULL), 0));
+}
+
+static void bch2_read_single_folio_end_io(struct bio *bio)
+{
+ complete(bio->bi_private);
+}
+
+int bch2_read_single_folio(struct folio *folio, struct address_space *mapping)
+{
+ struct bch_inode_info *inode = to_bch_ei(mapping->host);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch_read_bio *rbio;
+ struct bch_io_opts opts;
+ int ret;
+ DECLARE_COMPLETION_ONSTACK(done);
+
+ bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+ rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
+ opts);
+ rbio->bio.bi_private = &done;
+ rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
+
+ __bchfs_readfolio(c, rbio, inode_inum(inode), folio);
+ wait_for_completion(&done);
+
+ ret = blk_status_to_errno(rbio->bio.bi_status);
+ bio_put(&rbio->bio);
+
+ if (ret < 0)
+ return ret;
+
+ folio_mark_uptodate(folio);
+ return 0;
+}
+
+int bch2_read_folio(struct file *file, struct folio *folio)
+{
+ int ret;
+
+ ret = bch2_read_single_folio(folio, folio->mapping);
+ folio_unlock(folio);
+ return bch2_err_class(ret);
+}
+
+/* writepages: */
+
+struct bch_writepage_io {
+ struct bch_inode_info *inode;
+
+ /* must be last: */
+ struct bch_write_op op;
+};
+
+struct bch_writepage_state {
+ struct bch_writepage_io *io;
+ struct bch_io_opts opts;
+ struct bch_folio_sector *tmp;
+ unsigned tmp_sectors;
+};
+
+static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
+ struct bch_inode_info *inode)
+{
+ struct bch_writepage_state ret = { 0 };
+
+ bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
+ return ret;
+}
+
+/*
+ * Determine when a writepage io is full. We have to limit writepage bios to a
+ * single page per bvec (i.e. 1MB with 4k pages) because that is the limit to
+ * what the bounce path in bch2_write_extent() can handle. In theory we could
+ * loosen this restriction for non-bounce I/O, but we don't have that context
+ * here. Ideally, we can up this limit and make it configurable in the future
+ * when the bounce path can be enhanced to accommodate larger source bios.
+ */
+static inline bool bch_io_full(struct bch_writepage_io *io, unsigned len)
+{
+ struct bio *bio = &io->op.wbio.bio;
+ return bio_full(bio, len) ||
+ (bio->bi_iter.bi_size + len > BIO_MAX_VECS * PAGE_SIZE);
+}
+
+static void bch2_writepage_io_done(struct bch_write_op *op)
+{
+ struct bch_writepage_io *io =
+ container_of(op, struct bch_writepage_io, op);
+ struct bch_fs *c = io->op.c;
+ struct bio *bio = &io->op.wbio.bio;
+ struct folio_iter fi;
+ unsigned i;
+
+ if (io->op.error) {
+ set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
+
+ bio_for_each_folio_all(fi, bio) {
+ struct bch_folio *s;
+
+ folio_set_error(fi.folio);
+ mapping_set_error(fi.folio->mapping, -EIO);
+
+ s = __bch2_folio(fi.folio);
+ spin_lock(&s->lock);
+ for (i = 0; i < folio_sectors(fi.folio); i++)
+ s->s[i].nr_replicas = 0;
+ spin_unlock(&s->lock);
+ }
+ }
+
+ if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
+ bio_for_each_folio_all(fi, bio) {
+ struct bch_folio *s;
+
+ s = __bch2_folio(fi.folio);
+ spin_lock(&s->lock);
+ for (i = 0; i < folio_sectors(fi.folio); i++)
+ s->s[i].nr_replicas = 0;
+ spin_unlock(&s->lock);
+ }
+ }
+
+ /*
+ * racing with fallocate can cause us to add fewer sectors than
+ * expected - but we shouldn't add more sectors than expected:
+ */
+ WARN_ON_ONCE(io->op.i_sectors_delta > 0);
+
+ /*
+ * (error (due to going RO) halfway through a page can screw that up
+ * slightly)
+ * XXX wtf?
+ BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
+ */
+
+ /*
+ * PageWriteback is effectively our ref on the inode - fixup i_blocks
+ * before calling end_page_writeback:
+ */
+ bch2_i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
+
+ bio_for_each_folio_all(fi, bio) {
+ struct bch_folio *s = __bch2_folio(fi.folio);
+
+ if (atomic_dec_and_test(&s->write_count))
+ folio_end_writeback(fi.folio);
+ }
+
+ bio_put(&io->op.wbio.bio);
+}
+
+static void bch2_writepage_do_io(struct bch_writepage_state *w)
+{
+ struct bch_writepage_io *io = w->io;
+
+ w->io = NULL;
+ closure_call(&io->op.cl, bch2_write, NULL, NULL);
+}
+
+/*
+ * Get a bch_writepage_io and add @page to it - appending to an existing one if
+ * possible, else allocating a new one:
+ */
+static void bch2_writepage_io_alloc(struct bch_fs *c,
+ struct writeback_control *wbc,
+ struct bch_writepage_state *w,
+ struct bch_inode_info *inode,
+ u64 sector,
+ unsigned nr_replicas)
+{
+ struct bch_write_op *op;
+
+ w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
+ REQ_OP_WRITE,
+ GFP_KERNEL,
+ &c->writepage_bioset),
+ struct bch_writepage_io, op.wbio.bio);
+
+ w->io->inode = inode;
+ op = &w->io->op;
+ bch2_write_op_init(op, c, w->opts);
+ op->target = w->opts.foreground_target;
+ op->nr_replicas = nr_replicas;
+ op->res.nr_replicas = nr_replicas;
+ op->write_point = writepoint_hashed(inode->ei_last_dirtied);
+ op->subvol = inode->ei_subvol;
+ op->pos = POS(inode->v.i_ino, sector);
+ op->end_io = bch2_writepage_io_done;
+ op->devs_need_flush = &inode->ei_devs_need_flush;
+ op->wbio.bio.bi_iter.bi_sector = sector;
+ op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
+}
+
+static int __bch2_writepage(struct folio *folio,
+ struct writeback_control *wbc,
+ void *data)
+{
+ struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch_writepage_state *w = data;
+ struct bch_folio *s;
+ unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
+ loff_t i_size = i_size_read(&inode->v);
+ int ret;
+
+ EBUG_ON(!folio_test_uptodate(folio));
+
+ /* Is the folio fully inside i_size? */
+ if (folio_end_pos(folio) <= i_size)
+ goto do_io;
+
+ /* Is the folio fully outside i_size? (truncate in progress) */
+ if (folio_pos(folio) >= i_size) {
+ folio_unlock(folio);
+ return 0;
+ }
+
+ /*
+ * The folio straddles i_size. It must be zeroed out on each and every
+ * writepage invocation because it may be mmapped. "A file is mapped
+ * in multiples of the folio size. For a file that is not a multiple of
+ * the folio size, the remaining memory is zeroed when mapped, and
+ * writes to that region are not written out to the file."
+ */
+ folio_zero_segment(folio,
+ i_size - folio_pos(folio),
+ folio_size(folio));
+do_io:
+ f_sectors = folio_sectors(folio);
+ s = bch2_folio(folio);
+
+ if (f_sectors > w->tmp_sectors) {
+ kfree(w->tmp);
+ w->tmp = kcalloc(f_sectors, sizeof(struct bch_folio_sector), __GFP_NOFAIL);
+ w->tmp_sectors = f_sectors;
+ }
+
+ /*
+ * Things get really hairy with errors during writeback:
+ */
+ ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
+ BUG_ON(ret);
+
+ /* Before unlocking the page, get copy of reservations: */
+ spin_lock(&s->lock);
+ memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
+
+ for (i = 0; i < f_sectors; i++) {
+ if (s->s[i].state < SECTOR_dirty)
+ continue;
+
+ nr_replicas_this_write =
+ min_t(unsigned, nr_replicas_this_write,
+ s->s[i].nr_replicas +
+ s->s[i].replicas_reserved);
+ }
+
+ for (i = 0; i < f_sectors; i++) {
+ if (s->s[i].state < SECTOR_dirty)
+ continue;
+
+ s->s[i].nr_replicas = w->opts.compression
+ ? 0 : nr_replicas_this_write;
+
+ s->s[i].replicas_reserved = 0;
+ bch2_folio_sector_set(folio, s, i, SECTOR_allocated);
+ }
+ spin_unlock(&s->lock);
+
+ BUG_ON(atomic_read(&s->write_count));
+ atomic_set(&s->write_count, 1);
+
+ BUG_ON(folio_test_writeback(folio));
+ folio_start_writeback(folio);
+
+ folio_unlock(folio);
+
+ offset = 0;
+ while (1) {
+ unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
+ u64 sector;
+
+ while (offset < f_sectors &&
+ w->tmp[offset].state < SECTOR_dirty)
+ offset++;
+
+ if (offset == f_sectors)
+ break;
+
+ while (offset + sectors < f_sectors &&
+ w->tmp[offset + sectors].state >= SECTOR_dirty) {
+ reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
+ dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
+ sectors++;
+ }
+ BUG_ON(!sectors);
+
+ sector = folio_sector(folio) + offset;
+
+ if (w->io &&
+ (w->io->op.res.nr_replicas != nr_replicas_this_write ||
+ bch_io_full(w->io, sectors << 9) ||
+ bio_end_sector(&w->io->op.wbio.bio) != sector))
+ bch2_writepage_do_io(w);
+
+ if (!w->io)
+ bch2_writepage_io_alloc(c, wbc, w, inode, sector,
+ nr_replicas_this_write);
+
+ atomic_inc(&s->write_count);
+
+ BUG_ON(inode != w->io->inode);
+ BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
+ sectors << 9, offset << 9));
+
+ /* Check for writing past i_size: */
+ WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
+ round_up(i_size, block_bytes(c)) &&
+ !test_bit(BCH_FS_EMERGENCY_RO, &c->flags),
+ "writing past i_size: %llu > %llu (unrounded %llu)\n",
+ bio_end_sector(&w->io->op.wbio.bio) << 9,
+ round_up(i_size, block_bytes(c)),
+ i_size);
+
+ w->io->op.res.sectors += reserved_sectors;
+ w->io->op.i_sectors_delta -= dirty_sectors;
+ w->io->op.new_i_size = i_size;
+
+ offset += sectors;
+ }
+
+ if (atomic_dec_and_test(&s->write_count))
+ folio_end_writeback(folio);
+
+ return 0;
+}
+
+int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
+{
+ struct bch_fs *c = mapping->host->i_sb->s_fs_info;
+ struct bch_writepage_state w =
+ bch_writepage_state_init(c, to_bch_ei(mapping->host));
+ struct blk_plug plug;
+ int ret;
+
+ blk_start_plug(&plug);
+ ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
+ if (w.io)
+ bch2_writepage_do_io(&w);
+ blk_finish_plug(&plug);
+ kfree(w.tmp);
+ return bch2_err_class(ret);
+}
+
+/* buffered writes: */
+
+int bch2_write_begin(struct file *file, struct address_space *mapping,
+ loff_t pos, unsigned len,
+ struct page **pagep, void **fsdata)
+{
+ struct bch_inode_info *inode = to_bch_ei(mapping->host);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch2_folio_reservation *res;
+ struct folio *folio;
+ unsigned offset;
+ int ret = -ENOMEM;
+
+ res = kmalloc(sizeof(*res), GFP_KERNEL);
+ if (!res)
+ return -ENOMEM;
+
+ bch2_folio_reservation_init(c, inode, res);
+ *fsdata = res;
+
+ bch2_pagecache_add_get(inode);
+
+ folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
+ FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE,
+ mapping_gfp_mask(mapping));
+ if (IS_ERR_OR_NULL(folio))
+ goto err_unlock;
+
+ offset = pos - folio_pos(folio);
+ len = min_t(size_t, len, folio_end_pos(folio) - pos);
+
+ if (folio_test_uptodate(folio))
+ goto out;
+
+ /* If we're writing entire folio, don't need to read it in first: */
+ if (!offset && len == folio_size(folio))
+ goto out;
+
+ if (!offset && pos + len >= inode->v.i_size) {
+ folio_zero_segment(folio, len, folio_size(folio));
+ flush_dcache_folio(folio);
+ goto out;
+ }
+
+ if (folio_pos(folio) >= inode->v.i_size) {
+ folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
+ flush_dcache_folio(folio);
+ goto out;
+ }
+readpage:
+ ret = bch2_read_single_folio(folio, mapping);
+ if (ret)
+ goto err;
+out:
+ ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
+ if (ret)
+ goto err;
+
+ ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
+ if (ret) {
+ if (!folio_test_uptodate(folio)) {
+ /*
+ * If the folio hasn't been read in, we won't know if we
+ * actually need a reservation - we don't actually need
+ * to read here, we just need to check if the folio is
+ * fully backed by uncompressed data:
+ */
+ goto readpage;
+ }
+
+ goto err;
+ }
+
+ *pagep = &folio->page;
+ return 0;
+err:
+ folio_unlock(folio);
+ folio_put(folio);
+ *pagep = NULL;
+err_unlock:
+ bch2_pagecache_add_put(inode);
+ kfree(res);
+ *fsdata = NULL;
+ return bch2_err_class(ret);
+}
+
+int bch2_write_end(struct file *file, struct address_space *mapping,
+ loff_t pos, unsigned len, unsigned copied,
+ struct page *page, void *fsdata)
+{
+ struct bch_inode_info *inode = to_bch_ei(mapping->host);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch2_folio_reservation *res = fsdata;
+ struct folio *folio = page_folio(page);
+ unsigned offset = pos - folio_pos(folio);
+
+ lockdep_assert_held(&inode->v.i_rwsem);
+ BUG_ON(offset + copied > folio_size(folio));
+
+ if (unlikely(copied < len && !folio_test_uptodate(folio))) {
+ /*
+ * The folio needs to be read in, but that would destroy
+ * our partial write - simplest thing is to just force
+ * userspace to redo the write:
+ */
+ folio_zero_range(folio, 0, folio_size(folio));
+ flush_dcache_folio(folio);
+ copied = 0;
+ }
+
+ spin_lock(&inode->v.i_lock);
+ if (pos + copied > inode->v.i_size)
+ i_size_write(&inode->v, pos + copied);
+ spin_unlock(&inode->v.i_lock);
+
+ if (copied) {
+ if (!folio_test_uptodate(folio))
+ folio_mark_uptodate(folio);
+
+ bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
+
+ inode->ei_last_dirtied = (unsigned long) current;
+ }
+
+ folio_unlock(folio);
+ folio_put(folio);
+ bch2_pagecache_add_put(inode);
+
+ bch2_folio_reservation_put(c, inode, res);
+ kfree(res);
+
+ return copied;
+}
+
+static noinline void folios_trunc(folios *fs, struct folio **fi)
+{
+ while (fs->data + fs->nr > fi) {
+ struct folio *f = darray_pop(fs);
+
+ folio_unlock(f);
+ folio_put(f);
+ }
+}
+
+static int __bch2_buffered_write(struct bch_inode_info *inode,
+ struct address_space *mapping,
+ struct iov_iter *iter,
+ loff_t pos, unsigned len)
+{
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch2_folio_reservation res;
+ folios fs;
+ struct folio **fi, *f;
+ unsigned copied = 0, f_offset, f_copied;
+ u64 end = pos + len, f_pos, f_len;
+ loff_t last_folio_pos = inode->v.i_size;
+ int ret = 0;
+
+ BUG_ON(!len);
+
+ bch2_folio_reservation_init(c, inode, &res);
+ darray_init(&fs);
+
+ ret = bch2_filemap_get_contig_folios_d(mapping, pos, end,
+ FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT,
+ mapping_gfp_mask(mapping),
+ &fs);
+ if (ret)
+ goto out;
+
+ BUG_ON(!fs.nr);
+
+ f = darray_first(fs);
+ if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
+ ret = bch2_read_single_folio(f, mapping);
+ if (ret)
+ goto out;
+ }
+
+ f = darray_last(fs);
+ end = min(end, folio_end_pos(f));
+ last_folio_pos = folio_pos(f);
+ if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
+ if (end >= inode->v.i_size) {
+ folio_zero_range(f, 0, folio_size(f));
+ } else {
+ ret = bch2_read_single_folio(f, mapping);
+ if (ret)
+ goto out;
+ }
+ }
+
+ ret = bch2_folio_set(c, inode_inum(inode), fs.data, fs.nr);
+ if (ret)
+ goto out;
+
+ f_pos = pos;
+ f_offset = pos - folio_pos(darray_first(fs));
+ darray_for_each(fs, fi) {
+ f = *fi;
+ f_len = min(end, folio_end_pos(f)) - f_pos;
+
+ /*
+ * XXX: per POSIX and fstests generic/275, on -ENOSPC we're
+ * supposed to write as much as we have disk space for.
+ *
+ * On failure here we should still write out a partial page if
+ * we aren't completely out of disk space - we don't do that
+ * yet:
+ */
+ ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len);
+ if (unlikely(ret)) {
+ folios_trunc(&fs, fi);
+ if (!fs.nr)
+ goto out;
+
+ end = min(end, folio_end_pos(darray_last(fs)));
+ break;
+ }
+
+ f_pos = folio_end_pos(f);
+ f_offset = 0;
+ }
+
+ if (mapping_writably_mapped(mapping))
+ darray_for_each(fs, fi)
+ flush_dcache_folio(*fi);
+
+ f_pos = pos;
+ f_offset = pos - folio_pos(darray_first(fs));
+ darray_for_each(fs, fi) {
+ f = *fi;
+ f_len = min(end, folio_end_pos(f)) - f_pos;
+ f_copied = copy_page_from_iter_atomic(&f->page, f_offset, f_len, iter);
+ if (!f_copied) {
+ folios_trunc(&fs, fi);
+ break;
+ }
+
+ if (!folio_test_uptodate(f) &&
+ f_copied != folio_size(f) &&
+ pos + copied + f_copied < inode->v.i_size) {
+ iov_iter_revert(iter, f_copied);
+ folio_zero_range(f, 0, folio_size(f));
+ folios_trunc(&fs, fi);
+ break;
+ }
+
+ flush_dcache_folio(f);
+ copied += f_copied;
+
+ if (f_copied != f_len) {
+ folios_trunc(&fs, fi + 1);
+ break;
+ }
+
+ f_pos = folio_end_pos(f);
+ f_offset = 0;
+ }
+
+ if (!copied)
+ goto out;
+
+ end = pos + copied;
+
+ spin_lock(&inode->v.i_lock);
+ if (end > inode->v.i_size)
+ i_size_write(&inode->v, end);
+ spin_unlock(&inode->v.i_lock);
+
+ f_pos = pos;
+ f_offset = pos - folio_pos(darray_first(fs));
+ darray_for_each(fs, fi) {
+ f = *fi;
+ f_len = min(end, folio_end_pos(f)) - f_pos;
+
+ if (!folio_test_uptodate(f))
+ folio_mark_uptodate(f);
+
+ bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
+
+ f_pos = folio_end_pos(f);
+ f_offset = 0;
+ }
+
+ inode->ei_last_dirtied = (unsigned long) current;
+out:
+ darray_for_each(fs, fi) {
+ folio_unlock(*fi);
+ folio_put(*fi);
+ }
+
+ /*
+ * If the last folio added to the mapping starts beyond current EOF, we
+ * performed a short write but left around at least one post-EOF folio.
+ * Clean up the mapping before we return.
+ */
+ if (last_folio_pos >= inode->v.i_size)
+ truncate_pagecache(&inode->v, inode->v.i_size);
+
+ darray_exit(&fs);
+ bch2_folio_reservation_put(c, inode, &res);
+
+ return copied ?: ret;
+}
+
+static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
+{
+ struct file *file = iocb->ki_filp;
+ struct address_space *mapping = file->f_mapping;
+ struct bch_inode_info *inode = file_bch_inode(file);
+ loff_t pos = iocb->ki_pos;
+ ssize_t written = 0;
+ int ret = 0;
+
+ bch2_pagecache_add_get(inode);
+
+ do {
+ unsigned offset = pos & (PAGE_SIZE - 1);
+ unsigned bytes = iov_iter_count(iter);
+again:
+ /*
+ * Bring in the user page that we will copy from _first_.
+ * Otherwise there's a nasty deadlock on copying from the
+ * same page as we're writing to, without it being marked
+ * up-to-date.
+ *
+ * Not only is this an optimisation, but it is also required
+ * to check that the address is actually valid, when atomic
+ * usercopies are used, below.
+ */
+ if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
+ bytes = min_t(unsigned long, iov_iter_count(iter),
+ PAGE_SIZE - offset);
+
+ if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
+ ret = -EFAULT;
+ break;
+ }
+ }
+
+ if (unlikely(fatal_signal_pending(current))) {
+ ret = -EINTR;
+ break;
+ }
+
+ ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
+ if (unlikely(ret < 0))
+ break;
+
+ cond_resched();
+
+ if (unlikely(ret == 0)) {
+ /*
+ * If we were unable to copy any data at all, we must
+ * fall back to a single segment length write.
+ *
+ * If we didn't fallback here, we could livelock
+ * because not all segments in the iov can be copied at
+ * once without a pagefault.
+ */
+ bytes = min_t(unsigned long, PAGE_SIZE - offset,
+ iov_iter_single_seg_count(iter));
+ goto again;
+ }
+ pos += ret;
+ written += ret;
+ ret = 0;
+
+ balance_dirty_pages_ratelimited(mapping);
+ } while (iov_iter_count(iter));
+
+ bch2_pagecache_add_put(inode);
+
+ return written ? written : ret;
+}
+
+ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
+{
+ struct file *file = iocb->ki_filp;
+ struct bch_inode_info *inode = file_bch_inode(file);
+ ssize_t ret;
+
+ if (iocb->ki_flags & IOCB_DIRECT) {
+ ret = bch2_direct_write(iocb, from);
+ goto out;
+ }
+
+ inode_lock(&inode->v);
+
+ ret = generic_write_checks(iocb, from);
+ if (ret <= 0)
+ goto unlock;
+
+ ret = file_remove_privs(file);
+ if (ret)
+ goto unlock;
+
+ ret = file_update_time(file);
+ if (ret)
+ goto unlock;
+
+ ret = bch2_buffered_write(iocb, from);
+ if (likely(ret > 0))
+ iocb->ki_pos += ret;
+unlock:
+ inode_unlock(&inode->v);
+
+ if (ret > 0)
+ ret = generic_write_sync(iocb, ret);
+out:
+ return bch2_err_class(ret);
+}
+
+void bch2_fs_fs_io_buffered_exit(struct bch_fs *c)
+{
+ bioset_exit(&c->writepage_bioset);
+}
+
+int bch2_fs_fs_io_buffered_init(struct bch_fs *c)
+{
+ if (bioset_init(&c->writepage_bioset,
+ 4, offsetof(struct bch_writepage_io, op.wbio.bio),
+ BIOSET_NEED_BVECS))
+ return -BCH_ERR_ENOMEM_writepage_bioset_init;
+
+ return 0;
+}
+
+#endif /* NO_BCACHEFS_FS */
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_FS_IO_BUFFERED_H
+#define _BCACHEFS_FS_IO_BUFFERED_H
+
+#ifndef NO_BCACHEFS_FS
+
+int bch2_read_single_folio(struct folio *, struct address_space *);
+int bch2_read_folio(struct file *, struct folio *);
+
+int bch2_writepages(struct address_space *, struct writeback_control *);
+void bch2_readahead(struct readahead_control *);
+
+int bch2_write_begin(struct file *, struct address_space *, loff_t,
+ unsigned, struct page **, void **);
+int bch2_write_end(struct file *, struct address_space *, loff_t,
+ unsigned, unsigned, struct page *, void *);
+
+ssize_t bch2_write_iter(struct kiocb *, struct iov_iter *);
+
+void bch2_fs_fs_io_buffered_exit(struct bch_fs *);
+int bch2_fs_fs_io_buffered_init(struct bch_fs *);
+#else
+static inline void bch2_fs_fs_io_buffered_exit(struct bch_fs *c) {}
+static inline int bch2_fs_fs_io_buffered_init(struct bch_fs *c) { return 0; }
+#endif
+
+#endif /* _BCACHEFS_FS_IO_BUFFERED_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+#ifndef NO_BCACHEFS_FS
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "fs.h"
+#include "fs-io.h"
+#include "fs-io-direct.h"
+#include "fs-io-pagecache.h"
+#include "io_read.h"
+#include "io_write.h"
+
+#include <linux/kthread.h>
+#include <linux/pagemap.h>
+#include <linux/prefetch.h>
+#include <linux/task_io_accounting_ops.h>
+
+/* O_DIRECT reads */
+
+struct dio_read {
+ struct closure cl;
+ struct kiocb *req;
+ long ret;
+ bool should_dirty;
+ struct bch_read_bio rbio;
+};
+
+static void bio_check_or_release(struct bio *bio, bool check_dirty)
+{
+ if (check_dirty) {
+ bio_check_pages_dirty(bio);
+ } else {
+ bio_release_pages(bio, false);
+ bio_put(bio);
+ }
+}
+
+static void bch2_dio_read_complete(struct closure *cl)
+{
+ struct dio_read *dio = container_of(cl, struct dio_read, cl);
+
+ dio->req->ki_complete(dio->req, dio->ret);
+ bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
+}
+
+static void bch2_direct_IO_read_endio(struct bio *bio)
+{
+ struct dio_read *dio = bio->bi_private;
+
+ if (bio->bi_status)
+ dio->ret = blk_status_to_errno(bio->bi_status);
+
+ closure_put(&dio->cl);
+}
+
+static void bch2_direct_IO_read_split_endio(struct bio *bio)
+{
+ struct dio_read *dio = bio->bi_private;
+ bool should_dirty = dio->should_dirty;
+
+ bch2_direct_IO_read_endio(bio);
+ bio_check_or_release(bio, should_dirty);
+}
+
+static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
+{
+ struct file *file = req->ki_filp;
+ struct bch_inode_info *inode = file_bch_inode(file);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch_io_opts opts;
+ struct dio_read *dio;
+ struct bio *bio;
+ loff_t offset = req->ki_pos;
+ bool sync = is_sync_kiocb(req);
+ size_t shorten;
+ ssize_t ret;
+
+ bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+ if ((offset|iter->count) & (block_bytes(c) - 1))
+ return -EINVAL;
+
+ ret = min_t(loff_t, iter->count,
+ max_t(loff_t, 0, i_size_read(&inode->v) - offset));
+
+ if (!ret)
+ return ret;
+
+ shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
+ iter->count -= shorten;
+
+ bio = bio_alloc_bioset(NULL,
+ bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
+ REQ_OP_READ,
+ GFP_KERNEL,
+ &c->dio_read_bioset);
+
+ bio->bi_end_io = bch2_direct_IO_read_endio;
+
+ dio = container_of(bio, struct dio_read, rbio.bio);
+ closure_init(&dio->cl, NULL);
+
+ /*
+ * this is a _really_ horrible hack just to avoid an atomic sub at the
+ * end:
+ */
+ if (!sync) {
+ set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
+ atomic_set(&dio->cl.remaining,
+ CLOSURE_REMAINING_INITIALIZER -
+ CLOSURE_RUNNING +
+ CLOSURE_DESTRUCTOR);
+ } else {
+ atomic_set(&dio->cl.remaining,
+ CLOSURE_REMAINING_INITIALIZER + 1);
+ dio->cl.closure_get_happened = true;
+ }
+
+ dio->req = req;
+ dio->ret = ret;
+ /*
+ * This is one of the sketchier things I've encountered: we have to skip
+ * the dirtying of requests that are internal from the kernel (i.e. from
+ * loopback), because we'll deadlock on page_lock.
+ */
+ dio->should_dirty = iter_is_iovec(iter);
+
+ goto start;
+ while (iter->count) {
+ bio = bio_alloc_bioset(NULL,
+ bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
+ REQ_OP_READ,
+ GFP_KERNEL,
+ &c->bio_read);
+ bio->bi_end_io = bch2_direct_IO_read_split_endio;
+start:
+ bio->bi_opf = REQ_OP_READ|REQ_SYNC;
+ bio->bi_iter.bi_sector = offset >> 9;
+ bio->bi_private = dio;
+
+ ret = bio_iov_iter_get_pages(bio, iter);
+ if (ret < 0) {
+ /* XXX: fault inject this path */
+ bio->bi_status = BLK_STS_RESOURCE;
+ bio_endio(bio);
+ break;
+ }
+
+ offset += bio->bi_iter.bi_size;
+
+ if (dio->should_dirty)
+ bio_set_pages_dirty(bio);
+
+ if (iter->count)
+ closure_get(&dio->cl);
+
+ bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
+ }
+
+ iter->count += shorten;
+
+ if (sync) {
+ closure_sync(&dio->cl);
+ closure_debug_destroy(&dio->cl);
+ ret = dio->ret;
+ bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
+ return ret;
+ } else {
+ return -EIOCBQUEUED;
+ }
+}
+
+ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
+{
+ struct file *file = iocb->ki_filp;
+ struct bch_inode_info *inode = file_bch_inode(file);
+ struct address_space *mapping = file->f_mapping;
+ size_t count = iov_iter_count(iter);
+ ssize_t ret;
+
+ if (!count)
+ return 0; /* skip atime */
+
+ if (iocb->ki_flags & IOCB_DIRECT) {
+ struct blk_plug plug;
+
+ if (unlikely(mapping->nrpages)) {
+ ret = filemap_write_and_wait_range(mapping,
+ iocb->ki_pos,
+ iocb->ki_pos + count - 1);
+ if (ret < 0)
+ goto out;
+ }
+
+ file_accessed(file);
+
+ blk_start_plug(&plug);
+ ret = bch2_direct_IO_read(iocb, iter);
+ blk_finish_plug(&plug);
+
+ if (ret >= 0)
+ iocb->ki_pos += ret;
+ } else {
+ bch2_pagecache_add_get(inode);
+ ret = generic_file_read_iter(iocb, iter);
+ bch2_pagecache_add_put(inode);
+ }
+out:
+ return bch2_err_class(ret);
+}
+
+/* O_DIRECT writes */
+
+struct dio_write {
+ struct kiocb *req;
+ struct address_space *mapping;
+ struct bch_inode_info *inode;
+ struct mm_struct *mm;
+ unsigned loop:1,
+ extending:1,
+ sync:1,
+ flush:1,
+ free_iov:1;
+ struct quota_res quota_res;
+ u64 written;
+
+ struct iov_iter iter;
+ struct iovec inline_vecs[2];
+
+ /* must be last: */
+ struct bch_write_op op;
+};
+
+static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
+ u64 offset, u64 size,
+ unsigned nr_replicas, bool compressed)
+{
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ u64 end = offset + size;
+ u32 snapshot;
+ bool ret = true;
+ int err;
+retry:
+ bch2_trans_begin(trans);
+
+ err = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (err)
+ goto err;
+
+ for_each_btree_key_norestart(trans, iter, BTREE_ID_extents,
+ SPOS(inum.inum, offset, snapshot),
+ BTREE_ITER_SLOTS, k, err) {
+ if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end)))
+ break;
+
+ if (k.k->p.snapshot != snapshot ||
+ nr_replicas > bch2_bkey_replicas(c, k) ||
+ (!compressed && bch2_bkey_sectors_compressed(k))) {
+ ret = false;
+ break;
+ }
+ }
+
+ offset = iter.pos.offset;
+ bch2_trans_iter_exit(trans, &iter);
+err:
+ if (bch2_err_matches(err, BCH_ERR_transaction_restart))
+ goto retry;
+ bch2_trans_put(trans);
+
+ return err ? false : ret;
+}
+
+static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio)
+{
+ struct bch_fs *c = dio->op.c;
+ struct bch_inode_info *inode = dio->inode;
+ struct bio *bio = &dio->op.wbio.bio;
+
+ return bch2_check_range_allocated(c, inode_inum(inode),
+ dio->op.pos.offset, bio_sectors(bio),
+ dio->op.opts.data_replicas,
+ dio->op.opts.compression != 0);
+}
+
+static void bch2_dio_write_loop_async(struct bch_write_op *);
+static __always_inline long bch2_dio_write_done(struct dio_write *dio);
+
+/*
+ * We're going to return -EIOCBQUEUED, but we haven't finished consuming the
+ * iov_iter yet, so we need to stash a copy of the iovec: it might be on the
+ * caller's stack, we're not guaranteed that it will live for the duration of
+ * the IO:
+ */
+static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
+{
+ struct iovec *iov = dio->inline_vecs;
+
+ /*
+ * iov_iter has a single embedded iovec - nothing to do:
+ */
+ if (iter_is_ubuf(&dio->iter))
+ return 0;
+
+ /*
+ * We don't currently handle non-iovec iov_iters here - return an error,
+ * and we'll fall back to doing the IO synchronously:
+ */
+ if (!iter_is_iovec(&dio->iter))
+ return -1;
+
+ if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
+ iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
+ GFP_KERNEL);
+ if (unlikely(!iov))
+ return -ENOMEM;
+
+ dio->free_iov = true;
+ }
+
+ memcpy(iov, dio->iter.__iov, dio->iter.nr_segs * sizeof(*iov));
+ dio->iter.__iov = iov;
+ return 0;
+}
+
+static void bch2_dio_write_flush_done(struct closure *cl)
+{
+ struct dio_write *dio = container_of(cl, struct dio_write, op.cl);
+ struct bch_fs *c = dio->op.c;
+
+ closure_debug_destroy(cl);
+
+ dio->op.error = bch2_journal_error(&c->journal);
+
+ bch2_dio_write_done(dio);
+}
+
+static noinline void bch2_dio_write_flush(struct dio_write *dio)
+{
+ struct bch_fs *c = dio->op.c;
+ struct bch_inode_unpacked inode;
+ int ret;
+
+ dio->flush = 0;
+
+ closure_init(&dio->op.cl, NULL);
+
+ if (!dio->op.error) {
+ ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode);
+ if (ret) {
+ dio->op.error = ret;
+ } else {
+ bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq,
+ &dio->op.cl);
+ bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl);
+ }
+ }
+
+ if (dio->sync) {
+ closure_sync(&dio->op.cl);
+ closure_debug_destroy(&dio->op.cl);
+ } else {
+ continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL);
+ }
+}
+
+static __always_inline long bch2_dio_write_done(struct dio_write *dio)
+{
+ struct kiocb *req = dio->req;
+ struct bch_inode_info *inode = dio->inode;
+ bool sync = dio->sync;
+ long ret;
+
+ if (unlikely(dio->flush)) {
+ bch2_dio_write_flush(dio);
+ if (!sync)
+ return -EIOCBQUEUED;
+ }
+
+ bch2_pagecache_block_put(inode);
+
+ if (dio->free_iov)
+ kfree(dio->iter.__iov);
+
+ ret = dio->op.error ?: ((long) dio->written << 9);
+ bio_put(&dio->op.wbio.bio);
+
+ /* inode->i_dio_count is our ref on inode and thus bch_fs */
+ inode_dio_end(&inode->v);
+
+ if (ret < 0)
+ ret = bch2_err_class(ret);
+
+ if (!sync) {
+ req->ki_complete(req, ret);
+ ret = -EIOCBQUEUED;
+ }
+ return ret;
+}
+
+static __always_inline void bch2_dio_write_end(struct dio_write *dio)
+{
+ struct bch_fs *c = dio->op.c;
+ struct kiocb *req = dio->req;
+ struct bch_inode_info *inode = dio->inode;
+ struct bio *bio = &dio->op.wbio.bio;
+
+ req->ki_pos += (u64) dio->op.written << 9;
+ dio->written += dio->op.written;
+
+ if (dio->extending) {
+ spin_lock(&inode->v.i_lock);
+ if (req->ki_pos > inode->v.i_size)
+ i_size_write(&inode->v, req->ki_pos);
+ spin_unlock(&inode->v.i_lock);
+ }
+
+ if (dio->op.i_sectors_delta || dio->quota_res.sectors) {
+ mutex_lock(&inode->ei_quota_lock);
+ __bch2_i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta);
+ __bch2_quota_reservation_put(c, inode, &dio->quota_res);
+ mutex_unlock(&inode->ei_quota_lock);
+ }
+
+ bio_release_pages(bio, false);
+
+ if (unlikely(dio->op.error))
+ set_bit(EI_INODE_ERROR, &inode->ei_flags);
+}
+
+static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
+{
+ struct bch_fs *c = dio->op.c;
+ struct kiocb *req = dio->req;
+ struct address_space *mapping = dio->mapping;
+ struct bch_inode_info *inode = dio->inode;
+ struct bch_io_opts opts;
+ struct bio *bio = &dio->op.wbio.bio;
+ unsigned unaligned, iter_count;
+ bool sync = dio->sync, dropped_locks;
+ long ret;
+
+ bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+ while (1) {
+ iter_count = dio->iter.count;
+
+ EBUG_ON(current->faults_disabled_mapping);
+ current->faults_disabled_mapping = mapping;
+
+ ret = bio_iov_iter_get_pages(bio, &dio->iter);
+
+ dropped_locks = fdm_dropped_locks();
+
+ current->faults_disabled_mapping = NULL;
+
+ /*
+ * If the fault handler returned an error but also signalled
+ * that it dropped & retook ei_pagecache_lock, we just need to
+ * re-shoot down the page cache and retry:
+ */
+ if (dropped_locks && ret)
+ ret = 0;
+
+ if (unlikely(ret < 0))
+ goto err;
+
+ if (unlikely(dropped_locks)) {
+ ret = bch2_write_invalidate_inode_pages_range(mapping,
+ req->ki_pos,
+ req->ki_pos + iter_count - 1);
+ if (unlikely(ret))
+ goto err;
+
+ if (!bio->bi_iter.bi_size)
+ continue;
+ }
+
+ unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
+ bio->bi_iter.bi_size -= unaligned;
+ iov_iter_revert(&dio->iter, unaligned);
+
+ if (!bio->bi_iter.bi_size) {
+ /*
+ * bio_iov_iter_get_pages was only able to get <
+ * blocksize worth of pages:
+ */
+ ret = -EFAULT;
+ goto err;
+ }
+
+ bch2_write_op_init(&dio->op, c, opts);
+ dio->op.end_io = sync
+ ? NULL
+ : bch2_dio_write_loop_async;
+ dio->op.target = dio->op.opts.foreground_target;
+ dio->op.write_point = writepoint_hashed((unsigned long) current);
+ dio->op.nr_replicas = dio->op.opts.data_replicas;
+ dio->op.subvol = inode->ei_subvol;
+ dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
+ dio->op.devs_need_flush = &inode->ei_devs_need_flush;
+
+ if (sync)
+ dio->op.flags |= BCH_WRITE_SYNC;
+ dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
+
+ ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
+ bio_sectors(bio), true);
+ if (unlikely(ret))
+ goto err;
+
+ ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
+ dio->op.opts.data_replicas, 0);
+ if (unlikely(ret) &&
+ !bch2_dio_write_check_allocated(dio))
+ goto err;
+
+ task_io_account_write(bio->bi_iter.bi_size);
+
+ if (unlikely(dio->iter.count) &&
+ !dio->sync &&
+ !dio->loop &&
+ bch2_dio_write_copy_iov(dio))
+ dio->sync = sync = true;
+
+ dio->loop = true;
+ closure_call(&dio->op.cl, bch2_write, NULL, NULL);
+
+ if (!sync)
+ return -EIOCBQUEUED;
+
+ bch2_dio_write_end(dio);
+
+ if (likely(!dio->iter.count) || dio->op.error)
+ break;
+
+ bio_reset(bio, NULL, REQ_OP_WRITE);
+ }
+out:
+ return bch2_dio_write_done(dio);
+err:
+ dio->op.error = ret;
+
+ bio_release_pages(bio, false);
+
+ bch2_quota_reservation_put(c, inode, &dio->quota_res);
+ goto out;
+}
+
+static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
+{
+ struct mm_struct *mm = dio->mm;
+
+ bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
+
+ if (mm)
+ kthread_use_mm(mm);
+ bch2_dio_write_loop(dio);
+ if (mm)
+ kthread_unuse_mm(mm);
+}
+
+static void bch2_dio_write_loop_async(struct bch_write_op *op)
+{
+ struct dio_write *dio = container_of(op, struct dio_write, op);
+
+ bch2_dio_write_end(dio);
+
+ if (likely(!dio->iter.count) || dio->op.error)
+ bch2_dio_write_done(dio);
+ else
+ bch2_dio_write_continue(dio);
+}
+
+ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
+{
+ struct file *file = req->ki_filp;
+ struct address_space *mapping = file->f_mapping;
+ struct bch_inode_info *inode = file_bch_inode(file);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct dio_write *dio;
+ struct bio *bio;
+ bool locked = true, extending;
+ ssize_t ret;
+
+ prefetch(&c->opts);
+ prefetch((void *) &c->opts + 64);
+ prefetch(&inode->ei_inode);
+ prefetch((void *) &inode->ei_inode + 64);
+
+ inode_lock(&inode->v);
+
+ ret = generic_write_checks(req, iter);
+ if (unlikely(ret <= 0))
+ goto err;
+
+ ret = file_remove_privs(file);
+ if (unlikely(ret))
+ goto err;
+
+ ret = file_update_time(file);
+ if (unlikely(ret))
+ goto err;
+
+ if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
+ goto err;
+
+ inode_dio_begin(&inode->v);
+ bch2_pagecache_block_get(inode);
+
+ extending = req->ki_pos + iter->count > inode->v.i_size;
+ if (!extending) {
+ inode_unlock(&inode->v);
+ locked = false;
+ }
+
+ bio = bio_alloc_bioset(NULL,
+ bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
+ REQ_OP_WRITE,
+ GFP_KERNEL,
+ &c->dio_write_bioset);
+ dio = container_of(bio, struct dio_write, op.wbio.bio);
+ dio->req = req;
+ dio->mapping = mapping;
+ dio->inode = inode;
+ dio->mm = current->mm;
+ dio->loop = false;
+ dio->extending = extending;
+ dio->sync = is_sync_kiocb(req) || extending;
+ dio->flush = iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
+ dio->free_iov = false;
+ dio->quota_res.sectors = 0;
+ dio->written = 0;
+ dio->iter = *iter;
+ dio->op.c = c;
+
+ if (unlikely(mapping->nrpages)) {
+ ret = bch2_write_invalidate_inode_pages_range(mapping,
+ req->ki_pos,
+ req->ki_pos + iter->count - 1);
+ if (unlikely(ret))
+ goto err_put_bio;
+ }
+
+ ret = bch2_dio_write_loop(dio);
+err:
+ if (locked)
+ inode_unlock(&inode->v);
+ return ret;
+err_put_bio:
+ bch2_pagecache_block_put(inode);
+ bio_put(bio);
+ inode_dio_end(&inode->v);
+ goto err;
+}
+
+void bch2_fs_fs_io_direct_exit(struct bch_fs *c)
+{
+ bioset_exit(&c->dio_write_bioset);
+ bioset_exit(&c->dio_read_bioset);
+}
+
+int bch2_fs_fs_io_direct_init(struct bch_fs *c)
+{
+ if (bioset_init(&c->dio_read_bioset,
+ 4, offsetof(struct dio_read, rbio.bio),
+ BIOSET_NEED_BVECS))
+ return -BCH_ERR_ENOMEM_dio_read_bioset_init;
+
+ if (bioset_init(&c->dio_write_bioset,
+ 4, offsetof(struct dio_write, op.wbio.bio),
+ BIOSET_NEED_BVECS))
+ return -BCH_ERR_ENOMEM_dio_write_bioset_init;
+
+ return 0;
+}
+
+#endif /* NO_BCACHEFS_FS */
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_FS_IO_DIRECT_H
+#define _BCACHEFS_FS_IO_DIRECT_H
+
+#ifndef NO_BCACHEFS_FS
+ssize_t bch2_direct_write(struct kiocb *, struct iov_iter *);
+ssize_t bch2_read_iter(struct kiocb *, struct iov_iter *);
+
+void bch2_fs_fs_io_direct_exit(struct bch_fs *);
+int bch2_fs_fs_io_direct_init(struct bch_fs *);
+#else
+static inline void bch2_fs_fs_io_direct_exit(struct bch_fs *c) {}
+static inline int bch2_fs_fs_io_direct_init(struct bch_fs *c) { return 0; }
+#endif
+
+#endif /* _BCACHEFS_FS_IO_DIRECT_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+#ifndef NO_BCACHEFS_FS
+
+#include "bcachefs.h"
+#include "btree_iter.h"
+#include "extents.h"
+#include "fs-io.h"
+#include "fs-io-pagecache.h"
+#include "subvolume.h"
+
+#include <linux/pagevec.h>
+#include <linux/writeback.h>
+
+int bch2_filemap_get_contig_folios_d(struct address_space *mapping,
+ loff_t start, u64 end,
+ fgf_t fgp_flags, gfp_t gfp,
+ folios *fs)
+{
+ struct folio *f;
+ u64 pos = start;
+ int ret = 0;
+
+ while (pos < end) {
+ if ((u64) pos >= (u64) start + (1ULL << 20))
+ fgp_flags &= ~FGP_CREAT;
+
+ ret = darray_make_room_gfp(fs, 1, gfp & GFP_KERNEL);
+ if (ret)
+ break;
+
+ f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp);
+ if (IS_ERR_OR_NULL(f))
+ break;
+
+ BUG_ON(fs->nr && folio_pos(f) != pos);
+
+ pos = folio_end_pos(f);
+ darray_push(fs, f);
+ }
+
+ if (!fs->nr && !ret && (fgp_flags & FGP_CREAT))
