bcachefs: Ensure journal reclaim runs when btree key cache is too dirty

Ensuring the key cache isn't too dirty is critical for ensuring that the
shrinker can reclaim memory.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/btree_key_cache.c

@@ -461,6 +461,7 @@ bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 	struct bkey_cached *ck = (void *) iter->l[0].b;
+	bool kick_reclaim = false;
 
 	BUG_ON(insert->u64s > ck->u64s);
 
@@ -485,11 +486,18 @@ bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 
 		set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 		c->btree_key_cache.nr_dirty++;
+
+		if (bch2_nr_btree_keys_need_flush(c))
+			kick_reclaim = true;
+
 		mutex_unlock(&c->btree_key_cache.lock);
 	}
 
 	bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 				&ck->journal, btree_key_cache_journal_flush);
+
+	if (kick_reclaim)
+		mod_delayed_work(c->journal_reclaim_wq, &c->journal.reclaim_work, 0);
 	return true;
 }
 

fs/bcachefs/btree_key_cache.h

 #ifndef _BCACHEFS_BTREE_KEY_CACHE_H
 #define _BCACHEFS_BTREE_KEY_CACHE_H
 
+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;
+
+	return max_t(ssize_t, 0, nr_dirty - max_dirty);
+}
+
 struct bkey_cached *
 bch2_btree_key_cache_find(struct bch_fs *, enum btree_id, struct bpos);
 

fs/bcachefs/journal_reclaim.c

 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
+#include "btree_key_cache.h"
 #include "journal.h"
 #include "journal_io.h"
 #include "journal_reclaim.h"
 #include "replicas.h"
 #include "super.h"
+#include "trace.h"
 
 /* Free space calculations: */
 
@@ -432,7 +434,6 @@ journal_get_next_pin(struct journal *j, u64 max_seq, u64 *seq)
 		list_move(&ret->list, &pin_list->flushed);
 		BUG_ON(j->flush_in_progress);
 		j->flush_in_progress = ret;
-		j->last_flushed = jiffies;
 	}
 
 	spin_unlock(&j->lock);
@@ -441,17 +442,24 @@ journal_get_next_pin(struct journal *j, u64 max_seq, u64 *seq)
 }
 
 /* returns true if we did work */
-static bool journal_flush_pins(struct journal *j, u64 seq_to_flush,
-			       unsigned min_nr)
+static u64 journal_flush_pins(struct journal *j, u64 seq_to_flush,
+			      unsigned min_nr)
 {
 	struct journal_entry_pin *pin;
-	bool ret = false;
-	u64 seq;
+	u64 seq, ret = 0;
 
 	lockdep_assert_held(&j->reclaim_lock);
 
-	while ((pin = journal_get_next_pin(j, min_nr
-				? U64_MAX : seq_to_flush, &seq))) {
+	while (1) {
+		cond_resched();
+
+		j->last_flushed = jiffies;
+
+		pin = journal_get_next_pin(j, min_nr
+				? U64_MAX : seq_to_flush, &seq);
+		if (!pin)
+			break;
+
 		if (min_nr)
 			min_nr--;
 
@@ -460,7 +468,7 @@ static bool journal_flush_pins(struct journal *j, u64 seq_to_flush,
 		BUG_ON(j->flush_in_progress != pin);
 		j->flush_in_progress = NULL;
 		wake_up(&j->pin_flush_wait);
-		ret = true;
+		ret++;
 	}
 
 	return ret;
@@ -527,8 +535,8 @@ static u64 journal_seq_to_flush(struct journal *j)
 void bch2_journal_reclaim(struct journal *j)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
-	unsigned min_nr = 0;
-	u64 seq_to_flush = 0;
+	u64 seq_to_flush, nr_flushed = 0;
+	size_t min_nr;
 
 	lockdep_assert_held(&j->reclaim_lock);
 
@@ -549,12 +557,25 @@ void bch2_journal_reclaim(struct journal *j)
 		if (j->prereserved.reserved * 2 > j->prereserved.remaining)
 			min_nr = 1;
 
