bcachefs: switch replicas to mark_lock

prep work for upcoming disk accounting changes

5 files changed, 124 insertions, 138 deletions

diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index 4a7a58f0eeeb..e8428e0fe59e 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -525,8 +525,8 @@ struct bch_fs {
 
 	struct bch_dev __rcu	*devs[BCH_SB_MEMBERS_MAX];
 
-	struct bch_replicas_cpu __rcu *replicas;
-	struct bch_replicas_cpu __rcu *replicas_gc;
+	struct bch_replicas_cpu replicas;
+	struct bch_replicas_cpu replicas_gc;
 	struct mutex		replicas_gc_lock;
 
 	struct bch_disk_groups_cpu __rcu *disk_groups;
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index a9fcf0524e11..eae38ea7d5be 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -130,8 +130,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 	int ret;
 
 	mutex_lock(&c->sb_lock);
-	if (!rcu_dereference_protected(c->replicas,
-			lockdep_is_held(&c->sb_lock))->nr) {
+	if (!c->replicas.entries) {
 		bch_info(c, "building replicas info");
 		set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
 	}
diff --git a/fs/bcachefs/replicas.c b/fs/bcachefs/replicas.c
index 6ab4e36ea9ee..1d3161ee13fb 100644
--- a/fs/bcachefs/replicas.c
+++ b/fs/bcachefs/replicas.c
@@ -148,35 +148,31 @@ static inline void devlist_to_replicas(struct bch_devs_list devs,
 	replicas_entry_sort(e);
 }
 
-static struct bch_replicas_cpu *
+static struct bch_replicas_cpu
 cpu_replicas_add_entry(struct bch_replicas_cpu *old,
 		       struct bch_replicas_entry *new_entry)
 {
-	struct bch_replicas_cpu *new;
-	unsigned i, nr, entry_size;
-
-	entry_size = max_t(unsigned, old->entry_size,
-			   replicas_entry_bytes(new_entry));
-	nr = old->nr + 1;
-
-	new = kzalloc(sizeof(struct bch_replicas_cpu) +
-		      nr * entry_size, GFP_NOIO);
-	if (!new)
-		return NULL;
+	unsigned i;
+	struct bch_replicas_cpu new = {
+		.nr		= old->nr + 1,
+		.entry_size	= max_t(unsigned, old->entry_size,
+					replicas_entry_bytes(new_entry)),
+	};
 
-	new->nr		= nr;
-	new->entry_size	= entry_size;
+	new.entries = kcalloc(new.nr, new.entry_size, GFP_NOIO);
+	if (!new.entries)
+		return new;
 
 	for (i = 0; i < old->nr; i++)
-		memcpy(cpu_replicas_entry(new, i),
+		memcpy(cpu_replicas_entry(&new, i),
 		       cpu_replicas_entry(old, i),
 		       old->entry_size);
 
-	memcpy(cpu_replicas_entry(new, old->nr),
+	memcpy(cpu_replicas_entry(&new, old->nr),
 	       new_entry,
 	       replicas_entry_bytes(new_entry));
 
-	bch2_cpu_replicas_sort(new);
+	bch2_cpu_replicas_sort(&new);
 	return new;
 }
 
@@ -193,16 +189,14 @@ static bool replicas_has_entry(struct bch_fs *c,
 			       struct bch_replicas_entry *search,
 			       bool check_gc_replicas)
 {
-	struct bch_replicas_cpu *r, *gc_r;
 	bool marked;
 
-	rcu_read_lock();
-	r = rcu_dereference(c->replicas);
-	marked = __replicas_has_entry(r, search) &&
+	percpu_down_read_preempt_disable(&c->mark_lock);
+	marked = __replicas_has_entry(&c->replicas, search) &&
 		(!check_gc_replicas ||
-		 likely(!(gc_r = rcu_dereference(c->replicas_gc))) ||
-		 __replicas_has_entry(gc_r, search));
-	rcu_read_unlock();
+		 likely((!c->replicas_gc.entries)) ||
+		 __replicas_has_entry(&c->replicas_gc, search));
+	percpu_up_read_preempt_enable(&c->mark_lock);
 
 	return marked;
 }
@@ -211,54 +205,55 @@ noinline
 static int bch2_mark_replicas_slowpath(struct bch_fs *c,
 				struct bch_replicas_entry *new_entry)
 {
-	struct bch_replicas_cpu *old_gc, *new_gc = NULL, *old_r, *new_r = NULL;
+	struct bch_replicas_cpu new_r, new_gc;
 	int ret = -ENOMEM;
 
