-e1f6739c4a9fee1db7d94a5087a253041542cb62
+c9b4a210f946889f56654dda24dd8ced3b1aac24
get_alloc_field(a.v, &d, i));
}
-int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
+static int bch2_alloc_read_fn(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bkey_s_c k)
{
- struct btree_trans trans;
- struct btree_and_journal_iter iter;
- struct bkey_s_c k;
- struct bch_dev *ca;
- unsigned i;
- int ret = 0;
-
- bch2_trans_init(&trans, c, 0, 0);
-
- bch2_btree_and_journal_iter_init(&iter, &trans, journal_keys,
- BTREE_ID_ALLOC, POS_MIN);
-
- while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
+ if (!level)
bch2_mark_key(c, k, 0, 0, NULL, 0,
BTREE_TRIGGER_ALLOC_READ|
BTREE_TRIGGER_NOATOMIC);
- bch2_btree_and_journal_iter_advance(&iter);
- }
+ return 0;
+}
+
+int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
+{
+ struct bch_dev *ca;
+ unsigned i;
+ int ret = 0;
- ret = bch2_trans_exit(&trans) ?: ret;
+ ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_ALLOC,
+ NULL, bch2_alloc_read_fn);
if (ret) {
bch_err(c, "error reading alloc info: %i", ret);
return ret;
struct bkey_s_c k;
bool invalidating_cached_data;
size_t b;
- int ret;
+ int ret = 0;
BUG_ON(!ca->alloc_heap.used ||
!ca->alloc_heap.data[0].nr);
BUG_ON(!fifo_push(&ca->free_inc, b));
+ g = bucket(ca, b);
+ m = READ_ONCE(g->mark);
+
bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0);
spin_unlock(&c->freelist_lock);
percpu_up_read(&c->mark_lock);
+ invalidating_cached_data = m.cached_sectors != 0;
+ if (!invalidating_cached_data)
+ goto out;
+
+ /*
+ * If the read-only path is trying to shut down, we can't be generating
+ * new btree updates:
+ */
+ if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags)) {
+ ret = 1;
+ goto out;
+ }
+
BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
flags);
if (ret == -EINTR)
goto retry;
-
+out:
if (!ret) {
/* remove from alloc_heap: */
struct alloc_heap_entry e, *top = ca->alloc_heap.data;
percpu_up_read(&c->mark_lock);
}
- return ret;
+ return ret < 0 ? ret : 0;
}
static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
}
rcu_read_unlock();
- if (c->btree_roots_dirty) {
- bch2_journal_meta(&c->journal);
- goto again;
- }
-
return !nodes_unwritten &&
!bch2_btree_interior_updates_nr_pending(c);
}
enum {
/* startup: */
BCH_FS_ALLOC_READ_DONE,
+ BCH_FS_ALLOC_CLEAN,
BCH_FS_ALLOCATOR_STARTED,
BCH_FS_ALLOCATOR_RUNNING,
+ BCH_FS_ALLOCATOR_STOPPING,
BCH_FS_INITIAL_GC_DONE,
BCH_FS_FSCK_DONE,
BCH_FS_STARTED,
struct bio_set btree_bio;
struct btree_root btree_roots[BTREE_ID_NR];
- bool btree_roots_dirty;
struct mutex btree_root_lock;
struct btree_cache btree_cache;
- mempool_t btree_reserve_pool;
-
/*
* Cache of allocated btree nodes - if we allocate a btree node and
* don't use it, if we free it that space can't be reused until going
struct mutex btree_interior_update_lock;
struct closure_waitlist btree_interior_update_wait;
+ struct workqueue_struct *btree_interior_update_worker;
+ struct work_struct btree_interior_update_work;
+
+ /* btree_iter.c: */
+ struct mutex btree_trans_lock;
+ struct list_head btree_trans_list;
mempool_t btree_iters_pool;
struct workqueue_struct *wq;
LE64_BITMASK(BCH_SB_HAS_ERRORS, struct bch_sb, flags[0], 60, 61);
+LE64_BITMASK(BCH_SB_REFLINK, struct bch_sb, flags[0], 61, 62);
+
/* 61-64 unused */
LE64_BITMASK(BCH_SB_STR_HASH_TYPE, struct bch_sb, flags[1], 0, 4);
void bch2_bkey_to_text(struct printbuf *out, const struct bkey *k)
{
- pr_buf(out, "u64s %u type %s ", k->u64s,
- bch2_bkey_types[k->type]);
+ if (k) {
+ pr_buf(out, "u64s %u type %s ", k->u64s,
+ bch2_bkey_types[k->type]);
- bch2_bpos_to_text(out, k->p);
+ bch2_bpos_to_text(out, k->p);
- pr_buf(out, " snap %u len %u ver %llu",
- k->p.snapshot, k->size, k->version.lo);
+ pr_buf(out, " snap %u len %u ver %llu",
+ k->p.snapshot, k->size, k->version.lo);
+ } else {
+ pr_buf(out, "(null)");
+ }
}
void bch2_val_to_text(struct printbuf *out, struct bch_fs *c,
struct bkey_s_c k)
{
bch2_bkey_to_text(out, k.k);
- pr_buf(out, ": ");
- bch2_val_to_text(out, c, k);
+
+ if (k.k) {
+ pr_buf(out, ": ");
+ bch2_val_to_text(out, c, k);
+ }
}
void bch2_bkey_swab_val(struct bkey_s k)
list_del_init(&b->list);
mutex_unlock(&bc->lock);
- memalloc_nofs_restore(flags);
out:
b->flags = 0;
b->written = 0;
bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
start_time);
+ memalloc_nofs_restore(flags);
return b;
err:
/* Try to cannibalize another cached btree node: */
}
mutex_unlock(&bc->lock);
+ memalloc_nofs_restore(flags);
return ERR_PTR(-ENOMEM);
}
if (!parent)
return NULL;
+ /*
+ * There's a corner case where a btree_iter might have a node locked
+ * that is just outside its current pos - when
+ * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
+ *
+ * But the lock ordering checks in __bch2_btree_node_lock() go off of
+ * iter->pos, not the node's key: so if the iterator is marked as
+ * needing to be traversed, we risk deadlock if we don't bail out here:
+ */
+ if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
+ return ERR_PTR(-EINTR);
+
if (!bch2_btree_node_relock(iter, level + 1)) {
ret = ERR_PTR(-EINTR);
goto out;
mutex_unlock(&c->sb_lock);
}
+#if 0
/* Also see bch2_pending_btree_node_free_insert_done() */
static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
{
mutex_unlock(&c->btree_interior_update_lock);
}
+#endif
static void bch2_mark_allocator_buckets(struct bch_fs *c)
{
#define copy_bucket_field(_f) \
if (dst->b[b].mark._f != src->b[b].mark._f) { \
if (verify) \
- fsck_err(c, "dev %u bucket %zu has wrong " #_f \
+ fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f \
": got %u, should be %u", i, b, \
+ dst->b[b].mark.gen, \
+ bch2_data_types[dst->b[b].mark.data_type],\
dst->b[b].mark._f, src->b[b].mark._f); \
dst->b[b]._mark._f = src->b[b].mark._f; \
}
trace_gc_start(c);
down_write(&c->gc_lock);
+
+ /* flush interior btree updates: */
+ closure_wait_event(&c->btree_interior_update_wait,
+ !bch2_btree_interior_updates_nr_pending(c));
again:
ret = bch2_gc_start(c, metadata_only);
if (ret)
if (ret)
goto out;
+#if 0
bch2_mark_pending_btree_node_frees(c);
+#endif
bch2_mark_allocator_buckets(c);
c->gc_count++;
btree_node_reset_sib_u64s(n);
bch2_btree_build_aux_trees(n);
+
+ bch2_btree_update_add_new_node(as, n);
six_unlock_write(&n->lock);
bch2_btree_node_write(c, n, SIX_LOCK_intent);
bch2_btree_iter_node_replace(iter, new_nodes[0]);
for (i = 0; i < nr_new_nodes; i++)
- bch2_open_buckets_put(c, &new_nodes[i]->ob);
+ bch2_btree_update_get_open_buckets(as, new_nodes[i]);
/* Free the old nodes and update our sliding window */
for (i = 0; i < nr_old_nodes; i++) {
struct btree *b, struct bset *i,
unsigned offset, int write)
{
- pr_buf(out, "error validating btree node %s"
- "at btree %u level %u/%u\n"
- "pos %llu:%llu node offset %u",
+ pr_buf(out, "error validating btree node %sat btree %u level %u/%u\n"
+ "pos ",
write ? "before write " : "",
b->btree_id, b->level,
- c->btree_roots[b->btree_id].level,
- b->key.k.p.inode, b->key.k.p.offset,
- b->written);
+ c->btree_roots[b->btree_id].level);
+ bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
+
+ pr_buf(out, " node offset %u", b->written);
if (i)
pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
}
btree_err_on(b->data->keys.seq != bp->seq,
BTREE_ERR_MUST_RETRY, c, b, NULL,
- "got wrong btree node");
+ "got wrong btree node (seq %llx want %llx)",
+ b->data->keys.seq, bp->seq);
}
while (b->written < c->opts.btree_node_size) {
if (!linked->nodes_locked)
continue;
- /* * Must lock btree nodes in key order: */
- if (__btree_iter_cmp(iter->btree_id, pos, linked) < 0)
+ /* Must lock btree nodes in key order: */
+ if ((cmp_int(iter->btree_id, linked->btree_id) ?:
+ bkey_cmp(pos, linked->pos)) < 0)
ret = false;
/*
btree_iter_advance_to_pos(iter, l, -1);
+ /*
+ * XXX:
+ * keeping a node locked that's outside (even just outside) iter->pos
+ * breaks __bch2_btree_node_lock(). This seems to only affect
+ * bch2_btree_node_get_sibling so for now it's fixed there, but we
+ * should try to get rid of this corner case.
+ *
+ * (this behaviour is currently needed for BTREE_INSERT_NOUNLOCK)
+ */
+
if (bch2_btree_node_iter_end(&l->iter) &&
btree_iter_pos_after_node(iter, l->b))
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
struct btree_iter *iter;
trans_for_each_iter(trans, iter) {
- pr_err("iter: btree %s pos %llu:%llu%s%s%s %pf",
+ pr_err("iter: btree %s pos %llu:%llu%s%s%s %ps",
bch2_btree_ids[iter->btree_id],
iter->pos.inode,
iter->pos.offset,
trans->nr_updates2 = 0;
trans->mem_top = 0;
+ trans->extra_journal_entries = NULL;
+ trans->extra_journal_entry_u64s = 0;
+
if (trans->fs_usage_deltas) {
trans->fs_usage_deltas->used = 0;
memset(&trans->fs_usage_deltas->memset_start, 0,
if (expected_mem_bytes)
bch2_trans_preload_mem(trans, expected_mem_bytes);
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trans->pid = current->pid;
+ mutex_lock(&c->btree_trans_lock);
+ list_add(&trans->list, &c->btree_trans_list);
+ mutex_unlock(&c->btree_trans_lock);
+#endif
}
int bch2_trans_exit(struct btree_trans *trans)
{
bch2_trans_unlock(trans);
+#ifdef CONFIG_BCACHEFS_DEBUG
+ mutex_lock(&trans->c->btree_trans_lock);
+ list_del(&trans->list);
+ mutex_unlock(&trans->c->btree_trans_lock);
+#endif
+
kfree(trans->fs_usage_deltas);
kfree(trans->mem);
if (trans->used_mempool)
return trans->error ? -EIO : 0;
}
+void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
+{
+#ifdef CONFIG_BCACHEFS_DEBUG
+ struct btree_trans *trans;
+ struct btree_iter *iter;
+ struct btree *b;
+ unsigned l;
+
+ mutex_lock(&c->btree_trans_lock);
+ list_for_each_entry(trans, &c->btree_trans_list, list) {
+ pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
+
+ trans_for_each_iter(trans, iter) {
+ if (!iter->nodes_locked)
+ continue;
+
+ pr_buf(out, " iter %s:", bch2_btree_ids[iter->btree_id]);
+ bch2_bpos_to_text(out, iter->pos);
+ pr_buf(out, "\n");
+
+ for (l = 0; l < BTREE_MAX_DEPTH; l++) {
+ if (btree_node_locked(iter, l)) {
+ b = iter->l[l].b;
+
+ pr_buf(out, " %p l=%u %s ",
+ b, l, btree_node_intent_locked(iter, l) ? "i" : "r");
+ bch2_bpos_to_text(out, b->key.k.p);
+ pr_buf(out, "\n");
+ }
+ }
+ }
+
+ b = READ_ONCE(trans->locking);
+ if (b) {
+ pr_buf(out, " locking %px l=%u %s:",
+ b, b->level,
+ bch2_btree_ids[b->btree_id]);
+ bch2_bpos_to_text(out, b->key.k.p);
+ pr_buf(out, "\n");
+ }
+ }
+ mutex_unlock(&c->btree_trans_lock);
+#endif
+}
+
void bch2_fs_btree_iter_exit(struct bch_fs *c)
{
mempool_exit(&c->btree_iters_pool);
{
unsigned nr = BTREE_ITER_MAX;
+ INIT_LIST_HEAD(&c->btree_trans_list);
+ mutex_init(&c->btree_trans_lock);
+
return mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
sizeof(struct btree_iter) * nr +
sizeof(struct btree_insert_entry) * nr +
void __bch2_btree_iter_set_pos(struct btree_iter *, struct bpos, bool);
void bch2_btree_iter_set_pos(struct btree_iter *, struct bpos);
-static inline int __btree_iter_cmp(enum btree_id id,
- struct bpos pos,
- const struct btree_iter *r)
-{
- return cmp_int(id, r->btree_id) ?: bkey_cmp(pos, r->pos);
-}
-
static inline int btree_iter_cmp(const struct btree_iter *l,
const struct btree_iter *r)
{
- return __btree_iter_cmp(l->btree_id, l->pos, r);
+ return cmp_int(l->btree_id, r->btree_id) ?: bkey_cmp(l->pos, r->pos);
}
/*
void bch2_trans_init(struct btree_trans *, struct bch_fs *, unsigned, size_t);
int bch2_trans_exit(struct btree_trans *);
+void bch2_btree_trans_to_text(struct printbuf *, struct bch_fs *);
+
void bch2_fs_btree_iter_exit(struct bch_fs *);
int bch2_fs_btree_iter_init(struct bch_fs *);
struct btree_iter *iter,
enum six_lock_type type)
{
+ bool ret;
+
EBUG_ON(level >= BTREE_MAX_DEPTH);
+#ifdef CONFIG_BCACHEFS_DEBUG
+ iter->trans->locking = b;
+#endif
- return likely(six_trylock_type(&b->lock, type)) ||
+ ret = likely(six_trylock_type(&b->lock, type)) ||
btree_node_lock_increment(iter, b, level, type) ||
__bch2_btree_node_lock(b, pos, level, iter, type);
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+ iter->trans->locking = NULL;
+#endif
+ return ret;
}
bool __bch2_btree_node_relock(struct btree_iter *, unsigned);
struct btree_trans {
struct bch_fs *c;
+#ifdef CONFIG_BCACHEFS_DEBUG
+ struct list_head list;
+ struct btree *locking;
+ pid_t pid;
+#endif
unsigned long ip;
u64 iters_linked;
struct btree_insert_entry *updates2;
/* update path: */
+ struct jset_entry *extra_journal_entries;
+ unsigned extra_journal_entry_u64s;
+ struct journal_entry_pin *journal_pin;
+
struct journal_res journal_res;
struct journal_preres journal_preres;
u64 *journal_seq;
#include <linux/random.h>
#include <trace/events/bcachefs.h>
-static void btree_node_will_make_reachable(struct btree_update *,
- struct btree *);
-static void btree_update_drop_new_node(struct bch_fs *, struct btree *);
-
/* Debug code: */
/*
/* Btree node freeing/allocation: */
-static bool btree_key_matches(struct bch_fs *c,
- struct bkey_s_c l,
- struct bkey_s_c r)
-{
- struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(l);
- struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(r);
- const struct bch_extent_ptr *ptr1, *ptr2;
-
- bkey_for_each_ptr(ptrs1, ptr1)
- bkey_for_each_ptr(ptrs2, ptr2)
- if (ptr1->dev == ptr2->dev &&
- ptr1->gen == ptr2->gen &&
- ptr1->offset == ptr2->offset)
- return true;
-
- return false;
-}
-
-/*
- * We're doing the index update that makes @b unreachable, update stuff to
- * reflect that:
- *
- * Must be called _before_ btree_update_updated_root() or
- * btree_update_updated_node:
- */
-static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b,
- struct bkey_s_c k,
- struct bch_fs_usage *stats)
-{
- struct bch_fs *c = as->c;
- struct pending_btree_node_free *d;
-
- for (d = as->pending; d < as->pending + as->nr_pending; d++)
- if (!bkey_cmp(k.k->p, d->key.k.p) &&
- btree_key_matches(c, k, bkey_i_to_s_c(&d->key)))
- goto found;
- BUG();
-found:
- BUG_ON(d->index_update_done);
- d->index_update_done = true;
-
- /*
- * We're dropping @k from the btree, but it's still live until the
- * index update is persistent so we need to keep a reference around for
- * mark and sweep to find - that's primarily what the
- * btree_node_pending_free list is for.
- *
- * So here (when we set index_update_done = true), we're moving an
- * existing reference to a different part of the larger "gc keyspace" -
- * and the new position comes after the old position, since GC marks
- * the pending free list after it walks the btree.
- *
- * If we move the reference while mark and sweep is _between_ the old
- * and the new position, mark and sweep will see the reference twice
- * and it'll get double accounted - so check for that here and subtract
- * to cancel out one of mark and sweep's markings if necessary:
- */
-
- if (gc_pos_cmp(c->gc_pos, b
- ? gc_pos_btree_node(b)
- : gc_pos_btree_root(as->btree_id)) >= 0 &&
- gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0)
- bch2_mark_key_locked(c, bkey_i_to_s_c(&d->key),
- 0, 0, NULL, 0,
- BTREE_TRIGGER_OVERWRITE|
- BTREE_TRIGGER_GC);
-}
-
static void __btree_node_free(struct bch_fs *c, struct btree *b)
{
trace_btree_node_free(c, b);
{
struct open_buckets ob = b->ob;
- btree_update_drop_new_node(c, b);
-
b->ob.nr = 0;
clear_btree_node_dirty(b);
trans_for_each_iter(iter->trans, linked)
BUG_ON(linked->l[b->level].b == b);
- /*
- * Is this a node that isn't reachable on disk yet?
- *
- * Nodes that aren't reachable yet have writes blocked until they're
- * reachable - now that we've cancelled any pending writes and moved
- * things waiting on that write to wait on this update, we can drop this
- * node from the list of nodes that the other update is making
- * reachable, prior to freeing it:
- */
- btree_update_drop_new_node(c, b);
-
six_lock_write(&b->lock);
__btree_node_free(c, b);
six_unlock_write(&b->lock);
six_unlock_intent(&b->lock);
}
-static void bch2_btree_node_free_ondisk(struct bch_fs *c,
- struct pending_btree_node_free *pending,
- u64 journal_seq)
-{
- BUG_ON(!pending->index_update_done);
-
- bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
- 0, 0, NULL, journal_seq, BTREE_TRIGGER_OVERWRITE);
-
- if (gc_visited(c, gc_phase(GC_PHASE_PENDING_DELETE)))
- bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
- 0, 0, NULL, journal_seq,
- BTREE_TRIGGER_OVERWRITE|
- BTREE_TRIGGER_GC);
-}
-
static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
struct disk_reservation *res,
struct closure *cl,
int ret;
BUG_ON(level >= BTREE_MAX_DEPTH);
- BUG_ON(!as->reserve->nr);
+ BUG_ON(!as->nr_prealloc_nodes);
- b = as->reserve->b[--as->reserve->nr];
+ b = as->prealloc_nodes[--as->nr_prealloc_nodes];
set_btree_node_accessed(b);
set_btree_node_dirty(b);
bch2_btree_build_aux_trees(b);
- btree_node_will_make_reachable(as, b);
-
ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
BUG_ON(ret);
btree_node_set_format(b, b->data->format);
bch2_btree_build_aux_trees(b);
+ bch2_btree_update_add_new_node(as, b);
six_unlock_write(&b->lock);
return b;
}
-static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reserve)
+static void bch2_btree_reserve_put(struct btree_update *as)
{
- bch2_disk_reservation_put(c, &reserve->disk_res);
+ struct bch_fs *c = as->c;
mutex_lock(&c->btree_reserve_cache_lock);
- while (reserve->nr) {
- struct btree *b = reserve->b[--reserve->nr];
+ while (as->nr_prealloc_nodes) {
+ struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
six_unlock_write(&b->lock);
}
mutex_unlock(&c->btree_reserve_cache_lock);
-
- mempool_free(reserve, &c->btree_reserve_pool);
}
-static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c,
- unsigned nr_nodes,
- unsigned flags,
- struct closure *cl)
+static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
+ unsigned flags, struct closure *cl)
{
- struct btree_reserve *reserve;
+ struct bch_fs *c = as->c;
struct btree *b;
- struct disk_reservation disk_res = { 0, 0 };
- unsigned sectors = nr_nodes * c->opts.btree_node_size;
- int ret, disk_res_flags = 0;
-
- if (flags & BTREE_INSERT_NOFAIL)
- disk_res_flags |= BCH_DISK_RESERVATION_NOFAIL;
-
- /*
- * This check isn't necessary for correctness - it's just to potentially
- * prevent us from doing a lot of work that'll end up being wasted:
- */
- ret = bch2_journal_error(&c->journal);
- if (ret)
- return ERR_PTR(ret);
-
- if (bch2_disk_reservation_get(c, &disk_res, sectors,
- c->opts.metadata_replicas,
- disk_res_flags))
- return ERR_PTR(-ENOSPC);
+ int ret;
BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
* open bucket reserve:
*/
ret = bch2_btree_cache_cannibalize_lock(c, cl);
- if (ret) {
- bch2_disk_reservation_put(c, &disk_res);
- return ERR_PTR(ret);
- }
-
- reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO);
-
- reserve->disk_res = disk_res;
- reserve->nr = 0;
+ if (ret)
+ return ret;
- while (reserve->nr < nr_nodes) {
- b = __bch2_btree_node_alloc(c, &disk_res,
+ while (as->nr_prealloc_nodes < nr_nodes) {
+ b = __bch2_btree_node_alloc(c, &as->disk_res,
flags & BTREE_INSERT_NOWAIT
? NULL : cl, flags);
if (IS_ERR(b)) {
if (ret)
goto err_free;
- reserve->b[reserve->nr++] = b;
+ as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
}
bch2_btree_cache_cannibalize_unlock(c);
- return reserve;
+ return 0;
err_free:
- bch2_btree_reserve_put(c, reserve);
bch2_btree_cache_cannibalize_unlock(c);
trace_btree_reserve_get_fail(c, nr_nodes, cl);
- return ERR_PTR(ret);
+ return ret;
}
/* Asynchronous interior node update machinery */
-static void __bch2_btree_update_free(struct btree_update *as)
+static void bch2_btree_update_free(struct btree_update *as)
{
struct bch_fs *c = as->c;
bch2_journal_pin_drop(&c->journal, &as->journal);
bch2_journal_pin_flush(&c->journal, &as->journal);
+ bch2_disk_reservation_put(c, &as->disk_res);
+ bch2_btree_reserve_put(as);
- BUG_ON(as->nr_new_nodes || as->nr_pending);
-
- if (as->reserve)
- bch2_btree_reserve_put(c, as->reserve);
-
+ mutex_lock(&c->btree_interior_update_lock);
list_del(&as->unwritten_list);
list_del(&as->list);
+ mutex_unlock(&c->btree_interior_update_lock);
closure_debug_destroy(&as->cl);
mempool_free(as, &c->btree_interior_update_pool);
closure_wake_up(&c->btree_interior_update_wait);
}
-static void bch2_btree_update_free(struct btree_update *as)
+static void btree_update_will_delete_key(struct btree_update *as,
+ struct bkey_i *k)
{
- struct bch_fs *c = as->c;
+ BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
+ ARRAY_SIZE(as->_old_keys));
+ bch2_keylist_add(&as->old_keys, k);
+}
- mutex_lock(&c->btree_interior_update_lock);
- __bch2_btree_update_free(as);
- mutex_unlock(&c->btree_interior_update_lock);
+static void btree_update_will_add_key(struct btree_update *as,
+ struct bkey_i *k)
+{
+ BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
+ ARRAY_SIZE(as->_new_keys));
+ bch2_keylist_add(&as->new_keys, k);
}
-static inline bool six_trylock_intentwrite(struct six_lock *lock)
+/*
+ * The transactional part of an interior btree node update, where we journal the
+ * update we did to the interior node and update alloc info:
+ */
+static int btree_update_nodes_written_trans(struct btree_trans *trans,
+ struct btree_update *as)
{
- if (!six_trylock_intent(lock))
- return false;
+ struct bkey_i *k;
+ int ret;
+
+ trans->extra_journal_entries = (void *) &as->journal_entries[0];
+ trans->extra_journal_entry_u64s = as->journal_u64s;
+ trans->journal_pin = &as->journal;
- if (!six_trylock_write(lock)) {
- six_unlock_intent(lock);
- return false;
+ for_each_keylist_key(&as->new_keys, k) {
+ ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
+ 0, 0, BTREE_TRIGGER_INSERT);
+ if (ret)
+ return ret;
}
- return true;
+ for_each_keylist_key(&as->old_keys, k) {
+ ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
+ 0, 0, BTREE_TRIGGER_OVERWRITE);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
}
-static void btree_update_nodes_written(struct closure *cl)
+static void btree_update_nodes_written(struct btree_update *as)
{
- struct btree_update *as = container_of(cl, struct btree_update, cl);
- struct btree *nodes_need_write[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES + 1];
- unsigned nr_nodes_need_write;
- struct journal_res res = { 0 };
struct bch_fs *c = as->c;
- struct btree_root *r;
- struct btree *b;
+ struct btree *b = as->b;
+ u64 journal_seq = 0;
+ unsigned i;
int ret;
/*
* to child nodes that weren't written yet: now, the child nodes have
* been written so we can write out the update to the interior node.
*/
- mutex_lock(&c->btree_interior_update_lock);
- as->nodes_written = true;
-again:
- nr_nodes_need_write = 0;
- as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
- struct btree_update, unwritten_list);
- if (!as || !as->nodes_written) {
- mutex_unlock(&c->btree_interior_update_lock);
- return;
- }
-
- b = as->b;
- if (b && !six_trylock_intentwrite(&b->lock)) {
- mutex_unlock(&c->btree_interior_update_lock);
-
- btree_node_lock_type(c, b, SIX_LOCK_intent);
- six_lock_write(&b->lock);
-
- six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
-
- mutex_lock(&c->btree_interior_update_lock);
- goto again;
- }
-
- ret = bch2_journal_res_get(&c->journal, &res, as->journal_u64s,
- JOURNAL_RES_GET_NONBLOCK|
- JOURNAL_RES_GET_RESERVED);
- if (ret == -EAGAIN) {
- unsigned u64s = as->journal_u64s;
-
- if (b) {
- six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
- }
-
- mutex_unlock(&c->btree_interior_update_lock);
-
- ret = bch2_journal_res_get(&c->journal, &res, u64s,
- JOURNAL_RES_GET_CHECK|
- JOURNAL_RES_GET_RESERVED);
- if (!ret) {
- mutex_lock(&c->btree_interior_update_lock);
- goto again;
- }
- }
-
- if (!ret) {
- struct journal_buf *buf = &c->journal.buf[res.idx];
- struct jset_entry *entry = vstruct_idx(buf->data, res.offset);
-
- res.offset += as->journal_u64s;
- res.u64s -= as->journal_u64s;
- memcpy_u64s(entry, as->journal_entries, as->journal_u64s);
- } else {
- /*
- * On journal error we have to run most of the normal path so
- * that shutdown works - unblocking btree node writes in
- * particular and writing them if needed - except for
- * journalling the update:
- */
-
- BUG_ON(!bch2_journal_error(&c->journal));
- }
-
- switch (as->mode) {
- case BTREE_INTERIOR_NO_UPDATE:
- BUG();
- case BTREE_INTERIOR_UPDATING_NODE:
- /* @b is the node we did the final insert into: */
-
+ ret = bch2_trans_do(c, &as->disk_res, &journal_seq,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_NOCHECK_RW|
+ BTREE_INSERT_JOURNAL_RESERVED,
+ btree_update_nodes_written_trans(&trans, as));
+ BUG_ON(ret && !bch2_journal_error(&c->journal));
+
+ if (b) {
/*
+ * @b is the node we did the final insert into:
+ *
* On failure to get a journal reservation, we still have to
* unblock the write and allow most of the write path to happen
* so that shutdown works, but the i->journal_seq mechanism
* we're in journal error state:
*/
+ btree_node_lock_type(c, b, SIX_LOCK_intent);
+ btree_node_lock_type(c, b, SIX_LOCK_write);
+ mutex_lock(&c->btree_interior_update_lock);
+
list_del(&as->write_blocked_list);
- if (!ret) {
+ if (!ret && as->b == b) {
struct bset *i = btree_bset_last(b);
+ BUG_ON(!b->level);
+ BUG_ON(!btree_node_dirty(b));
+
i->journal_seq = cpu_to_le64(
- max(res.seq,
+ max(journal_seq,
le64_to_cpu(i->journal_seq)));
- bch2_btree_add_journal_pin(c, b, res.seq);
+ bch2_btree_add_journal_pin(c, b, journal_seq);
}
- nodes_need_write[nr_nodes_need_write++] = b;
-
+ mutex_unlock(&c->btree_interior_update_lock);
six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
- break;
-
- case BTREE_INTERIOR_UPDATING_AS:
- BUG_ON(b);
- break;
- case BTREE_INTERIOR_UPDATING_ROOT:
- r = &c->btree_roots[as->btree_id];
-
- BUG_ON(b);
-
- mutex_lock(&c->btree_root_lock);
- bkey_copy(&r->key, as->parent_keys.keys);
- r->level = as->level;
- r->alive = true;
- c->btree_roots_dirty = true;
- mutex_unlock(&c->btree_root_lock);
- break;
+ btree_node_write_if_need(c, b, SIX_LOCK_intent);
+ six_unlock_intent(&b->lock);
}
bch2_journal_pin_drop(&c->journal, &as->journal);
- bch2_journal_res_put(&c->journal, &res);
bch2_journal_preres_put(&c->journal, &as->journal_preres);
- while (as->nr_new_nodes) {
- b = as->new_nodes[--as->nr_new_nodes];
+ mutex_lock(&c->btree_interior_update_lock);
+ for (i = 0; i < as->nr_new_nodes; i++) {
+ b = as->new_nodes[i];
BUG_ON(b->will_make_reachable != (unsigned long) as);
b->will_make_reachable = 0;
+ }
+ mutex_unlock(&c->btree_interior_update_lock);
+
+ for (i = 0; i < as->nr_new_nodes; i++) {
+ b = as->new_nodes[i];
- nodes_need_write[nr_nodes_need_write++] = b;
+ btree_node_lock_type(c, b, SIX_LOCK_read);
+ btree_node_write_if_need(c, b, SIX_LOCK_read);
+ six_unlock_read(&b->lock);
}
- while (as->nr_pending)
- bch2_btree_node_free_ondisk(c,
- &as->pending[--as->nr_pending], res.seq);
+ for (i = 0; i < as->nr_open_buckets; i++)
+ bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
- __bch2_btree_update_free(as);
- /*
- * for flush_held_btree_writes() waiting on updates to flush or
- * nodes to be writeable:
- */
- closure_wake_up(&c->btree_interior_update_wait);
+ bch2_btree_update_free(as);
+}
- /*
- * Can't take btree node locks while holding btree_interior_update_lock:
- * */
- mutex_unlock(&c->btree_interior_update_lock);
+static void btree_interior_update_work(struct work_struct *work)
+{
+ struct bch_fs *c =
+ container_of(work, struct bch_fs, btree_interior_update_work);
+ struct btree_update *as;
- /* Do btree writes after dropping journal res/locks: */
- while (nr_nodes_need_write) {
- b = nodes_need_write[--nr_nodes_need_write];
+ while (1) {
+ mutex_lock(&c->btree_interior_update_lock);
+ as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
+ struct btree_update, unwritten_list);
+ if (as && !as->nodes_written)
+ as = NULL;
+ mutex_unlock(&c->btree_interior_update_lock);
- btree_node_lock_type(c, b, SIX_LOCK_read);
- bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
- six_unlock_read(&b->lock);
+ if (!as)
+ break;
+
+ btree_update_nodes_written(as);
}
+}
+
+static void btree_update_set_nodes_written(struct closure *cl)
+{
+ struct btree_update *as = container_of(cl, struct btree_update, cl);
+ struct bch_fs *c = as->c;
mutex_lock(&c->btree_interior_update_lock);
- goto again;
+ as->nodes_written = true;
+ mutex_unlock(&c->btree_interior_update_lock);
+
+ queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
}
/*
as->mode = BTREE_INTERIOR_UPDATING_NODE;
as->b = b;
- as->level = b->level;
list_add(&as->write_blocked_list, &b->write_blocked);
mutex_unlock(&c->btree_interior_update_lock);
static void btree_update_updated_root(struct btree_update *as, struct btree *b)
{
+ struct bkey_i *insert = &b->key;
struct bch_fs *c = as->c;
BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
- BUG_ON(!bch2_keylist_empty(&as->parent_keys));
+
+ BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
+ ARRAY_SIZE(as->journal_entries));
+
+ as->journal_u64s +=
+ journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
+ BCH_JSET_ENTRY_btree_root,
+ b->btree_id, b->level,
+ insert, insert->k.u64s);
mutex_lock(&c->btree_interior_update_lock);
list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
as->mode = BTREE_INTERIOR_UPDATING_ROOT;
- as->level = b->level;
- bch2_keylist_add(&as->parent_keys, &b->key);
mutex_unlock(&c->btree_interior_update_lock);
}
-static void btree_node_will_make_reachable(struct btree_update *as,
- struct btree *b)
+/*
+ * bch2_btree_update_add_new_node:
+ *
+ * This causes @as to wait on @b to be written, before it gets to
+ * bch2_btree_update_nodes_written
+ *
+ * Additionally, it sets b->will_make_reachable to prevent any additional writes
+ * to @b from happening besides the first until @b is reachable on disk
+ *
+ * And it adds @b to the list of @as's new nodes, so that we can update sector
+ * counts in bch2_btree_update_nodes_written:
+ */
+void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
{
struct bch_fs *c = as->c;
+ closure_get(&as->cl);
+
mutex_lock(&c->btree_interior_update_lock);
BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
BUG_ON(b->will_make_reachable);
as->new_nodes[as->nr_new_nodes++] = b;
b->will_make_reachable = 1UL|(unsigned long) as;
- closure_get(&as->cl);
mutex_unlock(&c->btree_interior_update_lock);
+
+ btree_update_will_add_key(as, &b->key);
}
+/*
+ * returns true if @b was a new node
+ */
static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
{
struct btree_update *as;
unsigned i;
mutex_lock(&c->btree_interior_update_lock);
+ /*
+ * When b->will_make_reachable != 0, it owns a ref on as->cl that's
+ * dropped when it gets written by bch2_btree_complete_write - the
+ * xchg() is for synchronization with bch2_btree_complete_write:
+ */
v = xchg(&b->will_make_reachable, 0);
as = (struct btree_update *) (v & ~1UL);
closure_put(&as->cl);
}
-static void btree_interior_update_add_node_reference(struct btree_update *as,
- struct btree *b)
+void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
{
- struct bch_fs *c = as->c;
- struct pending_btree_node_free *d;
-
- mutex_lock(&c->btree_interior_update_lock);
-
- /* Add this node to the list of nodes being freed: */
- BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending));
-
- d = &as->pending[as->nr_pending++];
- d->index_update_done = false;
- d->seq = b->data->keys.seq;
- d->btree_id = b->btree_id;
- d->level = b->level;
- bkey_copy(&d->key, &b->key);
-
- mutex_unlock(&c->btree_interior_update_lock);
+ while (b->ob.nr)
+ as->open_buckets[as->nr_open_buckets++] =
+ b->ob.v[--b->ob.nr];
}
/*
if (btree_node_fake(b))
return;
- btree_interior_update_add_node_reference(as, b);
-
mutex_lock(&c->btree_interior_update_lock);
/*
bch2_journal_pin_drop(&c->journal, &w->journal);
mutex_unlock(&c->btree_interior_update_lock);
+
+ /*
+ * Is this a node that isn't reachable on disk yet?
+ *
+ * Nodes that aren't reachable yet have writes blocked until they're
+ * reachable - now that we've cancelled any pending writes and moved
+ * things waiting on that write to wait on this update, we can drop this
+ * node from the list of nodes that the other update is making
+ * reachable, prior to freeing it:
+ */
+ btree_update_drop_new_node(c, b);
+
+ btree_update_will_delete_key(as, &b->key);
}
void bch2_btree_update_done(struct btree_update *as)
{
BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
- bch2_btree_reserve_put(as->c, as->reserve);
- as->reserve = NULL;
+ bch2_btree_reserve_put(as);
- continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq);
+ continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
}
struct btree_update *
struct closure *cl)
{
struct bch_fs *c = trans->c;
- struct journal_preres journal_preres = { 0 };
- struct btree_reserve *reserve;
struct btree_update *as;
- int ret;
+ int ret, disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
+ ? BCH_DISK_RESERVATION_NOFAIL : 0;
+
+ /*
+ * This check isn't necessary for correctness - it's just to potentially
+ * prevent us from doing a lot of work that'll end up being wasted:
+ */
+ ret = bch2_journal_error(&c->journal);
+ if (ret)
+ return ERR_PTR(ret);
- ret = bch2_journal_preres_get(&c->journal, &journal_preres,
+ as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
+ memset(as, 0, sizeof(*as));
+ closure_init(&as->cl, NULL);
+ as->c = c;
+ as->mode = BTREE_INTERIOR_NO_UPDATE;
+ as->btree_id = id;
+ INIT_LIST_HEAD(&as->list);
+ INIT_LIST_HEAD(&as->unwritten_list);
+ INIT_LIST_HEAD(&as->write_blocked_list);
+ bch2_keylist_init(&as->old_keys, as->_old_keys);
+ bch2_keylist_init(&as->new_keys, as->_new_keys);
+ bch2_keylist_init(&as->parent_keys, as->inline_keys);
+
+ ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
BTREE_UPDATE_JOURNAL_RES,
JOURNAL_RES_GET_NONBLOCK);
if (ret == -EAGAIN) {
bch2_trans_unlock(trans);
- ret = bch2_journal_preres_get(&c->journal, &journal_preres,
+ ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
BTREE_UPDATE_JOURNAL_RES, 0);
if (ret)
return ERR_PTR(ret);
if (!bch2_trans_relock(trans)) {
- bch2_journal_preres_put(&c->journal, &journal_preres);
- return ERR_PTR(-EINTR);
+ ret = -EINTR;
+ goto err;
}
}
- reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
- if (IS_ERR(reserve)) {
- bch2_journal_preres_put(&c->journal, &journal_preres);
- return ERR_CAST(reserve);
- }
-
- as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
- memset(as, 0, sizeof(*as));
- closure_init(&as->cl, NULL);
- as->c = c;
- as->mode = BTREE_INTERIOR_NO_UPDATE;
- as->btree_id = id;
- as->reserve = reserve;
- INIT_LIST_HEAD(&as->write_blocked_list);
- INIT_LIST_HEAD(&as->unwritten_list);
- as->journal_preres = journal_preres;
+ ret = bch2_disk_reservation_get(c, &as->disk_res,
+ nr_nodes * c->opts.btree_node_size,
+ c->opts.metadata_replicas,
+ disk_res_flags);
+ if (ret)
+ goto err;
- bch2_keylist_init(&as->parent_keys, as->inline_keys);
+ ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
+ if (ret)
+ goto err;
mutex_lock(&c->btree_interior_update_lock);
list_add_tail(&as->list, &c->btree_interior_update_list);
mutex_unlock(&c->btree_interior_update_lock);
return as;
+err:
+ bch2_btree_update_free(as);
+ return ERR_PTR(ret);
}
/* Btree root updates: */
-static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
+static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
{
/* Root nodes cannot be reaped */
mutex_lock(&c->btree_cache.lock);
bch2_recalc_btree_reserve(c);
}
-static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
-{
- struct bch_fs *c = as->c;
- struct btree *old = btree_node_root(c, b);
- struct bch_fs_usage *fs_usage;
-
- __bch2_btree_set_root_inmem(c, b);
-
- mutex_lock(&c->btree_interior_update_lock);
- percpu_down_read(&c->mark_lock);
- fs_usage = bch2_fs_usage_scratch_get(c);
-
- bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
- 0, 0, fs_usage, 0,
- BTREE_TRIGGER_INSERT);
- if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
- bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
- 0, 0, NULL, 0,
- BTREE_TRIGGER_INSERT|
- BTREE_TRIGGER_GC);
-
- if (old && !btree_node_fake(old))
- bch2_btree_node_free_index(as, NULL,
- bkey_i_to_s_c(&old->key),
- fs_usage);
- bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
-
- bch2_fs_usage_scratch_put(c, fs_usage);
- percpu_up_read(&c->mark_lock);
- mutex_unlock(&c->btree_interior_update_lock);
-}
-
/**
* bch_btree_set_root - update the root in memory and on disk
*
*/
bch2_btree_node_lock_write(old, iter);
- bch2_btree_set_root_inmem(as, b);
+ bch2_btree_set_root_inmem(c, b);
btree_update_updated_root(as, b);
struct bkey_i *insert,
struct btree_node_iter *node_iter)
{
- struct bch_fs *c = as->c;
- struct bch_fs_usage *fs_usage;
- struct jset_entry *entry;
struct bkey_packed *k;
- struct bkey tmp;
BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
ARRAY_SIZE(as->journal_entries));
- entry = (void *) &as->journal_entries[as->journal_u64s];
- memset(entry, 0, sizeof(*entry));
- entry->u64s = cpu_to_le16(insert->k.u64s);
- entry->type = BCH_JSET_ENTRY_btree_keys;
- entry->btree_id = b->btree_id;
- entry->level = b->level;
- memcpy_u64s_small(entry->_data, insert, insert->k.u64s);
- as->journal_u64s += jset_u64s(insert->k.u64s);
-
- mutex_lock(&c->btree_interior_update_lock);
- percpu_down_read(&c->mark_lock);
- fs_usage = bch2_fs_usage_scratch_get(c);
-
- bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
- 0, 0, fs_usage, 0,
- BTREE_TRIGGER_INSERT);
-
- if (gc_visited(c, gc_pos_btree_node(b)))
- bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
- 0, 0, NULL, 0,
- BTREE_TRIGGER_INSERT|
- BTREE_TRIGGER_GC);
+ as->journal_u64s +=
+ journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
+ BCH_JSET_ENTRY_btree_keys,
+ b->btree_id, b->level,
+ insert, insert->k.u64s);
while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
bch2_btree_node_iter_advance(node_iter, b);
- /*
- * If we're overwriting, look up pending delete and mark so that gc
- * marks it on the pending delete list:
- */
- if (k && !bkey_cmp_packed(b, k, &insert->k))
- bch2_btree_node_free_index(as, b,
- bkey_disassemble(b, k, &tmp),
- fs_usage);
-
- bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
-
- bch2_fs_usage_scratch_put(c, fs_usage);
- percpu_up_read(&c->mark_lock);
- mutex_unlock(&c->btree_interior_update_lock);
-
bch2_btree_bset_insert_key(iter, b, node_iter, insert);
set_btree_node_dirty(b);
set_btree_node_need_write(b);
struct bkey_packed *k, *prev = NULL;
n2 = bch2_btree_node_alloc(as, n1->level);
+ bch2_btree_update_add_new_node(as, n2);
n2->data->max_key = n1->data->max_key;
n2->data->format = n1->format;
struct bkey_packed *src, *dst, *n;
struct bset *i;
- /*
- * XXX
- *
- * these updates must be journalled
- *
- * oops
- */
-
BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
bch2_btree_interior_update_will_free_node(as, b);
n1 = bch2_btree_node_alloc_replacement(as, b);
+ bch2_btree_update_add_new_node(as, n1);
if (keys)
btree_split_insert_keys(as, n1, iter, keys);
bch2_btree_set_root(as, n1, iter);
}
- bch2_open_buckets_put(c, &n1->ob);
+ bch2_btree_update_get_open_buckets(as, n1);
if (n2)
- bch2_open_buckets_put(c, &n2->ob);
+ bch2_btree_update_get_open_buckets(as, n2);
if (n3)
- bch2_open_buckets_put(c, &n3->ob);
+ bch2_btree_update_get_open_buckets(as, n3);
/* Successful split, update the iterator to point to the new nodes: */
bch2_btree_node_lock_for_insert(c, b, iter);
- if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
+ if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
bch2_btree_node_unlock_write(b, iter);
goto split;
}
bch2_btree_interior_update_will_free_node(as, m);
n = bch2_btree_node_alloc(as, b->level);
+ bch2_btree_update_add_new_node(as, n);
btree_set_min(n, prev->data->min_key);
btree_set_max(n, next->data->max_key);
bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
- bch2_open_buckets_put(c, &n->ob);
+ bch2_btree_update_get_open_buckets(as, n);
six_lock_increment(&b->lock, SIX_LOCK_intent);
bch2_btree_iter_node_drop(iter, b);
bch2_btree_interior_update_will_free_node(as, b);
n = bch2_btree_node_alloc_replacement(as, b);
+ bch2_btree_update_add_new_node(as, n);
bch2_btree_build_aux_trees(n);
six_unlock_write(&n->lock);
bch2_btree_set_root(as, n, iter);
}
- bch2_open_buckets_put(c, &n->ob);
+ bch2_btree_update_get_open_buckets(as, n);
six_lock_increment(&b->lock, SIX_LOCK_intent);
bch2_btree_iter_node_drop(iter, b);
struct btree *parent;
int ret;
- /*
- * Two corner cases that need to be thought about here:
- *
- * @b may not be reachable yet - there might be another interior update
- * operation waiting on @b to be written, and we're gonna deliver the
- * write completion to that interior update operation _before_
- * persisting the new_key update
- *
- * That ends up working without us having to do anything special here:
- * the reason is, we do kick off (and do the in memory updates) for the
- * update for @new_key before we return, creating a new interior_update
- * operation here.
- *
- * The new interior update operation here will in effect override the
- * previous one. The previous one was going to terminate - make @b
- * reachable - in one of two ways:
- * - updating the btree root pointer
- * In that case,
- * no, this doesn't work. argh.
- */
-
- if (b->will_make_reachable)
- as->must_rewrite = true;
-
- btree_interior_update_add_node_reference(as, b);
-
- /*
- * XXX: the rest of the update path treats this like we're actually
- * inserting a new node and deleting the existing node, so the
- * reservation needs to include enough space for @b
- *
- * that is actually sketch as fuck though and I am surprised the code
- * seems to work like that, definitely need to go back and rework it
- * into something saner.
