bcachefs: Persist alloc info on clean shutdown

 - Does not persist alloc info for stripes yet
 - Also does not yet include filesystem block/sector counts yet, from
struct fs_usage
 - Not made use of just yet
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/alloc_background.c

@@ -250,6 +250,9 @@ int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
 				bch2_alloc_read_key(c, bkey_i_to_s_c(k));
 	}
 
+	for_each_member_device(ca, c, i)
+		bch2_dev_usage_from_buckets(c, ca);
+
 	mutex_lock(&c->bucket_clock[READ].lock);
 	for_each_member_device(ca, c, i) {
 		down_read(&ca->bucket_lock);
@@ -281,35 +284,51 @@ static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
 #endif
 	struct bkey_i_alloc *a = bkey_alloc_init(&alloc_key.k);
 	struct bucket *g;
-	struct bucket_mark m;
+	struct bucket_mark m, new;
 	int ret;
 
 	BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 
 	a->k.p = POS(ca->dev_idx, b);
 
+	bch2_btree_iter_set_pos(iter, a->k.p);
+
+	ret = bch2_btree_iter_traverse(iter);
+	if (ret)
+		return ret;
+
 	percpu_down_read(&c->mark_lock);
 	g = bucket(ca, b);
-	m = bucket_cmpxchg(g, m, m.dirty = false);
+	m = READ_ONCE(g->mark);
+
+	if (!m.dirty) {
+		percpu_up_read(&c->mark_lock);
+		return 0;
+	}
 
 	__alloc_write_key(a, g, m);
 	percpu_up_read(&c->mark_lock);
 
 	bch2_btree_iter_cond_resched(iter);
 
-	bch2_btree_iter_set_pos(iter, a->k.p);
-
 	ret = bch2_btree_insert_at(c, NULL, journal_seq,
+				   BTREE_INSERT_NOCHECK_RW|
 				   BTREE_INSERT_NOFAIL|
 				   BTREE_INSERT_USE_RESERVE|
 				   BTREE_INSERT_USE_ALLOC_RESERVE|
 				   flags,
 				   BTREE_INSERT_ENTRY(iter, &a->k_i));
+	if (ret)
+		return ret;
 
-	if (!ret && ca->buckets_written)
+	new = m;
+	new.dirty = false;
+	atomic64_cmpxchg(&g->_mark.v, m.v.counter, new.v.counter);
+
+	if (ca->buckets_written)
 		set_bit(b, ca->buckets_written);
 
-	return ret;
+	return 0;
 }
 
 int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
@@ -899,10 +918,19 @@ static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t
 		for (i = 0; i < RESERVE_NR; i++)
 			if (fifo_push(&ca->free[i], bucket)) {
 				fifo_pop(&ca->free_inc, bucket);
+
 				closure_wake_up(&c->freelist_wait);
+				ca->allocator_blocked_full = false;
+
 				spin_unlock(&c->freelist_lock);
 				goto out;
 			}
+
+		if (!ca->allocator_blocked_full) {
+			ca->allocator_blocked_full = true;
+			closure_wake_up(&c->freelist_wait);
+		}
+
 		spin_unlock(&c->freelist_lock);
 
 		if ((current->flags & PF_KTHREAD) &&
@@ -1227,6 +1255,11 @@ void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
 			set_bit(ca->dev_idx, c->rw_devs[i].d);
 }
 
+void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
+{
+	closure_wait_event(&c->freelist_wait, ca->allocator_blocked_full);
+}
+
 /* stop allocator thread: */
 void bch2_dev_allocator_stop(struct bch_dev *ca)
 {

fs/bcachefs/alloc_background.h

@@ -52,6 +52,7 @@ void bch2_recalc_capacity(struct bch_fs *);
 void bch2_dev_allocator_remove(struct bch_fs *, struct bch_dev *);
 void bch2_dev_allocator_add(struct bch_fs *, struct bch_dev *);
 
+void bch2_dev_allocator_quiesce(struct bch_fs *, struct bch_dev *);
 void bch2_dev_allocator_stop(struct bch_dev *);
 int bch2_dev_allocator_start(struct bch_dev *);
 

fs/bcachefs/bcachefs.h

@@ -431,7 +431,13 @@ struct bch_dev {
 
 	size_t			inc_gen_needs_gc;
 	size_t			inc_gen_really_needs_gc;
+
+	/*
+	 * XXX: this should be an enum for allocator state, so as to include
+	 * error state
+	 */
 	bool			allocator_blocked;
+	bool			allocator_blocked_full;
 
 	alloc_heap		alloc_heap;
 

fs/bcachefs/btree_update.h

@@ -78,6 +78,7 @@ enum {
 	__BTREE_INSERT_ATOMIC,
 	__BTREE_INSERT_NOUNLOCK,
 	__BTREE_INSERT_NOFAIL,
+	__BTREE_INSERT_NOCHECK_RW,
 	__BTREE_INSERT_USE_RESERVE,
 	__BTREE_INSERT_USE_ALLOC_RESERVE,
 	__BTREE_INSERT_JOURNAL_REPLAY,
@@ -101,6 +102,8 @@ enum {
 /* Don't check for -ENOSPC: */
 #define BTREE_INSERT_NOFAIL		(1 << __BTREE_INSERT_NOFAIL)
 