+	memset(&new_r, 0, sizeof(new_r));
+	memset(&new_gc, 0, sizeof(new_gc));
+
 	mutex_lock(&c->sb_lock);
 
-	old_gc = rcu_dereference_protected(c->replicas_gc,
-					   lockdep_is_held(&c->sb_lock));
-	if (old_gc && !__replicas_has_entry(old_gc, new_entry)) {
-		new_gc = cpu_replicas_add_entry(old_gc, new_entry);
-		if (!new_gc)
+	if (c->replicas_gc.entries &&
+	    !__replicas_has_entry(&c->replicas_gc, new_entry)) {
+		new_gc = cpu_replicas_add_entry(&c->replicas_gc, new_entry);
+		if (!new_gc.entries)
 			goto err;
 	}
 
-	old_r = rcu_dereference_protected(c->replicas,
-					  lockdep_is_held(&c->sb_lock));
-	if (!__replicas_has_entry(old_r, new_entry)) {
-		new_r = cpu_replicas_add_entry(old_r, new_entry);
-		if (!new_r)
+	if (!__replicas_has_entry(&c->replicas, new_entry)) {
+		new_r = cpu_replicas_add_entry(&c->replicas, new_entry);
+		if (!new_r.entries)
 			goto err;
 
-		ret = bch2_cpu_replicas_to_sb_replicas(c, new_r);
+		ret = bch2_cpu_replicas_to_sb_replicas(c, &new_r);
 		if (ret)
 			goto err;
 	}
 
+	if (!new_r.entries &&
+	    !new_gc.entries)
+		goto out;
+
 	/* allocations done, now commit: */
 
-	if (new_r)
+	if (new_r.entries)
 		bch2_write_super(c);
 
 	/* don't update in memory replicas until changes are persistent */
-
-	if (new_gc) {
-		rcu_assign_pointer(c->replicas_gc, new_gc);
-		kfree_rcu(old_gc, rcu);
-	}
-
-	if (new_r) {
-		rcu_assign_pointer(c->replicas, new_r);
-		kfree_rcu(old_r, rcu);
-	}
-
-	mutex_unlock(&c->sb_lock);
-	return 0;
+	percpu_down_write(&c->mark_lock);
+	if (new_r.entries)
+		swap(new_r, c->replicas);
+	if (new_gc.entries)
+		swap(new_gc, c->replicas_gc);
+	percpu_up_write(&c->mark_lock);
+out:
+	ret = 0;
 err:
 	mutex_unlock(&c->sb_lock);
-	kfree(new_gc);
-	kfree(new_r);
+
+	kfree(new_r.entries);
+	kfree(new_gc.entries);
+
 	return ret;
 }
 
@@ -311,20 +306,14 @@ int bch2_mark_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 
 int bch2_replicas_gc_end(struct bch_fs *c, int ret)
 {
-	struct bch_replicas_cpu *new_r, *old_r;
-
 	lockdep_assert_held(&c->replicas_gc_lock);
 
 	mutex_lock(&c->sb_lock);
 
-	new_r = rcu_dereference_protected(c->replicas_gc,
-					  lockdep_is_held(&c->sb_lock));
-	rcu_assign_pointer(c->replicas_gc, NULL);
-
 	if (ret)
 		goto err;
 
-	if (bch2_cpu_replicas_to_sb_replicas(c, new_r)) {
+	if (bch2_cpu_replicas_to_sb_replicas(c, &c->replicas_gc)) {
 		ret = -ENOSPC;
 		goto err;
 	}
@@ -332,51 +321,54 @@ int bch2_replicas_gc_end(struct bch_fs *c, int ret)
 	bch2_write_super(c);
 
 	/* don't update in memory replicas until changes are persistent */
+err:
+	percpu_down_write(&c->mark_lock);
+	if (!ret)
+		swap(c->replicas, c->replicas_gc);
 
-	old_r = rcu_dereference_protected(c->replicas,
-					  lockdep_is_held(&c->sb_lock));
+	kfree(c->replicas_gc.entries);
+	c->replicas_gc.entries = NULL;
+	percpu_up_write(&c->mark_lock);
 
-	rcu_assign_pointer(c->replicas, new_r);
-	kfree_rcu(old_r, rcu);
-out:
 	mutex_unlock(&c->sb_lock);
 	return ret;
-err:
-	kfree_rcu(new_r, rcu);
-	goto out;
 }
 
 int bch2_replicas_gc_start(struct bch_fs *c, unsigned typemask)
 {
-	struct bch_replicas_cpu *dst, *src;
 	struct bch_replicas_entry *e;
+	unsigned i = 0;
 
 	lockdep_assert_held(&c->replicas_gc_lock);
 
 	mutex_lock(&c->sb_lock);
-	BUG_ON(c->replicas_gc);
+	BUG_ON(c->replicas_gc.entries);
 
-	src = rcu_dereference_protected(c->replicas,
-					lockdep_is_held(&c->sb_lock));
+	c->replicas_gc.nr		= 0;
+	c->replicas_gc.entry_size	= 0;
 
-	dst = kzalloc(sizeof(struct bch_replicas_cpu) +
-		      src->nr * src->entry_size, GFP_NOIO);
-	if (!dst) {
+	for_each_cpu_replicas_entry(&c->replicas, e)
+		if (!((1 << e->data_type) & typemask)) {
+			c->replicas_gc.nr++;
+			c->replicas_gc.entry_size =
+				max_t(unsigned, c->replicas_gc.entry_size,
+				      replicas_entry_bytes(e));
+		}
+
+	c->replicas_gc.entries = kcalloc(c->replicas_gc.nr,
+					 c->replicas_gc.entry_size,
+					 GFP_NOIO);
+	if (!c->replicas_gc.entries) {
 		mutex_unlock(&c->sb_lock);
 		return -ENOMEM;
 	}
 
-	dst->nr		= 0;
-	dst->entry_size	= src->entry_size;
-
-	for_each_cpu_replicas_entry(src, e)
+	for_each_cpu_replicas_entry(&c->replicas, e)
 		if (!((1 << e->data_type) & typemask))
-			memcpy(cpu_replicas_entry(dst, dst->nr++),
-			       e, dst->entry_size);
-
-	bch2_cpu_replicas_sort(dst);
+			memcpy(cpu_replicas_entry(&c->replicas_gc, i++),
+			       e, c->replicas_gc.entry_size);
 
-	rcu_assign_pointer(c->replicas_gc, dst);
+	bch2_cpu_replicas_sort(&c->replicas_gc);
 	mutex_unlock(&c->sb_lock);
 
 	return 0;
@@ -384,11 +376,11 @@ int bch2_replicas_gc_start(struct bch_fs *c, unsigned typemask)
 
 /* Replicas tracking - superblock: */
 
-static struct bch_replicas_cpu *
-__bch2_sb_replicas_to_cpu_replicas(struct bch_sb_field_replicas *sb_r)
+static int
+__bch2_sb_replicas_to_cpu_replicas(struct bch_sb_field_replicas *sb_r,
+				   struct bch_replicas_cpu *cpu_r)
 {
 	struct bch_replicas_entry *e, *dst;
-	struct bch_replicas_cpu *cpu_r;
 	unsigned nr = 0, entry_size = 0, idx = 0;
 
 	for_each_replicas_entry(sb_r, e) {
@@ -397,10 +389,9 @@ __bch2_sb_replicas_to_cpu_replicas(struct bch_sb_field_replicas *sb_r)
 		nr++;
 	}
 
-	cpu_r = kzalloc(sizeof(struct bch_replicas_cpu) +
-			nr * entry_size, GFP_NOIO);
-	if (!cpu_r)
-		return NULL;
+	cpu_r->entries = kcalloc(nr, entry_size, GFP_NOIO);
+	if (!cpu_r->entries)
+		return -ENOMEM;
 
 	cpu_r->nr		= nr;
 	cpu_r->entry_size	= entry_size;
@@ -411,14 +402,14 @@ __bch2_sb_replicas_to_cpu_replicas(struct bch_sb_field_replicas *sb_r)
 		replicas_entry_sort(dst);
 	}
 
-	return cpu_r;
+	return 0;
 }
 
-static struct bch_replicas_cpu *
-__bch2_sb_replicas_v0_to_cpu_replicas(struct bch_sb_field_replicas_v0 *sb_r)
+static int
+__bch2_sb_replicas_v0_to_cpu_replicas(struct bch_sb_field_replicas_v0 *sb_r,
+				      struct bch_replicas_cpu *cpu_r)
 {
 	struct bch_replicas_entry_v0 *e;
-	struct bch_replicas_cpu *cpu_r;
 	unsigned nr = 0, entry_size = 0, idx = 0;
 
 	for_each_replicas_entry(sb_r, e) {
@@ -430,10 +421,9 @@ __bch2_sb_replicas_v0_to_cpu_replicas(struct bch_sb_field_replicas_v0 *sb_r)
 	entry_size += sizeof(struct bch_replicas_entry) -
 		sizeof(struct bch_replicas_entry_v0);
 
