bcachefs: More allocator startup improvements

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

fs/bcachefs/alloc_background.c

@@ -347,12 +347,14 @@ int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
 	return ret;
 }
 
-int bch2_alloc_write(struct bch_fs *c)
+int bch2_alloc_write(struct bch_fs *c, bool nowait, bool *wrote)
 {
 	struct bch_dev *ca;
 	unsigned i;
 	int ret = 0;
 
+	*wrote = false;
+
 	for_each_rw_member(ca, c, i) {
 		struct btree_iter iter;
 		struct bucket_array *buckets;
@@ -370,9 +372,14 @@ int bch2_alloc_write(struct bch_fs *c)
 			if (!buckets->b[b].mark.dirty)
 				continue;
 
-			ret = __bch2_alloc_write_key(c, ca, b, &iter, NULL, 0);
+			ret = __bch2_alloc_write_key(c, ca, b, &iter, NULL,
+						     nowait
+						     ? BTREE_INSERT_NOWAIT
+						     : 0);
 			if (ret)
 				break;
+
+			*wrote = true;
 		}
 		up_read(&ca->bucket_lock);
 		bch2_btree_iter_unlock(&iter);
@@ -1270,20 +1277,23 @@ static void flush_held_btree_writes(struct bch_fs *c)
 	struct bucket_table *tbl;
 	struct rhash_head *pos;
 	struct btree *b;
-	bool flush_updates;
-	size_t i, nr_pending_updates;
+	bool nodes_blocked;
+	size_t i;
+	struct closure cl;
+
+	closure_init_stack(&cl);
 
 	clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
 again:
 	pr_debug("flushing dirty btree nodes");
 	cond_resched();
+	closure_wait(&c->btree_interior_update_wait, &cl);
 
-	flush_updates = false;
-	nr_pending_updates = bch2_btree_interior_updates_nr_pending(c);
+	nodes_blocked = false;
 
 	rcu_read_lock();
 	for_each_cached_btree(b, c, tbl, i, pos)
-		if (btree_node_dirty(b) && (!b->written || b->level)) {
+		if (btree_node_need_write(b)) {
 			if (btree_node_may_write(b)) {
 				rcu_read_unlock();
 				btree_node_lock_type(c, b, SIX_LOCK_read);
@@ -1291,7 +1301,7 @@ static void flush_held_btree_writes(struct bch_fs *c)
 				six_unlock_read(&b->lock);
 				goto again;
 			} else {
-				flush_updates = true;
+				nodes_blocked = true;
 			}
 		}
 	rcu_read_unlock();
@@ -1299,17 +1309,16 @@ static void flush_held_btree_writes(struct bch_fs *c)
 	if (c->btree_roots_dirty)
 		bch2_journal_meta(&c->journal);
 
-	/*
-	 * This is ugly, but it's needed to flush btree node writes
-	 * without spinning...
-	 */
-	if (flush_updates) {
-		closure_wait_event(&c->btree_interior_update_wait,
-				   bch2_btree_interior_updates_nr_pending(c) <
-				   nr_pending_updates);
+	if (nodes_blocked) {
+		closure_sync(&cl);
 		goto again;
 	}
 
+	closure_wake_up(&c->btree_interior_update_wait);
+	closure_sync(&cl);
+
+	closure_wait_event(&c->btree_interior_update_wait,
+			   !bch2_btree_interior_updates_nr_pending(c));
 }
 
 static void allocator_start_issue_discards(struct bch_fs *c)
@@ -1331,13 +1340,10 @@ static int __bch2_fs_allocator_start(struct bch_fs *c)
 	unsigned dev_iter;
 	u64 journal_seq = 0;
 	long bu;
-	bool invalidating_data = false;
 	int ret = 0;
 
-	if (test_alloc_startup(c)) {
-		invalidating_data = true;
+	if (test_alloc_startup(c))
 		goto not_enough;
-	}
 
 	/* Scan for buckets that are already invalidated: */
 	for_each_rw_member(ca, c, dev_iter) {
@@ -1384,21 +1390,6 @@ static int __bch2_fs_allocator_start(struct bch_fs *c)
 not_enough:
 	pr_debug("not enough empty buckets; scanning for reclaimable buckets");
 
