#include "btree_gc.h"
#include "btree_io.h"
#include "btree_iter.h"
+#include "btree_key_cache.h"
#include "btree_locking.h"
+#include "btree_write_buffer.h"
#include "buckets.h"
#include "debug.h"
+#include "errcode.h"
#include "error.h"
-#include "extents.h"
+#include "extent_update.h"
#include "journal.h"
#include "journal_reclaim.h"
#include "keylist.h"
+#include "recovery.h"
+#include "subvolume.h"
#include "replicas.h"
+#include <linux/prefetch.h>
#include <linux/sort.h>
#include <trace/events/bcachefs.h>
-inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
- struct btree_iter *iter)
+/*
+ * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
+ * different snapshot:
+ */
+struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
{
- bch2_btree_node_lock_write(b, iter);
+ struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
- if (btree_node_just_written(b) &&
- bch2_btree_post_write_cleanup(c, b))
- bch2_btree_iter_reinit_node(iter, b);
+ if (k.k && bpos_eq(path->pos, k.k->p))
+ return k;
- /*
- * 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(c, b, iter);
+ bkey_init(u);
+ u->p = path->pos;
+ return (struct bkey_s_c) { u, NULL };
}
-static void btree_trans_lock_write(struct bch_fs *c, struct btree_trans *trans)
+static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
{
- 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 int __must_check
+bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
+ struct bkey_i *, enum btree_update_flags);
- trans_for_each_update_leaf(trans, i)
- bch2_btree_node_lock_for_insert(c, i->iter->l[0].b, i->iter);
+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 void btree_trans_unlock_write(struct btree_trans *trans)
+static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
{
- 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;
+}
- trans_for_each_update_leaf(trans, i)
- bch2_btree_node_unlock_write(i->iter->l[0].b, i->iter);
+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;
}
-static inline int btree_trans_cmp(struct btree_insert_entry l,
- struct btree_insert_entry r)
+inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree *b)
{
- return cmp_int(l.deferred, r.deferred) ?:
- btree_iter_cmp(l.iter, r.iter);
+ 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);
}
/* 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_iter *iter,
+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)
{
- const struct bkey_format *f = &b->format;
struct bkey_packed *k;
- unsigned clobber_u64s;
+ 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(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
- bkey_cmp(insert->k.p, b->data->max_key) > 0);
+ 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));
k = bch2_btree_node_iter_peek_all(node_iter, b);
- if (k && !bkey_cmp_packed(b, k, &insert->k)) {
- BUG_ON(bkey_whiteout(k));
-
- if (!bkey_written(b, k) &&
- bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k) &&
- !bkey_whiteout(&insert->k)) {
- k->type = insert->k.type;
- memcpy_u64s(bkeyp_val(f, k), &insert->v,
- bkey_val_u64s(&insert->k));
- return true;
- }
+ if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
+ k = NULL;
- insert->k.needs_whiteout = k->needs_whiteout;
+ /* @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;
-
- /*
- * If we're deleting, and the key we're deleting doesn't
- * need a whiteout (it wasn't overwriting a key that had
- * been written to disk) - just delete it:
- */
- if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
- bch2_bset_delete(b, k, clobber_u64s);
- bch2_btree_node_iter_fix(iter, b, node_iter,
- k, clobber_u64s, 0);
- bch2_btree_iter_verify(iter, b);
- return true;
- }
-
- goto overwrite;
+ 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;
- bch2_btree_node_iter_fix(iter, b, node_iter, k,
- k->u64s, k->u64s);
- bch2_btree_iter_verify(iter, b);
- if (bkey_whiteout(&insert->k)) {
- reserve_whiteout(b, k);
- return true;
+ 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 {
- k->needs_whiteout = false;
+ bch2_btree_path_fix_key_modified(trans, b, k);
}
- } else {
- /*
- * Deleting, but the key to delete wasn't found - nothing to do:
- */
- if (bkey_whiteout(&insert->k))
- return false;
-
- insert->k.needs_whiteout = false;
}
k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
- clobber_u64s = 0;
overwrite:
bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
- if (k->u64s != clobber_u64s || bkey_whiteout(&insert->k))
- bch2_btree_node_iter_fix(iter, b, node_iter, k,
- clobber_u64s, k->u64s);
- bch2_btree_iter_verify(iter, b);
+ 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 void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
+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;
+ unsigned long old, new, v;
+ unsigned idx = w - b->writes;
- btree_node_lock_type(c, b, SIX_LOCK_read);
- bch2_btree_node_write_cond(c, b,
- (btree_current_write(b) == w && w->journal.seq == seq));
- six_unlock_read(&b->lock);
-}
+ bch2_trans_init(&trans, c, 0, 0);
-static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
-{
- return __btree_node_flush(j, pin, 0, seq);
-}
+ btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
+ v = READ_ONCE(b->flags);
-static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
-{
- return __btree_node_flush(j, pin, 1, seq);
-}
+ do {
+ old = new = v;
-static inline void __btree_journal_key(struct btree_trans *trans,
- enum btree_id btree_id,
- struct bkey_i *insert)
-{
- struct journal *j = &trans->c->journal;
- u64 seq = trans->journal_res.seq;
- bool needs_whiteout = insert->k.needs_whiteout;
+ if (!(old & (1 << BTREE_NODE_dirty)) ||
+ !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
+ w->journal.seq != seq)
+ break;
- /* ick */
- insert->k.needs_whiteout = false;
- bch2_journal_add_keys(j, &trans->journal_res,
- btree_id, insert);
- insert->k.needs_whiteout = needs_whiteout;
+ new &= ~BTREE_WRITE_TYPE_MASK;
+ new |= BTREE_WRITE_journal_reclaim;
+ new |= 1 << BTREE_NODE_need_write;
+ } while ((v = cmpxchg(&b->flags, old, new)) != old);
- bch2_journal_set_has_inode(j, &trans->journal_res,
- insert->k.p.inode);
+ btree_node_write_if_need(c, b, SIX_LOCK_read);
+ six_unlock_read(&b->c.lock);
- if (trans->journal_seq)
- *trans->journal_seq = seq;
+ bch2_trans_exit(&trans);
+ return 0;
}
-void bch2_btree_journal_key(struct btree_trans *trans,
- struct btree_iter *iter,
- struct bkey_i *insert)
+static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
{
- struct bch_fs *c = trans->c;
- struct journal *j = &c->journal;
- struct btree *b = iter->l[0].b;
- struct btree_write *w = btree_current_write(b);
-
- EBUG_ON(iter->level || b->level);
- EBUG_ON(trans->journal_res.ref !=
- !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
-
- if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
- __btree_journal_key(trans, iter->btree_id, insert);
- btree_bset_last(b)->journal_seq =
- cpu_to_le64(trans->journal_res.seq);
- }
-
- if (unlikely(!journal_pin_active(&w->journal))) {
- u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
- ? trans->journal_res.seq
- : j->replay_journal_seq;
-
- bch2_journal_pin_add(j, seq, &w->journal,
- btree_node_write_idx(b) == 0
- ? btree_node_flush0
- : btree_node_flush1);
- }
-
- if (unlikely(!btree_node_dirty(b)))
- set_btree_node_dirty(b);
+ return __btree_node_flush(j, pin, 0, seq);
}
-static void bch2_insert_fixup_key(struct btree_trans *trans,
- struct btree_insert_entry *insert)
+static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
{
- struct btree_iter *iter = insert->iter;
- struct btree_iter_level *l = &iter->l[0];
+ return __btree_node_flush(j, pin, 1, seq);
+}
- EBUG_ON(iter->level);
- EBUG_ON(insert->k->k.u64s >
- bch_btree_keys_u64s_remaining(trans->c, l->b));
+inline void bch2_btree_add_journal_pin(struct bch_fs *c,
+ struct btree *b, u64 seq)
+{
+ struct btree_write *w = btree_current_write(b);
- if (bch2_btree_bset_insert_key(iter, l->b, &l->iter,
- insert->k))
- bch2_btree_journal_key(trans, iter, insert->k);
+ bch2_journal_pin_add(&c->journal, seq, &w->journal,
+ btree_node_write_idx(b) == 0
+ ? btree_node_flush0
+ : btree_node_flush1);
}
/**
* btree_insert_key - insert a key one key into a leaf node
*/
-static void btree_insert_key_leaf(struct btree_trans *trans,
- struct btree_insert_entry *insert)
+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_iter *iter = insert->iter;
- struct btree *b = iter->l[0].b;
- int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
+ 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 (!btree_node_is_extents(b))
- bch2_insert_fixup_key(trans, insert);
- else
- bch2_insert_fixup_extent(trans, insert);
+ 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)))
+ set_btree_node_dirty_acct(c, b);
live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
- u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_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 (u64s_added > live_u64s_added &&
bch2_maybe_compact_whiteouts(c, b))
- bch2_btree_iter_reinit_node(iter, b);
-
- trace_btree_insert_key(c, b, insert->k);
+ bch2_trans_node_reinit_iter(trans, b);
}
-/* Deferred btree updates: */
-
-static void deferred_update_flush(struct journal *j,
- struct journal_entry_pin *pin,
- u64 seq)
+static void btree_insert_key_leaf(struct btree_trans *trans,
+ struct btree_insert_entry *insert)
{
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
- struct deferred_update *d =
- container_of(pin, struct deferred_update, journal);
- struct journal_preres res = { 0 };
- u64 tmp[32];
- struct bkey_i *k = (void *) tmp;
- int ret;
-
- if (d->allocated_u64s > ARRAY_SIZE(tmp)) {
- k = kmalloc(d->allocated_u64s * sizeof(u64), GFP_NOFS);
+ bch2_btree_insert_key_leaf(trans, insert->path, insert->k, trans->journal_res.seq);
+}
- BUG_ON(!k); /* XXX */
- }
+/* Cached btree updates: */
- spin_lock(&d->lock);
- if (d->dirty) {
- BUG_ON(jset_u64s(d->k.k.u64s) > d->res.u64s);
+/* Normal update interface: */
- swap(res, d->res);
+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 &&
+ !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
+ test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
+ i->k->k.p.snapshot &&
+ bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
+}
- BUG_ON(d->k.k.u64s > d->allocated_u64s);
+static noinline int
+bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
+ unsigned long trace_ip)
+{
+ struct bch_fs *c = trans->c;
+ int ret;
- bkey_copy(k, &d->k);
- d->dirty = false;
- spin_unlock(&d->lock);
+ bch2_trans_unlock(trans);
- ret = bch2_btree_insert(c, d->btree_id, k, NULL, NULL,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|
- BTREE_INSERT_JOURNAL_RESERVED);
- bch2_fs_fatal_err_on(ret && !bch2_journal_error(j),
- c, "error flushing deferred btree update: %i", ret);
+ ret = bch2_journal_preres_get(&c->journal,
+ &trans->journal_preres,
+ trans->journal_preres_u64s,
+ (flags & JOURNAL_WATERMARK_MASK));
+ if (ret)
+ return ret;
- spin_lock(&d->lock);
+ ret = bch2_trans_relock(trans);
+ if (ret) {
+ trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
+ return ret;
}
- if (!d->dirty)
- bch2_journal_pin_drop(j, &d->journal);
- spin_unlock(&d->lock);
+ return 0;
+}
- bch2_journal_preres_put(j, &res);
- if (k != (void *) tmp)
- kfree(k);
+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);
}
-static void btree_insert_key_deferred(struct btree_trans *trans,
- struct btree_insert_entry *insert)
+#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 deferred_update *d = insert->d;
- int difference;
-
- BUG_ON(trans->flags & BTREE_INSERT_JOURNAL_REPLAY);
- BUG_ON(insert->k->u64s > d->allocated_u64s);
-
- __btree_journal_key(trans, d->btree_id, insert->k);
-
- spin_lock(&d->lock);
- BUG_ON(jset_u64s(insert->k->u64s) >
- trans->journal_preres.u64s);
-
- difference = jset_u64s(insert->k->u64s) - d->res.u64s;
- if (difference > 0) {
- trans->journal_preres.u64s -= difference;
- d->res.u64s += difference;
- }
-
- bkey_copy(&d->k, insert->k);
- d->dirty = true;
-
- bch2_journal_pin_update(j, trans->journal_res.seq, &d->journal,
- deferred_update_flush);
- spin_unlock(&d->lock);
+ 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));
}
-void bch2_deferred_update_free(struct bch_fs *c,
- struct deferred_update *d)
+static inline int btree_key_can_insert(struct btree_trans *trans,
+ struct btree *b, unsigned u64s)
{
- deferred_update_flush(&c->journal, &d->journal, 0);
+ struct bch_fs *c = trans->c;
- BUG_ON(journal_pin_active(&d->journal));
+ if (!bch2_btree_node_insert_fits(c, b, u64s))
+ return -BCH_ERR_btree_insert_btree_node_full;
- bch2_journal_pin_flush(&c->journal, &d->journal);
- kfree(d);
+ return 0;
}
-struct deferred_update *
-bch2_deferred_update_alloc(struct bch_fs *c,
- enum btree_id btree_id,
- unsigned u64s)
+static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
+ struct btree_path *path, unsigned u64s)
{
- struct deferred_update *d;
-
- BUG_ON(u64s > U8_MAX);
-
- d = kmalloc(offsetof(struct deferred_update, k) +
- u64s * sizeof(u64), GFP_NOFS);
- BUG_ON(!d);
+ 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;
- memset(d, 0, offsetof(struct deferred_update, k));
+ EBUG_ON(path->level);
- spin_lock_init(&d->lock);
- d->allocated_u64s = u64s;
- d->btree_id = btree_id;
+ if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+ bch2_btree_key_cache_must_wait(c) &&
+ !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
+ return -BCH_ERR_btree_insert_need_journal_reclaim;
- return d;
-}
+ /*
+ * bch2_varint_decode can read past the end of the buffer by at most 7
+ * bytes (it won't be used):
+ */
+ u64s += 1;
-/* Normal update interface: */
+ if (u64s <= ck->u64s)
+ return 0;
-static inline void btree_insert_entry_checks(struct btree_trans *trans,
- struct btree_insert_entry *i)
-{
- struct bch_fs *c = trans->c;
- enum btree_id btree_id = !i->deferred
- ? i->iter->btree_id
- : i->d->btree_id;
-
- if (!i->deferred) {
- BUG_ON(i->iter->level);
- BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
- EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
- bkey_cmp(i->k->k.p, i->iter->l[0].b->key.k.p) > 0);
- EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
- !(trans->flags & BTREE_INSERT_ATOMIC));
+ new_u64s = roundup_pow_of_two(u64s);
+ new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
+ if (!new_k) {
+ bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
+ bch2_btree_ids[path->btree_id], new_u64s);
+ return -ENOMEM;
}
- BUG_ON(debug_check_bkeys(c) &&
- !bkey_deleted(&i->k->k) &&
- bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), btree_id));
+ 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;
}
-static int bch2_trans_journal_preres_get(struct btree_trans *trans)
+/* Triggers: */
+
+static int run_one_mem_trigger(struct btree_trans *trans,
+ struct btree_insert_entry *i,
+ unsigned flags)
{
- struct bch_fs *c = trans->c;
- struct btree_insert_entry *i;
- unsigned u64s = 0;
+ struct bkey_s_c old = { &i->old_k, i->old_v };
+ struct bkey_i *new = i->k;
int ret;
- trans_for_each_update(trans, i)
- if (i->deferred)
- u64s += jset_u64s(i->k->k.u64s);
+ verify_update_old_key(trans, i);
- if (!u64s)
+ if (unlikely(flags & BTREE_TRIGGER_NORUN))
return 0;
- ret = bch2_journal_preres_get(&c->journal,
- &trans->journal_preres, u64s,
- JOURNAL_RES_GET_NONBLOCK);
- if (ret != -EAGAIN)
- return ret;
+ if (!btree_node_type_needs_gc(i->btree_id))
+ return 0;
- bch2_trans_unlock(trans);
+ if (bch2_bkey_ops[old.k->type].atomic_trigger ==
+ bch2_bkey_ops[i->k->k.type].atomic_trigger &&
+ ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
+ 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 };
- ret = bch2_journal_preres_get(&c->journal,
- &trans->journal_preres, u64s, 0);
- if (ret)
- return ret;
+ _deleted.p = i->path->pos;
- if (!bch2_trans_relock(trans)) {
- trace_trans_restart_journal_preres_get(trans->ip);
- return -EINTR;
+ 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 0;
+ return ret;
}
-static int bch2_trans_journal_res_get(struct btree_trans *trans,
- unsigned flags)
+static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
+ bool overwrite)
{
- struct bch_fs *c = trans->c;
- int ret;
+ /*
+ * 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 };
- if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
- flags |= JOURNAL_RES_GET_RESERVED;
+ verify_update_old_key(trans, i);
- ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
- trans->journal_u64s, flags);
+ if ((i->flags & BTREE_TRIGGER_NORUN) ||
+ !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
+ return 0;
- return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
+ if (!i->insert_trigger_run &&
+ !i->overwrite_trigger_run &&
+ bch2_bkey_ops[old.k->type].trans_trigger ==
+ bch2_bkey_ops[i->k->k.type].trans_trigger &&
+ ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
+ 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 enum btree_insert_ret
-btree_key_can_insert(struct btree_trans *trans,
- struct btree_insert_entry *insert,
- unsigned *u64s)
+static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
+ struct btree_insert_entry *btree_id_start)
{
- struct bch_fs *c = trans->c;
- struct btree *b = insert->iter->l[0].b;
- static enum btree_insert_ret ret;
-
- if (unlikely(btree_node_fake(b)))
- return BTREE_INSERT_BTREE_NODE_FULL;
+ struct btree_insert_entry *i;
+ bool trans_trigger_run;
+ int ret, overwrite;
- ret = !btree_node_is_extents(b)
- ? BTREE_INSERT_OK
- : bch2_extent_can_insert(trans, insert, u64s);
- if (ret)
- return ret;
+ for (overwrite = 1; overwrite >= 0; --overwrite) {
- if (*u64s > bch_btree_keys_u64s_remaining(c, b))
- return BTREE_INSERT_BTREE_NODE_FULL;
+ /*
+ * 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 BTREE_INSERT_OK;
+ return 0;
}
-static int btree_trans_check_can_insert(struct btree_trans *trans,
- struct btree_insert_entry **stopped_at)
+static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
{
- struct btree_insert_entry *i;
- unsigned u64s = 0;
- int ret;
+ struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
+ unsigned btree_id = 0;
+ int ret = 0;
- trans_for_each_update_iter(trans, i) {
- /* Multiple inserts might go to same leaf: */
- if (!same_leaf_as_prev(trans, i))
- u64s = 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;
- u64s += i->k->k.u64s;
- ret = btree_key_can_insert(trans, i, &u64s);
- if (ret) {
- *stopped_at = i;
+ 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 inline void do_btree_insert_one(struct btree_trans *trans,
- struct btree_insert_entry *insert)
+static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
{
- if (likely(!insert->deferred))
- btree_insert_key_leaf(trans, insert);
- else
- btree_insert_key_deferred(trans, insert);
-}
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ int ret = 0;
-static inline bool update_triggers_transactional(struct btree_trans *trans,
- struct btree_insert_entry *i)
-{
- return likely(!(trans->flags & BTREE_INSERT_MARK_INMEM)) &&
- (i->iter->btree_id == BTREE_ID_EXTENTS ||
- i->iter->btree_id == BTREE_ID_INODES ||
- i->iter->btree_id == BTREE_ID_REFLINK);
-}
+ 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);
-static inline bool update_has_triggers(struct btree_trans *trans,
- struct btree_insert_entry *i)
-{
- return likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
- !i->deferred &&
- btree_node_type_needs_gc(i->iter->btree_id);
+ 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;
}
-/*
- * Get journal reservation, take write locks, and attempt to do btree update(s):
- */
-static inline int do_btree_insert_at(struct btree_trans *trans,
- struct btree_insert_entry **stopped_at)
+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 bch_fs_usage *fs_usage = NULL;
struct btree_insert_entry *i;
- bool saw_non_marked;
- unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
- ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
- : 0;
+ struct btree_write_buffered_key *wb;
+ struct btree_trans_commit_hook *h;
+ unsigned u64s = 0;
+ bool marking = false;
int ret;
- trans_for_each_update_iter(trans, i)
- BUG_ON(i->iter->uptodate >= BTREE_ITER_NEED_RELOCK);
-
- trans_for_each_update_iter(trans, i)
- i->marked = false;
-
- do {
- saw_non_marked = false;
-
- trans_for_each_update_iter(trans, i) {
- if (i->marked)
- continue;
-
- saw_non_marked = true;
- i->marked = true;
-
- if (update_has_triggers(trans, i) &&
- update_triggers_transactional(trans, i)) {
- ret = bch2_trans_mark_update(trans, i->iter, i->k);
- if (ret == -EINTR)
- trace_trans_restart_mark(trans->ip);
- if (ret)
- goto out_clear_replicas;
- }
- }
- } while (saw_non_marked);
-
- trans_for_each_update(trans, i)
- btree_insert_entry_checks(trans, i);
- bch2_btree_trans_verify_locks(trans);
-
- btree_trans_lock_write(c, trans);
-
if (race_fault()) {
- ret = -EINTR;
- trace_trans_restart_fault_inject(trans->ip);
- goto out;
+ trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
+ return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
}
/*
* held, otherwise another thread could write the node changing the
* amount of space available:
*/
- ret = btree_trans_check_can_insert(trans, stopped_at);
- if (ret)
- goto out;
- trans_for_each_update_iter(trans, i) {
- if (i->deferred ||
- !btree_node_type_needs_gc(i->iter->btree_id))
- continue;
+ prefetch(&trans->c->journal.flags);
- if (!fs_usage) {
- percpu_down_read(&c->mark_lock);
- fs_usage = bch2_fs_usage_scratch_get(c);
- }
+ h = trans->hooks;
+ while (h) {
+ ret = h->fn(trans, h);
+ if (ret)
+ return ret;
+ h = h->next;
+ }
- if (!bch2_bkey_replicas_marked_locked(c,
- bkey_i_to_s_c(i->k), true)) {
- ret = BTREE_INSERT_NEED_MARK_REPLICAS;
- goto out;
+ 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;
}
+
+ if (btree_node_type_needs_gc(i->bkey_type))
+ marking = true;
}
+ 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(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
- trans->journal_u64s = 0;
-
- trans_for_each_update(trans, i)
- trans->journal_u64s += jset_u64s(i->k->k.u64s);
-
- ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_NONBLOCK);
+ if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+ ret = bch2_trans_journal_res_get(trans,
+ (flags & JOURNAL_WATERMARK_MASK)|
+ JOURNAL_RES_GET_NONBLOCK);
if (ret)
- goto out;
+ return ret;
+
+ if (unlikely(trans->journal_transaction_names))
+ journal_transaction_name(trans);
+ } else {
+ trans->journal_res.seq = c->journal.replay_journal_seq;
}
- if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
- if (journal_seq_verify(c))
+ /*
+ * Not allowed to fail after we've gotten our journal reservation - we
+ * have to use it:
+ */
+
+ if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
+ !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
+ if (bch2_journal_seq_verify)
trans_for_each_update(trans, i)
i->k->k.version.lo = trans->journal_res.seq;
- else if (inject_invalid_keys(c))
+ else if (bch2_inject_invalid_keys)
trans_for_each_update(trans, i)
i->k->k.version = MAX_VERSION;
}
- trans_for_each_update_iter(trans, i)
- if (update_has_triggers(trans, i) &&
- !update_triggers_transactional(trans, i))
- bch2_mark_update(trans, i, fs_usage, mark_flags);
-
- if (fs_usage && trans->fs_usage_deltas)
- bch2_replicas_delta_list_apply(c, fs_usage,
- trans->fs_usage_deltas);
+ if (trans->fs_usage_deltas &&
+ bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
+ return -BCH_ERR_btree_insert_need_mark_replicas;
- if (fs_usage)
- bch2_trans_fs_usage_apply(trans, fs_usage);
+ if (trans->nr_wb_updates) {
+ EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
- if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
- unlikely(c->gc_pos.phase))
- trans_for_each_update_iter(trans, i)
- if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
- bch2_mark_update(trans, i, NULL,
- mark_flags|
- BCH_BUCKET_MARK_GC);
+ ret = bch2_btree_insert_keys_write_buffer(trans);
+ if (ret)
+ goto revert_fs_usage;
+ }
trans_for_each_update(trans, i)
- do_btree_insert_one(trans, i);
-out:
- BUG_ON(ret &&
- (trans->flags & BTREE_INSERT_JOURNAL_RESERVED) &&
- trans->journal_res.ref);
+ 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;
+ }
- btree_trans_unlock_write(trans);
+ 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);
- if (fs_usage) {
- bch2_fs_usage_scratch_put(c, fs_usage);
- percpu_up_read(&c->mark_lock);
+ trans->journal_res.offset += trans->extra_journal_entries.nr;
+ trans->journal_res.u64s -= trans->extra_journal_entries.nr;
}
- bch2_journal_res_put(&c->journal, &trans->journal_res);
-out_clear_replicas:
- if (trans->fs_usage_deltas) {
- memset(&trans->fs_usage_deltas->fs_usage, 0,
- sizeof(trans->fs_usage_deltas->fs_usage));
- trans->fs_usage_deltas->used = 0;
+ if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+ 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(&entry->start[0],
+ (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(&entry->start[0], 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(&entry->start[0], &wb->k);
+ }
+
+ if (trans->journal_seq)
+ *trans->journal_seq = trans->journal_res.seq;
+ }
+
+ trans_for_each_update(trans, i) {
+ i->k->k.needs_whiteout = false;
+
+ if (!i->cached)
+ btree_insert_key_leaf(trans, i);
+ else if (!i->key_cache_already_flushed)
+ bch2_btree_insert_key_cached(trans, flags, i->path, i->k);
+ 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
-int bch2_trans_commit_error(struct btree_trans *trans,
- struct btree_insert_entry *i,
- int ret)
+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 trans_lock_write(struct btree_trans *trans)
+{
+ struct btree_insert_entry *i;
+
+ 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);
+ }
+
+ return 0;
+}
+
+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);
+}
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
+ struct btree_insert_entry *i,
+ struct printbuf *err)
+{
+ struct bch_fs *c = trans->c;
+ int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
+
+ 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, rw, err);
+ bch2_print_string_as_lines(KERN_ERR, err->buf);
+
+ bch2_inconsistent_error(c);
+ bch2_dump_trans_updates(trans);
+ printbuf_exit(err);
+
+ return -EINVAL;
+}
+#endif
+
+/*
+ * 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;
- unsigned flags = trans->flags;
- struct btree_insert_entry *src, *dst;
+ struct btree_insert_entry *i;
+ int ret, u64s_delta = 0;
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+ struct printbuf buf = PRINTBUF;
+
+ trans_for_each_update(trans, i) {
+ int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
+
+ if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
+ i->bkey_type, rw, &buf)))
+ return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
+ btree_insert_entry_checks(trans, i);
+ }
+ printbuf_exit(&buf);
+#endif
- src = dst = trans->updates;
+ 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;
+ }
- while (src < trans->updates + trans->nr_updates) {
- if (!src->triggered) {
- *dst = *src;
- dst++;
+ u64s_delta = 0;
}
- src++;
}
- trans->nr_updates = dst - trans->updates;
+ ret = bch2_journal_preres_get(&c->journal,
+ &trans->journal_preres, trans->journal_preres_u64s,
+ (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
+ if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
+ ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
+ if (unlikely(ret))
+ return ret;
+
+ ret = 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);
+
+ 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);
+
+ if (!ret && trans->journal_pin)
+ bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
+ trans->journal_pin, NULL);
/*
- * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
- * update; if we haven't done anything yet it doesn't apply
+ * Drop journal reservation after dropping write locks, since dropping
+ * the journal reservation may kick off a journal write:
*/
- flags &= ~BTREE_INSERT_NOUNLOCK;
+ bch2_journal_res_put(&c->journal, &trans->journal_res);
- switch (ret) {
- case BTREE_INSERT_BTREE_NODE_FULL:
- ret = bch2_btree_split_leaf(c, i->iter, flags);
+ if (unlikely(ret))
+ return ret;
- /*
- * if the split succeeded without dropping locks the insert will
- * still be atomic (in the BTREE_INSERT_ATOMIC sense, what the
- * caller peeked() and is overwriting won't have changed)
- */
-#if 0
- /*
- * XXX:
- * split -> btree node merging (of parent node) might still drop
- * locks when we're not passing it BTREE_INSERT_NOUNLOCK
- *
- * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
- * will inhibit merging - but we don't have a reliable way yet
- * (do we?) of checking if we dropped locks in this path
- */
- if (!ret)
- goto retry;
-#endif
+ bch2_trans_downgrade(trans);
- /*
- * don't care if we got ENOSPC because we told split it
- * couldn't block:
- */
- if (!ret ||
- ret == -EINTR ||
- (flags & BTREE_INSERT_NOUNLOCK)) {
- trace_trans_restart_btree_node_split(trans->ip);
- ret = -EINTR;
- }
+ return 0;
+}
+
+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 BTREE_INSERT_ENOSPC:
- ret = -ENOSPC;
+ case -BCH_ERR_btree_insert_need_mark_replicas:
+ bch2_trans_unlock(trans);
+
+ ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
+ if (ret)
+ break;
+
+ ret = bch2_trans_relock(trans);
+ if (ret)
+ trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
break;
- case BTREE_INSERT_NEED_MARK_REPLICAS:
+ case -BCH_ERR_journal_res_get_blocked:
bch2_trans_unlock(trans);
- trans_for_each_update_iter(trans, i) {
- ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
- if (ret)
- return ret;
+ if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
+ !(flags & JOURNAL_WATERMARK_reserved)) {
+ ret = -BCH_ERR_journal_reclaim_would_deadlock;
+ break;
}
- if (bch2_trans_relock(trans))
- return 0;
+ ret = bch2_trans_journal_res_get(trans,
+ (flags & JOURNAL_WATERMARK_MASK)|
+ JOURNAL_RES_GET_CHECK);
+ if (ret)
+ break;
- trace_trans_restart_mark_replicas(trans->ip);
- ret = -EINTR;
+ ret = bch2_trans_relock(trans);
+ if (ret)
+ trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
break;
- case BTREE_INSERT_NEED_JOURNAL_RES:
+ case -BCH_ERR_btree_insert_need_journal_reclaim:
bch2_trans_unlock(trans);
- ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
+ 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);
if (ret)
- return ret;
+ trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
+ break;
+ case -BCH_ERR_btree_insert_need_flush_buffer: {
+ struct btree_write_buffer *wb = &c->btree_write_buffer;
- if (bch2_trans_relock(trans))
- return 0;
+ ret = 0;
+
+ if (wb->state.nr > wb->size * 3 / 4) {
+ bch2_trans_reset_updates(trans);
+ bch2_trans_unlock(trans);
- trace_trans_restart_journal_res_get(trans->ip);
- ret = -EINTR;
+ mutex_lock(&wb->flush_lock);
+
+ if (wb->state.nr > wb->size * 3 / 4)
+ ret = __bch2_btree_write_buffer_flush(trans,
+ flags|BTREE_INSERT_NOCHECK_RW, true);
+ else
+ mutex_unlock(&wb->flush_lock);
+
+ 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);
+ }
+ }
break;
+ }
default:
BUG_ON(ret >= 0);
break;
}
- if (ret == -EINTR) {
- int ret2 = bch2_btree_iter_traverse_all(trans);
-
- if (ret2) {
- trace_trans_restart_traverse(trans->ip);
- return ret2;
- }
-
- /*
- * BTREE_ITER_ATOMIC means we have to return -EINTR if we
- * dropped locks:
- */
- if (!(flags & BTREE_INSERT_ATOMIC))
- return 0;
+ BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
- trace_trans_restart_atomic(trans->ip);
- }
+ bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
+ !(flags & BTREE_INSERT_NOWAIT) &&
+ (flags & BTREE_INSERT_NOFAIL), c,
+ "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
return ret;
}
-/**
- * __bch_btree_insert_at - insert keys at given iterator positions
- *
- * This is main entry point for btree updates.
- *
- * Return values:
- * -EINTR: locking changed, this function should be called again. Only returned
- * if passed BTREE_INSERT_ATOMIC.
- * -EROFS: filesystem read only
- * -EIO: journal or btree node IO error
- */
-static int __bch2_trans_commit(struct btree_trans *trans,
- struct btree_insert_entry **stopped_at)
+static noinline int
+bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
{
struct bch_fs *c = trans->c;
- struct btree_insert_entry *i;
int ret;
- trans_for_each_update_iter(trans, i) {
- if (!bch2_btree_iter_upgrade(i->iter, 1)) {
- trace_trans_restart_upgrade(trans->ip);
- ret = -EINTR;
- goto err;
- }
-
- ret = btree_iter_err(i->iter);
- if (ret)
- goto err;
- }
+ if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
+ test_bit(BCH_FS_STARTED, &c->flags))
+ return -BCH_ERR_erofs_trans_commit;
- ret = do_btree_insert_at(trans, stopped_at);
- if (unlikely(ret))
- goto err;
+ bch2_trans_unlock(trans);
- if (trans->flags & BTREE_INSERT_NOUNLOCK)
- trans->nounlock = true;
+ ret = bch2_fs_read_write_early(c) ?:
+ bch2_trans_relock(trans);
+ if (ret)
+ return ret;
- trans_for_each_update_leaf(trans, i)
- bch2_foreground_maybe_merge(c, i->iter, 0, trans->flags);
+ bch2_write_ref_get(c, BCH_WRITE_REF_trans);
+ return 0;
+}
- trans->nounlock = false;
+/*
+ * 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_iter(trans, i)
- bch2_btree_iter_downgrade(i->iter);
-err:
- /* make sure we didn't drop or screw up locks: */
- bch2_btree_trans_verify_locks(trans);
+ 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,
- struct disk_reservation *disk_res,
- u64 *journal_seq,
- unsigned flags)
+int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
{
struct bch_fs *c = trans->c;
struct btree_insert_entry *i = NULL;
- unsigned orig_mem_top = trans->mem_top;
+ struct btree_write_buffered_key *wb;
+ unsigned u64s;
int ret = 0;
- if (!trans->nr_updates)
- goto out_noupdates;
-
- /* for the sake of sanity: */
- BUG_ON(trans->nr_updates > 1 && !(flags & BTREE_INSERT_ATOMIC));
+ if (!trans->nr_updates &&
+ !trans->nr_wb_updates &&
+ !trans->extra_journal_entries.nr)
+ goto out_reset;
if (flags & BTREE_INSERT_GC_LOCK_HELD)
lockdep_assert_held(&c->gc_lock);
- if (!trans->commit_start)
- trans->commit_start = local_clock();
+ ret = bch2_trans_commit_run_triggers(trans);
+ if (ret)
+ goto out_reset;
- memset(&trans->journal_res, 0, sizeof(trans->journal_res));
- memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
- trans->disk_res = disk_res;
- trans->journal_seq = journal_seq;
- trans->flags = flags;
+ if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
+ ret = do_bch2_trans_commit_to_journal_replay(trans);
+ goto out_reset;
+ }
- if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
- !percpu_ref_tryget(&c->writes))) {
- if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
- return -EROFS;
+ if (!(flags & BTREE_INSERT_NOCHECK_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_fs_read_write_early(c);
- if (ret)
- return ret;
+ ret = __bch2_btree_write_buffer_flush(trans,
+ flags|BTREE_INSERT_NOCHECK_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));
+
+ memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
+
+ trans->journal_u64s = trans->extra_journal_entries.nr;
+ trans->journal_preres_u64s = 0;
+
+ 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));
- percpu_ref_get(&c->writes);
+ if (i->key_cache_already_flushed)
+ continue;
+
+ /* we're going to journal the key being updated: */
+ u64s = jset_u64s(i->k->k.u64s);
+ if (i->cached &&
+ likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
+ trans->journal_preres_u64s += u64s;
+
+ if (i->flags & BTREE_UPDATE_NOJOURNAL)
+ continue;
+
+ trans->journal_u64s += 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 (!bch2_trans_relock(trans)) {
- ret = -EINTR;
+ if (trans->extra_journal_res) {
+ ret = bch2_disk_reservation_add(c, trans->disk_res,
+ trans->extra_journal_res,
+ (flags & BTREE_INSERT_NOFAIL)
+ ? BCH_DISK_RESERVATION_NOFAIL : 0);
+ if (ret)
goto err;
- }
}
retry:
- ret = bch2_trans_journal_preres_get(trans);
- if (ret)
- goto err;
+ bch2_trans_verify_not_in_restart(trans);
+ 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);
- ret = __bch2_trans_commit(trans, &i);
if (ret)
goto err;
+
+ trace_and_count(c, transaction_commit, trans, _RET_IP_);
out:
bch2_journal_preres_put(&c->journal, &trans->journal_preres);
- if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
- percpu_ref_put(&c->writes);
-out_noupdates:
- if (!ret && trans->commit_start) {
- bch2_time_stats_update(&c->times[BCH_TIME_btree_update],
- trans->commit_start);
- trans->commit_start = 0;
+ if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
+ bch2_write_ref_put(c, BCH_WRITE_REF_trans);
+out_reset:
+ bch2_trans_reset_updates(trans);
+
+ return ret;
+err:
+ ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
+ if (ret)
+ goto out;
+
+ goto retry;
+}
+
+static noinline int __check_pos_snapshot_overwritten(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 inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
+ enum btree_id id,
+ struct bpos pos)
+{
+ if (!btree_type_has_snapshots(id) ||
+ pos.snapshot == U32_MAX ||
+ !snapshot_t(trans->c, pos.snapshot)->children[0])
+ return 0;
+
+ return __check_pos_snapshot_overwritten(trans, id, pos);
+}
+
+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(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 = check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
+ check_pos_snapshot_overwritten(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 = check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
+ check_pos_snapshot_overwritten(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_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 bpos start = bkey_start_pos(&insert->k);
+ struct bkey_i *update;
+ struct bkey_s_c k;
+ enum btree_id btree_id = orig_iter->btree_id;
+ int ret = 0, compressed_sectors;
+
+ bch2_trans_iter_init(trans, &iter, btree_id, start,
+ 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;
}
- BUG_ON(!(trans->flags & BTREE_INSERT_ATOMIC) && ret == -EINTR);
+ while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
+ bool front_split = bkey_lt(bkey_start_pos(k.k), start);
+ bool back_split = bkey_gt(k.k->p, insert->k.p);
+
+ /*
+ * 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) && k.k->p.snapshot != insert->k.p.snapshot)) &&
+ (compressed_sectors = bch2_bkey_sectors_compressed(k)))
+ trans->extra_journal_res += compressed_sectors;
+
+ if (front_split) {
+ update = bch2_bkey_make_mut(trans, k);
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ bch2_cut_back(start, update);
+
+ ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+ if (ret)
+ goto err;
+ }
+
+ if (k.k->p.snapshot != insert->k.p.snapshot &&
+ (front_split || back_split)) {
+ update = bch2_bkey_make_mut(trans, k);
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ bch2_cut_front(start, update);
+ bch2_cut_back(insert->k.p, update);
+
+ ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+ if (ret)
+ goto err;
+ }
+
+ if (bkey_le(k.k->p, insert->k.p)) {
+ update = bch2_trans_kmalloc(trans, sizeof(*update));
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ bkey_init(&update->k);
+ update->k.p = k.k->p;
+ update->k.p.snapshot = insert->k.p.snapshot;
+
+ if (insert->k.p.snapshot != k.k->p.snapshot ||
+ (btree_type_has_snapshots(btree_id) &&
+ need_whiteout_for_snapshot(trans, btree_id, update->k.p)))
+ update->k.type = KEY_TYPE_whiteout;
+
+ ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+ if (ret)
+ goto err;
+ }
+
+ if (back_split) {
+ update = bch2_bkey_make_mut(trans, k);
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ bch2_cut_front(insert->k.p, update);
+
+ ret = bch2_trans_update_by_path(trans, iter.path, update,
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+ flags);
+ if (ret)
+ goto err;
+ 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 (!ret) {
- bch2_trans_unlink_iters(trans, ~trans->iters_touched|
- trans->iters_unlink_on_commit);
- trans->iters_touched = 0;
+ 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)) {
+ /*
+ * Rewinding iterators is expensive: get a new one and the one
+ * that points to the start of insert will be cloned from:
+ */
+ bch2_trans_iter_exit(trans, &iter);
+ bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
+ BTREE_ITER_NOT_EXTENTS|
+ BTREE_ITER_INTENT);
+ ret = bch2_btree_iter_traverse(&iter) ?:
+ bch2_trans_update(trans, &iter, insert, flags);
}
- trans->nr_updates = 0;
- trans->mem_top = 0;
+err:
+ bch2_trans_iter_exit(trans, &iter);
return ret;
+}
+
+static int __must_check
+bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
+ struct bkey_i *k, enum btree_update_flags flags,
+ unsigned long ip);
+
+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;
+ int ret;
+
+ i->key_cache_already_flushed = true;
+ i->flags |= BTREE_TRIGGER_NORUN;
+
+ 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 err;
+
+ btree_path_set_should_be_locked(btree_path);
+ ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
err:
- ret = bch2_trans_commit_error(trans, i, ret);
+ bch2_path_put(trans, btree_path, true);
+ return ret;
+}
+
+static int __must_check
+bch2_trans_update_by_path_trace(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;
- /* can't loop if it was passed in and we changed it: */
- if (unlikely(trans->flags & BTREE_INSERT_NO_CLEAR_REPLICAS) && !ret)
- ret = -EINTR;
+ 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 (!ret) {
- /* free memory used by triggers, they'll be reexecuted: */
- trans->mem_top = orig_mem_top;
- goto retry;
+ if (j_k) {
+ i->old_k = j_k->k;
+ i->old_v = &j_k->v;