+#define BTREE_INSERT_NOCHECK_RW		(1 << __BTREE_INSERT_NOCHECK_RW)
+
 /* for copygc, or when merging btree nodes */
 #define BTREE_INSERT_USE_RESERVE	(1 << __BTREE_INSERT_USE_RESERVE)
 #define BTREE_INSERT_USE_ALLOC_RESERVE	(1 << __BTREE_INSERT_USE_ALLOC_RESERVE)

fs/bcachefs/btree_update_interior.c

@@ -1172,6 +1172,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 
 	mutex_lock(&c->btree_interior_update_lock);
 	percpu_down_read(&c->mark_lock);
+	preempt_disable();
 	fs_usage = bch2_fs_usage_get_scratch(c);
 
 	bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
@@ -1194,6 +1195,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 	bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res,
 			    gc_pos_btree_node(b));
 
+	preempt_enable();
 	percpu_up_read(&c->mark_lock);
 	mutex_unlock(&c->btree_interior_update_lock);
 

fs/bcachefs/btree_update_leaf.c

@@ -629,7 +629,8 @@ int __bch2_btree_insert_at(struct btree_insert *trans)
 	trans_for_each_entry(trans, i)
 		btree_insert_entry_checks(c, i);
 
-	if (unlikely(!percpu_ref_tryget(&c->writes)))
+	if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
+		     !percpu_ref_tryget(&c->writes)))
 		return -EROFS;
 retry:
 	trans_for_each_iter(trans, i) {
@@ -659,7 +660,8 @@ int __bch2_btree_insert_at(struct btree_insert *trans)
 	trans_for_each_iter(trans, i)
 		bch2_btree_iter_downgrade(i->iter);
 out:
-	percpu_ref_put(&c->writes);
+	if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
+		percpu_ref_put(&c->writes);
 
 	/* make sure we didn't drop or screw up locks: */
 	trans_for_each_iter(trans, i) {

fs/bcachefs/buckets.c

@@ -387,7 +387,8 @@ static void __bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 	*old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
 		BUG_ON(!is_available_bucket(new));
 
-		new.owned_by_allocator	= 1;
+		new.owned_by_allocator	= true;
+		new.dirty		= true;
 		new.data_type		= 0;
 		new.cached_sectors	= 0;
 		new.dirty_sectors	= 0;
@@ -460,6 +461,7 @@ static void __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 	       type != BCH_DATA_JOURNAL);
 
 	bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
+		new.dirty	= true;
 		new.data_type	= type;
 		checked_add(new.dirty_sectors, sectors);
 	}));
@@ -487,13 +489,14 @@ void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 						    true);
 	} else {
 		struct bucket *g;
-		struct bucket_mark old, new;
+		struct bucket_mark new;
 
 		rcu_read_lock();
 
 		g = bucket(ca, b);
-		old = bucket_cmpxchg(g, new, ({
-			new.data_type = type;
+		bucket_cmpxchg(g, new, ({
+			new.dirty	= true;
+			new.data_type	= type;
 			checked_add(new.dirty_sectors, sectors);
 		}));
 
@@ -546,6 +549,8 @@ static void bch2_mark_pointer(struct bch_fs *c,
 	do {
 		new.v.counter = old.v.counter = v;
 
+		new.dirty = true;
+
 		/*
 		 * Check this after reading bucket mark to guard against
 		 * the allocator invalidating a bucket after we've already
@@ -709,6 +714,7 @@ static void bucket_set_stripe(struct bch_fs *c,
 		BUG_ON(ptr_stale(ca, ptr));
 
 		old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
+			new.dirty			= true;
 			new.stripe			= enabled;
 			if (journal_seq) {
 				new.journal_seq_valid	= 1;

fs/bcachefs/buckets.h

@@ -182,6 +182,8 @@ static inline bool bucket_needs_journal_commit(struct bucket_mark m,
 
 struct bch_dev_usage bch2_dev_usage_read(struct bch_fs *, struct bch_dev *);
 
+void bch2_dev_usage_from_buckets(struct bch_fs *, struct bch_dev *);
+
 static inline u64 __dev_buckets_available(struct bch_dev *ca,
 					  struct bch_dev_usage stats)
 {

fs/bcachefs/super.c

@@ -174,7 +174,9 @@ 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;
 	unsigned i;
+	int ret;
 
 	bch2_rebalance_stop(c);
 
@@ -189,23 +191,36 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	 */
 	bch2_journal_flush_all_pins(&c->journal);
 
-	for_each_member_device(ca, c, i)
-		bch2_dev_allocator_stop(ca);
+	do {
+		ret = bch2_alloc_write(c, false, &wrote);
+		if (ret) {
+			bch2_fs_inconsistent(c, "error writing out alloc info %i", ret);
+			break;
+		}
 
-	bch2_journal_flush_all_pins(&c->journal);
+		for_each_member_device(ca, c, i)
+			bch2_dev_allocator_quiesce(c, ca);
 
-	/*
-	 * 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:
-	 */
-	closure_wait_event(&c->btree_interior_update_wait,
-			   !bch2_btree_interior_updates_nr_pending(c));
+		bch2_journal_flush_all_pins(&c->journal);
+
+		/*
+		 * 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:
+		 */
+		closure_wait_event(&c->btree_interior_update_wait,
+				   !bch2_btree_interior_updates_nr_pending(c));
+	} while (wrote);
+
+	for_each_member_device(ca, c, i)
+		bch2_dev_allocator_stop(ca);
 
 	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: