bcachefs: Centralize marking of replicas in btree update path

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

11 files changed, 84 insertions, 91 deletions

diff --git a/fs/bcachefs/btree_gc.c b/fs/bcachefs/btree_gc.c
index 0fb89e03fac8..b0d04ed5f2a6 100644
--- a/fs/bcachefs/btree_gc.c
+++ b/fs/bcachefs/btree_gc.c
@@ -155,7 +155,7 @@ static int bch2_btree_mark_ptrs_initial(struct bch_fs *c, enum bkey_type type,
 	       k.k->version.lo > journal_cur_seq(&c->journal));
 
 	if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
-	    fsck_err_on(!bch2_bkey_replicas_marked(c, type, k), c,
+	    fsck_err_on(!bch2_bkey_replicas_marked(c, type, k, false), c,
 			"superblock not marked as containing replicas (type %u)",
 			data_type)) {
 		ret = bch2_mark_bkey_replicas(c, type, k);
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index 467c619f7f6d..7e9ba60288aa 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -440,11 +440,11 @@ enum btree_insert_ret {
 	BTREE_INSERT_OK,
 	/* extent spanned multiple leaf nodes: have to traverse to next node: */
 	BTREE_INSERT_NEED_TRAVERSE,
-	/* write lock held for too long */
 	/* leaf node needs to be split */
 	BTREE_INSERT_BTREE_NODE_FULL,
 	BTREE_INSERT_ENOSPC,
 	BTREE_INSERT_NEED_GC_LOCK,
+	BTREE_INSERT_NEED_MARK_REPLICAS,
 };
 
 enum btree_gc_coalesce_fail_reason {
diff --git a/fs/bcachefs/btree_update_leaf.c b/fs/bcachefs/btree_update_leaf.c
index 44501e98a4ac..093e480977c7 100644
--- a/fs/bcachefs/btree_update_leaf.c
+++ b/fs/bcachefs/btree_update_leaf.c
@@ -12,6 +12,7 @@
 #include "journal.h"
 #include "journal_reclaim.h"
 #include "keylist.h"
+#include "replicas.h"
 #include "trace.h"
 
 #include <linux/sort.h>
@@ -301,8 +302,8 @@ static inline int btree_trans_cmp(struct btree_insert_entry l,
 
 static enum btree_insert_ret
 btree_key_can_insert(struct btree_insert *trans,
-		      struct btree_insert_entry *insert,
-		      unsigned *u64s)
+		     struct btree_insert_entry *insert,
+		     unsigned *u64s)
 {
 	struct bch_fs *c = trans->c;
 	struct btree *b = insert->iter->l[0].b;
@@ -311,6 +312,12 @@ btree_key_can_insert(struct btree_insert *trans,
 	if (unlikely(btree_node_fake(b)))
 		return BTREE_INSERT_BTREE_NODE_FULL;
 
+	if (!bch2_bkey_replicas_marked(c,
+			insert->iter->btree_id,
+			bkey_i_to_s_c(insert->k),
+			true))
+		return BTREE_INSERT_NEED_MARK_REPLICAS;
+
 	ret = !btree_node_is_extents(b)
 		? BTREE_INSERT_OK
 		: bch2_extent_can_insert(trans, insert, u64s);
@@ -327,8 +334,7 @@ btree_key_can_insert(struct btree_insert *trans,
  * Get journal reservation, take write locks, and attempt to do btree update(s):
  */
 static inline int do_btree_insert_at(struct btree_insert *trans,
-				     struct btree_iter **split,
-				     bool *cycle_gc_lock)
+				     struct btree_insert_entry **stopped_at)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
@@ -372,22 +378,10 @@ static inline int do_btree_insert_at(struct btree_insert *trans,
 			u64s = 0;
 
 		u64s += i->k->k.u64s;
-		switch (btree_key_can_insert(trans, i, &u64s)) {
-		case BTREE_INSERT_OK:
-			break;
-		case BTREE_INSERT_BTREE_NODE_FULL:
-			ret = -EINTR;
-			*split = i->iter;
-			goto out;
-		case BTREE_INSERT_ENOSPC:
-			ret = -ENOSPC;
+		ret = btree_key_can_insert(trans, i, &u64s);
+		if (ret) {
+			*stopped_at = i;
 			goto out;
-		case BTREE_INSERT_NEED_GC_LOCK:
-			ret = -EINTR;
-			*cycle_gc_lock = true;
-			goto out;
-		default:
-			BUG();
 		}
 	}
 
@@ -445,8 +439,7 @@ int __bch2_btree_insert_at(struct btree_insert *trans)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
-	struct btree_iter *linked, *split = NULL;
-	bool cycle_gc_lock = false;
+	struct btree_iter *linked;
 	unsigned flags;
 	int ret;
 
@@ -466,9 +459,6 @@ int __bch2_btree_insert_at(struct btree_insert *trans)
 	if (unlikely(!percpu_ref_tryget(&c->writes)))
 		return -EROFS;
 retry:
-	split = NULL;
-	cycle_gc_lock = false;
-
 	trans_for_each_entry(trans, i) {
 		unsigned old_locks_want = i->iter->locks_want;
 		unsigned old_uptodate = i->iter->uptodate;
@@ -486,7 +476,7 @@ retry:
 		}
 	}
 
-	ret = do_btree_insert_at(trans, &split, &cycle_gc_lock);
+	ret = do_btree_insert_at(trans, &i);
 	if (unlikely(ret))
 		goto err;
 
@@ -521,8 +511,9 @@ err:
 	if (!trans->did_work)
 		flags &= ~BTREE_INSERT_NOUNLOCK;
 
-	if (split) {
-		ret = bch2_btree_split_leaf(c, split, flags);
+	switch (ret) {
+	case BTREE_INSERT_BTREE_NODE_FULL:
+		ret = bch2_btree_split_leaf(c, i->iter, flags);
 
 		/*
 		 * if the split succeeded without dropping locks the insert will
@@ -547,9 +538,10 @@ err:
 			trans_restart(" (split)");
 			ret = -EINTR;
 		}
-	}
+		break;
+	case BTREE_INSERT_NEED_GC_LOCK:
+		ret = -EINTR;
 
-	if (cycle_gc_lock) {
 		if (!down_read_trylock(&c->gc_lock)) {
 			if (flags & BTREE_INSERT_NOUNLOCK)
 				goto out;
@@ -558,6 +550,24 @@ err:
 			down_read(&c->gc_lock);
 		}
 		up_read(&c->gc_lock);
+		break;
+	case BTREE_INSERT_ENOSPC:
+		ret = -ENOSPC;
+		break;
+	case BTREE_INSERT_NEED_MARK_REPLICAS:
+		if (flags & BTREE_INSERT_NOUNLOCK) {
+			ret = -EINTR;
+			goto out;
+		}
+
+		bch2_btree_iter_unlock(trans->entries[0].iter);
+		ret = bch2_mark_bkey_replicas(c, i->iter->btree_id,
+					      bkey_i_to_s_c(i->k))
+			?: -EINTR;
+		break;
+	default:
+		BUG_ON(ret >= 0);
+		break;
 	}
 
 	if (ret == -EINTR) {
diff --git a/fs/bcachefs/extents.c b/fs/bcachefs/extents.c
index 0cf343624793..df0ca1fcf2e8 100644
--- a/fs/bcachefs/extents.c
+++ b/fs/bcachefs/extents.c
@@ -675,7 +675,8 @@ void bch2_btree_ptr_debugcheck(struct bch_fs *c, struct btree *b,
 	}
 
 	if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
-	    !bch2_bkey_replicas_marked(c, btree_node_type(b), e.s_c)) {
+	    !bch2_bkey_replicas_marked(c, btree_node_type(b),
+				       e.s_c, false)) {
 		bch2_bkey_val_to_text(&PBUF(buf), c, btree_node_type(b), k);
 		bch2_fs_bug(c,
 			"btree key bad (replicas not marked in superblock):\n%s",
@@ -1635,7 +1636,8 @@ static void bch2_extent_debugcheck_extent(struct bch_fs *c, struct btree *b,
 	}
 
 	if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
-	    !bch2_bkey_replicas_marked(c, btree_node_type(b), e.s_c)) {
+	    !bch2_bkey_replicas_marked(c, btree_node_type(b),
+				       e.s_c, false)) {
 		bch2_bkey_val_to_text(&PBUF(buf), c, btree_node_type(b),
 				      e.s_c);
 		bch2_fs_bug(c,
diff --git a/fs/bcachefs/io.c b/fs/bcachefs/io.c
index 133b702299dd..fbd0a82fdeac 100644
--- a/fs/bcachefs/io.c
+++ b/fs/bcachefs/io.c
@@ -23,7 +23,6 @@
 #include "keylist.h"
 #include "move.h"
 #include "rebalance.h"
-#include "replicas.h"
 #include "super.h"
 #include "super-io.h"
 #include "trace.h"
@@ -336,13 +335,6 @@ static void __bch2_write_index(struct bch_write_op *op)
 			goto err;
 		}
 
