Update bcachefs sources to 09d4c2acbf4c bcachefs: reconstruct_inode()

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

.bcachefs_revision

@@ -1 +1 @@
-794723fc10c4a1ff28d4b11c436277ba783f47e6
+09d4c2acbf4c864fef0f520bbcba256c9a19102e

Makefile.compiler

@@ -38,7 +38,7 @@ as-option = $(call try-run,\
 # Usage: aflags-y += $(call as-instr,instr,option1,option2)
 
 as-instr = $(call try-run,\
-	printf "%b\n" "$(1)" | $(CC) -Werror $(CLANG_FLAGS) $(KBUILD_AFLAGS) -c -x assembler-with-cpp -o "$$TMP" -,$(2),$(3))
+	printf "%b\n" "$(1)" | $(CC) -Werror $(CLANG_FLAGS) $(KBUILD_AFLAGS) -Wa$(comma)--fatal-warnings -c -x assembler-with-cpp -o "$$TMP" -,$(2),$(3))
 
 # __cc-option
 # Usage: MY_CFLAGS += $(call __cc-option,$(CC),$(MY_CFLAGS),-march=winchip-c6,-march=i586)

c_src/cmd_migrate.c

@@ -676,12 +676,12 @@ static int migrate_fs(const char		*fs_path,
 	struct dev_opts dev = dev_opts_default();
 
 	dev.path = dev_t_to_path(stat.st_dev);
-	dev.handle = bdev_open_by_path(dev.path, BLK_OPEN_READ|BLK_OPEN_WRITE, &dev, NULL);
+	dev.file = bdev_file_open_by_path(dev.path, BLK_OPEN_READ|BLK_OPEN_WRITE, &dev, NULL);
 
-	int ret = PTR_ERR_OR_ZERO(dev.handle);
+	int ret = PTR_ERR_OR_ZERO(dev.file);
 	if (ret < 0)
 		die("Error opening device to format %s: %s", dev.path, strerror(-ret));
-	dev.bdev = dev.handle->bdev;
+	dev.bdev = file_bdev(dev.file);
 
 	opt_set(fs_opts, block_size, get_blocksize(dev.bdev->bd_fd));
 

c_src/libbcachefs.h

@@ -52,7 +52,7 @@ static inline struct format_opts format_opts_default()
 }
 
 struct dev_opts {
-	struct bdev_handle *handle;
+	struct file	*file;
 	struct block_device *bdev;
 	char		*path;
 	u64		size;		/* bytes*/

c_src/tools-util.c

@@ -189,13 +189,13 @@ int open_for_format(struct dev_opts *dev, bool force)
 	const char *fs_type = NULL, *fs_label = NULL;
 	size_t fs_type_len, fs_label_len;
 
-	dev->handle = bdev_open_by_path(dev->path,
+	dev->file = bdev_file_open_by_path(dev->path,
 				BLK_OPEN_READ|BLK_OPEN_WRITE|BLK_OPEN_EXCL|BLK_OPEN_BUFFERED,
 				dev, NULL);
-	int ret = PTR_ERR_OR_ZERO(dev->handle);
+	int ret = PTR_ERR_OR_ZERO(dev->file);
 	if (ret < 0)
 		die("Error opening device to format %s: %s", dev->path, strerror(-ret));
-	dev->bdev = dev->handle->bdev;
+	dev->bdev = file_bdev(dev->file);
 
 	if (!(pr = blkid_new_probe()))
 		die("blkid error 1");

include/linux/blk_types.h

@@ -16,6 +16,28 @@ typedef void (bio_end_io_t) (struct bio *);
 
 #define BDEVNAME_SIZE	32
 
+typedef unsigned int __bitwise blk_mode_t;
+
+/* open for reading */
+#define BLK_OPEN_READ		((__force blk_mode_t)(1 << 0))
+/* open for writing */
+#define BLK_OPEN_WRITE		((__force blk_mode_t)(1 << 1))
+/* open exclusively (vs other exclusive openers */
+#define BLK_OPEN_EXCL		((__force blk_mode_t)(1 << 2))
+/* opened with O_NDELAY */
+#define BLK_OPEN_NDELAY		((__force blk_mode_t)(1 << 3))
+/* open for "writes" only for ioctls (specialy hack for floppy.c) */
+#define BLK_OPEN_WRITE_IOCTL	((__force blk_mode_t)(1 << 4))
+
+#define BLK_OPEN_BUFFERED	((__force blk_mode_t)(1 << 5))
+
+struct inode {
+	unsigned long		i_ino;
+	loff_t			i_size;
+	struct super_block	*i_sb;
+	blk_mode_t		mode;
+};
+
 struct request_queue {
 	struct backing_dev_info *backing_dev_info;
 };
@@ -34,6 +56,7 @@ struct block_device {
 	dev_t			bd_dev;
 	char			name[BDEVNAME_SIZE];
 	struct inode		*bd_inode;
+	struct inode		__bd_inode;
 	struct request_queue	queue;
 	void			*bd_holder;
 	struct gendisk *	bd_disk;

include/linux/blkdev.h

@@ -23,27 +23,6 @@ struct user_namespace;
 #define MINOR(dev)	((unsigned int) ((dev) & MINORMASK))
 #define MKDEV(ma,mi)	(((ma) << MINORBITS) | (mi))
 
-typedef unsigned int __bitwise blk_mode_t;
-
-/* open for reading */
-#define BLK_OPEN_READ		((__force blk_mode_t)(1 << 0))
-/* open for writing */
-#define BLK_OPEN_WRITE		((__force blk_mode_t)(1 << 1))
-/* open exclusively (vs other exclusive openers */
-#define BLK_OPEN_EXCL		((__force blk_mode_t)(1 << 2))
-/* opened with O_NDELAY */
-#define BLK_OPEN_NDELAY		((__force blk_mode_t)(1 << 3))
-/* open for "writes" only for ioctls (specialy hack for floppy.c) */
-#define BLK_OPEN_WRITE_IOCTL	((__force blk_mode_t)(1 << 4))
-
-#define BLK_OPEN_BUFFERED	((__force blk_mode_t)(1 << 5))
-
-struct inode {
-	unsigned long		i_ino;
-	loff_t			i_size;
-	struct super_block	*i_sb;
-};
-
 struct file {
 	struct inode		*f_inode;
 };
@@ -89,15 +68,14 @@ struct blk_holder_ops {
         void (*mark_dead)(struct block_device *bdev);
 };
 
-struct bdev_handle {
+static inline struct block_device *file_bdev(struct file *file)
-	struct block_device *bdev;
+{
-	void *holder;
+	return container_of(file->f_inode, struct block_device, __bd_inode);
-	blk_mode_t mode;
+}
-};
 
-void bdev_release(struct bdev_handle *);
+void fput(struct file *);
-struct bdev_handle *bdev_open_by_path(const char *, blk_mode_t, void *,
+struct file *bdev_file_open_by_path(const char *, blk_mode_t, void *,
-				      const struct blk_holder_ops *);
+				    const struct blk_holder_ops *);
 int lookup_bdev(const char *path, dev_t *);
 
 struct super_block {

include/linux/poison.h

@@ -92,4 +92,7 @@
 /********** VFS **********/
 #define VFS_PTR_POISON ((void *)(0xF5 + POISON_POINTER_DELTA))
 
+/********** lib/stackdepot.c **********/
+#define STACK_DEPOT_POISON ((void *)(0xD390 + POISON_POINTER_DELTA))
+
 #endif

libbcachefs/alloc_background.c

@@ -1713,34 +1713,37 @@ static int bch2_discard_one_bucket(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
+	if (a->v.dirty_sectors) {
-		a->v.gen++;
+		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
-		SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
+					       trans, "attempting to discard bucket with dirty data\n%s",
-		goto write;
+					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
-	}
-
-	if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
-		if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
-			bch2_trans_inconsistent(trans,
-				"clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
-				"%s",
-				a->v.journal_seq,
-				c->journal.flushed_seq_ondisk,
-				(bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 			ret = -EIO;
-		}
 		goto out;
 	}
 
 	if (a->v.data_type != BCH_DATA_need_discard) {
-		if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
+		if (data_type_is_empty(a->v.data_type) &&
-			bch2_trans_inconsistent(trans,
+		    BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
-				"bucket incorrectly set in need_discard btree\n"
+			a->v.gen++;
-				"%s",
+			SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
-				(bch2_bkey_val_to_text(&buf, c, k), buf.buf));
+			goto write;
-			ret = -EIO;
 		}
 
+		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
+					       trans, "bucket incorrectly set in need_discard btree\n"
+					       "%s",
+					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
+			ret = -EIO;
+		goto out;
+	}
+
+	if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
+		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
+					       trans, "clearing need_discard but journal_seq %llu > flushed_seq %llu\n%s",
+					       a->v.journal_seq,
+					       c->journal.flushed_seq_ondisk,
+					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
+			ret = -EIO;
 		goto out;
 	}
 
@@ -1835,6 +1838,7 @@ static int bch2_clear_bucket_needs_discard(struct btree_trans *trans, struct bpo
 	if (ret)
 		goto err;
 
+	BUG_ON(a->v.dirty_sectors);
 	SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
 	a->v.data_type = alloc_data_type(a->v, a->v.data_type);
 
@@ -1942,6 +1946,7 @@ static int invalidate_one_bucket(struct btree_trans *trans,
 		goto out;
 
 	BUG_ON(a->v.data_type != BCH_DATA_cached);
+	BUG_ON(a->v.dirty_sectors);
 
 	if (!a->v.cached_sectors)
 		bch_err(c, "invalidating empty bucket, confused");

libbcachefs/alloc_foreground.c

@@ -188,8 +188,10 @@ long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 {
 	switch (watermark) {
-	case BCH_WATERMARK_reclaim:
+	case BCH_WATERMARK_interior_updates:
 		return 0;
+	case BCH_WATERMARK_reclaim:
+		return OPEN_BUCKETS_COUNT / 6;
 	case BCH_WATERMARK_btree:
 	case BCH_WATERMARK_btree_copygc:
 		return OPEN_BUCKETS_COUNT / 4;

libbcachefs/alloc_types.h

@@ -22,7 +22,8 @@ struct bucket_alloc_state {
 	x(copygc)			\
 	x(btree)			\
 	x(btree_copygc)			\
-	x(reclaim)
+	x(reclaim)			\
+	x(interior_updates)
 
 enum bch_watermark {
 #define x(name)	BCH_WATERMARK_##name,

libbcachefs/backpointers.c

@@ -8,6 +8,7 @@
 #include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_write_buffer.h"
+#include "checksum.h"
 #include "error.h"
 
 #include <linux/mm.h>
@@ -418,6 +419,84 @@ struct extents_to_bp_state {
 	struct bkey_buf last_flushed;
 };
 
+static int drop_dev_and_update(struct btree_trans *trans, enum btree_id btree,
+			       struct bkey_s_c extent, unsigned dev)
+{
+	struct bkey_i *n = bch2_bkey_make_mut_noupdate(trans, extent);
+	int ret = PTR_ERR_OR_ZERO(n);
+	if (ret)
+		return ret;
+
+	bch2_bkey_drop_device(bkey_i_to_s(n), dev);
+	return bch2_btree_insert_trans(trans, btree, n, 0);
+}
+
+static int check_extent_checksum(struct btree_trans *trans,
+				 enum btree_id btree, struct bkey_s_c extent,
+				 enum btree_id o_btree, struct bkey_s_c extent2, unsigned dev)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(extent);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	struct printbuf buf = PRINTBUF;
+	void *data_buf = NULL;
+	struct bio *bio = NULL;
+	size_t bytes;
+	int ret = 0;
+
+	if (bkey_is_btree_ptr(extent.k))
+		return false;
+
+	bkey_for_each_ptr_decode(extent.k, ptrs, p, entry)
+		if (p.ptr.dev == dev)
+			goto found;
+	BUG();
+found:
+	if (!p.crc.csum_type)
+		return false;
+
+	bytes = p.crc.compressed_size << 9;
+
+	struct bch_dev *ca = bch_dev_bkey_exists(c, dev);
+	if (!bch2_dev_get_ioref(ca, READ))
+		return false;
+
+	data_buf = kvmalloc(bytes, GFP_KERNEL);
+	if (!data_buf) {
+		ret = -ENOMEM;
+		goto err;
+	}
+
+	bio = bio_alloc(ca->disk_sb.bdev, 1, REQ_OP_READ, GFP_KERNEL);
+	bio->bi_iter.bi_sector = p.ptr.offset;
+	bch2_bio_map(bio, data_buf, bytes);
+	ret = submit_bio_wait(bio);
+	if (ret)
+		goto err;
+
+	prt_str(&buf, "extents pointing to same space, but first extent checksum bad:");
+	prt_printf(&buf, "\n  %s ", bch2_btree_id_str(btree));
+	bch2_bkey_val_to_text(&buf, c, extent);
+	prt_printf(&buf, "\n  %s ", bch2_btree_id_str(o_btree));
+	bch2_bkey_val_to_text(&buf, c, extent2);
+
+	struct nonce nonce = extent_nonce(extent.k->version, p.crc);
+	struct bch_csum csum = bch2_checksum(c, p.crc.csum_type, nonce, data_buf, bytes);
+	if (fsck_err_on(bch2_crc_cmp(csum, p.crc.csum),
+			c, dup_backpointer_to_bad_csum_extent,
+			"%s", buf.buf))
+		ret = drop_dev_and_update(trans, btree, extent, dev) ?: 1;
+fsck_err:
+err:
+	if (bio)
+		bio_put(bio);
+	kvfree(data_buf);
+	percpu_ref_put(&ca->io_ref);
+	printbuf_exit(&buf);
+	return ret;
+}
+
 static int check_bp_exists(struct btree_trans *trans,
 			   struct extents_to_bp_state *s,
 			   struct bpos bucket,
@@ -425,7 +504,8 @@ static int check_bp_exists(struct btree_trans *trans,
 			   struct bkey_s_c orig_k)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter bp_iter = { NULL };
+	struct btree_iter bp_iter = {};
+	struct btree_iter other_extent_iter = {};
 	struct printbuf buf = PRINTBUF;
 	struct bkey_s_c bp_k;
 	struct bkey_buf tmp;
@@ -433,13 +513,19 @@ static int check_bp_exists(struct btree_trans *trans,
 
 	bch2_bkey_buf_init(&tmp);
 
+	if (!bch2_dev_bucket_exists(c, bucket)) {
+		prt_str(&buf, "extent for nonexistent device:bucket ");
+		bch2_bpos_to_text(&buf, bucket);
+		prt_str(&buf, "\n  ");
+		bch2_bkey_val_to_text(&buf, c, orig_k);
+		bch_err(c, "%s", buf.buf);
+		return -BCH_ERR_fsck_repair_unimplemented;
+	}
+
 	if (bpos_lt(bucket, s->bucket_start) ||
 	    bpos_gt(bucket, s->bucket_end))
 		return 0;
 
-	if (!bch2_dev_bucket_exists(c, bucket))
-		goto missing;
-
 	bp_k = bch2_bkey_get_iter(trans, &bp_iter, BTREE_ID_backpointers,
 				  bucket_pos_to_bp(c, bucket, bp.bucket_offset),
 				  0);
@@ -465,21 +551,94 @@ static int check_bp_exists(struct btree_trans *trans,
 			ret = -BCH_ERR_transaction_restart_write_buffer_flush;
 			goto out;
 		}
-		goto missing;
+
+		goto check_existing_bp;
 	}
 out:
 err:
 fsck_err:
+	bch2_trans_iter_exit(trans, &other_extent_iter);
 	bch2_trans_iter_exit(trans, &bp_iter);
 	bch2_bkey_buf_exit(&tmp, c);
 	printbuf_exit(&buf);
 	return ret;
+check_existing_bp:
+	/* Do we have a backpointer for a different extent? */
+	if (bp_k.k->type != KEY_TYPE_backpointer)
+		goto missing;
+
+	struct bch_backpointer other_bp = *bkey_s_c_to_backpointer(bp_k).v;
+
+	struct bkey_s_c other_extent =
+		bch2_backpointer_get_key(trans, &other_extent_iter, bp_k.k->p, other_bp, 0);
+	ret = bkey_err(other_extent);
+	if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
+		ret = 0;
+	if (ret)
+		goto err;
+
+	if (!other_extent.k)
+		goto missing;
+
+	if (bch2_extents_match(orig_k, other_extent)) {
+		printbuf_reset(&buf);
+		prt_printf(&buf, "duplicate versions of same extent, deleting smaller\n  ");
+		bch2_bkey_val_to_text(&buf, c, orig_k);
+		prt_str(&buf, "\n  ");
+		bch2_bkey_val_to_text(&buf, c, other_extent);
+		bch_err(c, "%s", buf.buf);
+
+		if (other_extent.k->size <= orig_k.k->size) {
+			ret = drop_dev_and_update(trans, other_bp.btree_id, other_extent, bucket.inode);
+			if (ret)
+				goto err;
+			goto out;
+		} else {
+			ret = drop_dev_and_update(trans, bp.btree_id, orig_k, bucket.inode);
+			if (ret)
+				goto err;
+			goto missing;
+		}
+	}
+
+	ret = check_extent_checksum(trans, other_bp.btree_id, other_extent, bp.btree_id, orig_k, bucket.inode);
+	if (ret < 0)
+		goto err;
+	if (ret) {
+		ret = 0;
+		goto missing;
+	}
+
+	ret = check_extent_checksum(trans, bp.btree_id, orig_k, other_bp.btree_id, other_extent, bucket.inode);
+	if (ret < 0)
+		goto err;
+	if (ret) {
+		ret = 0;
+		goto out;
+	}
+
+	printbuf_reset(&buf);
+	prt_printf(&buf, "duplicate extents pointing to same space on dev %llu\n  ", bucket.inode);
+	bch2_bkey_val_to_text(&buf, c, orig_k);
+	prt_str(&buf, "\n  ");
+	bch2_bkey_val_to_text(&buf, c, other_extent);
+	bch_err(c, "%s", buf.buf);
+	ret = -BCH_ERR_fsck_repair_unimplemented;
+	goto err;
 missing:
+	printbuf_reset(&buf);
 	prt_printf(&buf, "missing backpointer for btree=%s l=%u ",
 	       bch2_btree_id_str(bp.btree_id), bp.level);
 	bch2_bkey_val_to_text(&buf, c, orig_k);
-	prt_printf(&buf, "\nbp pos ");
+	prt_printf(&buf, "\n  got:   ");
-	bch2_bpos_to_text(&buf, bp_iter.pos);
+	bch2_bkey_val_to_text(&buf, c, bp_k);
+
+	struct bkey_i_backpointer n_bp_k;
+	bkey_backpointer_init(&n_bp_k.k_i);
+	n_bp_k.k.p = bucket_pos_to_bp(trans->c, bucket, bp.bucket_offset);
+	n_bp_k.v = bp;
+	prt_printf(&buf, "\n  want:  ");
+	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&n_bp_k.k_i));
 
 	if (fsck_err(c, ptr_to_missing_backpointer, "%s", buf.buf))
 		ret = bch2_bucket_backpointer_mod(trans, bucket, bp, orig_k, true);

libbcachefs/bcachefs.h

@@ -615,6 +615,7 @@ struct bch_dev {
  */
 
 #define BCH_FS_FLAGS()			\
+	x(new_fs)			\
 	x(started)			\
 	x(may_go_rw)			\
 	x(rw)				\
@@ -797,6 +798,7 @@ struct bch_fs {
 		u64		features;
 		u64		compat;
 		unsigned long	errors_silent[BITS_TO_LONGS(BCH_SB_ERR_MAX)];
+		u64		btrees_lost_data;
 	}			sb;
 
 
@@ -826,7 +828,6 @@ struct bch_fs {
 	struct btree_root	btree_roots_known[BTREE_ID_NR];
 	DARRAY(struct btree_root) btree_roots_extra;
 	struct mutex		btree_root_lock;
-	unsigned long		btrees_lost_data; /* bitmask */
 
 	struct btree_cache	btree_cache;
 

libbcachefs/bcachefs_format.h

@@ -818,6 +818,7 @@ struct bch_sb_field_ext {
 	struct bch_sb_field	field;
 	__le64			recovery_passes_required[2];
 	__le64			errors_silent[8];
+	__le64			btrees_lost_data;
 };
 
 struct bch_sb_field_downgrade_entry {

libbcachefs/btree_io.c

@@ -1264,11 +1264,11 @@ out:
 	return retry_read;
 fsck_err:
 	if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
-	    ret == -BCH_ERR_btree_node_read_err_must_retry)
+	    ret == -BCH_ERR_btree_node_read_err_must_retry) {
 		retry_read = 1;
-	else {
+	} else {
 		set_btree_node_read_error(b);
-		set_bit(b->c.btree_id, &c->btrees_lost_data);
+		bch2_btree_lost_data(c, b->c.btree_id);
 	}
 	goto out;
 }
@@ -1330,7 +1330,7 @@ start:
 
 		if (!can_retry) {
 			set_btree_node_read_error(b);
-			set_bit(b->c.btree_id, &c->btrees_lost_data);
+			bch2_btree_lost_data(c, b->c.btree_id);
 			break;
 		}
 	}
@@ -1532,7 +1532,7 @@ fsck_err:
 
 	if (ret) {
 		set_btree_node_read_error(b);
-		set_bit(b->c.btree_id, &c->btrees_lost_data);
+		bch2_btree_lost_data(c, b->c.btree_id);
 	} else if (*saw_error)
 		bch2_btree_node_rewrite_async(c, b);
 
@@ -1669,7 +1669,7 @@ void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
 			bch2_fatal_error(c);
 
 		set_btree_node_read_error(b);
-		set_bit(b->c.btree_id, &c->btrees_lost_data);
+		bch2_btree_lost_data(c, b->c.btree_id);
 		clear_btree_node_read_in_flight(b);
 		wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
 		printbuf_exit(&buf);
@@ -1866,7 +1866,7 @@ static void btree_node_write_work(struct work_struct *work)
 	} else {
 		ret = bch2_trans_do(c, NULL, NULL, 0,
 			bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
-					BCH_WATERMARK_reclaim|
+					BCH_WATERMARK_interior_updates|
 					BCH_TRANS_COMMIT_journal_reclaim|
 					BCH_TRANS_COMMIT_no_enospc|
 					BCH_TRANS_COMMIT_no_check_rw,

libbcachefs/btree_trans_commit.c

@@ -887,6 +887,7 @@ int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 			    int ret, unsigned long trace_ip)
 {
 	struct bch_fs *c = trans->c;
+	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 
 	switch (ret) {
 	case -BCH_ERR_btree_insert_btree_node_full:
@@ -905,7 +906,7 @@ int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 		 * flag
 		 */
 		if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
-		    (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
+		    watermark < BCH_WATERMARK_reclaim) {
 			ret = -BCH_ERR_journal_reclaim_would_deadlock;
 			break;
 		}

libbcachefs/btree_update_interior.c

@@ -26,6 +26,13 @@
 
 #include <linux/random.h>
 
+const char * const bch2_btree_update_modes[] = {
+#define x(t) #t,
+	BCH_WATERMARKS()
+#undef x
+	NULL
+};
+
 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
 				  btree_path_idx_t, struct btree *, struct keylist *);
 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
@@ -303,7 +310,7 @@ static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 	struct open_buckets obs = { .nr = 0 };
 	struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
-	unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
+	unsigned nr_reserve = watermark < BCH_WATERMARK_reclaim
 		? BTREE_NODE_RESERVE
 		: 0;
 	int ret;
@@ -687,7 +694,7 @@ static void btree_update_nodes_written(struct btree_update *as)
 	 * which may require allocations as well.
 	 */
 	ret = commit_do(trans, &as->disk_res, &journal_seq,
-			BCH_WATERMARK_reclaim|
+			BCH_WATERMARK_interior_updates|
 			BCH_TRANS_COMMIT_no_enospc|
 			BCH_TRANS_COMMIT_no_check_rw|
 			BCH_TRANS_COMMIT_journal_reclaim,
@@ -846,11 +853,11 @@ static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 	mutex_lock(&c->btree_interior_update_lock);
 	list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 
-	BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
+	BUG_ON(as->mode != BTREE_UPDATE_none);
 	BUG_ON(!btree_node_dirty(b));
 	BUG_ON(!b->c.level);
 
-	as->mode	= BTREE_INTERIOR_UPDATING_NODE;
+	as->mode	= BTREE_UPDATE_node;
 	as->b		= b;
 
 	set_btree_node_write_blocked(b);
@@ -873,7 +880,7 @@ static void btree_update_reparent(struct btree_update *as,
 	lockdep_assert_held(&c->btree_interior_update_lock);
 
 	child->b = NULL;
-	child->mode = BTREE_INTERIOR_UPDATING_AS;
+	child->mode = BTREE_UPDATE_update;
 
 	bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
 			      bch2_update_reparent_journal_pin_flush);
@@ -884,7 +891,7 @@ static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 	struct bkey_i *insert = &b->key;
 	struct bch_fs *c = as->c;
 
-	BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
+	BUG_ON(as->mode != BTREE_UPDATE_none);
 
 	BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 	       ARRAY_SIZE(as->journal_entries));
@@ -898,7 +905,7 @@ static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 	mutex_lock(&c->btree_interior_update_lock);
 	list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 
-	as->mode	= BTREE_INTERIOR_UPDATING_ROOT;
+	as->mode	= BTREE_UPDATE_root;
 	mutex_unlock(&c->btree_interior_update_lock);
 }
 
@@ -1076,7 +1083,7 @@ static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *
 	struct bch_fs *c = as->c;
 	u64 start_time = as->start_time;
 
-	BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
+	BUG_ON(as->mode == BTREE_UPDATE_none);
 
 	if (as->took_gc_lock)
 		up_read(&as->c->gc_lock);
@@ -1121,7 +1128,7 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 		unsigned journal_flags = watermark|JOURNAL_RES_GET_CHECK;
 
 		if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
-		    watermark != BCH_WATERMARK_reclaim)
+		    watermark < BCH_WATERMARK_reclaim)
 			journal_flags |= JOURNAL_RES_GET_NONBLOCK;
 
 		ret = drop_locks_do(trans,
@@ -1172,7 +1179,8 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 	as->c		= c;
 	as->start_time	= start_time;
 	as->ip_started	= _RET_IP_;
-	as->mode	= BTREE_INTERIOR_NO_UPDATE;
+	as->mode	= BTREE_UPDATE_none;
+	as->watermark	= watermark;
 	as->took_gc_lock = true;
 	as->btree_id	= path->btree_id;
 	as->update_level = update_level;
@@ -1217,7 +1225,7 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 		 */
 		if (bch2_err_matches(ret, ENOSPC) &&
 		    (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
-		    watermark != BCH_WATERMARK_reclaim) {
+		    watermark < BCH_WATERMARK_reclaim) {
 			ret = -BCH_ERR_journal_reclaim_would_deadlock;
 			goto err;
 		}
@@ -2509,18 +2517,25 @@ void bch2_btree_root_alloc_fake(struct bch_fs *c, enum btree_id id, unsigned lev
 	bch2_trans_run(c, __bch2_btree_root_alloc_fake(trans, id, level));
 }
 
+static void bch2_btree_update_to_text(struct printbuf *out, struct btree_update *as)
+{
+	prt_printf(out, "%ps: btree=%s watermark=%s mode=%s nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
+		   (void *) as->ip_started,
+		   bch2_btree_id_str(as->btree_id),
+		   bch2_watermarks[as->watermark],
+		   bch2_btree_update_modes[as->mode],
+		   as->nodes_written,
+		   closure_nr_remaining(&as->cl),
+		   as->journal.seq);
+}
+
 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
 {
 	struct btree_update *as;
 
 	mutex_lock(&c->btree_interior_update_lock);
 	list_for_each_entry(as, &c->btree_interior_update_list, list)
-		prt_printf(out, "%ps: mode=%u nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
+		bch2_btree_update_to_text(out, as);
-			   (void *) as->ip_started,
-			   as->mode,
-			   as->nodes_written,
-			   closure_nr_remaining(&as->cl),
-			   as->journal.seq);
 	mutex_unlock(&c->btree_interior_update_lock);
 }
 

libbcachefs/btree_update_interior.h

@@ -12,6 +12,18 @@
 
 int bch2_btree_node_check_topology(struct btree_trans *, struct btree *);
 
+#define BTREE_UPDATE_MODES()	\
+	x(none)			\
+	x(node)			\
+	x(root)			\
+	x(update)
+
+enum btree_update_mode {
+#define x(n)	BTREE_UPDATE_##n,
+	BTREE_UPDATE_MODES()
+#undef x
+};
+
 /*
  * Tracks an in progress split/rewrite of a btree node and the update to the
  * parent node:
@@ -39,14 +51,8 @@ struct btree_update {
 	struct list_head		list;
 	struct list_head		unwritten_list;
 
-	/* What kind of update are we doing? */
+	enum btree_update_mode		mode;
-	enum {
+	enum bch_watermark		watermark;
-		BTREE_INTERIOR_NO_UPDATE,
-		BTREE_INTERIOR_UPDATING_NODE,
-		BTREE_INTERIOR_UPDATING_ROOT,
-		BTREE_INTERIOR_UPDATING_AS,
-	} mode;
-
 	unsigned			nodes_written:1;
 	unsigned			took_gc_lock:1;
 
@@ -56,7 +62,7 @@ struct btree_update {
 	struct disk_reservation		disk_res;
 
 	/*
-	 * BTREE_INTERIOR_UPDATING_NODE:
+	 * BTREE_UPDATE_node:
 	 * The update that made the new nodes visible was a regular update to an
 	 * existing interior node - @b. We can't write out the update to @b
 	 * until the new nodes we created are finished writing, so we block @b

libbcachefs/buckets.h

@@ -226,6 +226,7 @@ static inline u64 bch2_dev_buckets_reserved(struct bch_dev *ca, enum bch_waterma
 		fallthrough;
 	case BCH_WATERMARK_btree_copygc:
 	case BCH_WATERMARK_reclaim:
+	case BCH_WATERMARK_interior_updates:
 		break;
 	}
 

libbcachefs/data_update.c

@@ -580,8 +580,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 				move_ctxt_wait_event(ctxt,
 						(locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 									  PTR_BUCKET_POS(c, &p.ptr), 0)) ||
-						(!atomic_read(&ctxt->read_sectors) &&
+						list_empty(&ctxt->ios));
-						 !atomic_read(&ctxt->write_sectors)));
 
 				if (!locked)
 					bch2_bucket_nocow_lock(&c->nocow_locks,

libbcachefs/error.c

@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "bcachefs.h"
 #include "error.h"
+#include "journal.h"
 #include "recovery_passes.h"
 #include "super.h"
 #include "thread_with_file.h"
@@ -16,7 +17,8 @@ bool bch2_inconsistent_error(struct bch_fs *c)
 		return false;
 	case BCH_ON_ERROR_ro:
 		if (bch2_fs_emergency_read_only(c))
-			bch_err(c, "inconsistency detected - emergency read only");
+			bch_err(c, "inconsistency detected - emergency read only at journal seq %llu",
+				journal_cur_seq(&c->journal));
 		return true;
 	case BCH_ON_ERROR_panic:
 		panic(bch2_fmt(c, "panic after error"));

libbcachefs/eytzinger.c

@@ -115,7 +115,7 @@ static void swap_bytes(void *a, void *b, size_t n)
 
 struct wrapper {
 	cmp_func_t cmp;
-	swap_func_t swap_f;
+	swap_func_t swap;
 };
 
 /*
@@ -125,7 +125,7 @@ struct wrapper {
 static void do_swap(void *a, void *b, size_t size, swap_r_func_t swap_func, const void *priv)
 {
 	if (swap_func == SWAP_WRAPPER) {
-		((const struct wrapper *)priv)->swap_f(a, b, (int)size);
+		((const struct wrapper *)priv)->swap(a, b, (int)size);
 		return;
 	}
 
@@ -174,7 +174,7 @@ void eytzinger0_sort_r(void *base, size_t n, size_t size,
 	int i, c, r;
 
 	/* called from 'sort' without swap function, let's pick the default */
-	if (swap_func == SWAP_WRAPPER && !((struct wrapper *)priv)->swap_f)
+	if (swap_func == SWAP_WRAPPER && !((struct wrapper *)priv)->swap)
 		swap_func = NULL;
 
 	if (!swap_func) {
@@ -227,7 +227,7 @@ void eytzinger0_sort(void *base, size_t n, size_t size,
 {
 	struct wrapper w = {
 		.cmp  = cmp_func,
-		.swap_f = swap_func,
+		.swap = swap_func,
 	};
 
 	return eytzinger0_sort_r(base, n, size, _CMP_WRAPPER, SWAP_WRAPPER, &w);

libbcachefs/fsck.c

@@ -63,9 +63,7 @@ static int subvol_lookup(struct btree_trans *trans, u32 subvol,
 			 u32 *snapshot, u64 *inum)
 {
 	struct bch_subvolume s;
-	int ret;
+	int ret = bch2_subvolume_get(trans, subvol, false, 0, &s);
-
-	ret = bch2_subvolume_get(trans, subvol, false, 0, &s);
 
 	*snapshot = le32_to_cpu(s.snapshot);
 	*inum = le64_to_cpu(s.inode);
@@ -158,9 +156,10 @@ static int __remove_dirent(struct btree_trans *trans, struct bpos pos)
 
 	bch2_trans_iter_init(trans, &iter, BTREE_ID_dirents, pos, BTREE_ITER_INTENT);
 
-	ret = bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
+	ret =   bch2_btree_iter_traverse(&iter) ?:
-				  &dir_hash_info, &iter,
+		bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
-				  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+				    &dir_hash_info, &iter,
+				    BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 	bch2_trans_iter_exit(trans, &iter);
 err:
 	bch_err_fn(c, ret);
@@ -169,7 +168,8 @@ err:
 
 /* Get lost+found, create if it doesn't exist: */
 static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
-			    struct bch_inode_unpacked *lostfound)
+			    struct bch_inode_unpacked *lostfound,
+			    u64 reattaching_inum)
 {
 	struct bch_fs *c = trans->c;
 	struct qstr lostfound_str = QSTR("lost+found");
@@ -184,19 +184,36 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 		return ret;
 
 	subvol_inum root_inum = { .subvol = le32_to_cpu(st.master_subvol) };
-	u32 subvol_snapshot;
 
-	ret = subvol_lookup(trans, le32_to_cpu(st.master_subvol),
+	struct bch_subvolume subvol;
-			    &subvol_snapshot, &root_inum.inum);
+	ret = bch2_subvolume_get(trans, le32_to_cpu(st.master_subvol),
-	bch_err_msg(c, ret, "looking up root subvol");
+				 false, 0, &subvol);
+	bch_err_msg(c, ret, "looking up root subvol %u for snapshot %u",
+		    le32_to_cpu(st.master_subvol), snapshot);
 	if (ret)
 		return ret;
 
+	if (!subvol.inode) {
+		struct btree_iter iter;
+		struct bkey_i_subvolume *subvol = bch2_bkey_get_mut_typed(trans, &iter,
+				BTREE_ID_subvolumes, POS(0, le32_to_cpu(st.master_subvol)),
+				0, subvolume);
+		ret = PTR_ERR_OR_ZERO(subvol);
+		if (ret)
+			return ret;
+
+		subvol->v.inode = cpu_to_le64(reattaching_inum);
+		bch2_trans_iter_exit(trans, &iter);
+	}
+
+	root_inum.inum = le64_to_cpu(subvol.inode);
+
 	struct bch_inode_unpacked root_inode;
 	struct bch_hash_info root_hash_info;
 	u32 root_inode_snapshot = snapshot;
 	ret = lookup_inode(trans, root_inum.inum, &root_inode, &root_inode_snapshot);
-	bch_err_msg(c, ret, "looking up root inode");
+	bch_err_msg(c, ret, "looking up root inode %llu for subvol %u",
+		    root_inum.inum, le32_to_cpu(st.master_subvol));
 	if (ret)
 		return ret;
 
@@ -292,7 +309,7 @@ static int reattach_inode(struct btree_trans *trans,
 		snprintf(name_buf, sizeof(name_buf), "%llu", inode->bi_inum);
 	}
 
-	ret = lookup_lostfound(trans, dirent_snapshot, &lostfound);
+	ret = lookup_lostfound(trans, dirent_snapshot, &lostfound, inode->bi_inum);
 	if (ret)
 		return ret;
 
@@ -363,6 +380,112 @@ static int reattach_subvol(struct btree_trans *trans, struct bkey_s_c_subvolume
 	return ret;
 }
 
+static int reconstruct_subvol(struct btree_trans *trans, u32 snapshotid, u32 subvolid, u64 inum)
+{
+	struct bch_fs *c = trans->c;
+
+	if (!bch2_snapshot_is_leaf(c, snapshotid)) {
+		bch_err(c, "need to reconstruct subvol, but have interior node snapshot");
+		return -BCH_ERR_fsck_repair_unimplemented;
+	}
+
+	/*
+	 * If inum isn't set, that means we're being called from check_dirents,
+	 * not check_inodes - the root of this subvolume doesn't exist or we
+	 * would have found it there:
+	 */
+	if (!inum) {
+		struct btree_iter inode_iter = {};
+		struct bch_inode_unpacked new_inode;
+		u64 cpu = raw_smp_processor_id();
+
+		bch2_inode_init_early(c, &new_inode);
+		bch2_inode_init_late(&new_inode, bch2_current_time(c), 0, 0, S_IFDIR|0755, 0, NULL);
+
+		new_inode.bi_subvol = subvolid;
+
+		int ret = bch2_inode_create(trans, &inode_iter, &new_inode, snapshotid, cpu) ?:
+			  bch2_btree_iter_traverse(&inode_iter) ?:
+			  bch2_inode_write(trans, &inode_iter, &new_inode);
+		bch2_trans_iter_exit(trans, &inode_iter);
+		if (ret)
+			return ret;
+
+		inum = new_inode.bi_inum;
+	}
+
+	bch_info(c, "reconstructing subvol %u with root inode %llu", subvolid, inum);
+
+	struct bkey_i_subvolume *new_subvol = bch2_trans_kmalloc(trans, sizeof(*new_subvol));
+	int ret = PTR_ERR_OR_ZERO(new_subvol);
+	if (ret)
+		return ret;
+
+	bkey_subvolume_init(&new_subvol->k_i);
+	new_subvol->k.p.offset	= subvolid;
+	new_subvol->v.snapshot	= cpu_to_le32(snapshotid);
+	new_subvol->v.inode	= cpu_to_le64(inum);
+	ret = bch2_btree_insert_trans(trans, BTREE_ID_subvolumes, &new_subvol->k_i, 0);
+	if (ret)
+		return ret;
+
+	struct btree_iter iter;
+	struct bkey_i_snapshot *s = bch2_bkey_get_mut_typed(trans, &iter,
+			BTREE_ID_snapshots, POS(0, snapshotid),
+			0, snapshot);
+	ret = PTR_ERR_OR_ZERO(s);
+	bch_err_msg(c, ret, "getting snapshot %u", snapshotid);
+	if (ret)
+		return ret;
+
+	u32 snapshot_tree = le32_to_cpu(s->v.tree);
+
+	s->v.subvol = cpu_to_le32(subvolid);
+	SET_BCH_SNAPSHOT_SUBVOL(&s->v, true);
+	bch2_trans_iter_exit(trans, &iter);
+
+	struct bkey_i_snapshot_tree *st = bch2_bkey_get_mut_typed(trans, &iter,
+			BTREE_ID_snapshot_trees, POS(0, snapshot_tree),
+			0, snapshot_tree);
+	ret = PTR_ERR_OR_ZERO(st);
+	bch_err_msg(c, ret, "getting snapshot tree %u", snapshot_tree);
+	if (ret)
+		return ret;
+
+	if (!st->v.master_subvol)
+		st->v.master_subvol = cpu_to_le32(subvolid);
+
+	bch2_trans_iter_exit(trans, &iter);
+	return 0;
+}
+
+static int reconstruct_inode(struct btree_trans *trans, u32 snapshot, u64 inum, u64 size, unsigned mode)
+{
+	struct bch_fs *c = trans->c;
+	struct bch_inode_unpacked new_inode;
+
+	bch2_inode_init_early(c, &new_inode);
+	bch2_inode_init_late(&new_inode, bch2_current_time(c), 0, 0, mode|0755, 0, NULL);
+	new_inode.bi_size = size;
+	new_inode.bi_inum = inum;
+
+	return __bch2_fsck_write_inode(trans, &new_inode, snapshot);
+}
+
+static int reconstruct_reg_inode(struct btree_trans *trans, u32 snapshot, u64 inum)
+{
+	struct btree_iter iter = {};
+
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents, SPOS(inum, U64_MAX, snapshot), 0);
+	struct bkey_s_c k = bch2_btree_iter_peek_prev(&iter);
+	bch2_trans_iter_exit(trans, &iter);
+	int ret = bkey_err(k);
+	if (ret)
+		return ret;
+
+	return reconstruct_inode(trans, snapshot, inum, k.k->p.offset << 9, S_IFREG);
+}
+
 struct snapshots_seen_entry {
 	u32				id;
 	u32				equiv;
@@ -1064,6 +1187,11 @@ static int check_inode(struct btree_trans *trans,
 		if (ret && !bch2_err_matches(ret, ENOENT))
 			goto err;
 
+		if (ret && (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_subvolumes))) {
+			ret = reconstruct_subvol(trans, k.k->p.snapshot, u.bi_subvol, u.bi_inum);
+			goto do_update;
+		}
+
 		if (fsck_err_on(ret,
 				c, inode_bi_subvol_missing,
 				"inode %llu:%u bi_subvol points to missing subvolume %u",
@@ -1081,7 +1209,7 @@ static int check_inode(struct btree_trans *trans,
 			do_update = true;
 		}
 	}
-
+do_update:
 	if (do_update) {
 		ret = __bch2_fsck_write_inode(trans, &u, iter->pos.snapshot);
 		bch_err_msg(c, ret, "in fsck updating inode");
@@ -1130,8 +1258,8 @@ static int check_i_sectors_notnested(struct btree_trans *trans, struct inode_wal
 			i->count = count2;
 
 		if (i->count != count2) {
-			bch_err(c, "fsck counted i_sectors wrong for inode %llu:%u: got %llu should be %llu",
+			bch_err_ratelimited(c, "fsck counted i_sectors wrong for inode %llu:%u: got %llu should be %llu",
-				w->last_pos.inode, i->snapshot, i->count, count2);
+					    w->last_pos.inode, i->snapshot, i->count, count2);
 			return -BCH_ERR_internal_fsck_err;
 		}
 
@@ -1371,10 +1499,6 @@ static int check_overlapping_extents(struct btree_trans *trans,
 			goto err;
 	}
 
-	ret = extent_ends_at(c, extent_ends, seen, k);
-	if (ret)
-		goto err;
-
 	extent_ends->last_pos = k.k->p;
 err:
 	return ret;
@@ -1438,6 +1562,17 @@ static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
 		goto err;
 
 	if (k.k->type != KEY_TYPE_whiteout) {
+		if (!i && (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_inodes))) {
+			ret =   reconstruct_reg_inode(trans, k.k->p.snapshot, k.k->p.inode) ?:
+				bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
+			if (ret)
+				goto err;
+
+			inode->last_pos.inode--;
+			ret = -BCH_ERR_transaction_restart_nested;
+			goto err;
+		}
+
 		if (fsck_err_on(!i, c, extent_in_missing_inode,
 				"extent in missing inode:\n  %s",
 				(printbuf_reset(&buf),
@@ -1504,6 +1639,12 @@ static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
 
 		i->seen_this_pos = true;
 	}
+
+	if (k.k->type != KEY_TYPE_whiteout) {
+		ret = extent_ends_at(c, extent_ends, s, k);
+		if (ret)
+			goto err;
+	}
 out:
 err:
 fsck_err:
@@ -1584,8 +1725,8 @@ static int check_subdir_count_notnested(struct btree_trans *trans, struct inode_
 			return count2;
 
 		if (i->count != count2) {
-			bch_err(c, "fsck counted subdirectories wrong: got %llu should be %llu",
+			bch_err_ratelimited(c, "fsck counted subdirectories wrong for inum %llu:%u: got %llu should be %llu",
-				i->count, count2);
+					    w->last_pos.inode, i->snapshot, i->count, count2);
 			i->count = count2;
 			if (i->inode.bi_nlink == i->count)
 				continue;
@@ -1782,6 +1923,7 @@ static int check_dirent_to_subvol(struct btree_trans *trans, struct btree_iter *
 	u32 parent_subvol = le32_to_cpu(d.v->d_parent_subvol);
 	u32 target_subvol = le32_to_cpu(d.v->d_child_subvol);
 	u32 parent_snapshot;
+	u32 new_parent_subvol = 0;
 	u64 parent_inum;
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
@@ -1790,6 +1932,27 @@ static int check_dirent_to_subvol(struct btree_trans *trans, struct btree_iter *
 	if (ret && !bch2_err_matches(ret, ENOENT))
 		return ret;
 
+	if (ret ||
+	    (!ret && !bch2_snapshot_is_ancestor(c, parent_snapshot, d.k->p.snapshot))) {
+		int ret2 = find_snapshot_subvol(trans, d.k->p.snapshot, &new_parent_subvol);
+		if (ret2 && !bch2_err_matches(ret, ENOENT))
+			return ret2;
+	}
+
+	if (ret &&
+	    !new_parent_subvol &&
+	    (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_subvolumes))) {
+		/*
+		 * Couldn't find a subvol for dirent's snapshot - but we lost
+		 * subvols, so we need to reconstruct:
+		 */
+		ret = reconstruct_subvol(trans, d.k->p.snapshot, parent_subvol, 0);
+		if (ret)
+			return ret;
+
+		parent_snapshot = d.k->p.snapshot;
+	}
+
 	if (fsck_err_on(ret, c, dirent_to_missing_parent_subvol,
 			"dirent parent_subvol points to missing subvolume\n%s",
 			(bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf)) ||
@@ -1798,10 +1961,10 @@ static int check_dirent_to_subvol(struct btree_trans *trans, struct btree_iter *
 			"dirent not visible in parent_subvol (not an ancestor of subvol snap %u)\n%s",
 			parent_snapshot,
 			(bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf))) {
-		u32 new_parent_subvol;
+		if (!new_parent_subvol) {
-		ret = find_snapshot_subvol(trans, d.k->p.snapshot, &new_parent_subvol);
+			bch_err(c, "could not find a subvol for snapshot %u", d.k->p.snapshot);
-		if (ret)
+			return -BCH_ERR_fsck_repair_unimplemented;
-			goto err;
+		}
 
 		struct bkey_i_dirent *new_dirent = bch2_bkey_make_mut_typed(trans, iter, &d.s_c, 0, dirent);
 		ret = PTR_ERR_OR_ZERO(new_dirent);
@@ -1847,9 +2010,16 @@ static int check_dirent_to_subvol(struct btree_trans *trans, struct btree_iter *
 
 	ret = lookup_inode(trans, target_inum, &subvol_root, &target_snapshot);
 	if (ret && !bch2_err_matches(ret, ENOENT))
-		return ret;
+		goto err;
 
-	if (fsck_err_on(parent_subvol != subvol_root.bi_parent_subvol,
+	if (ret) {
+		bch_err(c, "subvol %u points to missing inode root %llu", target_subvol, target_inum);
+		ret = -BCH_ERR_fsck_repair_unimplemented;
+		ret = 0;
+		goto err;
+	}
+
+	if (fsck_err_on(!ret && parent_subvol != subvol_root.bi_parent_subvol,
 			c, inode_bi_parent_wrong,
 			"subvol root %llu has wrong bi_parent_subvol: got %u, should be %u",
 			target_inum,
@@ -1857,13 +2027,13 @@ static int check_dirent_to_subvol(struct btree_trans *trans, struct btree_iter *
 		subvol_root.bi_parent_subvol = parent_subvol;
 		ret = __bch2_fsck_write_inode(trans, &subvol_root, target_snapshot);
 		if (ret)
-			return ret;
+			goto err;
 	}
 
 	ret = check_dirent_target(trans, iter, d, &subvol_root,
 				  target_snapshot);
 	if (ret)
-		return ret;
+		goto err;
 out:
 err:
 fsck_err:
@@ -1880,7 +2050,6 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 			struct snapshots_seen *s)
 {
 	struct bch_fs *c = trans->c;
-	struct bkey_s_c_dirent d;
 	struct inode_walker_entry *i;
 	struct printbuf buf = PRINTBUF;
 	struct bpos equiv;
@@ -1919,6 +2088,17 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &dir->inodes.data[0].inode);
 	dir->first_this_inode = false;
 
+	if (!i && (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_inodes))) {
+		ret =   reconstruct_inode(trans, k.k->p.snapshot, k.k->p.inode, 0, S_IFDIR) ?:
+			bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
+		if (ret)
+			goto err;
+
+		dir->last_pos.inode--;
+		ret = -BCH_ERR_transaction_restart_nested;
+		goto err;
+	}
+
 	if (fsck_err_on(!i, c, dirent_in_missing_dir_inode,
 			"dirent in nonexisting directory:\n%s",
 			(printbuf_reset(&buf),
@@ -1953,7 +2133,7 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 	if (k.k->type != KEY_TYPE_dirent)
 		goto out;
 
-	d = bkey_s_c_to_dirent(k);
+	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
 
 	if (d.v->d_type == DT_SUBVOL) {
 		ret = check_dirent_to_subvol(trans, iter, d);

libbcachefs/logged_ops.c

@@ -37,7 +37,6 @@ static int resume_logged_op(struct btree_trans *trans, struct btree_iter *iter,
 	const struct bch_logged_op_fn *fn = logged_op_fn(k.k->type);
 	struct bkey_buf sk;
 	u32 restart_count = trans->restart_count;
-	int ret;
 
 	if (!fn)
 		return 0;
@@ -45,11 +44,11 @@ static int resume_logged_op(struct btree_trans *trans, struct btree_iter *iter,
 	bch2_bkey_buf_init(&sk);
 	bch2_bkey_buf_reassemble(&sk, c, k);
 
-	ret =   drop_locks_do(trans, (bch2_fs_lazy_rw(c), 0)) ?:
+	fn->resume(trans, sk.k);
-		fn->resume(trans, sk.k) ?: trans_was_restarted(trans, restart_count);
 
 	bch2_bkey_buf_exit(&sk, c);
-	return ret;
+
+	return trans_was_restarted(trans, restart_count);
 }
 
 int bch2_resume_logged_ops(struct bch_fs *c)

libbcachefs/opts.c

@@ -7,6 +7,7 @@
 #include "disk_groups.h"
 #include "error.h"
 #include "opts.h"
+#include "recovery_passes.h"
 #include "super-io.h"
 #include "util.h"
 
@@ -205,6 +206,9 @@ const struct bch_option bch2_opt_table[] = {
 #define OPT_STR(_choices)	.type = BCH_OPT_STR,			\
 				.min = 0, .max = ARRAY_SIZE(_choices),	\
 				.choices = _choices
+#define OPT_STR_NOLIMIT(_choices)	.type = BCH_OPT_STR,		\
+				.min = 0, .max = U64_MAX,		\
+				.choices = _choices
 #define OPT_FN(_fn)		.type = BCH_OPT_FN, .fn	= _fn
 
 #define x(_name, _bits, _flags, _type, _sb_opt, _default, _hint, _help)	\

libbcachefs/opts.h

@@ -362,7 +362,12 @@ enum fsck_err_opts {
 	  OPT_FS|OPT_MOUNT,						\
 	  OPT_BOOL(),							\
 	  BCH2_NO_SB_OPT,		false,				\
-	  NULL,		"Don't replay the journal")			\
+	  NULL,		"Exit recovery immediately prior to journal replay")\
+	x(recovery_pass_last,		u8,				\
+	  OPT_FS|OPT_MOUNT,						\
+	  OPT_STR_NOLIMIT(bch2_recovery_passes),			\
+	  BCH2_NO_SB_OPT,		0,				\
+	  NULL,		"Exit recovery after specified pass")		\
 	x(retain_recovery_info,		u8,				\
 	  0,								\
 	  OPT_BOOL(),							\

libbcachefs/recovery.c

@@ -33,6 +33,20 @@
 
 #define QSTR(n) { { { .len = strlen(n) } }, .name = n }
 
+void bch2_btree_lost_data(struct bch_fs *c, enum btree_id btree)
+{
+	u64 b = BIT_ULL(btree);
+
+	if (!(c->sb.btrees_lost_data & b)) {
+		bch_err(c, "flagging btree %s lost data", bch2_btree_id_str(btree));
+
+		mutex_lock(&c->sb_lock);
+		bch2_sb_field_get(c->disk_sb.sb, ext)->btrees_lost_data |= cpu_to_le64(b);
+		bch2_write_super(c);
+		mutex_unlock(&c->sb_lock);
+	}
+}
+
 static bool btree_id_is_alloc(enum btree_id id)
 {
 	switch (id) {
@@ -272,7 +286,8 @@ int bch2_journal_replay(struct bch_fs *c)
 	bch2_trans_put(trans);
 	trans = NULL;
 
-	if (!c->opts.retain_recovery_info)
+	if (!c->opts.retain_recovery_info &&
+	    c->recovery_pass_done >= BCH_RECOVERY_PASS_journal_replay)
 		bch2_journal_keys_put_initial(c);
 
 	replay_now_at(j, j->replay_journal_seq_end);
@@ -468,8 +483,8 @@ static int read_btree_roots(struct bch_fs *c)
 				c->recovery_passes_explicit |= BIT_ULL(BCH_RECOVERY_PASS_check_topology);
 			}
 
-			set_bit(i, &c->btrees_lost_data);
 			ret = 0;
+			bch2_btree_lost_data(c, i);
 		}
 	}
 
@@ -590,27 +605,14 @@ int bch2_fs_recovery(struct bch_fs *c)
 		goto err;
 	}
 
-	if (c->opts.fsck && c->opts.norecovery) {
+	if (c->opts.norecovery)
-		bch_err(c, "cannot select both norecovery and fsck");
+		c->opts.recovery_pass_last = BCH_RECOVERY_PASS_journal_replay - 1;
-		ret = -EINVAL;
-		goto err;
-	}
-
-	c->opts.retain_recovery_info	|= c->opts.norecovery;
-	c->opts.nochanges		|= c->opts.norecovery;
 
 	if (!c->opts.nochanges) {
 		mutex_lock(&c->sb_lock);
+		struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
 		bool write_sb = false;
 
-		struct bch_sb_field_ext *ext =
-			bch2_sb_field_get_minsize(&c->disk_sb, ext, sizeof(*ext) / sizeof(u64));
-		if (!ext) {
-			ret = -BCH_ERR_ENOSPC_sb;
-			mutex_unlock(&c->sb_lock);
-			goto err;
-		}
-
 		if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb)) {
 			ext->recovery_passes_required[0] |=
 				cpu_to_le64(bch2_recovery_passes_to_stable(BIT_ULL(BCH_RECOVERY_PASS_check_topology)));
@@ -841,6 +843,7 @@ use_clean:
 	}
 
 	mutex_lock(&c->sb_lock);
+	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
 	bool write_sb = false;
 
 	if (BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb) != le16_to_cpu(c->disk_sb.sb->version)) {
@@ -854,15 +857,18 @@ use_clean:
 		write_sb = true;
 	}
 
-	if (!test_bit(BCH_FS_error, &c->flags)) {
+	if (!test_bit(BCH_FS_error, &c->flags) &&
-		struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+	    !bch2_is_zero(ext->errors_silent, sizeof(ext->errors_silent))) {
-		if (ext &&
+		memset(ext->errors_silent, 0, sizeof(ext->errors_silent));
-		    (!bch2_is_zero(ext->recovery_passes_required, sizeof(ext->recovery_passes_required)) ||
+		write_sb = true;
-		     !bch2_is_zero(ext->errors_silent, sizeof(ext->errors_silent)))) {
+	}
-			memset(ext->recovery_passes_required, 0, sizeof(ext->recovery_passes_required));
+
-			memset(ext->errors_silent, 0, sizeof(ext->errors_silent));
+	if (c->opts.fsck &&
-			write_sb = true;
+	    !test_bit(BCH_FS_error, &c->flags) &&
-		}
+	    c->recovery_pass_done == BCH_RECOVERY_PASS_NR - 1 &&
+	    ext->btrees_lost_data) {
+		ext->btrees_lost_data = 0;
+		write_sb = true;
 	}
 
 	if (c->opts.fsck &&
@@ -932,6 +938,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	int ret;
 
 	bch_notice(c, "initializing new filesystem");
+	set_bit(BCH_FS_new_fs, &c->flags);
 
 	mutex_lock(&c->sb_lock);
 	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);

libbcachefs/recovery.h

@@ -2,6 +2,8 @@
 #ifndef _BCACHEFS_RECOVERY_H
 #define _BCACHEFS_RECOVERY_H
 
+void bch2_btree_lost_data(struct bch_fs *, enum btree_id);
+
 int bch2_journal_replay(struct bch_fs *);
 
 int bch2_fs_recovery(struct bch_fs *);

libbcachefs/recovery_passes.c

@@ -17,6 +17,7 @@
 #include "snapshot.h"
 #include "subvolume.h"
 #include "super.h"
+#include "super-io.h"
 
 const char * const bch2_recovery_passes[] = {
 #define x(_fn, ...)	#_fn,
@@ -27,7 +28,7 @@ const char * const bch2_recovery_passes[] = {
 
 static int bch2_check_allocations(struct bch_fs *c)
 {
-	return bch2_gc(c, true, c->opts.norecovery);
+	return bch2_gc(c, true, false);
 }
 
 static int bch2_set_may_go_rw(struct bch_fs *c)
@@ -59,18 +60,23 @@ static struct recovery_pass_fn recovery_pass_fns[] = {
 #undef x
 };
 
-u64 bch2_recovery_passes_to_stable(u64 v)
+static const u8 passes_to_stable_map[] = {
-{
-	static const u8 map[] = {
 #define x(n, id, ...)	[BCH_RECOVERY_PASS_##n] = BCH_RECOVERY_PASS_STABLE_##n,
 	BCH_RECOVERY_PASSES()
 #undef x
-	};
+};
 
+static enum bch_recovery_pass_stable bch2_recovery_pass_to_stable(enum bch_recovery_pass pass)
+{
+	return passes_to_stable_map[pass];
+}
+
+u64 bch2_recovery_passes_to_stable(u64 v)
+{
 	u64 ret = 0;
-	for (unsigned i = 0; i < ARRAY_SIZE(map); i++)
+	for (unsigned i = 0; i < ARRAY_SIZE(passes_to_stable_map); i++)
 		if (v & BIT_ULL(i))
-			ret |= BIT_ULL(map[i]);
+			ret |= BIT_ULL(passes_to_stable_map[i]);
 	return ret;
 }
 
@@ -113,6 +119,38 @@ int bch2_run_explicit_recovery_pass(struct bch_fs *c,
 	}
 }
 
+int bch2_run_explicit_recovery_pass_persistent(struct bch_fs *c,
+					       enum bch_recovery_pass pass)
+{
+	enum bch_recovery_pass_stable s = bch2_recovery_pass_to_stable(pass);
+
+	mutex_lock(&c->sb_lock);
+	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+
+	if (!test_bit_le64(s, ext->recovery_passes_required)) {
+		__set_bit_le64(s, ext->recovery_passes_required);
+		bch2_write_super(c);
+	}
+	mutex_unlock(&c->sb_lock);
+
+	return bch2_run_explicit_recovery_pass(c, pass);
+}
+
+static void bch2_clear_recovery_pass_required(struct bch_fs *c,
+					      enum bch_recovery_pass pass)
+{
+	enum bch_recovery_pass_stable s = bch2_recovery_pass_to_stable(pass);
+
+	mutex_lock(&c->sb_lock);
+	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+
+	if (test_bit_le64(s, ext->recovery_passes_required)) {
+		__clear_bit_le64(s, ext->recovery_passes_required);
+		bch2_write_super(c);
+	}
+	mutex_unlock(&c->sb_lock);
+}
+
 u64 bch2_fsck_recovery_passes(void)
 {
 	u64 ret = 0;
@@ -127,8 +165,6 @@ static bool should_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pa
 {
 	struct recovery_pass_fn *p = recovery_pass_fns + pass;
 
-	if (c->opts.norecovery && pass > BCH_RECOVERY_PASS_snapshots_read)
-		return false;
 	if (c->recovery_passes_explicit & BIT_ULL(pass))
 		return true;
 	if ((p->when & PASS_FSCK) && c->opts.fsck)
@@ -184,6 +220,10 @@ int bch2_run_recovery_passes(struct bch_fs *c)
 	int ret = 0;
 
 	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns)) {
+		if (c->opts.recovery_pass_last &&
+		    c->curr_recovery_pass > c->opts.recovery_pass_last)
+			break;
+
 		if (should_run_recovery_pass(c, c->curr_recovery_pass)) {
 			unsigned pass = c->curr_recovery_pass;
 
@@ -196,8 +236,13 @@ int bch2_run_recovery_passes(struct bch_fs *c)
 
 			c->recovery_passes_complete |= BIT_ULL(c->curr_recovery_pass);
 		}
-		c->curr_recovery_pass++;
+
 		c->recovery_pass_done = max(c->recovery_pass_done, c->curr_recovery_pass);
+
+		if (!test_bit(BCH_FS_error, &c->flags))
+			bch2_clear_recovery_pass_required(c, c->curr_recovery_pass);
+
+		c->curr_recovery_pass++;
 	}
 
 	return ret;

libbcachefs/recovery_passes.h

@@ -9,6 +9,7 @@ u64 bch2_recovery_passes_from_stable(u64 v);
 u64 bch2_fsck_recovery_passes(void);
 
 int bch2_run_explicit_recovery_pass(struct bch_fs *, enum bch_recovery_pass);
+int bch2_run_explicit_recovery_pass_persistent(struct bch_fs *, enum bch_recovery_pass);
 
 int bch2_run_online_recovery_passes(struct bch_fs *);
 int bch2_run_recovery_passes(struct bch_fs *);

libbcachefs/recovery_passes_types.h

@@ -32,6 +32,7 @@
 	x(check_alloc_to_lru_refs,		15, PASS_ONLINE|PASS_FSCK)	\
 	x(fs_freespace_init,			16, PASS_ALWAYS|PASS_SILENT)	\
 	x(bucket_gens_init,			17, 0)				\
+	x(reconstruct_snapshots,		38, 0)				\
 	x(check_snapshot_trees,			18, PASS_ONLINE|PASS_FSCK)	\
 	x(check_snapshots,			19, PASS_ONLINE|PASS_FSCK)	\
 	x(check_subvols,			20, PASS_ONLINE|PASS_FSCK)	\

libbcachefs/sb-errors_types.h

@@ -268,7 +268,9 @@
 	x(btree_node_bkey_bad_u64s,				260)	\
 	x(btree_node_topology_empty_interior_node,		261)	\
 	x(btree_ptr_v2_min_key_bad,				262)	\
-	x(btree_root_unreadable_and_scan_found_nothing,		263)
+	x(btree_root_unreadable_and_scan_found_nothing,		263)	\
+	x(snapshot_node_missing,				264)	\
+	x(dup_backpointer_to_bad_csum_extent,			265)
 
 enum bch_sb_error_id {
 #define x(t, n) BCH_FSCK_ERR_##t = n,

libbcachefs/snapshot.c

@@ -8,6 +8,7 @@
 #include "errcode.h"
 #include "error.h"
 #include "fs.h"
+#include "recovery_passes.h"
 #include "snapshot.h"
 
 #include <linux/random.h>
@@ -131,7 +132,7 @@ bool __bch2_snapshot_is_ancestor(struct bch_fs *c, u32 id, u32 ancestor)
 	rcu_read_lock();
 	struct snapshot_table *t = rcu_dereference(c->snapshots);
 
-	if (unlikely(c->recovery_pass_done <= BCH_RECOVERY_PASS_check_snapshots)) {
+	if (unlikely(c->recovery_pass_done < BCH_RECOVERY_PASS_check_snapshots)) {
 		ret = __bch2_snapshot_is_ancestor_early(t, id, ancestor);
 		goto out;
 	}
@@ -574,6 +575,13 @@ static int check_snapshot_tree(struct btree_trans *trans,
 		u32 subvol_id;
 
 		ret = bch2_snapshot_tree_master_subvol(trans, root_id, &subvol_id);
+		bch_err_fn(c, ret);
+
+		if (bch2_err_matches(ret, ENOENT)) { /* nothing to be done here */
+			ret = 0;
+			goto err;
+		}
+
 		if (ret)
 			goto err;
 
@@ -731,7 +739,6 @@ static int check_snapshot(struct btree_trans *trans,
 	u32 parent_id = bch2_snapshot_parent_early(c, k.k->p.offset);
 	u32 real_depth;
 	struct printbuf buf = PRINTBUF;
-	bool should_have_subvol;
 	u32 i, id;
 	int ret = 0;
 
@@ -777,7 +784,7 @@ static int check_snapshot(struct btree_trans *trans,
 		}
 	}
 
-	should_have_subvol = BCH_SNAPSHOT_SUBVOL(&s) &&
+	bool should_have_subvol = BCH_SNAPSHOT_SUBVOL(&s) &&
 		!BCH_SNAPSHOT_DELETED(&s);
 
 	if (should_have_subvol) {
@@ -879,6 +886,154 @@ int bch2_check_snapshots(struct bch_fs *c)
 	return ret;
 }
 
+static int check_snapshot_exists(struct btree_trans *trans, u32 id)
+{
+	struct bch_fs *c = trans->c;
+
+	if (bch2_snapshot_equiv(c, id))
+		return 0;
+
+	u32 tree_id;
+	int ret = bch2_snapshot_tree_create(trans, id, 0, &tree_id);
+	if (ret)
+		return ret;
+
+	struct bkey_i_snapshot *snapshot = bch2_trans_kmalloc(trans, sizeof(*snapshot));
+	ret = PTR_ERR_OR_ZERO(snapshot);
+	if (ret)
+		return ret;
+
+	bkey_snapshot_init(&snapshot->k_i);
+	snapshot->k.p		= POS(0, id);
+	snapshot->v.tree	= cpu_to_le32(tree_id);
+	snapshot->v.btime.lo	= cpu_to_le64(bch2_current_time(c));
+
+	return  bch2_btree_insert_trans(trans, BTREE_ID_snapshots, &snapshot->k_i, 0) ?:
+		bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0,
+				   bkey_s_c_null, bkey_i_to_s(&snapshot->k_i), 0) ?:
+		bch2_snapshot_set_equiv(trans, bkey_i_to_s_c(&snapshot->k_i));
+}
+
+/* Figure out which snapshot nodes belong in the same tree: */
+struct snapshot_tree_reconstruct {
+	enum btree_id			btree;
+	struct bpos			cur_pos;
+	snapshot_id_list		cur_ids;
+	DARRAY(snapshot_id_list)	trees;
+};
+
+static void snapshot_tree_reconstruct_exit(struct snapshot_tree_reconstruct *r)
+{
+	darray_for_each(r->trees, i)
+		darray_exit(i);
+	darray_exit(&r->trees);
+	darray_exit(&r->cur_ids);
+}
+
+static inline bool same_snapshot(struct snapshot_tree_reconstruct *r, struct bpos pos)
+{
+	return r->btree == BTREE_ID_inodes
+		? r->cur_pos.offset == pos.offset
+		: r->cur_pos.inode == pos.inode;
+}
+
+static inline bool snapshot_id_lists_have_common(snapshot_id_list *l, snapshot_id_list *r)
+{
+	darray_for_each(*l, i)
+		if (snapshot_list_has_id(r, *i))
+			return true;
+	return false;
+}
+
+static void snapshot_id_list_to_text(struct printbuf *out, snapshot_id_list *s)
+{
+	bool first = true;
+	darray_for_each(*s, i) {
+		if (!first)
+			prt_char(out, ' ');
+		first = false;
+		prt_printf(out, "%u", *i);
+	}
+}
+
+static int snapshot_tree_reconstruct_next(struct bch_fs *c, struct snapshot_tree_reconstruct *r)
+{
+	if (r->cur_ids.nr) {
+		darray_for_each(r->trees, i)
+			if (snapshot_id_lists_have_common(i, &r->cur_ids)) {
+				int ret = snapshot_list_merge(c, i, &r->cur_ids);
+				if (ret)
+					return ret;
+				goto out;
+			}
+		darray_push(&r->trees, r->cur_ids);
+		darray_init(&r->cur_ids);
+	}
+out:
+	r->cur_ids.nr = 0;
+	return 0;
+}
+
+static int get_snapshot_trees(struct bch_fs *c, struct snapshot_tree_reconstruct *r, struct bpos pos)
+{
+	if (!same_snapshot(r, pos))
+		snapshot_tree_reconstruct_next(c, r);
+	r->cur_pos = pos;
+	return snapshot_list_add_nodup(c, &r->cur_ids, pos.snapshot);
+}
+
+int bch2_reconstruct_snapshots(struct bch_fs *c)
+{
+	struct btree_trans *trans = bch2_trans_get(c);
+	struct printbuf buf = PRINTBUF;
+	struct snapshot_tree_reconstruct r = {};
+	int ret = 0;
+
+	for (unsigned btree = 0; btree < BTREE_ID_NR; btree++) {
+		if (btree_type_has_snapshots(btree)) {
+			r.btree = btree;
+
+			ret = for_each_btree_key(trans, iter, btree, POS_MIN,
+					BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_PREFETCH, k, ({
+				get_snapshot_trees(c, &r, k.k->p);
+			}));
+			if (ret)
+				goto err;
+
+			snapshot_tree_reconstruct_next(c, &r);
+		}
+	}
+
+	darray_for_each(r.trees, t) {
+		printbuf_reset(&buf);
+		snapshot_id_list_to_text(&buf, t);
+
+		darray_for_each(*t, id) {
+			if (fsck_err_on(!bch2_snapshot_equiv(c, *id),
+					c, snapshot_node_missing,
+					"snapshot node %u from tree %s missing", *id, buf.buf)) {
+				if (t->nr > 1) {
+					bch_err(c, "cannot reconstruct snapshot trees with multiple nodes");
+					ret = -BCH_ERR_fsck_repair_unimplemented;
+					goto err;
+				}
+
+				ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+						check_snapshot_exists(trans, *id));
+				if (ret)
+					goto err;
+			}
+		}
+	}
+fsck_err:
+err:
+	bch2_trans_put(trans);
+	snapshot_tree_reconstruct_exit(&r);
+	printbuf_exit(&buf);
+	bch_err_fn(c, ret);
+	return ret;
+}
+
 /*
  * Mark a snapshot as deleted, for future cleanup:
  */
@@ -1689,6 +1844,20 @@ int bch2_snapshots_read(struct bch_fs *c)
 				   POS_MIN, 0, k,
 			   (set_is_ancestor_bitmap(c, k.k->p.offset), 0)));
 	bch_err_fn(c, ret);
+
+	/*
+	 * It's important that we check if we need to reconstruct snapshots
+	 * before going RW, so we mark that pass as required in the superblock -
+	 * otherwise, we could end up deleting keys with missing snapshot nodes
+	 * instead
+	 */
+	BUG_ON(!test_bit(BCH_FS_new_fs, &c->flags) &&
+	       test_bit(BCH_FS_may_go_rw, &c->flags));
+
+	if (bch2_err_matches(ret, EIO) ||
+	    (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_snapshots)))
+		ret = bch2_run_explicit_recovery_pass_persistent(c, BCH_RECOVERY_PASS_reconstruct_snapshots);
+
 	return ret;
 }
 

libbcachefs/snapshot.h

@@ -209,15 +209,34 @@ static inline bool snapshot_list_has_ancestor(struct bch_fs *c, snapshot_id_list
 
 static inline int snapshot_list_add(struct bch_fs *c, snapshot_id_list *s, u32 id)
 {
-	int ret;
-
 	BUG_ON(snapshot_list_has_id(s, id));
-	ret = darray_push(s, id);
+	int ret = darray_push(s, id);
 	if (ret)
 		bch_err(c, "error reallocating snapshot_id_list (size %zu)", s->size);
 	return ret;
 }
 
+static inline int snapshot_list_add_nodup(struct bch_fs *c, snapshot_id_list *s, u32 id)
+{
+	int ret = snapshot_list_has_id(s, id)
+		? 0
+		: darray_push(s, id);
+	if (ret)
+		bch_err(c, "error reallocating snapshot_id_list (size %zu)", s->size);
+	return ret;
+}
+
+static inline int snapshot_list_merge(struct bch_fs *c, snapshot_id_list *dst, snapshot_id_list *src)
+{
+	darray_for_each(*src, i) {
+		int ret = snapshot_list_add_nodup(c, dst, *i);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
 int bch2_snapshot_lookup(struct btree_trans *trans, u32 id,
 			 struct bch_snapshot *s);
 int bch2_snapshot_get_subvol(struct btree_trans *, u32,
@@ -229,6 +248,7 @@ int bch2_snapshot_node_create(struct btree_trans *, u32,
 
 int bch2_check_snapshot_trees(struct bch_fs *);
 int bch2_check_snapshots(struct bch_fs *);
+int bch2_reconstruct_snapshots(struct bch_fs *);
 
 int bch2_snapshot_node_set_deleted(struct btree_trans *, u32);
 void bch2_delete_dead_snapshots_work(struct work_struct *);

libbcachefs/super-io.c

@@ -142,8 +142,8 @@ void bch2_sb_field_delete(struct bch_sb_handle *sb,
 void bch2_free_super(struct bch_sb_handle *sb)
 {
 	kfree(sb->bio);
-	if (!IS_ERR_OR_NULL(sb->bdev_handle))
+	if (!IS_ERR_OR_NULL(sb->s_bdev_file))
-		bdev_release(sb->bdev_handle);
+		fput(sb->s_bdev_file);
 	kfree(sb->holder);
 	kfree(sb->sb_name);
 
@@ -527,9 +527,11 @@ static void bch2_sb_update(struct bch_fs *c)
 	memset(c->sb.errors_silent, 0, sizeof(c->sb.errors_silent));
 
 	struct bch_sb_field_ext *ext = bch2_sb_field_get(src, ext);
-	if (ext)
+	if (ext) {
 		le_bitvector_to_cpu(c->sb.errors_silent, (void *) ext->errors_silent,
 				    sizeof(c->sb.errors_silent) * 8);
+		c->sb.btrees_lost_data = le64_to_cpu(ext->btrees_lost_data);
+	}
 
 	for_each_member_device(c, ca) {
 		struct bch_member m = bch2_sb_member_get(src, ca->dev_idx);
@@ -712,23 +714,23 @@ retry:
 	if (!opt_get(*opts, nochanges))
 		sb->mode |= BLK_OPEN_WRITE;
 
-	sb->bdev_handle = bdev_open_by_path(path, sb->mode, sb->holder, &bch2_sb_handle_bdev_ops);
+	sb->s_bdev_file = bdev_file_open_by_path(path, sb->mode, sb->holder, &bch2_sb_handle_bdev_ops);
-	if (IS_ERR(sb->bdev_handle) &&
+	if (IS_ERR(sb->s_bdev_file) &&
-	    PTR_ERR(sb->bdev_handle) == -EACCES &&
+	    PTR_ERR(sb->s_bdev_file) == -EACCES &&
 	    opt_get(*opts, read_only)) {
 		sb->mode &= ~BLK_OPEN_WRITE;
 
-		sb->bdev_handle = bdev_open_by_path(path, sb->mode, sb->holder, &bch2_sb_handle_bdev_ops);
+		sb->s_bdev_file = bdev_file_open_by_path(path, sb->mode, sb->holder, &bch2_sb_handle_bdev_ops);
-		if (!IS_ERR(sb->bdev_handle))
+		if (!IS_ERR(sb->s_bdev_file))
 			opt_set(*opts, nochanges, true);
 	}
 
-	if (IS_ERR(sb->bdev_handle)) {
+	if (IS_ERR(sb->s_bdev_file)) {
-		ret = PTR_ERR(sb->bdev_handle);
+		ret = PTR_ERR(sb->s_bdev_file);
 		prt_printf(&err, "error opening %s: %s", path, bch2_err_str(ret));
 		goto err;
 	}
-	sb->bdev = sb->bdev_handle->bdev;
+	sb->bdev = file_bdev(sb->s_bdev_file);
 
 	ret = bch2_sb_realloc(sb, 0);
 	if (ret) {
@@ -1162,6 +1164,11 @@ static void bch2_sb_ext_to_text(struct printbuf *out, struct bch_sb *sb,
 
 		kfree(errors_silent);
 	}
+
+	prt_printf(out, "Btrees with missing data:");
+	prt_tab(out);
+	prt_bitflags(out, __bch2_btree_ids, le64_to_cpu(e->btrees_lost_data));
+	prt_newline(out);
 }
 
 static const struct bch_sb_field_ops bch_sb_field_ops_ext = {

libbcachefs/super.c

@@ -366,7 +366,7 @@ void bch2_fs_read_only(struct bch_fs *c)
 	    !test_bit(BCH_FS_emergency_ro, &c->flags) &&
 	    test_bit(BCH_FS_started, &c->flags) &&
 	    test_bit(BCH_FS_clean_shutdown, &c->flags) &&
-	    !c->opts.norecovery) {
+	    c->recovery_pass_done >= BCH_RECOVERY_PASS_journal_replay) {
 		BUG_ON(c->journal.last_empty_seq != journal_cur_seq(&c->journal));
 		BUG_ON(atomic_read(&c->btree_cache.dirty));
 		BUG_ON(atomic_long_read(&c->btree_key_cache.nr_dirty));
@@ -511,7 +511,8 @@ err:
 
 int bch2_fs_read_write(struct bch_fs *c)
 {
-	if (c->opts.norecovery)
+	if (c->opts.recovery_pass_last &&
+	    c->opts.recovery_pass_last < BCH_RECOVERY_PASS_journal_replay)
 		return -BCH_ERR_erofs_norecovery;
 
 	if (c->opts.nochanges)
@@ -1018,8 +1019,16 @@ int bch2_fs_start(struct bch_fs *c)
 	for_each_online_member(c, ca)
 		bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount = cpu_to_le64(now);
 
+	struct bch_sb_field_ext *ext =
+		bch2_sb_field_get_minsize(&c->disk_sb, ext, sizeof(*ext) / sizeof(u64));
 	mutex_unlock(&c->sb_lock);
 
+	if (!ext) {
+		bch_err(c, "insufficient space in superblock for sb_field_ext");
+		ret = -BCH_ERR_ENOSPC_sb;
+		goto err;
+	}
+
 	for_each_rw_member(c, ca)
 		bch2_dev_allocator_add(c, ca);
 	bch2_recalc_capacity(c);

libbcachefs/super_types.h

@@ -4,7 +4,7 @@
 
 struct bch_sb_handle {
 	struct bch_sb		*sb;
-	struct bdev_handle	*bdev_handle;
+	struct file		*s_bdev_file;
 	struct block_device	*bdev;
 	char			*sb_name;
 	struct bio		*bio;

libbcachefs/util.h

@@ -793,4 +793,14 @@ static inline void __set_bit_le64(size_t bit, __le64 *addr)
 	addr[bit / 64] |= cpu_to_le64(BIT_ULL(bit % 64));
 }
 
+static inline void __clear_bit_le64(size_t bit, __le64 *addr)
+{
+	addr[bit / 64] &= !cpu_to_le64(BIT_ULL(bit % 64));
+}
+
+static inline bool test_bit_le64(size_t bit, __le64 *addr)
+{
+	return (addr[bit / 64] & cpu_to_le64(BIT_ULL(bit % 64))) != 0;
+}
+
 #endif /* _BCACHEFS_UTIL_H */

linux/blkdev.c

@@ -162,16 +162,18 @@ sector_t get_capacity(struct gendisk *disk)
 	return bytes >> 9;
 }
 
-void bdev_release(struct bdev_handle *handle)
+void fput(struct file *file)
 {
-	fdatasync(handle->bdev->bd_fd);
+	struct block_device *bdev = file_bdev(file);
-	close(handle->bdev->bd_fd);
+
-	free(handle->bdev);
+	fdatasync(bdev->bd_fd);
-	free(handle);
+	close(bdev->bd_fd);
+	free(bdev);
+	free(file);
 }
 
-struct bdev_handle *bdev_open_by_path(const char *path, blk_mode_t mode,
+struct file *bdev_file_open_by_path(const char *path, blk_mode_t mode,
-				      void *holder, const struct blk_holder_ops *hop)
+				    void *holder, const struct blk_holder_ops *hop)
 {
 	int fd, flags = 0;
 
@@ -204,13 +206,12 @@ struct bdev_handle *bdev_open_by_path(const char *path, blk_mode_t mode,
 	bdev->bd_disk		= &bdev->__bd_disk;
 	bdev->bd_disk->bdi	= &bdev->bd_disk->__bdi;
 	bdev->queue.backing_dev_info = bdev->bd_disk->bdi;
+	bdev->bd_inode		= &bdev->__bd_inode;
 
-	struct bdev_handle *handle = calloc(sizeof(*handle), 1);
+	struct file *file = calloc(sizeof(*file), 1);
-	handle->bdev	= bdev;
+	file->f_inode = bdev->bd_inode;
-	handle->holder	= holder;
-	handle->mode	= mode;
 
-	return handle;
+	return file;
 }
 
 int lookup_bdev(const char *path, dev_t *dev)