-		if ((atomic_read(&c->btree_cache.dirty) * 4 >
-		     c->btree_cache.used  * 3) ||
-		    (c->btree_key_cache.nr_dirty * 4 >
-		     c->btree_key_cache.nr_keys))
+		if (atomic_read(&c->btree_cache.dirty) * 4 >
+		    c->btree_cache.used  * 3)
 			min_nr = 1;
-	} while (journal_flush_pins(j, seq_to_flush, min_nr));
+
+		min_nr = max(min_nr, bch2_nr_btree_keys_need_flush(c));
+
+		trace_journal_reclaim_start(c,
+				min_nr,
+				j->prereserved.reserved,
+				j->prereserved.remaining,
+				atomic_read(&c->btree_cache.dirty),
+				c->btree_cache.used,
+				c->btree_key_cache.nr_dirty,
+				c->btree_key_cache.nr_keys);
+
+		nr_flushed += journal_flush_pins(j, seq_to_flush, min_nr);
+	} while (min_nr);
+
+	trace_journal_reclaim_finish(c, nr_flushed);
 
 	if (!bch2_journal_error(j))
 		queue_delayed_work(c->journal_reclaim_wq, &j->reclaim_work,
@@ -582,7 +603,7 @@ static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 
 	mutex_lock(&j->reclaim_lock);
 
-	*did_work = journal_flush_pins(j, seq_to_flush, 0);
+	*did_work = journal_flush_pins(j, seq_to_flush, 0) != 0;
 
 	spin_lock(&j->lock);
 	/*

fs/bcachefs/trace.h

@@ -121,6 +121,65 @@ DEFINE_EVENT(bio, journal_write,
 	TP_ARGS(bio)
 );
 
+TRACE_EVENT(journal_reclaim_start,
+	TP_PROTO(struct bch_fs *c, u64 min_nr,
+		 u64 prereserved, u64 prereserved_total,
+		 u64 btree_cache_dirty, u64 btree_cache_total,
+		 u64 btree_key_cache_dirty, u64 btree_key_cache_total),
+	TP_ARGS(c, min_nr, prereserved, prereserved_total,
+		btree_cache_dirty, btree_cache_total,
+		btree_key_cache_dirty, btree_key_cache_total),
+
+	TP_STRUCT__entry(
+		__array(char,		uuid,	16		)
+		__field(u64,		min_nr			)
+		__field(u64,		prereserved		)
+		__field(u64,		prereserved_total	)
+		__field(u64,		btree_cache_dirty	)
+		__field(u64,		btree_cache_total	)
+		__field(u64,		btree_key_cache_dirty	)
+		__field(u64,		btree_key_cache_total	)
+	),
+
+	TP_fast_assign(
+		memcpy(__entry->uuid, c->sb.user_uuid.b, 16);
+		__entry->min_nr			= min_nr;
+		__entry->prereserved		= prereserved;
+		__entry->prereserved_total	= prereserved_total;
+		__entry->btree_cache_dirty	= btree_cache_dirty;
+		__entry->btree_cache_total	= btree_cache_total;
+		__entry->btree_key_cache_dirty	= btree_key_cache_dirty;
+		__entry->btree_key_cache_total	= btree_key_cache_total;
+	),
+
+	TP_printk("%pU min %llu prereserved %llu/%llu btree cache %llu/%llu key cache %llu/%llu",
+		  __entry->uuid,
+		  __entry->min_nr,
+		  __entry->prereserved,
+		  __entry->prereserved_total,
+		  __entry->btree_cache_dirty,
+		  __entry->btree_cache_total,
+		  __entry->btree_key_cache_dirty,
+		  __entry->btree_key_cache_total)
+);
+
+TRACE_EVENT(journal_reclaim_finish,
+	TP_PROTO(struct bch_fs *c, u64 nr_flushed),
+	TP_ARGS(c, nr_flushed),
+
+	TP_STRUCT__entry(
+		__array(char,		uuid,	16 )
+		__field(u64,		nr_flushed )
+	),
+
+	TP_fast_assign(
+		memcpy(__entry->uuid, c->sb.user_uuid.b, 16);
+		__entry->nr_flushed = nr_flushed;
+	),
+
+	TP_printk("%pU flushed %llu", __entry->uuid, __entry->nr_flushed)
+);
+
 /* bset.c: */
 
 DEFINE_EVENT(bpos, bkey_pack_pos_fail,