-	cpu_r = kzalloc(sizeof(struct bch_replicas_cpu) +
-			nr * entry_size, GFP_NOIO);
-	if (!cpu_r)
-		return NULL;
+	cpu_r->entries = kcalloc(nr, entry_size, GFP_NOIO);
+	if (!cpu_r->entries)
+		return -ENOMEM;
 
 	cpu_r->nr		= nr;
 	cpu_r->entry_size	= entry_size;
@@ -449,31 +439,31 @@ __bch2_sb_replicas_v0_to_cpu_replicas(struct bch_sb_field_replicas_v0 *sb_r)
 		replicas_entry_sort(dst);
 	}
 
-	return cpu_r;
+	return 0;
 }
 
 int bch2_sb_replicas_to_cpu_replicas(struct bch_fs *c)
 {
 	struct bch_sb_field_replicas *sb_v1;
 	struct bch_sb_field_replicas_v0 *sb_v0;
-	struct bch_replicas_cpu *cpu_r, *old_r;
+	struct bch_replicas_cpu new_r = { 0, 0, NULL };
+	int ret = 0;
 
 	if ((sb_v1 = bch2_sb_get_replicas(c->disk_sb.sb)))
-		cpu_r = __bch2_sb_replicas_to_cpu_replicas(sb_v1);
+		ret = __bch2_sb_replicas_to_cpu_replicas(sb_v1, &new_r);
 	else if ((sb_v0 = bch2_sb_get_replicas_v0(c->disk_sb.sb)))
-		cpu_r = __bch2_sb_replicas_v0_to_cpu_replicas(sb_v0);
-	else
-		cpu_r = kzalloc(sizeof(struct bch_replicas_cpu), GFP_NOIO);
+		ret = __bch2_sb_replicas_v0_to_cpu_replicas(sb_v0, &new_r);
 
-	if (!cpu_r)
+	if (ret)
 		return -ENOMEM;
 
-	bch2_cpu_replicas_sort(cpu_r);
+	bch2_cpu_replicas_sort(&new_r);
+
+	percpu_down_write(&c->mark_lock);
+	swap(c->replicas, new_r);
+	percpu_up_write(&c->mark_lock);
 
-	old_r = rcu_dereference_check(c->replicas, lockdep_is_held(&c->sb_lock));
-	rcu_assign_pointer(c->replicas, cpu_r);
-	if (old_r)
-		kfree_rcu(old_r, rcu);
+	kfree(new_r.entries);
 
 	return 0;
 }
@@ -588,7 +578,7 @@ static const char *bch2_sb_validate_replicas(struct bch_sb *sb, struct bch_sb_fi
 {
 	struct bch_sb_field_replicas *sb_r = field_to_type(f, replicas);
 	struct bch_sb_field_members *mi = bch2_sb_get_members(sb);
-	struct bch_replicas_cpu *cpu_r = NULL;
+	struct bch_replicas_cpu cpu_r = { .entries = NULL };
 	struct bch_replicas_entry *e;
 	const char *err;
 	unsigned i;
@@ -615,13 +605,12 @@ static const char *bch2_sb_validate_replicas(struct bch_sb *sb, struct bch_sb_fi
 	}
 
 	err = "cannot allocate memory";
-	cpu_r = __bch2_sb_replicas_to_cpu_replicas(sb_r);
-	if (!cpu_r)
+	if (__bch2_sb_replicas_to_cpu_replicas(sb_r, &cpu_r))
 		goto err;
 
-	err = check_dup_replicas_entries(cpu_r);
+	err = check_dup_replicas_entries(&cpu_r);
 err:
-	kfree(cpu_r);
+	kfree(cpu_r.entries);
 	return err;
 }
 
@@ -651,7 +640,7 @@ static const char *bch2_sb_validate_replicas_v0(struct bch_sb *sb, struct bch_sb
 {
 	struct bch_sb_field_replicas_v0 *sb_r = field_to_type(f, replicas_v0);
 	struct bch_sb_field_members *mi = bch2_sb_get_members(sb);
-	struct bch_replicas_cpu *cpu_r = NULL;
+	struct bch_replicas_cpu cpu_r = { .entries = NULL };
 	struct bch_replicas_entry_v0 *e;
 	const char *err;
 	unsigned i;
@@ -672,13 +661,12 @@ static const char *bch2_sb_validate_replicas_v0(struct bch_sb *sb, struct bch_sb
 	}
 
 	err = "cannot allocate memory";
-	cpu_r = __bch2_sb_replicas_v0_to_cpu_replicas(sb_r);
-	if (!cpu_r)
+	if (__bch2_sb_replicas_v0_to_cpu_replicas(sb_r, &cpu_r))
 		goto err;
 
