Update bcachefs sources to 5b8c4a1366 bcachefs: bcachefs_metadata_version_deleted_inodes

11 files changed, 349 insertions, 107 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index 7fb9f9a3..b2d87414 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-33a60d9b05f523be93973b25e0df1ab2d65fa4fc
+5b8c4a1366df20bc043404cb882230ce86296590
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index 82b0706a..e1f1e8e8 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -208,6 +208,7 @@
 #include "fifo.h"
 #include "nocow_locking_types.h"
 #include "opts.h"
+#include "recovery_types.h"
 #include "seqmutex.h"
 #include "util.h"
 
@@ -452,6 +453,7 @@ enum gc_phase {
 	GC_PHASE_BTREE_backpointers,
 	GC_PHASE_BTREE_bucket_gens,
 	GC_PHASE_BTREE_snapshot_trees,
+	GC_PHASE_BTREE_deleted_inodes,
 
 	GC_PHASE_PENDING_DELETE,
 };
@@ -655,48 +657,6 @@ enum bch_write_ref {
 	BCH_WRITE_REF_NR,
 };
 
-#define PASS_SILENT		BIT(0)
-#define PASS_FSCK		BIT(1)
-#define PASS_UNCLEAN		BIT(2)
-#define PASS_ALWAYS		BIT(3)
-
-#define BCH_RECOVERY_PASSES()									\
-	x(alloc_read,			PASS_ALWAYS)						\
-	x(stripes_read,			PASS_ALWAYS)						\
-	x(initialize_subvolumes,	0)							\
-	x(snapshots_read,		PASS_ALWAYS)						\
-	x(check_topology,		0)							\
-	x(check_allocations,		PASS_FSCK)						\
-	x(set_may_go_rw,		PASS_ALWAYS|PASS_SILENT)				\
-	x(journal_replay,		PASS_ALWAYS)						\
-	x(check_alloc_info,		PASS_FSCK)						\
-	x(check_lrus,			PASS_FSCK)						\
-	x(check_btree_backpointers,	PASS_FSCK)						\
-	x(check_backpointers_to_extents,PASS_FSCK)						\
-	x(check_extents_to_backpointers,PASS_FSCK)						\
-	x(check_alloc_to_lru_refs,	PASS_FSCK)						\
-	x(fs_freespace_init,		PASS_ALWAYS|PASS_SILENT)				\
-	x(bucket_gens_init,		0)							\
-	x(check_snapshot_trees,		PASS_FSCK)						\
-	x(check_snapshots,		PASS_FSCK)						\
-	x(check_subvols,		PASS_FSCK)						\
-	x(delete_dead_snapshots,	PASS_FSCK|PASS_UNCLEAN)					\
-	x(fs_upgrade_for_subvolumes,	0)							\
-	x(check_inodes,			PASS_FSCK|PASS_UNCLEAN)					\
-	x(check_extents,		PASS_FSCK)						\
-	x(check_dirents,		PASS_FSCK)						\
-	x(check_xattrs,			PASS_FSCK)						\
-	x(check_root,			PASS_FSCK)						\
-	x(check_directory_structure,	PASS_FSCK)						\
-	x(check_nlinks,			PASS_FSCK)						\
-	x(fix_reflink_p,		0)							\
-
-enum bch_recovery_pass {
-#define x(n, when)	BCH_RECOVERY_PASS_##n,
-	BCH_RECOVERY_PASSES()
-#undef x
-};
-
 struct bch_fs {
 	struct closure		cl;
 
diff --git a/libbcachefs/bcachefs_format.h b/libbcachefs/bcachefs_format.h
index b771d80f..5ec218ee 100644
--- a/libbcachefs/bcachefs_format.h
+++ b/libbcachefs/bcachefs_format.h
@@ -1629,7 +1629,9 @@ struct bch_sb_field_journal_seq_blacklist {
 	x(major_minor,			BCH_VERSION(1,  0),		\
 	  0)								\
 	x(snapshot_skiplists,		BCH_VERSION(1,  1),		\
-	  BIT_ULL(BCH_RECOVERY_PASS_check_snapshots))
+	  BIT_ULL(BCH_RECOVERY_PASS_check_snapshots))			\
+	x(deleted_inodes,		BCH_VERSION(1,  2),		\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_inodes))
 
 enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min = 9,
@@ -2251,7 +2253,9 @@ enum btree_id_flags {
 	x(bucket_gens,		14,	0,					\
 	  BIT_ULL(KEY_TYPE_bucket_gens))					\
 	x(snapshot_trees,	15,	0,					\
-	  BIT_ULL(KEY_TYPE_snapshot_tree))
+	  BIT_ULL(KEY_TYPE_snapshot_tree))					\
+	x(deleted_inodes,	16,	BTREE_ID_SNAPSHOTS,			\
+	  BIT_ULL(KEY_TYPE_set))
 
 enum btree_id {
 #define x(name, nr, ...) BTREE_ID_##name = nr,
diff --git a/libbcachefs/fs-io.c b/libbcachefs/fs-io.c
index 6b691b2b..d433f4d5 100644
--- a/libbcachefs/fs-io.c
+++ b/libbcachefs/fs-io.c
@@ -35,7 +35,7 @@
 
 #include <trace/events/writeback.h>
 
-static void bch2_clamp_data_hole(struct inode *, u64 *, u64 *, unsigned);
+static int bch2_clamp_data_hole(struct inode *, u64 *, u64 *, unsigned, bool);
 
 struct folio_vec {
 	struct folio	*fv_folio;
@@ -3410,11 +3410,15 @@ static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		}
 
 		if (!(mode & FALLOC_FL_ZERO_RANGE)) {
-			ret = drop_locks_do(&trans,
-				(bch2_clamp_data_hole(&inode->v,
-						      &hole_start,
-						      &hole_end,
-						      opts.data_replicas), 0));
+			if (bch2_clamp_data_hole(&inode->v,
+						 &hole_start,
+						 &hole_end,
+						 opts.data_replicas, true))
+				ret = drop_locks_do(&trans,
+					(bch2_clamp_data_hole(&inode->v,
+							      &hole_start,
+							      &hole_end,
+							      opts.data_replicas, false), 0));
 			bch2_btree_iter_set_pos(&iter, POS(iter.pos.inode, hole_start));
 
 			if (ret)
@@ -3714,7 +3718,8 @@ static int folio_data_offset(struct folio *folio, loff_t pos,
 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
 				       loff_t start_offset,
 				       loff_t end_offset,
-				       unsigned min_replicas)
+				       unsigned min_replicas,
+				       bool nonblock)
 {
 	struct folio_batch fbatch;
 	pgoff_t start_index	= start_offset >> PAGE_SHIFT;
@@ -3731,7 +3736,13 @@ static loff_t bch2_seek_pagecache_data(struct inode *vinode,
 		for (i = 0; i < folio_batch_count(&fbatch); i++) {
 			struct folio *folio = fbatch.folios[i];
 
-			folio_lock(folio);
+			if (!nonblock) {
+				folio_lock(folio);
+			} else if (!folio_trylock(folio)) {
+				folio_batch_release(&fbatch);
+				return -EAGAIN;
+			}
+
 			offset = folio_data_offset(folio,
 					max(folio_pos(folio), start_offset),
 					min_replicas);
@@ -3796,7 +3807,7 @@ err:
 
 	if (next_data > offset)
 		next_data = bch2_seek_pagecache_data(&inode->v,
-						     offset, next_data, 0);
+					offset, next_data, 0, false);
 
 	if (next_data >= isize)
 		return -ENXIO;
@@ -3804,18 +3815,24 @@ err:
 	return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
 }
 
-static bool folio_hole_offset(struct address_space *mapping, loff_t *offset,
-			      unsigned min_replicas)
+static int folio_hole_offset(struct address_space *mapping, loff_t *offset,
+			      unsigned min_replicas, bool nonblock)
 {
 	struct folio *folio;
 	struct bch_folio *s;
 	unsigned i, sectors;
 	bool ret = true;
 
