Merge tag 'bcachefs-2024-10-14' of git://evilpiepirate.org/bcachefs

Pull bcachefs fixes from Kent Overstreet:

- New metadata version inode_has_child_snapshots

This fixes bugs with handling of unlinked inodes + snapshots, in
particular when an inode is reattached after taking a snapshot;
deleted inodes now get correctly cleaned up across snapshots.

- Disk accounting rewrite fixes
- validation fixes for when a device has been removed
- fix journal replay failing with "journal_reclaim_would_deadlock"

- Some more small fixes for erasure coding + device removal

- Assorted small syzbot fixes

* tag 'bcachefs-2024-10-14' of git://evilpiepirate.org/bcachefs: (27 commits)
bcachefs: Fix sysfs warning in fstests generic/730,731
bcachefs: Handle race between stripe reuse, invalidate_stripe_to_dev
bcachefs: Fix kasan splat in new_stripe_alloc_buckets()
bcachefs: Add missing validation for bch_stripe.csum_granularity_bits
bcachefs: Fix missing bounds checks in bch2_alloc_read()
bcachefs: fix uaf in bch2_dio_write_done()
bcachefs: Improve check_snapshot_exists()
bcachefs: Fix bkey_nocow_lock()
bcachefs: Fix accounting replay flags
bcachefs: Fix invalid shift in member_to_text()
bcachefs: Fix bch2_have_enough_devs() for BCH_SB_MEMBER_INVALID
bcachefs: __wait_for_freeing_inode: Switch to wait_bit_queue_entry
bcachefs: Check if stuck in journal_res_get()
closures: Add closure_wait_event_timeout()
bcachefs: Fix state lock involved deadlock
bcachefs: Fix NULL pointer dereference in bch2_opt_to_text
bcachefs: Release transaction before wake up
bcachefs: add check for btree id against max in try read node
bcachefs: Disk accounting device validation fixes
bcachefs: bch2_inode_or_descendents_is_open()
...

fs/bcachefs/alloc_background.c

···
 				continue;
 			}
 
+			if (k.k->p.offset < ca->mi.first_bucket) {
+				bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode, ca->mi.first_bucket));
+				continue;
+			}
+
+			if (k.k->p.offset >= ca->mi.nbuckets) {
+				bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
+				continue;
+			}
+
 			struct bch_alloc_v4 a;
 			*bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 			0;

fs/bcachefs/bcachefs_format.h

···
 	x(disk_accounting_v2,		BCH_VERSION(1,  9))		\
 	x(disk_accounting_v3,		BCH_VERSION(1, 10))		\
 	x(disk_accounting_inum,		BCH_VERSION(1, 11))		\
-	x(rebalance_work_acct_fix,	BCH_VERSION(1, 12))
+	x(rebalance_work_acct_fix,	BCH_VERSION(1, 12))		\
+	x(inode_has_child_snapshots,	BCH_VERSION(1, 13))
 
 enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min = 9,

fs/bcachefs/btree_gc.c

···
 	u64 b, start_time = local_clock();
 	int ret;
 
-	/*
-	 * Ideally we would be using state_lock and not gc_gens_lock here, but that
-	 * introduces a deadlock in the RO path - we currently take the state
-	 * lock at the start of going RO, thus the gc thread may get stuck:
-	 */
 	if (!mutex_trylock(&c->gc_gens_lock))
 		return 0;
 
 	trace_and_count(c, gc_gens_start, c);
 
-	down_read(&c->state_lock);
+	/*
+	 * We have to use trylock here. Otherwise, we would
+	 * introduce a deadlock in the RO path - we take the
+	 * state lock at the start of going RO.
+	 */
+	if (!down_read_trylock(&c->state_lock)) {
+		mutex_unlock(&c->gc_gens_lock);
+		return 0;
+	}
 
 	for_each_member_device(c, ca) {
 		struct bucket_gens *gens = bucket_gens(ca);

fs/bcachefs/btree_io.c

···
 	struct btree_trans *trans = bch2_trans_get(c);
 
 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
+
+	/* we don't need transaction context anymore after we got the lock. */
+	bch2_trans_put(trans);
 	__btree_node_write_done(c, b);
 	six_unlock_read(&b->c.lock);
-
-	bch2_trans_put(trans);
 }
 
 static void btree_node_write_work(struct work_struct *work)

fs/bcachefs/btree_iter.c

···
 		else
 			iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
 
-		if (unlikely(!(iter->flags & BTREE_ITER_is_extents)
-			     ? bkey_gt(iter_pos, end)
-			     : bkey_ge(iter_pos, end)))
+		if (unlikely(iter->flags & BTREE_ITER_all_snapshots	? bpos_gt(iter_pos, end) :
+			     iter->flags & BTREE_ITER_is_extents	? bkey_ge(iter_pos, end) :
+									  bkey_gt(iter_pos, end)))
 			goto end;
 
 		break;

fs/bcachefs/btree_iter.h

···
 	for_each_btree_key_upto_norestart(_trans, _iter, _btree_id, _start,\
 					  SPOS_MAX, _flags, _k, _ret)
 
+#define for_each_btree_key_reverse_norestart(_trans, _iter, _btree_id,	\
+					     _start, _flags, _k, _ret)	\
+	for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),	\
+				  (_start), (_flags));			\
+	     (_k) = bch2_btree_iter_peek_prev_type(&(_iter), _flags),	\
+	     !((_ret) = bkey_err(_k)) && (_k).k;			\
+	     bch2_btree_iter_rewind(&(_iter)))
+
 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)	\
 	for_each_btree_key_upto_continue_norestart(_iter, SPOS_MAX, _flags, _k, _ret)
 

fs/bcachefs/btree_node_scan.c

···
 	if (BTREE_NODE_LEVEL(bn) >= BTREE_MAX_DEPTH)
 		return;
 
+	if (BTREE_NODE_ID(bn) >= BTREE_ID_NR_MAX)
+		return;
+
 	rcu_read_lock();
 	struct found_btree_node n = {
 		.btree_id	= BTREE_NODE_ID(bn),

fs/bcachefs/data_update.c

···
 					if (ptr2 == ptr)
 						break;
 
+					ca = bch2_dev_have_ref(c, ptr2->dev);
 					bucket = PTR_BUCKET_POS(ca, ptr2);
 					bch2_bucket_nocow_unlock(&c->nocow_locks, bucket, 0);
 				}

fs/bcachefs/disk_accounting.c

···
 		*p = swab64(*p);
 }
 
+static inline void __accounting_to_replicas(struct bch_replicas_entry_v1 *r,
+					    struct disk_accounting_pos acc)
+{
+	unsafe_memcpy(r, &acc.replicas,
+		      replicas_entry_bytes(&acc.replicas),
+		      "variable length struct");
+}
+
 static inline bool accounting_to_replicas(struct bch_replicas_entry_v1 *r, struct bpos p)
 {
 	struct disk_accounting_pos acc_k;
···
 
 	switch (acc_k.type) {
 	case BCH_DISK_ACCOUNTING_replicas:
-		unsafe_memcpy(r, &acc_k.replicas,
-			      replicas_entry_bytes(&acc_k.replicas),
-			      "variable length struct");
+		__accounting_to_replicas(r, acc_k);
 		return true;
 	default:
 		return false;
···
 	return ret;
 }
 
+static int bch2_disk_accounting_validate_late(struct btree_trans *trans,
+					      struct disk_accounting_pos acc,
+					      u64 *v, unsigned nr)
+{
+	struct bch_fs *c = trans->c;
+	struct printbuf buf = PRINTBUF;
+	int ret = 0, invalid_dev = -1;
+
+	switch (acc.type) {
+	case BCH_DISK_ACCOUNTING_replicas: {
+		struct bch_replicas_padded r;
+		__accounting_to_replicas(&r.e, acc);
+
+		for (unsigned i = 0; i < r.e.nr_devs; i++)
+			if (r.e.devs[i] != BCH_SB_MEMBER_INVALID &&
+			    !bch2_dev_exists(c, r.e.devs[i])) {
+				invalid_dev = r.e.devs[i];
+				goto invalid_device;
+			}
+
+		/*
+		 * All replicas entry checks except for invalid device are done
+		 * in bch2_accounting_validate
+		 */
+		BUG_ON(bch2_replicas_entry_validate(&r.e, c, &buf));
+
+		if (fsck_err_on(!bch2_replicas_marked_locked(c, &r.e),
+				trans, accounting_replicas_not_marked,
+				"accounting not marked in superblock replicas\n  %s",
+				(printbuf_reset(&buf),
+				 bch2_accounting_key_to_text(&buf, &acc),
+				 buf.buf))) {
+			/*
+			 * We're not RW yet and still single threaded, dropping
+			 * and retaking lock is ok:
+			 */
+			percpu_up_write(&c->mark_lock);
+			ret = bch2_mark_replicas(c, &r.e);
+			if (ret)
+				goto fsck_err;
+			percpu_down_write(&c->mark_lock);
+		}
+		break;
+	}
+
+	case BCH_DISK_ACCOUNTING_dev_data_type:
+		if (!bch2_dev_exists(c, acc.dev_data_type.dev)) {
+			invalid_dev = acc.dev_data_type.dev;
+			goto invalid_device;
+		}
+		break;
+	}
+
+fsck_err:
+	printbuf_exit(&buf);
+	return ret;
+invalid_device:
+	if (fsck_err(trans, accounting_to_invalid_device,
+		     "accounting entry points to invalid device %i\n  %s",
+		     invalid_dev,
+		     (printbuf_reset(&buf),
+		      bch2_accounting_key_to_text(&buf, &acc),
+		      buf.buf))) {
+		for (unsigned i = 0; i < nr; i++)
+			v[i] = -v[i];
+
+		ret = commit_do(trans, NULL, NULL, 0,
+				bch2_disk_accounting_mod(trans, &acc, v, nr, false)) ?:
+			-BCH_ERR_remove_disk_accounting_entry;
+	} else {
+		ret = -BCH_ERR_remove_disk_accounting_entry;
+	}
+	goto fsck_err;
+}
+
 /*
  * At startup time, initialize the in memory accounting from the btree (and
  * journal)
···
 	}
 	keys->gap = keys->nr = dst - keys->data;
 
-	percpu_down_read(&c->mark_lock);
-	for (unsigned i = 0; i < acc->k.nr; i++) {
+	percpu_down_write(&c->mark_lock);
+	unsigned i = 0;
+	while (i < acc->k.nr) {
+		unsigned idx = inorder_to_eytzinger0(i, acc->k.nr);
+
+		struct disk_accounting_pos acc_k;
+		bpos_to_disk_accounting_pos(&acc_k, acc->k.data[idx].pos);
+
 		u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
-		bch2_accounting_mem_read_counters(acc, i, v, ARRAY_SIZE(v), false);
-
-		if (bch2_is_zero(v, sizeof(v[0]) * acc->k.data[i].nr_counters))
-			continue;
-
-		struct bch_replicas_padded r;
-		if (!accounting_to_replicas(&r.e, acc->k.data[i].pos))
-			continue;
+		bch2_accounting_mem_read_counters(acc, idx, v, ARRAY_SIZE(v), false);
 
 		/*
-		 * If the replicas entry is invalid it'll get cleaned up by
-		 * check_allocations:
+		 * If the entry counters are zeroed, it should be treated as
+		 * nonexistent - it might point to an invalid device.
+		 *
+		 * Remove it, so that if it's re-added it gets re-marked in the
+		 * superblock:
 		 */
-		if (bch2_replicas_entry_validate(&r.e, c, &buf))
+		ret = bch2_is_zero(v, sizeof(v[0]) * acc->k.data[idx].nr_counters)
+			? -BCH_ERR_remove_disk_accounting_entry
+			: bch2_disk_accounting_validate_late(trans, acc_k,
+							v, acc->k.data[idx].nr_counters);
+
+		if (ret == -BCH_ERR_remove_disk_accounting_entry) {
+			free_percpu(acc->k.data[idx].v[0]);
+			free_percpu(acc->k.data[idx].v[1]);
+			darray_remove_item(&acc->k, &acc->k.data[idx]);
+			eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
+					accounting_pos_cmp, NULL);
+			ret = 0;
 			continue;
-
-		struct disk_accounting_pos k;
-		bpos_to_disk_accounting_pos(&k, acc->k.data[i].pos);
-
-		if (fsck_err_on(!bch2_replicas_marked_locked(c, &r.e),
-				trans, accounting_replicas_not_marked,
-				"accounting not marked in superblock replicas\n  %s",
-				(printbuf_reset(&buf),
-				 bch2_accounting_key_to_text(&buf, &k),
-				 buf.buf))) {
-			/*
-			 * We're not RW yet and still single threaded, dropping
-			 * and retaking lock is ok:
-			 */
-			percpu_up_read(&c->mark_lock);
-			ret = bch2_mark_replicas(c, &r.e);
-			if (ret)
-				goto fsck_err;
-			percpu_down_read(&c->mark_lock);
 		}
+
+		if (ret)
+			goto fsck_err;
+		i++;
 	}
 
 	preempt_disable();
···
 	}
 	preempt_enable();
 fsck_err:
-	percpu_up_read(&c->mark_lock);
+	percpu_up_write(&c->mark_lock);
 err:
 	printbuf_exit(&buf);
 	bch2_trans_put(trans);

fs/bcachefs/ec.c

···
 			 "incorrect value size (%zu < %u)",
 			 bkey_val_u64s(k.k), stripe_val_u64s(s));
 
+	bkey_fsck_err_on(s->csum_granularity_bits >= 64,
+			 c, stripe_csum_granularity_bad,
+			 "invalid csum granularity (%u >= 64)",
+			 s->csum_granularity_bits);
+
 	ret = bch2_bkey_ptrs_validate(c, k, flags);
 fsck_err:
 	return ret;
···
 		   nr_data,
 		   s.nr_redundant);
 	bch2_prt_csum_type(out, s.csum_type);
-	prt_printf(out, " gran %u", 1U << s.csum_granularity_bits);
+	prt_str(out, " gran ");
+	if (s.csum_granularity_bits < 64)
+		prt_printf(out, "%llu", 1ULL << s.csum_granularity_bits);
+	else
+		prt_printf(out, "(invalid shift %u)", s.csum_granularity_bits);
 
 	if (s.disk_label) {
 		prt_str(out, " label");
···
 /* stripe creation: */
 
 static int ec_stripe_key_update(struct btree_trans *trans,
-				struct bkey_i_stripe *new,
-				bool create)
+				struct bkey_i_stripe *old,
+				struct bkey_i_stripe *new)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	int ret;
+	bool create = !old;
 
-	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
-			       new->k.p, BTREE_ITER_intent);
-	ret = bkey_err(k);
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
+					       new->k.p, BTREE_ITER_intent);
+	int ret = bkey_err(k);
 	if (ret)
 		goto err;
 
-	if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
-		bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
-				     create ? "creating" : "updating",
-				     bch2_bkey_types[k.k->type]);
+	if (bch2_fs_inconsistent_on(k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe),
+				    c, "error %s stripe: got existing key type %s",
+				    create ? "creating" : "updating",
+				    bch2_bkey_types[k.k->type])) {
 		ret = -EINVAL;
 		goto err;
 	}
 
 	if (k.k->type == KEY_TYPE_stripe) {
-		const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
-		unsigned i;
+		const struct bch_stripe *v = bkey_s_c_to_stripe(k).v;
 
-		if (old->nr_blocks != new->v.nr_blocks) {
-			bch_err(c, "error updating stripe: nr_blocks does not match");
-			ret = -EINVAL;
-			goto err;
-		}
+		BUG_ON(old->v.nr_blocks != new->v.nr_blocks);
+		BUG_ON(old->v.nr_blocks != v->nr_blocks);
 
-		for (i = 0; i < new->v.nr_blocks; i++) {
-			unsigned v = stripe_blockcount_get(old, i);
+		for (unsigned i = 0; i < new->v.nr_blocks; i++) {
+			unsigned sectors = stripe_blockcount_get(v, i);
 
-			BUG_ON(v &&
-			       (old->ptrs[i].dev != new->v.ptrs[i].dev ||
-				old->ptrs[i].gen != new->v.ptrs[i].gen ||
-				old->ptrs[i].offset != new->v.ptrs[i].offset));
+			if (!bch2_extent_ptr_eq(old->v.ptrs[i], new->v.ptrs[i]) && sectors) {
+				struct printbuf buf = PRINTBUF;
 
-			stripe_blockcount_set(&new->v, i, v);
+				prt_printf(&buf, "stripe changed nonempty block %u", i);
+				prt_str(&buf, "\nold: ");
+				bch2_bkey_val_to_text(&buf, c, k);
+				prt_str(&buf, "\nnew: ");
+				bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&new->k_i));
+				bch2_fs_inconsistent(c, "%s", buf.buf);
+				printbuf_exit(&buf);
+				ret = -EINVAL;
+				goto err;
+			}
+
+			/*
+			 * If the stripe ptr changed underneath us, it must have
+			 * been dev_remove_stripes() -> * invalidate_stripe_to_dev()
+			 */
+			if (!bch2_extent_ptr_eq(old->v.ptrs[i], v->ptrs[i])) {
+				BUG_ON(v->ptrs[i].dev != BCH_SB_MEMBER_INVALID);
+
+				if (bch2_extent_ptr_eq(old->v.ptrs[i], new->v.ptrs[i]))
+					new->v.ptrs[i].dev = BCH_SB_MEMBER_INVALID;
+			}
+
+			stripe_blockcount_set(&new->v, i, sectors);
 		}
 	}
 
···
 			    BCH_TRANS_COMMIT_no_check_rw|
 			    BCH_TRANS_COMMIT_no_enospc,
 			    ec_stripe_key_update(trans,
-					bkey_i_to_stripe(&s->new_stripe.key),
-					!s->have_existing_stripe));
+					s->have_existing_stripe
+					? bkey_i_to_stripe(&s->existing_stripe.key)
+					: NULL,
+					bkey_i_to_stripe(&s->new_stripe.key)));
 	bch_err_msg(c, ret, "creating stripe key");
 	if (ret) {
 		goto err;
···
 	bitmap_and(devs.d, devs.d, c->rw_devs[BCH_DATA_user].d, BCH_SB_MEMBERS_MAX);
 
 	for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) {
-		__clear_bit(v->ptrs[i].dev, devs.d);
+		/*
+		 * Note: we don't yet repair invalid blocks (failed/removed
+		 * devices) when reusing stripes - we still need a codepath to
+		 * walk backpointers and update all extents that point to that
+		 * block when updating the stripe
+		 */
+		if (v->ptrs[i].dev != BCH_SB_MEMBER_INVALID)
+			__clear_bit(v->ptrs[i].dev, devs.d);
+
 		if (i < h->s->nr_data)
 			nr_have_data++;
 		else

fs/bcachefs/errcode.h

···
 	x(BCH_ERR_nopromote,		nopromote_no_writes)			\
 	x(BCH_ERR_nopromote,		nopromote_enomem)			\
 	x(0,				invalid_snapshot_node)			\
-	x(0,				option_needs_open_fs)
+	x(0,				option_needs_open_fs)			\
+	x(0,				remove_disk_accounting_entry)
 
 enum bch_errcode {
 	BCH_ERR_START		= 2048,

fs/bcachefs/extents.h

···
 int bch2_bkey_ptrs_validate(struct bch_fs *, struct bkey_s_c,
 			    enum bch_validate_flags);
 
+static inline bool bch2_extent_ptr_eq(struct bch_extent_ptr ptr1,
+				      struct bch_extent_ptr ptr2)
+{
+	return (ptr1.cached	== ptr2.cached &&
+		ptr1.unwritten	== ptr2.unwritten &&
+		ptr1.offset	== ptr2.offset &&
+		ptr1.dev	== ptr2.dev &&
+		ptr1.dev	== ptr2.dev);
+}
+
 void bch2_ptr_swab(struct bkey_s);
 
 const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c);

fs/bcachefs/fs-io-direct.c

···
 
 static __always_inline long bch2_dio_write_done(struct dio_write *dio)
 {
+	struct bch_fs *c = dio->op.c;
 	struct kiocb *req = dio->req;
 	struct bch_inode_info *inode = dio->inode;
 	bool sync = dio->sync;
···
 	ret = dio->op.error ?: ((long) dio->written << 9);
 	bio_put(&dio->op.wbio.bio);
 
-	bch2_write_ref_put(dio->op.c, BCH_WRITE_REF_dio_write);
+	bch2_write_ref_put(c, BCH_WRITE_REF_dio_write);
 
 	/* inode->i_dio_count is our ref on inode and thus bch_fs */
 	inode_dio_end(&inode->v);

fs/bcachefs/fs.c

···
 	return a.subvol == b.subvol && a.inum == b.inum;
 }
 
+static u32 bch2_vfs_inode_hash_fn(const void *data, u32 len, u32 seed)
+{
+	const subvol_inum *inum = data;
+
+	return jhash(&inum->inum, sizeof(inum->inum), seed);
+}
+
+static u32 bch2_vfs_inode_obj_hash_fn(const void *data, u32 len, u32 seed)
+{
+	const struct bch_inode_info *inode = data;
+
+	return bch2_vfs_inode_hash_fn(&inode->ei_inum, sizeof(inode->ei_inum), seed);
+}
+
 static int bch2_vfs_inode_cmp_fn(struct rhashtable_compare_arg *arg,
 				 const void *obj)
 {
···
 	.head_offset		= offsetof(struct bch_inode_info, hash),
 	.key_offset		= offsetof(struct bch_inode_info, ei_inum),
 	.key_len		= sizeof(subvol_inum),
+	.hashfn			= bch2_vfs_inode_hash_fn,
+	.obj_hashfn		= bch2_vfs_inode_obj_hash_fn,
 	.obj_cmpfn		= bch2_vfs_inode_cmp_fn,
 	.automatic_shrinking	= true,
 };
 
-struct bch_inode_info *__bch2_inode_hash_find(struct bch_fs *c, subvol_inum inum)
+int bch2_inode_or_descendents_is_open(struct btree_trans *trans, struct bpos p)
+{
+	struct bch_fs *c = trans->c;
+	struct rhashtable *ht = &c->vfs_inodes_table;
+	subvol_inum inum = (subvol_inum) { .inum = p.offset };
+	DARRAY(u32) subvols;
+	int ret = 0;
+
+	if (!test_bit(BCH_FS_started, &c->flags))
+		return false;
+
+	darray_init(&subvols);
+restart_from_top:
+
+	/*
+	 * Tweaked version of __rhashtable_lookup(); we need to get a list of
+	 * subvolumes in which the given inode number is open.
+	 *
+	 * For this to work, we don't include the subvolume ID in the key that
+	 * we hash - all inodes with the same inode number regardless of
+	 * subvolume will hash to the same slot.
+	 *
+	 * This will be less than ideal if the same file is ever open
+	 * simultaneously in many different snapshots:
+	 */
+	rcu_read_lock();
+	struct rhash_lock_head __rcu *const *bkt;
+	struct rhash_head *he;
+	unsigned int hash;
+	struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
+restart:
+	hash = rht_key_hashfn(ht, tbl, &inum, bch2_vfs_inodes_params);
+	bkt = rht_bucket(tbl, hash);
+	do {
+		struct bch_inode_info *inode;
+
+		rht_for_each_entry_rcu_from(inode, he, rht_ptr_rcu(bkt), tbl, hash, hash) {
+			if (inode->ei_inum.inum == inum.inum) {
+				ret = darray_push_gfp(&subvols, inode->ei_inum.subvol,
+						      GFP_NOWAIT|__GFP_NOWARN);
+				if (ret) {
+					rcu_read_unlock();
+					ret = darray_make_room(&subvols, 1);
+					if (ret)
+						goto err;
+					subvols.nr = 0;
+					goto restart_from_top;
+				}
+			}
+		}
+		/* An object might have been moved to a different hash chain,
+		 * while we walk along it - better check and retry.
+		 */
+	} while (he != RHT_NULLS_MARKER(bkt));
+
+	/* Ensure we see any new tables. */
+	smp_rmb();
+
+	tbl = rht_dereference_rcu(tbl->future_tbl, ht);
+	if (unlikely(tbl))
+		goto restart;
+	rcu_read_unlock();
+
+	darray_for_each(subvols, i) {
+		u32 snap;
+		ret = bch2_subvolume_get_snapshot(trans, *i, &snap);
+		if (ret)
+			goto err;
+
+		ret = bch2_snapshot_is_ancestor(c, snap, p.snapshot);
+		if (ret)
+			break;
+	}
+err:
+	darray_exit(&subvols);
+	return ret;
+}
+
+static struct bch_inode_info *__bch2_inode_hash_find(struct bch_fs *c, subvol_inum inum)
 {
 	return rhashtable_lookup_fast(&c->vfs_inodes_table, &inum, bch2_vfs_inodes_params);
 }
···
 				    subvol_inum inum)
 {
 	wait_queue_head_t *wq;
-	DEFINE_WAIT_BIT(wait, &inode->v.i_state, __I_NEW);
+	struct wait_bit_queue_entry wait;
+
 	wq = inode_bit_waitqueue(&wait, &inode->v, __I_NEW);
 	prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
 	spin_unlock(&inode->v.i_lock);
···
 
 	set_bit(EI_INODE_HASHED, &inode->ei_flags);
 retry:
-	if (unlikely(rhashtable_lookup_insert_fast(&c->vfs_inodes_table,
+	if (unlikely(rhashtable_lookup_insert_key(&c->vfs_inodes_table,
+					&inode->ei_inum,
 					&inode->hash,
 					bch2_vfs_inodes_params))) {
 		old = bch2_inode_hash_find(c, trans, inode->ei_inum);

fs/bcachefs/fs.h

···
 	return inode->ei_inum;
 }
 
-struct bch_inode_info *__bch2_inode_hash_find(struct bch_fs *, subvol_inum);
-
 /*
  * Set if we've gotten a btree error for this inode, and thus the vfs inode and
  * btree inode may be inconsistent:
···
 __bch2_create(struct mnt_idmap *, struct bch_inode_info *,
 	      struct dentry *, umode_t, dev_t, subvol_inum, unsigned);
 
+int bch2_inode_or_descendents_is_open(struct btree_trans *trans, struct bpos p);
+
 int bch2_fs_quota_transfer(struct bch_fs *,
 			   struct bch_inode_info *,
 			   struct bch_qid,
···
 
 #define bch2_inode_update_after_write(_trans, _inode, _inode_u, _fields)	({ do {} while (0); })
 
-static inline struct bch_inode_info *__bch2_inode_hash_find(struct bch_fs *c, subvol_inum inum)
-{
-	return NULL;
-}
+static inline int bch2_inode_or_descendents_is_open(struct btree_trans *trans, struct bpos p) { return 0; }
 
 static inline void bch2_evict_subvolume_inodes(struct bch_fs *c,
 					       snapshot_id_list *s) {}