-		if (!(op->flags & BCH_WRITE_NOMARK_REPLICAS)) {
-			ret = bch2_mark_bkey_replicas(c, BKEY_TYPE_EXTENTS,
-						      e.s_c);
-			if (ret)
-				goto err;
-		}
-
 		dst = bkey_next(dst);
 	}
 
diff --git a/fs/bcachefs/io.h b/fs/bcachefs/io.h
index 8a7f246e8823..84070b674187 100644
--- a/fs/bcachefs/io.h
+++ b/fs/bcachefs/io.h
@@ -35,10 +35,9 @@ enum bch_write_flags {
 	BCH_WRITE_PAGES_OWNED		= (1 << 5),
 	BCH_WRITE_ONLY_SPECIFIED_DEVS	= (1 << 6),
 	BCH_WRITE_NOPUT_RESERVATION	= (1 << 7),
-	BCH_WRITE_NOMARK_REPLICAS	= (1 << 8),
 
 	/* Internal: */
-	BCH_WRITE_JOURNAL_SEQ_PTR	= (1 << 9),
+	BCH_WRITE_JOURNAL_SEQ_PTR	= (1 << 8),
 };
 
 static inline u64 *op_journal_seq(struct bch_write_op *op)
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index 4555d55b23dd..b1f6433cf9e9 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -785,7 +785,7 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 		if (!degraded &&
 		    (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 		     fsck_err_on(!bch2_replicas_marked(c, BCH_DATA_JOURNAL,
-						       i->devs), c,
+						       i->devs, false), c,
 				 "superblock not marked as containing replicas (type %u)",
 				 BCH_DATA_JOURNAL))) {
 			ret = bch2_mark_replicas(c, BCH_DATA_JOURNAL, i->devs);
diff --git a/fs/bcachefs/migrate.c b/fs/bcachefs/migrate.c
index 38b392472521..46878590327d 100644
--- a/fs/bcachefs/migrate.c
+++ b/fs/bcachefs/migrate.c
@@ -72,11 +72,6 @@ static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 		 */
 		bch2_extent_normalize(c, e.s);
 
-		ret = bch2_mark_bkey_replicas(c, BKEY_TYPE_EXTENTS,
-					      bkey_i_to_s_c(&tmp.key));
-		if (ret)
-			break;
-
 		iter.pos = bkey_start_pos(&tmp.key.k);
 
 		ret = bch2_btree_insert_at(c, NULL, NULL,
diff --git a/fs/bcachefs/move.c b/fs/bcachefs/move.c
index 1f6bad1ae388..7de3c6c475be 100644
--- a/fs/bcachefs/move.c
+++ b/fs/bcachefs/move.c
@@ -150,11 +150,6 @@ static int bch2_migrate_index_update(struct bch_write_op *op)
 			goto next;
 		}
 
-		ret = bch2_mark_bkey_replicas(c, BKEY_TYPE_EXTENTS,
-					      extent_i_to_s_c(insert).s_c);
-		if (ret)
-			break;
-
 		ret = bch2_btree_insert_at(c, &op->res,
 				op_journal_seq(op),
 				BTREE_INSERT_ATOMIC|
@@ -239,8 +234,7 @@ int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 	m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
 		BCH_WRITE_PAGES_STABLE|
 		BCH_WRITE_PAGES_OWNED|
-		BCH_WRITE_DATA_ENCODED|
-		BCH_WRITE_NOMARK_REPLICAS;
+		BCH_WRITE_DATA_ENCODED;
 
 	m->op.nr_replicas	= 1;
 	m->op.nr_replicas_required = 1;
diff --git a/fs/bcachefs/replicas.c b/fs/bcachefs/replicas.c
index ef62756e8908..83fc9c93d295 100644
--- a/fs/bcachefs/replicas.c
+++ b/fs/bcachefs/replicas.c
@@ -160,8 +160,8 @@ cpu_replicas_add_entry(struct bch_replicas_cpu *old,
 	return new;
 }
 
-static bool replicas_has_entry(struct bch_replicas_cpu *r,
-			       struct bch_replicas_entry *search)
+static bool __replicas_has_entry(struct bch_replicas_cpu *r,
+				 struct bch_replicas_entry *search)
 {
 	return replicas_entry_bytes(search) <= r->entry_size &&
 		eytzinger0_find(r->entries, r->nr,
@@ -169,6 +169,24 @@ static bool replicas_has_entry(struct bch_replicas_cpu *r,
 				memcmp, search) < r->nr;
 }
 