-	folio = filemap_lock_folio(mapping, *offset >> PAGE_SHIFT);
+	folio = __filemap_get_folio(mapping, *offset >> PAGE_SHIFT,
+				    !nonblock ? FGP_LOCK : 0, 0);
 	if (IS_ERR_OR_NULL(folio))
 		return true;
 
+	if (nonblock && !folio_trylock(folio)) {
+		folio_put(folio);
+		return -EAGAIN;
+	}
+
 	s = bch2_folio(folio);
 	if (!s)
 		goto unlock;
@@ -3833,37 +3850,51 @@ static bool folio_hole_offset(struct address_space *mapping, loff_t *offset,
 	ret = false;
 unlock:
 	folio_unlock(folio);
+	folio_put(folio);
 	return ret;
 }
 
 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
 				       loff_t start_offset,
 				       loff_t end_offset,
-				       unsigned min_replicas)
+				       unsigned min_replicas,
+				       bool nonblock)
 {
 	struct address_space *mapping = vinode->i_mapping;
 	loff_t offset = start_offset;
 
 	while (offset < end_offset &&
-	       !folio_hole_offset(mapping, &offset, min_replicas))
+	       !folio_hole_offset(mapping, &offset, min_replicas, nonblock))
 		;
 
 	return min(offset, end_offset);
 }
 
-static void bch2_clamp_data_hole(struct inode *inode,
-				 u64 *hole_start,
-				 u64 *hole_end,
-				 unsigned min_replicas)
+static int bch2_clamp_data_hole(struct inode *inode,
+				u64 *hole_start,
+				u64 *hole_end,
+				unsigned min_replicas,
+				bool nonblock)
 {
-	*hole_start = bch2_seek_pagecache_hole(inode,
-		*hole_start << 9, *hole_end << 9, min_replicas) >> 9;
+	loff_t ret;
+
+	ret = bch2_seek_pagecache_hole(inode,
+		*hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
+	if (ret < 0)
+		return ret;
+
+	*hole_start = ret;
 
 	if (*hole_start == *hole_end)
-		return;
+		return 0;
 
-	*hole_end = bch2_seek_pagecache_data(inode,
-		*hole_start << 9, *hole_end << 9, min_replicas) >> 9;
+	ret = bch2_seek_pagecache_data(inode,
+		*hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
+	if (ret < 0)
+		return ret;
+
+	*hole_end = ret;
+	return 0;
 }
 
 static loff_t bch2_seek_hole(struct file *file, u64 offset)
@@ -3895,12 +3926,12 @@ retry:
 			   BTREE_ITER_SLOTS, k, ret) {
 		if (k.k->p.inode != inode->v.i_ino) {
 			next_hole = bch2_seek_pagecache_hole(&inode->v,
-					offset, MAX_LFS_FILESIZE, 0);
+					offset, MAX_LFS_FILESIZE, 0, false);
 			break;
 		} else if (!bkey_extent_is_data(k.k)) {
 			next_hole = bch2_seek_pagecache_hole(&inode->v,
 					max(offset, bkey_start_offset(k.k) << 9),
-					k.k->p.offset << 9, 0);
+					k.k->p.offset << 9, 0, false);
 
 			if (next_hole < k.k->p.offset << 9)
 				break;
diff --git a/libbcachefs/fsck.c b/libbcachefs/fsck.c
index 8c186acc..0852dbe9 100644
--- a/libbcachefs/fsck.c
+++ b/libbcachefs/fsck.c
@@ -1123,73 +1123,100 @@ static int extent_ends_at(struct bch_fs *c,
 static int overlapping_extents_found(struct btree_trans *trans,
 				     enum btree_id btree,
 				     struct bpos pos1, struct bkey pos2,
-				     bool *fixed)
+				     bool *fixed,
+				     struct extent_end *extent_end)
 {
 	struct bch_fs *c = trans->c;
 	struct printbuf buf = PRINTBUF;
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u32 snapshot = min(pos1.snapshot, pos2.p.snapshot);
+	struct btree_iter iter1, iter2 = { NULL };
+	struct bkey_s_c k1, k2;
 	int ret;
 
 	BUG_ON(bkey_le(pos1, bkey_start_pos(&pos2)));
 
-	bch2_trans_iter_init(trans, &iter, btree, SPOS(pos1.inode, pos1.offset - 1, snapshot), 0);
-	k = bch2_btree_iter_peek_upto(&iter, POS(pos1.inode, U64_MAX));
-	ret = bkey_err(k);
+	bch2_trans_iter_init(trans, &iter1, btree, pos1,
+			     BTREE_ITER_ALL_SNAPSHOTS|
+			     BTREE_ITER_NOT_EXTENTS);
+	k1 = bch2_btree_iter_peek_upto(&iter1, POS(pos1.inode, U64_MAX));
+	ret = bkey_err(k1);
 	if (ret)
 		goto err;
 
 	prt_str(&buf, "\n  ");
-	bch2_bkey_val_to_text(&buf, c, k);
+	bch2_bkey_val_to_text(&buf, c, k1);
 
-	if (!bpos_eq(pos1, k.k->p)) {
-		bch_err(c, "%s: error finding first overlapping extent when repairing%s",
+	if (!bpos_eq(pos1, k1.k->p)) {
+		prt_str(&buf, "\n  wanted\n  ");
+		bch2_bpos_to_text(&buf, pos1);
+		prt_str(&buf, "\n  ");
+		bch2_bkey_to_text(&buf, &pos2);
+
+		bch_err(c, "%s: error finding first overlapping extent when repairing, got%s",
 			__func__, buf.buf);
 		ret = -BCH_ERR_internal_fsck_err;
 		goto err;
 	}
 
+	bch2_trans_copy_iter(&iter2, &iter1);
+
 	while (1) {
-		bch2_btree_iter_advance(&iter);
+		bch2_btree_iter_advance(&iter2);
 
-		k = bch2_btree_iter_peek_upto(&iter, POS(pos1.inode, U64_MAX));
-		ret = bkey_err(k);
+		k2 = bch2_btree_iter_peek_upto(&iter2, POS(pos1.inode, U64_MAX));
+		ret = bkey_err(k2);
 		if (ret)
 			goto err;
 
-		if (bkey_ge(k.k->p, pos2.p))
+		if (bpos_ge(k2.k->p, pos2.p))
 			break;
-
 	}
 
 	prt_str(&buf, "\n  ");
-	bch2_bkey_val_to_text(&buf, c, k);
+	bch2_bkey_val_to_text(&buf, c, k2);
 
-	if (bkey_gt(k.k->p, pos2.p) ||
-	    pos2.size != k.k->size) {
+	if (bpos_gt(k2.k->p, pos2.p) ||
+	    pos2.size != k2.k->size) {
 		bch_err(c, "%s: error finding seconding overlapping extent when repairing%s",
 			__func__, buf.buf);
 		ret = -BCH_ERR_internal_fsck_err;
 		goto err;
 	}
 
+	prt_printf(&buf, "\n  overwriting %s extent",
+		   pos1.snapshot >= pos2.p.snapshot ? "first" : "second");
+
 	if (fsck_err(c, "overlapping extents%s", buf.buf)) {
-		struct bpos update_pos = pos1.snapshot < pos2.p.snapshot ? pos1 : pos2.p;
-		struct btree_iter update_iter;
+		struct btree_iter *old_iter = &iter1;
+		struct disk_reservation res = { 0 };
 
-		struct bkey_i *update = bch2_bkey_get_mut(trans, &update_iter,
-						btree, update_pos,
-						BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
-		bch2_trans_iter_exit(trans, &update_iter);
-		if ((ret = PTR_ERR_OR_ZERO(update)))
+		if (pos1.snapshot < pos2.p.snapshot) {
+			old_iter = &iter2;
+			swap(k1, k2);
+		}
+
+		trans->extra_journal_res += bch2_bkey_sectors_compressed(k2);
+
+		ret =   bch2_trans_update_extent_overwrite(trans, old_iter,
+				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE,
+				k1, k2) ?:
+			bch2_trans_commit(trans, &res, NULL,
+				BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL);
+		bch2_disk_reservation_put(c, &res);
+
+		if (ret)
 			goto err;
 
 		*fixed = true;
+
+		if (pos1.snapshot == pos2.p.snapshot)
+			extent_end->offset = bkey_start_offset(&pos2);
+		else
+			ret = -BCH_ERR_transaction_restart_nested;
 	}
 fsck_err:
 err:
-	bch2_trans_iter_exit(trans, &iter);
+	bch2_trans_iter_exit(trans, &iter2);
+	bch2_trans_iter_exit(trans, &iter1);
 	printbuf_exit(&buf);
 	return ret;
 }
@@ -1199,11 +1226,11 @@ static int check_overlapping_extents(struct btree_trans *trans,
 			      struct extent_ends *extent_ends,
 			      struct bkey_s_c k,
 			      u32 equiv,
-			      struct btree_iter *iter)
+			      struct btree_iter *iter,
+			      bool *fixed)
 {
 	struct bch_fs *c = trans->c;
 	struct extent_end *i;
-	bool fixed = false;
 	int ret = 0;
 
 	/* transaction restart, running again */
@@ -1226,7 +1253,7 @@ static int check_overlapping_extents(struct btree_trans *trans,
 						SPOS(iter->pos.inode,
 						     i->offset,
 						     i->snapshot),
-						*k.k, &fixed);
+						*k.k, fixed, i);
 		if (ret)
 			goto err;
 	}
@@ -1237,7 +1264,7 @@ static int check_overlapping_extents(struct btree_trans *trans,
 
 	extent_ends->last_pos = k.k->p;
 err:
-	return ret ?: fixed;
+	return ret;
 }
 
 static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
@@ -1292,13 +1319,10 @@ static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
 			goto delete;
 
 		ret = check_overlapping_extents(trans, s, extent_ends, k,
-						equiv.snapshot, iter);
-		if (ret < 0)
-			goto err;
-
+						equiv.snapshot, iter,
+						&inode->recalculate_sums);
 		if (ret)
-			inode->recalculate_sums = true;
-		ret = 0;
+			goto err;
 	}
 
 	/*
@@ -1373,7 +1397,7 @@ int bch2_check_extents(struct bch_fs *c)
 
 	snapshots_seen_init(&s);
 	extent_ends_init(&extent_ends);
-	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 
 	ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_extents,
 			POS(BCACHEFS_ROOT_INO, 0),
diff --git a/libbcachefs/inode.c b/libbcachefs/inode.c
index 755cf7d1..fea21e1e 100644
--- a/libbcachefs/inode.c
+++ b/libbcachefs/inode.c
@@ -2,6 +2,7 @@
 
 #include "bcachefs.h"
 #include "btree_key_cache.h"
+#include "btree_write_buffer.h"
 #include "bkey_methods.h"
 #include "btree_update.h"
 #include "buckets.h"
@@ -519,6 +520,23 @@ void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c
 	__bch2_inode_unpacked_to_text(out, &inode);
 }
 
+static inline bool bkey_is_deleted_inode(struct bkey_s_c k)
+{
+	switch (k.k->type) {
+	case KEY_TYPE_inode:
+		return bkey_s_c_to_inode(k).v->bi_flags &
+			cpu_to_le32(BCH_INODE_UNLINKED);
+	case KEY_TYPE_inode_v2:
+		return bkey_s_c_to_inode_v2(k).v->bi_flags &
+			cpu_to_le32(BCH_INODE_UNLINKED);
+	case KEY_TYPE_inode_v3:
+		return bkey_s_c_to_inode_v3(k).v->bi_flags &
+			cpu_to_le64(BCH_INODE_UNLINKED);
+	default:
+		return false;
+	}
+}
+
 int bch2_trans_mark_inode(struct btree_trans *trans,
 			  enum btree_id btree_id, unsigned level,
 			  struct bkey_s_c old,
@@ -526,6 +544,8 @@ int bch2_trans_mark_inode(struct btree_trans *trans,
 			  unsigned flags)
 {
 	int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
+	bool old_deleted = bkey_is_deleted_inode(old);
+	bool new_deleted = bkey_is_deleted_inode(bkey_i_to_s_c(new));
 
 	if (nr) {
 		int ret = bch2_replicas_deltas_realloc(trans, 0);
@@ -537,6 +557,12 @@ int bch2_trans_mark_inode(struct btree_trans *trans,
 		d->nr_inodes += nr;
 	}
 
+	if (old_deleted != new_deleted) {
+		int ret = bch2_btree_bit_mod(trans, BTREE_ID_deleted_inodes, new->k.p, new_deleted);
+		if (ret)
+			return ret;
+	}
+
 	return 0;
 }
 
@@ -986,3 +1012,90 @@ err:
 
 	return ret ?: -BCH_ERR_transaction_restart_nested;
 }
+
+static int may_delete_deleted_inode(struct btree_trans *trans, struct bpos pos)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	struct bch_inode_unpacked inode;
+	int ret;
+
+	if (bch2_snapshot_is_internal_node(c, pos.snapshot))
+		return 0;
+
+	if (!fsck_err_on(c->sb.clean, c,
+			 "filesystem marked as clean but have deleted inode %llu:%u",
+			 pos.offset, pos.snapshot))
+		return 0;
+
+	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes, pos, BTREE_ITER_CACHED);
+	ret = bkey_err(k);
+	if (ret)
+		return ret;
+
+	ret = bkey_is_inode(k.k) ? 0 : -BCH_ERR_ENOENT_inode;
+	if (fsck_err_on(!bkey_is_inode(k.k), c,
+			"nonexistent inode %llu:%u in deleted_inodes btree",
+			pos.offset, pos.snapshot))
+		goto delete;
+
+	ret = bch2_inode_unpack(k, &inode);
+	if (ret)
+		goto err;
+
+	if (fsck_err_on(!(inode.bi_flags & BCH_INODE_UNLINKED), c,
+			"non-deleted inode %llu:%u in deleted_inodes btree",
+			pos.offset, pos.snapshot))
+		goto delete;
+
+	return 1;
+err:
+fsck_err:
+	return ret;
+delete:
+	return bch2_btree_bit_mod(trans, BTREE_ID_deleted_inodes, pos, false);
+}
+
+int bch2_delete_dead_inodes(struct bch_fs *c)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	ret = bch2_btree_write_buffer_flush_sync(&trans);
+	if (ret)
+		goto err;
+
+	/*
+	 * Weird transaction restart handling here because on successful delete,
+	 * bch2_inode_rm_snapshot() will return a nested transaction restart,
+	 * but we can't retry because the btree write buffer won't have been
+	 * flushed and we'd spin:
+	 */
+	for_each_btree_key(&trans, iter, BTREE_ID_deleted_inodes, POS_MIN,
+			   BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
+		ret = lockrestart_do(&trans, may_delete_deleted_inode(&trans, k.k->p));
+		if (ret < 0)
+			break;
+
+		if (ret) {
+			if (!test_bit(BCH_FS_RW, &c->flags)) {
+				bch2_trans_unlock(&trans);
+				bch2_fs_lazy_rw(c);
+			}
+
+			ret = bch2_inode_rm_snapshot(&trans, k.k->p.offset, k.k->p.snapshot);
+			if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart))
+				break;
+		}
+	}
+	bch2_trans_iter_exit(&trans, &iter);
+err:
+	bch2_trans_exit(&trans);
+
+	return ret;
+}
diff --git a/libbcachefs/inode.h b/libbcachefs/inode.h
index 1b9dc27e..22b24405 100644
--- a/libbcachefs/inode.h
+++ b/libbcachefs/inode.h
@@ -199,5 +199,6 @@ void bch2_inode_opts_get(struct bch_io_opts *, struct bch_fs *,
 			 struct bch_inode_unpacked *);
 
 int bch2_inode_rm_snapshot(struct btree_trans *, u64, u32);
+int bch2_delete_dead_inodes(struct bch_fs *);
 