+ ret = -ENOMEM;
+
+ return fs->nr ? 0 : ret;
+}
+
+/* pagecache_block must be held */
+int bch2_write_invalidate_inode_pages_range(struct address_space *mapping,
+ loff_t start, loff_t end)
+{
+ int ret;
+
+ /*
+ * XXX: the way this is currently implemented, we can spin if a process
+ * is continually redirtying a specific page
+ */
+ do {
+ if (!mapping->nrpages)
+ return 0;
+
+ ret = filemap_write_and_wait_range(mapping, start, end);
+ if (ret)
+ break;
+
+ if (!mapping->nrpages)
+ return 0;
+
+ ret = invalidate_inode_pages2_range(mapping,
+ start >> PAGE_SHIFT,
+ end >> PAGE_SHIFT);
+ } while (ret == -EBUSY);
+
+ return ret;
+}
+
+#if 0
+/* Useful for debug tracing: */
+static const char * const bch2_folio_sector_states[] = {
+#define x(n) #n,
+ BCH_FOLIO_SECTOR_STATE()
+#undef x
+ NULL
+};
+#endif
+
+static inline enum bch_folio_sector_state
+folio_sector_dirty(enum bch_folio_sector_state state)
+{
+ switch (state) {
+ case SECTOR_unallocated:
+ return SECTOR_dirty;
+ case SECTOR_reserved:
+ return SECTOR_dirty_reserved;
+ default:
+ return state;
+ }
+}
+
+static inline enum bch_folio_sector_state
+folio_sector_undirty(enum bch_folio_sector_state state)
+{
+ switch (state) {
+ case SECTOR_dirty:
+ return SECTOR_unallocated;
+ case SECTOR_dirty_reserved:
+ return SECTOR_reserved;
+ default:
+ return state;
+ }
+}
+
+static inline enum bch_folio_sector_state
+folio_sector_reserve(enum bch_folio_sector_state state)
+{
+ switch (state) {
+ case SECTOR_unallocated:
+ return SECTOR_reserved;
+ case SECTOR_dirty:
+ return SECTOR_dirty_reserved;
+ default:
+ return state;
+ }
+}
+
+/* for newly allocated folios: */
+struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp)
+{
+ struct bch_folio *s;
+
+ s = kzalloc(sizeof(*s) +
+ sizeof(struct bch_folio_sector) *
+ folio_sectors(folio), gfp);
+ if (!s)
+ return NULL;
+
+ spin_lock_init(&s->lock);
+ folio_attach_private(folio, s);
+ return s;
+}
+
+struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp)
+{
+ return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp);
+}
+
+static unsigned bkey_to_sector_state(struct bkey_s_c k)
+{
+ if (bkey_extent_is_reservation(k))
+ return SECTOR_reserved;
+ if (bkey_extent_is_allocation(k.k))
+ return SECTOR_allocated;
+ return SECTOR_unallocated;
+}
+
+static void __bch2_folio_set(struct folio *folio,
+ unsigned pg_offset, unsigned pg_len,
+ unsigned nr_ptrs, unsigned state)
+{
+ struct bch_folio *s = bch2_folio(folio);
+ unsigned i, sectors = folio_sectors(folio);
+
+ BUG_ON(pg_offset >= sectors);
+ BUG_ON(pg_offset + pg_len > sectors);
+
+ spin_lock(&s->lock);
+
+ for (i = pg_offset; i < pg_offset + pg_len; i++) {
+ s->s[i].nr_replicas = nr_ptrs;
+ bch2_folio_sector_set(folio, s, i, state);
+ }
+
+ if (i == sectors)
+ s->uptodate = true;
+
+ spin_unlock(&s->lock);
+}
+
+/*
+ * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the
+ * extents btree:
+ */
+int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
+ struct folio **fs, unsigned nr_folios)
+{
+ struct btree_trans *trans;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ struct bch_folio *s;
+ u64 offset = folio_sector(fs[0]);
+ unsigned folio_idx;
+ u32 snapshot;
+ bool need_set = false;
+ int ret;
+
+ for (folio_idx = 0; folio_idx < nr_folios; folio_idx++) {
+ s = bch2_folio_create(fs[folio_idx], GFP_KERNEL);
+ if (!s)
+ return -ENOMEM;
+
+ need_set |= !s->uptodate;
+ }
+
+ if (!need_set)
+ return 0;
+
+ folio_idx = 0;
+ trans = bch2_trans_get(c);
+retry:
+ bch2_trans_begin(trans);
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (ret)
+ goto err;
+
+ for_each_btree_key_norestart(trans, iter, BTREE_ID_extents,
+ SPOS(inum.inum, offset, snapshot),
+ BTREE_ITER_SLOTS, k, ret) {
+ unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
+ unsigned state = bkey_to_sector_state(k);
+
+ while (folio_idx < nr_folios) {
+ struct folio *folio = fs[folio_idx];
+ u64 folio_start = folio_sector(folio);
+ u64 folio_end = folio_end_sector(folio);
+ unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) -
+ folio_start;
+ unsigned folio_len = min(k.k->p.offset, folio_end) -
+ folio_offset - folio_start;
+
+ BUG_ON(k.k->p.offset < folio_start);
+ BUG_ON(bkey_start_offset(k.k) > folio_end);
+
+ if (!bch2_folio(folio)->uptodate)
+ __bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state);
+
+ if (k.k->p.offset < folio_end)
+ break;
+ folio_idx++;
+ }
+
+ if (folio_idx == nr_folios)
+ break;
+ }
+
+ offset = iter.pos.offset;
+ bch2_trans_iter_exit(trans, &iter);
+err:
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ goto retry;
+ bch2_trans_put(trans);
+
+ return ret;
+}
+
+void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
+{
+ struct bvec_iter iter;
+ struct folio_vec fv;
+ unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
+ ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
+ unsigned state = bkey_to_sector_state(k);
+
+ bio_for_each_folio(fv, bio, iter)
+ __bch2_folio_set(fv.fv_folio,
+ fv.fv_offset >> 9,
+ fv.fv_len >> 9,
+ nr_ptrs, state);
+}
+
+void bch2_mark_pagecache_unallocated(struct bch_inode_info *inode,
+ u64 start, u64 end)
+{
+ pgoff_t index = start >> PAGE_SECTORS_SHIFT;
+ pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
+ struct folio_batch fbatch;
+ unsigned i, j;
+
+ if (end <= start)
+ return;
+
+ folio_batch_init(&fbatch);
+
+ while (filemap_get_folios(inode->v.i_mapping,
+ &index, end_index, &fbatch)) {
+ for (i = 0; i < folio_batch_count(&fbatch); i++) {
+ struct folio *folio = fbatch.folios[i];
+ u64 folio_start = folio_sector(folio);
+ u64 folio_end = folio_end_sector(folio);
+ unsigned folio_offset = max(start, folio_start) - folio_start;
+ unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
+ struct bch_folio *s;
+
+ BUG_ON(end <= folio_start);
+
+ folio_lock(folio);
+ s = bch2_folio(folio);
+
+ if (s) {
+ spin_lock(&s->lock);
+ for (j = folio_offset; j < folio_offset + folio_len; j++)
+ s->s[j].nr_replicas = 0;
+ spin_unlock(&s->lock);
+ }
+
+ folio_unlock(folio);
+ }
+ folio_batch_release(&fbatch);
+ cond_resched();
+ }
+}
+
+void bch2_mark_pagecache_reserved(struct bch_inode_info *inode,
+ u64 start, u64 end)
+{
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ pgoff_t index = start >> PAGE_SECTORS_SHIFT;
+ pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
+ struct folio_batch fbatch;
+ s64 i_sectors_delta = 0;
+ unsigned i, j;
+
+ if (end <= start)
+ return;
+
+ folio_batch_init(&fbatch);
+
+ while (filemap_get_folios(inode->v.i_mapping,
+ &index, end_index, &fbatch)) {
+ for (i = 0; i < folio_batch_count(&fbatch); i++) {
+ struct folio *folio = fbatch.folios[i];
+ u64 folio_start = folio_sector(folio);
+ u64 folio_end = folio_end_sector(folio);
+ unsigned folio_offset = max(start, folio_start) - folio_start;
+ unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
+ struct bch_folio *s;
+
+ BUG_ON(end <= folio_start);
+
+ folio_lock(folio);
+ s = bch2_folio(folio);
+
+ if (s) {
+ spin_lock(&s->lock);
+ for (j = folio_offset; j < folio_offset + folio_len; j++) {
+ i_sectors_delta -= s->s[j].state == SECTOR_dirty;
+ bch2_folio_sector_set(folio, s, j,
+ folio_sector_reserve(s->s[j].state));
+ }
+ spin_unlock(&s->lock);
+ }
+
+ folio_unlock(folio);
+ }
+ folio_batch_release(&fbatch);
+ cond_resched();
+ }
+
+ bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
+}
+
+static inline unsigned sectors_to_reserve(struct bch_folio_sector *s,
+ unsigned nr_replicas)
+{
+ return max(0, (int) nr_replicas -
+ s->nr_replicas -
+ s->replicas_reserved);
+}
+
+int bch2_get_folio_disk_reservation(struct bch_fs *c,
+ struct bch_inode_info *inode,
+ struct folio *folio, bool check_enospc)
+{
+ struct bch_folio *s = bch2_folio_create(folio, 0);
+ unsigned nr_replicas = inode_nr_replicas(c, inode);
+ struct disk_reservation disk_res = { 0 };
+ unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0;
+ int ret;
+
+ if (!s)
+ return -ENOMEM;
+
+ for (i = 0; i < sectors; i++)
+ disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
+
+ if (!disk_res_sectors)
+ return 0;
+
+ ret = bch2_disk_reservation_get(c, &disk_res,
+ disk_res_sectors, 1,
+ !check_enospc
+ ? BCH_DISK_RESERVATION_NOFAIL
+ : 0);
+ if (unlikely(ret))
+ return ret;
+
+ for (i = 0; i < sectors; i++)
+ s->s[i].replicas_reserved +=
+ sectors_to_reserve(&s->s[i], nr_replicas);
+
+ return 0;
+}
+
+void bch2_folio_reservation_put(struct bch_fs *c,
+ struct bch_inode_info *inode,
+ struct bch2_folio_reservation *res)
+{
+ bch2_disk_reservation_put(c, &res->disk);
+ bch2_quota_reservation_put(c, inode, &res->quota);
+}
+
+int bch2_folio_reservation_get(struct bch_fs *c,
+ struct bch_inode_info *inode,
+ struct folio *folio,
+ struct bch2_folio_reservation *res,
+ unsigned offset, unsigned len)
+{
+ struct bch_folio *s = bch2_folio_create(folio, 0);
+ unsigned i, disk_sectors = 0, quota_sectors = 0;
+ int ret;
+
+ if (!s)
+ return -ENOMEM;
+
+ BUG_ON(!s->uptodate);
+
+ for (i = round_down(offset, block_bytes(c)) >> 9;
+ i < round_up(offset + len, block_bytes(c)) >> 9;
+ i++) {
+ disk_sectors += sectors_to_reserve(&s->s[i],
+ res->disk.nr_replicas);
+ quota_sectors += s->s[i].state == SECTOR_unallocated;
+ }
+
+ if (disk_sectors) {
+ ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
+ if (unlikely(ret))
+ return ret;
+ }
+
+ if (quota_sectors) {
+ ret = bch2_quota_reservation_add(c, inode, &res->quota,
+ quota_sectors, true);
+ if (unlikely(ret)) {
+ struct disk_reservation tmp = {
+ .sectors = disk_sectors
+ };
+
+ bch2_disk_reservation_put(c, &tmp);
+ res->disk.sectors -= disk_sectors;
+ return ret;
+ }
+ }
+
+ return 0;
+}
+
+static void bch2_clear_folio_bits(struct folio *folio)
+{
+ struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch_folio *s = bch2_folio(folio);
+ struct disk_reservation disk_res = { 0 };
+ int i, sectors = folio_sectors(folio), dirty_sectors = 0;
+
+ if (!s)
+ return;
+
+ EBUG_ON(!folio_test_locked(folio));
+ EBUG_ON(folio_test_writeback(folio));
+
+ for (i = 0; i < sectors; i++) {
+ disk_res.sectors += s->s[i].replicas_reserved;
+ s->s[i].replicas_reserved = 0;
+
+ dirty_sectors -= s->s[i].state == SECTOR_dirty;
+ bch2_folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state));
+ }
+
+ bch2_disk_reservation_put(c, &disk_res);
+
+ bch2_i_sectors_acct(c, inode, NULL, dirty_sectors);
+
+ bch2_folio_release(folio);
+}
+
+void bch2_set_folio_dirty(struct bch_fs *c,
+ struct bch_inode_info *inode,
+ struct folio *folio,
+ struct bch2_folio_reservation *res,
+ unsigned offset, unsigned len)
+{
+ struct bch_folio *s = bch2_folio(folio);
+ unsigned i, dirty_sectors = 0;
+
+ WARN_ON((u64) folio_pos(folio) + offset + len >
+ round_up((u64) i_size_read(&inode->v), block_bytes(c)));
+
+ BUG_ON(!s->uptodate);
+
+ spin_lock(&s->lock);
+
+ for (i = round_down(offset, block_bytes(c)) >> 9;
+ i < round_up(offset + len, block_bytes(c)) >> 9;
+ i++) {
+ unsigned sectors = sectors_to_reserve(&s->s[i],
+ res->disk.nr_replicas);
+
+ /*
+ * This can happen if we race with the error path in
+ * bch2_writepage_io_done():
+ */
+ sectors = min_t(unsigned, sectors, res->disk.sectors);
+
+ s->s[i].replicas_reserved += sectors;
+ res->disk.sectors -= sectors;
+
+ dirty_sectors += s->s[i].state == SECTOR_unallocated;
+
+ bch2_folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state));
+ }
+
+ spin_unlock(&s->lock);
+
+ bch2_i_sectors_acct(c, inode, &res->quota, dirty_sectors);
+
+ if (!folio_test_dirty(folio))
+ filemap_dirty_folio(inode->v.i_mapping, folio);
+}
+
+vm_fault_t bch2_page_fault(struct vm_fault *vmf)
+{
+ struct file *file = vmf->vma->vm_file;
+ struct address_space *mapping = file->f_mapping;
+ struct address_space *fdm = faults_disabled_mapping();
+ struct bch_inode_info *inode = file_bch_inode(file);
+ vm_fault_t ret;
+
+ if (fdm == mapping)
+ return VM_FAULT_SIGBUS;
+
+ /* Lock ordering: */
+ if (fdm > mapping) {
+ struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
+
+ if (bch2_pagecache_add_tryget(inode))
+ goto got_lock;
+
+ bch2_pagecache_block_put(fdm_host);
+
+ bch2_pagecache_add_get(inode);
+ bch2_pagecache_add_put(inode);
+
+ bch2_pagecache_block_get(fdm_host);
+
+ /* Signal that lock has been dropped: */
+ set_fdm_dropped_locks();
+ return VM_FAULT_SIGBUS;
+ }
+
+ bch2_pagecache_add_get(inode);
+got_lock:
+ ret = filemap_fault(vmf);
+ bch2_pagecache_add_put(inode);
+
+ return ret;
+}
+
+vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
+{
+ struct folio *folio = page_folio(vmf->page);
+ struct file *file = vmf->vma->vm_file;
+ struct bch_inode_info *inode = file_bch_inode(file);
+ struct address_space *mapping = file->f_mapping;
+ struct bch_fs *c = inode->v.i_sb->s_fs_info;
+ struct bch2_folio_reservation res;
+ unsigned len;
+ loff_t isize;
+ vm_fault_t ret;
+
+ bch2_folio_reservation_init(c, inode, &res);
+
+ sb_start_pagefault(inode->v.i_sb);
+ file_update_time(file);
+
+ /*
+ * Not strictly necessary, but helps avoid dio writes livelocking in
+ * bch2_write_invalidate_inode_pages_range() - can drop this if/when we get
+ * a bch2_write_invalidate_inode_pages_range() that works without dropping
+ * page lock before invalidating page
+ */
+ bch2_pagecache_add_get(inode);
+
+ folio_lock(folio);
+ isize = i_size_read(&inode->v);
+
+ if (folio->mapping != mapping || folio_pos(folio) >= isize) {
+ folio_unlock(folio);
+ ret = VM_FAULT_NOPAGE;
+ goto out;
+ }
+
+ len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
+
+ if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?:
+ bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
+ folio_unlock(folio);
+ ret = VM_FAULT_SIGBUS;
+ goto out;
+ }
+
+ bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
+ bch2_folio_reservation_put(c, inode, &res);
+
+ folio_wait_stable(folio);
+ ret = VM_FAULT_LOCKED;
+out:
+ bch2_pagecache_add_put(inode);
+ sb_end_pagefault(inode->v.i_sb);
+
+ return ret;
+}
+
+void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
+{
+ if (offset || length < folio_size(folio))
+ return;
+
+ bch2_clear_folio_bits(folio);
+}
+
+bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
+{
+ if (folio_test_dirty(folio) || folio_test_writeback(folio))
+ return false;
+
+ bch2_clear_folio_bits(folio);
+ return true;
+}
+
+/* fseek: */
+
+static int folio_data_offset(struct folio *folio, loff_t pos,
+ unsigned min_replicas)
+{
+ struct bch_folio *s = bch2_folio(folio);
+ unsigned i, sectors = folio_sectors(folio);
+
+ if (s)
+ for (i = folio_pos_to_s(folio, pos); i < sectors; i++)
+ if (s->s[i].state >= SECTOR_dirty &&
+ s->s[i].nr_replicas + s->s[i].replicas_reserved >= min_replicas)
+ return i << SECTOR_SHIFT;
+
+ return -1;
+}
+
+loff_t bch2_seek_pagecache_data(struct inode *vinode,
+ loff_t start_offset,
+ loff_t end_offset,
+ unsigned min_replicas,
+ bool nonblock)
+{
+ struct folio_batch fbatch;
+ pgoff_t start_index = start_offset >> PAGE_SHIFT;
+ pgoff_t end_index = end_offset >> PAGE_SHIFT;
+ pgoff_t index = start_index;
+ unsigned i;
+ loff_t ret;
+ int offset;
+
+ folio_batch_init(&fbatch);
+
+ while (filemap_get_folios(vinode->i_mapping,
+ &index, end_index, &fbatch)) {
+ for (i = 0; i < folio_batch_count(&fbatch); i++) {
+ struct folio *folio = fbatch.folios[i];
+
+ if (!nonblock) {
+ folio_lock(folio);
+ } else if (!folio_trylock(folio)) {
+ folio_batch_release(&fbatch);
+ return -EAGAIN;
+ }
+
+ offset = folio_data_offset(folio,
+ max(folio_pos(folio), start_offset),
+ min_replicas);
+ if (offset >= 0) {
+ ret = clamp(folio_pos(folio) + offset,
+ start_offset, end_offset);
+ folio_unlock(folio);
+ folio_batch_release(&fbatch);
+ return ret;
+ }
+ folio_unlock(folio);
+ }
+ folio_batch_release(&fbatch);
+ cond_resched();
+ }
+
+ return end_offset;
+}
+
+/*
+ * Search for a hole in a folio.
+ *
+ * The filemap layer returns -ENOENT if no folio exists, so reuse the same error
+ * code to indicate a pagecache hole exists at the returned offset. Otherwise
+ * return 0 if the folio is filled with data, or an error code. This function
+ * can return -EAGAIN if nonblock is specified.
+ */
+static int folio_hole_offset(struct address_space *mapping, loff_t *offset,
+ unsigned min_replicas, bool nonblock)
+{
+ struct folio *folio;
+ struct bch_folio *s;
+ unsigned i, sectors;
+ int ret = -ENOENT;
+
+ folio = __filemap_get_folio(mapping, *offset >> PAGE_SHIFT,
+ FGP_LOCK|(nonblock ? FGP_NOWAIT : 0), 0);
+ if (IS_ERR(folio))
+ return PTR_ERR(folio);
+
+ s = bch2_folio(folio);
+ if (!s)
+ goto unlock;
+
+ sectors = folio_sectors(folio);
+ for (i = folio_pos_to_s(folio, *offset); i < sectors; i++)
+ if (s->s[i].state < SECTOR_dirty ||
+ s->s[i].nr_replicas + s->s[i].replicas_reserved < min_replicas) {
+ *offset = max(*offset,
+ folio_pos(folio) + (i << SECTOR_SHIFT));
+ goto unlock;
+ }
+
+ *offset = folio_end_pos(folio);
+ ret = 0;
+unlock:
+ folio_unlock(folio);
+ folio_put(folio);
+ return ret;
+}
+
+loff_t bch2_seek_pagecache_hole(struct inode *vinode,
+ loff_t start_offset,
+ loff_t end_offset,
+ unsigned min_replicas,
+ bool nonblock)
+{
+ struct address_space *mapping = vinode->i_mapping;
+ loff_t offset = start_offset;
+ loff_t ret = 0;
+
+ while (!ret && offset < end_offset)
+ ret = folio_hole_offset(mapping, &offset, min_replicas, nonblock);
+
+ if (ret && ret != -ENOENT)
+ return ret;
+ return min(offset, end_offset);
+}
+
+int bch2_clamp_data_hole(struct inode *inode,
+ u64 *hole_start,
+ u64 *hole_end,
+ unsigned min_replicas,
+ bool nonblock)
+{
+ loff_t ret;
+
+ ret = bch2_seek_pagecache_hole(inode,
+ *hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
+ if (ret < 0)
+ return ret;
+
+ *hole_start = ret;
+
+ if (*hole_start == *hole_end)
+ return 0;
+
+ ret = bch2_seek_pagecache_data(inode,
+ *hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
+ if (ret < 0)
+ return ret;
+
+ *hole_end = ret;
+ return 0;
+}
+
+#endif /* NO_BCACHEFS_FS */
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_FS_IO_PAGECACHE_H
+#define _BCACHEFS_FS_IO_PAGECACHE_H
+
+#include <linux/pagemap.h>
+
+typedef DARRAY(struct folio *) folios;
+
+int bch2_filemap_get_contig_folios_d(struct address_space *, loff_t,
+ u64, fgf_t, gfp_t, folios *);
+int bch2_write_invalidate_inode_pages_range(struct address_space *, loff_t, loff_t);
+
+/*
+ * Use u64 for the end pos and sector helpers because if the folio covers the
+ * max supported range of the mapping, the start offset of the next folio
+ * overflows loff_t. This breaks much of the range based processing in the
+ * buffered write path.
+ */
+static inline u64 folio_end_pos(struct folio *folio)
+{
+ return folio_pos(folio) + folio_size(folio);
+}
+
+static inline size_t folio_sectors(struct folio *folio)
+{
+ return PAGE_SECTORS << folio_order(folio);
+}
+
+static inline loff_t folio_sector(struct folio *folio)
+{
+ return folio_pos(folio) >> 9;
+}
+
+static inline u64 folio_end_sector(struct folio *folio)
+{
+ return folio_end_pos(folio) >> 9;
+}
+
+#define BCH_FOLIO_SECTOR_STATE() \
+ x(unallocated) \
+ x(reserved) \
+ x(dirty) \
+ x(dirty_reserved) \
+ x(allocated)
+
+enum bch_folio_sector_state {
+#define x(n) SECTOR_##n,
+ BCH_FOLIO_SECTOR_STATE()
+#undef x
+};
+
+struct bch_folio_sector {
+ /* Uncompressed, fully allocated replicas (or on disk reservation): */
+ unsigned nr_replicas:4;
+
+ /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
+ unsigned replicas_reserved:4;
+
+ /* i_sectors: */
+ enum bch_folio_sector_state state:8;
+};
+
+struct bch_folio {
+ spinlock_t lock;
+ atomic_t write_count;
+ /*
+ * Is the sector state up to date with the btree?
+ * (Not the data itself)
+ */
+ bool uptodate;
+ struct bch_folio_sector s[];
+};
+
+/* Helper for when we need to add debug instrumentation: */
+static inline void bch2_folio_sector_set(struct folio *folio,
+ struct bch_folio *s,
+ unsigned i, unsigned n)
+{
+ s->s[i].state = n;
+}
+
+/* file offset (to folio offset) to bch_folio_sector index */
+static inline int folio_pos_to_s(struct folio *folio, loff_t pos)
+{
+ u64 f_offset = pos - folio_pos(folio);
+
+ BUG_ON(pos < folio_pos(folio) || pos >= folio_end_pos(folio));
+ return f_offset >> SECTOR_SHIFT;
+}
+
+/* for newly allocated folios: */
+static inline void __bch2_folio_release(struct folio *folio)
+{
+ kfree(folio_detach_private(folio));
+}
+
+static inline void bch2_folio_release(struct folio *folio)
+{
+ EBUG_ON(!folio_test_locked(folio));
+ __bch2_folio_release(folio);
+}
+
+static inline struct bch_folio *__bch2_folio(struct folio *folio)
+{
+ return folio_has_private(folio)
+ ? (struct bch_folio *) folio_get_private(folio)
+ : NULL;
+}
+
+static inline struct bch_folio *bch2_folio(struct folio *folio)
+{
+ EBUG_ON(!folio_test_locked(folio));
+
+ return __bch2_folio(folio);
+}
+
+struct bch_folio *__bch2_folio_create(struct folio *, gfp_t);
+struct bch_folio *bch2_folio_create(struct folio *, gfp_t);
+
+struct bch2_folio_reservation {
+ struct disk_reservation disk;
+ struct quota_res quota;
+};
+
+static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
+{
+ /* XXX: this should not be open coded */
+ return inode->ei_inode.bi_data_replicas
+ ? inode->ei_inode.bi_data_replicas - 1
+ : c->opts.data_replicas;
+}
+
+static inline void bch2_folio_reservation_init(struct bch_fs *c,
+ struct bch_inode_info *inode,
+ struct bch2_folio_reservation *res)
+{
+ memset(res, 0, sizeof(*res));
+
+ res->disk.nr_replicas = inode_nr_replicas(c, inode);
+}
+
+int bch2_folio_set(struct bch_fs *, subvol_inum, struct folio **, unsigned);
+void bch2_bio_page_state_set(struct bio *, struct bkey_s_c);
+
+void bch2_mark_pagecache_unallocated(struct bch_inode_info *, u64, u64);
+void bch2_mark_pagecache_reserved(struct bch_inode_info *, u64, u64);
+
+int bch2_get_folio_disk_reservation(struct bch_fs *,
+ struct bch_inode_info *,
+ struct folio *, bool);
+
+void bch2_folio_reservation_put(struct bch_fs *,
+ struct bch_inode_info *,
+ struct bch2_folio_reservation *);
+int bch2_folio_reservation_get(struct bch_fs *,
+ struct bch_inode_info *,
+ struct folio *,
+ struct bch2_folio_reservation *,
+ unsigned, unsigned);
+
+void bch2_set_folio_dirty(struct bch_fs *,
+ struct bch_inode_info *,
+ struct folio *,
+ struct bch2_folio_reservation *,
+ unsigned, unsigned);
+
+vm_fault_t bch2_page_fault(struct vm_fault *);
+vm_fault_t bch2_page_mkwrite(struct vm_fault *);
+void bch2_invalidate_folio(struct folio *, size_t, size_t);
+bool bch2_release_folio(struct folio *, gfp_t);
+
+loff_t bch2_seek_pagecache_data(struct inode *, loff_t, loff_t, unsigned, bool);
+loff_t bch2_seek_pagecache_hole(struct inode *, loff_t, loff_t, unsigned, bool);
+int bch2_clamp_data_hole(struct inode *, u64 *, u64 *, unsigned, bool);
+
+#endif /* _BCACHEFS_FS_IO_PAGECACHE_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * io_misc.c - fallocate, fpunch, truncate:
+ */
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "bkey_buf.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "clock.h"
+#include "error.h"
+#include "extents.h"
+#include "extent_update.h"
+#include "inode.h"
+#include "io_misc.h"
+#include "io_write.h"
+#include "logged_ops.h"
+#include "rebalance.h"
+#include "subvolume.h"
+
+/* Overwrites whatever was present with zeroes: */
+int bch2_extent_fallocate(struct btree_trans *trans,
+ subvol_inum inum,
+ struct btree_iter *iter,
+ u64 sectors,
+ struct bch_io_opts opts,
+ s64 *i_sectors_delta,
+ struct write_point_specifier write_point)
+{
+ struct bch_fs *c = trans->c;
+ struct disk_reservation disk_res = { 0 };
+ struct closure cl;
+ struct open_buckets open_buckets = { 0 };
+ struct bkey_s_c k;
+ struct bkey_buf old, new;
+ unsigned sectors_allocated = 0;
+ bool have_reservation = false;
+ bool unwritten = opts.nocow &&
+ c->sb.version >= bcachefs_metadata_version_unwritten_extents;
+ int ret;
+
+ bch2_bkey_buf_init(&old);
+ bch2_bkey_buf_init(&new);
+ closure_init_stack(&cl);
+
+ k = bch2_btree_iter_peek_slot(iter);
+ ret = bkey_err(k);
+ if (ret)
+ return ret;
+
+ sectors = min_t(u64, sectors, k.k->p.offset - iter->pos.offset);
+
+ if (!have_reservation) {
+ unsigned new_replicas =
+ max(0, (int) opts.data_replicas -
+ (int) bch2_bkey_nr_ptrs_fully_allocated(k));
+ /*
+ * Get a disk reservation before (in the nocow case) calling
+ * into the allocator:
+ */
+ ret = bch2_disk_reservation_get(c, &disk_res, sectors, new_replicas, 0);
+ if (unlikely(ret))
+ goto err;
+
+ bch2_bkey_buf_reassemble(&old, c, k);
+ }
+
+ if (have_reservation) {
+ if (!bch2_extents_match(k, bkey_i_to_s_c(old.k)))
+ goto err;
+
+ bch2_key_resize(&new.k->k, sectors);
+ } else if (!unwritten) {
+ struct bkey_i_reservation *reservation;
+
+ bch2_bkey_buf_realloc(&new, c, sizeof(*reservation) / sizeof(u64));
+ reservation = bkey_reservation_init(new.k);
+ reservation->k.p = iter->pos;
+ bch2_key_resize(&reservation->k, sectors);
+ reservation->v.nr_replicas = opts.data_replicas;
+ } else {
+ struct bkey_i_extent *e;
+ struct bch_devs_list devs_have;
+ struct write_point *wp;
+ struct bch_extent_ptr *ptr;
+
+ devs_have.nr = 0;
+
+ bch2_bkey_buf_realloc(&new, c, BKEY_EXTENT_U64s_MAX);
+
+ e = bkey_extent_init(new.k);
+ e->k.p = iter->pos;
+
+ ret = bch2_alloc_sectors_start_trans(trans,
+ opts.foreground_target,
+ false,
+ write_point,
+ &devs_have,
+ opts.data_replicas,
+ opts.data_replicas,
+ BCH_WATERMARK_normal, 0, &cl, &wp);
+ if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
+ ret = -BCH_ERR_transaction_restart_nested;
+ if (ret)
+ goto err;
+
+ sectors = min_t(u64, sectors, wp->sectors_free);
+ sectors_allocated = sectors;
+
+ bch2_key_resize(&e->k, sectors);
+
+ bch2_open_bucket_get(c, wp, &open_buckets);
+ bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
+ bch2_alloc_sectors_done(c, wp);
+
+ extent_for_each_ptr(extent_i_to_s(e), ptr)
+ ptr->unwritten = true;
+ }
+
+ have_reservation = true;
+
+ ret = bch2_extent_update(trans, inum, iter, new.k, &disk_res,
+ 0, i_sectors_delta, true);
+err:
+ if (!ret && sectors_allocated)
+ bch2_increment_clock(c, sectors_allocated, WRITE);
+
+ bch2_open_buckets_put(c, &open_buckets);
+ bch2_disk_reservation_put(c, &disk_res);
+ bch2_bkey_buf_exit(&new, c);
+ bch2_bkey_buf_exit(&old, c);
+
+ if (closure_nr_remaining(&cl) != 1) {
+ bch2_trans_unlock(trans);
+ closure_sync(&cl);
+ }
+
+ return ret;
+}
+
+/*
+ * Returns -BCH_ERR_transacton_restart if we had to drop locks:
+ */
+int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
+ subvol_inum inum, u64 end,
+ s64 *i_sectors_delta)
+{
+ struct bch_fs *c = trans->c;
+ unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
+ struct bpos end_pos = POS(inum.inum, end);
+ struct bkey_s_c k;
+ int ret = 0, ret2 = 0;
+ u32 snapshot;
+
+ while (!ret ||
+ bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
+ struct disk_reservation disk_res =
+ bch2_disk_reservation_init(c, 0);
+ struct bkey_i delete;
+
+ if (ret)
+ ret2 = ret;
+
+ bch2_trans_begin(trans);
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (ret)
+ continue;
+
+ bch2_btree_iter_set_snapshot(iter, snapshot);
+
+ /*
+ * peek_upto() doesn't have ideal semantics for extents:
+ */
+ k = bch2_btree_iter_peek_upto(iter, end_pos);
+ if (!k.k)
+ break;
+
+ ret = bkey_err(k);
+ if (ret)
+ continue;
+
+ bkey_init(&delete.k);
+ delete.k.p = iter->pos;
+
+ /* create the biggest key we can */
+ bch2_key_resize(&delete.k, max_sectors);
+ bch2_cut_back(end_pos, &delete);
+
+ ret = bch2_extent_update(trans, inum, iter, &delete,
+ &disk_res, 0, i_sectors_delta, false);
+ bch2_disk_reservation_put(c, &disk_res);
+ }
+
+ return ret ?: ret2;
+}
+
+int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
+ s64 *i_sectors_delta)
+{
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct btree_iter iter;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+ POS(inum.inum, start),
+ BTREE_ITER_INTENT);
+
+ ret = bch2_fpunch_at(trans, &iter, inum, end, i_sectors_delta);
+
+ bch2_trans_iter_exit(trans, &iter);
+ bch2_trans_put(trans);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ ret = 0;
+
+ return ret;
+}
+
+/* truncate: */
+
+void bch2_logged_op_truncate_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
+{
+ struct bkey_s_c_logged_op_truncate op = bkey_s_c_to_logged_op_truncate(k);
+
+ prt_printf(out, "subvol=%u", le32_to_cpu(op.v->subvol));
+ prt_printf(out, " inum=%llu", le64_to_cpu(op.v->inum));
+ prt_printf(out, " new_i_size=%llu", le64_to_cpu(op.v->new_i_size));
+}
+
+static int truncate_set_isize(struct btree_trans *trans,
+ subvol_inum inum,
+ u64 new_i_size)
+{
+ struct btree_iter iter = { NULL };
+ struct bch_inode_unpacked inode_u;
+ int ret;
+
+ ret = bch2_inode_peek(trans, &iter, &inode_u, inum, BTREE_ITER_INTENT) ?:
+ (inode_u.bi_size = new_i_size, 0) ?:
+ bch2_inode_write(trans, &iter, &inode_u);
+
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static int __bch2_resume_logged_op_truncate(struct btree_trans *trans,
+ struct bkey_i *op_k,
+ u64 *i_sectors_delta)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter fpunch_iter;
+ struct bkey_i_logged_op_truncate *op = bkey_i_to_logged_op_truncate(op_k);
+ subvol_inum inum = { le32_to_cpu(op->v.subvol), le64_to_cpu(op->v.inum) };
+ u64 new_i_size = le64_to_cpu(op->v.new_i_size);
+ int ret;
+
+ ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ truncate_set_isize(trans, inum, new_i_size));
+ if (ret)
+ goto err;
+
+ bch2_trans_iter_init(trans, &fpunch_iter, BTREE_ID_extents,
+ POS(inum.inum, round_up(new_i_size, block_bytes(c)) >> 9),
+ BTREE_ITER_INTENT);
+ ret = bch2_fpunch_at(trans, &fpunch_iter, inum, U64_MAX, i_sectors_delta);
+ bch2_trans_iter_exit(trans, &fpunch_iter);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ ret = 0;
+err:
+ bch2_logged_op_finish(trans, op_k);
+ return ret;
+}
+
+int bch2_resume_logged_op_truncate(struct btree_trans *trans, struct bkey_i *op_k)
+{
+ return __bch2_resume_logged_op_truncate(trans, op_k, NULL);
+}
+
+int bch2_truncate(struct bch_fs *c, subvol_inum inum, u64 new_i_size, u64 *i_sectors_delta)
+{
+ struct bkey_i_logged_op_truncate op;
+
+ bkey_logged_op_truncate_init(&op.k_i);
+ op.v.subvol = cpu_to_le32(inum.subvol);
+ op.v.inum = cpu_to_le64(inum.inum);
+ op.v.new_i_size = cpu_to_le64(new_i_size);
+
+ /*
+ * Logged ops aren't atomic w.r.t. snapshot creation: creating a
+ * snapshot while they're in progress, then crashing, will result in the
+ * resume only proceeding in one of the snapshots
+ */
+ down_read(&c->snapshot_create_lock);
+ int ret = bch2_trans_run(c,
+ bch2_logged_op_start(trans, &op.k_i) ?:
+ __bch2_resume_logged_op_truncate(trans, &op.k_i, i_sectors_delta));
+ up_read(&c->snapshot_create_lock);
+
+ return ret;
+}
+
+/* finsert/fcollapse: */
+
+void bch2_logged_op_finsert_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
+{
+ struct bkey_s_c_logged_op_finsert op = bkey_s_c_to_logged_op_finsert(k);
+
+ prt_printf(out, "subvol=%u", le32_to_cpu(op.v->subvol));
+ prt_printf(out, " inum=%llu", le64_to_cpu(op.v->inum));
+ prt_printf(out, " dst_offset=%lli", le64_to_cpu(op.v->dst_offset));
+ prt_printf(out, " src_offset=%llu", le64_to_cpu(op.v->src_offset));
+}
+
+static int adjust_i_size(struct btree_trans *trans, subvol_inum inum, u64 offset, s64 len)
+{
+ struct btree_iter iter;
+ struct bch_inode_unpacked inode_u;
+ int ret;
+
+ offset <<= 9;
+ len <<= 9;
+
+ ret = bch2_inode_peek(trans, &iter, &inode_u, inum, BTREE_ITER_INTENT);
+ if (ret)
+ return ret;
+
+ if (len > 0) {
+ if (MAX_LFS_FILESIZE - inode_u.bi_size < len) {
+ ret = -EFBIG;
+ goto err;
+ }
+
+ if (offset >= inode_u.bi_size) {
+ ret = -EINVAL;
+ goto err;
+ }
+ }
+
+ inode_u.bi_size += len;
+ inode_u.bi_mtime = inode_u.bi_ctime = bch2_current_time(trans->c);
+
+ ret = bch2_inode_write(trans, &iter, &inode_u);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static int __bch2_resume_logged_op_finsert(struct btree_trans *trans,
+ struct bkey_i *op_k,
+ u64 *i_sectors_delta)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_i_logged_op_finsert *op = bkey_i_to_logged_op_finsert(op_k);
+ subvol_inum inum = { le32_to_cpu(op->v.subvol), le64_to_cpu(op->v.inum) };
+ struct bch_io_opts opts;
+ u64 dst_offset = le64_to_cpu(op->v.dst_offset);
+ u64 src_offset = le64_to_cpu(op->v.src_offset);
+ s64 shift = dst_offset - src_offset;
+ u64 len = abs(shift);
+ u64 pos = le64_to_cpu(op->v.pos);
+ bool insert = shift > 0;
+ int ret = 0;
+
+ ret = bch2_inum_opts_get(trans, inum, &opts);
+ if (ret)
+ return ret;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+ POS(inum.inum, 0),
+ BTREE_ITER_INTENT);
+
+ switch (op->v.state) {
+case LOGGED_OP_FINSERT_start:
+ op->v.state = LOGGED_OP_FINSERT_shift_extents;
+
+ if (insert) {
+ ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ adjust_i_size(trans, inum, src_offset, len) ?:
+ bch2_logged_op_update(trans, &op->k_i));
+ if (ret)
+ goto err;
+ } else {
+ bch2_btree_iter_set_pos(&iter, POS(inum.inum, src_offset));
+
+ ret = bch2_fpunch_at(trans, &iter, inum, src_offset + len, i_sectors_delta);
+ if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ goto err;
+
+ ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ bch2_logged_op_update(trans, &op->k_i));
+ }
+
+ fallthrough;
+case LOGGED_OP_FINSERT_shift_extents:
+ while (1) {
+ struct disk_reservation disk_res =
+ bch2_disk_reservation_init(c, 0);
+ struct bkey_i delete, *copy;
+ struct bkey_s_c k;
+ struct bpos src_pos = POS(inum.inum, src_offset);
+ u32 snapshot;
+
+ bch2_trans_begin(trans);
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (ret)
+ goto btree_err;
+
+ bch2_btree_iter_set_snapshot(&iter, snapshot);
+ bch2_btree_iter_set_pos(&iter, SPOS(inum.inum, pos, snapshot));
+
+ k = insert
+ ? bch2_btree_iter_peek_prev(&iter)
+ : bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX));
+ if ((ret = bkey_err(k)))
+ goto btree_err;
+
+ if (!k.k ||
+ k.k->p.inode != inum.inum ||
+ bkey_le(k.k->p, POS(inum.inum, src_offset)))
+ break;
+
+ copy = bch2_bkey_make_mut_noupdate(trans, k);
+ if ((ret = PTR_ERR_OR_ZERO(copy)))
+ goto btree_err;
+
+ if (insert &&
+ bkey_lt(bkey_start_pos(k.k), src_pos)) {
+ bch2_cut_front(src_pos, copy);
+
+ /* Splitting compressed extent? */
+ bch2_disk_reservation_add(c, &disk_res,
+ copy->k.size *
+ bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy)),
+ BCH_DISK_RESERVATION_NOFAIL);
+ }
+
+ bkey_init(&delete.k);
+ delete.k.p = copy->k.p;
+ delete.k.p.snapshot = snapshot;
+ delete.k.size = copy->k.size;
+
+ copy->k.p.offset += shift;
+ copy->k.p.snapshot = snapshot;
+
+ op->v.pos = cpu_to_le64(insert ? bkey_start_offset(&delete.k) : delete.k.p.offset);
+
+ ret = bch2_bkey_set_needs_rebalance(c, copy,
+ opts.background_target,
+ opts.background_compression) ?:
+ bch2_btree_insert_trans(trans, BTREE_ID_extents, &delete, 0) ?:
+ bch2_btree_insert_trans(trans, BTREE_ID_extents, copy, 0) ?:
+ bch2_logged_op_update(trans, &op->k_i) ?:
+ bch2_trans_commit(trans, &disk_res, NULL, BCH_TRANS_COMMIT_no_enospc);
+btree_err:
+ bch2_disk_reservation_put(c, &disk_res);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ goto err;
+
+ pos = le64_to_cpu(op->v.pos);
+ }
+
+ op->v.state = LOGGED_OP_FINSERT_finish;
+
+ if (!insert) {
+ ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ adjust_i_size(trans, inum, src_offset, shift) ?:
+ bch2_logged_op_update(trans, &op->k_i));
+ } else {
+ /* We need an inode update to update bi_journal_seq for fsync: */
+ ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ adjust_i_size(trans, inum, 0, 0) ?:
+ bch2_logged_op_update(trans, &op->k_i));
+ }
+
+ break;
+case LOGGED_OP_FINSERT_finish:
+ break;
+ }
+err:
+ bch2_logged_op_finish(trans, op_k);
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+int bch2_resume_logged_op_finsert(struct btree_trans *trans, struct bkey_i *op_k)
+{
+ return __bch2_resume_logged_op_finsert(trans, op_k, NULL);
+}
+
+int bch2_fcollapse_finsert(struct bch_fs *c, subvol_inum inum,
+ u64 offset, u64 len, bool insert,
+ s64 *i_sectors_delta)
+{
+ struct bkey_i_logged_op_finsert op;
+ s64 shift = insert ? len : -len;
+
+ bkey_logged_op_finsert_init(&op.k_i);
+ op.v.subvol = cpu_to_le32(inum.subvol);
+ op.v.inum = cpu_to_le64(inum.inum);
+ op.v.dst_offset = cpu_to_le64(offset + shift);
+ op.v.src_offset = cpu_to_le64(offset);
+ op.v.pos = cpu_to_le64(insert ? U64_MAX : offset);
+
+ /*
+ * Logged ops aren't atomic w.r.t. snapshot creation: creating a
+ * snapshot while they're in progress, then crashing, will result in the
+ * resume only proceeding in one of the snapshots
+ */
+ down_read(&c->snapshot_create_lock);
+ int ret = bch2_trans_run(c,
+ bch2_logged_op_start(trans, &op.k_i) ?:
+ __bch2_resume_logged_op_finsert(trans, &op.k_i, i_sectors_delta));
+ up_read(&c->snapshot_create_lock);
+
+ return ret;
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_MISC_H
+#define _BCACHEFS_IO_MISC_H
+
+int bch2_extent_fallocate(struct btree_trans *, subvol_inum, struct btree_iter *,
+ u64, struct bch_io_opts, s64 *,
+ struct write_point_specifier);
+int bch2_fpunch_at(struct btree_trans *, struct btree_iter *,
+ subvol_inum, u64, s64 *);
+int bch2_fpunch(struct bch_fs *c, subvol_inum, u64, u64, s64 *);
+
+void bch2_logged_op_truncate_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+
+#define bch2_bkey_ops_logged_op_truncate ((struct bkey_ops) { \
+ .val_to_text = bch2_logged_op_truncate_to_text, \
+ .min_val_size = 24, \
+})
+
+int bch2_resume_logged_op_truncate(struct btree_trans *, struct bkey_i *);
+
+int bch2_truncate(struct bch_fs *, subvol_inum, u64, u64 *);
+
+void bch2_logged_op_finsert_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+
+#define bch2_bkey_ops_logged_op_finsert ((struct bkey_ops) { \
+ .val_to_text = bch2_logged_op_finsert_to_text, \
+ .min_val_size = 24, \
+})
+
+int bch2_resume_logged_op_finsert(struct btree_trans *, struct bkey_i *);
+
+int bch2_fcollapse_finsert(struct bch_fs *, subvol_inum, u64, u64, bool, s64 *);
+
+#endif /* _BCACHEFS_IO_MISC_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Some low level IO code, and hacks for various block layer limitations
+ *
+ * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
+ * Copyright 2012 Google, Inc.
+ */
+
+#include "bcachefs.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "checksum.h"
+#include "clock.h"
+#include "compress.h"
+#include "data_update.h"
+#include "disk_groups.h"
+#include "ec.h"
+#include "error.h"
+#include "io_read.h"
+#include "io_misc.h"
+#include "io_write.h"
+#include "subvolume.h"
+#include "trace.h"
+
+#include <linux/sched/mm.h>
+
+#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
+
+static bool bch2_target_congested(struct bch_fs *c, u16 target)
+{
+ const struct bch_devs_mask *devs;
+ unsigned d, nr = 0, total = 0;
+ u64 now = local_clock(), last;
+ s64 congested;
+ struct bch_dev *ca;
+
+ if (!target)
+ return false;
+
+ rcu_read_lock();
+ devs = bch2_target_to_mask(c, target) ?:
+ &c->rw_devs[BCH_DATA_user];
+
+ for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
+ ca = rcu_dereference(c->devs[d]);
+ if (!ca)
+ continue;
+
+ congested = atomic_read(&ca->congested);
+ last = READ_ONCE(ca->congested_last);
+ if (time_after64(now, last))
+ congested -= (now - last) >> 12;
+
+ total += max(congested, 0LL);
+ nr++;
+ }
+ rcu_read_unlock();
+
+ return bch2_rand_range(nr * CONGESTED_MAX) < total;
+}
+
+#else
+
+static bool bch2_target_congested(struct bch_fs *c, u16 target)
+{
+ return false;
+}
+
+#endif
+
+/* Cache promotion on read */
+
+struct promote_op {
+ struct rcu_head rcu;
+ u64 start_time;
+
+ struct rhash_head hash;
+ struct bpos pos;
+
+ struct data_update write;
+ struct bio_vec bi_inline_vecs[0]; /* must be last */
+};
+
+static const struct rhashtable_params bch_promote_params = {
+ .head_offset = offsetof(struct promote_op, hash),
+ .key_offset = offsetof(struct promote_op, pos),
+ .key_len = sizeof(struct bpos),
+};
+
+static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,
+ struct bpos pos,
+ struct bch_io_opts opts,
+ unsigned flags)
+{
+ BUG_ON(!opts.promote_target);
+
+ if (!(flags & BCH_READ_MAY_PROMOTE))
+ return -BCH_ERR_nopromote_may_not;
+
+ if (bch2_bkey_has_target(c, k, opts.promote_target))
+ return -BCH_ERR_nopromote_already_promoted;
+
+ if (bkey_extent_is_unwritten(k))
+ return -BCH_ERR_nopromote_unwritten;
+
+ if (bch2_target_congested(c, opts.promote_target))
+ return -BCH_ERR_nopromote_congested;
+
+ if (rhashtable_lookup_fast(&c->promote_table, &pos,
+ bch_promote_params))
+ return -BCH_ERR_nopromote_in_flight;
+
+ return 0;
+}
+
+static void promote_free(struct bch_fs *c, struct promote_op *op)
+{
+ int ret;
+
+ bch2_data_update_exit(&op->write);
+
+ ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
+ bch_promote_params);
+ BUG_ON(ret);
+ bch2_write_ref_put(c, BCH_WRITE_REF_promote);
+ kfree_rcu(op, rcu);
+}
+
+static void promote_done(struct bch_write_op *wop)
+{
+ struct promote_op *op =
+ container_of(wop, struct promote_op, write.op);
+ struct bch_fs *c = op->write.op.c;
+
+ bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
+ op->start_time);
+ promote_free(c, op);
+}
+
+static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
+{
+ struct bio *bio = &op->write.op.wbio.bio;
+
+ trace_and_count(op->write.op.c, read_promote, &rbio->bio);
+
+ /* we now own pages: */
+ BUG_ON(!rbio->bounce);
+ BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
+
+ memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
+ sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
+ swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
+
+ bch2_data_update_read_done(&op->write, rbio->pick.crc);
+}
+
+static struct promote_op *__promote_alloc(struct btree_trans *trans,
+ enum btree_id btree_id,
+ struct bkey_s_c k,
+ struct bpos pos,
+ struct extent_ptr_decoded *pick,
+ struct bch_io_opts opts,
+ unsigned sectors,
+ struct bch_read_bio **rbio)
+{
+ struct bch_fs *c = trans->c;
+ struct promote_op *op = NULL;
+ struct bio *bio;
+ unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
+ int ret;
+
+ if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
+ return NULL;
+
+ op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOFS);
+ if (!op)
+ goto err;
+
+ op->start_time = local_clock();
+ op->pos = pos;
+
+ /*
+ * We don't use the mempool here because extents that aren't
+ * checksummed or compressed can be too big for the mempool:
+ */
+ *rbio = kzalloc(sizeof(struct bch_read_bio) +
+ sizeof(struct bio_vec) * pages,
+ GFP_NOFS);
+ if (!*rbio)
+ goto err;
+
+ rbio_init(&(*rbio)->bio, opts);
+ bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
+
+ if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
+ GFP_NOFS))
+ goto err;
+
+ (*rbio)->bounce = true;
+ (*rbio)->split = true;
+ (*rbio)->kmalloc = true;
+
+ if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
+ bch_promote_params))
+ goto err;
+
+ bio = &op->write.op.wbio.bio;
+ bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
+
+ ret = bch2_data_update_init(trans, NULL, &op->write,
+ writepoint_hashed((unsigned long) current),
+ opts,
+ (struct data_update_opts) {
+ .target = opts.promote_target,
+ .extra_replicas = 1,
+ .write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
+ },
+ btree_id, k);
+ /*
+ * possible errors: -BCH_ERR_nocow_lock_blocked,
+ * -BCH_ERR_ENOSPC_disk_reservation:
+ */
+ if (ret) {
+ ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
+ bch_promote_params);
+ BUG_ON(ret);
+ goto err;
+ }
+
+ op->write.op.end_io = promote_done;
+
+ return op;
+err:
+ if (*rbio)
+ bio_free_pages(&(*rbio)->bio);
+ kfree(*rbio);
+ *rbio = NULL;
+ kfree(op);
+ bch2_write_ref_put(c, BCH_WRITE_REF_promote);
+ return NULL;
+}
+
+noinline
+static struct promote_op *promote_alloc(struct btree_trans *trans,
+ struct bvec_iter iter,
+ struct bkey_s_c k,
+ struct extent_ptr_decoded *pick,
+ struct bch_io_opts opts,
+ unsigned flags,
+ struct bch_read_bio **rbio,
+ bool *bounce,
+ bool *read_full)
+{
+ struct bch_fs *c = trans->c;
+ bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
+ /* data might have to be decompressed in the write path: */
+ unsigned sectors = promote_full
+ ? max(pick->crc.compressed_size, pick->crc.live_size)
+ : bvec_iter_sectors(iter);
+ struct bpos pos = promote_full
+ ? bkey_start_pos(k.k)
+ : POS(k.k->p.inode, iter.bi_sector);
+ struct promote_op *promote;
+ int ret;
+
+ ret = should_promote(c, k, pos, opts, flags);
+ if (ret)
+ goto nopromote;
+
+ promote = __promote_alloc(trans,
+ k.k->type == KEY_TYPE_reflink_v
+ ? BTREE_ID_reflink
+ : BTREE_ID_extents,
+ k, pos, pick, opts, sectors, rbio);
+ if (!promote) {
+ ret = -BCH_ERR_nopromote_enomem;
+ goto nopromote;
+ }
+
+ *bounce = true;
+ *read_full = promote_full;
+ return promote;
+nopromote:
+ trace_read_nopromote(c, ret);
+ return NULL;
+}
+
+/* Read */
+
+#define READ_RETRY_AVOID 1
+#define READ_RETRY 2
+#define READ_ERR 3
+
+enum rbio_context {
+ RBIO_CONTEXT_NULL,
+ RBIO_CONTEXT_HIGHPRI,
+ RBIO_CONTEXT_UNBOUND,
+};
+
+static inline struct bch_read_bio *
+bch2_rbio_parent(struct bch_read_bio *rbio)
+{
+ return rbio->split ? rbio->parent : rbio;
+}
+
+__always_inline
+static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
+ enum rbio_context context,
+ struct workqueue_struct *wq)
+{
+ if (context <= rbio->context) {
+ fn(&rbio->work);
+ } else {
+ rbio->work.func = fn;
+ rbio->context = context;
+ queue_work(wq, &rbio->work);
+ }
+}
+
+static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
+{
+ BUG_ON(rbio->bounce && !rbio->split);
+
+ if (rbio->promote)
+ promote_free(rbio->c, rbio->promote);
+ rbio->promote = NULL;
+
+ if (rbio->bounce)
+ bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
+
+ if (rbio->split) {
+ struct bch_read_bio *parent = rbio->parent;
+
+ if (rbio->kmalloc)
+ kfree(rbio);
+ else
+ bio_put(&rbio->bio);
+
+ rbio = parent;
+ }
+
+ return rbio;
+}
+
+/*
+ * Only called on a top level bch_read_bio to complete an entire read request,
+ * not a split:
+ */
+static void bch2_rbio_done(struct bch_read_bio *rbio)
+{
+ if (rbio->start_time)
+ bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
+ rbio->start_time);
+ bio_endio(&rbio->bio);
+}
+
+static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
+ struct bvec_iter bvec_iter,
+ struct bch_io_failures *failed,
+ unsigned flags)
+{
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct btree_iter iter;
+ struct bkey_buf sk;
+ struct bkey_s_c k;
+ int ret;
+
+ flags &= ~BCH_READ_LAST_FRAGMENT;
+ flags |= BCH_READ_MUST_CLONE;
+
+ bch2_bkey_buf_init(&sk);
+
+ bch2_trans_iter_init(trans, &iter, rbio->data_btree,
+ rbio->read_pos, BTREE_ITER_SLOTS);
+retry:
+ rbio->bio.bi_status = 0;
+
+ k = bch2_btree_iter_peek_slot(&iter);
+ if (bkey_err(k))
+ goto err;
+
+ bch2_bkey_buf_reassemble(&sk, c, k);
+ k = bkey_i_to_s_c(sk.k);
+ bch2_trans_unlock(trans);
+
+ if (!bch2_bkey_matches_ptr(c, k,
+ rbio->pick.ptr,
+ rbio->data_pos.offset -
+ rbio->pick.crc.offset)) {
+ /* extent we wanted to read no longer exists: */
+ rbio->hole = true;
+ goto out;
+ }
+
+ ret = __bch2_read_extent(trans, rbio, bvec_iter,
+ rbio->read_pos,
+ rbio->data_btree,
+ k, 0, failed, flags);
+ if (ret == READ_RETRY)
+ goto retry;
+ if (ret)
+ goto err;
+out:
+ bch2_rbio_done(rbio);
+ bch2_trans_iter_exit(trans, &iter);
+ bch2_trans_put(trans);
+ bch2_bkey_buf_exit(&sk, c);
+ return;
+err:
+ rbio->bio.bi_status = BLK_STS_IOERR;
+ goto out;
+}
+
+static void bch2_rbio_retry(struct work_struct *work)
+{
+ struct bch_read_bio *rbio =
+ container_of(work, struct bch_read_bio, work);
+ struct bch_fs *c = rbio->c;
+ struct bvec_iter iter = rbio->bvec_iter;
+ unsigned flags = rbio->flags;
+ subvol_inum inum = {
+ .subvol = rbio->subvol,
+ .inum = rbio->read_pos.inode,
+ };
+ struct bch_io_failures failed = { .nr = 0 };
+
+ trace_and_count(c, read_retry, &rbio->bio);
+
+ if (rbio->retry == READ_RETRY_AVOID)
+ bch2_mark_io_failure(&failed, &rbio->pick);
+
+ rbio->bio.bi_status = 0;
+
+ rbio = bch2_rbio_free(rbio);
+
+ flags |= BCH_READ_IN_RETRY;
+ flags &= ~BCH_READ_MAY_PROMOTE;
+
+ if (flags & BCH_READ_NODECODE) {
+ bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
+ } else {
+ flags &= ~BCH_READ_LAST_FRAGMENT;
+ flags |= BCH_READ_MUST_CLONE;
+
+ __bch2_read(c, rbio, iter, inum, &failed, flags);
+ }
+}
+
+static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
+ blk_status_t error)
+{
+ rbio->retry = retry;
+
+ if (rbio->flags & BCH_READ_IN_RETRY)
+ return;
+
+ if (retry == READ_ERR) {
+ rbio = bch2_rbio_free(rbio);
+
+ rbio->bio.bi_status = error;
+ bch2_rbio_done(rbio);
+ } else {
+ bch2_rbio_punt(rbio, bch2_rbio_retry,
+ RBIO_CONTEXT_UNBOUND, system_unbound_wq);
+ }
+}
+
+static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
+ struct bch_read_bio *rbio)
+{
+ struct bch_fs *c = rbio->c;
+ u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
+ struct bch_extent_crc_unpacked new_crc;
+ struct btree_iter iter;
+ struct bkey_i *new;
+ struct bkey_s_c k;
+ int ret = 0;
+
+ if (crc_is_compressed(rbio->pick.crc))
+ return 0;
+
+ k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+ if ((ret = bkey_err(k)))
+ goto out;
+
+ if (bversion_cmp(k.k->version, rbio->version) ||
+ !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
+ goto out;
+
+ /* Extent was merged? */
+ if (bkey_start_offset(k.k) < data_offset ||
+ k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
+ goto out;
+
+ if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
+ rbio->pick.crc, NULL, &new_crc,
+ bkey_start_offset(k.k) - data_offset, k.k->size,
+ rbio->pick.crc.csum_type)) {
+ bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
+ ret = 0;
+ goto out;
+ }
+
+ /*
+ * going to be temporarily appending another checksum entry:
+ */
+ new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
+ sizeof(struct bch_extent_crc128));
+ if ((ret = PTR_ERR_OR_ZERO(new)))
+ goto out;
+
+ bkey_reassemble(new, k);
+
+ if (!bch2_bkey_narrow_crcs(new, new_crc))
+ goto out;
+
+ ret = bch2_trans_update(trans, &iter, new,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+out:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
+{
+ bch2_trans_do(rbio->c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ __bch2_rbio_narrow_crcs(trans, rbio));
+}
+
+/* Inner part that may run in process context */
+static void __bch2_read_endio(struct work_struct *work)
+{
+ struct bch_read_bio *rbio =
+ container_of(work, struct bch_read_bio, work);
+ struct bch_fs *c = rbio->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
+ struct bio *src = &rbio->bio;
+ struct bio *dst = &bch2_rbio_parent(rbio)->bio;
+ struct bvec_iter dst_iter = rbio->bvec_iter;
+ struct bch_extent_crc_unpacked crc = rbio->pick.crc;
+ struct nonce nonce = extent_nonce(rbio->version, crc);
+ unsigned nofs_flags;
+ struct bch_csum csum;
+ int ret;
+
+ nofs_flags = memalloc_nofs_save();
+
+ /* Reset iterator for checksumming and copying bounced data: */
+ if (rbio->bounce) {
+ src->bi_iter.bi_size = crc.compressed_size << 9;
+ src->bi_iter.bi_idx = 0;
+ src->bi_iter.bi_bvec_done = 0;
+ } else {
+ src->bi_iter = rbio->bvec_iter;
+ }
+
+ csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
+ if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
+ goto csum_err;
+
+ /*
+ * XXX
+ * We need to rework the narrow_crcs path to deliver the read completion
+ * first, and then punt to a different workqueue, otherwise we're
+ * holding up reads while doing btree updates which is bad for memory
+ * reclaim.
+ */
+ if (unlikely(rbio->narrow_crcs))
+ bch2_rbio_narrow_crcs(rbio);
+
+ if (rbio->flags & BCH_READ_NODECODE)
+ goto nodecode;
+
+ /* Adjust crc to point to subset of data we want: */
+ crc.offset += rbio->offset_into_extent;
+ crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
+
+ if (crc_is_compressed(crc)) {
+ ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+ if (ret)
+ goto decrypt_err;
+
+ if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
+ !c->opts.no_data_io)
+ goto decompression_err;
+ } else {
+ /* don't need to decrypt the entire bio: */
+ nonce = nonce_add(nonce, crc.offset << 9);
+ bio_advance(src, crc.offset << 9);
+
+ BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
+ src->bi_iter.bi_size = dst_iter.bi_size;
+
+ ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+ if (ret)
+ goto decrypt_err;
+
+ if (rbio->bounce) {
+ struct bvec_iter src_iter = src->bi_iter;
+
+ bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
+ }
+ }
+
+ if (rbio->promote) {
+ /*
+ * Re encrypt data we decrypted, so it's consistent with
+ * rbio->crc:
+ */
+ ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
+ if (ret)
+ goto decrypt_err;
+
+ promote_start(rbio->promote, rbio);
+ rbio->promote = NULL;
+ }
+nodecode:
+ if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
+ rbio = bch2_rbio_free(rbio);
+ bch2_rbio_done(rbio);
+ }
+out:
+ memalloc_nofs_restore(nofs_flags);
+ return;
+csum_err:
+ /*
+ * Checksum error: if the bio wasn't bounced, we may have been
+ * reading into buffers owned by userspace (that userspace can
+ * scribble over) - retry the read, bouncing it this time:
+ */
+ if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
+ rbio->flags |= BCH_READ_MUST_BOUNCE;
+ bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
+ goto out;
+ }
+
+ bch_err_inum_offset_ratelimited(ca,
+ rbio->read_pos.inode,
+ rbio->read_pos.offset << 9,
+ "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
+ rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
+ csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
+ bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+ goto out;
+decompression_err:
+ bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
+ rbio->read_pos.offset << 9,
+ "decompression error");
+ bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+ goto out;
+decrypt_err:
+ bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
+ rbio->read_pos.offset << 9,
+ "decrypt error");
+ bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+ goto out;
+}
+
+static void bch2_read_endio(struct bio *bio)
+{
+ struct bch_read_bio *rbio =
+ container_of(bio, struct bch_read_bio, bio);
+ struct bch_fs *c = rbio->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
+ struct workqueue_struct *wq = NULL;
+ enum rbio_context context = RBIO_CONTEXT_NULL;
+
+ if (rbio->have_ioref) {
+ bch2_latency_acct(ca, rbio->submit_time, READ);
+ percpu_ref_put(&ca->io_ref);
+ }
+
+ if (!rbio->split)
+ rbio->bio.bi_end_io = rbio->end_io;
+
+ if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
+ rbio->read_pos.inode,
+ rbio->read_pos.offset,
+ "data read error: %s",
+ bch2_blk_status_to_str(bio->bi_status))) {
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
+ return;
+ }
+
+ if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
+ ptr_stale(ca, &rbio->pick.ptr)) {
+ trace_and_count(c, read_reuse_race, &rbio->bio);
+
+ if (rbio->flags & BCH_READ_RETRY_IF_STALE)
+ bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
+ else
+ bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
+ return;
+ }
+
+ if (rbio->narrow_crcs ||
+ rbio->promote ||
+ crc_is_compressed(rbio->pick.crc) ||
+ bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
+ context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
+ else if (rbio->pick.crc.csum_type)
+ context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
+
+ bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
+}
+
+int __bch2_read_indirect_extent(struct btree_trans *trans,
+ unsigned *offset_into_extent,
+ struct bkey_buf *orig_k)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ u64 reflink_offset;
+ int ret;
+
+ reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
+ *offset_into_extent;
+
+ k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
+ POS(0, reflink_offset), 0);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ if (k.k->type != KEY_TYPE_reflink_v &&
+ k.k->type != KEY_TYPE_indirect_inline_data) {
+ bch_err_inum_offset_ratelimited(trans->c,
+ orig_k->k->k.p.inode,
+ orig_k->k->k.p.offset << 9,
+ "%llu len %u points to nonexistent indirect extent %llu",
+ orig_k->k->k.p.offset,
+ orig_k->k->k.size,
+ reflink_offset);
+ bch2_inconsistent_error(trans->c);
+ ret = -EIO;
+ goto err;
+ }
+
+ *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
+ bch2_bkey_buf_reassemble(orig_k, trans->c, k);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
+ struct bkey_s_c k,
+ struct bch_extent_ptr ptr)
+{
+ struct bch_fs *c = trans->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
+ struct btree_iter iter;
+ struct printbuf buf = PRINTBUF;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
+ PTR_BUCKET_POS(c, &ptr),
+ BTREE_ITER_CACHED);
+
+ prt_printf(&buf, "Attempting to read from stale dirty pointer:");
+ printbuf_indent_add(&buf, 2);
+ prt_newline(&buf);
+
+ bch2_bkey_val_to_text(&buf, c, k);
+ prt_newline(&buf);
+
+ prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
+
+ ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
+ if (!ret) {
+ prt_newline(&buf);
+ bch2_bkey_val_to_text(&buf, c, k);
+ }
+
+ bch2_fs_inconsistent(c, "%s", buf.buf);
+
+ bch2_trans_iter_exit(trans, &iter);
+ printbuf_exit(&buf);
+}
+
+int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
+ struct bvec_iter iter, struct bpos read_pos,
+ enum btree_id data_btree, struct bkey_s_c k,
+ unsigned offset_into_extent,
+ struct bch_io_failures *failed, unsigned flags)
+{
+ struct bch_fs *c = trans->c;
+ struct extent_ptr_decoded pick;
+ struct bch_read_bio *rbio = NULL;
+ struct bch_dev *ca = NULL;
+ struct promote_op *promote = NULL;
+ bool bounce = false, read_full = false, narrow_crcs = false;
+ struct bpos data_pos = bkey_start_pos(k.k);
+ int pick_ret;
+
+ if (bkey_extent_is_inline_data(k.k)) {
+ unsigned bytes = min_t(unsigned, iter.bi_size,
+ bkey_inline_data_bytes(k.k));
+
+ swap(iter.bi_size, bytes);
+ memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
+ swap(iter.bi_size, bytes);
+ bio_advance_iter(&orig->bio, &iter, bytes);
+ zero_fill_bio_iter(&orig->bio, iter);
+ goto out_read_done;
+ }
+retry_pick:
+ pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
+
+ /* hole or reservation - just zero fill: */
+ if (!pick_ret)
+ goto hole;
+
+ if (pick_ret < 0) {
+ bch_err_inum_offset_ratelimited(c,
+ read_pos.inode, read_pos.offset << 9,
+ "no device to read from");
+ goto err;
+ }
+
+ ca = bch_dev_bkey_exists(c, pick.ptr.dev);
+
+ /*
+ * Stale dirty pointers are treated as IO errors, but @failed isn't
+ * allocated unless we're in the retry path - so if we're not in the
+ * retry path, don't check here, it'll be caught in bch2_read_endio()
+ * and we'll end up in the retry path:
+ */
+ if ((flags & BCH_READ_IN_RETRY) &&
+ !pick.ptr.cached &&
+ unlikely(ptr_stale(ca, &pick.ptr))) {
+ read_from_stale_dirty_pointer(trans, k, pick.ptr);
+ bch2_mark_io_failure(failed, &pick);
+ goto retry_pick;
+ }
+
+ /*
+ * Unlock the iterator while the btree node's lock is still in
+ * cache, before doing the IO:
+ */
+ bch2_trans_unlock(trans);
+
+ if (flags & BCH_READ_NODECODE) {
+ /*
+ * can happen if we retry, and the extent we were going to read
+ * has been merged in the meantime:
+ */
+ if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
+ goto hole;
+
+ iter.bi_size = pick.crc.compressed_size << 9;
+ goto get_bio;
+ }
+
+ if (!(flags & BCH_READ_LAST_FRAGMENT) ||
+ bio_flagged(&orig->bio, BIO_CHAIN))
+ flags |= BCH_READ_MUST_CLONE;
+
+ narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
+ bch2_can_narrow_extent_crcs(k, pick.crc);
+
+ if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
+ flags |= BCH_READ_MUST_BOUNCE;
+
+ EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
+
+ if (crc_is_compressed(pick.crc) ||
+ (pick.crc.csum_type != BCH_CSUM_none &&
+ (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+ (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
+ (flags & BCH_READ_USER_MAPPED)) ||
+ (flags & BCH_READ_MUST_BOUNCE)))) {
+ read_full = true;
+ bounce = true;
+ }
+
+ if (orig->opts.promote_target)
+ promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
+ &rbio, &bounce, &read_full);
+
+ if (!read_full) {
+ EBUG_ON(crc_is_compressed(pick.crc));
+ EBUG_ON(pick.crc.csum_type &&
+ (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+ bvec_iter_sectors(iter) != pick.crc.live_size ||
+ pick.crc.offset ||
+ offset_into_extent));
+
+ data_pos.offset += offset_into_extent;
+ pick.ptr.offset += pick.crc.offset +
+ offset_into_extent;
+ offset_into_extent = 0;
+ pick.crc.compressed_size = bvec_iter_sectors(iter);
+ pick.crc.uncompressed_size = bvec_iter_sectors(iter);
+ pick.crc.offset = 0;
+ pick.crc.live_size = bvec_iter_sectors(iter);
+ }
+get_bio:
+ if (rbio) {
+ /*
+ * promote already allocated bounce rbio:
+ * promote needs to allocate a bio big enough for uncompressing
+ * data in the write path, but we're not going to use it all
+ * here:
+ */
+ EBUG_ON(rbio->bio.bi_iter.bi_size <
+ pick.crc.compressed_size << 9);
+ rbio->bio.bi_iter.bi_size =
+ pick.crc.compressed_size << 9;
+ } else if (bounce) {
+ unsigned sectors = pick.crc.compressed_size;
+
+ rbio = rbio_init(bio_alloc_bioset(NULL,
+ DIV_ROUND_UP(sectors, PAGE_SECTORS),
+ 0,
+ GFP_NOFS,
+ &c->bio_read_split),
+ orig->opts);
+
+ bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
+ rbio->bounce = true;
+ rbio->split = true;
+ } else if (flags & BCH_READ_MUST_CLONE) {
+ /*
+ * Have to clone if there were any splits, due to error
+ * reporting issues (if a split errored, and retrying didn't
+ * work, when it reports the error to its parent (us) we don't
+ * know if the error was from our bio, and we should retry, or
+ * from the whole bio, in which case we don't want to retry and
+ * lose the error)
+ */
+ rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
+ &c->bio_read_split),
+ orig->opts);
+ rbio->bio.bi_iter = iter;
+ rbio->split = true;
+ } else {
+ rbio = orig;
+ rbio->bio.bi_iter = iter;
+ EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
+ }
+
+ EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
+
+ rbio->c = c;
+ rbio->submit_time = local_clock();
+ if (rbio->split)
+ rbio->parent = orig;
+ else
+ rbio->end_io = orig->bio.bi_end_io;
+ rbio->bvec_iter = iter;
+ rbio->offset_into_extent= offset_into_extent;
+ rbio->flags = flags;
+ rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
+ rbio->narrow_crcs = narrow_crcs;
+ rbio->hole = 0;
+ rbio->retry = 0;
+ rbio->context = 0;
+ /* XXX: only initialize this if needed */
+ rbio->devs_have = bch2_bkey_devs(k);
+ rbio->pick = pick;
+ rbio->subvol = orig->subvol;
+ rbio->read_pos = read_pos;
+ rbio->data_btree = data_btree;
+ rbio->data_pos = data_pos;
+ rbio->version = k.k->version;
+ rbio->promote = promote;
+ INIT_WORK(&rbio->work, NULL);
+
+ rbio->bio.bi_opf = orig->bio.bi_opf;
+ rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
+ rbio->bio.bi_end_io = bch2_read_endio;
+
+ if (rbio->bounce)
+ trace_and_count(c, read_bounce, &rbio->bio);
+
+ this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
+ bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
+
+ /*
+ * If it's being moved internally, we don't want to flag it as a cache
+ * hit:
+ */
+ if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
+ bch2_bucket_io_time_reset(trans, pick.ptr.dev,
+ PTR_BUCKET_NR(ca, &pick.ptr), READ);
+
+ if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
+ bio_inc_remaining(&orig->bio);
+ trace_and_count(c, read_split, &orig->bio);
+ }
+
+ if (!rbio->pick.idx) {
+ if (!rbio->have_ioref) {
+ bch_err_inum_offset_ratelimited(c,
+ read_pos.inode,
+ read_pos.offset << 9,
+ "no device to read from");
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+ goto out;
+ }
+
+ this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
+ bio_sectors(&rbio->bio));
+ bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
+
+ if (unlikely(c->opts.no_data_io)) {
+ if (likely(!(flags & BCH_READ_IN_RETRY)))
+ bio_endio(&rbio->bio);
+ } else {
+ if (likely(!(flags & BCH_READ_IN_RETRY)))
+ submit_bio(&rbio->bio);
+ else
+ submit_bio_wait(&rbio->bio);
+ }
+
+ /*
+ * We just submitted IO which may block, we expect relock fail
+ * events and shouldn't count them:
+ */
+ trans->notrace_relock_fail = true;
+ } else {
+ /* Attempting reconstruct read: */
+ if (bch2_ec_read_extent(trans, rbio)) {
+ bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+ goto out;
+ }
+
+ if (likely(!(flags & BCH_READ_IN_RETRY)))
+ bio_endio(&rbio->bio);
+ }
+out:
+ if (likely(!(flags & BCH_READ_IN_RETRY))) {
+ return 0;
+ } else {
+ int ret;
+
+ rbio->context = RBIO_CONTEXT_UNBOUND;
+ bch2_read_endio(&rbio->bio);
+
+ ret = rbio->retry;
+ rbio = bch2_rbio_free(rbio);
+
+ if (ret == READ_RETRY_AVOID) {
+ bch2_mark_io_failure(failed, &pick);
+ ret = READ_RETRY;
+ }
+
+ if (!ret)
+ goto out_read_done;
+
+ return ret;
+ }
+
+err:
+ if (flags & BCH_READ_IN_RETRY)
+ return READ_ERR;
+
+ orig->bio.bi_status = BLK_STS_IOERR;
+ goto out_read_done;
+
+hole:
+ /*
+ * won't normally happen in the BCH_READ_NODECODE
+ * (bch2_move_extent()) path, but if we retry and the extent we wanted
+ * to read no longer exists we have to signal that:
+ */
+ if (flags & BCH_READ_NODECODE)
+ orig->hole = true;
+
+ zero_fill_bio_iter(&orig->bio, iter);
+out_read_done:
+ if (flags & BCH_READ_LAST_FRAGMENT)
+ bch2_rbio_done(orig);
+ return 0;
+}
+
+void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
+ struct bvec_iter bvec_iter, subvol_inum inum,
+ struct bch_io_failures *failed, unsigned flags)
+{
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct btree_iter iter;
+ struct bkey_buf sk;
+ struct bkey_s_c k;
+ u32 snapshot;
+ int ret;
+
+ BUG_ON(flags & BCH_READ_NODECODE);
+
+ bch2_bkey_buf_init(&sk);
+retry:
+ bch2_trans_begin(trans);
+ iter = (struct btree_iter) { NULL };
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+ if (ret)
+ goto err;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+ SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
+ BTREE_ITER_SLOTS);
+ while (1) {
+ unsigned bytes, sectors, offset_into_extent;
+ enum btree_id data_btree = BTREE_ID_extents;
+
+ /*
+ * read_extent -> io_time_reset may cause a transaction restart
+ * without returning an error, we need to check for that here:
+ */
+ ret = bch2_trans_relock(trans);
+ if (ret)
+ break;
+
+ bch2_btree_iter_set_pos(&iter,
+ POS(inum.inum, bvec_iter.bi_sector));
+
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ break;
+
+ offset_into_extent = iter.pos.offset -
+ bkey_start_offset(k.k);
+ sectors = k.k->size - offset_into_extent;
+
+ bch2_bkey_buf_reassemble(&sk, c, k);
+
+ ret = bch2_read_indirect_extent(trans, &data_btree,
+ &offset_into_extent, &sk);
+ if (ret)
+ break;
+
+ k = bkey_i_to_s_c(sk.k);
+
+ /*
+ * With indirect extents, the amount of data to read is the min
+ * of the original extent and the indirect extent:
+ */
+ sectors = min(sectors, k.k->size - offset_into_extent);
+
+ bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
+ swap(bvec_iter.bi_size, bytes);
+
+ if (bvec_iter.bi_size == bytes)
+ flags |= BCH_READ_LAST_FRAGMENT;
+
+ ret = __bch2_read_extent(trans, rbio, bvec_iter, iter.pos,
+ data_btree, k,
+ offset_into_extent, failed, flags);
+ if (ret)
+ break;
+
+ if (flags & BCH_READ_LAST_FRAGMENT)
+ break;
+
+ swap(bvec_iter.bi_size, bytes);
+ bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
+
+ ret = btree_trans_too_many_iters(trans);
+ if (ret)
+ break;
+ }
+err:
+ bch2_trans_iter_exit(trans, &iter);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
+ ret == READ_RETRY ||
+ ret == READ_RETRY_AVOID)
+ goto retry;
+
+ bch2_trans_put(trans);
+ bch2_bkey_buf_exit(&sk, c);
+
+ if (ret) {
+ bch_err_inum_offset_ratelimited(c, inum.inum,
+ bvec_iter.bi_sector << 9,
+ "read error %i from btree lookup", ret);
+ rbio->bio.bi_status = BLK_STS_IOERR;
+ bch2_rbio_done(rbio);
+ }
+}
+
+void bch2_fs_io_read_exit(struct bch_fs *c)
+{
+ if (c->promote_table.tbl)
+ rhashtable_destroy(&c->promote_table);
+ bioset_exit(&c->bio_read_split);
+ bioset_exit(&c->bio_read);
+}
+
+int bch2_fs_io_read_init(struct bch_fs *c)
+{
+ if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
+ BIOSET_NEED_BVECS))
+ return -BCH_ERR_ENOMEM_bio_read_init;
+
+ if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
+ BIOSET_NEED_BVECS))
+ return -BCH_ERR_ENOMEM_bio_read_split_init;
+
+ if (rhashtable_init(&c->promote_table, &bch_promote_params))
+ return -BCH_ERR_ENOMEM_promote_table_init;
+
+ return 0;
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_READ_H
+#define _BCACHEFS_IO_READ_H
+
+#include "bkey_buf.h"
+
+struct bch_read_bio {
+ struct bch_fs *c;
+ u64 start_time;
+ u64 submit_time;
+
+ /*
+ * Reads will often have to be split, and if the extent being read from
+ * was checksummed or compressed we'll also have to allocate bounce
+ * buffers and copy the data back into the original bio.
+ *
+ * If we didn't have to split, we have to save and restore the original
+ * bi_end_io - @split below indicates which:
+ */
+ union {
+ struct bch_read_bio *parent;
+ bio_end_io_t *end_io;
+ };
+
+ /*
+ * Saved copy of bio->bi_iter, from submission time - allows us to
+ * resubmit on IO error, and also to copy data back to the original bio
+ * when we're bouncing:
+ */
+ struct bvec_iter bvec_iter;
+
+ unsigned offset_into_extent;
+
+ u16 flags;
+ union {
+ struct {
+ u16 bounce:1,
+ split:1,
+ kmalloc:1,
+ have_ioref:1,
+ narrow_crcs:1,
+ hole:1,
+ retry:2,
+ context:2;
+ };
+ u16 _state;
+ };
+
+ struct bch_devs_list devs_have;
+
+ struct extent_ptr_decoded pick;
+
+ /*
+ * pos we read from - different from data_pos for indirect extents:
+ */
+ u32 subvol;
+ struct bpos read_pos;
+
+ /*
+ * start pos of data we read (may not be pos of data we want) - for
+ * promote, narrow extents paths:
+ */
+ enum btree_id data_btree;
+ struct bpos data_pos;
+ struct bversion version;
+
+ struct promote_op *promote;
+
+ struct bch_io_opts opts;
+
+ struct work_struct work;
+
+ struct bio bio;
+};
+
+#define to_rbio(_bio) container_of((_bio), struct bch_read_bio, bio)
+
+struct bch_devs_mask;
+struct cache_promote_op;
+struct extent_ptr_decoded;
+
+int __bch2_read_indirect_extent(struct btree_trans *, unsigned *,
+ struct bkey_buf *);
+
+static inline int bch2_read_indirect_extent(struct btree_trans *trans,
+ enum btree_id *data_btree,
+ unsigned *offset_into_extent,
+ struct bkey_buf *k)
+{
+ if (k->k->k.type != KEY_TYPE_reflink_p)
+ return 0;
+
+ *data_btree = BTREE_ID_reflink;
+ return __bch2_read_indirect_extent(trans, offset_into_extent, k);
+}
+
+enum bch_read_flags {
+ BCH_READ_RETRY_IF_STALE = 1 << 0,
+ BCH_READ_MAY_PROMOTE = 1 << 1,
+ BCH_READ_USER_MAPPED = 1 << 2,
+ BCH_READ_NODECODE = 1 << 3,
+ BCH_READ_LAST_FRAGMENT = 1 << 4,
+
+ /* internal: */
+ BCH_READ_MUST_BOUNCE = 1 << 5,
+ BCH_READ_MUST_CLONE = 1 << 6,
+ BCH_READ_IN_RETRY = 1 << 7,
+};
+
+int __bch2_read_extent(struct btree_trans *, struct bch_read_bio *,
+ struct bvec_iter, struct bpos, enum btree_id,
+ struct bkey_s_c, unsigned,
+ struct bch_io_failures *, unsigned);
+
+static inline void bch2_read_extent(struct btree_trans *trans,
+ struct bch_read_bio *rbio, struct bpos read_pos,
+ enum btree_id data_btree, struct bkey_s_c k,
+ unsigned offset_into_extent, unsigned flags)
+{
+ __bch2_read_extent(trans, rbio, rbio->bio.bi_iter, read_pos,
+ data_btree, k, offset_into_extent, NULL, flags);
+}
+
+void __bch2_read(struct bch_fs *, struct bch_read_bio *, struct bvec_iter,
+ subvol_inum, struct bch_io_failures *, unsigned flags);
+
+static inline void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
+ subvol_inum inum)
+{
+ struct bch_io_failures failed = { .nr = 0 };
+
+ BUG_ON(rbio->_state);
+
+ rbio->c = c;
+ rbio->start_time = local_clock();
+ rbio->subvol = inum.subvol;
+
+ __bch2_read(c, rbio, rbio->bio.bi_iter, inum, &failed,
+ BCH_READ_RETRY_IF_STALE|
+ BCH_READ_MAY_PROMOTE|
+ BCH_READ_USER_MAPPED);
+}
+
+static inline struct bch_read_bio *rbio_init(struct bio *bio,
+ struct bch_io_opts opts)
+{
+ struct bch_read_bio *rbio = to_rbio(bio);
+
+ rbio->_state = 0;
+ rbio->promote = NULL;
+ rbio->opts = opts;
+ return rbio;
+}
+
+void bch2_fs_io_read_exit(struct bch_fs *);
+int bch2_fs_io_read_init(struct bch_fs *);
+
+#endif /* _BCACHEFS_IO_READ_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
+ * Copyright 2012 Google, Inc.
+ */
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "bkey_buf.h"
+#include "bset.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "checksum.h"
+#include "clock.h"
+#include "compress.h"
+#include "debug.h"
+#include "ec.h"
+#include "error.h"
+#include "extent_update.h"
+#include "inode.h"
+#include "io_write.h"
+#include "journal.h"
+#include "keylist.h"
+#include "move.h"
+#include "nocow_locking.h"
+#include "rebalance.h"
+#include "subvolume.h"
+#include "super.h"
+#include "super-io.h"
+#include "trace.h"
+
+#include <linux/blkdev.h>
+#include <linux/prefetch.h>
+#include <linux/random.h>
+#include <linux/sched/mm.h>
+
+#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
+
+static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
+ u64 now, int rw)
+{
+ u64 latency_capable =
+ ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
+ /* ideally we'd be taking into account the device's variance here: */
+ u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
+ s64 latency_over = io_latency - latency_threshold;
+
+ if (latency_threshold && latency_over > 0) {
+ /*
+ * bump up congested by approximately latency_over * 4 /
+ * latency_threshold - we don't need much accuracy here so don't
+ * bother with the divide:
+ */
+ if (atomic_read(&ca->congested) < CONGESTED_MAX)
+ atomic_add(latency_over >>
+ max_t(int, ilog2(latency_threshold) - 2, 0),
+ &ca->congested);
+
+ ca->congested_last = now;
+ } else if (atomic_read(&ca->congested) > 0) {
+ atomic_dec(&ca->congested);
+ }
+}
+
+void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
+{
+ atomic64_t *latency = &ca->cur_latency[rw];
+ u64 now = local_clock();
+ u64 io_latency = time_after64(now, submit_time)
+ ? now - submit_time
+ : 0;
+ u64 old, new, v = atomic64_read(latency);
+
+ do {
+ old = v;
+
+ /*
+ * If the io latency was reasonably close to the current
+ * latency, skip doing the update and atomic operation - most of
+ * the time:
+ */
+ if (abs((int) (old - io_latency)) < (old >> 1) &&
+ now & ~(~0U << 5))
+ break;
+
+ new = ewma_add(old, io_latency, 5);
+ } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
+
+ bch2_congested_acct(ca, io_latency, now, rw);
+
+ __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
+}
+
+#endif
+
+/* Allocate, free from mempool: */
+
+void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
+{
+ struct bvec_iter_all iter;
+ struct bio_vec *bv;
+
+ bio_for_each_segment_all(bv, bio, iter)
+ if (bv->bv_page != ZERO_PAGE(0))
+ mempool_free(bv->bv_page, &c->bio_bounce_pages);
+ bio->bi_vcnt = 0;
+}
+
+static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
+{
+ struct page *page;
+
+ if (likely(!*using_mempool)) {
+ page = alloc_page(GFP_NOFS);
+ if (unlikely(!page)) {
+ mutex_lock(&c->bio_bounce_pages_lock);
+ *using_mempool = true;
+ goto pool_alloc;
+
+ }
+ } else {
+pool_alloc:
+ page = mempool_alloc(&c->bio_bounce_pages, GFP_NOFS);
+ }
+
+ return page;
+}
+
+void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
+ size_t size)
+{
+ bool using_mempool = false;
+
+ while (size) {
+ struct page *page = __bio_alloc_page_pool(c, &using_mempool);
+ unsigned len = min_t(size_t, PAGE_SIZE, size);
+
+ BUG_ON(!bio_add_page(bio, page, len, 0));
+ size -= len;
+ }
+
+ if (using_mempool)
+ mutex_unlock(&c->bio_bounce_pages_lock);
+}
+
+/* Extent update path: */
+
+int bch2_sum_sector_overwrites(struct btree_trans *trans,
+ struct btree_iter *extent_iter,
+ struct bkey_i *new,
+ bool *usage_increasing,
+ s64 *i_sectors_delta,
+ s64 *disk_sectors_delta)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_s_c old;
+ unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
+ bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
+ int ret = 0;
+
+ *usage_increasing = false;
+ *i_sectors_delta = 0;
+ *disk_sectors_delta = 0;
+
+ bch2_trans_copy_iter(&iter, extent_iter);
+
+ for_each_btree_key_upto_continue_norestart(iter,
+ new->k.p, BTREE_ITER_SLOTS, old, ret) {
+ s64 sectors = min(new->k.p.offset, old.k->p.offset) -
+ max(bkey_start_offset(&new->k),
+ bkey_start_offset(old.k));
+
+ *i_sectors_delta += sectors *
+ (bkey_extent_is_allocation(&new->k) -
+ bkey_extent_is_allocation(old.k));
+
+ *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
+ *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
+ ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
+ : 0;
+
+ if (!*usage_increasing &&
+ (new->k.p.snapshot != old.k->p.snapshot ||
+ new_replicas > bch2_bkey_replicas(c, old) ||
+ (!new_compressed && bch2_bkey_sectors_compressed(old))))
+ *usage_increasing = true;
+
+ if (bkey_ge(old.k->p, new->k.p))
+ break;
+ }
+
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
+ struct btree_iter *extent_iter,
+ u64 new_i_size,
+ s64 i_sectors_delta)
+{
+ struct btree_iter iter;
+ struct bkey_i *k;
+ struct bkey_i_inode_v3 *inode;
+ /*
+ * Crazy performance optimization:
+ * Every extent update needs to also update the inode: the inode trigger
+ * will set bi->journal_seq to the journal sequence number of this
+ * transaction - for fsync.
+ *
+ * But if that's the only reason we're updating the inode (we're not
+ * updating bi_size or bi_sectors), then we don't need the inode update
+ * to be journalled - if we crash, the bi_journal_seq update will be
+ * lost, but that's fine.
+ */
+ unsigned inode_update_flags = BTREE_UPDATE_NOJOURNAL;
+ int ret;
+
+ k = bch2_bkey_get_mut_noupdate(trans, &iter, BTREE_ID_inodes,
+ SPOS(0,
+ extent_iter->pos.inode,
+ extent_iter->snapshot),
+ BTREE_ITER_CACHED);
+ ret = PTR_ERR_OR_ZERO(k);
+ if (unlikely(ret))
+ return ret;
+
+ if (unlikely(k->k.type != KEY_TYPE_inode_v3)) {
+ k = bch2_inode_to_v3(trans, k);
+ ret = PTR_ERR_OR_ZERO(k);
+ if (unlikely(ret))
+ goto err;
+ }
+
+ inode = bkey_i_to_inode_v3(k);
+
+ if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_i_size_dirty) &&
+ new_i_size > le64_to_cpu(inode->v.bi_size)) {
+ inode->v.bi_size = cpu_to_le64(new_i_size);
+ inode_update_flags = 0;
+ }
+
+ if (i_sectors_delta) {
+ le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
+ inode_update_flags = 0;
+ }
+
+ if (inode->k.p.snapshot != iter.snapshot) {
+ inode->k.p.snapshot = iter.snapshot;
+ inode_update_flags = 0;
+ }
+
+ ret = bch2_trans_update(trans, &iter, &inode->k_i,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+ inode_update_flags);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+int bch2_extent_update(struct btree_trans *trans,
+ subvol_inum inum,
+ struct btree_iter *iter,
+ struct bkey_i *k,
+ struct disk_reservation *disk_res,
+ u64 new_i_size,
+ s64 *i_sectors_delta_total,
+ bool check_enospc)
+{
+ struct bpos next_pos;
+ bool usage_increasing;
+ s64 i_sectors_delta = 0, disk_sectors_delta = 0;
+ int ret;
+
+ /*
+ * This traverses us the iterator without changing iter->path->pos to
+ * search_key() (which is pos + 1 for extents): we want there to be a
+ * path already traversed at iter->pos because
+ * bch2_trans_extent_update() will use it to attempt extent merging
+ */
+ ret = __bch2_btree_iter_traverse(iter);
+ if (ret)
+ return ret;
+
+ ret = bch2_extent_trim_atomic(trans, iter, k);
+ if (ret)
+ return ret;
+
+ next_pos = k->k.p;
+
+ ret = bch2_sum_sector_overwrites(trans, iter, k,
+ &usage_increasing,
+ &i_sectors_delta,
+ &disk_sectors_delta);
+ if (ret)
+ return ret;
+
+ if (disk_res &&
+ disk_sectors_delta > (s64) disk_res->sectors) {
+ ret = bch2_disk_reservation_add(trans->c, disk_res,
+ disk_sectors_delta - disk_res->sectors,
+ !check_enospc || !usage_increasing
+ ? BCH_DISK_RESERVATION_NOFAIL : 0);
+ if (ret)
+ return ret;
+ }
+
+ /*
+ * Note:
+ * We always have to do an inode update - even when i_size/i_sectors
+ * aren't changing - for fsync to work properly; fsync relies on
+ * inode->bi_journal_seq which is updated by the trigger code:
+ */
+ ret = bch2_extent_update_i_size_sectors(trans, iter,
+ min(k->k.p.offset << 9, new_i_size),
+ i_sectors_delta) ?:
+ bch2_trans_update(trans, iter, k, 0) ?:
+ bch2_trans_commit(trans, disk_res, NULL,
+ BCH_TRANS_COMMIT_no_check_rw|
+ BCH_TRANS_COMMIT_no_enospc);
+ if (unlikely(ret))
+ return ret;
+
+ if (i_sectors_delta_total)
+ *i_sectors_delta_total += i_sectors_delta;
+ bch2_btree_iter_set_pos(iter, next_pos);
+ return 0;
+}
+
+static int bch2_write_index_default(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct bkey_buf sk;
+ struct keylist *keys = &op->insert_keys;
+ struct bkey_i *k = bch2_keylist_front(keys);
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct btree_iter iter;
+ subvol_inum inum = {
+ .subvol = op->subvol,
+ .inum = k->k.p.inode,
+ };
+ int ret;
+
+ BUG_ON(!inum.subvol);
+
+ bch2_bkey_buf_init(&sk);
+
+ do {
+ bch2_trans_begin(trans);
+
+ k = bch2_keylist_front(keys);
+ bch2_bkey_buf_copy(&sk, c, k);
+
+ ret = bch2_subvolume_get_snapshot(trans, inum.subvol,
+ &sk.k->k.p.snapshot);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ break;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+ bkey_start_pos(&sk.k->k),
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+
+ ret = bch2_bkey_set_needs_rebalance(c, sk.k,
+ op->opts.background_target,
+ op->opts.background_compression) ?:
+ bch2_extent_update(trans, inum, &iter, sk.k,
+ &op->res,
+ op->new_i_size, &op->i_sectors_delta,
+ op->flags & BCH_WRITE_CHECK_ENOSPC);
+ bch2_trans_iter_exit(trans, &iter);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ break;
+
+ if (bkey_ge(iter.pos, k->k.p))
+ bch2_keylist_pop_front(&op->insert_keys);
+ else
+ bch2_cut_front(iter.pos, k);
+ } while (!bch2_keylist_empty(keys));
+
+ bch2_trans_put(trans);
+ bch2_bkey_buf_exit(&sk, c);
+
+ return ret;
+}
+
+/* Writes */
+
+void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
+ enum bch_data_type type,
+ const struct bkey_i *k,
+ bool nocow)
+{
+ struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
+ const struct bch_extent_ptr *ptr;
+ struct bch_write_bio *n;
+ struct bch_dev *ca;
+
+ BUG_ON(c->opts.nochanges);
+
+ bkey_for_each_ptr(ptrs, ptr) {
+ BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
+ !c->devs[ptr->dev]);
+
+ ca = bch_dev_bkey_exists(c, ptr->dev);
+
+ if (to_entry(ptr + 1) < ptrs.end) {
+ n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
+ GFP_NOFS, &ca->replica_set));
+
+ n->bio.bi_end_io = wbio->bio.bi_end_io;
+ n->bio.bi_private = wbio->bio.bi_private;
+ n->parent = wbio;
+ n->split = true;
+ n->bounce = false;
+ n->put_bio = true;
+ n->bio.bi_opf = wbio->bio.bi_opf;
+ bio_inc_remaining(&wbio->bio);
+ } else {
+ n = wbio;
+ n->split = false;
+ }
+
+ n->c = c;
+ n->dev = ptr->dev;
+ n->have_ioref = nocow || bch2_dev_get_ioref(ca,
+ type == BCH_DATA_btree ? READ : WRITE);
+ n->nocow = nocow;
+ n->submit_time = local_clock();
+ n->inode_offset = bkey_start_offset(&k->k);
+ n->bio.bi_iter.bi_sector = ptr->offset;
+
+ if (likely(n->have_ioref)) {
+ this_cpu_add(ca->io_done->sectors[WRITE][type],
+ bio_sectors(&n->bio));
+
+ bio_set_dev(&n->bio, ca->disk_sb.bdev);
+
+ if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
+ bio_endio(&n->bio);
+ continue;
+ }
+
+ submit_bio(&n->bio);
+ } else {
+ n->bio.bi_status = BLK_STS_REMOVED;
+ bio_endio(&n->bio);
+ }
+ }
+}
+
+static void __bch2_write(struct bch_write_op *);
+
+static void bch2_write_done(struct closure *cl)
+{
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct bch_fs *c = op->c;
+
+ EBUG_ON(op->open_buckets.nr);
+
+ bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
+ bch2_disk_reservation_put(c, &op->res);
+
+ if (!(op->flags & BCH_WRITE_MOVE))
+ bch2_write_ref_put(c, BCH_WRITE_REF_write);
+ bch2_keylist_free(&op->insert_keys, op->inline_keys);
+
+ EBUG_ON(cl->parent);
+ closure_debug_destroy(cl);
+ if (op->end_io)
+ op->end_io(op);
+}
+
+static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
+{
+ struct keylist *keys = &op->insert_keys;
+ struct bch_extent_ptr *ptr;
+ struct bkey_i *src, *dst = keys->keys, *n;
+
+ for (src = keys->keys; src != keys->top; src = n) {
+ n = bkey_next(src);
+
+ if (bkey_extent_is_direct_data(&src->k)) {
+ bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
+ test_bit(ptr->dev, op->failed.d));
+
+ if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
+ return -EIO;
+ }
+
+ if (dst != src)
+ memmove_u64s_down(dst, src, src->k.u64s);
+ dst = bkey_next(dst);
+ }
+
+ keys->top = dst;
+ return 0;
+}
+
+/**
+ * __bch2_write_index - after a write, update index to point to new data
+ * @op: bch_write_op to process
+ */
+static void __bch2_write_index(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct keylist *keys = &op->insert_keys;
+ unsigned dev;
+ int ret = 0;
+
+ if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
+ ret = bch2_write_drop_io_error_ptrs(op);
+ if (ret)
+ goto err;
+ }
+
+ if (!bch2_keylist_empty(keys)) {
+ u64 sectors_start = keylist_sectors(keys);
+
+ ret = !(op->flags & BCH_WRITE_MOVE)
+ ? bch2_write_index_default(op)
+ : bch2_data_update_index_update(op);
+
+ BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
+ BUG_ON(keylist_sectors(keys) && !ret);
+
+ op->written += sectors_start - keylist_sectors(keys);
+
+ if (ret && !bch2_err_matches(ret, EROFS)) {
+ struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
+
+ bch_err_inum_offset_ratelimited(c,
+ insert->k.p.inode, insert->k.p.offset << 9,
+ "write error while doing btree update: %s",
+ bch2_err_str(ret));
+ }
+
+ if (ret)
+ goto err;
+ }
+out:
+ /* If some a bucket wasn't written, we can't erasure code it: */
+ for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
+ bch2_open_bucket_write_error(c, &op->open_buckets, dev);
+
+ bch2_open_buckets_put(c, &op->open_buckets);
+ return;
+err:
+ keys->top = keys->keys;
+ op->error = ret;
+ op->flags |= BCH_WRITE_DONE;
+ goto out;
+}
+
+static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
+{
+ if (state != wp->state) {
+ u64 now = ktime_get_ns();
+
+ if (wp->last_state_change &&
+ time_after64(now, wp->last_state_change))
+ wp->time[wp->state] += now - wp->last_state_change;
+ wp->state = state;
+ wp->last_state_change = now;
+ }
+}
+
+static inline void wp_update_state(struct write_point *wp, bool running)
+{
+ enum write_point_state state;
+
+ state = running ? WRITE_POINT_running :
+ !list_empty(&wp->writes) ? WRITE_POINT_waiting_io
+ : WRITE_POINT_stopped;
+
+ __wp_update_state(wp, state);
+}
+
+static void bch2_write_index(struct closure *cl)
+{
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct write_point *wp = op->wp;
+ struct workqueue_struct *wq = index_update_wq(op);
+ unsigned long flags;
+
+ if ((op->flags & BCH_WRITE_DONE) &&
+ (op->flags & BCH_WRITE_MOVE))
+ bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
+
+ spin_lock_irqsave(&wp->writes_lock, flags);
+ if (wp->state == WRITE_POINT_waiting_io)
+ __wp_update_state(wp, WRITE_POINT_waiting_work);
+ list_add_tail(&op->wp_list, &wp->writes);
+ spin_unlock_irqrestore (&wp->writes_lock, flags);
+
+ queue_work(wq, &wp->index_update_work);
+}
+
+static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
+{
+ op->wp = wp;
+
+ if (wp->state == WRITE_POINT_stopped) {
+ spin_lock_irq(&wp->writes_lock);
+ __wp_update_state(wp, WRITE_POINT_waiting_io);
+ spin_unlock_irq(&wp->writes_lock);
+ }
+}
+
+void bch2_write_point_do_index_updates(struct work_struct *work)
+{
+ struct write_point *wp =
+ container_of(work, struct write_point, index_update_work);
+ struct bch_write_op *op;
+
+ while (1) {
+ spin_lock_irq(&wp->writes_lock);
+ op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
+ if (op)
+ list_del(&op->wp_list);
+ wp_update_state(wp, op != NULL);
+ spin_unlock_irq(&wp->writes_lock);
+
+ if (!op)
+ break;
+
+ op->flags |= BCH_WRITE_IN_WORKER;
+
+ __bch2_write_index(op);
+
+ if (!(op->flags & BCH_WRITE_DONE))
+ __bch2_write(op);
+ else
+ bch2_write_done(&op->cl);
+ }
+}
+
+static void bch2_write_endio(struct bio *bio)
+{
+ struct closure *cl = bio->bi_private;
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct bch_write_bio *wbio = to_wbio(bio);
+ struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
+ struct bch_fs *c = wbio->c;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
+
+ if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
+ op->pos.inode,
+ wbio->inode_offset << 9,
+ "data write error: %s",
+ bch2_blk_status_to_str(bio->bi_status))) {
+ set_bit(wbio->dev, op->failed.d);
+ op->flags |= BCH_WRITE_IO_ERROR;
+ }
+
+ if (wbio->nocow)
+ set_bit(wbio->dev, op->devs_need_flush->d);
+
+ if (wbio->have_ioref) {
+ bch2_latency_acct(ca, wbio->submit_time, WRITE);
+ percpu_ref_put(&ca->io_ref);
+ }
+
+ if (wbio->bounce)
+ bch2_bio_free_pages_pool(c, bio);
+
+ if (wbio->put_bio)
+ bio_put(bio);
+
+ if (parent)
+ bio_endio(&parent->bio);
+ else
+ closure_put(cl);
+}
+
+static void init_append_extent(struct bch_write_op *op,
+ struct write_point *wp,
+ struct bversion version,
+ struct bch_extent_crc_unpacked crc)
+{
+ struct bkey_i_extent *e;
+
+ op->pos.offset += crc.uncompressed_size;
+
+ e = bkey_extent_init(op->insert_keys.top);
+ e->k.p = op->pos;
+ e->k.size = crc.uncompressed_size;
+ e->k.version = version;
+
+ if (crc.csum_type ||
+ crc.compression_type ||
+ crc.nonce)
+ bch2_extent_crc_append(&e->k_i, crc);
+
+ bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
+ op->flags & BCH_WRITE_CACHED);
+
+ bch2_keylist_push(&op->insert_keys);
+}
+
+static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
+ struct write_point *wp,
+ struct bio *src,
+ bool *page_alloc_failed,
+ void *buf)
+{
+ struct bch_write_bio *wbio;
+ struct bio *bio;
+ unsigned output_available =
+ min(wp->sectors_free << 9, src->bi_iter.bi_size);
+ unsigned pages = DIV_ROUND_UP(output_available +
+ (buf
+ ? ((unsigned long) buf & (PAGE_SIZE - 1))
+ : 0), PAGE_SIZE);
+
+ pages = min(pages, BIO_MAX_VECS);
+
+ bio = bio_alloc_bioset(NULL, pages, 0,
+ GFP_NOFS, &c->bio_write);
+ wbio = wbio_init(bio);
+ wbio->put_bio = true;
+ /* copy WRITE_SYNC flag */
+ wbio->bio.bi_opf = src->bi_opf;
+
+ if (buf) {
+ bch2_bio_map(bio, buf, output_available);
+ return bio;
+ }
+
+ wbio->bounce = true;
+
+ /*
+ * We can't use mempool for more than c->sb.encoded_extent_max
+ * worth of pages, but we'd like to allocate more if we can:
+ */
+ bch2_bio_alloc_pages_pool(c, bio,
+ min_t(unsigned, output_available,
+ c->opts.encoded_extent_max));
+
+ if (bio->bi_iter.bi_size < output_available)
+ *page_alloc_failed =
+ bch2_bio_alloc_pages(bio,
+ output_available -
+ bio->bi_iter.bi_size,
+ GFP_NOFS) != 0;
+
+ return bio;
+}
+
+static int bch2_write_rechecksum(struct bch_fs *c,
+ struct bch_write_op *op,
+ unsigned new_csum_type)
+{
+ struct bio *bio = &op->wbio.bio;
+ struct bch_extent_crc_unpacked new_crc;
+ int ret;
+
+ /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
+
+ if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
+ bch2_csum_type_is_encryption(new_csum_type))
+ new_csum_type = op->crc.csum_type;
+
+ ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
+ NULL, &new_crc,
+ op->crc.offset, op->crc.live_size,
+ new_csum_type);
+ if (ret)
+ return ret;
+
+ bio_advance(bio, op->crc.offset << 9);
+ bio->bi_iter.bi_size = op->crc.live_size << 9;
+ op->crc = new_crc;
+ return 0;
+}
+
+static int bch2_write_decrypt(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct nonce nonce = extent_nonce(op->version, op->crc);
+ struct bch_csum csum;
+ int ret;
+
+ if (!bch2_csum_type_is_encryption(op->crc.csum_type))
+ return 0;
+
+ /*
+ * If we need to decrypt data in the write path, we'll no longer be able
+ * to verify the existing checksum (poly1305 mac, in this case) after
+ * it's decrypted - this is the last point we'll be able to reverify the
+ * checksum:
+ */
+ csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
+ if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
+ return -EIO;
+
+ ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
+ op->crc.csum_type = 0;
+ op->crc.csum = (struct bch_csum) { 0, 0 };
+ return ret;
+}
+
+static enum prep_encoded_ret {
+ PREP_ENCODED_OK,
+ PREP_ENCODED_ERR,
+ PREP_ENCODED_CHECKSUM_ERR,
+ PREP_ENCODED_DO_WRITE,
+} bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
+{
+ struct bch_fs *c = op->c;
+ struct bio *bio = &op->wbio.bio;
+
+ if (!(op->flags & BCH_WRITE_DATA_ENCODED))
+ return PREP_ENCODED_OK;
+
+ BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
+
+ /* Can we just write the entire extent as is? */
+ if (op->crc.uncompressed_size == op->crc.live_size &&
+ op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
+ op->crc.compressed_size <= wp->sectors_free &&
+ (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
+ op->incompressible)) {
+ if (!crc_is_compressed(op->crc) &&
+ op->csum_type != op->crc.csum_type &&
+ bch2_write_rechecksum(c, op, op->csum_type) &&
+ !c->opts.no_data_io)
+ return PREP_ENCODED_CHECKSUM_ERR;
+
+ return PREP_ENCODED_DO_WRITE;
+ }
+
+ /*
+ * If the data is compressed and we couldn't write the entire extent as
+ * is, we have to decompress it:
+ */
+ if (crc_is_compressed(op->crc)) {
+ struct bch_csum csum;
+
+ if (bch2_write_decrypt(op))
+ return PREP_ENCODED_CHECKSUM_ERR;
+
+ /* Last point we can still verify checksum: */
+ csum = bch2_checksum_bio(c, op->crc.csum_type,
+ extent_nonce(op->version, op->crc),
+ bio);
+ if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
+ return PREP_ENCODED_CHECKSUM_ERR;
+
+ if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
+ return PREP_ENCODED_ERR;
+ }
+
+ /*
+ * No longer have compressed data after this point - data might be
+ * encrypted:
+ */
+
+ /*
+ * If the data is checksummed and we're only writing a subset,
+ * rechecksum and adjust bio to point to currently live data:
+ */
+ if ((op->crc.live_size != op->crc.uncompressed_size ||
+ op->crc.csum_type != op->csum_type) &&
+ bch2_write_rechecksum(c, op, op->csum_type) &&
+ !c->opts.no_data_io)
+ return PREP_ENCODED_CHECKSUM_ERR;
+
+ /*
+ * If we want to compress the data, it has to be decrypted:
+ */
+ if ((op->compression_opt ||
+ bch2_csum_type_is_encryption(op->crc.csum_type) !=
+ bch2_csum_type_is_encryption(op->csum_type)) &&
+ bch2_write_decrypt(op))
+ return PREP_ENCODED_CHECKSUM_ERR;
+
+ return PREP_ENCODED_OK;
+}
+
+static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
+ struct bio **_dst)
+{
+ struct bch_fs *c = op->c;
+ struct bio *src = &op->wbio.bio, *dst = src;
+ struct bvec_iter saved_iter;
+ void *ec_buf;
+ unsigned total_output = 0, total_input = 0;
+ bool bounce = false;
+ bool page_alloc_failed = false;
+ int ret, more = 0;
+
+ BUG_ON(!bio_sectors(src));
+
+ ec_buf = bch2_writepoint_ec_buf(c, wp);
+
+ switch (bch2_write_prep_encoded_data(op, wp)) {
+ case PREP_ENCODED_OK:
+ break;
+ case PREP_ENCODED_ERR:
+ ret = -EIO;
+ goto err;
+ case PREP_ENCODED_CHECKSUM_ERR:
+ goto csum_err;
+ case PREP_ENCODED_DO_WRITE:
+ /* XXX look for bug here */
+ if (ec_buf) {
+ dst = bch2_write_bio_alloc(c, wp, src,
+ &page_alloc_failed,
+ ec_buf);
+ bio_copy_data(dst, src);
+ bounce = true;
+ }
+ init_append_extent(op, wp, op->version, op->crc);
+ goto do_write;
+ }
+
+ if (ec_buf ||
+ op->compression_opt ||
+ (op->csum_type &&
+ !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
+ (bch2_csum_type_is_encryption(op->csum_type) &&
+ !(op->flags & BCH_WRITE_PAGES_OWNED))) {
+ dst = bch2_write_bio_alloc(c, wp, src,
+ &page_alloc_failed,
+ ec_buf);
+ bounce = true;
+ }
+
+ saved_iter = dst->bi_iter;
+
+ do {
+ struct bch_extent_crc_unpacked crc = { 0 };
+ struct bversion version = op->version;
+ size_t dst_len = 0, src_len = 0;
+
+ if (page_alloc_failed &&
+ dst->bi_iter.bi_size < (wp->sectors_free << 9) &&
+ dst->bi_iter.bi_size < c->opts.encoded_extent_max)
+ break;
+
+ BUG_ON(op->compression_opt &&
+ (op->flags & BCH_WRITE_DATA_ENCODED) &&
+ bch2_csum_type_is_encryption(op->crc.csum_type));
+ BUG_ON(op->compression_opt && !bounce);
+
+ crc.compression_type = op->incompressible
+ ? BCH_COMPRESSION_TYPE_incompressible
+ : op->compression_opt
+ ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
+ op->compression_opt)
+ : 0;
+ if (!crc_is_compressed(crc)) {
+ dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
+ dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
+
+ if (op->csum_type)
+ dst_len = min_t(unsigned, dst_len,
+ c->opts.encoded_extent_max);
+
+ if (bounce) {
+ swap(dst->bi_iter.bi_size, dst_len);
+ bio_copy_data(dst, src);
+ swap(dst->bi_iter.bi_size, dst_len);
+ }
+
+ src_len = dst_len;
+ }
+
+ BUG_ON(!src_len || !dst_len);
+
+ if (bch2_csum_type_is_encryption(op->csum_type)) {
+ if (bversion_zero(version)) {
+ version.lo = atomic64_inc_return(&c->key_version);
+ } else {
+ crc.nonce = op->nonce;
+ op->nonce += src_len >> 9;
+ }
+ }
+
+ if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
+ !crc_is_compressed(crc) &&
+ bch2_csum_type_is_encryption(op->crc.csum_type) ==
+ bch2_csum_type_is_encryption(op->csum_type)) {
+ u8 compression_type = crc.compression_type;
+ u16 nonce = crc.nonce;
+ /*
+ * Note: when we're using rechecksum(), we need to be
+ * checksumming @src because it has all the data our
+ * existing checksum covers - if we bounced (because we
+ * were trying to compress), @dst will only have the
+ * part of the data the new checksum will cover.
+ *
+ * But normally we want to be checksumming post bounce,
+ * because part of the reason for bouncing is so the
+ * data can't be modified (by userspace) while it's in
+ * flight.
+ */
+ if (bch2_rechecksum_bio(c, src, version, op->crc,
+ &crc, &op->crc,
+ src_len >> 9,
+ bio_sectors(src) - (src_len >> 9),
+ op->csum_type))
+ goto csum_err;
+ /*
+ * rchecksum_bio sets compression_type on crc from op->crc,
+ * this isn't always correct as sometimes we're changing
+ * an extent from uncompressed to incompressible.
+ */
+ crc.compression_type = compression_type;
+ crc.nonce = nonce;
+ } else {
+ if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
+ bch2_rechecksum_bio(c, src, version, op->crc,
+ NULL, &op->crc,
+ src_len >> 9,
+ bio_sectors(src) - (src_len >> 9),
+ op->crc.csum_type))
+ goto csum_err;
+
+ crc.compressed_size = dst_len >> 9;
+ crc.uncompressed_size = src_len >> 9;
+ crc.live_size = src_len >> 9;
+
+ swap(dst->bi_iter.bi_size, dst_len);
+ ret = bch2_encrypt_bio(c, op->csum_type,
+ extent_nonce(version, crc), dst);
+ if (ret)
+ goto err;
+
+ crc.csum = bch2_checksum_bio(c, op->csum_type,
+ extent_nonce(version, crc), dst);
+ crc.csum_type = op->csum_type;
+ swap(dst->bi_iter.bi_size, dst_len);
+ }
+
+ init_append_extent(op, wp, version, crc);
+
+ if (dst != src)
+ bio_advance(dst, dst_len);
+ bio_advance(src, src_len);
+ total_output += dst_len;
+ total_input += src_len;
+ } while (dst->bi_iter.bi_size &&
+ src->bi_iter.bi_size &&
+ wp->sectors_free &&
+ !bch2_keylist_realloc(&op->insert_keys,
+ op->inline_keys,
+ ARRAY_SIZE(op->inline_keys),
+ BKEY_EXTENT_U64s_MAX));
+
+ more = src->bi_iter.bi_size != 0;
+
+ dst->bi_iter = saved_iter;
+
+ if (dst == src && more) {
+ BUG_ON(total_output != total_input);
+
+ dst = bio_split(src, total_input >> 9,
+ GFP_NOFS, &c->bio_write);
+ wbio_init(dst)->put_bio = true;
+ /* copy WRITE_SYNC flag */
+ dst->bi_opf = src->bi_opf;
+ }
+
+ dst->bi_iter.bi_size = total_output;
+do_write:
+ *_dst = dst;
+ return more;
+csum_err:
+ bch_err(c, "error verifying existing checksum while rewriting existing data (memory corruption?)");
+ ret = -EIO;
+err:
+ if (to_wbio(dst)->bounce)
+ bch2_bio_free_pages_pool(c, dst);
+ if (to_wbio(dst)->put_bio)
+ bio_put(dst);
+
+ return ret;
+}
+
+static bool bch2_extent_is_writeable(struct bch_write_op *op,
+ struct bkey_s_c k)
+{
+ struct bch_fs *c = op->c;
+ struct bkey_s_c_extent e;
+ struct extent_ptr_decoded p;
+ const union bch_extent_entry *entry;
+ unsigned replicas = 0;
+
+ if (k.k->type != KEY_TYPE_extent)
+ return false;
+
+ e = bkey_s_c_to_extent(k);
+ extent_for_each_ptr_decode(e, p, entry) {
+ if (crc_is_encoded(p.crc) || p.has_ec)
+ return false;
+
+ replicas += bch2_extent_ptr_durability(c, &p);
+ }
+
+ return replicas >= op->opts.data_replicas;
+}
+
+static inline void bch2_nocow_write_unlock(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ const struct bch_extent_ptr *ptr;
+ struct bkey_i *k;
+
+ for_each_keylist_key(&op->insert_keys, k) {
+ struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
+
+ bkey_for_each_ptr(ptrs, ptr)
+ bch2_bucket_nocow_unlock(&c->nocow_locks,
+ PTR_BUCKET_POS(c, ptr),
+ BUCKET_NOCOW_LOCK_UPDATE);
+ }
+}
+
+static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_i *orig,
+ struct bkey_s_c k,
+ u64 new_i_size)
+{
+ struct bkey_i *new;
+ struct bkey_ptrs ptrs;
+ struct bch_extent_ptr *ptr;
+ int ret;
+
+ if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
+ /* trace this */
+ return 0;
+ }
+
+ new = bch2_bkey_make_mut_noupdate(trans, k);
+ ret = PTR_ERR_OR_ZERO(new);
+ if (ret)
+ return ret;
+
+ bch2_cut_front(bkey_start_pos(&orig->k), new);
+ bch2_cut_back(orig->k.p, new);
+
+ ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
+ bkey_for_each_ptr(ptrs, ptr)
+ ptr->unwritten = 0;
+
+ /*
+ * Note that we're not calling bch2_subvol_get_snapshot() in this path -
+ * that was done when we kicked off the write, and here it's important
+ * that we update the extent that we wrote to - even if a snapshot has
+ * since been created. The write is still outstanding, so we're ok
+ * w.r.t. snapshot atomicity:
+ */
+ return bch2_extent_update_i_size_sectors(trans, iter,
+ min(new->k.p.offset << 9, new_i_size), 0) ?:
+ bch2_trans_update(trans, iter, new,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+}
+
+static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct btree_trans *trans = bch2_trans_get(c);
+ struct btree_iter iter;
+ struct bkey_i *orig;
+ struct bkey_s_c k;
+ int ret;
+
+ for_each_keylist_key(&op->insert_keys, orig) {
+ ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
+ bkey_start_pos(&orig->k), orig->k.p,
+ BTREE_ITER_INTENT, k,
+ NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
+ bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
+ }));
+
+ if (ret && !bch2_err_matches(ret, EROFS)) {
+ struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
+
+ bch_err_inum_offset_ratelimited(c,
+ insert->k.p.inode, insert->k.p.offset << 9,
+ "write error while doing btree update: %s",
+ bch2_err_str(ret));
+ }
+
+ if (ret) {
+ op->error = ret;
+ break;
+ }
+ }
+
+ bch2_trans_put(trans);
+}
+
+static void __bch2_nocow_write_done(struct bch_write_op *op)
+{
+ bch2_nocow_write_unlock(op);
+
+ if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
+ op->error = -EIO;
+ } else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
+ bch2_nocow_write_convert_unwritten(op);
+}
+
+static void bch2_nocow_write_done(struct closure *cl)
+{
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+
+ __bch2_nocow_write_done(op);
+ bch2_write_done(cl);
+}
+
+static void bch2_nocow_write(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct btree_trans *trans;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ struct bkey_ptrs_c ptrs;
+ const struct bch_extent_ptr *ptr;
+ struct {
+ struct bpos b;
+ unsigned gen;
+ struct nocow_lock_bucket *l;
+ } buckets[BCH_REPLICAS_MAX];
+ unsigned nr_buckets = 0;
+ u32 snapshot;
+ int ret, i;
+
+ if (op->flags & BCH_WRITE_MOVE)
+ return;
+
+ trans = bch2_trans_get(c);
+retry:
+ bch2_trans_begin(trans);
+
+ ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
+ if (unlikely(ret))
+ goto err;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+ SPOS(op->pos.inode, op->pos.offset, snapshot),
+ BTREE_ITER_SLOTS);
+ while (1) {
+ struct bio *bio = &op->wbio.bio;
+
+ nr_buckets = 0;
+
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ break;
+
+ /* fall back to normal cow write path? */
+ if (unlikely(k.k->p.snapshot != snapshot ||
+ !bch2_extent_is_writeable(op, k)))
+ break;
+
+ if (bch2_keylist_realloc(&op->insert_keys,
+ op->inline_keys,
+ ARRAY_SIZE(op->inline_keys),
+ k.k->u64s))
+ break;
+
+ /* Get iorefs before dropping btree locks: */
+ ptrs = bch2_bkey_ptrs_c(k);
+ bkey_for_each_ptr(ptrs, ptr) {
+ buckets[nr_buckets].b = PTR_BUCKET_POS(c, ptr);
+ buckets[nr_buckets].gen = ptr->gen;
+ buckets[nr_buckets].l =
+ bucket_nocow_lock(&c->nocow_locks,
+ bucket_to_u64(buckets[nr_buckets].b));
+
+ prefetch(buckets[nr_buckets].l);
+
+ if (unlikely(!bch2_dev_get_ioref(bch_dev_bkey_exists(c, ptr->dev), WRITE)))
+ goto err_get_ioref;
+
+ nr_buckets++;
+
+ if (ptr->unwritten)
+ op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
+ }
+
+ /* Unlock before taking nocow locks, doing IO: */
+ bkey_reassemble(op->insert_keys.top, k);
+ bch2_trans_unlock(trans);
+
+ bch2_cut_front(op->pos, op->insert_keys.top);
+ if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
+ bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
+
+ for (i = 0; i < nr_buckets; i++) {
+ struct bch_dev *ca = bch_dev_bkey_exists(c, buckets[i].b.inode);
+ struct nocow_lock_bucket *l = buckets[i].l;
+ bool stale;
+
+ __bch2_bucket_nocow_lock(&c->nocow_locks, l,
+ bucket_to_u64(buckets[i].b),
+ BUCKET_NOCOW_LOCK_UPDATE);
+
+ rcu_read_lock();
+ stale = gen_after(*bucket_gen(ca, buckets[i].b.offset), buckets[i].gen);
+ rcu_read_unlock();
+
+ if (unlikely(stale))
+ goto err_bucket_stale;
+ }
+
+ bio = &op->wbio.bio;
+ if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
+ bio = bio_split(bio, k.k->p.offset - op->pos.offset,
+ GFP_KERNEL, &c->bio_write);
+ wbio_init(bio)->put_bio = true;
+ bio->bi_opf = op->wbio.bio.bi_opf;
+ } else {
+ op->flags |= BCH_WRITE_DONE;
+ }
+
+ op->pos.offset += bio_sectors(bio);
+ op->written += bio_sectors(bio);
+
+ bio->bi_end_io = bch2_write_endio;
+ bio->bi_private = &op->cl;
+ bio->bi_opf |= REQ_OP_WRITE;
+ closure_get(&op->cl);
+ bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
+ op->insert_keys.top, true);
+
+ bch2_keylist_push(&op->insert_keys);
+ if (op->flags & BCH_WRITE_DONE)
+ break;
+ bch2_btree_iter_advance(&iter);
+ }
+out:
+ bch2_trans_iter_exit(trans, &iter);
+err:
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ goto retry;
+
+ if (ret) {
+ bch_err_inum_offset_ratelimited(c,
+ op->pos.inode,
+ op->pos.offset << 9,
+ "%s: btree lookup error %s",
+ __func__, bch2_err_str(ret));
+ op->error = ret;
+ op->flags |= BCH_WRITE_DONE;
+ }
+
+ bch2_trans_put(trans);
+
+ /* fallback to cow write path? */
+ if (!(op->flags & BCH_WRITE_DONE)) {
+ closure_sync(&op->cl);
+ __bch2_nocow_write_done(op);
+ op->insert_keys.top = op->insert_keys.keys;
+ } else if (op->flags & BCH_WRITE_SYNC) {
+ closure_sync(&op->cl);
+ bch2_nocow_write_done(&op->cl);
+ } else {
+ /*
+ * XXX
+ * needs to run out of process context because ei_quota_lock is
+ * a mutex
+ */
+ continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
+ }
+ return;
+err_get_ioref:
+ for (i = 0; i < nr_buckets; i++)
+ percpu_ref_put(&bch_dev_bkey_exists(c, buckets[i].b.inode)->io_ref);
+
+ /* Fall back to COW path: */
+ goto out;
+err_bucket_stale:
+ while (i >= 0) {
+ bch2_bucket_nocow_unlock(&c->nocow_locks,
+ buckets[i].b,
+ BUCKET_NOCOW_LOCK_UPDATE);
+ --i;
+ }
+ for (i = 0; i < nr_buckets; i++)
+ percpu_ref_put(&bch_dev_bkey_exists(c, buckets[i].b.inode)->io_ref);
+
+ /* We can retry this: */
+ ret = -BCH_ERR_transaction_restart;
+ goto out;
+}
+
+static void __bch2_write(struct bch_write_op *op)
+{
+ struct bch_fs *c = op->c;
+ struct write_point *wp = NULL;
+ struct bio *bio = NULL;
+ unsigned nofs_flags;
+ int ret;
+
+ nofs_flags = memalloc_nofs_save();
+
+ if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
+ bch2_nocow_write(op);
+ if (op->flags & BCH_WRITE_DONE)
+ goto out_nofs_restore;
+ }
+again:
+ memset(&op->failed, 0, sizeof(op->failed));
+
+ do {
+ struct bkey_i *key_to_write;
+ unsigned key_to_write_offset = op->insert_keys.top_p -
+ op->insert_keys.keys_p;
+
+ /* +1 for possible cache device: */
+ if (op->open_buckets.nr + op->nr_replicas + 1 >
+ ARRAY_SIZE(op->open_buckets.v))
+ break;
+
+ if (bch2_keylist_realloc(&op->insert_keys,
+ op->inline_keys,
+ ARRAY_SIZE(op->inline_keys),
+ BKEY_EXTENT_U64s_MAX))
+ break;
+
+ /*
+ * The copygc thread is now global, which means it's no longer
+ * freeing up space on specific disks, which means that
+ * allocations for specific disks may hang arbitrarily long:
+ */
+ ret = bch2_trans_do(c, NULL, NULL, 0,
+ bch2_alloc_sectors_start_trans(trans,
+ op->target,
+ op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
+ op->write_point,
+ &op->devs_have,
+ op->nr_replicas,
+ op->nr_replicas_required,
+ op->watermark,
+ op->flags,
+ (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
+ BCH_WRITE_ONLY_SPECIFIED_DEVS))
+ ? NULL : &op->cl, &wp));
+ if (unlikely(ret)) {
+ if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
+ break;
+
+ goto err;
+ }
+
+ EBUG_ON(!wp);
+
+ bch2_open_bucket_get(c, wp, &op->open_buckets);
+ ret = bch2_write_extent(op, wp, &bio);
+
+ bch2_alloc_sectors_done_inlined(c, wp);
+err:
+ if (ret <= 0) {
+ op->flags |= BCH_WRITE_DONE;
+
+ if (ret < 0) {
+ op->error = ret;
+ break;
+ }
+ }
+
+ bio->bi_end_io = bch2_write_endio;
+ bio->bi_private = &op->cl;
+ bio->bi_opf |= REQ_OP_WRITE;
+
+ closure_get(bio->bi_private);
+
+ key_to_write = (void *) (op->insert_keys.keys_p +
+ key_to_write_offset);
+
+ bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
+ key_to_write, false);
+ } while (ret);
+
+ /*
+ * Sync or no?
+ *
+ * If we're running asynchronously, wne may still want to block
+ * synchronously here if we weren't able to submit all of the IO at
+ * once, as that signals backpressure to the caller.
+ */
+ if ((op->flags & BCH_WRITE_SYNC) ||
+ (!(op->flags & BCH_WRITE_DONE) &&
+ !(op->flags & BCH_WRITE_IN_WORKER))) {
+ closure_sync(&op->cl);
+ __bch2_write_index(op);
+
+ if (!(op->flags & BCH_WRITE_DONE))
+ goto again;
+ bch2_write_done(&op->cl);
+ } else {
+ bch2_write_queue(op, wp);
+ continue_at(&op->cl, bch2_write_index, NULL);
+ }
+out_nofs_restore:
+ memalloc_nofs_restore(nofs_flags);
+}
+
+static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
+{
+ struct bio *bio = &op->wbio.bio;
+ struct bvec_iter iter;
+ struct bkey_i_inline_data *id;
+ unsigned sectors;
+ int ret;
+
+ op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
+ op->flags |= BCH_WRITE_DONE;
+
+ bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
+
+ ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
+ ARRAY_SIZE(op->inline_keys),
+ BKEY_U64s + DIV_ROUND_UP(data_len, 8));
+ if (ret) {
+ op->error = ret;
+ goto err;
+ }
+
+ sectors = bio_sectors(bio);
+ op->pos.offset += sectors;
+
+ id = bkey_inline_data_init(op->insert_keys.top);
+ id->k.p = op->pos;
+ id->k.version = op->version;
+ id->k.size = sectors;
+
+ iter = bio->bi_iter;
+ iter.bi_size = data_len;
+ memcpy_from_bio(id->v.data, bio, iter);
+
+ while (data_len & 7)
+ id->v.data[data_len++] = '\0';
+ set_bkey_val_bytes(&id->k, data_len);
+ bch2_keylist_push(&op->insert_keys);
+
+ __bch2_write_index(op);
+err:
+ bch2_write_done(&op->cl);
+}
+
+/**
+ * bch2_write() - handle a write to a cache device or flash only volume
+ * @cl: &bch_write_op->cl
+ *
+ * This is the starting point for any data to end up in a cache device; it could
+ * be from a normal write, or a writeback write, or a write to a flash only
+ * volume - it's also used by the moving garbage collector to compact data in
+ * mostly empty buckets.
+ *
+ * It first writes the data to the cache, creating a list of keys to be inserted
+ * (if the data won't fit in a single open bucket, there will be multiple keys);
+ * after the data is written it calls bch_journal, and after the keys have been
+ * added to the next journal write they're inserted into the btree.
+ *
+ * If op->discard is true, instead of inserting the data it invalidates the
+ * region of the cache represented by op->bio and op->inode.
+ */
+void bch2_write(struct closure *cl)
+{
+ struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
+ struct bio *bio = &op->wbio.bio;
+ struct bch_fs *c = op->c;
+ unsigned data_len;
+
+ EBUG_ON(op->cl.parent);
+ BUG_ON(!op->nr_replicas);
+ BUG_ON(!op->write_point.v);
+ BUG_ON(bkey_eq(op->pos, POS_MAX));
+
+ op->start_time = local_clock();
+ bch2_keylist_init(&op->insert_keys, op->inline_keys);
+ wbio_init(bio)->put_bio = false;
+
+ if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
+ bch_err_inum_offset_ratelimited(c,
+ op->pos.inode,
+ op->pos.offset << 9,
+ "misaligned write");
+ op->error = -EIO;
+ goto err;
+ }
+
+ if (c->opts.nochanges) {
+ op->error = -BCH_ERR_erofs_no_writes;
+ goto err;
+ }
+
+ if (!(op->flags & BCH_WRITE_MOVE) &&
+ !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
+ op->error = -BCH_ERR_erofs_no_writes;
+ goto err;
+ }
+
+ this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
+ bch2_increment_clock(c, bio_sectors(bio), WRITE);
+
+ data_len = min_t(u64, bio->bi_iter.bi_size,
+ op->new_i_size - (op->pos.offset << 9));
+
+ if (c->opts.inline_data &&
+ data_len <= min(block_bytes(c) / 2, 1024U)) {
+ bch2_write_data_inline(op, data_len);
+ return;
+ }
+
+ __bch2_write(op);
+ return;
+err:
+ bch2_disk_reservation_put(c, &op->res);
+
+ closure_debug_destroy(&op->cl);
+ if (op->end_io)
+ op->end_io(op);
+}
+
+static const char * const bch2_write_flags[] = {
+#define x(f) #f,
+ BCH_WRITE_FLAGS()
+#undef x
+ NULL
+};
+
+void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
+{
+ prt_str(out, "pos: ");
+ bch2_bpos_to_text(out, op->pos);
+ prt_newline(out);
+ printbuf_indent_add(out, 2);
+
+ prt_str(out, "started: ");
+ bch2_pr_time_units(out, local_clock() - op->start_time);
+ prt_newline(out);
+
+ prt_str(out, "flags: ");
+ prt_bitflags(out, bch2_write_flags, op->flags);
+ prt_newline(out);
+
+ prt_printf(out, "ref: %u", closure_nr_remaining(&op->cl));
+ prt_newline(out);
+
+ printbuf_indent_sub(out, 2);
+}
+
+void bch2_fs_io_write_exit(struct bch_fs *c)
+{
+ mempool_exit(&c->bio_bounce_pages);
+ bioset_exit(&c->bio_write);
+}
+
+int bch2_fs_io_write_init(struct bch_fs *c)
+{
+ if (bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
+ BIOSET_NEED_BVECS))
+ return -BCH_ERR_ENOMEM_bio_write_init;
+
+ if (mempool_init_page_pool(&c->bio_bounce_pages,
+ max_t(unsigned,
+ c->opts.btree_node_size,
+ c->opts.encoded_extent_max) /
+ PAGE_SIZE, 0))
+ return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
+
+ return 0;
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_WRITE_H
+#define _BCACHEFS_IO_WRITE_H
+
+#include "checksum.h"
+#include "io_write_types.h"
+
+#define to_wbio(_bio) \
+ container_of((_bio), struct bch_write_bio, bio)
+
+void bch2_bio_free_pages_pool(struct bch_fs *, struct bio *);
+void bch2_bio_alloc_pages_pool(struct bch_fs *, struct bio *, size_t);
+
+#ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
+void bch2_latency_acct(struct bch_dev *, u64, int);
+#else
+static inline void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw) {}
+#endif
+
+void bch2_submit_wbio_replicas(struct bch_write_bio *, struct bch_fs *,
+ enum bch_data_type, const struct bkey_i *, bool);
+
+#define BCH_WRITE_FLAGS() \
+ x(ALLOC_NOWAIT) \
+ x(CACHED) \
+ x(DATA_ENCODED) \
+ x(PAGES_STABLE) \
+ x(PAGES_OWNED) \
+ x(ONLY_SPECIFIED_DEVS) \
+ x(WROTE_DATA_INLINE) \
+ x(FROM_INTERNAL) \
+ x(CHECK_ENOSPC) \
+ x(SYNC) \
+ x(MOVE) \
+ x(IN_WORKER) \
+ x(DONE) \
+ x(IO_ERROR) \
+ x(CONVERT_UNWRITTEN)
+
+enum __bch_write_flags {
+#define x(f) __BCH_WRITE_##f,
+ BCH_WRITE_FLAGS()
+#undef x
+};
+
+enum bch_write_flags {
+#define x(f) BCH_WRITE_##f = BIT(__BCH_WRITE_##f),
+ BCH_WRITE_FLAGS()
+#undef x
+};
+
+static inline struct workqueue_struct *index_update_wq(struct bch_write_op *op)
+{
+ return op->watermark == BCH_WATERMARK_copygc
+ ? op->c->copygc_wq
+ : op->c->btree_update_wq;
+}
+
+int bch2_sum_sector_overwrites(struct btree_trans *, struct btree_iter *,
+ struct bkey_i *, bool *, s64 *, s64 *);
+int bch2_extent_update(struct btree_trans *, subvol_inum,
+ struct btree_iter *, struct bkey_i *,
+ struct disk_reservation *, u64, s64 *, bool);
+
+static inline void bch2_write_op_init(struct bch_write_op *op, struct bch_fs *c,
+ struct bch_io_opts opts)
+{
+ op->c = c;
+ op->end_io = NULL;
+ op->flags = 0;
+ op->written = 0;
+ op->error = 0;
+ op->csum_type = bch2_data_checksum_type(c, opts);
+ op->compression_opt = opts.compression;
+ op->nr_replicas = 0;
+ op->nr_replicas_required = c->opts.data_replicas_required;
+ op->watermark = BCH_WATERMARK_normal;
+ op->incompressible = 0;
+ op->open_buckets.nr = 0;
+ op->devs_have.nr = 0;
+ op->target = 0;
+ op->opts = opts;
+ op->subvol = 0;
+ op->pos = POS_MAX;
+ op->version = ZERO_VERSION;
+ op->write_point = (struct write_point_specifier) { 0 };
+ op->res = (struct disk_reservation) { 0 };
+ op->new_i_size = U64_MAX;
+ op->i_sectors_delta = 0;
+ op->devs_need_flush = NULL;
+}
+
+void bch2_write(struct closure *);
+
+void bch2_write_point_do_index_updates(struct work_struct *);
+
+static inline struct bch_write_bio *wbio_init(struct bio *bio)
+{
+ struct bch_write_bio *wbio = to_wbio(bio);
+
+ memset(&wbio->wbio, 0, sizeof(wbio->wbio));
+ return wbio;
+}
+
+void bch2_write_op_to_text(struct printbuf *, struct bch_write_op *);
+
+void bch2_fs_io_write_exit(struct bch_fs *);
+int bch2_fs_io_write_init(struct bch_fs *);
+
+#endif /* _BCACHEFS_IO_WRITE_H */
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_IO_WRITE_TYPES_H
+#define _BCACHEFS_IO_WRITE_TYPES_H
+
+#include "alloc_types.h"
+#include "btree_types.h"
+#include "buckets_types.h"
+#include "extents_types.h"
+#include "keylist_types.h"
+#include "opts.h"
+#include "super_types.h"
+
+#include <linux/llist.h>
+#include <linux/workqueue.h>
+
+struct bch_write_bio {
+ struct_group(wbio,
+ struct bch_fs *c;
+ struct bch_write_bio *parent;
+
+ u64 submit_time;
+ u64 inode_offset;
+
+ struct bch_devs_list failed;
+ u8 dev;
+
+ unsigned split:1,
+ bounce:1,
+ put_bio:1,
+ have_ioref:1,
+ nocow:1,
+ used_mempool:1,
+ first_btree_write:1;
+ );
+
+ struct bio bio;
+};
+
+struct bch_write_op {
+ struct closure cl;
+ struct bch_fs *c;
+ void (*end_io)(struct bch_write_op *);
+ u64 start_time;
+
+ unsigned written; /* sectors */
+ u16 flags;
+ s16 error; /* dio write path expects it to hold -ERESTARTSYS... */
+
+ unsigned compression_opt:8;
+ unsigned csum_type:4;
+ unsigned nr_replicas:4;
+ unsigned nr_replicas_required:4;
+ unsigned watermark:3;
+ unsigned incompressible:1;
+ unsigned stripe_waited:1;
+
+ struct bch_devs_list devs_have;
+ u16 target;
+ u16 nonce;
+ struct bch_io_opts opts;
+
+ u32 subvol;
+ struct bpos pos;
+ struct bversion version;
+
+ /* For BCH_WRITE_DATA_ENCODED: */
+ struct bch_extent_crc_unpacked crc;
+
+ struct write_point_specifier write_point;
+
+ struct write_point *wp;
+ struct list_head wp_list;
+
+ struct disk_reservation res;
+
+ struct open_buckets open_buckets;
+
+ u64 new_i_size;
+ s64 i_sectors_delta;
+
+ struct bch_devs_mask failed;
+
+ struct keylist insert_keys;
+ u64 inline_keys[BKEY_EXTENT_U64s_MAX * 2];
+
+ /*
+ * Bitmask of devices that have had nocow writes issued to them since
+ * last flush:
+ */
+ struct bch_devs_mask *devs_need_flush;
+
+ /* Must be last: */
+ struct bch_write_bio wbio;
+};
+
+#endif /* _BCACHEFS_IO_WRITE_TYPES_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "bkey_buf.h"
+#include "btree_update.h"
+#include "error.h"
+#include "io_misc.h"
+#include "logged_ops.h"
+#include "super.h"
+
+struct bch_logged_op_fn {
+ u8 type;
+ int (*resume)(struct btree_trans *, struct bkey_i *);
+};
+
+static const struct bch_logged_op_fn logged_op_fns[] = {
+#define x(n) { \
+ .type = KEY_TYPE_logged_op_##n, \
+ .resume = bch2_resume_logged_op_##n, \
+},
+ BCH_LOGGED_OPS()
+#undef x
+};
+
+static const struct bch_logged_op_fn *logged_op_fn(enum bch_bkey_type type)
+{
+ for (unsigned i = 0; i < ARRAY_SIZE(logged_op_fns); i++)
+ if (logged_op_fns[i].type == type)
+ return logged_op_fns + i;
+ return NULL;
+}
+
+static int resume_logged_op(struct btree_trans *trans, struct btree_iter *iter,
+ struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ const struct bch_logged_op_fn *fn = logged_op_fn(k.k->type);
+ struct bkey_buf sk;
+ u32 restart_count = trans->restart_count;
+ int ret;
+
+ if (!fn)
+ return 0;
+
+ bch2_bkey_buf_init(&sk);
+ bch2_bkey_buf_reassemble(&sk, c, k);
+
+ ret = drop_locks_do(trans, (bch2_fs_lazy_rw(c), 0)) ?:
+ fn->resume(trans, sk.k) ?: trans_was_restarted(trans, restart_count);
+
+ bch2_bkey_buf_exit(&sk, c);
+ return ret;
+}
+
+int bch2_resume_logged_ops(struct bch_fs *c)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret;
+
+ ret = bch2_trans_run(c,
+ for_each_btree_key2(trans, iter,
+ BTREE_ID_logged_ops, POS_MIN, BTREE_ITER_PREFETCH, k,
+ resume_logged_op(trans, &iter, k)));
+ if (ret)
+ bch_err_fn(c, ret);
+ return ret;
+}
+
+static int __bch2_logged_op_start(struct btree_trans *trans, struct bkey_i *k)
+{
+ struct btree_iter iter;
+ int ret;
+
+ ret = bch2_bkey_get_empty_slot(trans, &iter, BTREE_ID_logged_ops, POS_MAX);
+ if (ret)
+ return ret;
+
+ k->k.p = iter.pos;
+
+ ret = bch2_trans_update(trans, &iter, k, 0);
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+int bch2_logged_op_start(struct btree_trans *trans, struct bkey_i *k)
+{
+ return commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ __bch2_logged_op_start(trans, k));
+}
+
+void bch2_logged_op_finish(struct btree_trans *trans, struct bkey_i *k)
+{
+ int ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ bch2_btree_delete(trans, BTREE_ID_logged_ops, k->k.p, 0));
+ /*
+ * This needs to be a fatal error because we've left an unfinished
+ * operation in the logged ops btree.
+ *
+ * We should only ever see an error here if the filesystem has already
+ * been shut down, but make sure of that here:
+ */
+ if (ret) {
+ struct bch_fs *c = trans->c;
+ struct printbuf buf = PRINTBUF;
+
+ bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
+ bch2_fs_fatal_error(c, "%s: error deleting logged operation %s: %s",
+ __func__, buf.buf, bch2_err_str(ret));
+ printbuf_exit(&buf);
+ }
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_LOGGED_OPS_H
+#define _BCACHEFS_LOGGED_OPS_H
+
+#include "bkey.h"
+
+#define BCH_LOGGED_OPS() \
+ x(truncate) \
+ x(finsert)
+
+static inline int bch2_logged_op_update(struct btree_trans *trans, struct bkey_i *op)
+{
+ return bch2_btree_insert_nonextent(trans, BTREE_ID_logged_ops, op, 0);
+}
+
+int bch2_resume_logged_ops(struct bch_fs *);
+int bch2_logged_op_start(struct btree_trans *, struct bkey_i *);
+void bch2_logged_op_finish(struct btree_trans *, struct bkey_i *);
+
+#endif /* _BCACHEFS_LOGGED_OPS_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_update_interior.h"
+#include "buckets.h"
+#include "error.h"
+#include "journal_io.h"
+#include "replicas.h"
+#include "sb-clean.h"
+#include "super-io.h"
+
+/*
+ * BCH_SB_FIELD_clean:
+ *
+ * Btree roots, and a few other things, are recovered from the journal after an
+ * unclean shutdown - but after a clean shutdown, to avoid having to read the
+ * journal, we can store them in the superblock.
+ *
+ * bch_sb_field_clean simply contains a list of journal entries, stored exactly
+ * as they would be in the journal:
+ */
+
+int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *clean,
+ int write)
+{
+ struct jset_entry *entry;
+ int ret;
+
+ for (entry = clean->start;
+ entry < (struct jset_entry *) vstruct_end(&clean->field);
+ entry = vstruct_next(entry)) {
+ ret = bch2_journal_entry_validate(c, NULL, entry,
+ le16_to_cpu(c->disk_sb.sb->version),
+ BCH_SB_BIG_ENDIAN(c->disk_sb.sb),
+ write);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
+static struct bkey_i *btree_root_find(struct bch_fs *c,
+ struct bch_sb_field_clean *clean,
+ struct jset *j,
+ enum btree_id id, unsigned *level)
+{
+ struct bkey_i *k;
+ struct jset_entry *entry, *start, *end;
+
+ if (clean) {
+ start = clean->start;
+ end = vstruct_end(&clean->field);
+ } else {
+ start = j->start;
+ end = vstruct_last(j);
+ }
+
+ for (entry = start; entry < end; entry = vstruct_next(entry))
+ if (entry->type == BCH_JSET_ENTRY_btree_root &&
+ entry->btree_id == id)
+ goto found;
+
+ return NULL;
+found:
+ if (!entry->u64s)
+ return ERR_PTR(-EINVAL);
+
+ k = entry->start;
+ *level = entry->level;
+ return k;
+}
+
+int bch2_verify_superblock_clean(struct bch_fs *c,
+ struct bch_sb_field_clean **cleanp,
+ struct jset *j)
+{
+ unsigned i;
+ struct bch_sb_field_clean *clean = *cleanp;
+ struct printbuf buf1 = PRINTBUF;
+ struct printbuf buf2 = PRINTBUF;
+ int ret = 0;
+
+ if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
+ sb_clean_journal_seq_mismatch,
+ "superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
+ le64_to_cpu(clean->journal_seq),
+ le64_to_cpu(j->seq))) {
+ kfree(clean);
+ *cleanp = NULL;
+ return 0;
+ }
+
+ for (i = 0; i < BTREE_ID_NR; i++) {
+ struct bkey_i *k1, *k2;
+ unsigned l1 = 0, l2 = 0;
+
+ k1 = btree_root_find(c, clean, NULL, i, &l1);
+ k2 = btree_root_find(c, NULL, j, i, &l2);
+
+ if (!k1 && !k2)
+ continue;
+
+ printbuf_reset(&buf1);
+ printbuf_reset(&buf2);
+
+ if (k1)
+ bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(k1));
+ else
+ prt_printf(&buf1, "(none)");
+
+ if (k2)
+ bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(k2));
+ else
+ prt_printf(&buf2, "(none)");
+
+ mustfix_fsck_err_on(!k1 || !k2 ||
+ IS_ERR(k1) ||
+ IS_ERR(k2) ||
+ k1->k.u64s != k2->k.u64s ||
+ memcmp(k1, k2, bkey_bytes(&k1->k)) ||
+ l1 != l2, c,
+ sb_clean_btree_root_mismatch,
+ "superblock btree root %u doesn't match journal after clean shutdown\n"
+ "sb: l=%u %s\n"
+ "journal: l=%u %s\n", i,
+ l1, buf1.buf,
+ l2, buf2.buf);
+ }
+fsck_err:
+ printbuf_exit(&buf2);
+ printbuf_exit(&buf1);
+ return ret;
+}
+
+struct bch_sb_field_clean *bch2_read_superblock_clean(struct bch_fs *c)
+{
+ struct bch_sb_field_clean *clean, *sb_clean;
+ int ret;
+
+ mutex_lock(&c->sb_lock);
+ sb_clean = bch2_sb_field_get(c->disk_sb.sb, clean);
+
+ if (fsck_err_on(!sb_clean, c,
+ sb_clean_missing,
+ "superblock marked clean but clean section not present")) {
+ SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+ c->sb.clean = false;
+ mutex_unlock(&c->sb_lock);
+ return NULL;
+ }
+
+ clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
+ GFP_KERNEL);
+ if (!clean) {
+ mutex_unlock(&c->sb_lock);
+ return ERR_PTR(-BCH_ERR_ENOMEM_read_superblock_clean);
+ }
+
+ ret = bch2_sb_clean_validate_late(c, clean, READ);
+ if (ret) {
+ mutex_unlock(&c->sb_lock);
+ return ERR_PTR(ret);
+ }
+
+ mutex_unlock(&c->sb_lock);
+
+ return clean;
+fsck_err:
+ mutex_unlock(&c->sb_lock);
+ return ERR_PTR(ret);
+}
+
+static struct jset_entry *jset_entry_init(struct jset_entry **end, size_t size)
+{
+ struct jset_entry *entry = *end;
+ unsigned u64s = DIV_ROUND_UP(size, sizeof(u64));
+
+ memset(entry, 0, u64s * sizeof(u64));
+ /*
+ * The u64s field counts from the start of data, ignoring the shared
+ * fields.
+ */
+ entry->u64s = cpu_to_le16(u64s - 1);
+
+ *end = vstruct_next(*end);
+ return entry;
+}
+
+void bch2_journal_super_entries_add_common(struct bch_fs *c,
+ struct jset_entry **end,
+ u64 journal_seq)
+{
+ struct bch_dev *ca;
+ unsigned i, dev;
+
+ percpu_down_read(&c->mark_lock);
+
+ if (!journal_seq) {
+ for (i = 0; i < ARRAY_SIZE(c->usage); i++)
+ bch2_fs_usage_acc_to_base(c, i);
+ } else {
+ bch2_fs_usage_acc_to_base(c, journal_seq & JOURNAL_BUF_MASK);
+ }
+
+ {
+ struct jset_entry_usage *u =
+ container_of(jset_entry_init(end, sizeof(*u)),
+ struct jset_entry_usage, entry);
+
+ u->entry.type = BCH_JSET_ENTRY_usage;
+ u->entry.btree_id = BCH_FS_USAGE_inodes;
+ u->v = cpu_to_le64(c->usage_base->nr_inodes);
+ }
+
+ {
+ struct jset_entry_usage *u =
+ container_of(jset_entry_init(end, sizeof(*u)),
+ struct jset_entry_usage, entry);
+
+ u->entry.type = BCH_JSET_ENTRY_usage;
+ u->entry.btree_id = BCH_FS_USAGE_key_version;
+ u->v = cpu_to_le64(atomic64_read(&c->key_version));
+ }
+
+ for (i = 0; i < BCH_REPLICAS_MAX; i++) {
+ struct jset_entry_usage *u =
+ container_of(jset_entry_init(end, sizeof(*u)),
+ struct jset_entry_usage, entry);
+
+ u->entry.type = BCH_JSET_ENTRY_usage;
+ u->entry.btree_id = BCH_FS_USAGE_reserved;
+ u->entry.level = i;
+ u->v = cpu_to_le64(c->usage_base->persistent_reserved[i]);
+ }
+
+ for (i = 0; i < c->replicas.nr; i++) {
+ struct bch_replicas_entry *e =
+ cpu_replicas_entry(&c->replicas, i);
+ struct jset_entry_data_usage *u =
+ container_of(jset_entry_init(end, sizeof(*u) + e->nr_devs),
+ struct jset_entry_data_usage, entry);
+
+ u->entry.type = BCH_JSET_ENTRY_data_usage;
+ u->v = cpu_to_le64(c->usage_base->replicas[i]);
+ unsafe_memcpy(&u->r, e, replicas_entry_bytes(e),
+ "embedded variable length struct");
+ }
+
+ for_each_member_device(ca, c, dev) {
+ unsigned b = sizeof(struct jset_entry_dev_usage) +
+ sizeof(struct jset_entry_dev_usage_type) * BCH_DATA_NR;
+ struct jset_entry_dev_usage *u =
+ container_of(jset_entry_init(end, b),
+ struct jset_entry_dev_usage, entry);
+
+ u->entry.type = BCH_JSET_ENTRY_dev_usage;
+ u->dev = cpu_to_le32(dev);
+ u->buckets_ec = cpu_to_le64(ca->usage_base->buckets_ec);
+
+ for (i = 0; i < BCH_DATA_NR; i++) {
+ u->d[i].buckets = cpu_to_le64(ca->usage_base->d[i].buckets);
+ u->d[i].sectors = cpu_to_le64(ca->usage_base->d[i].sectors);
+ u->d[i].fragmented = cpu_to_le64(ca->usage_base->d[i].fragmented);
+ }
+ }
+
+ percpu_up_read(&c->mark_lock);
+
+ for (i = 0; i < 2; i++) {
+ struct jset_entry_clock *clock =
+ container_of(jset_entry_init(end, sizeof(*clock)),
+ struct jset_entry_clock, entry);
+
+ clock->entry.type = BCH_JSET_ENTRY_clock;
+ clock->rw = i;
+ clock->time = cpu_to_le64(atomic64_read(&c->io_clock[i].now));
+ }
+}
+
+static int bch2_sb_clean_validate(struct bch_sb *sb,
+ struct bch_sb_field *f,
+ struct printbuf *err)
+{
+ struct bch_sb_field_clean *clean = field_to_type(f, clean);
+
+ if (vstruct_bytes(&clean->field) < sizeof(*clean)) {
+ prt_printf(err, "wrong size (got %zu should be %zu)",
+ vstruct_bytes(&clean->field), sizeof(*clean));
+ return -BCH_ERR_invalid_sb_clean;
+ }
+
+ return 0;
+}
+
+static void bch2_sb_clean_to_text(struct printbuf *out, struct bch_sb *sb,
+ struct bch_sb_field *f)
+{
+ struct bch_sb_field_clean *clean = field_to_type(f, clean);
+ struct jset_entry *entry;
+
+ prt_printf(out, "flags: %x", le32_to_cpu(clean->flags));
+ prt_newline(out);
+ prt_printf(out, "journal_seq: %llu", le64_to_cpu(clean->journal_seq));
+ prt_newline(out);
+
+ for (entry = clean->start;
+ entry != vstruct_end(&clean->field);
+ entry = vstruct_next(entry)) {
+ if (entry->type == BCH_JSET_ENTRY_btree_keys &&
+ !entry->u64s)
+ continue;
+
+ bch2_journal_entry_to_text(out, NULL, entry);
+ prt_newline(out);
+ }
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_clean = {
+ .validate = bch2_sb_clean_validate,
+ .to_text = bch2_sb_clean_to_text,
+};
+
+int bch2_fs_mark_dirty(struct bch_fs *c)
+{
+ int ret;
+
+ /*
+ * Unconditionally write superblock, to verify it hasn't changed before
+ * we go rw:
+ */
+
+ mutex_lock(&c->sb_lock);
+ SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+
+ bch2_sb_maybe_downgrade(c);
+ c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALWAYS);
+
+ ret = bch2_write_super(c);
+ mutex_unlock(&c->sb_lock);
+
+ return ret;
+}
+
+void bch2_fs_mark_clean(struct bch_fs *c)
+{
+ struct bch_sb_field_clean *sb_clean;
+ struct jset_entry *entry;
+ unsigned u64s;
+ int ret;
+
+ mutex_lock(&c->sb_lock);
+ if (BCH_SB_CLEAN(c->disk_sb.sb))
+ goto out;
+
+ SET_BCH_SB_CLEAN(c->disk_sb.sb, true);
+
+ c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_info);
+ c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_metadata);
+ c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_extents_above_btree_updates));
+ c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_btree_updates_journalled));
+
+ u64s = sizeof(*sb_clean) / sizeof(u64) + c->journal.entry_u64s_reserved;
+
+ sb_clean = bch2_sb_field_resize(&c->disk_sb, clean, u64s);
+ if (!sb_clean) {
+ bch_err(c, "error resizing superblock while setting filesystem clean");
+ goto out;
+ }
+
+ sb_clean->flags = 0;
+ sb_clean->journal_seq = cpu_to_le64(atomic64_read(&c->journal.seq));
+
+ /* Trying to catch outstanding bug: */
+ BUG_ON(le64_to_cpu(sb_clean->journal_seq) > S64_MAX);
+
+ entry = sb_clean->start;
+ bch2_journal_super_entries_add_common(c, &entry, 0);
+ entry = bch2_btree_roots_to_journal_entries(c, entry, 0);
+ BUG_ON((void *) entry > vstruct_end(&sb_clean->field));
+
+ memset(entry, 0,
+ vstruct_end(&sb_clean->field) - (void *) entry);
+
+ /*
+ * this should be in the write path, and we should be validating every
+ * superblock section:
+ */
+ ret = bch2_sb_clean_validate_late(c, sb_clean, WRITE);
+ if (ret) {
+ bch_err(c, "error writing marking filesystem clean: validate error");
+ goto out;
+ }
+
+ bch2_write_super(c);
+out:
+ mutex_unlock(&c->sb_lock);
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SB_CLEAN_H
+#define _BCACHEFS_SB_CLEAN_H
+
+int bch2_sb_clean_validate_late(struct bch_fs *, struct bch_sb_field_clean *, int);
+int bch2_verify_superblock_clean(struct bch_fs *, struct bch_sb_field_clean **,
+ struct jset *);
+struct bch_sb_field_clean *bch2_read_superblock_clean(struct bch_fs *);
+void bch2_journal_super_entries_add_common(struct bch_fs *, struct jset_entry **, u64);
+
+extern const struct bch_sb_field_ops bch_sb_field_ops_clean;
+
+int bch2_fs_mark_dirty(struct bch_fs *);
+void bch2_fs_mark_clean(struct bch_fs *);
+
+#endif /* _BCACHEFS_SB_CLEAN_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "sb-errors.h"
+#include "super-io.h"
+
+static const char * const bch2_sb_error_strs[] = {
+#define x(t, n, ...) [n] = #t,
+ BCH_SB_ERRS()
+ NULL
+};
+
+static void bch2_sb_error_id_to_text(struct printbuf *out, enum bch_sb_error_id id)
+{
+ if (id < BCH_SB_ERR_MAX)
+ prt_str(out, bch2_sb_error_strs[id]);
+ else
+ prt_printf(out, "(unknown error %u)", id);
+}
+
+static inline unsigned bch2_sb_field_errors_nr_entries(struct bch_sb_field_errors *e)
+{
+ return e
+ ? (bch2_sb_field_bytes(&e->field) - sizeof(*e)) / sizeof(e->entries[0])
+ : 0;
+}
+
+static inline unsigned bch2_sb_field_errors_u64s(unsigned nr)
+{
+ return (sizeof(struct bch_sb_field_errors) +
+ sizeof(struct bch_sb_field_error_entry) * nr) / sizeof(u64);
+}
+
+static int bch2_sb_errors_validate(struct bch_sb *sb, struct bch_sb_field *f,
+ struct printbuf *err)
+{
+ struct bch_sb_field_errors *e = field_to_type(f, errors);
+ unsigned i, nr = bch2_sb_field_errors_nr_entries(e);
+
+ for (i = 0; i < nr; i++) {
+ if (!BCH_SB_ERROR_ENTRY_NR(&e->entries[i])) {
+ prt_printf(err, "entry with count 0 (id ");
+ bch2_sb_error_id_to_text(err, BCH_SB_ERROR_ENTRY_ID(&e->entries[i]));
+ prt_printf(err, ")");
+ return -BCH_ERR_invalid_sb_errors;
+ }
+
+ if (i + 1 < nr &&
+ BCH_SB_ERROR_ENTRY_ID(&e->entries[i]) >=
+ BCH_SB_ERROR_ENTRY_ID(&e->entries[i + 1])) {
+ prt_printf(err, "entries out of order");
+ return -BCH_ERR_invalid_sb_errors;
+ }
+ }
+
+ return 0;
+}
+
+static void bch2_sb_errors_to_text(struct printbuf *out, struct bch_sb *sb,
+ struct bch_sb_field *f)
+{
+ struct bch_sb_field_errors *e = field_to_type(f, errors);
+ unsigned i, nr = bch2_sb_field_errors_nr_entries(e);
+
+ if (out->nr_tabstops <= 1)
+ printbuf_tabstop_push(out, 16);
+
+ for (i = 0; i < nr; i++) {
+ bch2_sb_error_id_to_text(out, BCH_SB_ERROR_ENTRY_ID(&e->entries[i]));
+ prt_tab(out);
+ prt_u64(out, BCH_SB_ERROR_ENTRY_NR(&e->entries[i]));
+ prt_tab(out);
+ bch2_prt_datetime(out, le64_to_cpu(e->entries[i].last_error_time));
+ prt_newline(out);
+ }
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_errors = {
+ .validate = bch2_sb_errors_validate,
+ .to_text = bch2_sb_errors_to_text,
+};
+
+void bch2_sb_error_count(struct bch_fs *c, enum bch_sb_error_id err)
+{
+ bch_sb_errors_cpu *e = &c->fsck_error_counts;
+ struct bch_sb_error_entry_cpu n = {
+ .id = err,
+ .nr = 1,
+ .last_error_time = ktime_get_real_seconds()
+ };
+ unsigned i;
+
+ mutex_lock(&c->fsck_error_counts_lock);
+ for (i = 0; i < e->nr; i++) {
+ if (err == e->data[i].id) {
+ e->data[i].nr++;
+ e->data[i].last_error_time = n.last_error_time;
+ goto out;
+ }
+ if (err < e->data[i].id)
+ break;
+ }
+
+ if (darray_make_room(e, 1))
+ goto out;
+
+ darray_insert_item(e, i, n);
+out:
+ mutex_unlock(&c->fsck_error_counts_lock);
+}
+
+void bch2_sb_errors_from_cpu(struct bch_fs *c)
+{
+ bch_sb_errors_cpu *src = &c->fsck_error_counts;
+ struct bch_sb_field_errors *dst =
+ bch2_sb_field_resize(&c->disk_sb, errors,
+ bch2_sb_field_errors_u64s(src->nr));
+ unsigned i;
+
+ if (!dst)
+ return;
+
+ for (i = 0; i < src->nr; i++) {
+ SET_BCH_SB_ERROR_ENTRY_ID(&dst->entries[i], src->data[i].id);
+ SET_BCH_SB_ERROR_ENTRY_NR(&dst->entries[i], src->data[i].nr);
+ dst->entries[i].last_error_time = cpu_to_le64(src->data[i].last_error_time);
+ }
+}
+
+static int bch2_sb_errors_to_cpu(struct bch_fs *c)
+{
+ struct bch_sb_field_errors *src = bch2_sb_field_get(c->disk_sb.sb, errors);
+ bch_sb_errors_cpu *dst = &c->fsck_error_counts;
+ unsigned i, nr = bch2_sb_field_errors_nr_entries(src);
+ int ret;
+
+ if (!nr)
+ return 0;
+
+ mutex_lock(&c->fsck_error_counts_lock);
+ ret = darray_make_room(dst, nr);
+ if (ret)
+ goto err;
+
+ dst->nr = nr;
+
+ for (i = 0; i < nr; i++) {
+ dst->data[i].id = BCH_SB_ERROR_ENTRY_ID(&src->entries[i]);
+ dst->data[i].nr = BCH_SB_ERROR_ENTRY_NR(&src->entries[i]);
+ dst->data[i].last_error_time = le64_to_cpu(src->entries[i].last_error_time);
+ }
+err:
+ mutex_unlock(&c->fsck_error_counts_lock);
+
+ return ret;
+}
+
+void bch2_fs_sb_errors_exit(struct bch_fs *c)
+{
+ darray_exit(&c->fsck_error_counts);
+}
+
+void bch2_fs_sb_errors_init_early(struct bch_fs *c)
+{
+ mutex_init(&c->fsck_error_counts_lock);
+ darray_init(&c->fsck_error_counts);
+}
+
+int bch2_fs_sb_errors_init(struct bch_fs *c)
+{
+ return bch2_sb_errors_to_cpu(c);
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SB_ERRORS_H
+#define _BCACHEFS_SB_ERRORS_H
+
+#include "sb-errors_types.h"
+
+#define BCH_SB_ERRS() \
+ x(clean_but_journal_not_empty, 0) \
+ x(dirty_but_no_journal_entries, 1) \
+ x(dirty_but_no_journal_entries_post_drop_nonflushes, 2) \
+ x(sb_clean_journal_seq_mismatch, 3) \
+ x(sb_clean_btree_root_mismatch, 4) \
+ x(sb_clean_missing, 5) \
+ x(jset_unsupported_version, 6) \
+ x(jset_unknown_csum, 7) \
+ x(jset_last_seq_newer_than_seq, 8) \
+ x(jset_past_bucket_end, 9) \
+ x(jset_seq_blacklisted, 10) \
+ x(journal_entries_missing, 11) \
+ x(journal_entry_replicas_not_marked, 12) \
+ x(journal_entry_past_jset_end, 13) \
+ x(journal_entry_replicas_data_mismatch, 14) \
+ x(journal_entry_bkey_u64s_0, 15) \
+ x(journal_entry_bkey_past_end, 16) \
+ x(journal_entry_bkey_bad_format, 17) \
+ x(journal_entry_bkey_invalid, 18) \
+ x(journal_entry_btree_root_bad_size, 19) \
+ x(journal_entry_blacklist_bad_size, 20) \
+ x(journal_entry_blacklist_v2_bad_size, 21) \
+ x(journal_entry_blacklist_v2_start_past_end, 22) \
+ x(journal_entry_usage_bad_size, 23) \
+ x(journal_entry_data_usage_bad_size, 24) \
+ x(journal_entry_clock_bad_size, 25) \
+ x(journal_entry_clock_bad_rw, 26) \
+ x(journal_entry_dev_usage_bad_size, 27) \
+ x(journal_entry_dev_usage_bad_dev, 28) \
+ x(journal_entry_dev_usage_bad_pad, 29) \
+ x(btree_node_unreadable, 30) \
+ x(btree_node_fault_injected, 31) \
+ x(btree_node_bad_magic, 32) \
+ x(btree_node_bad_seq, 33) \
+ x(btree_node_unsupported_version, 34) \
+ x(btree_node_bset_older_than_sb_min, 35) \
+ x(btree_node_bset_newer_than_sb, 36) \
+ x(btree_node_data_missing, 37) \
+ x(btree_node_bset_after_end, 38) \
+ x(btree_node_replicas_sectors_written_mismatch, 39) \
+ x(btree_node_replicas_data_mismatch, 40) \
+ x(bset_unknown_csum, 41) \
+ x(bset_bad_csum, 42) \
+ x(bset_past_end_of_btree_node, 43) \
+ x(bset_wrong_sector_offset, 44) \
+ x(bset_empty, 45) \
+ x(bset_bad_seq, 46) \
+ x(bset_blacklisted_journal_seq, 47) \
+ x(first_bset_blacklisted_journal_seq, 48) \
+ x(btree_node_bad_btree, 49) \
+ x(btree_node_bad_level, 50) \
+ x(btree_node_bad_min_key, 51) \
+ x(btree_node_bad_max_key, 52) \
+ x(btree_node_bad_format, 53) \
+ x(btree_node_bkey_past_bset_end, 54) \
+ x(btree_node_bkey_bad_format, 55) \
+ x(btree_node_bad_bkey, 56) \
+ x(btree_node_bkey_out_of_order, 57) \
+ x(btree_root_bkey_invalid, 58) \
+ x(btree_root_read_error, 59) \
+ x(btree_root_bad_min_key, 50) \
+ x(btree_root_bad_max_key, 61) \
+ x(btree_node_read_error, 62) \
+ x(btree_node_topology_bad_min_key, 63) \
+ x(btree_node_topology_bad_max_key, 64) \
+ x(btree_node_topology_overwritten_by_prev_node, 65) \
+ x(btree_node_topology_overwritten_by_next_node, 66) \
+ x(btree_node_topology_interior_node_empty, 67) \
+ x(fs_usage_hidden_wrong, 68) \
+ x(fs_usage_btree_wrong, 69) \
+ x(fs_usage_data_wrong, 70) \
+ x(fs_usage_cached_wrong, 71) \
+ x(fs_usage_reserved_wrong, 72) \
+ x(fs_usage_persistent_reserved_wrong, 73) \
+ x(fs_usage_nr_inodes_wrong, 74) \
+ x(fs_usage_replicas_wrong, 75) \
+ x(dev_usage_buckets_wrong, 76) \
+ x(dev_usage_sectors_wrong, 77) \
+ x(dev_usage_fragmented_wrong, 78) \
+ x(dev_usage_buckets_ec_wrong, 79) \
+ x(bkey_version_in_future, 80) \
+ x(bkey_u64s_too_small, 81) \
+ x(bkey_invalid_type_for_btree, 82) \
+ x(bkey_extent_size_zero, 83) \
+ x(bkey_extent_size_greater_than_offset, 84) \
+ x(bkey_size_nonzero, 85) \
+ x(bkey_snapshot_nonzero, 86) \
+ x(bkey_snapshot_zero, 87) \
+ x(bkey_at_pos_max, 88) \
+ x(bkey_before_start_of_btree_node, 89) \
+ x(bkey_after_end_of_btree_node, 90) \
+ x(bkey_val_size_nonzero, 91) \
+ x(bkey_val_size_too_small, 92) \
+ x(alloc_v1_val_size_bad, 93) \
+ x(alloc_v2_unpack_error, 94) \
+ x(alloc_v3_unpack_error, 95) \
+ x(alloc_v4_val_size_bad, 96) \
+ x(alloc_v4_backpointers_start_bad, 97) \
+ x(alloc_key_data_type_bad, 98) \
+ x(alloc_key_empty_but_have_data, 99) \
+ x(alloc_key_dirty_sectors_0, 100) \
+ x(alloc_key_data_type_inconsistency, 101) \
+ x(alloc_key_to_missing_dev_bucket, 102) \
+ x(alloc_key_cached_inconsistency, 103) \
+ x(alloc_key_cached_but_read_time_zero, 104) \
+ x(alloc_key_to_missing_lru_entry, 105) \
+ x(alloc_key_data_type_wrong, 106) \
+ x(alloc_key_gen_wrong, 107) \
+ x(alloc_key_dirty_sectors_wrong, 108) \
+ x(alloc_key_cached_sectors_wrong, 109) \
+ x(alloc_key_stripe_wrong, 110) \
+ x(alloc_key_stripe_redundancy_wrong, 111) \
+ x(bucket_sector_count_overflow, 112) \
+ x(bucket_metadata_type_mismatch, 113) \
+ x(need_discard_key_wrong, 114) \
+ x(freespace_key_wrong, 115) \
+ x(freespace_hole_missing, 116) \
+ x(bucket_gens_val_size_bad, 117) \
+ x(bucket_gens_key_wrong, 118) \
+ x(bucket_gens_hole_wrong, 119) \
+ x(bucket_gens_to_invalid_dev, 120) \
+ x(bucket_gens_to_invalid_buckets, 121) \
+ x(bucket_gens_nonzero_for_invalid_buckets, 122) \
+ x(need_discard_freespace_key_to_invalid_dev_bucket, 123) \
+ x(need_discard_freespace_key_bad, 124) \
+ x(backpointer_pos_wrong, 125) \
+ x(backpointer_to_missing_device, 126) \
+ x(backpointer_to_missing_alloc, 127) \
+ x(backpointer_to_missing_ptr, 128) \
+ x(lru_entry_at_time_0, 129) \
+ x(lru_entry_to_invalid_bucket, 130) \
+ x(lru_entry_bad, 131) \
+ x(btree_ptr_val_too_big, 132) \
+ x(btree_ptr_v2_val_too_big, 133) \
+ x(btree_ptr_has_non_ptr, 134) \
+ x(extent_ptrs_invalid_entry, 135) \
+ x(extent_ptrs_no_ptrs, 136) \
+ x(extent_ptrs_too_many_ptrs, 137) \
+ x(extent_ptrs_redundant_crc, 138) \
+ x(extent_ptrs_redundant_stripe, 139) \
+ x(extent_ptrs_unwritten, 140) \
+ x(extent_ptrs_written_and_unwritten, 141) \
+ x(ptr_to_invalid_device, 142) \
+ x(ptr_to_duplicate_device, 143) \
+ x(ptr_after_last_bucket, 144) \
+ x(ptr_before_first_bucket, 145) \
+ x(ptr_spans_multiple_buckets, 146) \
+ x(ptr_to_missing_backpointer, 147) \
+ x(ptr_to_missing_alloc_key, 148) \
+ x(ptr_to_missing_replicas_entry, 149) \
+ x(ptr_to_missing_stripe, 150) \
+ x(ptr_to_incorrect_stripe, 151) \
+ x(ptr_gen_newer_than_bucket_gen, 152) \
+ x(ptr_too_stale, 153) \
+ x(stale_dirty_ptr, 154) \
+ x(ptr_bucket_data_type_mismatch, 155) \
+ x(ptr_cached_and_erasure_coded, 156) \
+ x(ptr_crc_uncompressed_size_too_small, 157) \
+ x(ptr_crc_csum_type_unknown, 158) \
+ x(ptr_crc_compression_type_unknown, 159) \
+ x(ptr_crc_redundant, 160) \
+ x(ptr_crc_uncompressed_size_too_big, 161) \
+ x(ptr_crc_nonce_mismatch, 162) \
+ x(ptr_stripe_redundant, 163) \
+ x(reservation_key_nr_replicas_invalid, 164) \
+ x(reflink_v_refcount_wrong, 165) \
+ x(reflink_p_to_missing_reflink_v, 166) \
+ x(stripe_pos_bad, 167) \
+ x(stripe_val_size_bad, 168) \
+ x(stripe_sector_count_wrong, 169) \
+ x(snapshot_tree_pos_bad, 170) \
+ x(snapshot_tree_to_missing_snapshot, 171) \
+ x(snapshot_tree_to_missing_subvol, 172) \
+ x(snapshot_tree_to_wrong_subvol, 173) \
+ x(snapshot_tree_to_snapshot_subvol, 174) \
+ x(snapshot_pos_bad, 175) \
+ x(snapshot_parent_bad, 176) \
+ x(snapshot_children_not_normalized, 177) \
+ x(snapshot_child_duplicate, 178) \
+ x(snapshot_child_bad, 179) \
+ x(snapshot_skiplist_not_normalized, 180) \
+ x(snapshot_skiplist_bad, 181) \
+ x(snapshot_should_not_have_subvol, 182) \
+ x(snapshot_to_bad_snapshot_tree, 183) \
+ x(snapshot_bad_depth, 184) \
+ x(snapshot_bad_skiplist, 185) \
+ x(subvol_pos_bad, 186) \
+ x(subvol_not_master_and_not_snapshot, 187) \
+ x(subvol_to_missing_root, 188) \
+ x(subvol_root_wrong_bi_subvol, 189) \
+ x(bkey_in_missing_snapshot, 190) \
+ x(inode_pos_inode_nonzero, 191) \
+ x(inode_pos_blockdev_range, 192) \
+ x(inode_unpack_error, 193) \
+ x(inode_str_hash_invalid, 194) \
+ x(inode_v3_fields_start_bad, 195) \
+ x(inode_snapshot_mismatch, 196) \
+ x(inode_unlinked_but_clean, 197) \
+ x(inode_unlinked_but_nlink_nonzero, 198) \
+ x(inode_checksum_type_invalid, 199) \
+ x(inode_compression_type_invalid, 200) \
+ x(inode_subvol_root_but_not_dir, 201) \
+ x(inode_i_size_dirty_but_clean, 202) \
+ x(inode_i_sectors_dirty_but_clean, 203) \
+ x(inode_i_sectors_wrong, 204) \
+ x(inode_dir_wrong_nlink, 205) \
+ x(inode_dir_multiple_links, 206) \
+ x(inode_multiple_links_but_nlink_0, 207) \
+ x(inode_wrong_backpointer, 208) \
+ x(inode_wrong_nlink, 209) \
+ x(inode_unreachable, 210) \
+ x(deleted_inode_but_clean, 211) \
+ x(deleted_inode_missing, 212) \
+ x(deleted_inode_is_dir, 213) \
+ x(deleted_inode_not_unlinked, 214) \
+ x(extent_overlapping, 215) \
+ x(extent_in_missing_inode, 216) \
+ x(extent_in_non_reg_inode, 217) \
+ x(extent_past_end_of_inode, 218) \
+ x(dirent_empty_name, 219) \
+ x(dirent_val_too_big, 220) \
+ x(dirent_name_too_long, 221) \
+ x(dirent_name_embedded_nul, 222) \
+ x(dirent_name_dot_or_dotdot, 223) \
+ x(dirent_name_has_slash, 224) \
+ x(dirent_d_type_wrong, 225) \
+ x(dirent_d_parent_subvol_wrong, 226) \
+ x(dirent_in_missing_dir_inode, 227) \
+ x(dirent_in_non_dir_inode, 228) \
+ x(dirent_to_missing_inode, 229) \
+ x(dirent_to_missing_subvol, 230) \
+ x(dirent_to_itself, 231) \
+ x(quota_type_invalid, 232) \
+ x(xattr_val_size_too_small, 233) \
+ x(xattr_val_size_too_big, 234) \
+ x(xattr_invalid_type, 235) \
+ x(xattr_name_invalid_chars, 236) \
+ x(xattr_in_missing_inode, 237) \
+ x(root_subvol_missing, 238) \
+ x(root_dir_missing, 239) \
+ x(root_inode_not_dir, 240) \
+ x(dir_loop, 241) \
+ x(hash_table_key_duplicate, 242) \
+ x(hash_table_key_wrong_offset, 243)
+
+enum bch_sb_error_id {
+#define x(t, n) BCH_FSCK_ERR_##t = n,
+ BCH_SB_ERRS()
+#undef x
+ BCH_SB_ERR_MAX
+};
+
+extern const struct bch_sb_field_ops bch_sb_field_ops_errors;
+
+void bch2_sb_error_count(struct bch_fs *, enum bch_sb_error_id);
+
+void bch2_sb_errors_from_cpu(struct bch_fs *);
+
+void bch2_fs_sb_errors_exit(struct bch_fs *);
+void bch2_fs_sb_errors_init_early(struct bch_fs *);
+int bch2_fs_sb_errors_init(struct bch_fs *);
+
+#endif /* _BCACHEFS_SB_ERRORS_H */
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SB_ERRORS_TYPES_H
+#define _BCACHEFS_SB_ERRORS_TYPES_H
+
+#include "darray.h"
+
+struct bch_sb_error_entry_cpu {
+ u64 id:16,
+ nr:48;
+ u64 last_error_time;
+};
+
+typedef DARRAY(struct bch_sb_error_entry_cpu) bch_sb_errors_cpu;
+
+#endif /* _BCACHEFS_SB_ERRORS_TYPES_H */
+
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "disk_groups.h"
+#include "opts.h"
+#include "replicas.h"
+#include "sb-members.h"
+#include "super-io.h"
+
+#define x(t, n, ...) [n] = #t,
+static const char * const bch2_iops_measurements[] = {
+ BCH_IOPS_MEASUREMENTS()
+ NULL
+};
+
+char * const bch2_member_error_strs[] = {
+ BCH_MEMBER_ERROR_TYPES()
+ NULL
+};
+#undef x
+
+/* Code for bch_sb_field_members_v1: */
+
+struct bch_member *bch2_members_v2_get_mut(struct bch_sb *sb, int i)
+{