-	for_each_rw_member(ca, c, dev_iter) {
-		find_reclaimable_buckets(c, ca);
-
-		while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
-		       (bu = next_alloc_bucket(ca)) >= 0) {
-			invalidating_data |=
-				bch2_invalidate_one_bucket(c, ca, bu, &journal_seq);
-
-			fifo_push(&ca->free[RESERVE_BTREE], bu);
-			bucket_set_dirty(ca, bu);
-		}
-	}
-
-	pr_debug("done scanning for reclaimable buckets");
-
 	/*
 	 * We're moving buckets to freelists _before_ they've been marked as
 	 * invalidated on disk - we have to so that we can allocate new btree
@@ -1408,38 +1399,59 @@ static int __bch2_fs_allocator_start(struct bch_fs *c)
 	 * have cached data in them, which is live until they're marked as
 	 * invalidated on disk:
 	 */
-	if (invalidating_data) {
-		pr_debug("invalidating existing data");
-		set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
-	} else {
-		pr_debug("issuing discards");
-		allocator_start_issue_discards(c);
-	}
+	set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
 
-	/*
-	 * XXX: it's possible for this to deadlock waiting on journal reclaim,
-	 * since we're holding btree writes. What then?
-	 */
-	ret = bch2_alloc_write(c);
-	if (ret)
-		return ret;
+	while (1) {
+		bool wrote = false;
 
-	if (invalidating_data) {
-		pr_debug("flushing journal");
+		for_each_rw_member(ca, c, dev_iter) {
+			find_reclaimable_buckets(c, ca);
 
-		ret = bch2_journal_flush_seq(&c->journal, journal_seq);
-		if (ret)
-			return ret;
+			while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
+			       (bu = next_alloc_bucket(ca)) >= 0) {
+				bch2_invalidate_one_bucket(c, ca, bu,
+							   &journal_seq);
+
+				fifo_push(&ca->free[RESERVE_BTREE], bu);
+				bucket_set_dirty(ca, bu);
+			}
+		}
+
+		pr_debug("done scanning for reclaimable buckets");
+
+		/*
+		 * XXX: it's possible for this to deadlock waiting on journal reclaim,
+		 * since we're holding btree writes. What then?
+		 */
+		ret = bch2_alloc_write(c, true, &wrote);
 
-		pr_debug("issuing discards");
-		allocator_start_issue_discards(c);
+		/*
+		 * If bch2_alloc_write() did anything, it may have used some
+		 * buckets, and we need the RESERVE_BTREE freelist full - so we
+		 * need to loop and scan again.
+		 * And if it errored, it may have been because there weren't
+		 * enough buckets, so just scan and loop again as long as it
+		 * made some progress:
+		 */
+		if (!wrote && ret)
+			return ret;
+		if (!wrote && !ret)
+			break;
 	}
 
+	pr_debug("flushing journal");
+
+	ret = bch2_journal_flush(&c->journal);
+	if (ret)
+		return ret;
+
+	pr_debug("issuing discards");
+	allocator_start_issue_discards(c);
+
 	set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
 
 	/* now flush dirty btree nodes: */
-	if (invalidating_data)
-		flush_held_btree_writes(c);
+	flush_held_btree_writes(c);
 
 	return 0;
 }
@@ -1448,6 +1460,7 @@ int bch2_fs_allocator_start(struct bch_fs *c)
 {
 	struct bch_dev *ca;
 	unsigned i;
+	bool wrote;
 	int ret;
 
 	down_read(&c->gc_lock);
@@ -1465,7 +1478,7 @@ int bch2_fs_allocator_start(struct bch_fs *c)
 		}
 	}
 
-	return bch2_alloc_write(c);
+	return bch2_alloc_write(c, false, &wrote);
 }
 
 void bch2_fs_allocator_background_init(struct bch_fs *c)

fs/bcachefs/alloc_background.h

@@ -55,7 +55,7 @@ void bch2_dev_allocator_add(struct bch_fs *, struct bch_dev *);
 void bch2_dev_allocator_stop(struct bch_dev *);
 int bch2_dev_allocator_start(struct bch_dev *);
 