+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) &&
+		(!check_gc_replicas ||
+		 likely(!(gc_r = rcu_dereference(c->replicas_gc))) ||
+		 __replicas_has_entry(gc_r, search));
+	rcu_read_unlock();
+
+	return marked;
+}
+
 noinline
 static int bch2_mark_replicas_slowpath(struct bch_fs *c,
 				struct bch_replicas_entry *new_entry)
@@ -180,7 +198,7 @@ static int bch2_mark_replicas_slowpath(struct bch_fs *c,
 
 	old_gc = rcu_dereference_protected(c->replicas_gc,
 					   lockdep_is_held(&c->sb_lock));
-	if (old_gc && !replicas_has_entry(old_gc, new_entry)) {
+	if (old_gc && !__replicas_has_entry(old_gc, new_entry)) {
 		new_gc = cpu_replicas_add_entry(old_gc, new_entry);
 		if (!new_gc)
 			goto err;
@@ -188,7 +206,7 @@ static int bch2_mark_replicas_slowpath(struct bch_fs *c,
 
 	old_r = rcu_dereference_protected(c->replicas,
 					  lockdep_is_held(&c->sb_lock));
-	if (!replicas_has_entry(old_r, new_entry)) {
+	if (!__replicas_has_entry(old_r, new_entry)) {
 		new_r = cpu_replicas_add_entry(old_r, new_entry);
 		if (!new_r)
 			goto err;
@@ -227,17 +245,8 @@ err:
 static int __bch2_mark_replicas(struct bch_fs *c,
 				struct bch_replicas_entry *devs)
 {
-	struct bch_replicas_cpu *r, *gc_r;
-	bool marked;
-
-	rcu_read_lock();
-	r = rcu_dereference(c->replicas);
-	gc_r = rcu_dereference(c->replicas_gc);
-	marked = replicas_has_entry(r, devs) &&
-		(!likely(gc_r) || replicas_has_entry(gc_r, devs));
-	rcu_read_unlock();
-
-	return likely(marked) ? 0
+	return likely(replicas_has_entry(c, devs, true))
+		? 0
 		: bch2_mark_replicas_slowpath(c, devs);
 }
 
@@ -666,10 +675,10 @@ const struct bch_sb_field_ops bch_sb_field_ops_replicas_v0 = {
 
 bool bch2_replicas_marked(struct bch_fs *c,
 			  enum bch_data_type data_type,
-			  struct bch_devs_list devs)
+			  struct bch_devs_list devs,
+			  bool check_gc_replicas)
 {
 	struct bch_replicas_entry_padded search;
-	bool ret;
 
 	if (!devs.nr)
 		return true;
@@ -678,19 +687,15 @@ bool bch2_replicas_marked(struct bch_fs *c,
 
 	devlist_to_replicas(devs, data_type, &search.e);
 
-	rcu_read_lock();
-	ret = replicas_has_entry(rcu_dereference(c->replicas), &search.e);
-	rcu_read_unlock();
-
-	return ret;
+	return replicas_has_entry(c, &search.e, check_gc_replicas);
 }
 
 bool bch2_bkey_replicas_marked(struct bch_fs *c,
 			       enum bkey_type type,
-			       struct bkey_s_c k)
+			       struct bkey_s_c k,
+			       bool check_gc_replicas)
 {
 	struct bch_replicas_entry_padded search;
-	bool ret;
 
 	memset(&search, 0, sizeof(search));
 
@@ -700,20 +705,16 @@ bool bch2_bkey_replicas_marked(struct bch_fs *c,
 
 		for (i = 0; i < cached.nr; i++)
 			if (!bch2_replicas_marked(c, BCH_DATA_CACHED,
-					bch2_dev_list_single(cached.devs[i])))
+					bch2_dev_list_single(cached.devs[i]),
+					check_gc_replicas))
 				return false;
 	}
 
 	bkey_to_replicas(type, k, &search.e);
 
-	if (!search.e.nr_devs)
-		return true;
-
-	rcu_read_lock();
-	ret = replicas_has_entry(rcu_dereference(c->replicas), &search.e);
-	rcu_read_unlock();
-
-	return ret;
+	return search.e.nr_devs
+		? replicas_has_entry(c, &search.e, check_gc_replicas)
+		: true;
 }
 
 struct replicas_status __bch2_replicas_status(struct bch_fs *c,
diff --git a/fs/bcachefs/replicas.h b/fs/bcachefs/replicas.h
index a343dd9cd97f..e22d2d7cd08a 100644
--- a/fs/bcachefs/replicas.h
+++ b/fs/bcachefs/replicas.h
@@ -5,9 +5,9 @@
 #include "replicas_types.h"
 
 bool bch2_replicas_marked(struct bch_fs *, enum bch_data_type,
-			  struct bch_devs_list);
+			  struct bch_devs_list, bool);
 bool bch2_bkey_replicas_marked(struct bch_fs *, enum bkey_type,
-			       struct bkey_s_c);
+			       struct bkey_s_c, bool);
 int bch2_mark_replicas(struct bch_fs *, enum bch_data_type,
 		       struct bch_devs_list);
 int bch2_mark_bkey_replicas(struct bch_fs *, enum bkey_type,