+ return __bch2_members_v2_get_mut(bch2_sb_field_get(sb, members_v2), i);
+}
+
+static struct bch_member members_v2_get(struct bch_sb_field_members_v2 *mi, int i)
+{
+ struct bch_member ret, *p = __bch2_members_v2_get_mut(mi, i);
+ memset(&ret, 0, sizeof(ret));
+ memcpy(&ret, p, min_t(size_t, le16_to_cpu(mi->member_bytes), sizeof(ret)));
+ return ret;
+}
+
+static struct bch_member *members_v1_get_mut(struct bch_sb_field_members_v1 *mi, int i)
+{
+ return (void *) mi->_members + (i * BCH_MEMBER_V1_BYTES);
+}
+
+static struct bch_member members_v1_get(struct bch_sb_field_members_v1 *mi, int i)
+{
+ struct bch_member ret, *p = members_v1_get_mut(mi, i);
+ memset(&ret, 0, sizeof(ret));
+ memcpy(&ret, p, min_t(size_t, BCH_MEMBER_V1_BYTES, sizeof(ret)));
+ return ret;
+}
+
+struct bch_member bch2_sb_member_get(struct bch_sb *sb, int i)
+{
+ struct bch_sb_field_members_v2 *mi2 = bch2_sb_field_get(sb, members_v2);
+ if (mi2)
+ return members_v2_get(mi2, i);
+ struct bch_sb_field_members_v1 *mi1 = bch2_sb_field_get(sb, members_v1);
+ return members_v1_get(mi1, i);
+}
+
+static int sb_members_v2_resize_entries(struct bch_fs *c)
+{
+ struct bch_sb_field_members_v2 *mi = bch2_sb_field_get(c->disk_sb.sb, members_v2);
+
+ if (le16_to_cpu(mi->member_bytes) < sizeof(struct bch_member)) {
+ unsigned u64s = DIV_ROUND_UP((sizeof(*mi) + sizeof(mi->_members[0]) *
+ c->disk_sb.sb->nr_devices), 8);
+
+ mi = bch2_sb_field_resize(&c->disk_sb, members_v2, u64s);
+ if (!mi)
+ return -BCH_ERR_ENOSPC_sb_members_v2;
+
+ for (int i = c->disk_sb.sb->nr_devices - 1; i >= 0; --i) {
+ void *dst = (void *) mi->_members + (i * sizeof(struct bch_member));
+ memmove(dst, __bch2_members_v2_get_mut(mi, i), le16_to_cpu(mi->member_bytes));
+ memset(dst + le16_to_cpu(mi->member_bytes),
+ 0, (sizeof(struct bch_member) - le16_to_cpu(mi->member_bytes)));
+ }
+ mi->member_bytes = cpu_to_le16(sizeof(struct bch_member));
+ }
+ return 0;
+}
+
+int bch2_sb_members_v2_init(struct bch_fs *c)
+{
+ struct bch_sb_field_members_v1 *mi1;
+ struct bch_sb_field_members_v2 *mi2;
+
+ if (!bch2_sb_field_get(c->disk_sb.sb, members_v2)) {
+ mi2 = bch2_sb_field_resize(&c->disk_sb, members_v2,
+ DIV_ROUND_UP(sizeof(*mi2) +
+ sizeof(struct bch_member) * c->sb.nr_devices,
+ sizeof(u64)));
+ mi1 = bch2_sb_field_get(c->disk_sb.sb, members_v1);
+ memcpy(&mi2->_members[0], &mi1->_members[0],
+ BCH_MEMBER_V1_BYTES * c->sb.nr_devices);
+ memset(&mi2->pad[0], 0, sizeof(mi2->pad));
+ mi2->member_bytes = cpu_to_le16(BCH_MEMBER_V1_BYTES);
+ }
+
+ return sb_members_v2_resize_entries(c);
+}
+
+int bch2_sb_members_cpy_v2_v1(struct bch_sb_handle *disk_sb)
+{
+ struct bch_sb_field_members_v1 *mi1;
+ struct bch_sb_field_members_v2 *mi2;
+
+ mi1 = bch2_sb_field_resize(disk_sb, members_v1,
+ DIV_ROUND_UP(sizeof(*mi1) + BCH_MEMBER_V1_BYTES *
+ disk_sb->sb->nr_devices, sizeof(u64)));
+ if (!mi1)
+ return -BCH_ERR_ENOSPC_sb_members;
+
+ mi2 = bch2_sb_field_get(disk_sb->sb, members_v2);
+
+ for (unsigned i = 0; i < disk_sb->sb->nr_devices; i++)
+ memcpy(members_v1_get_mut(mi1, i), __bch2_members_v2_get_mut(mi2, i), BCH_MEMBER_V1_BYTES);
+
+ return 0;
+}
+
+static int validate_member(struct printbuf *err,
+ struct bch_member m,
+ struct bch_sb *sb,
+ int i)
+{
+ if (le64_to_cpu(m.nbuckets) > LONG_MAX) {
+ prt_printf(err, "device %u: too many buckets (got %llu, max %lu)",
+ i, le64_to_cpu(m.nbuckets), LONG_MAX);
+ return -BCH_ERR_invalid_sb_members;
+ }
+
+ if (le64_to_cpu(m.nbuckets) -
+ le16_to_cpu(m.first_bucket) < BCH_MIN_NR_NBUCKETS) {
+ prt_printf(err, "device %u: not enough buckets (got %llu, max %u)",
+ i, le64_to_cpu(m.nbuckets), BCH_MIN_NR_NBUCKETS);
+ return -BCH_ERR_invalid_sb_members;
+ }
+
+ if (le16_to_cpu(m.bucket_size) <
+ le16_to_cpu(sb->block_size)) {
+ prt_printf(err, "device %u: bucket size %u smaller than block size %u",
+ i, le16_to_cpu(m.bucket_size), le16_to_cpu(sb->block_size));
+ return -BCH_ERR_invalid_sb_members;
+ }
+
+ if (le16_to_cpu(m.bucket_size) <
+ BCH_SB_BTREE_NODE_SIZE(sb)) {
+ prt_printf(err, "device %u: bucket size %u smaller than btree node size %llu",
+ i, le16_to_cpu(m.bucket_size), BCH_SB_BTREE_NODE_SIZE(sb));
+ return -BCH_ERR_invalid_sb_members;
+ }
+
+ return 0;
+}
+
+static void member_to_text(struct printbuf *out,
+ struct bch_member m,
+ struct bch_sb_field_disk_groups *gi,
+ struct bch_sb *sb,
+ int i)
+{
+ unsigned data_have = bch2_sb_dev_has_data(sb, i);
+ u64 bucket_size = le16_to_cpu(m.bucket_size);
+ u64 device_size = le64_to_cpu(m.nbuckets) * bucket_size;
+
+ if (!bch2_member_exists(&m))
+ return;
+
+ prt_printf(out, "Device:");
+ prt_tab(out);
+ prt_printf(out, "%u", i);
+ prt_newline(out);
+
+ printbuf_indent_add(out, 2);
+
+ prt_printf(out, "Label:");
+ prt_tab(out);
+ if (BCH_MEMBER_GROUP(&m)) {
+ unsigned idx = BCH_MEMBER_GROUP(&m) - 1;
+
+ if (idx < disk_groups_nr(gi))
+ prt_printf(out, "%s (%u)",
+ gi->entries[idx].label, idx);
+ else
+ prt_printf(out, "(bad disk labels section)");
+ } else {
+ prt_printf(out, "(none)");
+ }
+ prt_newline(out);
+
+ prt_printf(out, "UUID:");
+ prt_tab(out);
+ pr_uuid(out, m.uuid.b);
+ prt_newline(out);
+
+ prt_printf(out, "Size:");
+ prt_tab(out);
+ prt_units_u64(out, device_size << 9);
+ prt_newline(out);
+
+ for (unsigned i = 0; i < BCH_MEMBER_ERROR_NR; i++) {
+ prt_printf(out, "%s errors:", bch2_member_error_strs[i]);
+ prt_tab(out);
+ prt_u64(out, le64_to_cpu(m.errors[i]));
+ prt_newline(out);
+ }
+
+ for (unsigned i = 0; i < BCH_IOPS_NR; i++) {
+ prt_printf(out, "%s iops:", bch2_iops_measurements[i]);
+ prt_tab(out);
+ prt_printf(out, "%u", le32_to_cpu(m.iops[i]));
+ prt_newline(out);
+ }
+
+ prt_printf(out, "Bucket size:");
+ prt_tab(out);
+ prt_units_u64(out, bucket_size << 9);
+ prt_newline(out);
+
+ prt_printf(out, "First bucket:");
+ prt_tab(out);
+ prt_printf(out, "%u", le16_to_cpu(m.first_bucket));
+ prt_newline(out);
+
+ prt_printf(out, "Buckets:");
+ prt_tab(out);
+ prt_printf(out, "%llu", le64_to_cpu(m.nbuckets));
+ prt_newline(out);
+
+ prt_printf(out, "Last mount:");
+ prt_tab(out);
+ if (m.last_mount)
+ bch2_prt_datetime(out, le64_to_cpu(m.last_mount));
+ else
+ prt_printf(out, "(never)");
+ prt_newline(out);
+
+ prt_printf(out, "State:");
+ prt_tab(out);
+ prt_printf(out, "%s",
+ BCH_MEMBER_STATE(&m) < BCH_MEMBER_STATE_NR
+ ? bch2_member_states[BCH_MEMBER_STATE(&m)]
+ : "unknown");
+ prt_newline(out);
+
+ prt_printf(out, "Data allowed:");
+ prt_tab(out);
+ if (BCH_MEMBER_DATA_ALLOWED(&m))
+ prt_bitflags(out, bch2_data_types, BCH_MEMBER_DATA_ALLOWED(&m));
+ else
+ prt_printf(out, "(none)");
+ prt_newline(out);
+
+ prt_printf(out, "Has data:");
+ prt_tab(out);
+ if (data_have)
+ prt_bitflags(out, bch2_data_types, data_have);
+ else
+ prt_printf(out, "(none)");
+ prt_newline(out);
+
+ prt_printf(out, "Discard:");
+ prt_tab(out);
+ prt_printf(out, "%llu", BCH_MEMBER_DISCARD(&m));
+ prt_newline(out);
+
+ prt_printf(out, "Freespace initialized:");
+ prt_tab(out);
+ prt_printf(out, "%llu", BCH_MEMBER_FREESPACE_INITIALIZED(&m));
+ prt_newline(out);
+
+ printbuf_indent_sub(out, 2);
+}
+
+static int bch2_sb_members_v1_validate(struct bch_sb *sb,
+ struct bch_sb_field *f,
+ struct printbuf *err)
+{
+ struct bch_sb_field_members_v1 *mi = field_to_type(f, members_v1);
+ unsigned i;
+
+ if ((void *) members_v1_get_mut(mi, sb->nr_devices) > vstruct_end(&mi->field)) {
+ prt_printf(err, "too many devices for section size");
+ return -BCH_ERR_invalid_sb_members;
+ }
+
+ for (i = 0; i < sb->nr_devices; i++) {
+ struct bch_member m = members_v1_get(mi, i);
+
+ int ret = validate_member(err, m, sb, i);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
+static void bch2_sb_members_v1_to_text(struct printbuf *out, struct bch_sb *sb,
+ struct bch_sb_field *f)
+{
+ struct bch_sb_field_members_v1 *mi = field_to_type(f, members_v1);
+ struct bch_sb_field_disk_groups *gi = bch2_sb_field_get(sb, disk_groups);
+ unsigned i;
+
+ for (i = 0; i < sb->nr_devices; i++)
+ member_to_text(out, members_v1_get(mi, i), gi, sb, i);
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_members_v1 = {
+ .validate = bch2_sb_members_v1_validate,
+ .to_text = bch2_sb_members_v1_to_text,
+};
+
+static void bch2_sb_members_v2_to_text(struct printbuf *out, struct bch_sb *sb,
+ struct bch_sb_field *f)
+{
+ struct bch_sb_field_members_v2 *mi = field_to_type(f, members_v2);
+ struct bch_sb_field_disk_groups *gi = bch2_sb_field_get(sb, disk_groups);
+ unsigned i;
+
+ for (i = 0; i < sb->nr_devices; i++)
+ member_to_text(out, members_v2_get(mi, i), gi, sb, i);
+}
+
+static int bch2_sb_members_v2_validate(struct bch_sb *sb,
+ struct bch_sb_field *f,
+ struct printbuf *err)
+{
+ struct bch_sb_field_members_v2 *mi = field_to_type(f, members_v2);
+ size_t mi_bytes = (void *) __bch2_members_v2_get_mut(mi, sb->nr_devices) -
+ (void *) mi;
+
+ if (mi_bytes > vstruct_bytes(&mi->field)) {
+ prt_printf(err, "section too small (%zu > %zu)",
+ mi_bytes, vstruct_bytes(&mi->field));
+ return -BCH_ERR_invalid_sb_members;
+ }
+
+ for (unsigned i = 0; i < sb->nr_devices; i++) {
+ int ret = validate_member(err, members_v2_get(mi, i), sb, i);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_members_v2 = {
+ .validate = bch2_sb_members_v2_validate,
+ .to_text = bch2_sb_members_v2_to_text,
+};
+
+void bch2_sb_members_from_cpu(struct bch_fs *c)
+{
+ struct bch_sb_field_members_v2 *mi = bch2_sb_field_get(c->disk_sb.sb, members_v2);
+ struct bch_dev *ca;
+ unsigned i, e;
+
+ rcu_read_lock();
+ for_each_member_device_rcu(ca, c, i, NULL) {
+ struct bch_member *m = __bch2_members_v2_get_mut(mi, i);
+
+ for (e = 0; e < BCH_MEMBER_ERROR_NR; e++)
+ m->errors[e] = cpu_to_le64(atomic64_read(&ca->errors[e]));
+ }
+ rcu_read_unlock();
+}
+
+void bch2_dev_io_errors_to_text(struct printbuf *out, struct bch_dev *ca)
+{
+ struct bch_fs *c = ca->fs;
+ struct bch_member m;
+
+ mutex_lock(&ca->fs->sb_lock);
+ m = bch2_sb_member_get(c->disk_sb.sb, ca->dev_idx);
+ mutex_unlock(&ca->fs->sb_lock);
+
+ printbuf_tabstop_push(out, 12);
+
+ prt_str(out, "IO errors since filesystem creation");
+ prt_newline(out);
+
+ printbuf_indent_add(out, 2);
+ for (unsigned i = 0; i < BCH_MEMBER_ERROR_NR; i++) {
+ prt_printf(out, "%s:", bch2_member_error_strs[i]);
+ prt_tab(out);
+ prt_u64(out, atomic64_read(&ca->errors[i]));
+ prt_newline(out);
+ }
+ printbuf_indent_sub(out, 2);
+
+ prt_str(out, "IO errors since ");
+ bch2_pr_time_units(out, (ktime_get_real_seconds() - le64_to_cpu(m.errors_reset_time)) * NSEC_PER_SEC);
+ prt_str(out, " ago");
+ prt_newline(out);
+
+ printbuf_indent_add(out, 2);
+ for (unsigned i = 0; i < BCH_MEMBER_ERROR_NR; i++) {
+ prt_printf(out, "%s:", bch2_member_error_strs[i]);
+ prt_tab(out);
+ prt_u64(out, atomic64_read(&ca->errors[i]) - le64_to_cpu(m.errors_at_reset[i]));
+ prt_newline(out);
+ }
+ printbuf_indent_sub(out, 2);
+}
+
+void bch2_dev_errors_reset(struct bch_dev *ca)
+{
+ struct bch_fs *c = ca->fs;
+ struct bch_member *m;
+
+ mutex_lock(&c->sb_lock);
+ m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
+ for (unsigned i = 0; i < ARRAY_SIZE(m->errors_at_reset); i++)
+ m->errors_at_reset[i] = cpu_to_le64(atomic64_read(&ca->errors[i]));
+ m->errors_reset_time = ktime_get_real_seconds();
+
+ bch2_write_super(c);
+ mutex_unlock(&c->sb_lock);
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SB_MEMBERS_H
+#define _BCACHEFS_SB_MEMBERS_H
+
+extern char * const bch2_member_error_strs[];
+
+static inline struct bch_member *
+__bch2_members_v2_get_mut(struct bch_sb_field_members_v2 *mi, unsigned i)
+{
+ return (void *) mi->_members + (i * le16_to_cpu(mi->member_bytes));
+}
+
+int bch2_sb_members_v2_init(struct bch_fs *c);
+int bch2_sb_members_cpy_v2_v1(struct bch_sb_handle *disk_sb);
+struct bch_member *bch2_members_v2_get_mut(struct bch_sb *sb, int i);
+struct bch_member bch2_sb_member_get(struct bch_sb *sb, int i);
+
+static inline bool bch2_dev_is_online(struct bch_dev *ca)
+{
+ return !percpu_ref_is_zero(&ca->io_ref);
+}
+
+static inline bool bch2_dev_is_readable(struct bch_dev *ca)
+{
+ return bch2_dev_is_online(ca) &&
+ ca->mi.state != BCH_MEMBER_STATE_failed;
+}
+
+static inline bool bch2_dev_get_ioref(struct bch_dev *ca, int rw)
+{
+ if (!percpu_ref_tryget(&ca->io_ref))
+ return false;
+
+ if (ca->mi.state == BCH_MEMBER_STATE_rw ||
+ (ca->mi.state == BCH_MEMBER_STATE_ro && rw == READ))
+ return true;
+
+ percpu_ref_put(&ca->io_ref);
+ return false;
+}
+
+static inline unsigned dev_mask_nr(const struct bch_devs_mask *devs)
+{
+ return bitmap_weight(devs->d, BCH_SB_MEMBERS_MAX);
+}
+
+static inline bool bch2_dev_list_has_dev(struct bch_devs_list devs,
+ unsigned dev)
+{
+ unsigned i;
+
+ for (i = 0; i < devs.nr; i++)
+ if (devs.devs[i] == dev)
+ return true;
+
+ return false;
+}
+
+static inline void bch2_dev_list_drop_dev(struct bch_devs_list *devs,
+ unsigned dev)
+{
+ unsigned i;
+
+ for (i = 0; i < devs->nr; i++)
+ if (devs->devs[i] == dev) {
+ array_remove_item(devs->devs, devs->nr, i);
+ return;
+ }
+}
+
+static inline void bch2_dev_list_add_dev(struct bch_devs_list *devs,
+ unsigned dev)
+{
+ if (!bch2_dev_list_has_dev(*devs, dev)) {
+ BUG_ON(devs->nr >= ARRAY_SIZE(devs->devs));
+ devs->devs[devs->nr++] = dev;
+ }
+}
+
+static inline struct bch_devs_list bch2_dev_list_single(unsigned dev)
+{
+ return (struct bch_devs_list) { .nr = 1, .devs[0] = dev };
+}
+
+static inline struct bch_dev *__bch2_next_dev(struct bch_fs *c, unsigned *iter,
+ const struct bch_devs_mask *mask)
+{
+ struct bch_dev *ca = NULL;
+
+ while ((*iter = mask
+ ? find_next_bit(mask->d, c->sb.nr_devices, *iter)
+ : *iter) < c->sb.nr_devices &&
+ !(ca = rcu_dereference_check(c->devs[*iter],
+ lockdep_is_held(&c->state_lock))))
+ (*iter)++;
+
+ return ca;
+}
+
+#define for_each_member_device_rcu(ca, c, iter, mask) \
+ for ((iter) = 0; ((ca) = __bch2_next_dev((c), &(iter), mask)); (iter)++)
+
+static inline struct bch_dev *bch2_get_next_dev(struct bch_fs *c, unsigned *iter)
+{
+ struct bch_dev *ca;
+
+ rcu_read_lock();
+ if ((ca = __bch2_next_dev(c, iter, NULL)))
+ percpu_ref_get(&ca->ref);
+ rcu_read_unlock();
+
+ return ca;
+}
+
+/*
+ * If you break early, you must drop your ref on the current device
+ */
+#define for_each_member_device(ca, c, iter) \
+ for ((iter) = 0; \
+ (ca = bch2_get_next_dev(c, &(iter))); \
+ percpu_ref_put(&ca->ref), (iter)++)
+
+static inline struct bch_dev *bch2_get_next_online_dev(struct bch_fs *c,
+ unsigned *iter,
+ int state_mask)
+{
+ struct bch_dev *ca;
+
+ rcu_read_lock();
+ while ((ca = __bch2_next_dev(c, iter, NULL)) &&
+ (!((1 << ca->mi.state) & state_mask) ||
+ !percpu_ref_tryget(&ca->io_ref)))
+ (*iter)++;
+ rcu_read_unlock();
+
+ return ca;
+}
+
+#define __for_each_online_member(ca, c, iter, state_mask) \
+ for ((iter) = 0; \
+ (ca = bch2_get_next_online_dev(c, &(iter), state_mask)); \
+ percpu_ref_put(&ca->io_ref), (iter)++)
+
+#define for_each_online_member(ca, c, iter) \
+ __for_each_online_member(ca, c, iter, ~0)
+
+#define for_each_rw_member(ca, c, iter) \
+ __for_each_online_member(ca, c, iter, 1 << BCH_MEMBER_STATE_rw)
+
+#define for_each_readable_member(ca, c, iter) \
+ __for_each_online_member(ca, c, iter, \
+ (1 << BCH_MEMBER_STATE_rw)|(1 << BCH_MEMBER_STATE_ro))
+
+/*
+ * If a key exists that references a device, the device won't be going away and
+ * we can omit rcu_read_lock():
+ */
+static inline struct bch_dev *bch_dev_bkey_exists(const struct bch_fs *c, unsigned idx)
+{
+ EBUG_ON(idx >= c->sb.nr_devices || !c->devs[idx]);
+
+ return rcu_dereference_check(c->devs[idx], 1);
+}
+
+static inline struct bch_dev *bch_dev_locked(struct bch_fs *c, unsigned idx)
+{
+ EBUG_ON(idx >= c->sb.nr_devices || !c->devs[idx]);
+
+ return rcu_dereference_protected(c->devs[idx],
+ lockdep_is_held(&c->sb_lock) ||
+ lockdep_is_held(&c->state_lock));
+}
+
+/* XXX kill, move to struct bch_fs */
+static inline struct bch_devs_mask bch2_online_devs(struct bch_fs *c)
+{
+ struct bch_devs_mask devs;
+ struct bch_dev *ca;
+ unsigned i;
+
+ memset(&devs, 0, sizeof(devs));
+ for_each_online_member(ca, c, i)
+ __set_bit(ca->dev_idx, devs.d);
+ return devs;
+}
+
+extern const struct bch_sb_field_ops bch_sb_field_ops_members_v1;
+extern const struct bch_sb_field_ops bch_sb_field_ops_members_v2;
+
+static inline bool bch2_member_exists(struct bch_member *m)
+{
+ return !bch2_is_zero(&m->uuid, sizeof(m->uuid));
+}
+
+static inline bool bch2_dev_exists(struct bch_sb *sb, unsigned dev)
+{
+ if (dev < sb->nr_devices) {
+ struct bch_member m = bch2_sb_member_get(sb, dev);
+ return bch2_member_exists(&m);
+ }
+ return false;
+}
+
+static inline struct bch_member_cpu bch2_mi_to_cpu(struct bch_member *mi)
+{
+ return (struct bch_member_cpu) {
+ .nbuckets = le64_to_cpu(mi->nbuckets),
+ .first_bucket = le16_to_cpu(mi->first_bucket),
+ .bucket_size = le16_to_cpu(mi->bucket_size),
+ .group = BCH_MEMBER_GROUP(mi),
+ .state = BCH_MEMBER_STATE(mi),
+ .discard = BCH_MEMBER_DISCARD(mi),
+ .data_allowed = BCH_MEMBER_DATA_ALLOWED(mi),
+ .durability = BCH_MEMBER_DURABILITY(mi)
+ ? BCH_MEMBER_DURABILITY(mi) - 1
+ : 1,
+ .freespace_initialized = BCH_MEMBER_FREESPACE_INITIALIZED(mi),
+ .valid = bch2_member_exists(mi),
+ };
+}
+
+void bch2_sb_members_from_cpu(struct bch_fs *);
+
+void bch2_dev_io_errors_to_text(struct printbuf *, struct bch_dev *);
+void bch2_dev_errors_reset(struct bch_dev *);
+
+#endif /* _BCACHEFS_SB_MEMBERS_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/export.h>
+#include <linux/log2.h>
+#include <linux/percpu.h>
+#include <linux/preempt.h>
+#include <linux/rcupdate.h>
+#include <linux/sched.h>
+#include <linux/sched/clock.h>
+#include <linux/sched/rt.h>
+#include <linux/sched/task.h>
+#include <linux/slab.h>
+
+#include <trace/events/lock.h>
+
+#include "six.h"
+
+#ifdef DEBUG
+#define EBUG_ON(cond) BUG_ON(cond)
+#else
+#define EBUG_ON(cond) do {} while (0)
+#endif
+
+#define six_acquire(l, t, r, ip) lock_acquire(l, 0, t, r, 1, NULL, ip)
+#define six_release(l, ip) lock_release(l, ip)
+
+static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type);
+
+#define SIX_LOCK_HELD_read_OFFSET 0
+#define SIX_LOCK_HELD_read ~(~0U << 26)
+#define SIX_LOCK_HELD_intent (1U << 26)
+#define SIX_LOCK_HELD_write (1U << 27)
+#define SIX_LOCK_WAITING_read (1U << (28 + SIX_LOCK_read))
+#define SIX_LOCK_WAITING_write (1U << (28 + SIX_LOCK_write))
+#define SIX_LOCK_NOSPIN (1U << 31)
+
+struct six_lock_vals {
+ /* Value we add to the lock in order to take the lock: */
+ u32 lock_val;
+
+ /* If the lock has this value (used as a mask), taking the lock fails: */
+ u32 lock_fail;
+
+ /* Mask that indicates lock is held for this type: */
+ u32 held_mask;
+
+ /* Waitlist we wakeup when releasing the lock: */
+ enum six_lock_type unlock_wakeup;
+};
+
+static const struct six_lock_vals l[] = {
+ [SIX_LOCK_read] = {
+ .lock_val = 1U << SIX_LOCK_HELD_read_OFFSET,
+ .lock_fail = SIX_LOCK_HELD_write,
+ .held_mask = SIX_LOCK_HELD_read,
+ .unlock_wakeup = SIX_LOCK_write,
+ },
+ [SIX_LOCK_intent] = {
+ .lock_val = SIX_LOCK_HELD_intent,
+ .lock_fail = SIX_LOCK_HELD_intent,
+ .held_mask = SIX_LOCK_HELD_intent,
+ .unlock_wakeup = SIX_LOCK_intent,
+ },
+ [SIX_LOCK_write] = {
+ .lock_val = SIX_LOCK_HELD_write,
+ .lock_fail = SIX_LOCK_HELD_read,
+ .held_mask = SIX_LOCK_HELD_write,
+ .unlock_wakeup = SIX_LOCK_read,
+ },
+};
+
+static inline void six_set_bitmask(struct six_lock *lock, u32 mask)
+{
+ if ((atomic_read(&lock->state) & mask) != mask)
+ atomic_or(mask, &lock->state);
+}
+
+static inline void six_clear_bitmask(struct six_lock *lock, u32 mask)
+{
+ if (atomic_read(&lock->state) & mask)
+ atomic_and(~mask, &lock->state);
+}
+
+static inline void six_set_owner(struct six_lock *lock, enum six_lock_type type,
+ u32 old, struct task_struct *owner)
+{
+ if (type != SIX_LOCK_intent)
+ return;
+
+ if (!(old & SIX_LOCK_HELD_intent)) {
+ EBUG_ON(lock->owner);
+ lock->owner = owner;
+ } else {
+ EBUG_ON(lock->owner != current);
+ }
+}
+
+static inline unsigned pcpu_read_count(struct six_lock *lock)
+{
+ unsigned read_count = 0;
+ int cpu;
+
+ for_each_possible_cpu(cpu)
+ read_count += *per_cpu_ptr(lock->readers, cpu);
+ return read_count;
+}
+
+/*
+ * __do_six_trylock() - main trylock routine
+ *
+ * Returns 1 on success, 0 on failure
+ *
+ * In percpu reader mode, a failed trylock may cause a spurious trylock failure
+ * for anoter thread taking the competing lock type, and we may havve to do a
+ * wakeup: when a wakeup is required, we return -1 - wakeup_type.
+ */
+static int __do_six_trylock(struct six_lock *lock, enum six_lock_type type,
+ struct task_struct *task, bool try)
+{
+ int ret;
+ u32 old;
+
+ EBUG_ON(type == SIX_LOCK_write && lock->owner != task);
+ EBUG_ON(type == SIX_LOCK_write &&
+ (try != !(atomic_read(&lock->state) & SIX_LOCK_HELD_write)));
+
+ /*
+ * Percpu reader mode:
+ *
+ * The basic idea behind this algorithm is that you can implement a lock
+ * between two threads without any atomics, just memory barriers:
+ *
+ * For two threads you'll need two variables, one variable for "thread a
+ * has the lock" and another for "thread b has the lock".
+ *
+ * To take the lock, a thread sets its variable indicating that it holds
+ * the lock, then issues a full memory barrier, then reads from the
+ * other thread's variable to check if the other thread thinks it has
+ * the lock. If we raced, we backoff and retry/sleep.
+ *
+ * Failure to take the lock may cause a spurious trylock failure in
+ * another thread, because we temporarily set the lock to indicate that
+ * we held it. This would be a problem for a thread in six_lock(), when
+ * they are calling trylock after adding themself to the waitlist and
+ * prior to sleeping.
+ *
+ * Therefore, if we fail to get the lock, and there were waiters of the
+ * type we conflict with, we will have to issue a wakeup.
+ *
+ * Since we may be called under wait_lock (and by the wakeup code
+ * itself), we return that the wakeup has to be done instead of doing it
+ * here.
+ */
+ if (type == SIX_LOCK_read && lock->readers) {
+ preempt_disable();
+ this_cpu_inc(*lock->readers); /* signal that we own lock */
+
+ smp_mb();
+
+ old = atomic_read(&lock->state);
+ ret = !(old & l[type].lock_fail);
+
+ this_cpu_sub(*lock->readers, !ret);
+ preempt_enable();
+
+ if (!ret) {
+ smp_mb();
+ if (atomic_read(&lock->state) & SIX_LOCK_WAITING_write)
+ ret = -1 - SIX_LOCK_write;
+ }
+ } else if (type == SIX_LOCK_write && lock->readers) {
+ if (try) {
+ atomic_add(SIX_LOCK_HELD_write, &lock->state);
+ smp_mb__after_atomic();
+ }
+
+ ret = !pcpu_read_count(lock);
+
+ if (try && !ret) {
+ old = atomic_sub_return(SIX_LOCK_HELD_write, &lock->state);
+ if (old & SIX_LOCK_WAITING_read)
+ ret = -1 - SIX_LOCK_read;
+ }
+ } else {
+ old = atomic_read(&lock->state);
+ do {
+ ret = !(old & l[type].lock_fail);
+ if (!ret || (type == SIX_LOCK_write && !try)) {
+ smp_mb();
+ break;
+ }
+ } while (!atomic_try_cmpxchg_acquire(&lock->state, &old, old + l[type].lock_val));
+
+ EBUG_ON(ret && !(atomic_read(&lock->state) & l[type].held_mask));
+ }
+
+ if (ret > 0)
+ six_set_owner(lock, type, old, task);
+
+ EBUG_ON(type == SIX_LOCK_write && try && ret <= 0 &&
+ (atomic_read(&lock->state) & SIX_LOCK_HELD_write));
+
+ return ret;
+}
+
+static void __six_lock_wakeup(struct six_lock *lock, enum six_lock_type lock_type)
+{
+ struct six_lock_waiter *w, *next;
+ struct task_struct *task;
+ bool saw_one;
+ int ret;
+again:
+ ret = 0;
+ saw_one = false;
+ raw_spin_lock(&lock->wait_lock);
+
+ list_for_each_entry_safe(w, next, &lock->wait_list, list) {
+ if (w->lock_want != lock_type)
+ continue;
+
+ if (saw_one && lock_type != SIX_LOCK_read)
+ goto unlock;
+ saw_one = true;
+
+ ret = __do_six_trylock(lock, lock_type, w->task, false);
+ if (ret <= 0)
+ goto unlock;
+
+ /*
+ * Similar to percpu_rwsem_wake_function(), we need to guard
+ * against the wakee noticing w->lock_acquired, returning, and
+ * then exiting before we do the wakeup:
+ */
+ task = get_task_struct(w->task);
+ __list_del(w->list.prev, w->list.next);
+ /*
+ * The release barrier here ensures the ordering of the
+ * __list_del before setting w->lock_acquired; @w is on the
+ * stack of the thread doing the waiting and will be reused
+ * after it sees w->lock_acquired with no other locking:
+ * pairs with smp_load_acquire() in six_lock_slowpath()
+ */
+ smp_store_release(&w->lock_acquired, true);
+ wake_up_process(task);
+ put_task_struct(task);
+ }
+
+ six_clear_bitmask(lock, SIX_LOCK_WAITING_read << lock_type);
+unlock:
+ raw_spin_unlock(&lock->wait_lock);
+
+ if (ret < 0) {
+ lock_type = -ret - 1;
+ goto again;
+ }
+}
+
+__always_inline
+static void six_lock_wakeup(struct six_lock *lock, u32 state,
+ enum six_lock_type lock_type)
+{
+ if (lock_type == SIX_LOCK_write && (state & SIX_LOCK_HELD_read))
+ return;
+
+ if (!(state & (SIX_LOCK_WAITING_read << lock_type)))
+ return;
+
+ __six_lock_wakeup(lock, lock_type);
+}
+
+__always_inline
+static bool do_six_trylock(struct six_lock *lock, enum six_lock_type type, bool try)
+{
+ int ret;
+
+ ret = __do_six_trylock(lock, type, current, try);
+ if (ret < 0)
+ __six_lock_wakeup(lock, -ret - 1);
+
+ return ret > 0;
+}
+
+/**
+ * six_trylock_ip - attempt to take a six lock without blocking
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @ip: ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * Return: true on success, false on failure.
+ */
+bool six_trylock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip)
+{
+ if (!do_six_trylock(lock, type, true))
+ return false;
+
+ if (type != SIX_LOCK_write)
+ six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
+ return true;
+}
+EXPORT_SYMBOL_GPL(six_trylock_ip);
+
+/**
+ * six_relock_ip - attempt to re-take a lock that was held previously
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @seq: lock sequence number obtained from six_lock_seq() while lock was
+ * held previously
+ * @ip: ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * Return: true on success, false on failure.
+ */
+bool six_relock_ip(struct six_lock *lock, enum six_lock_type type,
+ unsigned seq, unsigned long ip)
+{
+ if (six_lock_seq(lock) != seq || !six_trylock_ip(lock, type, ip))
+ return false;
+
+ if (six_lock_seq(lock) != seq) {
+ six_unlock_ip(lock, type, ip);
+ return false;
+ }
+
+ return true;
+}
+EXPORT_SYMBOL_GPL(six_relock_ip);
+
+#ifdef CONFIG_LOCK_SPIN_ON_OWNER
+
+static inline bool six_owner_running(struct six_lock *lock)
+{
+ /*
+ * When there's no owner, we might have preempted between the owner
+ * acquiring the lock and setting the owner field. If we're an RT task
+ * that will live-lock because we won't let the owner complete.
+ */
+ rcu_read_lock();
+ struct task_struct *owner = READ_ONCE(lock->owner);
+ bool ret = owner ? owner_on_cpu(owner) : !rt_task(current);
+ rcu_read_unlock();
+
+ return ret;
+}
+
+static inline bool six_optimistic_spin(struct six_lock *lock,
+ struct six_lock_waiter *wait,
+ enum six_lock_type type)
+{
+ unsigned loop = 0;
+ u64 end_time;
+
+ if (type == SIX_LOCK_write)
+ return false;
+
+ if (lock->wait_list.next != &wait->list)
+ return false;
+
+ if (atomic_read(&lock->state) & SIX_LOCK_NOSPIN)
+ return false;
+
+ preempt_disable();
+ end_time = sched_clock() + 10 * NSEC_PER_USEC;
+
+ while (!need_resched() && six_owner_running(lock)) {
+ /*
+ * Ensures that writes to the waitlist entry happen after we see
+ * wait->lock_acquired: pairs with the smp_store_release in
+ * __six_lock_wakeup
+ */
+ if (smp_load_acquire(&wait->lock_acquired)) {
+ preempt_enable();
+ return true;
+ }
+
+ if (!(++loop & 0xf) && (time_after64(sched_clock(), end_time))) {
+ six_set_bitmask(lock, SIX_LOCK_NOSPIN);
+ break;
+ }
+
+ /*
+ * The cpu_relax() call is a compiler barrier which forces
+ * everything in this loop to be re-loaded. We don't need
+ * memory barriers as we'll eventually observe the right
+ * values at the cost of a few extra spins.
+ */
+ cpu_relax();
+ }
+
+ preempt_enable();
+ return false;
+}
+
+#else /* CONFIG_LOCK_SPIN_ON_OWNER */
+
+static inline bool six_optimistic_spin(struct six_lock *lock,
+ struct six_lock_waiter *wait,
+ enum six_lock_type type)
+{
+ return false;
+}
+
+#endif
+
+noinline
+static int six_lock_slowpath(struct six_lock *lock, enum six_lock_type type,
+ struct six_lock_waiter *wait,
+ six_lock_should_sleep_fn should_sleep_fn, void *p,
+ unsigned long ip)
+{
+ int ret = 0;
+
+ if (type == SIX_LOCK_write) {
+ EBUG_ON(atomic_read(&lock->state) & SIX_LOCK_HELD_write);
+ atomic_add(SIX_LOCK_HELD_write, &lock->state);
+ smp_mb__after_atomic();
+ }
+
+ trace_contention_begin(lock, 0);
+ lock_contended(&lock->dep_map, ip);
+
+ wait->task = current;
+ wait->lock_want = type;
+ wait->lock_acquired = false;
+
+ raw_spin_lock(&lock->wait_lock);
+ six_set_bitmask(lock, SIX_LOCK_WAITING_read << type);
+ /*
+ * Retry taking the lock after taking waitlist lock, in case we raced
+ * with an unlock:
+ */
+ ret = __do_six_trylock(lock, type, current, false);
+ if (ret <= 0) {
+ wait->start_time = local_clock();
+
+ if (!list_empty(&lock->wait_list)) {
+ struct six_lock_waiter *last =
+ list_last_entry(&lock->wait_list,
+ struct six_lock_waiter, list);
+
+ if (time_before_eq64(wait->start_time, last->start_time))
+ wait->start_time = last->start_time + 1;
+ }
+
+ list_add_tail(&wait->list, &lock->wait_list);
+ }
+ raw_spin_unlock(&lock->wait_lock);
+
+ if (unlikely(ret > 0)) {
+ ret = 0;
+ goto out;
+ }
+
+ if (unlikely(ret < 0)) {
+ __six_lock_wakeup(lock, -ret - 1);
+ ret = 0;
+ }
+
+ if (six_optimistic_spin(lock, wait, type))
+ goto out;
+
+ while (1) {
+ set_current_state(TASK_UNINTERRUPTIBLE);
+
+ /*
+ * Ensures that writes to the waitlist entry happen after we see
+ * wait->lock_acquired: pairs with the smp_store_release in
+ * __six_lock_wakeup
+ */
+ if (smp_load_acquire(&wait->lock_acquired))
+ break;
+
+ ret = should_sleep_fn ? should_sleep_fn(lock, p) : 0;
+ if (unlikely(ret)) {
+ bool acquired;
+
+ /*
+ * If should_sleep_fn() returns an error, we are
+ * required to return that error even if we already
+ * acquired the lock - should_sleep_fn() might have
+ * modified external state (e.g. when the deadlock cycle
+ * detector in bcachefs issued a transaction restart)
+ */
+ raw_spin_lock(&lock->wait_lock);
+ acquired = wait->lock_acquired;
+ if (!acquired)
+ list_del(&wait->list);
+ raw_spin_unlock(&lock->wait_lock);
+
+ if (unlikely(acquired))
+ do_six_unlock_type(lock, type);
+ break;
+ }
+
+ schedule();
+ }
+
+ __set_current_state(TASK_RUNNING);
+out:
+ if (ret && type == SIX_LOCK_write) {
+ six_clear_bitmask(lock, SIX_LOCK_HELD_write);
+ six_lock_wakeup(lock, atomic_read(&lock->state), SIX_LOCK_read);
+ }
+ trace_contention_end(lock, 0);
+
+ return ret;
+}
+
+/**
+ * six_lock_ip_waiter - take a lock, with full waitlist interface
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @wait: pointer to wait object, which will be added to lock's waitlist
+ * @should_sleep_fn: callback run after adding to waitlist, immediately prior
+ * to scheduling
+ * @p: passed through to @should_sleep_fn
+ * @ip: ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * This is the most general six_lock() variant, with parameters to support full
+ * cycle detection for deadlock avoidance.
+ *
+ * The code calling this function must implement tracking of held locks, and the
+ * @wait object should be embedded into the struct that tracks held locks -
+ * which must also be accessible in a thread-safe way.
+ *
+ * @should_sleep_fn should invoke the cycle detector; it should walk each
+ * lock's waiters, and for each waiter recursively walk their held locks.
+ *
+ * When this function must block, @wait will be added to @lock's waitlist before
+ * calling trylock, and before calling @should_sleep_fn, and @wait will not be
+ * removed from the lock waitlist until the lock has been successfully acquired,
+ * or we abort.
+ *
+ * @wait.start_time will be monotonically increasing for any given waitlist, and
+ * thus may be used as a loop cursor.
+ *
+ * Return: 0 on success, or the return code from @should_sleep_fn on failure.
+ */
+int six_lock_ip_waiter(struct six_lock *lock, enum six_lock_type type,
+ struct six_lock_waiter *wait,
+ six_lock_should_sleep_fn should_sleep_fn, void *p,
+ unsigned long ip)
+{
+ int ret;
+
+ wait->start_time = 0;
+
+ if (type != SIX_LOCK_write)
+ six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, ip);
+
+ ret = do_six_trylock(lock, type, true) ? 0
+ : six_lock_slowpath(lock, type, wait, should_sleep_fn, p, ip);
+
+ if (ret && type != SIX_LOCK_write)
+ six_release(&lock->dep_map, ip);
+ if (!ret)
+ lock_acquired(&lock->dep_map, ip);
+
+ return ret;
+}
+EXPORT_SYMBOL_GPL(six_lock_ip_waiter);
+
+__always_inline
+static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type)
+{
+ u32 state;
+
+ if (type == SIX_LOCK_intent)
+ lock->owner = NULL;
+
+ if (type == SIX_LOCK_read &&
+ lock->readers) {
+ smp_mb(); /* unlock barrier */
+ this_cpu_dec(*lock->readers);
+ smp_mb(); /* between unlocking and checking for waiters */
+ state = atomic_read(&lock->state);
+ } else {
+ u32 v = l[type].lock_val;
+
+ if (type != SIX_LOCK_read)
+ v += atomic_read(&lock->state) & SIX_LOCK_NOSPIN;
+
+ EBUG_ON(!(atomic_read(&lock->state) & l[type].held_mask));
+ state = atomic_sub_return_release(v, &lock->state);
+ }
+
+ six_lock_wakeup(lock, state, l[type].unlock_wakeup);
+}
+
+/**
+ * six_unlock_ip - drop a six lock
+ * @lock: lock to unlock
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @ip: ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * When a lock is held multiple times (because six_lock_incement()) was used),
+ * this decrements the 'lock held' counter by one.
+ *
+ * For example:
+ * six_lock_read(&foo->lock); read count 1
+ * six_lock_increment(&foo->lock, SIX_LOCK_read); read count 2
+ * six_lock_unlock(&foo->lock, SIX_LOCK_read); read count 1
+ * six_lock_unlock(&foo->lock, SIX_LOCK_read); read count 0
+ */
+void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip)
+{
+ EBUG_ON(type == SIX_LOCK_write &&
+ !(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
+ EBUG_ON((type == SIX_LOCK_write ||
+ type == SIX_LOCK_intent) &&
+ lock->owner != current);
+
+ if (type != SIX_LOCK_write)
+ six_release(&lock->dep_map, ip);
+ else
+ lock->seq++;
+
+ if (type == SIX_LOCK_intent &&
+ lock->intent_lock_recurse) {
+ --lock->intent_lock_recurse;
+ return;
+ }
+
+ do_six_unlock_type(lock, type);
+}
+EXPORT_SYMBOL_GPL(six_unlock_ip);
+
+/**
+ * six_lock_downgrade - convert an intent lock to a read lock
+ * @lock: lock to dowgrade
+ *
+ * @lock will have read count incremented and intent count decremented
+ */
+void six_lock_downgrade(struct six_lock *lock)
+{
+ six_lock_increment(lock, SIX_LOCK_read);
+ six_unlock_intent(lock);
+}
+EXPORT_SYMBOL_GPL(six_lock_downgrade);
+
+/**
+ * six_lock_tryupgrade - attempt to convert read lock to an intent lock
+ * @lock: lock to upgrade
+ *
+ * On success, @lock will have intent count incremented and read count
+ * decremented
+ *
+ * Return: true on success, false on failure
+ */
+bool six_lock_tryupgrade(struct six_lock *lock)
+{
+ u32 old = atomic_read(&lock->state), new;
+
+ do {
+ new = old;
+
+ if (new & SIX_LOCK_HELD_intent)
+ return false;
+
+ if (!lock->readers) {
+ EBUG_ON(!(new & SIX_LOCK_HELD_read));
+ new -= l[SIX_LOCK_read].lock_val;
+ }
+
+ new |= SIX_LOCK_HELD_intent;
+ } while (!atomic_try_cmpxchg_acquire(&lock->state, &old, new));
+
+ if (lock->readers)
+ this_cpu_dec(*lock->readers);
+
+ six_set_owner(lock, SIX_LOCK_intent, old, current);
+
+ return true;
+}
+EXPORT_SYMBOL_GPL(six_lock_tryupgrade);
+
+/**
+ * six_trylock_convert - attempt to convert a held lock from one type to another
+ * @lock: lock to upgrade
+ * @from: SIX_LOCK_read or SIX_LOCK_intent
+ * @to: SIX_LOCK_read or SIX_LOCK_intent
+ *
+ * On success, @lock will have intent count incremented and read count
+ * decremented
+ *
+ * Return: true on success, false on failure
+ */
+bool six_trylock_convert(struct six_lock *lock,
+ enum six_lock_type from,
+ enum six_lock_type to)
+{
+ EBUG_ON(to == SIX_LOCK_write || from == SIX_LOCK_write);
+
+ if (to == from)
+ return true;
+
+ if (to == SIX_LOCK_read) {
+ six_lock_downgrade(lock);
+ return true;
+ } else {
+ return six_lock_tryupgrade(lock);
+ }
+}
+EXPORT_SYMBOL_GPL(six_trylock_convert);
+
+/**
+ * six_lock_increment - increase held lock count on a lock that is already held
+ * @lock: lock to increment
+ * @type: SIX_LOCK_read or SIX_LOCK_intent
+ *
+ * @lock must already be held, with a lock type that is greater than or equal to
+ * @type
+ *
+ * A corresponding six_unlock_type() call will be required for @lock to be fully
+ * unlocked.
+ */
+void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
+{
+ six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, _RET_IP_);
+
+ /* XXX: assert already locked, and that we don't overflow: */
+
+ switch (type) {
+ case SIX_LOCK_read:
+ if (lock->readers) {
+ this_cpu_inc(*lock->readers);
+ } else {
+ EBUG_ON(!(atomic_read(&lock->state) &
+ (SIX_LOCK_HELD_read|
+ SIX_LOCK_HELD_intent)));
+ atomic_add(l[type].lock_val, &lock->state);
+ }
+ break;
+ case SIX_LOCK_intent:
+ EBUG_ON(!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
+ lock->intent_lock_recurse++;
+ break;
+ case SIX_LOCK_write:
+ BUG();
+ break;
+ }
+}
+EXPORT_SYMBOL_GPL(six_lock_increment);
+
+/**
+ * six_lock_wakeup_all - wake up all waiters on @lock
+ * @lock: lock to wake up waiters for
+ *
+ * Wakeing up waiters will cause them to re-run should_sleep_fn, which may then
+ * abort the lock operation.
+ *
+ * This function is never needed in a bug-free program; it's only useful in
+ * debug code, e.g. to determine if a cycle detector is at fault.
+ */
+void six_lock_wakeup_all(struct six_lock *lock)
+{
+ u32 state = atomic_read(&lock->state);
+ struct six_lock_waiter *w;
+
+ six_lock_wakeup(lock, state, SIX_LOCK_read);
+ six_lock_wakeup(lock, state, SIX_LOCK_intent);
+ six_lock_wakeup(lock, state, SIX_LOCK_write);
+
+ raw_spin_lock(&lock->wait_lock);
+ list_for_each_entry(w, &lock->wait_list, list)
+ wake_up_process(w->task);
+ raw_spin_unlock(&lock->wait_lock);
+}
+EXPORT_SYMBOL_GPL(six_lock_wakeup_all);
+
+/**
+ * six_lock_counts - return held lock counts, for each lock type
+ * @lock: lock to return counters for
+ *
+ * Return: the number of times a lock is held for read, intent and write.
+ */
+struct six_lock_count six_lock_counts(struct six_lock *lock)
+{
+ struct six_lock_count ret;
+
+ ret.n[SIX_LOCK_read] = !lock->readers
+ ? atomic_read(&lock->state) & SIX_LOCK_HELD_read
+ : pcpu_read_count(lock);
+ ret.n[SIX_LOCK_intent] = !!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent) +
+ lock->intent_lock_recurse;
+ ret.n[SIX_LOCK_write] = !!(atomic_read(&lock->state) & SIX_LOCK_HELD_write);
+
+ return ret;
+}
+EXPORT_SYMBOL_GPL(six_lock_counts);
+
+/**
+ * six_lock_readers_add - directly manipulate reader count of a lock
+ * @lock: lock to add/subtract readers for
+ * @nr: reader count to add/subtract
+ *
+ * When an upper layer is implementing lock reentrency, we may have both read
+ * and intent locks on the same lock.
+ *
+ * When we need to take a write lock, the read locks will cause self-deadlock,
+ * because six locks themselves do not track which read locks are held by the
+ * current thread and which are held by a different thread - it does no
+ * per-thread tracking of held locks.
+ *
+ * The upper layer that is tracking held locks may however, if trylock() has
+ * failed, count up its own read locks, subtract them, take the write lock, and
+ * then re-add them.
+ *
+ * As in any other situation when taking a write lock, @lock must be held for
+ * intent one (or more) times, so @lock will never be left unlocked.
+ */
+void six_lock_readers_add(struct six_lock *lock, int nr)
+{
+ if (lock->readers) {
+ this_cpu_add(*lock->readers, nr);
+ } else {
+ EBUG_ON((int) (atomic_read(&lock->state) & SIX_LOCK_HELD_read) + nr < 0);
+ /* reader count starts at bit 0 */
+ atomic_add(nr, &lock->state);
+ }
+}
+EXPORT_SYMBOL_GPL(six_lock_readers_add);
+
+/**
+ * six_lock_exit - release resources held by a lock prior to freeing
+ * @lock: lock to exit
+ *
+ * When a lock was initialized in percpu mode (SIX_OLCK_INIT_PCPU), this is
+ * required to free the percpu read counts.
+ */
+void six_lock_exit(struct six_lock *lock)
+{
+ WARN_ON(lock->readers && pcpu_read_count(lock));
+ WARN_ON(atomic_read(&lock->state) & SIX_LOCK_HELD_read);
+
+ free_percpu(lock->readers);
+ lock->readers = NULL;
+}
+EXPORT_SYMBOL_GPL(six_lock_exit);
+
+void __six_lock_init(struct six_lock *lock, const char *name,
+ struct lock_class_key *key, enum six_lock_init_flags flags)
+{
+ atomic_set(&lock->state, 0);
+ raw_spin_lock_init(&lock->wait_lock);
+ INIT_LIST_HEAD(&lock->wait_list);
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+ debug_check_no_locks_freed((void *) lock, sizeof(*lock));
+ lockdep_init_map(&lock->dep_map, name, key, 0);
+#endif
+
+ /*
+ * Don't assume that we have real percpu variables available in
+ * userspace:
+ */
+#ifdef __KERNEL__
+ if (flags & SIX_LOCK_INIT_PCPU) {
+ /*
+ * We don't return an error here on memory allocation failure
+ * since percpu is an optimization, and locks will work with the
+ * same semantics in non-percpu mode: callers can check for
+ * failure if they wish by checking lock->readers, but generally
+ * will not want to treat it as an error.
+ */
+ lock->readers = alloc_percpu(unsigned);
+ }
+#endif
+}
+EXPORT_SYMBOL_GPL(__six_lock_init);
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef _LINUX_SIX_H
+#define _LINUX_SIX_H
+
+/**
+ * DOC: SIX locks overview
+ *
+ * Shared/intent/exclusive locks: sleepable read/write locks, like rw semaphores
+ * but with an additional state: read/shared, intent, exclusive/write
+ *
+ * The purpose of the intent state is to allow for greater concurrency on tree
+ * structures without deadlocking. In general, a read can't be upgraded to a
+ * write lock without deadlocking, so an operation that updates multiple nodes
+ * will have to take write locks for the full duration of the operation.
+ *
+ * But by adding an intent state, which is exclusive with other intent locks but
+ * not with readers, we can take intent locks at thte start of the operation,
+ * and then take write locks only for the actual update to each individual
+ * nodes, without deadlocking.
+ *
+ * Example usage:
+ * six_lock_read(&foo->lock);
+ * six_unlock_read(&foo->lock);
+ *
+ * An intent lock must be held before taking a write lock:
+ * six_lock_intent(&foo->lock);
+ * six_lock_write(&foo->lock);
+ * six_unlock_write(&foo->lock);
+ * six_unlock_intent(&foo->lock);
+ *
+ * Other operations:
+ * six_trylock_read()
+ * six_trylock_intent()
+ * six_trylock_write()
+ *
+ * six_lock_downgrade() convert from intent to read
+ * six_lock_tryupgrade() attempt to convert from read to intent, may fail
+ *
+ * There are also interfaces that take the lock type as an enum:
+ *
+ * six_lock_type(&foo->lock, SIX_LOCK_read);
+ * six_trylock_convert(&foo->lock, SIX_LOCK_read, SIX_LOCK_intent)
+ * six_lock_type(&foo->lock, SIX_LOCK_write);
+ * six_unlock_type(&foo->lock, SIX_LOCK_write);
+ * six_unlock_type(&foo->lock, SIX_LOCK_intent);
+ *
+ * Lock sequence numbers - unlock(), relock():
+ *
+ * Locks embed sequences numbers, which are incremented on write lock/unlock.
+ * This allows locks to be dropped and the retaken iff the state they protect
+ * hasn't changed; this makes it much easier to avoid holding locks while e.g.
+ * doing IO or allocating memory.
+ *
+ * Example usage:
+ * six_lock_read(&foo->lock);
+ * u32 seq = six_lock_seq(&foo->lock);
+ * six_unlock_read(&foo->lock);
+ *
+ * some_operation_that_may_block();
+ *
+ * if (six_relock_read(&foo->lock, seq)) { ... }
+ *
+ * If the relock operation succeeds, it is as if the lock was never unlocked.
+ *
+ * Reentrancy:
+ *
+ * Six locks are not by themselves reentrent, but have counters for both the
+ * read and intent states that can be used to provide reentrency by an upper
+ * layer that tracks held locks. If a lock is known to already be held in the
+ * read or intent state, six_lock_increment() can be used to bump the "lock
+ * held in this state" counter, increasing the number of unlock calls that
+ * will be required to fully unlock it.
+ *
+ * Example usage:
+ * six_lock_read(&foo->lock);
+ * six_lock_increment(&foo->lock, SIX_LOCK_read);
+ * six_unlock_read(&foo->lock);
+ * six_unlock_read(&foo->lock);
+ * foo->lock is now fully unlocked.
+ *
+ * Since the intent state supercedes read, it's legal to increment the read
+ * counter when holding an intent lock, but not the reverse.
+ *
+ * A lock may only be held once for write: six_lock_increment(.., SIX_LOCK_write)
+ * is not legal.
+ *
+ * should_sleep_fn:
+ *
+ * There is a six_lock() variant that takes a function pointer that is called
+ * immediately prior to schedule() when blocking, and may return an error to
+ * abort.
+ *
+ * One possible use for this feature is when objects being locked are part of
+ * a cache and may reused, and lock ordering is based on a property of the
+ * object that will change when the object is reused - i.e. logical key order.
+ *
+ * If looking up an object in the cache may race with object reuse, and lock
+ * ordering is required to prevent deadlock, object reuse may change the
+ * correct lock order for that object and cause a deadlock. should_sleep_fn
+ * can be used to check if the object is still the object we want and avoid
+ * this deadlock.
+ *
+ * Wait list entry interface:
+ *
+ * There is a six_lock() variant, six_lock_waiter(), that takes a pointer to a
+ * wait list entry. By embedding six_lock_waiter into another object, and by
+ * traversing lock waitlists, it is then possible for an upper layer to
+ * implement full cycle detection for deadlock avoidance.
+ *
+ * should_sleep_fn should be used for invoking the cycle detector, walking the
+ * graph of held locks to check for a deadlock. The upper layer must track
+ * held locks for each thread, and each thread's held locks must be reachable
+ * from its six_lock_waiter object.
+ *
+ * six_lock_waiter() will add the wait object to the waitlist re-trying taking
+ * the lock, and before calling should_sleep_fn, and the wait object will not
+ * be removed from the waitlist until either the lock has been successfully
+ * acquired, or we aborted because should_sleep_fn returned an error.
+ *
+ * Also, six_lock_waiter contains a timestamp, and waiters on a waitlist will
+ * have timestamps in strictly ascending order - this is so the timestamp can
+ * be used as a cursor for lock graph traverse.
+ */
+
+#include <linux/lockdep.h>
+#include <linux/sched.h>
+#include <linux/types.h>
+
+enum six_lock_type {
+ SIX_LOCK_read,
+ SIX_LOCK_intent,
+ SIX_LOCK_write,
+};
+
+struct six_lock {
+ atomic_t state;
+ u32 seq;
+ unsigned intent_lock_recurse;
+ struct task_struct *owner;
+ unsigned __percpu *readers;
+ raw_spinlock_t wait_lock;
+ struct list_head wait_list;
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+ struct lockdep_map dep_map;
+#endif
+};
+
+struct six_lock_waiter {
+ struct list_head list;
+ struct task_struct *task;
+ enum six_lock_type lock_want;
+ bool lock_acquired;
+ u64 start_time;
+};
+
+typedef int (*six_lock_should_sleep_fn)(struct six_lock *lock, void *);
+
+void six_lock_exit(struct six_lock *lock);
+
+enum six_lock_init_flags {
+ SIX_LOCK_INIT_PCPU = 1U << 0,
+};
+
+void __six_lock_init(struct six_lock *lock, const char *name,
+ struct lock_class_key *key, enum six_lock_init_flags flags);
+
+/**
+ * six_lock_init - initialize a six lock
+ * @lock: lock to initialize
+ * @flags: optional flags, i.e. SIX_LOCK_INIT_PCPU
+ */
+#define six_lock_init(lock, flags) \
+do { \
+ static struct lock_class_key __key; \
+ \
+ __six_lock_init((lock), #lock, &__key, flags); \
+} while (0)
+
+/**
+ * six_lock_seq - obtain current lock sequence number
+ * @lock: six_lock to obtain sequence number for
+ *
+ * @lock should be held for read or intent, and not write
+ *
+ * By saving the lock sequence number, we can unlock @lock and then (typically
+ * after some blocking operation) attempt to relock it: the relock will succeed
+ * if the sequence number hasn't changed, meaning no write locks have been taken
+ * and state corresponding to what @lock protects is still valid.
+ */
+static inline u32 six_lock_seq(const struct six_lock *lock)
+{
+ return lock->seq;
+}
+
+bool six_trylock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip);
+
+/**
+ * six_trylock_type - attempt to take a six lock without blocking
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ *
+ * Return: true on success, false on failure.
+ */
+static inline bool six_trylock_type(struct six_lock *lock, enum six_lock_type type)
+{
+ return six_trylock_ip(lock, type, _THIS_IP_);
+}
+
+int six_lock_ip_waiter(struct six_lock *lock, enum six_lock_type type,
+ struct six_lock_waiter *wait,
+ six_lock_should_sleep_fn should_sleep_fn, void *p,
+ unsigned long ip);
+
+/**
+ * six_lock_waiter - take a lock, with full waitlist interface
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @wait: pointer to wait object, which will be added to lock's waitlist
+ * @should_sleep_fn: callback run after adding to waitlist, immediately prior
+ * to scheduling
+ * @p: passed through to @should_sleep_fn
+ *
+ * This is a convenience wrapper around six_lock_ip_waiter(), see that function
+ * for full documentation.
+ *
+ * Return: 0 on success, or the return code from @should_sleep_fn on failure.
+ */
+static inline int six_lock_waiter(struct six_lock *lock, enum six_lock_type type,
+ struct six_lock_waiter *wait,
+ six_lock_should_sleep_fn should_sleep_fn, void *p)
+{
+ return six_lock_ip_waiter(lock, type, wait, should_sleep_fn, p, _THIS_IP_);
+}
+
+/**
+ * six_lock_ip - take a six lock lock
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @should_sleep_fn: callback run after adding to waitlist, immediately prior
+ * to scheduling
+ * @p: passed through to @should_sleep_fn
+ * @ip: ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * Return: 0 on success, or the return code from @should_sleep_fn on failure.
+ */
+static inline int six_lock_ip(struct six_lock *lock, enum six_lock_type type,
+ six_lock_should_sleep_fn should_sleep_fn, void *p,
+ unsigned long ip)
+{
+ struct six_lock_waiter wait;
+
+ return six_lock_ip_waiter(lock, type, &wait, should_sleep_fn, p, ip);
+}
+
+/**
+ * six_lock_type - take a six lock lock
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @should_sleep_fn: callback run after adding to waitlist, immediately prior
+ * to scheduling
+ * @p: passed through to @should_sleep_fn
+ *
+ * Return: 0 on success, or the return code from @should_sleep_fn on failure.
+ */
+static inline int six_lock_type(struct six_lock *lock, enum six_lock_type type,
+ six_lock_should_sleep_fn should_sleep_fn, void *p)
+{
+ struct six_lock_waiter wait;
+
+ return six_lock_ip_waiter(lock, type, &wait, should_sleep_fn, p, _THIS_IP_);
+}
+
+bool six_relock_ip(struct six_lock *lock, enum six_lock_type type,
+ unsigned seq, unsigned long ip);
+
+/**
+ * six_relock_type - attempt to re-take a lock that was held previously
+ * @lock: lock to take
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @seq: lock sequence number obtained from six_lock_seq() while lock was
+ * held previously
+ *
+ * Return: true on success, false on failure.
+ */
+static inline bool six_relock_type(struct six_lock *lock, enum six_lock_type type,
+ unsigned seq)
+{
+ return six_relock_ip(lock, type, seq, _THIS_IP_);
+}
+
+void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip);
+
+/**
+ * six_unlock_type - drop a six lock
+ * @lock: lock to unlock
+ * @type: SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ *
+ * When a lock is held multiple times (because six_lock_incement()) was used),
+ * this decrements the 'lock held' counter by one.
+ *
+ * For example:
+ * six_lock_read(&foo->lock); read count 1
+ * six_lock_increment(&foo->lock, SIX_LOCK_read); read count 2
+ * six_lock_unlock(&foo->lock, SIX_LOCK_read); read count 1
+ * six_lock_unlock(&foo->lock, SIX_LOCK_read); read count 0
+ */
+static inline void six_unlock_type(struct six_lock *lock, enum six_lock_type type)
+{
+ six_unlock_ip(lock, type, _THIS_IP_);
+}
+
+#define __SIX_LOCK(type) \
+static inline bool six_trylock_ip_##type(struct six_lock *lock, unsigned long ip)\
+{ \
+ return six_trylock_ip(lock, SIX_LOCK_##type, ip); \
+} \
+ \
+static inline bool six_trylock_##type(struct six_lock *lock) \
+{ \
+ return six_trylock_ip(lock, SIX_LOCK_##type, _THIS_IP_); \
+} \
+ \
+static inline int six_lock_ip_waiter_##type(struct six_lock *lock, \
+ struct six_lock_waiter *wait, \
+ six_lock_should_sleep_fn should_sleep_fn, void *p,\
+ unsigned long ip) \
+{ \
+ return six_lock_ip_waiter(lock, SIX_LOCK_##type, wait, should_sleep_fn, p, ip);\
+} \
+ \
+static inline int six_lock_ip_##type(struct six_lock *lock, \
+ six_lock_should_sleep_fn should_sleep_fn, void *p, \
+ unsigned long ip) \
+{ \
+ return six_lock_ip(lock, SIX_LOCK_##type, should_sleep_fn, p, ip);\
+} \
+ \
+static inline bool six_relock_ip_##type(struct six_lock *lock, u32 seq, unsigned long ip)\
+{ \
+ return six_relock_ip(lock, SIX_LOCK_##type, seq, ip); \
+} \
+ \
+static inline bool six_relock_##type(struct six_lock *lock, u32 seq) \
+{ \
+ return six_relock_ip(lock, SIX_LOCK_##type, seq, _THIS_IP_); \
+} \
+ \
+static inline int six_lock_##type(struct six_lock *lock, \
+ six_lock_should_sleep_fn fn, void *p)\
+{ \
+ return six_lock_ip_##type(lock, fn, p, _THIS_IP_); \
+} \
+ \
+static inline void six_unlock_ip_##type(struct six_lock *lock, unsigned long ip) \
+{ \
+ six_unlock_ip(lock, SIX_LOCK_##type, ip); \
+} \
+ \
+static inline void six_unlock_##type(struct six_lock *lock) \
+{ \
+ six_unlock_ip(lock, SIX_LOCK_##type, _THIS_IP_); \
+}
+
+__SIX_LOCK(read)
+__SIX_LOCK(intent)
+__SIX_LOCK(write)
+#undef __SIX_LOCK
+
+void six_lock_downgrade(struct six_lock *);
+bool six_lock_tryupgrade(struct six_lock *);
+bool six_trylock_convert(struct six_lock *, enum six_lock_type,
+ enum six_lock_type);
+
+void six_lock_increment(struct six_lock *, enum six_lock_type);
+
+void six_lock_wakeup_all(struct six_lock *);
+
+struct six_lock_count {
+ unsigned n[3];
+};
+
+struct six_lock_count six_lock_counts(struct six_lock *);
+void six_lock_readers_add(struct six_lock *, int);
+
+#endif /* _LINUX_SIX_H */
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "bkey_buf.h"
+#include "btree_key_cache.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "errcode.h"
+#include "error.h"
+#include "fs.h"
+#include "snapshot.h"
+
+#include <linux/random.h>
+
+/*
+ * Snapshot trees:
+ *
+ * Keys in BTREE_ID_snapshot_trees identify a whole tree of snapshot nodes; they
+ * exist to provide a stable identifier for the whole lifetime of a snapshot
+ * tree.
+ */
+
+void bch2_snapshot_tree_to_text(struct printbuf *out, struct bch_fs *c,
+ struct bkey_s_c k)
+{
+ struct bkey_s_c_snapshot_tree t = bkey_s_c_to_snapshot_tree(k);
+
+ prt_printf(out, "subvol %u root snapshot %u",
+ le32_to_cpu(t.v->master_subvol),
+ le32_to_cpu(t.v->root_snapshot));
+}
+
+int bch2_snapshot_tree_invalid(struct bch_fs *c, struct bkey_s_c k,
+ enum bkey_invalid_flags flags,
+ struct printbuf *err)
+{
+ int ret = 0;
+
+ bkey_fsck_err_on(bkey_gt(k.k->p, POS(0, U32_MAX)) ||
+ bkey_lt(k.k->p, POS(0, 1)), c, err,
+ snapshot_tree_pos_bad,
+ "bad pos");
+fsck_err:
+ return ret;
+}
+
+int bch2_snapshot_tree_lookup(struct btree_trans *trans, u32 id,
+ struct bch_snapshot_tree *s)
+{
+ int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_snapshot_trees, POS(0, id),
+ BTREE_ITER_WITH_UPDATES, snapshot_tree, s);
+
+ if (bch2_err_matches(ret, ENOENT))
+ ret = -BCH_ERR_ENOENT_snapshot_tree;
+ return ret;
+}
+
+struct bkey_i_snapshot_tree *
+__bch2_snapshot_tree_create(struct btree_trans *trans)
+{
+ struct btree_iter iter;
+ int ret = bch2_bkey_get_empty_slot(trans, &iter,
+ BTREE_ID_snapshot_trees, POS(0, U32_MAX));
+ struct bkey_i_snapshot_tree *s_t;
+
+ if (ret == -BCH_ERR_ENOSPC_btree_slot)
+ ret = -BCH_ERR_ENOSPC_snapshot_tree;
+ if (ret)
+ return ERR_PTR(ret);
+
+ s_t = bch2_bkey_alloc(trans, &iter, 0, snapshot_tree);
+ ret = PTR_ERR_OR_ZERO(s_t);
+ bch2_trans_iter_exit(trans, &iter);
+ return ret ? ERR_PTR(ret) : s_t;
+}
+
+static int bch2_snapshot_tree_create(struct btree_trans *trans,
+ u32 root_id, u32 subvol_id, u32 *tree_id)
+{
+ struct bkey_i_snapshot_tree *n_tree =
+ __bch2_snapshot_tree_create(trans);
+
+ if (IS_ERR(n_tree))
+ return PTR_ERR(n_tree);
+
+ n_tree->v.master_subvol = cpu_to_le32(subvol_id);
+ n_tree->v.root_snapshot = cpu_to_le32(root_id);
+ *tree_id = n_tree->k.p.offset;
+ return 0;
+}
+
+/* Snapshot nodes: */
+
+static bool bch2_snapshot_is_ancestor_early(struct bch_fs *c, u32 id, u32 ancestor)
+{
+ struct snapshot_table *t;
+
+ rcu_read_lock();
+ t = rcu_dereference(c->snapshots);
+
+ while (id && id < ancestor)
+ id = __snapshot_t(t, id)->parent;
+ rcu_read_unlock();
+
+ return id == ancestor;
+}
+
+static inline u32 get_ancestor_below(struct snapshot_table *t, u32 id, u32 ancestor)
+{
+ const struct snapshot_t *s = __snapshot_t(t, id);
+
+ if (s->skip[2] <= ancestor)
+ return s->skip[2];
+ if (s->skip[1] <= ancestor)
+ return s->skip[1];
+ if (s->skip[0] <= ancestor)
+ return s->skip[0];
+ return s->parent;
+}
+
+bool __bch2_snapshot_is_ancestor(struct bch_fs *c, u32 id, u32 ancestor)
+{
+ struct snapshot_table *t;
+ bool ret;
+
+ EBUG_ON(c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_snapshots);
+
+ rcu_read_lock();
+ t = rcu_dereference(c->snapshots);
+
+ while (id && id < ancestor - IS_ANCESTOR_BITMAP)
+ id = get_ancestor_below(t, id, ancestor);
+
+ if (id && id < ancestor) {
+ ret = test_bit(ancestor - id - 1, __snapshot_t(t, id)->is_ancestor);
+
+ EBUG_ON(ret != bch2_snapshot_is_ancestor_early(c, id, ancestor));
+ } else {
+ ret = id == ancestor;
+ }
+
+ rcu_read_unlock();
+
+ return ret;
+}
+
+static noinline struct snapshot_t *__snapshot_t_mut(struct bch_fs *c, u32 id)
+{
+ size_t idx = U32_MAX - id;
+ size_t new_size;
+ struct snapshot_table *new, *old;
+
+ new_size = max(16UL, roundup_pow_of_two(idx + 1));
+
+ new = kvzalloc(struct_size(new, s, new_size), GFP_KERNEL);
+ if (!new)
+ return NULL;
+
+ old = rcu_dereference_protected(c->snapshots, true);
+ if (old)
+ memcpy(new->s,
+ rcu_dereference_protected(c->snapshots, true)->s,
+ sizeof(new->s[0]) * c->snapshot_table_size);
+
+ rcu_assign_pointer(c->snapshots, new);
+ c->snapshot_table_size = new_size;
+ kvfree_rcu_mightsleep(old);
+
+ return &rcu_dereference_protected(c->snapshots, true)->s[idx];
+}
+
+static inline struct snapshot_t *snapshot_t_mut(struct bch_fs *c, u32 id)
+{
+ size_t idx = U32_MAX - id;
+
+ lockdep_assert_held(&c->snapshot_table_lock);
+
+ if (likely(idx < c->snapshot_table_size))
+ return &rcu_dereference_protected(c->snapshots, true)->s[idx];
+
+ return __snapshot_t_mut(c, id);
+}
+
+void bch2_snapshot_to_text(struct printbuf *out, struct bch_fs *c,
+ struct bkey_s_c k)
+{
+ struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(k);
+
+ prt_printf(out, "is_subvol %llu deleted %llu parent %10u children %10u %10u subvol %u tree %u",
+ BCH_SNAPSHOT_SUBVOL(s.v),
+ BCH_SNAPSHOT_DELETED(s.v),
+ le32_to_cpu(s.v->parent),
+ le32_to_cpu(s.v->children[0]),
+ le32_to_cpu(s.v->children[1]),
+ le32_to_cpu(s.v->subvol),
+ le32_to_cpu(s.v->tree));
+
+ if (bkey_val_bytes(k.k) > offsetof(struct bch_snapshot, depth))
+ prt_printf(out, " depth %u skiplist %u %u %u",
+ le32_to_cpu(s.v->depth),
+ le32_to_cpu(s.v->skip[0]),
+ le32_to_cpu(s.v->skip[1]),
+ le32_to_cpu(s.v->skip[2]));
+}
+
+int bch2_snapshot_invalid(struct bch_fs *c, struct bkey_s_c k,
+ enum bkey_invalid_flags flags,
+ struct printbuf *err)
+{
+ struct bkey_s_c_snapshot s;
+ u32 i, id;
+ int ret = 0;
+
+ bkey_fsck_err_on(bkey_gt(k.k->p, POS(0, U32_MAX)) ||
+ bkey_lt(k.k->p, POS(0, 1)), c, err,
+ snapshot_pos_bad,
+ "bad pos");
+
+ s = bkey_s_c_to_snapshot(k);
+
+ id = le32_to_cpu(s.v->parent);
+ bkey_fsck_err_on(id && id <= k.k->p.offset, c, err,
+ snapshot_parent_bad,
+ "bad parent node (%u <= %llu)",
+ id, k.k->p.offset);
+
+ bkey_fsck_err_on(le32_to_cpu(s.v->children[0]) < le32_to_cpu(s.v->children[1]), c, err,
+ snapshot_children_not_normalized,
+ "children not normalized");
+
+ bkey_fsck_err_on(s.v->children[0] && s.v->children[0] == s.v->children[1], c, err,
+ snapshot_child_duplicate,
+ "duplicate child nodes");
+
+ for (i = 0; i < 2; i++) {
+ id = le32_to_cpu(s.v->children[i]);
+
+ bkey_fsck_err_on(id >= k.k->p.offset, c, err,
+ snapshot_child_bad,
+ "bad child node (%u >= %llu)",
+ id, k.k->p.offset);
+ }
+
+ if (bkey_val_bytes(k.k) > offsetof(struct bch_snapshot, skip)) {
+ bkey_fsck_err_on(le32_to_cpu(s.v->skip[0]) > le32_to_cpu(s.v->skip[1]) ||
+ le32_to_cpu(s.v->skip[1]) > le32_to_cpu(s.v->skip[2]), c, err,
+ snapshot_skiplist_not_normalized,
+ "skiplist not normalized");
+
+ for (i = 0; i < ARRAY_SIZE(s.v->skip); i++) {
+ id = le32_to_cpu(s.v->skip[i]);
+
+ bkey_fsck_err_on(id && id < le32_to_cpu(s.v->parent), c, err,
+ snapshot_skiplist_bad,
+ "bad skiplist node %u", id);
+ }
+ }
+fsck_err:
+ return ret;
+}
+
+static void __set_is_ancestor_bitmap(struct bch_fs *c, u32 id)
+{
+ struct snapshot_t *t = snapshot_t_mut(c, id);
+ u32 parent = id;
+
+ while ((parent = bch2_snapshot_parent_early(c, parent)) &&
+ parent - id - 1 < IS_ANCESTOR_BITMAP)
+ __set_bit(parent - id - 1, t->is_ancestor);
+}
+
+static void set_is_ancestor_bitmap(struct bch_fs *c, u32 id)
+{
+ mutex_lock(&c->snapshot_table_lock);
+ __set_is_ancestor_bitmap(c, id);
+ mutex_unlock(&c->snapshot_table_lock);
+}
+
+int bch2_mark_snapshot(struct btree_trans *trans,
+ enum btree_id btree, unsigned level,
+ struct bkey_s_c old, struct bkey_s_c new,
+ unsigned flags)
+{
+ struct bch_fs *c = trans->c;
+ struct snapshot_t *t;
+ u32 id = new.k->p.offset;
+ int ret = 0;
+
+ mutex_lock(&c->snapshot_table_lock);
+
+ t = snapshot_t_mut(c, id);
+ if (!t) {
+ ret = -BCH_ERR_ENOMEM_mark_snapshot;
+ goto err;
+ }
+
+ if (new.k->type == KEY_TYPE_snapshot) {
+ struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(new);
+
+ t->parent = le32_to_cpu(s.v->parent);
+ t->children[0] = le32_to_cpu(s.v->children[0]);
+ t->children[1] = le32_to_cpu(s.v->children[1]);
+ t->subvol = BCH_SNAPSHOT_SUBVOL(s.v) ? le32_to_cpu(s.v->subvol) : 0;
+ t->tree = le32_to_cpu(s.v->tree);
+
+ if (bkey_val_bytes(s.k) > offsetof(struct bch_snapshot, depth)) {
+ t->depth = le32_to_cpu(s.v->depth);
+ t->skip[0] = le32_to_cpu(s.v->skip[0]);
+ t->skip[1] = le32_to_cpu(s.v->skip[1]);
+ t->skip[2] = le32_to_cpu(s.v->skip[2]);
+ } else {
+ t->depth = 0;
+ t->skip[0] = 0;
+ t->skip[1] = 0;
+ t->skip[2] = 0;
+ }
+
+ __set_is_ancestor_bitmap(c, id);
+
+ if (BCH_SNAPSHOT_DELETED(s.v)) {
+ set_bit(BCH_FS_NEED_DELETE_DEAD_SNAPSHOTS, &c->flags);
+ if (c->curr_recovery_pass > BCH_RECOVERY_PASS_delete_dead_snapshots)
+ bch2_delete_dead_snapshots_async(c);
+ }
+ } else {
+ memset(t, 0, sizeof(*t));
+ }
+err:
+ mutex_unlock(&c->snapshot_table_lock);
+ return ret;
+}
+
+int bch2_snapshot_lookup(struct btree_trans *trans, u32 id,
+ struct bch_snapshot *s)
+{
+ return bch2_bkey_get_val_typed(trans, BTREE_ID_snapshots, POS(0, id),
+ BTREE_ITER_WITH_UPDATES, snapshot, s);
+}
+
+static int bch2_snapshot_live(struct btree_trans *trans, u32 id)
+{
+ struct bch_snapshot v;
+ int ret;
+
+ if (!id)
+ return 0;
+
+ ret = bch2_snapshot_lookup(trans, id, &v);
+ if (bch2_err_matches(ret, ENOENT))
+ bch_err(trans->c, "snapshot node %u not found", id);
+ if (ret)
+ return ret;
+
+ return !BCH_SNAPSHOT_DELETED(&v);
+}
+
+/*
+ * If @k is a snapshot with just one live child, it's part of a linear chain,
+ * which we consider to be an equivalence class: and then after snapshot
+ * deletion cleanup, there should only be a single key at a given position in
+ * this equivalence class.
+ *
+ * This sets the equivalence class of @k to be the child's equivalence class, if
+ * it's part of such a linear chain: this correctly sets equivalence classes on
+ * startup if we run leaf to root (i.e. in natural key order).
+ */
+static int bch2_snapshot_set_equiv(struct btree_trans *trans, struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ unsigned i, nr_live = 0, live_idx = 0;
+ struct bkey_s_c_snapshot snap;
+ u32 id = k.k->p.offset, child[2];
+
+ if (k.k->type != KEY_TYPE_snapshot)
+ return 0;
+
+ snap = bkey_s_c_to_snapshot(k);
+
+ child[0] = le32_to_cpu(snap.v->children[0]);
+ child[1] = le32_to_cpu(snap.v->children[1]);
+
+ for (i = 0; i < 2; i++) {
+ int ret = bch2_snapshot_live(trans, child[i]);
+
+ if (ret < 0)
+ return ret;
+
+ if (ret)
+ live_idx = i;
+ nr_live += ret;
+ }
+
+ mutex_lock(&c->snapshot_table_lock);
+
+ snapshot_t_mut(c, id)->equiv = nr_live == 1
+ ? snapshot_t_mut(c, child[live_idx])->equiv
+ : id;
+
+ mutex_unlock(&c->snapshot_table_lock);
+
+ return 0;
+}
+
+/* fsck: */
+
+static u32 bch2_snapshot_child(struct bch_fs *c, u32 id, unsigned child)
+{
+ return snapshot_t(c, id)->children[child];
+}
+
+static u32 bch2_snapshot_left_child(struct bch_fs *c, u32 id)
+{
+ return bch2_snapshot_child(c, id, 0);
+}
+
+static u32 bch2_snapshot_right_child(struct bch_fs *c, u32 id)
+{
+ return bch2_snapshot_child(c, id, 1);
+}
+
+static u32 bch2_snapshot_tree_next(struct bch_fs *c, u32 id)
+{
+ u32 n, parent;
+
+ n = bch2_snapshot_left_child(c, id);
+ if (n)
+ return n;
+
+ while ((parent = bch2_snapshot_parent(c, id))) {
+ n = bch2_snapshot_right_child(c, parent);
+ if (n && n != id)
+ return n;
+ id = parent;
+ }
+
+ return 0;
+}
+
+static u32 bch2_snapshot_tree_oldest_subvol(struct bch_fs *c, u32 snapshot_root)
+{
+ u32 id = snapshot_root;
+ u32 subvol = 0, s;
+
+ while (id) {
+ s = snapshot_t(c, id)->subvol;
+
+ if (s && (!subvol || s < subvol))
+ subvol = s;
+
+ id = bch2_snapshot_tree_next(c, id);
+ }
+
+ return subvol;
+}
+
+static int bch2_snapshot_tree_master_subvol(struct btree_trans *trans,
+ u32 snapshot_root, u32 *subvol_id)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ struct bkey_s_c_subvolume s;
+ bool found = false;
+ int ret;
+
+ for_each_btree_key_norestart(trans, iter, BTREE_ID_subvolumes, POS_MIN,
+ 0, k, ret) {
+ if (k.k->type != KEY_TYPE_subvolume)
+ continue;
+
+ s = bkey_s_c_to_subvolume(k);
+ if (!bch2_snapshot_is_ancestor(c, le32_to_cpu(s.v->snapshot), snapshot_root))
+ continue;
+ if (!BCH_SUBVOLUME_SNAP(s.v)) {
+ *subvol_id = s.k->p.offset;
+ found = true;
+ break;
+ }
+ }
+
+ bch2_trans_iter_exit(trans, &iter);
+
+ if (!ret && !found) {
+ struct bkey_i_subvolume *u;
+
+ *subvol_id = bch2_snapshot_tree_oldest_subvol(c, snapshot_root);
+
+ u = bch2_bkey_get_mut_typed(trans, &iter,
+ BTREE_ID_subvolumes, POS(0, *subvol_id),
+ 0, subvolume);
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ return ret;
+
+ SET_BCH_SUBVOLUME_SNAP(&u->v, false);
+ }
+
+ return ret;
+}
+
+static int check_snapshot_tree(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ struct bkey_s_c_snapshot_tree st;
+ struct bch_snapshot s;
+ struct bch_subvolume subvol;
+ struct printbuf buf = PRINTBUF;
+ u32 root_id;
+ int ret;
+
+ if (k.k->type != KEY_TYPE_snapshot_tree)
+ return 0;
+
+ st = bkey_s_c_to_snapshot_tree(k);
+ root_id = le32_to_cpu(st.v->root_snapshot);
+
+ ret = bch2_snapshot_lookup(trans, root_id, &s);
+ if (ret && !bch2_err_matches(ret, ENOENT))
+ goto err;
+
+ if (fsck_err_on(ret ||
+ root_id != bch2_snapshot_root(c, root_id) ||
+ st.k->p.offset != le32_to_cpu(s.tree),
+ c, snapshot_tree_to_missing_snapshot,
+ "snapshot tree points to missing/incorrect snapshot:\n %s",
+ (bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf))) {
+ ret = bch2_btree_delete_at(trans, iter, 0);
+ goto err;
+ }
+
+ ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol),
+ false, 0, &subvol);
+ if (ret && !bch2_err_matches(ret, ENOENT))
+ goto err;
+
+ if (fsck_err_on(ret,
+ c, snapshot_tree_to_missing_subvol,
+ "snapshot tree points to missing subvolume:\n %s",
+ (printbuf_reset(&buf),
+ bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf)) ||
+ fsck_err_on(!bch2_snapshot_is_ancestor_early(c,
+ le32_to_cpu(subvol.snapshot),
+ root_id),
+ c, snapshot_tree_to_wrong_subvol,
+ "snapshot tree points to subvolume that does not point to snapshot in this tree:\n %s",
+ (printbuf_reset(&buf),
+ bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf)) ||
+ fsck_err_on(BCH_SUBVOLUME_SNAP(&subvol),
+ c, snapshot_tree_to_snapshot_subvol,
+ "snapshot tree points to snapshot subvolume:\n %s",
+ (printbuf_reset(&buf),
+ bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf))) {
+ struct bkey_i_snapshot_tree *u;
+ u32 subvol_id;
+
+ ret = bch2_snapshot_tree_master_subvol(trans, root_id, &subvol_id);
+ if (ret)
+ goto err;
+
+ u = bch2_bkey_make_mut_typed(trans, iter, &k, 0, snapshot_tree);
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ goto err;
+
+ u->v.master_subvol = cpu_to_le32(subvol_id);
+ st = snapshot_tree_i_to_s_c(u);
+ }
+err:
+fsck_err:
+ printbuf_exit(&buf);
+ return ret;
+}
+
+/*
+ * For each snapshot_tree, make sure it points to the root of a snapshot tree
+ * and that snapshot entry points back to it, or delete it.
+ *
+ * And, make sure it points to a subvolume within that snapshot tree, or correct
+ * it to point to the oldest subvolume within that snapshot tree.
+ */
+int bch2_check_snapshot_trees(struct bch_fs *c)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret;
+
+ ret = bch2_trans_run(c,
+ for_each_btree_key_commit(trans, iter,
+ BTREE_ID_snapshot_trees, POS_MIN,
+ BTREE_ITER_PREFETCH, k,
+ NULL, NULL, BCH_TRANS_COMMIT_lazy_rw|BCH_TRANS_COMMIT_no_enospc,
+ check_snapshot_tree(trans, &iter, k)));
+
+ if (ret)
+ bch_err(c, "error %i checking snapshot trees", ret);
+ return ret;
+}
+
+/*
+ * Look up snapshot tree for @tree_id and find root,
+ * make sure @snap_id is a descendent:
+ */
+static int snapshot_tree_ptr_good(struct btree_trans *trans,
+ u32 snap_id, u32 tree_id)
+{
+ struct bch_snapshot_tree s_t;
+ int ret = bch2_snapshot_tree_lookup(trans, tree_id, &s_t);
+
+ if (bch2_err_matches(ret, ENOENT))
+ return 0;
+ if (ret)
+ return ret;
+
+ return bch2_snapshot_is_ancestor_early(trans->c, snap_id, le32_to_cpu(s_t.root_snapshot));
+}
+
+u32 bch2_snapshot_skiplist_get(struct bch_fs *c, u32 id)
+{
+ const struct snapshot_t *s;
+
+ if (!id)
+ return 0;
+
+ rcu_read_lock();
+ s = snapshot_t(c, id);
+ if (s->parent)
+ id = bch2_snapshot_nth_parent(c, id, get_random_u32_below(s->depth));
+ rcu_read_unlock();
+
+ return id;
+}
+
+static int snapshot_skiplist_good(struct btree_trans *trans, u32 id, struct bch_snapshot s)
+{
+ unsigned i;
+
+ for (i = 0; i < 3; i++)
+ if (!s.parent) {
+ if (s.skip[i])
+ return false;
+ } else {
+ if (!bch2_snapshot_is_ancestor_early(trans->c, id, le32_to_cpu(s.skip[i])))
+ return false;
+ }
+
+ return true;
+}
+
+/*
+ * snapshot_tree pointer was incorrect: look up root snapshot node, make sure
+ * its snapshot_tree pointer is correct (allocate new one if necessary), then
+ * update this node's pointer to root node's pointer:
+ */
+static int snapshot_tree_ptr_repair(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k,
+ struct bch_snapshot *s)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter root_iter;
+ struct bch_snapshot_tree s_t;
+ struct bkey_s_c_snapshot root;
+ struct bkey_i_snapshot *u;
+ u32 root_id = bch2_snapshot_root(c, k.k->p.offset), tree_id;
+ int ret;
+
+ root = bch2_bkey_get_iter_typed(trans, &root_iter,
+ BTREE_ID_snapshots, POS(0, root_id),
+ BTREE_ITER_WITH_UPDATES, snapshot);
+ ret = bkey_err(root);
+ if (ret)
+ goto err;
+
+ tree_id = le32_to_cpu(root.v->tree);
+
+ ret = bch2_snapshot_tree_lookup(trans, tree_id, &s_t);
+ if (ret && !bch2_err_matches(ret, ENOENT))
+ return ret;
+
+ if (ret || le32_to_cpu(s_t.root_snapshot) != root_id) {
+ u = bch2_bkey_make_mut_typed(trans, &root_iter, &root.s_c, 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(u) ?:
+ bch2_snapshot_tree_create(trans, root_id,
+ bch2_snapshot_tree_oldest_subvol(c, root_id),
+ &tree_id);
+ if (ret)
+ goto err;
+
+ u->v.tree = cpu_to_le32(tree_id);
+ if (k.k->p.offset == root_id)
+ *s = u->v;
+ }
+
+ if (k.k->p.offset != root_id) {
+ u = bch2_bkey_make_mut_typed(trans, iter, &k, 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ goto err;
+
+ u->v.tree = cpu_to_le32(tree_id);
+ *s = u->v;
+ }
+err:
+ bch2_trans_iter_exit(trans, &root_iter);
+ return ret;
+}
+
+static int check_snapshot(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ struct bch_snapshot s;
+ struct bch_subvolume subvol;
+ struct bch_snapshot v;
+ struct bkey_i_snapshot *u;
+ u32 parent_id = bch2_snapshot_parent_early(c, k.k->p.offset);
+ u32 real_depth;
+ struct printbuf buf = PRINTBUF;
+ bool should_have_subvol;
+ u32 i, id;
+ int ret = 0;
+
+ if (k.k->type != KEY_TYPE_snapshot)
+ return 0;
+
+ memset(&s, 0, sizeof(s));
+ memcpy(&s, k.v, bkey_val_bytes(k.k));
+
+ id = le32_to_cpu(s.parent);
+ if (id) {
+ ret = bch2_snapshot_lookup(trans, id, &v);
+ if (bch2_err_matches(ret, ENOENT))
+ bch_err(c, "snapshot with nonexistent parent:\n %s",
+ (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
+ if (ret)
+ goto err;
+
+ if (le32_to_cpu(v.children[0]) != k.k->p.offset &&
+ le32_to_cpu(v.children[1]) != k.k->p.offset) {
+ bch_err(c, "snapshot parent %u missing pointer to child %llu",
+ id, k.k->p.offset);
+ ret = -EINVAL;
+ goto err;
+ }
+ }
+
+ for (i = 0; i < 2 && s.children[i]; i++) {
+ id = le32_to_cpu(s.children[i]);
+
+ ret = bch2_snapshot_lookup(trans, id, &v);
+ if (bch2_err_matches(ret, ENOENT))
+ bch_err(c, "snapshot node %llu has nonexistent child %u",
+ k.k->p.offset, id);
+ if (ret)
+ goto err;
+
+ if (le32_to_cpu(v.parent) != k.k->p.offset) {
+ bch_err(c, "snapshot child %u has wrong parent (got %u should be %llu)",
+ id, le32_to_cpu(v.parent), k.k->p.offset);
+ ret = -EINVAL;
+ goto err;
+ }
+ }
+
+ should_have_subvol = BCH_SNAPSHOT_SUBVOL(&s) &&
+ !BCH_SNAPSHOT_DELETED(&s);
+
+ if (should_have_subvol) {
+ id = le32_to_cpu(s.subvol);
+ ret = bch2_subvolume_get(trans, id, 0, false, &subvol);
+ if (bch2_err_matches(ret, ENOENT))
+ bch_err(c, "snapshot points to nonexistent subvolume:\n %s",
+ (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
+ if (ret)
+ goto err;
+
+ if (BCH_SNAPSHOT_SUBVOL(&s) != (le32_to_cpu(subvol.snapshot) == k.k->p.offset)) {
+ bch_err(c, "snapshot node %llu has wrong BCH_SNAPSHOT_SUBVOL",
+ k.k->p.offset);
+ ret = -EINVAL;
+ goto err;
+ }
+ } else {
+ if (fsck_err_on(s.subvol,
+ c, snapshot_should_not_have_subvol,
+ "snapshot should not point to subvol:\n %s",
+ (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
+ u = bch2_bkey_make_mut_typed(trans, iter, &k, 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ goto err;
+
+ u->v.subvol = 0;
+ s = u->v;
+ }
+ }
+
+ ret = snapshot_tree_ptr_good(trans, k.k->p.offset, le32_to_cpu(s.tree));
+ if (ret < 0)
+ goto err;
+
+ if (fsck_err_on(!ret, c, snapshot_to_bad_snapshot_tree,
+ "snapshot points to missing/incorrect tree:\n %s",
+ (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
+ ret = snapshot_tree_ptr_repair(trans, iter, k, &s);
+ if (ret)
+ goto err;
+ }
+ ret = 0;
+
+ real_depth = bch2_snapshot_depth(c, parent_id);
+
+ if (le32_to_cpu(s.depth) != real_depth &&
+ (c->sb.version_upgrade_complete < bcachefs_metadata_version_snapshot_skiplists ||
+ fsck_err(c, snapshot_bad_depth,
+ "snapshot with incorrect depth field, should be %u:\n %s",
+ real_depth, (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))) {
+ u = bch2_bkey_make_mut_typed(trans, iter, &k, 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ goto err;
+
+ u->v.depth = cpu_to_le32(real_depth);
+ s = u->v;
+ }
+
+ ret = snapshot_skiplist_good(trans, k.k->p.offset, s);
+ if (ret < 0)
+ goto err;
+
+ if (!ret &&
+ (c->sb.version_upgrade_complete < bcachefs_metadata_version_snapshot_skiplists ||
+ fsck_err(c, snapshot_bad_skiplist,
+ "snapshot with bad skiplist field:\n %s",
+ (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))) {
+ u = bch2_bkey_make_mut_typed(trans, iter, &k, 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(u);
+ if (ret)
+ goto err;
+
+ for (i = 0; i < ARRAY_SIZE(u->v.skip); i++)
+ u->v.skip[i] = cpu_to_le32(bch2_snapshot_skiplist_get(c, parent_id));
+
+ bubble_sort(u->v.skip, ARRAY_SIZE(u->v.skip), cmp_le32);
+ s = u->v;
+ }
+ ret = 0;
+err:
+fsck_err:
+ printbuf_exit(&buf);
+ return ret;
+}
+
+int bch2_check_snapshots(struct bch_fs *c)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret;
+
+ /*
+ * We iterate backwards as checking/fixing the depth field requires that
+ * the parent's depth already be correct:
+ */
+ ret = bch2_trans_run(c,
+ for_each_btree_key_reverse_commit(trans, iter,
+ BTREE_ID_snapshots, POS_MAX,
+ BTREE_ITER_PREFETCH, k,
+ NULL, NULL, BCH_TRANS_COMMIT_lazy_rw|BCH_TRANS_COMMIT_no_enospc,
+ check_snapshot(trans, &iter, k)));
+ if (ret)
+ bch_err_fn(c, ret);
+ return ret;
+}
+
+/*
+ * Mark a snapshot as deleted, for future cleanup:
+ */
+int bch2_snapshot_node_set_deleted(struct btree_trans *trans, u32 id)
+{
+ struct btree_iter iter;
+ struct bkey_i_snapshot *s;
+ int ret = 0;
+
+ s = bch2_bkey_get_mut_typed(trans, &iter,
+ BTREE_ID_snapshots, POS(0, id),
+ 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(s);
+ if (unlikely(ret)) {
+ bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT),
+ trans->c, "missing snapshot %u", id);
+ return ret;
+ }
+
+ /* already deleted? */
+ if (BCH_SNAPSHOT_DELETED(&s->v))
+ goto err;
+
+ SET_BCH_SNAPSHOT_DELETED(&s->v, true);
+ SET_BCH_SNAPSHOT_SUBVOL(&s->v, false);
+ s->v.subvol = 0;
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static inline void normalize_snapshot_child_pointers(struct bch_snapshot *s)
+{
+ if (le32_to_cpu(s->children[0]) < le32_to_cpu(s->children[1]))
+ swap(s->children[0], s->children[1]);
+}
+
+static int bch2_snapshot_node_delete(struct btree_trans *trans, u32 id)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter, p_iter = (struct btree_iter) { NULL };
+ struct btree_iter c_iter = (struct btree_iter) { NULL };
+ struct btree_iter tree_iter = (struct btree_iter) { NULL };
+ struct bkey_s_c_snapshot s;
+ u32 parent_id, child_id;
+ unsigned i;
+ int ret = 0;
+
+ s = bch2_bkey_get_iter_typed(trans, &iter, BTREE_ID_snapshots, POS(0, id),
+ BTREE_ITER_INTENT, snapshot);
+ ret = bkey_err(s);
+ bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
+ "missing snapshot %u", id);
+
+ if (ret)
+ goto err;
+
+ BUG_ON(s.v->children[1]);
+
+ parent_id = le32_to_cpu(s.v->parent);
+ child_id = le32_to_cpu(s.v->children[0]);
+
+ if (parent_id) {
+ struct bkey_i_snapshot *parent;
+
+ parent = bch2_bkey_get_mut_typed(trans, &p_iter,
+ BTREE_ID_snapshots, POS(0, parent_id),
+ 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(parent);
+ bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
+ "missing snapshot %u", parent_id);
+ if (unlikely(ret))
+ goto err;
+
+ /* find entry in parent->children for node being deleted */
+ for (i = 0; i < 2; i++)
+ if (le32_to_cpu(parent->v.children[i]) == id)
+ break;
+
+ if (bch2_fs_inconsistent_on(i == 2, c,
+ "snapshot %u missing child pointer to %u",
+ parent_id, id))
+ goto err;
+
+ parent->v.children[i] = cpu_to_le32(child_id);
+
+ normalize_snapshot_child_pointers(&parent->v);
+ }
+
+ if (child_id) {
+ struct bkey_i_snapshot *child;
+
+ child = bch2_bkey_get_mut_typed(trans, &c_iter,
+ BTREE_ID_snapshots, POS(0, child_id),
+ 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(child);
+ bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
+ "missing snapshot %u", child_id);
+ if (unlikely(ret))
+ goto err;
+
+ child->v.parent = cpu_to_le32(parent_id);
+
+ if (!child->v.parent) {
+ child->v.skip[0] = 0;
+ child->v.skip[1] = 0;
+ child->v.skip[2] = 0;
+ }
+ }
+
+ if (!parent_id) {
+ /*
+ * We're deleting the root of a snapshot tree: update the
+ * snapshot_tree entry to point to the new root, or delete it if
+ * this is the last snapshot ID in this tree:
+ */
+ struct bkey_i_snapshot_tree *s_t;
+
+ BUG_ON(s.v->children[1]);
+
+ s_t = bch2_bkey_get_mut_typed(trans, &tree_iter,
+ BTREE_ID_snapshot_trees, POS(0, le32_to_cpu(s.v->tree)),
+ 0, snapshot_tree);
+ ret = PTR_ERR_OR_ZERO(s_t);
+ if (ret)
+ goto err;
+
+ if (s.v->children[0]) {
+ s_t->v.root_snapshot = s.v->children[0];
+ } else {
+ s_t->k.type = KEY_TYPE_deleted;
+ set_bkey_val_u64s(&s_t->k, 0);
+ }
+ }
+
+ ret = bch2_btree_delete_at(trans, &iter, 0);
+err:
+ bch2_trans_iter_exit(trans, &tree_iter);
+ bch2_trans_iter_exit(trans, &p_iter);
+ bch2_trans_iter_exit(trans, &c_iter);
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static int create_snapids(struct btree_trans *trans, u32 parent, u32 tree,
+ u32 *new_snapids,
+ u32 *snapshot_subvols,
+ unsigned nr_snapids)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_i_snapshot *n;
+ struct bkey_s_c k;
+ unsigned i, j;
+ u32 depth = bch2_snapshot_depth(c, parent);
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_snapshots,
+ POS_MIN, BTREE_ITER_INTENT);
+ k = bch2_btree_iter_peek(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ for (i = 0; i < nr_snapids; i++) {
+ k = bch2_btree_iter_prev_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ if (!k.k || !k.k->p.offset) {
+ ret = -BCH_ERR_ENOSPC_snapshot_create;
+ goto err;
+ }
+
+ n = bch2_bkey_alloc(trans, &iter, 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(n);
+ if (ret)
+ goto err;
+
+ n->v.flags = 0;
+ n->v.parent = cpu_to_le32(parent);
+ n->v.subvol = cpu_to_le32(snapshot_subvols[i]);
+ n->v.tree = cpu_to_le32(tree);
+ n->v.depth = cpu_to_le32(depth);
+
+ for (j = 0; j < ARRAY_SIZE(n->v.skip); j++)
+ n->v.skip[j] = cpu_to_le32(bch2_snapshot_skiplist_get(c, parent));
+
+ bubble_sort(n->v.skip, ARRAY_SIZE(n->v.skip), cmp_le32);
+ SET_BCH_SNAPSHOT_SUBVOL(&n->v, true);
+
+ ret = bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0,
+ bkey_s_c_null, bkey_i_to_s_c(&n->k_i), 0);
+ if (ret)
+ goto err;
+
+ new_snapids[i] = iter.pos.offset;
+
+ mutex_lock(&c->snapshot_table_lock);
+ snapshot_t_mut(c, new_snapids[i])->equiv = new_snapids[i];
+ mutex_unlock(&c->snapshot_table_lock);
+ }
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+/*
+ * Create new snapshot IDs as children of an existing snapshot ID:
+ */
+static int bch2_snapshot_node_create_children(struct btree_trans *trans, u32 parent,
+ u32 *new_snapids,
+ u32 *snapshot_subvols,
+ unsigned nr_snapids)
+{
+ struct btree_iter iter;
+ struct bkey_i_snapshot *n_parent;
+ int ret = 0;
+
+ n_parent = bch2_bkey_get_mut_typed(trans, &iter,
+ BTREE_ID_snapshots, POS(0, parent),
+ 0, snapshot);
+ ret = PTR_ERR_OR_ZERO(n_parent);
+ if (unlikely(ret)) {
+ if (bch2_err_matches(ret, ENOENT))
+ bch_err(trans->c, "snapshot %u not found", parent);
+ return ret;
+ }
+
+ if (n_parent->v.children[0] || n_parent->v.children[1]) {
+ bch_err(trans->c, "Trying to add child snapshot nodes to parent that already has children");
+ ret = -EINVAL;
+ goto err;
+ }
+
+ ret = create_snapids(trans, parent, le32_to_cpu(n_parent->v.tree),
+ new_snapids, snapshot_subvols, nr_snapids);
+ if (ret)
+ goto err;
+
+ n_parent->v.children[0] = cpu_to_le32(new_snapids[0]);
+ n_parent->v.children[1] = cpu_to_le32(new_snapids[1]);
+ n_parent->v.subvol = 0;
+ SET_BCH_SNAPSHOT_SUBVOL(&n_parent->v, false);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+/*
+ * Create a snapshot node that is the root of a new tree:
+ */
+static int bch2_snapshot_node_create_tree(struct btree_trans *trans,
+ u32 *new_snapids,
+ u32 *snapshot_subvols,
+ unsigned nr_snapids)
+{
+ struct bkey_i_snapshot_tree *n_tree;
+ int ret;
+
+ n_tree = __bch2_snapshot_tree_create(trans);
+ ret = PTR_ERR_OR_ZERO(n_tree) ?:
+ create_snapids(trans, 0, n_tree->k.p.offset,
+ new_snapids, snapshot_subvols, nr_snapids);
+ if (ret)
+ return ret;
+
+ n_tree->v.master_subvol = cpu_to_le32(snapshot_subvols[0]);
+ n_tree->v.root_snapshot = cpu_to_le32(new_snapids[0]);
+ return 0;
+}
+
+int bch2_snapshot_node_create(struct btree_trans *trans, u32 parent,
+ u32 *new_snapids,
+ u32 *snapshot_subvols,
+ unsigned nr_snapids)
+{
+ BUG_ON((parent == 0) != (nr_snapids == 1));
+ BUG_ON((parent != 0) != (nr_snapids == 2));
+
+ return parent
+ ? bch2_snapshot_node_create_children(trans, parent,
+ new_snapids, snapshot_subvols, nr_snapids)
+ : bch2_snapshot_node_create_tree(trans,
+ new_snapids, snapshot_subvols, nr_snapids);
+
+}
+
+/*
+ * If we have an unlinked inode in an internal snapshot node, and the inode
+ * really has been deleted in all child snapshots, how does this get cleaned up?
+ *
+ * first there is the problem of how keys that have been overwritten in all
+ * child snapshots get deleted (unimplemented?), but inodes may perhaps be
+ * special?
+ *
+ * also: unlinked inode in internal snapshot appears to not be getting deleted
+ * correctly if inode doesn't exist in leaf snapshots
+ *
+ * solution:
+ *
+ * for a key in an interior snapshot node that needs work to be done that
+ * requires it to be mutated: iterate over all descendent leaf nodes and copy
+ * that key to snapshot leaf nodes, where we can mutate it
+ */
+
+static int snapshot_delete_key(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k,
+ snapshot_id_list *deleted,
+ snapshot_id_list *equiv_seen,
+ struct bpos *last_pos)
+{
+ struct bch_fs *c = trans->c;
+ u32 equiv = bch2_snapshot_equiv(c, k.k->p.snapshot);
+
+ if (!bkey_eq(k.k->p, *last_pos))
+ equiv_seen->nr = 0;
+ *last_pos = k.k->p;
+
+ if (snapshot_list_has_id(deleted, k.k->p.snapshot) ||
+ snapshot_list_has_id(equiv_seen, equiv)) {
+ return bch2_btree_delete_at(trans, iter,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+ } else {
+ return snapshot_list_add(c, equiv_seen, equiv);
+ }
+}
+
+static int move_key_to_correct_snapshot(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ u32 equiv = bch2_snapshot_equiv(c, k.k->p.snapshot);
+
+ /*
+ * When we have a linear chain of snapshot nodes, we consider
+ * those to form an equivalence class: we're going to collapse
+ * them all down to a single node, and keep the leaf-most node -
+ * which has the same id as the equivalence class id.
+ *
+ * If there are multiple keys in different snapshots at the same
+ * position, we're only going to keep the one in the newest
+ * snapshot - the rest have been overwritten and are redundant,
+ * and for the key we're going to keep we need to move it to the
+ * equivalance class ID if it's not there already.
+ */
+ if (equiv != k.k->p.snapshot) {
+ struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
+ struct btree_iter new_iter;
+ int ret;
+
+ ret = PTR_ERR_OR_ZERO(new);
+ if (ret)
+ return ret;
+
+ new->k.p.snapshot = equiv;
+
+ bch2_trans_iter_init(trans, &new_iter, iter->btree_id, new->k.p,
+ BTREE_ITER_ALL_SNAPSHOTS|
+ BTREE_ITER_CACHED|
+ BTREE_ITER_INTENT);
+
+ ret = bch2_btree_iter_traverse(&new_iter) ?:
+ bch2_trans_update(trans, &new_iter, new,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
+ bch2_btree_delete_at(trans, iter,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+ bch2_trans_iter_exit(trans, &new_iter);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
+static int bch2_snapshot_needs_delete(struct btree_trans *trans, struct bkey_s_c k)
+{
+ struct bkey_s_c_snapshot snap;
+ u32 children[2];
+ int ret;
+
+ if (k.k->type != KEY_TYPE_snapshot)
+ return 0;
+
+ snap = bkey_s_c_to_snapshot(k);
+ if (BCH_SNAPSHOT_DELETED(snap.v) ||
+ BCH_SNAPSHOT_SUBVOL(snap.v))
+ return 0;
+
+ children[0] = le32_to_cpu(snap.v->children[0]);
+ children[1] = le32_to_cpu(snap.v->children[1]);
+
+ ret = bch2_snapshot_live(trans, children[0]) ?:
+ bch2_snapshot_live(trans, children[1]);
+ if (ret < 0)
+ return ret;
+ return !ret;
+}
+
+/*
+ * For a given snapshot, if it doesn't have a subvolume that points to it, and
+ * it doesn't have child snapshot nodes - it's now redundant and we can mark it
+ * as deleted.
+ */
+static int bch2_delete_redundant_snapshot(struct btree_trans *trans, struct bkey_s_c k)
+{
+ int ret = bch2_snapshot_needs_delete(trans, k);
+
+ return ret <= 0
+ ? ret
+ : bch2_snapshot_node_set_deleted(trans, k.k->p.offset);
+}
+
+static inline u32 bch2_snapshot_nth_parent_skip(struct bch_fs *c, u32 id, u32 n,
+ snapshot_id_list *skip)
+{
+ rcu_read_lock();
+ while (snapshot_list_has_id(skip, id))
+ id = __bch2_snapshot_parent(c, id);
+
+ while (n--) {
+ do {
+ id = __bch2_snapshot_parent(c, id);
+ } while (snapshot_list_has_id(skip, id));
+ }
+ rcu_read_unlock();
+
+ return id;
+}
+
+static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
+ struct btree_iter *iter, struct bkey_s_c k,
+ snapshot_id_list *deleted)
+{
+ struct bch_fs *c = trans->c;
+ u32 nr_deleted_ancestors = 0;
+ struct bkey_i_snapshot *s;
+ u32 *i;
+ int ret;
+
+ if (k.k->type != KEY_TYPE_snapshot)
+ return 0;
+
+ if (snapshot_list_has_id(deleted, k.k->p.offset))
+ return 0;
+
+ s = bch2_bkey_make_mut_noupdate_typed(trans, k, snapshot);
+ ret = PTR_ERR_OR_ZERO(s);
+ if (ret)
+ return ret;
+
+ darray_for_each(*deleted, i)
+ nr_deleted_ancestors += bch2_snapshot_is_ancestor(c, s->k.p.offset, *i);
+
+ if (!nr_deleted_ancestors)
+ return 0;
+
+ le32_add_cpu(&s->v.depth, -nr_deleted_ancestors);
+
+ if (!s->v.depth) {
+ s->v.skip[0] = 0;
+ s->v.skip[1] = 0;
+ s->v.skip[2] = 0;
+ } else {
+ u32 depth = le32_to_cpu(s->v.depth);
+ u32 parent = bch2_snapshot_parent(c, s->k.p.offset);
+
+ for (unsigned j = 0; j < ARRAY_SIZE(s->v.skip); j++) {
+ u32 id = le32_to_cpu(s->v.skip[j]);
+
+ if (snapshot_list_has_id(deleted, id)) {
+ id = bch2_snapshot_nth_parent_skip(c,
+ parent,
+ depth > 1
+ ? get_random_u32_below(depth - 1)
+ : 0,
+ deleted);
+ s->v.skip[j] = cpu_to_le32(id);
+ }
+ }
+
+ bubble_sort(s->v.skip, ARRAY_SIZE(s->v.skip), cmp_le32);
+ }
+
+ return bch2_trans_update(trans, iter, &s->k_i, 0);
+}
+
+int bch2_delete_dead_snapshots(struct bch_fs *c)
+{
+ struct btree_trans *trans;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ struct bkey_s_c_snapshot snap;
+ snapshot_id_list deleted = { 0 };
+ snapshot_id_list deleted_interior = { 0 };
+ u32 *i, id;
+ int ret = 0;
+
+ if (!test_and_clear_bit(BCH_FS_NEED_DELETE_DEAD_SNAPSHOTS, &c->flags))
+ return 0;
+
+ if (!test_bit(BCH_FS_STARTED, &c->flags)) {
+ ret = bch2_fs_read_write_early(c);
+ if (ret) {
+ bch_err_msg(c, ret, "deleting dead snapshots: error going rw");
+ return ret;
+ }
+ }
+
+ trans = bch2_trans_get(c);
+
+ /*
+ * For every snapshot node: If we have no live children and it's not
+ * pointed to by a subvolume, delete it:
+ */
+ ret = for_each_btree_key_commit(trans, iter, BTREE_ID_snapshots,
+ POS_MIN, 0, k,
+ NULL, NULL, 0,
+ bch2_delete_redundant_snapshot(trans, k));
+ if (ret) {
+ bch_err_msg(c, ret, "deleting redundant snapshots");
+ goto err;
+ }
+
+ ret = for_each_btree_key2(trans, iter, BTREE_ID_snapshots,
+ POS_MIN, 0, k,
+ bch2_snapshot_set_equiv(trans, k));
+ if (ret) {
+ bch_err_msg(c, ret, "in bch2_snapshots_set_equiv");
+ goto err;
+ }
+
+ for_each_btree_key(trans, iter, BTREE_ID_snapshots,
+ POS_MIN, 0, k, ret) {
+ if (k.k->type != KEY_TYPE_snapshot)
+ continue;
+
+ snap = bkey_s_c_to_snapshot(k);
+ if (BCH_SNAPSHOT_DELETED(snap.v)) {
+ ret = snapshot_list_add(c, &deleted, k.k->p.offset);
+ if (ret)
+ break;
+ }
+ }
+ bch2_trans_iter_exit(trans, &iter);
+
+ if (ret) {
+ bch_err_msg(c, ret, "walking snapshots");
+ goto err;
+ }
+
+ for (id = 0; id < BTREE_ID_NR; id++) {
+ struct bpos last_pos = POS_MIN;
+ snapshot_id_list equiv_seen = { 0 };
+ struct disk_reservation res = { 0 };
+
+ if (!btree_type_has_snapshots(id))
+ continue;
+
+ /*
+ * deleted inodes btree is maintained by a trigger on the inodes
+ * btree - no work for us to do here, and it's not safe to scan
+ * it because we'll see out of date keys due to the btree write
+ * buffer:
+ */
+ if (id == BTREE_ID_deleted_inodes)
+ continue;
+
+ ret = for_each_btree_key_commit(trans, iter,
+ id, POS_MIN,
+ BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
+ &res, NULL, BCH_TRANS_COMMIT_no_enospc,
+ snapshot_delete_key(trans, &iter, k, &deleted, &equiv_seen, &last_pos)) ?:
+ for_each_btree_key_commit(trans, iter,
+ id, POS_MIN,
+ BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
+ &res, NULL, BCH_TRANS_COMMIT_no_enospc,
+ move_key_to_correct_snapshot(trans, &iter, k));
+
+ bch2_disk_reservation_put(c, &res);
+ darray_exit(&equiv_seen);
+
+ if (ret) {
+ bch_err_msg(c, ret, "deleting keys from dying snapshots");
+ goto err;
+ }
+ }
+
+ bch2_trans_unlock(trans);
+ down_write(&c->snapshot_create_lock);
+
+ for_each_btree_key(trans, iter, BTREE_ID_snapshots,
+ POS_MIN, 0, k, ret) {
+ u32 snapshot = k.k->p.offset;
+ u32 equiv = bch2_snapshot_equiv(c, snapshot);
+
+ if (equiv != snapshot)
+ snapshot_list_add(c, &deleted_interior, snapshot);
+ }
+ bch2_trans_iter_exit(trans, &iter);
+
+ if (ret)
+ goto err_create_lock;
+
+ /*
+ * Fixing children of deleted snapshots can't be done completely
+ * atomically, if we crash between here and when we delete the interior
+ * nodes some depth fields will be off:
+ */
+ ret = for_each_btree_key_commit(trans, iter, BTREE_ID_snapshots, POS_MIN,
+ BTREE_ITER_INTENT, k,
+ NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+ bch2_fix_child_of_deleted_snapshot(trans, &iter, k, &deleted_interior));
+ if (ret)
+ goto err_create_lock;
+
+ darray_for_each(deleted, i) {
+ ret = commit_do(trans, NULL, NULL, 0,
+ bch2_snapshot_node_delete(trans, *i));
+ if (ret) {
+ bch_err_msg(c, ret, "deleting snapshot %u", *i);
+ goto err_create_lock;
+ }
+ }
+
+ darray_for_each(deleted_interior, i) {
+ ret = commit_do(trans, NULL, NULL, 0,
+ bch2_snapshot_node_delete(trans, *i));
+ if (ret) {
+ bch_err_msg(c, ret, "deleting snapshot %u", *i);
+ goto err_create_lock;
+ }
+ }
+err_create_lock:
+ up_write(&c->snapshot_create_lock);
+err:
+ darray_exit(&deleted_interior);
+ darray_exit(&deleted);
+ bch2_trans_put(trans);
+ if (ret)
+ bch_err_fn(c, ret);
+ return ret;
+}
+
+void bch2_delete_dead_snapshots_work(struct work_struct *work)
+{
+ struct bch_fs *c = container_of(work, struct bch_fs, snapshot_delete_work);
+
+ bch2_delete_dead_snapshots(c);
+ bch2_write_ref_put(c, BCH_WRITE_REF_delete_dead_snapshots);
+}
+
+void bch2_delete_dead_snapshots_async(struct bch_fs *c)
+{
+ if (bch2_write_ref_tryget(c, BCH_WRITE_REF_delete_dead_snapshots) &&
+ !queue_work(c->write_ref_wq, &c->snapshot_delete_work))
+ bch2_write_ref_put(c, BCH_WRITE_REF_delete_dead_snapshots);
+}
+
+int __bch2_key_has_snapshot_overwrites(struct btree_trans *trans,
+ enum btree_id id,
+ struct bpos pos)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, id, pos,
+ BTREE_ITER_NOT_EXTENTS|
+ BTREE_ITER_ALL_SNAPSHOTS);
+ while (1) {
+ k = bch2_btree_iter_prev(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ break;
+
+ if (!k.k)
+ break;
+
+ if (!bkey_eq(pos, k.k->p))
+ break;
+
+ if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
+ ret = 1;
+ break;
+ }
+ }
+ bch2_trans_iter_exit(trans, &iter);
+
+ return ret;
+}
+
+static u32 bch2_snapshot_smallest_child(struct bch_fs *c, u32 id)
+{
+ const struct snapshot_t *s = snapshot_t(c, id);
+
+ return s->children[1] ?: s->children[0];
+}
+
+static u32 bch2_snapshot_smallest_descendent(struct bch_fs *c, u32 id)
+{
+ u32 child;
+
+ while ((child = bch2_snapshot_smallest_child(c, id)))
+ id = child;
+ return id;
+}
+
+static int bch2_propagate_key_to_snapshot_leaf(struct btree_trans *trans,
+ enum btree_id btree,
+ struct bkey_s_c interior_k,
+ u32 leaf_id, struct bpos *new_min_pos)
+{
+ struct btree_iter iter;
+ struct bpos pos = interior_k.k->p;
+ struct bkey_s_c k;
+ struct bkey_i *new;
+ int ret;
+
+ pos.snapshot = leaf_id;
+
+ bch2_trans_iter_init(trans, &iter, btree, pos, BTREE_ITER_INTENT);
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto out;
+
+ /* key already overwritten in this snapshot? */
+ if (k.k->p.snapshot != interior_k.k->p.snapshot)
+ goto out;
+
+ if (bpos_eq(*new_min_pos, POS_MIN)) {
+ *new_min_pos = k.k->p;
+ new_min_pos->snapshot = leaf_id;
+ }
+
+ new = bch2_bkey_make_mut_noupdate(trans, interior_k);
+ ret = PTR_ERR_OR_ZERO(new);
+ if (ret)
+ goto out;
+
+ new->k.p.snapshot = leaf_id;
+ ret = bch2_trans_update(trans, &iter, new, 0);
+out:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+int bch2_propagate_key_to_snapshot_leaves(struct btree_trans *trans,
+ enum btree_id btree,
+ struct bkey_s_c k,
+ struct bpos *new_min_pos)
+{
+ struct bch_fs *c = trans->c;
+ struct bkey_buf sk;
+ u32 restart_count = trans->restart_count;
+ int ret = 0;
+
+ bch2_bkey_buf_init(&sk);
+ bch2_bkey_buf_reassemble(&sk, c, k);
+ k = bkey_i_to_s_c(sk.k);
+
+ *new_min_pos = POS_MIN;
+
+ for (u32 id = bch2_snapshot_smallest_descendent(c, k.k->p.snapshot);
+ id < k.k->p.snapshot;
+ id++) {
+ if (!bch2_snapshot_is_ancestor(c, id, k.k->p.snapshot) ||
+ !bch2_snapshot_is_leaf(c, id))
+ continue;
+again:
+ ret = btree_trans_too_many_iters(trans) ?:
+ bch2_propagate_key_to_snapshot_leaf(trans, btree, k, id, new_min_pos) ?:
+ bch2_trans_commit(trans, NULL, NULL, 0);
+ if (ret && bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
+ bch2_trans_begin(trans);
+ goto again;
+ }
+
+ if (ret)
+ break;
+ }
+
+ bch2_bkey_buf_exit(&sk, c);
+
+ return ret ?: trans_was_restarted(trans, restart_count);
+}
+
+static int bch2_check_snapshot_needs_deletion(struct btree_trans *trans, struct bkey_s_c k)
+{
+ struct bch_fs *c = trans->c;
+ struct bkey_s_c_snapshot snap;
+ int ret = 0;
+
+ if (k.k->type != KEY_TYPE_snapshot)
+ return 0;
+
+ snap = bkey_s_c_to_snapshot(k);
+ if (BCH_SNAPSHOT_DELETED(snap.v) ||
+ bch2_snapshot_equiv(c, k.k->p.offset) != k.k->p.offset ||
+ (ret = bch2_snapshot_needs_delete(trans, k)) > 0) {
+ set_bit(BCH_FS_NEED_DELETE_DEAD_SNAPSHOTS, &c->flags);
+ return 0;
+ }
+
+ return ret;
+}
+
+int bch2_snapshots_read(struct bch_fs *c)
+{
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret = 0;
+
+ ret = bch2_trans_run(c,
+ for_each_btree_key2(trans, iter, BTREE_ID_snapshots,
+ POS_MIN, 0, k,
+ bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0, bkey_s_c_null, k, 0) ?:
+ bch2_snapshot_set_equiv(trans, k) ?:
+ bch2_check_snapshot_needs_deletion(trans, k)) ?:
+ for_each_btree_key2(trans, iter, BTREE_ID_snapshots,
+ POS_MIN, 0, k,
+ (set_is_ancestor_bitmap(c, k.k->p.offset), 0)));
+ if (ret)
+ bch_err_fn(c, ret);
+ return ret;
+}
+
+void bch2_fs_snapshots_exit(struct bch_fs *c)
+{
+ kfree(rcu_dereference_protected(c->snapshots, true));
+}
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SNAPSHOT_H
+#define _BCACHEFS_SNAPSHOT_H
+
+enum bkey_invalid_flags;
+
+void bch2_snapshot_tree_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+int bch2_snapshot_tree_invalid(struct bch_fs *, struct bkey_s_c,
+ enum bkey_invalid_flags, struct printbuf *);
+
+#define bch2_bkey_ops_snapshot_tree ((struct bkey_ops) { \
+ .key_invalid = bch2_snapshot_tree_invalid, \
+ .val_to_text = bch2_snapshot_tree_to_text, \
+ .min_val_size = 8, \
+})
+
+struct bkey_i_snapshot_tree *__bch2_snapshot_tree_create(struct btree_trans *);
+
+int bch2_snapshot_tree_lookup(struct btree_trans *, u32, struct bch_snapshot_tree *);
+
+void bch2_snapshot_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+int bch2_snapshot_invalid(struct bch_fs *, struct bkey_s_c,
+ enum bkey_invalid_flags, struct printbuf *);
+int bch2_mark_snapshot(struct btree_trans *, enum btree_id, unsigned,
+ struct bkey_s_c, struct bkey_s_c, unsigned);
+
+#define bch2_bkey_ops_snapshot ((struct bkey_ops) { \
+ .key_invalid = bch2_snapshot_invalid, \
+ .val_to_text = bch2_snapshot_to_text, \
+ .atomic_trigger = bch2_mark_snapshot, \
+ .min_val_size = 24, \
+})
+
+static inline struct snapshot_t *__snapshot_t(struct snapshot_table *t, u32 id)
+{
+ return &t->s[U32_MAX - id];
+}
+
+static inline const struct snapshot_t *snapshot_t(struct bch_fs *c, u32 id)
+{
+ return __snapshot_t(rcu_dereference(c->snapshots), id);
+}
+
+static inline u32 bch2_snapshot_tree(struct bch_fs *c, u32 id)
+{
+ rcu_read_lock();
+ id = snapshot_t(c, id)->tree;
+ rcu_read_unlock();
+
+ return id;
+}
+
+static inline u32 __bch2_snapshot_parent_early(struct bch_fs *c, u32 id)
+{
+ return snapshot_t(c, id)->parent;
+}
+
+static inline u32 bch2_snapshot_parent_early(struct bch_fs *c, u32 id)
+{
+ rcu_read_lock();
+ id = __bch2_snapshot_parent_early(c, id);
+ rcu_read_unlock();
+
+ return id;
+}
+
+static inline u32 __bch2_snapshot_parent(struct bch_fs *c, u32 id)
+{
+#ifdef CONFIG_BCACHEFS_DEBUG
+ u32 parent = snapshot_t(c, id)->parent;
+
+ if (parent &&
+ snapshot_t(c, id)->depth != snapshot_t(c, parent)->depth + 1)
+ panic("id %u depth=%u parent %u depth=%u\n",
+ id, snapshot_t(c, id)->depth,
+ parent, snapshot_t(c, parent)->depth);
+
+ return parent;
+#else
+ return snapshot_t(c, id)->parent;
+#endif
+}
+
+static inline u32 bch2_snapshot_parent(struct bch_fs *c, u32 id)
+{
+ rcu_read_lock();
+ id = __bch2_snapshot_parent(c, id);
+ rcu_read_unlock();
+
+ return id;
+}
+
+static inline u32 bch2_snapshot_nth_parent(struct bch_fs *c, u32 id, u32 n)
+{
+ rcu_read_lock();
+ while (n--)
+ id = __bch2_snapshot_parent(c, id);
+ rcu_read_unlock();
+
+ return id;
+}
+
+u32 bch2_snapshot_skiplist_get(struct bch_fs *, u32);
+
+static inline u32 bch2_snapshot_root(struct bch_fs *c, u32 id)
+{
+ u32 parent;
+
+ rcu_read_lock();
+ while ((parent = __bch2_snapshot_parent(c, id)))
+ id = parent;
+ rcu_read_unlock();
+
+ return id;
+}
+
+static inline u32 __bch2_snapshot_equiv(struct bch_fs *c, u32 id)
+{
+ return snapshot_t(c, id)->equiv;
+}
+
+static inline u32 bch2_snapshot_equiv(struct bch_fs *c, u32 id)
+{
+ rcu_read_lock();
+ id = __bch2_snapshot_equiv(c, id);
+ rcu_read_unlock();
+
+ return id;
+}
+
+static inline bool bch2_snapshot_is_equiv(struct bch_fs *c, u32 id)
+{
+ return id == bch2_snapshot_equiv(c, id);
+}
+
+static inline bool bch2_snapshot_is_internal_node(struct bch_fs *c, u32 id)
+{
+ const struct snapshot_t *s;
+ bool ret;
+
+ rcu_read_lock();
+ s = snapshot_t(c, id);
+ ret = s->children[0];
+ rcu_read_unlock();
+
+ return ret;
+}
+
+static inline u32 bch2_snapshot_is_leaf(struct bch_fs *c, u32 id)
+{
+ return !bch2_snapshot_is_internal_node(c, id);
+}
+
+static inline u32 bch2_snapshot_sibling(struct bch_fs *c, u32 id)
+{
+ const struct snapshot_t *s;
+ u32 parent = __bch2_snapshot_parent(c, id);
+
+ if (!parent)
+ return 0;
+
+ s = snapshot_t(c, __bch2_snapshot_parent(c, id));
+ if (id == s->children[0])
+ return s->children[1];
+ if (id == s->children[1])
+ return s->children[0];
+ return 0;
+}
+
+static inline u32 bch2_snapshot_depth(struct bch_fs *c, u32 parent)
+{
+ u32 depth;
+
+ rcu_read_lock();
+ depth = parent ? snapshot_t(c, parent)->depth + 1 : 0;
+ rcu_read_unlock();
+
+ return depth;
+}
+
+bool __bch2_snapshot_is_ancestor(struct bch_fs *, u32, u32);
+
+static inline bool bch2_snapshot_is_ancestor(struct bch_fs *c, u32 id, u32 ancestor)
+{
+ return id == ancestor
+ ? true
+ : __bch2_snapshot_is_ancestor(c, id, ancestor);
+}
+
+static inline bool bch2_snapshot_has_children(struct bch_fs *c, u32 id)
+{
+ const struct snapshot_t *t;
+ bool ret;
+
+ rcu_read_lock();
+ t = snapshot_t(c, id);
+ ret = (t->children[0]|t->children[1]) != 0;
+ rcu_read_unlock();
+
+ return ret;
+}
+
+static inline bool snapshot_list_has_id(snapshot_id_list *s, u32 id)
+{
+ u32 *i;
+
+ darray_for_each(*s, i)
+ if (*i == id)
+ return true;
+ return false;
+}
+
+static inline bool snapshot_list_has_ancestor(struct bch_fs *c, snapshot_id_list *s, u32 id)
+{
+ u32 *i;
+
+ darray_for_each(*s, i)
+ if (bch2_snapshot_is_ancestor(c, id, *i))
+ return true;
+ return false;
+}
+
+static inline int snapshot_list_add(struct bch_fs *c, snapshot_id_list *s, u32 id)
+{
+ int ret;
+
+ BUG_ON(snapshot_list_has_id(s, id));
+ ret = darray_push(s, id);
+ if (ret)
+ bch_err(c, "error reallocating snapshot_id_list (size %zu)", s->size);
+ return ret;
+}
+
+int bch2_snapshot_lookup(struct btree_trans *trans, u32 id,
+ struct bch_snapshot *s);
+int bch2_snapshot_get_subvol(struct btree_trans *, u32,
+ struct bch_subvolume *);
+
+/* only exported for tests: */
+int bch2_snapshot_node_create(struct btree_trans *, u32,
+ u32 *, u32 *, unsigned);
+
+int bch2_check_snapshot_trees(struct bch_fs *);
+int bch2_check_snapshots(struct bch_fs *);
+
+int bch2_snapshot_node_set_deleted(struct btree_trans *, u32);
+void bch2_delete_dead_snapshots_work(struct work_struct *);
+
+int __bch2_key_has_snapshot_overwrites(struct btree_trans *, enum btree_id, struct bpos);
+
+static inline int bch2_key_has_snapshot_overwrites(struct btree_trans *trans,
+ enum btree_id id,
+ struct bpos pos)
+{
+ if (!btree_type_has_snapshots(id) ||
+ bch2_snapshot_is_leaf(trans->c, pos.snapshot))
+ return 0;
+
+ return __bch2_key_has_snapshot_overwrites(trans, id, pos);
+}
+
+int bch2_propagate_key_to_snapshot_leaves(struct btree_trans *, enum btree_id,
+ struct bkey_s_c, struct bpos *);
+
+int bch2_snapshots_read(struct bch_fs *);
+void bch2_fs_snapshots_exit(struct bch_fs *);
+
+#endif /* _BCACHEFS_SNAPSHOT_H */
--- /dev/null
+#!/bin/bash
+
+set -o errexit
+
+version=$1
+
+git checkout v$version
+git clean -xfd
+
+(cd rust-src; cargo license) > COPYING.rust-dependencies
+
+git ls-files|
+ tar --create --file bcachefs-tools-$version.tar -T - \
+ --transform="s_^_bcachefs-tools-$version/_"
+
+tar --append --file bcachefs-tools-vendored-$version.tar \
+ --transform="s_^_bcachefs-tools-$version/_" \
+ COPYING.rust-dependencies
+
+zstd -z --ultra bcachefs-tools-$version.tar
+
+gpg --armor --detach-sign bcachefs-tools-$version.tar
+mv bcachefs-tools-$version.tar.asc bcachefs-tools-$version.tar.sign
+
+gpg --armor --sign bcachefs-tools-$version.tar
+
+scp bcachefs-tools-$version.tar.zst evilpiepirate.org:/var/www/htdocs/bcachefs-tools/
+scp bcachefs-tools-$version.tar.asc evilpiepirate.org:/var/www/htdocs/bcachefs-tools/
+scp bcachefs-tools-$version.tar.sign evilpiepirate.org:/var/www/htdocs/bcachefs-tools/
+
+cargo vendor --manifest-path rust-src/Cargo.toml
+
+mkdir .cargo
+cat > .cargo/config.toml <<-ZZ
+[source.crates-io]
+replace-with = "vendored-sources"
+
+[source."git+https://evilpiepirate.org/git/rust-bindgen.git"]
+git = "https://evilpiepirate.org/git/rust-bindgen.git"
+replace-with = "vendored-sources"
+
+[source.vendored-sources]
+directory = "vendor"
+ZZ
+
+cp bcachefs-tools-$version.tar bcachefs-tools-vendored-$version.tar
+tar --append --file bcachefs-tools-vendored-$version.tar \
+ --transform="s_^_bcachefs-tools-$version/_" \
+ .cargo vendor
+
+zstd -z --ultra bcachefs-tools-vendored-$version.tar
+
+gpg --armor --detach-sign bcachefs-tools-vendored-$version.tar
+mv bcachefs-tools-vendored-$version.tar.asc bcachefs-tools-vendored-$version.tar.sign
+
+gpg --armor --sign bcachefs-tools-vendored-$version.tar
+
+scp bcachefs-tools-vendored-$version.tar.zst evilpiepirate.org:/var/www/htdocs/bcachefs-tools/
+scp bcachefs-tools-vendored-$version.tar.asc evilpiepirate.org:/var/www/htdocs/bcachefs-tools/
+scp bcachefs-tools-vendored-$version.tar.sign evilpiepirate.org:/var/www/htdocs/bcachefs-tools/
--- /dev/null
+use crate::transform_c_args;
+use clap::{Command, CommandFactory, Parser};
+use clap_complete::{generate, Generator, Shell};
+use std::ffi::{c_char, c_int};
+use std::io;
+
+/// Generate shell completions
+#[derive(clap::Parser, Debug)]
+pub struct Cli {
+ shell: Shell,
+}
+
+fn print_completions<G: Generator>(gen: G, cmd: &mut Command) {
+ generate(gen, cmd, cmd.get_name().to_string(), &mut io::stdout());
+}
+
+#[no_mangle]
+#[allow(clippy::not_unsafe_ptr_arg_deref)]
+pub extern "C" fn cmd_completions(argc: c_int, argv: *const *const c_char) -> c_int {
+ transform_c_args!(argv, argc, argv);
+ let cli = Cli::parse_from(argv);
+ print_completions(cli.shell, &mut super::Cli::command());
+ 0
+}