-int bch2_alloc_write(struct bch_fs *);
+int bch2_alloc_write(struct bch_fs *, bool, bool *);
 int bch2_fs_allocator_start(struct bch_fs *);
 void bch2_fs_allocator_background_init(struct bch_fs *);
 

fs/bcachefs/btree_cache.c

@@ -171,6 +171,10 @@ static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 	if (!btree_node_may_write(b))
 		goto out_unlock;
 
+	if (btree_node_dirty(b) &&
+	    test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
+		goto out_unlock;
+
 	if (btree_node_dirty(b) ||
 	    btree_node_write_in_flight(b) ||
 	    btree_node_read_in_flight(b)) {

fs/bcachefs/btree_io.c

@@ -1330,8 +1330,7 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 		if (!(old & (1 << BTREE_NODE_dirty)))
 			return;
 
-		if (b->written &&
-		    !btree_node_may_write(b))
+		if (!btree_node_may_write(b))
 			return;
 
 		if (old & (1 << BTREE_NODE_write_in_flight)) {
@@ -1347,7 +1346,6 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 	} while (cmpxchg_acquire(&b->flags, old, new) != old);
 
 	BUG_ON(btree_node_fake(b));
-	BUG_ON(!list_empty(&b->write_blocked));
 	BUG_ON((b->will_make_reachable != 0) != !b->written);
 
 	BUG_ON(b->written >= c->opts.btree_node_size);
@@ -1685,15 +1683,13 @@ ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
 		unsigned long flags = READ_ONCE(b->flags);
 		unsigned idx = (flags & (1 << BTREE_NODE_write_idx)) != 0;
 
-		if (//!(flags & (1 << BTREE_NODE_dirty)) &&
-		    !b->writes[0].wait.list.first &&
-		    !b->writes[1].wait.list.first &&
-		    !(b->will_make_reachable & 1))
+		if (!(flags & (1 << BTREE_NODE_dirty)))
 			continue;
 
-		pr_buf(&out, "%p d %u l %u w %u b %u r %u:%lu c %u p %u\n",
+		pr_buf(&out, "%p d %u n %u l %u w %u b %u r %u:%lu c %u p %u\n",
 		       b,
 		       (flags & (1 << BTREE_NODE_dirty)) != 0,
+		       (flags & (1 << BTREE_NODE_need_write)) != 0,
 		       b->level,
 		       b->written,
 		       !list_empty_careful(&b->write_blocked),

fs/bcachefs/btree_io.h

@@ -3,6 +3,7 @@
 #define _BCACHEFS_BTREE_IO_H
 
 #include "bset.h"
+#include "btree_locking.h"
 #include "extents.h"
 #include "io_types.h"
 
@@ -48,7 +49,7 @@ static inline void btree_node_wait_on_io(struct btree *b)
 static inline bool btree_node_may_write(struct btree *b)
 {
 	return list_empty_careful(&b->write_blocked) &&
-		!b->will_make_reachable;
+		(!b->written || !b->will_make_reachable);
 }
 
 enum compact_mode {
@@ -100,42 +101,36 @@ bool bch2_btree_post_write_cleanup(struct bch_fs *, struct btree *);
 void bch2_btree_node_write(struct bch_fs *, struct btree *,
 			  enum six_lock_type);
 
-/*
- * btree_node_dirty() can be cleared with only a read lock,
- * and for bch2_btree_node_write_cond() we want to set need_write iff it's
- * still dirty:
- */
-static inline void set_btree_node_need_write_if_dirty(struct btree *b)
+static inline void btree_node_write_if_need(struct bch_fs *c, struct btree *b)
 {
-	unsigned long old, new, v = READ_ONCE(b->flags);
-
-	do {
-		old = new = v;
-
-		if (!(old & (1 << BTREE_NODE_dirty)))
-			return;
-
-		new |= (1 << BTREE_NODE_need_write);
-	} while ((v = cmpxchg(&b->flags, old, new)) != old);
+	while (b->written &&
+	       btree_node_need_write(b) &&
+	       btree_node_may_write(b)) {
+		if (!btree_node_write_in_flight(b)) {
+			bch2_btree_node_write(c, b, SIX_LOCK_read);
+			break;
+		}
+
+		six_unlock_read(&b->lock);
+		btree_node_wait_on_io(b);
+		btree_node_lock_type(c, b, SIX_LOCK_read);
+	}
 }
 