- *
- * (I think @b is just getting double counted until the btree update
- * finishes and "deletes" @b on disk)
- */
- ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
- c->opts.btree_node_size *
- bch2_bkey_nr_ptrs(bkey_i_to_s_c(new_key)),
- BCH_DISK_RESERVATION_NOFAIL);
- BUG_ON(ret);
+ btree_update_will_delete_key(as, &b->key);
+ btree_update_will_add_key(as, new_key);
parent = btree_node_parent(iter, b);
if (parent) {
bkey_copy(&b->key, new_key);
}
} else {
- struct bch_fs_usage *fs_usage;
-
BUG_ON(btree_node_root(c, b) != b);
bch2_btree_node_lock_write(b, iter);
+ bkey_copy(&b->key, new_key);
- mutex_lock(&c->btree_interior_update_lock);
- percpu_down_read(&c->mark_lock);
- fs_usage = bch2_fs_usage_scratch_get(c);
-
- bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
- 0, 0, fs_usage, 0,
- BTREE_TRIGGER_INSERT);
- if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
- bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
- 0, 0, NULL, 0,
- BTREE_TRIGGER_INSERT||
- BTREE_TRIGGER_GC);
-
- bch2_btree_node_free_index(as, NULL,
- bkey_i_to_s_c(&b->key),
- fs_usage);
- bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
-
- bch2_fs_usage_scratch_put(c, fs_usage);
- percpu_up_read(&c->mark_lock);
- mutex_unlock(&c->btree_interior_update_lock);
-
- if (btree_ptr_hash_val(new_key) != b->hash_val) {
+ if (btree_ptr_hash_val(&b->key) != b->hash_val) {
mutex_lock(&c->btree_cache.lock);
bch2_btree_node_hash_remove(&c->btree_cache, b);
- bkey_copy(&b->key, new_key);
ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
BUG_ON(ret);
mutex_unlock(&c->btree_cache.lock);
- } else {
- bkey_copy(&b->key, new_key);
}
btree_update_updated_root(as, b);
{
BUG_ON(btree_node_root(c, b));
- __bch2_btree_set_root_inmem(c, b);
+ bch2_btree_set_root_inmem(c, b);
}
void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
BUG_ON(ret);
- __bch2_btree_set_root_inmem(c, b);
+ bch2_btree_set_root_inmem(c, b);
six_unlock_write(&b->lock);
six_unlock_intent(&b->lock);
return ret;
}
+
+void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
+{
+ struct btree_root *r;
+ struct jset_entry *entry;
+
+ mutex_lock(&c->btree_root_lock);
+
+ vstruct_for_each(jset, entry)
+ if (entry->type == BCH_JSET_ENTRY_btree_root) {
+ r = &c->btree_roots[entry->btree_id];
+ r->level = entry->level;
+ r->alive = true;
+ bkey_copy(&r->key, &entry->start[0]);
+ }
+
+ mutex_unlock(&c->btree_root_lock);
+}
+
+struct jset_entry *
+bch2_btree_roots_to_journal_entries(struct bch_fs *c,
+ struct jset_entry *start,
+ struct jset_entry *end)
+{
+ struct jset_entry *entry;
+ unsigned long have = 0;
+ unsigned i;
+
+ for (entry = start; entry < end; entry = vstruct_next(entry))
+ if (entry->type == BCH_JSET_ENTRY_btree_root)
+ __set_bit(entry->btree_id, &have);
+
+ mutex_lock(&c->btree_root_lock);
+
+ for (i = 0; i < BTREE_ID_NR; i++)
+ if (c->btree_roots[i].alive && !test_bit(i, &have)) {
+ journal_entry_set(end,
+ BCH_JSET_ENTRY_btree_root,
+ i, c->btree_roots[i].level,
+ &c->btree_roots[i].key,
+ c->btree_roots[i].key.u64s);
+ end = vstruct_next(end);
+ }
+
+ mutex_unlock(&c->btree_root_lock);
+
+ return end;
+}
+
+void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
+{
+ if (c->btree_interior_update_worker)
+ destroy_workqueue(c->btree_interior_update_worker);
+ mempool_exit(&c->btree_interior_update_pool);
+}
+
+int bch2_fs_btree_interior_update_init(struct bch_fs *c)
+{
+ mutex_init(&c->btree_reserve_cache_lock);
+ INIT_LIST_HEAD(&c->btree_interior_update_list);
+ INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
+ mutex_init(&c->btree_interior_update_lock);
+ INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
+
+ c->btree_interior_update_worker =
+ alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
+ if (!c->btree_interior_update_worker)
+ return -ENOMEM;
+
+ return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
+ sizeof(struct btree_update));
+}
#include "btree_locking.h"
#include "btree_update.h"
-struct btree_reserve {
- struct disk_reservation disk_res;
- unsigned nr;
- struct btree *b[BTREE_RESERVE_MAX];
-};
-
void __bch2_btree_calc_format(struct bkey_format_state *, struct btree *);
bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *,
struct bkey_format *);
-/* Btree node freeing/allocation: */
-
-/*
- * Tracks a btree node that has been (or is about to be) freed in memory, but
- * has _not_ yet been freed on disk (because the write that makes the new
- * node(s) visible and frees the old hasn't completed yet)
- */
-struct pending_btree_node_free {
- bool index_update_done;
-
- __le64 seq;
- enum btree_id btree_id;
- unsigned level;
- __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
-};
+#define BTREE_UPDATE_NODES_MAX ((BTREE_MAX_DEPTH - 2) * 2 + GC_MERGE_NODES)
-#define BTREE_UPDATE_JOURNAL_RES \
- ((BKEY_BTREE_PTR_U64s_MAX + 1) * (BTREE_MAX_DEPTH - 1) * 2)
+#define BTREE_UPDATE_JOURNAL_RES (BTREE_UPDATE_NODES_MAX * (BKEY_BTREE_PTR_U64s_MAX + 1))
/*
* Tracks an in progress split/rewrite of a btree node and the update to the
unsigned nodes_written:1;
enum btree_id btree_id;
- u8 level;
- struct btree_reserve *reserve;
+ struct disk_reservation disk_res;
struct journal_preres journal_preres;
/*
*/
struct journal_entry_pin journal;
- /*
- * Nodes being freed:
- * Protected by c->btree_node_pending_free_lock
- */
- struct pending_btree_node_free pending[BTREE_MAX_DEPTH + GC_MERGE_NODES];
- unsigned nr_pending;
+ /* Preallocated nodes we reserve when we start the update: */
+ struct btree *prealloc_nodes[BTREE_UPDATE_NODES_MAX];
+ unsigned nr_prealloc_nodes;
+
+ /* Nodes being freed: */
+ struct keylist old_keys;
+ u64 _old_keys[BTREE_UPDATE_NODES_MAX *
+ BKEY_BTREE_PTR_VAL_U64s_MAX];
+
+ /* Nodes being added: */
+ struct keylist new_keys;
+ u64 _new_keys[BTREE_UPDATE_NODES_MAX *
+ BKEY_BTREE_PTR_VAL_U64s_MAX];
/* New nodes, that will be made reachable by this update: */
- struct btree *new_nodes[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES];
+ struct btree *new_nodes[BTREE_UPDATE_NODES_MAX];
unsigned nr_new_nodes;
+ u8 open_buckets[BTREE_UPDATE_NODES_MAX *
+ BCH_REPLICAS_MAX];
+ u8 nr_open_buckets;
+
unsigned journal_u64s;
u64 journal_entries[BTREE_UPDATE_JOURNAL_RES];
u64 inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
};
-#define for_each_pending_btree_node_free(c, as, p) \
- list_for_each_entry(as, &c->btree_interior_update_list, list) \
- for (p = as->pending; p < as->pending + as->nr_pending; p++)
-
void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
struct btree_iter *);
void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
+void bch2_btree_update_get_open_buckets(struct btree_update *, struct btree *);
+
struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
struct btree *,
struct bkey_format);
void bch2_btree_interior_update_will_free_node(struct btree_update *,
struct btree *);
+void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
void bch2_btree_insert_node(struct btree_update *, struct btree *,
struct btree_iter *, struct keylist *,
size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
+void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
+struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
+ struct jset_entry *, struct jset_entry *);
+
+void bch2_fs_btree_interior_update_exit(struct bch_fs *);
+int bch2_fs_btree_interior_update_init(struct bch_fs *);
+
#endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */
goto err;
}
+ if (unlikely(trans->extra_journal_entry_u64s)) {
+ memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
+ trans->extra_journal_entries,
+ trans->extra_journal_entry_u64s);
+
+ trans->journal_res.offset += trans->extra_journal_entry_u64s;
+ trans->journal_res.u64s -= trans->extra_journal_entry_u64s;
+ }
+
/*
* Not allowed to fail after we've gotten our journal reservation - we
* have to use it:
bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
i->iter);
+ if (!ret && trans->journal_pin)
+ bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
+ trans->journal_pin, NULL);
+
/*
* Drop journal reservation after dropping write locks, since dropping
* the journal reservation may kick off a journal write:
memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
- trans->journal_u64s = 0;
+ trans->journal_u64s = trans->extra_journal_entry_u64s;
trans->journal_preres_u64s = 0;
if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
})
static int __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
- size_t b, enum bch_data_type type,
+ size_t b, enum bch_data_type data_type,
unsigned sectors, bool gc)
{
struct bucket *g = __bucket(ca, b, gc);
struct bucket_mark old, new;
bool overflow;
- BUG_ON(type != BCH_DATA_SB &&
- type != BCH_DATA_JOURNAL);
+ BUG_ON(data_type != BCH_DATA_SB &&
+ data_type != BCH_DATA_JOURNAL);
old = bucket_cmpxchg(g, new, ({
- new.data_type = type;
+ new.data_type = data_type;
overflow = checked_add(new.dirty_sectors, sectors);
}));
bch2_fs_inconsistent_on(old.data_type &&
- old.data_type != type, c,
+ old.data_type != data_type, c,
"different types of data in same bucket: %s, %s",
bch2_data_types[old.data_type],
- bch2_data_types[type]);
+ bch2_data_types[data_type]);
bch2_fs_inconsistent_on(overflow, c,
- "bucket sector count overflow: %u + %u > U16_MAX",
+ "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
+ ca->dev_idx, b, new.gen,
+ bch2_data_types[old.data_type ?: data_type],
old.dirty_sectors, sectors);
if (c)
}
}
-static bool bch2_mark_pointer(struct bch_fs *c,
- struct extent_ptr_decoded p,
- s64 sectors, enum bch_data_type data_type,
- struct bch_fs_usage *fs_usage,
- u64 journal_seq, unsigned flags)
+static int __mark_pointer(struct bch_fs *c, struct bkey_s_c k,
+ struct extent_ptr_decoded p,
+ s64 sectors, enum bch_data_type ptr_data_type,
+ u8 bucket_gen, u8 *bucket_data_type,
+ u16 *dirty_sectors, u16 *cached_sectors)
+{
+ u16 *dst_sectors = !p.ptr.cached
+ ? dirty_sectors
+ : cached_sectors;
+ u16 orig_sectors = *dst_sectors;
+ char buf[200];
+
+ if (gen_after(p.ptr.gen, bucket_gen)) {
+ bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
+ "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
+ "while marking %s",
+ p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
+ bucket_gen,
+ bch2_data_types[*bucket_data_type ?: ptr_data_type],
+ p.ptr.gen,
+ (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
+ return -EIO;
+ }
+
+ if (gen_cmp(bucket_gen, p.ptr.gen) >= 96U) {
+ bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
+ "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
+ "while marking %s",
+ p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
+ bucket_gen,
+ bch2_data_types[*bucket_data_type ?: ptr_data_type],
+ p.ptr.gen,
+ (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
+ return -EIO;
+ }
+
+ if (bucket_gen != p.ptr.gen && !p.ptr.cached) {
+ bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
+ "bucket %u:%zu gen %u data type %s: stale dirty ptr (gen %u)\n"
+ "while marking %s",
+ p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
+ bucket_gen,
+ bch2_data_types[*bucket_data_type ?: ptr_data_type],
+ p.ptr.gen,
+ (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
+ return -EIO;
+ }
+
+ if (bucket_gen != p.ptr.gen)
+ return 1;
+
+ if (*bucket_data_type && *bucket_data_type != ptr_data_type) {
+ bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
+ "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
+ "while marking %s",
+ p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
+ bucket_gen,
+ bch2_data_types[*bucket_data_type],
+ bch2_data_types[ptr_data_type],
+ (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
+ return -EIO;
+ }
+
+ if (checked_add(*dst_sectors, sectors)) {
+ bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
+ "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
+ "while marking %s",
+ p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
+ bucket_gen,
+ bch2_data_types[*bucket_data_type ?: ptr_data_type],
+ orig_sectors, sectors,
+ (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
+ return -EIO;
+ }
+
+ *bucket_data_type = *dirty_sectors || *cached_sectors
+ ? ptr_data_type : 0;
+ return 0;
+}
+
+static int bch2_mark_pointer(struct bch_fs *c, struct bkey_s_c k,
+ struct extent_ptr_decoded p,
+ s64 sectors, enum bch_data_type data_type,
+ struct bch_fs_usage *fs_usage,
+ u64 journal_seq, unsigned flags)
{
bool gc = flags & BTREE_TRIGGER_GC;
struct bucket_mark old, new;
struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
- bool overflow;
+ u8 bucket_data_type;
u64 v;
+ int ret;
v = atomic64_read(&g->_mark.v);
do {
new.v.counter = old.v.counter = v;
+ bucket_data_type = new.data_type;
- /*
- * Check this after reading bucket mark to guard against
- * the allocator invalidating a bucket after we've already
- * checked the gen
- */
- if (gen_after(p.ptr.gen, new.gen)) {
- bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
- "pointer gen in the future");
- return true;
- }
-
- if (new.gen != p.ptr.gen) {
- /* XXX write repair code for this */
- if (!p.ptr.cached &&
- test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
- bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
- "stale dirty pointer");
- return true;
- }
-
- if (!p.ptr.cached)
- overflow = checked_add(new.dirty_sectors, sectors);
- else
- overflow = checked_add(new.cached_sectors, sectors);
+ ret = __mark_pointer(c, k, p, sectors, data_type, new.gen,
+ &bucket_data_type,
+ &new.dirty_sectors,
+ &new.cached_sectors);
+ if (ret)
+ return ret;
- if (!new.dirty_sectors &&
- !new.cached_sectors) {
- new.data_type = 0;
+ new.data_type = bucket_data_type;
- if (journal_seq) {
- new.journal_seq_valid = 1;
- new.journal_seq = journal_seq;
- }
- } else {
- new.data_type = data_type;
+ if (journal_seq) {
+ new.journal_seq_valid = 1;
+ new.journal_seq = journal_seq;
}
if (flags & BTREE_TRIGGER_NOATOMIC) {
old.v.counter,
new.v.counter)) != old.v.counter);
- if (old.data_type && old.data_type != data_type)
- bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
- "bucket %u:%zu gen %u different types of data in same bucket: %s, %s",
- p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
- new.gen,
- bch2_data_types[old.data_type],
- bch2_data_types[data_type]);
-
- bch2_fs_inconsistent_on(overflow, c,
- "bucket sector count overflow: %u + %lli > U16_MAX",
- !p.ptr.cached
- ? old.dirty_sectors
- : old.cached_sectors, sectors);
-
bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
BUG_ON(!gc && bucket_became_unavailable(old, new));
- return false;
+ return 0;
}
static int bch2_mark_stripe_ptr(struct bch_fs *c,
struct extent_ptr_decoded p;
struct bch_replicas_padded r;
s64 dirty_sectors = 0;
+ bool stale;
int ret;
r.e.data_type = data_type;
s64 disk_sectors = data_type == BCH_DATA_BTREE
? sectors
: ptr_disk_sectors_delta(p, offset, sectors, flags);
- bool stale = bch2_mark_pointer(c, p, disk_sectors, data_type,
- fs_usage, journal_seq, flags);
+
+ ret = bch2_mark_pointer(c, k, p, disk_sectors, data_type,
+ fs_usage, journal_seq, flags);
+ if (ret < 0)
+ return ret;
+
+ stale = ret > 0;
if (p.ptr.cached) {
if (!stale)
return 0;
}
-int bch2_mark_key_locked(struct bch_fs *c,
+static int bch2_mark_key_locked(struct bch_fs *c,
struct bkey_s_c k,
unsigned offset, s64 sectors,
struct bch_fs_usage *fs_usage,
}
static int bch2_trans_mark_pointer(struct btree_trans *trans,
- struct extent_ptr_decoded p,
+ struct bkey_s_c k, struct extent_ptr_decoded p,
s64 sectors, enum bch_data_type data_type)
{
struct bch_fs *c = trans->c;
struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
struct btree_iter *iter;
- struct bkey_s_c k;
+ struct bkey_s_c k_a;
struct bkey_alloc_unpacked u;
struct bkey_i_alloc *a;
- u16 *dst_sectors, orig_sectors;
int ret;
ret = trans_get_key(trans, BTREE_ID_ALLOC,
POS(p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr)),
- &iter, &k);
+ &iter, &k_a);
if (ret < 0)
return ret;
- if (!ret && unlikely(!test_bit(BCH_FS_ALLOC_WRITTEN, &c->flags))) {
+ if (k_a.k->type != KEY_TYPE_alloc ||
+ (!ret && unlikely(!test_bit(BCH_FS_ALLOC_WRITTEN, &c->flags)))) {
/*
* During journal replay, and if gc repairs alloc info at
* runtime, the alloc info in the btree might not be up to date
- * yet - so, trust the in memory mark:
+ * yet - so, trust the in memory mark - unless we're already
+ * updating that key:
*/
struct bucket *g;
struct bucket_mark m;
u = alloc_mem_to_key(g, m);
percpu_up_read(&c->mark_lock);
} else {
- /*
- * Unless we're already updating that key:
- */
- if (k.k->type != KEY_TYPE_alloc) {
- bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
- "pointer to nonexistent bucket %llu:%llu",
- iter->pos.inode, iter->pos.offset);
- ret = -1;
- goto out;
- }
-
- u = bch2_alloc_unpack(k);
- }
-
- if (gen_after(u.gen, p.ptr.gen)) {
- ret = 1;
- goto out;
+ u = bch2_alloc_unpack(k_a);
}
- if (u.data_type && u.data_type != data_type) {
- bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
- "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s",
- iter->pos.inode, iter->pos.offset,
- u.gen,
- bch2_data_types[u.data_type],
- bch2_data_types[data_type]);
- ret = -1;
- goto out;
- }
-
- dst_sectors = !p.ptr.cached
- ? &u.dirty_sectors
- : &u.cached_sectors;
- orig_sectors = *dst_sectors;
-
- if (checked_add(*dst_sectors, sectors)) {
- bch2_fs_inconsistent(c,
- "bucket sector count overflow: %u + %lli > U16_MAX",
- orig_sectors, sectors);
- /* return an error indicating that we need full fsck */
- ret = -EIO;
+ ret = __mark_pointer(c, k, p, sectors, data_type, u.gen, &u.data_type,
+ &u.dirty_sectors, &u.cached_sectors);
+ if (ret)
goto out;
- }
-
- u.data_type = u.dirty_sectors || u.cached_sectors
- ? data_type : 0;
a = bch2_trans_kmalloc(trans, BKEY_ALLOC_U64s_MAX * 8);
ret = PTR_ERR_OR_ZERO(a);
? sectors
: ptr_disk_sectors_delta(p, offset, sectors, flags);
- ret = bch2_trans_mark_pointer(trans, p, disk_sectors,
+ ret = bch2_trans_mark_pointer(trans, k, p, disk_sectors,
data_type);
if (ret < 0)
return ret;
size_t, enum bch_data_type, unsigned,
struct gc_pos, unsigned);
-int bch2_mark_key_locked(struct bch_fs *, struct bkey_s_c, unsigned, s64,
- struct bch_fs_usage *, u64, unsigned);
int bch2_mark_key(struct bch_fs *, struct bkey_s_c, unsigned, s64,
struct bch_fs_usage *, u64, unsigned);
int bch2_fs_usage_apply(struct bch_fs *, struct bch_fs_usage *,
now = atomic_long_read(&clock->now);
for (i = 0; i < clock->timers.used; i++)
- pr_buf(&out, "%pf:\t%li\n",
+ pr_buf(&out, "%ps:\t%li\n",
clock->timers.data[i]->fn,
clock->timers.data[i]->expire - now);
spin_unlock(&clock->timer_lock);
#include "super-io.h"
#include <linux/lz4.h>
+#include <linux/sched/mm.h>
#include <linux/zlib.h>
#include <linux/zstd.h>
struct bbuf ret;
struct bio_vec bv;
struct bvec_iter iter;
- unsigned nr_pages = 0;
+ unsigned nr_pages = 0, flags;
struct page *stack_pages[16];
struct page **pages = NULL;
void *data;
__bio_for_each_segment(bv, bio, iter, start)
pages[nr_pages++] = bv.bv_page;
+ flags = memalloc_nofs_save();
data = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
+ memalloc_nofs_restore(flags);
+
if (pages != stack_pages)
kfree(pages);
}
if (!mempool_initialized(&c->decompress_workspace)) {
- ret = mempool_init_kmalloc_pool(
+ ret = mempool_init_kvpmalloc_pool(
&c->decompress_workspace,
1, decompress_workspace_size);
if (ret)
bch_scnmemcpy(out, d.v->d_name,
bch2_dirent_name_bytes(d));
- pr_buf(out, " -> %llu", d.v->d_inum);
+ pr_buf(out, " -> %llu type %u", d.v->d_inum, d.v->d_type);
}
static struct bkey_i_dirent *dirent_create_key(struct btree_trans *trans,
return ret;
}
-int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
+static int bch2_stripes_read_fn(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bkey_s_c k)
{
- struct btree_trans trans;
- struct btree_and_journal_iter iter;
- struct bkey_s_c k;
- int ret;
-
- ret = bch2_fs_ec_start(c);
- if (ret)
- return ret;
-
- bch2_trans_init(&trans, c, 0, 0);
-
- bch2_btree_and_journal_iter_init(&iter, &trans, journal_keys,
- BTREE_ID_EC, POS_MIN);
-
+ int ret = 0;
- while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
- bch2_mark_key(c, k, 0, 0, NULL, 0,
- BTREE_TRIGGER_ALLOC_READ|
- BTREE_TRIGGER_NOATOMIC);
+ if (k.k->type == KEY_TYPE_stripe)
+ ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL) ?:
+ bch2_mark_key(c, k, 0, 0, NULL, 0,
+ BTREE_TRIGGER_ALLOC_READ|
+ BTREE_TRIGGER_NOATOMIC);
- bch2_btree_and_journal_iter_advance(&iter);
- }
+ return ret;
+}
- ret = bch2_trans_exit(&trans) ?: ret;
- if (ret) {
+int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
+{
+ int ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_EC,
+ NULL, bch2_stripes_read_fn);
+ if (ret)
bch_err(c, "error reading stripes: %i", ret);
- return ret;
- }
- return 0;
+ return ret;
}
int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
return 0;
}
-int bch2_fs_ec_start(struct bch_fs *c)
-{
- return bch2_ec_mem_alloc(c, false);
-}
-
void bch2_fs_ec_exit(struct bch_fs *c)
{
struct ec_stripe_head *h;
int bch2_ec_mem_alloc(struct bch_fs *, bool);
-int bch2_fs_ec_start(struct bch_fs *);
-
void bch2_fs_ec_exit(struct bch_fs *);
int bch2_fs_ec_init(struct bch_fs *);
if (s->fmt == fmt)
goto found;
- s = kzalloc(sizeof(*s), GFP_KERNEL);
+ s = kzalloc(sizeof(*s), GFP_NOFS);
if (!s) {
if (!c->fsck_alloc_err)
bch_err(c, "kmalloc err, cannot ratelimit fsck errs");
#define FSCK_CAN_IGNORE (1 << 1)
#define FSCK_NEED_FSCK (1 << 2)
+__printf(3, 4) __cold
enum fsck_err_ret bch2_fsck_err(struct bch_fs *,
unsigned, const char *, ...);
void bch2_flush_fsck_errs(struct bch_fs *);
{
struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
- pr_buf(out, "seq %llu sectors %u written %u min_key ",
+ pr_buf(out, "seq %llx sectors %u written %u min_key ",
le64_to_cpu(bp.v->seq),
le16_to_cpu(bp.v->sectors),
le16_to_cpu(bp.v->sectors_written));
sectors = k.k->size - offset_into_extent;
ret = bch2_read_indirect_extent(trans,
- &offset_into_extent, sk.k);
+ &offset_into_extent, &sk);
if (ret)
break;
u64 aligned_len;
loff_t ret = 0;
+ if (!c->opts.reflink)
+ return -EOPNOTSUPP;
+
if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
return -EINVAL;
sectors = k.k->size - offset_into_extent;
ret = bch2_read_indirect_extent(&trans,
- &offset_into_extent, cur.k);
+ &offset_into_extent, &cur);
if (ret)
break;
}
if (!S_ISDIR(u->bi_mode) && link->dir_count) {
- need_fsck_err(c, "non directory with subdirectories",
+ need_fsck_err(c, "non directory with subdirectories (inum %llu)",
u->bi_inum);
return 0;
}
sectors = k.k->size - offset_into_extent;
ret = bch2_read_indirect_extent(&trans,
- &offset_into_extent, sk.k);
+ &offset_into_extent, &sk);
if (ret)
break;
int __bch2_read_indirect_extent(struct btree_trans *trans,
unsigned *offset_into_extent,
- struct bkey_i *orig_k)
+ struct bkey_on_stack *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)->v.idx) +
+ reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
*offset_into_extent;
iter = bch2_trans_get_iter(trans, BTREE_ID_REFLINK,
}
*offset_into_extent = iter->pos.offset - bkey_start_offset(k.k);
- bkey_reassemble(orig_k, k);
+ bkey_on_stack_reassemble(orig_k, trans->c, k);
err:
bch2_trans_iter_put(trans, iter);
return ret;
k = bkey_i_to_s_c(sk.k);
ret = bch2_read_indirect_extent(&trans,
- &offset_into_extent, sk.k);
+ &offset_into_extent, &sk);
if (ret)
goto err;
#define _BCACHEFS_IO_H
#include "checksum.h"
+#include "bkey_on_stack.h"
#include "io_types.h"
#define to_wbio(_bio) \
struct extent_ptr_decoded;
int __bch2_read_indirect_extent(struct btree_trans *, unsigned *,
- struct bkey_i *);
+ struct bkey_on_stack *);
static inline int bch2_read_indirect_extent(struct btree_trans *trans,
unsigned *offset_into_extent,
- struct bkey_i *k)
+ struct bkey_on_stack *k)
{
- return k->k.type == KEY_TYPE_reflink_p
+ return k->k->k.type == KEY_TYPE_reflink_p
? __bch2_read_indirect_extent(trans, offset_into_extent, k)
: 0;
}
void bch2_fs_journal_stop(struct journal *j)
{
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
-
bch2_journal_flush_all_pins(j);
wait_event(j->wait, journal_entry_close(j));
/* do we need to write another journal entry? */
- if (test_bit(JOURNAL_NOT_EMPTY, &j->flags) ||
- c->btree_roots_dirty)
+ if (test_bit(JOURNAL_NOT_EMPTY, &j->flags))
bch2_journal_meta(j);
journal_quiesce(j);
i, atomic_read(&pin_list->count));
list_for_each_entry(pin, &pin_list->list, list)
- pr_buf(&out, "\t%p %pf\n",
+ pr_buf(&out, "\t%px %ps\n",
pin, pin->flush);
if (!list_empty(&pin_list->flushed))
pr_buf(&out, "flushed:\n");
list_for_each_entry(pin, &pin_list->flushed, list)
- pr_buf(&out, "\t%p %pf\n",
+ pr_buf(&out, "\t%px %ps\n",
pin, pin->flush);
}
spin_unlock(&j->lock);
return entry;
}
+static inline struct jset_entry *
+journal_res_entry(struct journal *j, struct journal_res *res)
+{
+ return vstruct_idx(j->buf[res->idx].data, res->offset);
+}
+
+static inline unsigned journal_entry_set(struct jset_entry *entry, unsigned type,
+ enum btree_id id, unsigned level,
+ const void *data, unsigned u64s)
+{
+ memset(entry, 0, sizeof(*entry));
+ entry->u64s = cpu_to_le16(u64s);
+ entry->type = type;
+ entry->btree_id = id;
+ entry->level = level;
+ memcpy_u64s_small(entry->_data, data, u64s);
+
+ return jset_u64s(u64s);
+}
+
static inline void bch2_journal_add_entry(struct journal *j, struct journal_res *res,
unsigned type, enum btree_id id,
unsigned level,
const void *data, unsigned u64s)
{
- struct journal_buf *buf = &j->buf[res->idx];
- struct jset_entry *entry = vstruct_idx(buf->data, res->offset);
- unsigned actual = jset_u64s(u64s);
+ unsigned actual = journal_entry_set(journal_res_entry(j, res),
+ type, id, level, data, u64s);
EBUG_ON(!res->ref);
EBUG_ON(actual > res->u64s);
res->offset += actual;
res->u64s -= actual;
-
- memset(entry, 0, sizeof(*entry));
- entry->u64s = cpu_to_le16(u64s);
- entry->type = type;
- entry->btree_id = id;
- entry->level = level;
- memcpy_u64s(entry->_data, data, u64s);
}
static inline void bch2_journal_add_keys(struct journal *j, struct journal_res *res,
#include "bcachefs.h"
#include "alloc_foreground.h"
#include "btree_io.h"
+#include "btree_update_interior.h"
#include "buckets.h"
#include "checksum.h"
#include "error.h"
j->write_start_time = local_clock();
- start = vstruct_last(jset);
- end = bch2_journal_super_entries_add_common(c, start,
+ /*
+ * New btree roots are set by journalling them; when the journal entry
+ * gets written we have to propagate them to c->btree_roots
+ *
+ * But, every journal entry we write has to contain all the btree roots
+ * (at least for now); so after we copy btree roots to c->btree_roots we
+ * have to get any missing btree roots and add them to this journal
+ * entry:
+ */
+
+ bch2_journal_entries_to_btree_roots(c, jset);
+
+ start = end = vstruct_last(jset);
+
+ end = bch2_btree_roots_to_journal_entries(c, jset->start, end);
+
+ end = bch2_journal_super_entries_add_common(c, end,
le64_to_cpu(jset->seq));
u64s = (u64 *) end - (u64 *) start;
BUG_ON(u64s > j->entry_u64s_reserved);
__journal_pin_drop(j, pin);
- BUG_ON(!atomic_read(&pin_list->count));
+ BUG_ON(!atomic_read(&pin_list->count) && seq == journal_last_seq(j));
atomic_inc(&pin_list->count);
pin->seq = seq;
return ret;
}
-static void journal_flush_pins(struct journal *j, u64 seq_to_flush,
+/* returns true if we did work */
+static bool journal_flush_pins(struct journal *j, u64 seq_to_flush,
unsigned min_nr)
{
struct journal_entry_pin *pin;
+ bool ret = false;
u64 seq;
lockdep_assert_held(&j->reclaim_lock);
BUG_ON(j->flush_in_progress != pin);
j->flush_in_progress = NULL;
wake_up(&j->pin_flush_wait);
+ ret = true;
}
+
+ return ret;
}
/**
mutex_unlock(&j->reclaim_lock);
}
-static int journal_flush_done(struct journal *j, u64 seq_to_flush)
+static int journal_flush_done(struct journal *j, u64 seq_to_flush,
+ bool *did_work)
{
int ret;
mutex_lock(&j->reclaim_lock);
- journal_flush_pins(j, seq_to_flush, 0);
+ *did_work = journal_flush_pins(j, seq_to_flush, 0);
spin_lock(&j->lock);
/*
return ret;
}
-void bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
+bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
{
+ bool did_work = false;
+
if (!test_bit(JOURNAL_STARTED, &j->flags))
- return;
+ return false;
+
+ closure_wait_event(&j->async_wait,
+ journal_flush_done(j, seq_to_flush, &did_work));
- closure_wait_event(&j->async_wait, journal_flush_done(j, seq_to_flush));
+ return did_work;
}
int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
struct journal_entry_pin *pin,
journal_pin_flush_fn flush_fn)
{
- if (unlikely(!journal_pin_active(pin)))
+ if (unlikely(!journal_pin_active(pin) || pin->seq > seq))
__bch2_journal_pin_add(j, seq, pin, flush_fn);
}
void bch2_journal_reclaim(struct journal *);
void bch2_journal_reclaim_work(struct work_struct *);
-void bch2_journal_flush_pins(struct journal *, u64);
+bool bch2_journal_flush_pins(struct journal *, u64);
-static inline void bch2_journal_flush_all_pins(struct journal *j)
+static inline bool bch2_journal_flush_all_pins(struct journal *j)
{
- bch2_journal_flush_pins(j, U64_MAX);
+ return bch2_journal_flush_pins(j, U64_MAX);
}
int bch2_journal_flush_device_pins(struct journal *, int);
int bch2_keylist_realloc(struct keylist *l, u64 *inline_u64s,
size_t nr_inline_u64s, size_t new_u64s)
{
- size_t oldsize = bch_keylist_u64s(l);
+ size_t oldsize = bch2_keylist_u64s(l);
size_t newsize = oldsize + new_u64s;
u64 *old_buf = l->keys_p == inline_u64s ? NULL : l->keys_p;
u64 *new_keys;
memmove_u64s_down(l->keys,
bkey_next(l->keys),
- bch_keylist_u64s(l));
+ bch2_keylist_u64s(l));
}
#ifdef CONFIG_BCACHEFS_DEBUG
return l->top == l->keys;
}
-static inline size_t bch_keylist_u64s(struct keylist *l)
+static inline size_t bch2_keylist_u64s(struct keylist *l)
{
return l->top_p - l->keys_p;
}
static inline size_t bch2_keylist_bytes(struct keylist *l)
{
- return bch_keylist_u64s(l) * sizeof(u64);
+ return bch2_keylist_u64s(l) * sizeof(u64);
}
static inline struct bkey_i *bch2_keylist_front(struct keylist *l)
}
/* flush relevant btree updates */
- while (1) {
- closure_wait_event(&c->btree_interior_update_wait,
- !bch2_btree_interior_updates_nr_pending(c) ||
- c->btree_roots_dirty);
- if (c->btree_roots_dirty)
- bch2_journal_meta(&c->journal);
- if (!bch2_btree_interior_updates_nr_pending(c))
- break;
- }
+ closure_wait_event(&c->btree_interior_update_wait,
+ !bch2_btree_interior_updates_nr_pending(c));
ret = 0;
err:
ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
- while (1) {
- closure_wait_event(&c->btree_interior_update_wait,
- !bch2_btree_interior_updates_nr_pending(c) ||
- c->btree_roots_dirty);
- if (!bch2_btree_interior_updates_nr_pending(c))
- break;
- bch2_journal_meta(&c->journal);
- }
+ closure_wait_event(&c->btree_interior_update_wait,
+ !bch2_btree_interior_updates_nr_pending(c));
ret = bch2_replicas_gc2(c) ?: ret;
OPT_BOOL(), \
BCH_SB_PRJQUOTA, false, \
NULL, "Enable project quotas") \
+ x(reflink, u8, \
+ OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME, \
+ OPT_BOOL(), \
+ BCH_SB_REFLINK, true, \
+ NULL, "Enable reflink support") \
x(degraded, u8, \
OPT_MOUNT, \
OPT_BOOL(), \
b->btree_id, b->level, b->data->min_key);
}
+/* Walk btree, overlaying keys from the journal: */
+
+static int bch2_btree_and_journal_walk_recurse(struct bch_fs *c, struct btree *b,
+ struct journal_keys *journal_keys,
+ enum btree_id btree_id,
+ btree_walk_node_fn node_fn,
+ btree_walk_key_fn key_fn)
+{
+ struct btree_and_journal_iter iter;
+ struct bkey_s_c k;
+ int ret = 0;
+
+ bch2_btree_and_journal_iter_init_node_iter(&iter, journal_keys, b);
+
+ while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
+ ret = key_fn(c, btree_id, b->level, k);
+ if (ret)
+ break;
+
+ if (b->level) {
+ struct btree *child;
+ BKEY_PADDED(k) tmp;
+
+ bkey_reassemble(&tmp.k, k);
+ k = bkey_i_to_s_c(&tmp.k);
+
+ bch2_btree_and_journal_iter_advance(&iter);
+
+ if (b->level > 0) {
+ child = bch2_btree_node_get_noiter(c, &tmp.k,
+ b->btree_id, b->level - 1);
+ ret = PTR_ERR_OR_ZERO(child);
+ if (ret)
+ break;
+
+ ret = (node_fn ? node_fn(c, b) : 0) ?:
+ bch2_btree_and_journal_walk_recurse(c, child,
+ journal_keys, btree_id, node_fn, key_fn);
+ six_unlock_read(&child->lock);
+
+ if (ret)
+ break;
+ }
+ } else {
+ bch2_btree_and_journal_iter_advance(&iter);
+ }
+ }
+
+ return ret;
+}
+
+int bch2_btree_and_journal_walk(struct bch_fs *c, struct journal_keys *journal_keys,
+ enum btree_id btree_id,
+ btree_walk_node_fn node_fn,
+ btree_walk_key_fn key_fn)
+{
+ struct btree *b = c->btree_roots[btree_id].b;
+ int ret = 0;
+
+ if (btree_node_fake(b))
+ return 0;
+
+ six_lock_read(&b->lock);
+ ret = (node_fn ? node_fn(c, b) : 0) ?:
+ bch2_btree_and_journal_walk_recurse(c, b, journal_keys, btree_id,
+ node_fn, key_fn) ?:
+ key_fn(c, btree_id, b->level + 1, bkey_i_to_s_c(&b->key));
+ six_unlock_read(&b->lock);
+
+ return ret;
+}
+
/* sort and dedup all keys in the journal: */
void bch2_journal_entries_free(struct list_head *list)
"superblock read clock doesn't match journal after clean shutdown");
for (i = 0; i < BTREE_ID_NR; i++) {
+ char buf1[200], buf2[200];
struct bkey_i *k1, *k2;
unsigned l1 = 0, l2 = 0;
k1->k.u64s != k2->k.u64s ||
memcmp(k1, k2, bkey_bytes(k1)) ||
l1 != l2, c,
- "superblock btree root doesn't match journal after clean shutdown");
+ "superblock btree root %u doesn't match journal after clean shutdown\n"
+ "sb: l=%u %s\n"
+ "journal: l=%u %s\n", i,
+ l1, (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(k1)), buf1),
+ l2, (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(k2)), buf2));
}
fsck_err:
return ret;
bch2_mark_dev_superblock(c, ca, 0);
mutex_unlock(&c->sb_lock);
+ mutex_lock(&c->sb_lock);
+ c->disk_sb.sb->version = c->disk_sb.sb->version_min =
+ le16_to_cpu(bcachefs_metadata_version_current);
+ c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_atomic_nlink;
+ c->disk_sb.sb->features[0] |= BCH_SB_FEATURES_ALL;
+
+ bch2_write_super(c);
+ mutex_unlock(&c->sb_lock);
+
set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
goto err;
mutex_lock(&c->sb_lock);
- c->disk_sb.sb->version = c->disk_sb.sb->version_min =
- le16_to_cpu(bcachefs_metadata_version_current);
- c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_atomic_nlink;
- c->disk_sb.sb->features[0] |= BCH_SB_FEATURES_ALL;
-
SET_BCH_SB_INITIALIZED(c->disk_sb.sb, true);
SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
struct journal_keys *,
struct btree *);
+typedef int (*btree_walk_node_fn)(struct bch_fs *c, struct btree *b);
+typedef int (*btree_walk_key_fn)(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bkey_s_c k);
+
+int bch2_btree_and_journal_walk(struct bch_fs *, struct journal_keys *, enum btree_id,
+ btree_walk_node_fn, btree_walk_key_fn);
+
void bch2_journal_keys_free(struct journal_keys *);
void bch2_journal_entries_free(struct list_head *);
u64 src_done, dst_done;
int ret = 0, ret2 = 0;
+ if (!c->opts.reflink)
+ return -EOPNOTSUPP;
+
if (!percpu_ref_tryget(&c->writes))
return -EROFS;
// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
+#include "btree_update_interior.h"
#include "buckets.h"
#include "checksum.h"
#include "disk_groups.h"
mutex_lock(&c->sb_lock);
SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
- c->disk_sb.sb->compat[0] &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA);
c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_new_extent_overwrite;
c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_extents_above_btree_updates;
c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_btree_updates_journalled;
struct jset_entry *entry,
u64 journal_seq)
{
- struct btree_root *r;
unsigned i;
- mutex_lock(&c->btree_root_lock);
-
- for (r = c->btree_roots;
- r < c->btree_roots + BTREE_ID_NR;
- r++)
- if (r->alive) {
- entry_init_u64s(entry, r->key.u64s + 1);
- entry->btree_id = r - c->btree_roots;
- entry->level = r->level;
- entry->type = BCH_JSET_ENTRY_btree_root;
- bkey_copy(&entry->start[0], &r->key);
-
- entry = vstruct_next(entry);
- }
- c->btree_roots_dirty = false;
-
- mutex_unlock(&c->btree_root_lock);
-
percpu_down_write(&c->mark_lock);
if (!journal_seq) {
entry = sb_clean->start;
entry = bch2_journal_super_entries_add_common(c, entry, 0);
+ entry = bch2_btree_roots_to_journal_entries(c, entry, entry);
BUG_ON((void *) entry > vstruct_end(&sb_clean->field));
memset(entry, 0,
static void __bch2_fs_read_only(struct bch_fs *c)
{
struct bch_dev *ca;
- bool wrote;
+ bool wrote = false;
unsigned i, clean_passes = 0;
int ret;
*/
bch2_journal_flush_all_pins(&c->journal);
+ /*
+ * If the allocator threads didn't all start up, the btree updates to
+ * write out alloc info aren't going to work:
+ */
if (!test_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags))
- goto allocator_not_running;
+ goto nowrote_alloc;
- do {
- wrote = false;
+ bch_verbose(c, "writing alloc info");
+ /*
+ * This should normally just be writing the bucket read/write clocks:
+ */
+ ret = bch2_stripes_write(c, BTREE_INSERT_NOCHECK_RW, &wrote) ?:
+ bch2_alloc_write(c, BTREE_INSERT_NOCHECK_RW, &wrote);
+ bch_verbose(c, "writing alloc info complete");
- ret = bch2_stripes_write(c, BTREE_INSERT_NOCHECK_RW, &wrote) ?:
- bch2_alloc_write(c, BTREE_INSERT_NOCHECK_RW, &wrote);
+ if (ret && !test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
+ bch2_fs_inconsistent(c, "error writing out alloc info %i", ret);
- if (ret && !test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
- bch2_fs_inconsistent(c, "error writing out alloc info %i", ret);
+ if (ret)
+ goto nowrote_alloc;
- if (ret)
- break;
+ bch_verbose(c, "flushing journal and stopping allocators");
- for_each_member_device(ca, c, i)
- bch2_dev_allocator_quiesce(c, ca);
+ bch2_journal_flush_all_pins(&c->journal);
+ set_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags);
- bch2_journal_flush_all_pins(&c->journal);
+ do {
+ clean_passes++;
+
+ if (bch2_journal_flush_all_pins(&c->journal))
+ clean_passes = 0;
/*
- * We need to explicitly wait on btree interior updates to complete
- * before stopping the journal, flushing all journal pins isn't
- * sufficient, because in the BTREE_INTERIOR_UPDATING_ROOT case btree
- * interior updates have to drop their journal pin before they're
- * fully complete:
+ * In flight interior btree updates will generate more journal
+ * updates and btree updates (alloc btree):
*/
- closure_wait_event(&c->btree_interior_update_wait,
- !bch2_btree_interior_updates_nr_pending(c));
+ if (bch2_btree_interior_updates_nr_pending(c)) {
+ closure_wait_event(&c->btree_interior_update_wait,
+ !bch2_btree_interior_updates_nr_pending(c));
+ clean_passes = 0;
+ }
+ flush_work(&c->btree_interior_update_work);
- clean_passes = wrote ? 0 : clean_passes + 1;
+ if (bch2_journal_flush_all_pins(&c->journal))
+ clean_passes = 0;
} while (clean_passes < 2);
-allocator_not_running:
+ bch_verbose(c, "flushing journal and stopping allocators complete");
+
+ set_bit(BCH_FS_ALLOC_CLEAN, &c->flags);
+nowrote_alloc:
+ closure_wait_event(&c->btree_interior_update_wait,
+ !bch2_btree_interior_updates_nr_pending(c));
+ flush_work(&c->btree_interior_update_work);
+
for_each_member_device(ca, c, i)
bch2_dev_allocator_stop(ca);
clear_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
+ clear_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags);
bch2_fs_journal_stop(&c->journal);
- /* XXX: mark super that alloc info is persistent */
-
/*
* the journal kicks off btree writes via reclaim - wait for in flight
* writes after stopping journal:
!test_bit(BCH_FS_ERROR, &c->flags) &&
!test_bit(BCH_FS_EMERGENCY_RO, &c->flags) &&
test_bit(BCH_FS_STARTED, &c->flags) &&
- !c->opts.norecovery)
+ test_bit(BCH_FS_ALLOC_CLEAN, &c->flags) &&
+ !c->opts.norecovery) {
+ bch_verbose(c, "marking filesystem clean");
bch2_fs_mark_clean(c);
+ }
clear_bit(BCH_FS_RW, &c->flags);
}
if (ret)
goto err;
+ clear_bit(BCH_FS_ALLOC_CLEAN, &c->flags);
+
for_each_rw_member(ca, c, i)
bch2_dev_allocator_add(c, ca);
bch2_recalc_capacity(c);
bch2_fs_ec_exit(c);
bch2_fs_encryption_exit(c);
bch2_fs_io_exit(c);
+ bch2_fs_btree_interior_update_exit(c);
bch2_fs_btree_iter_exit(c);
bch2_fs_btree_cache_exit(c);
bch2_fs_journal_exit(&c->journal);
mempool_exit(&c->large_bkey_pool);
mempool_exit(&c->btree_bounce_pool);
bioset_exit(&c->btree_bio);
- mempool_exit(&c->btree_interior_update_pool);
- mempool_exit(&c->btree_reserve_pool);
mempool_exit(&c->fill_iter);
percpu_ref_exit(&c->writes);
kfree(c->replicas.entries);
INIT_LIST_HEAD(&c->list);
- INIT_LIST_HEAD(&c->btree_interior_update_list);
- INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
- mutex_init(&c->btree_reserve_cache_lock);
- mutex_init(&c->btree_interior_update_lock);
-
mutex_init(&c->usage_scratch_lock);
mutex_init(&c->bio_bounce_pages_lock);
WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
percpu_ref_init(&c->writes, bch2_writes_disabled,
PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
- mempool_init_kmalloc_pool(&c->btree_reserve_pool, 1,
- sizeof(struct btree_reserve)) ||
- mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
- sizeof(struct btree_update)) ||
mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
bioset_init(&c->btree_bio, 1,
max(offsetof(struct btree_read_bio, bio),
bch2_fs_replicas_init(c) ||
bch2_fs_btree_cache_init(c) ||
bch2_fs_btree_iter_init(c) ||
+ bch2_fs_btree_interior_update_init(c) ||
bch2_fs_io_init(c) ||
bch2_fs_encryption_init(c) ||
bch2_fs_compress_init(c) ||
read_attribute(journal_pins);
read_attribute(btree_updates);
read_attribute(dirty_btree_nodes);
+read_attribute(btree_transactions);
read_attribute(internal_uuid);
if (attr == &sysfs_dirty_btree_nodes)
return bch2_dirty_btree_nodes_print(c, buf);
+ if (attr == &sysfs_btree_transactions) {
+ struct printbuf out = _PBUF(buf, PAGE_SIZE);
+
+ bch2_btree_trans_to_text(&out, c);
+ return out.pos - buf;
+ }
if (attr == &sysfs_compression_stats)
return bch2_compression_stats(c, buf);
&sysfs_journal_pins,
&sysfs_btree_updates,
&sysfs_dirty_btree_nodes,
+ &sysfs_btree_transactions,
&sysfs_read_realloc_races,
&sysfs_extent_migrate_done,
if (!do_six_trylock_type(lock, type))
return false;
- six_acquire(&lock->dep_map, 1);
+ if (type != SIX_LOCK_write)
+ six_acquire(&lock->dep_map, 1);
return true;
}
old.v + l[type].lock_val)) != old.v);
six_set_owner(lock, type, old);
- six_acquire(&lock->dep_map, 1);
+ if (type != SIX_LOCK_write)
+ six_acquire(&lock->dep_map, 1);
return true;
}
__always_inline
static void __six_lock_type(struct six_lock *lock, enum six_lock_type type)
{
- six_acquire(&lock->dep_map, 0);
+ if (type != SIX_LOCK_write)
+ six_acquire(&lock->dep_map, 0);
if (!do_six_trylock_type(lock, type))
__six_lock_type_slowpath(lock, type);
EBUG_ON(type == SIX_LOCK_write &&
!(lock->state.v & __SIX_LOCK_HELD_intent));
- six_release(&lock->dep_map);
+ if (type != SIX_LOCK_write)
+ six_release(&lock->dep_map);
if (type == SIX_LOCK_intent) {
EBUG_ON(lock->owner != current);
#include <linux/workqueue.h>
static pthread_mutex_t wq_lock = PTHREAD_MUTEX_INITIALIZER;
+static pthread_cond_t work_finished = PTHREAD_COND_INITIALIZER;
static LIST_HEAD(wq_list);
struct workqueue_struct {
struct work_struct *current_work;
struct list_head pending_work;
- pthread_cond_t work_finished;
-
struct task_struct *worker;
char name[24];
};
WORK_PENDING_BIT,
};
+static bool work_pending(struct work_struct *work)
+{
+ return test_bit(WORK_PENDING_BIT, work_data_bits(work));
+}
+
static void clear_work_pending(struct work_struct *work)
{
clear_bit(WORK_PENDING_BIT, work_data_bits(work));
static void __queue_work(struct workqueue_struct *wq,
struct work_struct *work)
{
- BUG_ON(!test_bit(WORK_PENDING_BIT, work_data_bits(work)));
+ BUG_ON(!work_pending(work));
BUG_ON(!list_empty(&work->entry));
list_add_tail(&work->entry, &wq->pending_work);
goto retry;
}
-static bool __flush_work(struct work_struct *work)
+static bool work_running(struct work_struct *work)
{
struct workqueue_struct *wq;
- bool ret = false;
-retry:
+
list_for_each_entry(wq, &wq_list, list)
- if (wq->current_work == work) {
- pthread_cond_wait(&wq->work_finished, &wq_lock);
- ret = true;
- goto retry;
- }
+ if (wq->current_work == work)
+ return true;
+
+ return false;
+}
+
+bool flush_work(struct work_struct *work)
+{
+ bool ret = false;
+
+ pthread_mutex_lock(&wq_lock);
+ while (work_pending(work) || work_running(work)) {
+ pthread_cond_wait(&work_finished, &wq_lock);
+ ret = true;
+ }
+ pthread_mutex_unlock(&wq_lock);
+
+ return ret;
+}
+
+static bool __flush_work(struct work_struct *work)
+{
+ bool ret = false;
+
+ while (work_running(work)) {
+ pthread_cond_wait(&work_finished, &wq_lock);
+ ret = true;
+ }
return ret;
}
continue;
}
- BUG_ON(!test_bit(WORK_PENDING_BIT, work_data_bits(work)));
+ BUG_ON(!work_pending(work));
list_del_init(&work->entry);
clear_work_pending(work);
work->func(work);
pthread_mutex_lock(&wq_lock);
- pthread_cond_broadcast(&wq->work_finished);
+ pthread_cond_broadcast(&work_finished);
}
pthread_mutex_unlock(&wq_lock);
INIT_LIST_HEAD(&wq->list);
INIT_LIST_HEAD(&wq->pending_work);
- pthread_cond_init(&wq->work_finished, NULL);
-
va_start(args, max_active);
vsnprintf(wq->name, sizeof(wq->name), fmt, args);
va_end(args);
projects / bcachefs-tools-debian / commitdiff