 #endif /* _BCACHEFS_INODE_H */
diff --git a/libbcachefs/opts.c b/libbcachefs/opts.c
index 4d0daeba..960bb247 100644
--- a/libbcachefs/opts.c
+++ b/libbcachefs/opts.c
@@ -445,6 +445,13 @@ int bch2_parse_mount_opts(struct bch_fs *c, struct bch_opts *opts,
 	if (!options)
 		return 0;
 
+	/*
+	 * sys_fsconfig() is now occasionally providing us with option lists
+	 * starting with a comma - weird.
+	 */
+	if (*options == ',')
+		options++;
+
 	copied_opts = kstrdup(options, GFP_KERNEL);
 	if (!copied_opts)
 		return -1;
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index dcd4f9f4..55a233c2 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -1458,6 +1458,29 @@ use_clean:
 	if (ret)
 		goto err;
 
+	/* If we fixed errors, verify that fs is actually clean now: */
+	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
+	    test_bit(BCH_FS_ERRORS_FIXED, &c->flags) &&
+	    !test_bit(BCH_FS_ERRORS_NOT_FIXED, &c->flags) &&
+	    !test_bit(BCH_FS_ERROR, &c->flags)) {
+		bch_info(c, "Fixed errors, running fsck a second time to verify fs is clean");
+		clear_bit(BCH_FS_ERRORS_FIXED, &c->flags);
+
+		c->curr_recovery_pass = BCH_RECOVERY_PASS_check_alloc_info;
+
+		ret = bch2_run_recovery_passes(c);
+		if (ret)
+			goto err;
+
+		if (test_bit(BCH_FS_ERRORS_FIXED, &c->flags) ||
+		    test_bit(BCH_FS_ERRORS_NOT_FIXED, &c->flags)) {
+			bch_err(c, "Second fsck run was not clean");
+			set_bit(BCH_FS_ERRORS_NOT_FIXED, &c->flags);
+		}
+
+		set_bit(BCH_FS_ERRORS_FIXED, &c->flags);
+	}
+
 	if (enabled_qtypes(c)) {
 		bch_verbose(c, "reading quotas");
 		ret = bch2_fs_quota_read(c);
diff --git a/libbcachefs/recovery_types.h b/libbcachefs/recovery_types.h
new file mode 100644
index 00000000..abf1f834
--- /dev/null
+++ b/libbcachefs/recovery_types.h
@@ -0,0 +1,48 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_RECOVERY_TYPES_H
+#define _BCACHEFS_RECOVERY_TYPES_H
+
+#define PASS_SILENT		BIT(0)
+#define PASS_FSCK		BIT(1)
+#define PASS_UNCLEAN		BIT(2)
+#define PASS_ALWAYS		BIT(3)
+
+#define BCH_RECOVERY_PASSES()									\
+	x(alloc_read,			PASS_ALWAYS)						\
+	x(stripes_read,			PASS_ALWAYS)						\
+	x(initialize_subvolumes,	0)							\
+	x(snapshots_read,		PASS_ALWAYS)						\
+	x(check_topology,		0)							\
+	x(check_allocations,		PASS_FSCK)						\
+	x(set_may_go_rw,		PASS_ALWAYS|PASS_SILENT)				\
+	x(journal_replay,		PASS_ALWAYS)						\
+	x(check_alloc_info,		PASS_FSCK)						\
+	x(check_lrus,			PASS_FSCK)						\
+	x(check_btree_backpointers,	PASS_FSCK)						\
+	x(check_backpointers_to_extents,PASS_FSCK)						\
+	x(check_extents_to_backpointers,PASS_FSCK)						\
+	x(check_alloc_to_lru_refs,	PASS_FSCK)						\
+	x(fs_freespace_init,		PASS_ALWAYS|PASS_SILENT)				\
+	x(bucket_gens_init,		0)							\
+	x(check_snapshot_trees,		PASS_FSCK)						\
+	x(check_snapshots,		PASS_FSCK)						\
+	x(check_subvols,		PASS_FSCK)						\
+	x(delete_dead_snapshots,	PASS_FSCK|PASS_UNCLEAN)					\
+	x(fs_upgrade_for_subvolumes,	0)							\
+	x(check_inodes,			PASS_FSCK)						\
+	x(check_extents,		PASS_FSCK)						\
+	x(check_dirents,		PASS_FSCK)						\
+	x(check_xattrs,			PASS_FSCK)						\
+	x(check_root,			PASS_FSCK)						\
+	x(check_directory_structure,	PASS_FSCK)						\
+	x(check_nlinks,			PASS_FSCK)						\
+	x(delete_dead_inodes,		PASS_FSCK|PASS_UNCLEAN)					\
+	x(fix_reflink_p,		0)							\
+
+enum bch_recovery_pass {
+#define x(n, when)	BCH_RECOVERY_PASS_##n,
+	BCH_RECOVERY_PASSES()
+#undef x
+};
+
+#endif /* _BCACHEFS_RECOVERY_TYPES_H */
diff --git a/libbcachefs/tests.c b/libbcachefs/tests.c
index cef23d2c..1d4b0a58 100644
--- a/libbcachefs/tests.c
+++ b/libbcachefs/tests.c
@@ -503,6 +503,36 @@ static int test_extent_overwrite_all(struct bch_fs *c, u64 nr)
 		__test_extent_overwrite(c, 32, 64, 32, 128);
 }
 
+static int insert_test_overlapping_extent(struct bch_fs *c, u64 inum, u64 start, u32 len, u32 snapid)
+{
+	struct bkey_i_cookie k;
+	int ret;
+
+	bkey_cookie_init(&k.k_i);
+	k.k_i.k.p.inode	= inum;
+	k.k_i.k.p.offset = start + len;
+	k.k_i.k.p.snapshot = snapid;
+	k.k_i.k.size = len;
+
+	ret = bch2_trans_do(c, NULL, NULL, 0,
+		bch2_btree_insert_nonextent(&trans, BTREE_ID_extents, &k.k_i,
+					    BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE));
+	if (ret)
+		bch_err_fn(c, ret);
+	return ret;
+}
+
+static int test_extent_create_overlapping(struct bch_fs *c, u64 inum)
+{
+	return  insert_test_overlapping_extent(c, inum,  0, 16, U32_MAX - 2) ?: /* overwrite entire */
+		insert_test_overlapping_extent(c, inum,  2,  8, U32_MAX - 2) ?:
+		insert_test_overlapping_extent(c, inum,  4,  4, U32_MAX) ?:
+		insert_test_overlapping_extent(c, inum, 32,  8, U32_MAX - 2) ?: /* overwrite front/back */
+		insert_test_overlapping_extent(c, inum, 36,  8, U32_MAX) ?:
+		insert_test_overlapping_extent(c, inum, 60,  8, U32_MAX - 2) ?:
+		insert_test_overlapping_extent(c, inum, 64,  8, U32_MAX);
+}
+
 /* snapshot unit tests */
 
 /* Test skipping over keys in unrelated snapshots: */
@@ -901,6 +931,7 @@ int bch2_btree_perf_test(struct bch_fs *c, const char *testname,
 	perf_test(test_extent_overwrite_back);
 	perf_test(test_extent_overwrite_middle);
 	perf_test(test_extent_overwrite_all);
+	perf_test(test_extent_create_overlapping);
 
 	perf_test(test_snapshots);