fs/bcachefs/fsck.c

···
 	return ret;
 }
 
+static inline bool inode_should_reattach(struct bch_inode_unpacked *inode)
+{
+	if (inode->bi_inum == BCACHEFS_ROOT_INO &&
+	    inode->bi_subvol == BCACHEFS_ROOT_SUBVOL)
+		return false;
+
+	return !inode->bi_dir && !(inode->bi_flags & BCH_INODE_unlinked);
+}
+
+static int maybe_delete_dirent(struct btree_trans *trans, struct bpos d_pos, u32 snapshot)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_dirents,
+					SPOS(d_pos.inode, d_pos.offset, snapshot),
+					BTREE_ITER_intent|
+					BTREE_ITER_with_updates);
+	int ret = bkey_err(k);
+	if (ret)
+		return ret;
+
+	if (bpos_eq(k.k->p, d_pos)) {
+		/*
+		 * delet_at() doesn't work because the update path doesn't
+		 * internally use BTREE_ITER_with_updates yet
+		 */
+		struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(*k));
+		ret = PTR_ERR_OR_ZERO(k);
+		if (ret)
+			goto err;
+
+		bkey_init(&k->k);
+		k->k.type = KEY_TYPE_whiteout;
+		k->k.p = iter.pos;
+		ret = bch2_trans_update(trans, &iter, k, BTREE_UPDATE_internal_snapshot_node);
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
 static int reattach_inode(struct btree_trans *trans, struct bch_inode_unpacked *inode)
 {
 	struct bch_fs *c = trans->c;
-	struct bch_hash_info dir_hash;
 	struct bch_inode_unpacked lostfound;
 	char name_buf[20];
-	struct qstr name;
-	u64 dir_offset = 0;
-	u32 dirent_snapshot = inode->bi_snapshot;
 	int ret;
 
+	u32 dirent_snapshot = inode->bi_snapshot;
 	if (inode->bi_subvol) {
 		inode->bi_parent_subvol = BCACHEFS_ROOT_SUBVOL;
 
···
 	if (ret)
 		return ret;
 
-	dir_hash = bch2_hash_info_init(c, &lostfound);
+	struct bch_hash_info dir_hash = bch2_hash_info_init(c, &lostfound);
+	struct qstr name = (struct qstr) QSTR(name_buf);
 
-	name = (struct qstr) QSTR(name_buf);
+	inode->bi_dir = lostfound.bi_inum;
 
 	ret = bch2_dirent_create_snapshot(trans,
 				inode->bi_parent_subvol, lostfound.bi_inum,
···
 				inode_d_type(inode),
 				&name,
 				inode->bi_subvol ?: inode->bi_inum,
-				&dir_offset,
+				&inode->bi_dir_offset,
 				STR_HASH_must_create);
 	if (ret) {
 		bch_err_msg(c, ret, "error creating dirent");
 		return ret;
 	}
 
-	inode->bi_dir		= lostfound.bi_inum;
-	inode->bi_dir_offset	= dir_offset;
+	ret = __bch2_fsck_write_inode(trans, inode);
+	if (ret)
+		return ret;
 
-	return __bch2_fsck_write_inode(trans, inode);
+	/*
+	 * Fix up inodes in child snapshots: if they should also be reattached
+	 * update the backpointer field, if they should not be we need to emit
+	 * whiteouts for the dirent we just created.
+	 */
+	if (!inode->bi_subvol && bch2_snapshot_is_leaf(c, inode->bi_snapshot) <= 0) {
+		snapshot_id_list whiteouts_done;
+		struct btree_iter iter;
+		struct bkey_s_c k;
+
+		darray_init(&whiteouts_done);
+
+		for_each_btree_key_reverse_norestart(trans, iter,
+				BTREE_ID_inodes, SPOS(0, inode->bi_inum, inode->bi_snapshot - 1),
+				BTREE_ITER_all_snapshots|BTREE_ITER_intent, k, ret) {
+			if (k.k->p.offset != inode->bi_inum)
+				break;
+
+			if (!bkey_is_inode(k.k) ||
+			    !bch2_snapshot_is_ancestor(c, k.k->p.snapshot, inode->bi_snapshot) ||
+			    snapshot_list_has_ancestor(c, &whiteouts_done, k.k->p.snapshot))
+				continue;
+
+			struct bch_inode_unpacked child_inode;
+			bch2_inode_unpack(k, &child_inode);
+
+			if (!inode_should_reattach(&child_inode)) {
+				ret = maybe_delete_dirent(trans,
+							  SPOS(lostfound.bi_inum, inode->bi_dir_offset,
+							       dirent_snapshot),
+							  k.k->p.snapshot);
+				if (ret)
+					break;
+
+				ret = snapshot_list_add(c, &whiteouts_done, k.k->p.snapshot);
+				if (ret)
+					break;
+			} else {
+				iter.snapshot = k.k->p.snapshot;
+				child_inode.bi_dir = inode->bi_dir;
+				child_inode.bi_dir_offset = inode->bi_dir_offset;
+
+				ret = bch2_inode_write_flags(trans, &iter, &child_inode,
+							     BTREE_UPDATE_internal_snapshot_node);
+				if (ret)
+					break;
+			}
+		}
+		darray_exit(&whiteouts_done);
+		bch2_trans_iter_exit(trans, &iter);
+	}
+
+	return ret;
 }
 
 static int remove_backpointer(struct btree_trans *trans,
···
 		 */
 		inode->bi_dir = 0;
 		inode->bi_dir_offset = 0;
-		inode->bi_flags &= ~BCH_INODE_backptr_untrusted;
 		*write_inode = true;
 	}
 