 #define bch2_btree_node_write_cond(_c, _b, cond)			\
 do {									\
-	while ((_b)->written && btree_node_dirty(_b) &&	(cond)) {	\
-		if (!btree_node_may_write(_b)) {			\
-			set_btree_node_need_write_if_dirty(_b);		\
-			break;						\
-		}							\
+	unsigned long old, new, v = READ_ONCE((_b)->flags);		\
+									\
+	do {								\
+		old = new = v;						\
 									\
-		if (!btree_node_write_in_flight(_b)) {			\
-			bch2_btree_node_write(_c, _b, SIX_LOCK_read);	\
+		if (!(old & (1 << BTREE_NODE_dirty)) || !(cond))	\
 			break;						\
-		}							\
 									\
-		six_unlock_read(&(_b)->lock);				\
-		btree_node_wait_on_io(_b);				\
-		btree_node_lock_type(c, b, SIX_LOCK_read);		\
-	}								\
+		new |= (1 << BTREE_NODE_need_write);			\
+	} while ((v = cmpxchg(&(_b)->flags, old, new)) != old);		\
+									\
+	btree_node_write_if_need(_c, _b);				\
 } while (0)
 
 void bch2_btree_flush_all_reads(struct bch_fs *);

fs/bcachefs/btree_iter.h

@@ -2,6 +2,7 @@
 #ifndef _BCACHEFS_BTREE_ITER_H
 #define _BCACHEFS_BTREE_ITER_H
 
+#include "bset.h"
 #include "btree_types.h"
 
 static inline void btree_iter_set_dirty(struct btree_iter *iter,

fs/bcachefs/btree_locking.h

@@ -11,7 +11,6 @@
  */
 
 #include "btree_iter.h"
-#include "btree_io.h"
 #include "six.h"
 
 /* matches six lock types */

fs/bcachefs/btree_update_interior.c

@@ -367,6 +367,7 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev
 
 	set_btree_node_accessed(b);
 	set_btree_node_dirty(b);
+	set_btree_node_need_write(b);
 
 	bch2_bset_init_first(b, &b->data->keys);
 	memset(&b->nr, 0, sizeof(b->nr));
@@ -655,6 +656,12 @@ static void btree_update_nodes_written(struct closure *cl)
 		closure_wait(&btree_current_write(b)->wait, cl);
 
 		list_del(&as->write_blocked_list);
+
+		/*
+		 * for flush_held_btree_writes() waiting on updates to flush or
+		 * nodes to be writeable:
+		 */
+		closure_wake_up(&c->btree_interior_update_wait);
 		mutex_unlock(&c->btree_interior_update_lock);
 
 		/*
@@ -958,6 +965,12 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 	list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 		list_del(&p->write_blocked_list);
 		btree_update_reparent(as, p);
+
+		/*
+		 * for flush_held_btree_writes() waiting on updates to flush or
+		 * nodes to be writeable:
+		 */
+		closure_wake_up(&c->btree_interior_update_wait);
 	}
 
 	clear_btree_node_dirty(b);

fs/bcachefs/buckets.c

@@ -1038,7 +1038,7 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	size_t reserve_none	= max_t(size_t, 1, nbuckets >> 9);
 	size_t copygc_reserve	= max_t(size_t, 2, nbuckets >> 7);
 	size_t free_inc_nr	= max(max_t(size_t, 1, nbuckets >> 12),
-				      btree_reserve);
+				      btree_reserve * 2);
 	bool resize = ca->buckets[0] != NULL,
 	     start_copygc = ca->copygc_thread != NULL;
 	int ret = -ENOMEM;

fs/bcachefs/util.c

@@ -25,9 +25,6 @@
 #include "eytzinger.h"
 #include "util.h"
 
-#define simple_strtoint(c, end, base)	simple_strtol(c, end, base)
-#define simple_strtouint(c, end, base)	simple_strtoul(c, end, base)
-
 static const char si_units[] = "?kMGTPEZY";
 
 static int __bch2_strtoh(const char *cp, u64 *res,