summaryrefslogtreecommitdiff
path: root/fs/bcachefs/data_update.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/bcachefs/data_update.c')
-rw-r--r--fs/bcachefs/data_update.c83
1 files changed, 58 insertions, 25 deletions
diff --git a/fs/bcachefs/data_update.c b/fs/bcachefs/data_update.c
index 4150feca42a2..0d807c2ce9c6 100644
--- a/fs/bcachefs/data_update.c
+++ b/fs/bcachefs/data_update.c
@@ -14,6 +14,7 @@
#include "move.h"
#include "nocow_locking.h"
#include "rebalance.h"
+#include "snapshot.h"
#include "subvolume.h"
#include "trace.h"
@@ -105,7 +106,7 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
bch2_trans_iter_init(trans, &iter, m->btree_id,
bkey_start_pos(&bch2_keylist_front(keys)->k),
- BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+ BTREE_ITER_slots|BTREE_ITER_intent);
while (1) {
struct bkey_s_c k;
@@ -202,6 +203,8 @@ restart_drop_conflicting_replicas:
/* Now, drop excess replicas: */
restart_drop_extra_replicas:
+
+ rcu_read_lock();
bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
@@ -213,6 +216,7 @@ restart_drop_extra_replicas:
goto restart_drop_extra_replicas;
}
}
+ rcu_read_unlock();
/* Finally, add the pointers we just wrote: */
extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
@@ -287,7 +291,7 @@ restart_drop_extra_replicas:
k.k->p, insert->k.p) ?:
bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
bch2_trans_update(trans, &iter, insert,
- BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
+ BTREE_UPDATE_internal_snapshot_node) ?:
bch2_trans_commit(trans, &op->res,
NULL,
BCH_TRANS_COMMIT_no_check_rw|
@@ -356,10 +360,11 @@ void bch2_data_update_exit(struct data_update *update)
bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
bkey_for_each_ptr(ptrs, ptr) {
+ struct bch_dev *ca = bch2_dev_have_ref(c, ptr->dev);
if (c->opts.nocow_enabled)
bch2_bucket_nocow_unlock(&c->nocow_locks,
- PTR_BUCKET_POS(c, ptr), 0);
- percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
+ PTR_BUCKET_POS(ca, ptr), 0);
+ bch2_dev_put(ca);
}
bch2_bkey_buf_exit(&update->k, c);
@@ -385,8 +390,10 @@ static void bch2_update_unwritten_extent(struct btree_trans *trans,
while (bio_sectors(bio)) {
unsigned sectors = bio_sectors(bio);
+ bch2_trans_begin(trans);
+
bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
- BTREE_ITER_SLOTS);
+ BTREE_ITER_slots);
ret = lockrestart_do(trans, ({
k = bch2_btree_iter_peek_slot(&iter);
bkey_err(k);
@@ -464,7 +471,6 @@ int bch2_extent_drop_ptrs(struct btree_trans *trans,
while (data_opts.kill_ptrs) {
unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
- struct bch_extent_ptr *ptr;
bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
data_opts.kill_ptrs ^= 1U << drop;
@@ -479,15 +485,15 @@ int bch2_extent_drop_ptrs(struct btree_trans *trans,
/*
* Since we're not inserting through an extent iterator
- * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
+ * (BTREE_ITER_all_snapshots iterators aren't extent iterators),
* we aren't using the extent overwrite path to delete, we're
* just using the normal key deletion path:
*/
- if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_IS_EXTENTS))
+ if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_is_extents))
n->k.size = 0;
return bch2_trans_relock(trans) ?:
- bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
+ bch2_trans_update(trans, iter, n, BTREE_UPDATE_internal_snapshot_node) ?:
bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
}
@@ -509,6 +515,14 @@ int bch2_data_update_init(struct btree_trans *trans,
unsigned ptrs_locked = 0;
int ret = 0;
+ /*
+ * fs is corrupt we have a key for a snapshot node that doesn't exist,
+ * and we have to check for this because we go rw before repairing the
+ * snapshots table - just skip it, we can move it later.
+ */
+ if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
+ return -BCH_ERR_data_update_done;
+
bch2_bkey_buf_init(&m->k);
bch2_bkey_buf_reassemble(&m->k, c, k);
m->btree_id = btree_id;
@@ -530,15 +544,26 @@ int bch2_data_update_init(struct btree_trans *trans,
m->op.compression_opt = background_compression(io_opts);
m->op.watermark = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
- bkey_for_each_ptr(ptrs, ptr)
- percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
+ bkey_for_each_ptr(ptrs, ptr) {
+ if (!bch2_dev_tryget(c, ptr->dev)) {
+ bkey_for_each_ptr(ptrs, ptr2) {
+ if (ptr2 == ptr)
+ break;
+ bch2_dev_put(bch2_dev_have_ref(c, ptr2->dev));
+ }
+ return -BCH_ERR_data_update_done;
+ }
+ }
unsigned durability_have = 0, durability_removing = 0;
i = 0;
bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+ struct bch_dev *ca = bch2_dev_have_ref(c, p.ptr.dev);
+ struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
bool locked;
+ rcu_read_lock();
if (((1U << i) & m->data_opts.rewrite_ptrs)) {
BUG_ON(p.ptr.cached);
@@ -552,6 +577,7 @@ int bch2_data_update_init(struct btree_trans *trans,
bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
durability_have += bch2_extent_ptr_durability(c, &p);
}
+ rcu_read_unlock();
/*
* op->csum_type is normally initialized from the fs/file's
@@ -570,16 +596,13 @@ int bch2_data_update_init(struct btree_trans *trans,
if (ctxt) {
move_ctxt_wait_event(ctxt,
(locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
- PTR_BUCKET_POS(c, &p.ptr), 0)) ||
- (!atomic_read(&ctxt->read_sectors) &&
- !atomic_read(&ctxt->write_sectors)));
+ bucket, 0)) ||
+ list_empty(&ctxt->ios));
if (!locked)
- bch2_bucket_nocow_lock(&c->nocow_locks,
- PTR_BUCKET_POS(c, &p.ptr), 0);
+ bch2_bucket_nocow_lock(&c->nocow_locks, bucket, 0);
} else {
- if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
- PTR_BUCKET_POS(c, &p.ptr), 0)) {
+ if (!bch2_bucket_nocow_trylock(&c->nocow_locks, bucket, 0)) {
ret = -BCH_ERR_nocow_lock_blocked;
goto err;
}
@@ -590,6 +613,8 @@ int bch2_data_update_init(struct btree_trans *trans,
i++;
}
+ unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
+
/*
* If current extent durability is less than io_opts.data_replicas,
* we're not trying to rereplicate the extent up to data_replicas here -
@@ -599,7 +624,7 @@ int bch2_data_update_init(struct btree_trans *trans,
* rereplicate, currently, so that users don't get an unexpected -ENOSPC
*/
if (!(m->data_opts.write_flags & BCH_WRITE_CACHED) &&
- durability_have >= io_opts.data_replicas) {
+ !durability_required) {
m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
m->data_opts.rewrite_ptrs = 0;
/* if iter == NULL, it's just a promote */
@@ -608,11 +633,18 @@ int bch2_data_update_init(struct btree_trans *trans,
goto done;
}
- m->op.nr_replicas = min(durability_removing, io_opts.data_replicas - durability_have) +
+ m->op.nr_replicas = min(durability_removing, durability_required) +
m->data_opts.extra_replicas;
- m->op.nr_replicas_required = m->op.nr_replicas;
- BUG_ON(!m->op.nr_replicas);
+ /*
+ * If device(s) were set to durability=0 after data was written to them
+ * we can end up with a duribilty=0 extent, and the normal algorithm
+ * that tries not to increase durability doesn't work:
+ */
+ if (!(durability_have + durability_removing))
+ m->op.nr_replicas = max((unsigned) m->op.nr_replicas, 1);
+
+ m->op.nr_replicas_required = m->op.nr_replicas;
if (reserve_sectors) {
ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
@@ -632,10 +664,11 @@ int bch2_data_update_init(struct btree_trans *trans,
err:
i = 0;
bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+ struct bch_dev *ca = bch2_dev_have_ref(c, p.ptr.dev);
+ struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
if ((1U << i) & ptrs_locked)
- bch2_bucket_nocow_unlock(&c->nocow_locks,
- PTR_BUCKET_POS(c, &p.ptr), 0);
- percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
+ bch2_bucket_nocow_unlock(&c->nocow_locks, bucket, 0);
+ bch2_dev_put(ca);
i++;
}
generated by cgit v1.2.3 (git 2.39.0) at 2024-09-02 05:09:30 +0300