bcachefs: Throttle updates when btree key cache is too dirty

This is needed to ensure we don't deadlock because journal reclaim and
thus memory reclaim isn't making forward progress.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

diff --git a/fs/bcachefs/btree_key_cache.h b/fs/bcachefs/btree_key_cache.h
index 7723a2178430f1b95c4bd997e7522815e0507b2e..d7d31a0662c366dafb431d8706c916f17098d51a 100644 (file)
--- a/fs/bcachefs/btree_key_cache.h
+++ b/fs/bcachefs/btree_key_cache.h
@@ -5,11 +5,20 @@ static inline size_t bch2_nr_btree_keys_need_flush(struct bch_fs *c)
 {
        size_t nr_dirty = READ_ONCE(c->btree_key_cache.nr_dirty);
        size_t nr_keys = READ_ONCE(c->btree_key_cache.nr_dirty);
-       size_t max_dirty = 1024 + (nr_keys * 3) / 4;
+       size_t max_dirty = 4096 + nr_keys  / 2;
 
        return max_t(ssize_t, 0, nr_dirty - max_dirty);
 }
 
+static inline bool bch2_btree_key_cache_must_wait(struct bch_fs *c)
+{
+       size_t nr_dirty = READ_ONCE(c->btree_key_cache.nr_dirty);
+       size_t nr_keys = READ_ONCE(c->btree_key_cache.nr_dirty);
+       size_t max_dirty = 4096 + (nr_keys * 3) / 4;
+
+       return nr_dirty > max_dirty;
+}
+
 struct bkey_cached *
 bch2_btree_key_cache_find(struct bch_fs *, enum btree_id, struct bpos);
 

diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index bf2fc979a2eb96e0ec2476d986f292f68a6392e3..d861d94242a46750cbc27983d1dba3bceddaf622 100644 (file)
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -649,6 +649,7 @@ enum btree_insert_ret {
        BTREE_INSERT_ENOSPC,
        BTREE_INSERT_NEED_MARK_REPLICAS,
        BTREE_INSERT_NEED_JOURNAL_RES,
+       BTREE_INSERT_NEED_JOURNAL_RECLAIM,
 };
 
 enum btree_gc_coalesce_fail_reason {

diff --git a/fs/bcachefs/btree_update_leaf.c b/fs/bcachefs/btree_update_leaf.c
index 08d08d2f1ea3c38f612da60147f5372879b04a7b..4504d7740a57a321f3d6659dd57c4badc7eb0bdc 100644 (file)
--- a/fs/bcachefs/btree_update_leaf.c
+++ b/fs/bcachefs/btree_update_leaf.c
@@ -286,6 +286,10 @@ btree_key_can_insert_cached(struct btree_trans *trans,
 
        BUG_ON(iter->level);
 
+       if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+           bch2_btree_key_cache_must_wait(trans->c))
+               return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
+
        if (u64s <= ck->u64s)
                return BTREE_INSERT_OK;
 
@@ -652,6 +656,21 @@ int bch2_trans_commit_error(struct btree_trans *trans,
                trace_trans_restart_journal_res_get(trans->ip);
                ret = -EINTR;
                break;
+       case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
+               bch2_trans_unlock(trans);
+
+               while (bch2_btree_key_cache_must_wait(c)) {
+                       mutex_lock(&c->journal.reclaim_lock);
+                       bch2_journal_reclaim(&c->journal);
+                       mutex_unlock(&c->journal.reclaim_lock);
+               }
+
+               if (bch2_trans_relock(trans))
+                       return 0;
+
+               trace_trans_restart_journal_reclaim(trans->ip);
+               ret = -EINTR;
+               break;
        default:
                BUG_ON(ret >= 0);
                break;

diff --git a/fs/bcachefs/trace.h b/fs/bcachefs/trace.h
index 9706b6a3b1b2ed043f1ddf0838a129f187a63a74..babb07e3acc478f3e6cda089885336c5ffb7f4c0 100644 (file)
--- a/fs/bcachefs/trace.h
+++ b/fs/bcachefs/trace.h
@@ -681,6 +681,11 @@ DEFINE_EVENT(transaction_restart,  trans_restart_journal_preres_get,
        TP_ARGS(ip)
 );
 
+DEFINE_EVENT(transaction_restart,      trans_restart_journal_reclaim,
+       TP_PROTO(unsigned long ip),
+       TP_ARGS(ip)
+);
+
 DEFINE_EVENT(transaction_restart,      trans_restart_mark_replicas,
        TP_PROTO(unsigned long ip),
        TP_ARGS(ip)