-	err = check_dup_replicas_entries(cpu_r);
+	err = check_dup_replicas_entries(&cpu_r);
 err:
-	kfree(cpu_r);
+	kfree(cpu_r.entries);
 	return err;
 }
 
@@ -733,7 +721,6 @@ struct replicas_status __bch2_replicas_status(struct bch_fs *c,
 {
 	struct bch_sb_field_members *mi;
 	struct bch_replicas_entry *e;
-	struct bch_replicas_cpu *r;
 	unsigned i, nr_online, nr_offline;
 	struct replicas_status ret;
 
@@ -743,10 +730,10 @@ struct replicas_status __bch2_replicas_status(struct bch_fs *c,
 		ret.replicas[i].redundancy = INT_MAX;
 
 	mi = bch2_sb_get_members(c->disk_sb.sb);
-	rcu_read_lock();
-	r = rcu_dereference(c->replicas);
 
-	for_each_cpu_replicas_entry(r, e) {
+	percpu_down_read_preempt_disable(&c->mark_lock);
+
+	for_each_cpu_replicas_entry(&c->replicas, e) {
 		if (e->data_type >= ARRAY_SIZE(ret.replicas))
 			panic("e %p data_type %u\n", e, e->data_type);
 
@@ -771,7 +758,7 @@ struct replicas_status __bch2_replicas_status(struct bch_fs *c,
 			    nr_offline);
 	}
 
-	rcu_read_unlock();
+	percpu_up_read_preempt_enable(&c->mark_lock);
 
 	for (i = 0; i < ARRAY_SIZE(ret.replicas); i++)
 		if (ret.replicas[i].redundancy == INT_MAX)
@@ -820,18 +807,16 @@ int bch2_replicas_online(struct bch_fs *c, bool meta)
 unsigned bch2_dev_has_data(struct bch_fs *c, struct bch_dev *ca)
 {
 	struct bch_replicas_entry *e;
-	struct bch_replicas_cpu *r;
 	unsigned i, ret = 0;
 
-	rcu_read_lock();
-	r = rcu_dereference(c->replicas);
+	percpu_down_read_preempt_disable(&c->mark_lock);
 
-	for_each_cpu_replicas_entry(r, e)
+	for_each_cpu_replicas_entry(&c->replicas, e)
 		for (i = 0; i < e->nr_devs; i++)
 			if (e->devs[i] == ca->dev_idx)
 				ret |= 1 << e->data_type;
 
-	rcu_read_unlock();
+	percpu_up_read_preempt_enable(&c->mark_lock);
 
 	return ret;
 }
diff --git a/fs/bcachefs/replicas_types.h b/fs/bcachefs/replicas_types.h
index 3061840b6a02..0535b1d3760e 100644
--- a/fs/bcachefs/replicas_types.h
+++ b/fs/bcachefs/replicas_types.h
@@ -2,10 +2,9 @@
 #define _BCACHEFS_REPLICAS_TYPES_H
 
 struct bch_replicas_cpu {
-	struct rcu_head		rcu;
 	unsigned		nr;
 	unsigned		entry_size;
-	struct bch_replicas_entry entries[];
+	struct bch_replicas_entry *entries;
 };
 
 #endif /* _BCACHEFS_REPLICAS_TYPES_H */
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index c151147f8945..89eee959222a 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -414,7 +414,8 @@ static void bch2_fs_free(struct bch_fs *c)
 	mempool_exit(&c->btree_reserve_pool);
 	mempool_exit(&c->fill_iter);
 	percpu_ref_exit(&c->writes);
-	kfree(rcu_dereference_protected(c->replicas, 1));
+	kfree(c->replicas.entries);
+	kfree(c->replicas_gc.entries);
 	kfree(rcu_dereference_protected(c->disk_groups, 1));
 
 	if (c->copygc_wq)
@@ -597,6 +598,9 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 
 	bch2_fs_btree_cache_init_early(&c->btree_cache);
 
+	if (percpu_init_rwsem(&c->mark_lock))
+		goto err;
+
 	mutex_lock(&c->sb_lock);
 
 	if (bch2_sb_to_fs(c, sb)) {
@@ -640,7 +644,6 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 			    offsetof(struct btree_write_bio, wbio.bio)),
 			BIOSET_NEED_BVECS) ||
 	    !(c->usage[0] = alloc_percpu(struct bch_fs_usage)) ||
-	    percpu_init_rwsem(&c->mark_lock) ||
 	    mempool_init_kvpmalloc_pool(&c->btree_bounce_pool, 1,
 					btree_bytes(c)) ||
 	    mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,