bcachefs: Fixes for going RO

Now that interior btree updates are fully transactional, we don't need
to write out alloc info in a loop. However, interior btree updates do
put more things in the journal, so we still need a loop in the RO
sequence.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/alloc_background.c

@@ -869,6 +869,15 @@ static int bch2_invalidate_one_bucket2(struct btree_trans *trans,
 	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));
@@ -956,7 +965,7 @@ static int bch2_invalidate_one_bucket2(struct btree_trans *trans,
 		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,

fs/bcachefs/bcachefs.h

@@ -482,6 +482,7 @@ enum {
 	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,

fs/bcachefs/journal_reclaim.c

@@ -413,10 +413,12 @@ journal_get_next_pin(struct journal *j, u64 max_seq, u64 *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);
@@ -431,7 +433,10 @@ static void 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;
 	}
+
+	return ret;
 }
 
 /**
@@ -523,7 +528,8 @@ void bch2_journal_reclaim_work(struct work_struct *work)
 	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;
 
@@ -533,7 +539,7 @@ static int journal_flush_done(struct journal *j, u64 seq_to_flush)
 
 	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);
 	/*
@@ -551,12 +557,17 @@ static int journal_flush_done(struct journal *j, u64 seq_to_flush)
 	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)

fs/bcachefs/journal_reclaim.h

@@ -53,11 +53,11 @@ void bch2_journal_do_discards(struct journal *);
 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);

fs/bcachefs/super.c

@@ -175,7 +175,7 @@ struct bch_fs *bch2_uuid_to_fs(__uuid_t uuid)
 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;
 
@@ -200,12 +200,12 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 		goto nowrote_alloc;
 
 	bch_verbose(c, "writing alloc info");
-
-	do {
-		wrote = false;
-
+	/*
+	 * 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");
 
 	if (ret && !test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
 		bch2_fs_inconsistent(c, "error writing out alloc info %i", ret);
@@ -213,26 +213,33 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	if (ret)
 		goto nowrote_alloc;
 
-		for_each_member_device(ca, c, i)
-			bch2_dev_allocator_quiesce(c, ca);
+	bch_verbose(c, "flushing journal and stopping allocators");
 
 	bch2_journal_flush_all_pins(&c->journal);
+	set_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags);
+
+	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):
 		 */
+		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);
+	bch_verbose(c, "flushing journal and stopping allocators complete");
 
-	bch_verbose(c, "writing alloc info complete");
 	set_bit(BCH_FS_ALLOC_CLEAN, &c->flags);
 nowrote_alloc:
 	closure_wait_event(&c->btree_interior_update_wait,
@@ -243,11 +250,10 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 		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: