bcachefs: Only check inode i_nlink during full fsck

Now that all filesystem operatinos that manipulate the filesystem
heirachy and i_nlink are fully atomic, we can add a feature bit to
indicate i_nlink doesn't need to be checked.

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

4 files changed, 40 insertions, 3 deletions

diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index b5e119d09a83..57132c79c4b9 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -472,6 +472,7 @@ enum {
 	/* misc: */
 	BCH_FS_BDEV_MOUNTED,
 	BCH_FS_FSCK_FIXED_ERRORS,
+	BCH_FS_FSCK_UNFIXED_ERRORS,
 	BCH_FS_FIXED_GENS,
 	BCH_FS_REBUILD_REPLICAS,
 	BCH_FS_HOLD_BTREE_WRITES,
diff --git a/fs/bcachefs/error.c b/fs/bcachefs/error.c
index e975fab43d49..08e79166dae4 100644
--- a/fs/bcachefs/error.c
+++ b/fs/bcachefs/error.c
@@ -132,8 +132,9 @@ print:
 
 	mutex_unlock(&c->fsck_error_lock);
 
-	if (fix)
-		set_bit(BCH_FS_FSCK_FIXED_ERRORS, &c->flags);
+	set_bit(fix
+		? BCH_FS_FSCK_FIXED_ERRORS
+		: BCH_FS_FSCK_UNFIXED_ERRORS, &c->flags);
 
 	return fix				? FSCK_ERR_FIX
 		: flags & FSCK_CAN_IGNORE	? FSCK_ERR_IGNORE
diff --git a/fs/bcachefs/fsck.c b/fs/bcachefs/fsck.c
index eb01284a841f..c352fa01fb5a 100644
--- a/fs/bcachefs/fsck.c
+++ b/fs/bcachefs/fsck.c
@@ -954,6 +954,23 @@ static int check_inode_nlink(struct bch_fs *c,
 		return 0;
 	}
 
+	if (!link->count &&
+	    !(u->bi_flags & BCH_INODE_UNLINKED) &&
+	    (c->sb.features & (1 << BCH_FEATURE_ATOMIC_NLINK))) {
+		if (fsck_err(c, "unreachable inode %llu not marked as unlinked (type %u)",
+			     u->bi_inum, mode_to_type(u->bi_mode)) ==
+		    FSCK_ERR_IGNORE)
+			return 0;
+
+		ret = reattach_inode(c, lostfound_inode, u->bi_inum);
+		if (ret)
+			return ret;
+
+		link->count = 1;
+		real_i_nlink = nlink_bias(u->bi_mode) + link->dir_count;
+		goto set_i_nlink;
+	}
+
 	if (i_nlink < link->count) {
 		if (fsck_err(c, "inode %llu i_link too small (%u < %u, type %i)",
 			     u->bi_inum, i_nlink, link->count,
@@ -973,6 +990,16 @@ static int check_inode_nlink(struct bch_fs *c,
 		goto set_i_nlink;
 	}
 
+	if (i_nlink != real_i_nlink &&
+	    (c->sb.features & (1 << BCH_FEATURE_ATOMIC_NLINK))) {
+		if (fsck_err(c, "inode %llu has wrong i_nlink "
+			     "(type %u i_nlink %u, should be %u)",
+			     u->bi_inum, mode_to_type(u->bi_mode),
+			     i_nlink, real_i_nlink) == FSCK_ERR_IGNORE)
+			return 0;
+		goto set_i_nlink;
+	}
+
 	if (real_i_nlink && i_nlink != real_i_nlink)
 		bch_verbose(c, "setting inode %llu nlink from %u to %u",
 			    u->bi_inum, i_nlink, real_i_nlink);
@@ -1299,7 +1326,8 @@ int bch2_fsck(struct bch_fs *c)
 	if (!c->opts.nofsck)
 		return bch2_fsck_full(c);
 
-	if (!c->sb.clean)
+	if (!c->sb.clean &&
+	    !(c->sb.features & (1 << BCH_FEATURE_ATOMIC_NLINK)))
 		return bch2_fsck_inode_nlink(c);
 
 	return bch2_fsck_walk_inodes_only(c);
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index 2596c3c26064..624d97dc4537 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -256,6 +256,12 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (ret)
 		goto err;
 
+	if (!test_bit(BCH_FS_FSCK_UNFIXED_ERRORS, &c->flags)) {
+		mutex_lock(&c->sb_lock);
+		c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_ATOMIC_NLINK;
+		mutex_unlock(&c->sb_lock);
+	}
+
 	if (enabled_qtypes(c)) {
 		bch_verbose(c, "reading quotas:");
 		ret = bch2_fs_quota_read(c);
@@ -366,6 +372,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	mutex_lock(&c->sb_lock);
 	SET_BCH_SB_INITIALIZED(c->disk_sb.sb, true);
 	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+	c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_ATOMIC_NLINK;
 
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);