+ }
+ }
}
- goto out;
+ __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) &&
+ !(flags & BTREE_UPDATE_NO_KEY_CACHE_COHERENCY))
+ return flush_new_cached_update(trans, path, i, flags, ip);
+
+ return 0;
}
-struct btree_insert_entry *bch2_trans_update(struct btree_trans *trans,
- struct btree_insert_entry entry)
+static inline 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)
{
- struct btree_insert_entry *i;
+ return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
+}
+
+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;
+ struct bkey_cached *ck;
+ int ret;
- BUG_ON(trans->nr_updates >= trans->nr_iters + 4);
+ if (iter->flags & BTREE_ITER_IS_EXTENTS)
+ return bch2_trans_update_extent(trans, iter, k, flags);
- for (i = trans->updates;
- i < trans->updates + trans->nr_updates;
- i++)
- if (btree_trans_cmp(entry, *i) < 0)
- break;
+ 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)) {
+ if (!iter->key_cache_path ||
+ !iter->key_cache_path->should_be_locked ||
+ !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
+ 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);
+ }
+
+ path = iter->key_cache_path;
+ }
+
+ return bch2_trans_update_by_path(trans, path, k, flags);
+}
+
+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;
+ }
- memmove(&i[1], &i[0],
- (void *) &trans->updates[trans->nr_updates] - (void *) i);
- trans->nr_updates++;
- *i = entry;
- return i;
+ 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;
+}
+
+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_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(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;
}
/**
* @hook: insert callback
*/
int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
- struct bkey_i *k,
- struct disk_reservation *disk_res,
- u64 *journal_seq, int flags)
+ struct bkey_i *k,
+ struct disk_reservation *disk_res,
+ u64 *journal_seq, int flags)
{
- struct btree_trans trans;
- struct btree_iter *iter;
- int ret;
-
- bch2_trans_init(&trans, c, 0, 0);
-retry:
- bch2_trans_begin(&trans);
+ return bch2_trans_do(c, disk_res, journal_seq, flags,
+ __bch2_btree_insert(&trans, id, k, 0));
+}
- iter = bch2_trans_get_iter(&trans, id, bkey_start_pos(&k->k),
- BTREE_ITER_INTENT);
+int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
+ unsigned len, unsigned update_flags)
+{
+ struct bkey_i *k;
- bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, k));
+ k = bch2_trans_kmalloc(trans, sizeof(*k));
+ if (IS_ERR(k))
+ return PTR_ERR(k);
- ret = bch2_trans_commit(&trans, disk_res, journal_seq, flags);
- if (ret == -EINTR)
- goto retry;
- bch2_trans_exit(&trans);
+ bkey_init(&k->k);
+ k->k.p = iter->pos;
+ bch2_key_resize(&k->k, len);
+ return bch2_trans_update(trans, iter, k, update_flags);
+}
- return ret;
+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_range(struct btree_trans *trans,
- struct btree_iter *iter,
- struct bpos end,
- u64 *journal_seq)
+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;
-retry:
- while ((k = bch2_btree_iter_peek(iter)).k &&
- !(ret = bkey_err(k)) &&
- bkey_cmp(iter->pos, end) < 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
* (bch2_btree_iter_peek() does guarantee that iter.pos >=
* bkey_start_pos(k.k)).
*/
- delete.k.p = iter->pos;
+ delete.k.p = iter.pos;
- if (iter->flags & BTREE_ITER_IS_EXTENTS) {
- unsigned max_sectors =
- KEY_SIZE_MAX & (~0 << trans->c->block_bits);
+ if (iter.flags & BTREE_ITER_IS_EXTENTS)
+ bch2_key_resize(&delete.k,
+ bpos_min(end, k.k->p).offset -
+ iter.pos.offset);
- /* create the biggest key we can */
- bch2_key_resize(&delete.k, max_sectors);
- bch2_cut_back(end, &delete.k);
-
- ret = bch2_extent_trim_atomic(&delete, iter);
- if (ret)
- break;
- }
+ ret = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
+ bch2_trans_commit(trans, &disk_res, journal_seq,
+ BTREE_INSERT_NOFAIL);
+ 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);
- bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &delete));
- ret = bch2_trans_commit(trans, NULL, journal_seq,
- BTREE_INSERT_ATOMIC|
- BTREE_INSERT_NOFAIL);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ ret = 0;
if (ret)
break;
-
- bch2_trans_cond_resched(trans);
- }
-
- if (ret == -EINTR) {
- ret = 0;
- goto retry;
}
+ bch2_trans_iter_exit(trans, &iter);
+ if (!ret && trans_was_restarted(trans, restart_count))
+ ret = -BCH_ERR_transaction_restart_nested;
return ret;
-
-}
-
-int bch2_btree_delete_at(struct btree_trans *trans,
- struct btree_iter *iter, unsigned flags)
-{
- struct bkey_i k;
-
- bkey_init(&k.k);
- k.k.p = iter->pos;
-
- bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &k));
- return bch2_trans_commit(trans, NULL, NULL,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|flags);
}
/*
*/
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)
{
- struct btree_trans trans;
- struct btree_iter *iter;
- int ret = 0;
+ 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;
+}
- /*
- * XXX: whether we need mem/more iters depends on whether this btree id
- * has triggers
- */
- bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
+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 ? -ENOMEM : 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;
+}
+
+int bch2_trans_log_msg(struct btree_trans *trans, const char *fmt, ...)
+{
+ va_list args;
+ int ret;
+
+ va_start(args, fmt);
+ ret = __bch2_trans_log_msg(&trans->extra_journal_entries, fmt, args);
+ va_end(args);
+
+ return ret;
+}
+
+int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
+{
+ va_list args;
+ int ret;
- iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
+ va_start(args, fmt);
- ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
- ret = bch2_trans_exit(&trans) ?: 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, BTREE_INSERT_LAZY_RW,
+ __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
+ }
+
+ va_end(args);
- BUG_ON(ret == -EINTR);
return ret;
+
}