···
 	return ret;
 }
 
-static bool bch2_inode_is_open(struct bch_fs *c, struct bpos p)
-{
-	subvol_inum inum = {
-		.subvol = snapshot_t(c, p.snapshot)->subvol,
-		.inum	= p.offset,
-	};
-
-	/* snapshot tree corruption, can't safely delete */
-	if (!inum.subvol) {
-		bch_warn_ratelimited(c, "%s(): snapshot %u has no subvol, unlinked but can't safely delete", __func__, p.snapshot);
-		return true;
-	}
-
-	return __bch2_inode_hash_find(c, inum) != NULL;
-}
-
 static int check_inode(struct btree_trans *trans,
 		       struct btree_iter *iter,
 		       struct bkey_s_c k,
 		       struct bch_inode_unpacked *prev,
-		       struct snapshots_seen *s,
-		       bool full)
+		       struct snapshots_seen *s)
 {
 	struct bch_fs *c = trans->c;
 	struct printbuf buf = PRINTBUF;
···
 		return 0;
 
 	BUG_ON(bch2_inode_unpack(k, &u));
-
-	if (!full &&
-	    !(u.bi_flags & (BCH_INODE_i_size_dirty|
-			    BCH_INODE_i_sectors_dirty|
-			    BCH_INODE_unlinked)))
-		return 0;
 
 	if (prev->bi_inum != u.bi_inum)
 		*prev = u;
···
 		ret = 0;
 	}
 
-	if ((u.bi_flags & (BCH_INODE_i_size_dirty|BCH_INODE_unlinked)) &&
-	    bch2_key_has_snapshot_overwrites(trans, BTREE_ID_inodes, k.k->p)) {
-		struct bpos new_min_pos;
+	ret = bch2_inode_has_child_snapshots(trans, k.k->p);
+	if (ret < 0)
+		goto err;
 
-		ret = bch2_propagate_key_to_snapshot_leaves(trans, iter->btree_id, k, &new_min_pos);
+	if (fsck_err_on(ret != !!(u.bi_flags & BCH_INODE_has_child_snapshot),
+			trans, inode_has_child_snapshots_wrong,
+			"inode has_child_snapshots flag wrong (should be %u)\n%s",
+			ret,
+			(printbuf_reset(&buf),
+			 bch2_inode_unpacked_to_text(&buf, &u),
+			 buf.buf))) {
 		if (ret)
-			goto err;
-
-		u.bi_flags &= ~BCH_INODE_i_size_dirty|BCH_INODE_unlinked;
-
-		ret = __bch2_fsck_write_inode(trans, &u);
-
-		bch_err_msg(c, ret, "in fsck updating inode");
-		if (ret)
-			goto err_noprint;
-
-		if (!bpos_eq(new_min_pos, POS_MIN))
-			bch2_btree_iter_set_pos(iter, bpos_predecessor(new_min_pos));
-		goto err_noprint;
+			u.bi_flags |= BCH_INODE_has_child_snapshot;
+		else
+			u.bi_flags &= ~BCH_INODE_has_child_snapshot;
+		do_update = true;
 	}
+	ret = 0;
 
-	if (u.bi_flags & BCH_INODE_unlinked) {
+	if ((u.bi_flags & BCH_INODE_unlinked) &&
+	    !(u.bi_flags & BCH_INODE_has_child_snapshot)) {
 		if (!test_bit(BCH_FS_started, &c->flags)) {
 			/*
 			 * If we're not in online fsck, don't delete unlinked
···
 			if (ret)
 				goto err;
 		} else {
-			if (fsck_err_on(!bch2_inode_is_open(c, k.k->p),
+			ret = bch2_inode_or_descendents_is_open(trans, k.k->p);
+			if (ret < 0)
+				goto err;
+
+			if (fsck_err_on(!ret,
 					trans, inode_unlinked_and_not_open,
 				      "inode %llu%u unlinked and not open",
 				      u.bi_inum, u.bi_snapshot)) {
···
 				bch_err_msg(c, ret, "in fsck deleting inode");
 				goto err_noprint;
 			}
+			ret = 0;
 		}
-	}
-
-	/* i_size_dirty is vestigal, since we now have logged ops for truncate * */
-	if (u.bi_flags & BCH_INODE_i_size_dirty &&
-	    (!test_bit(BCH_FS_clean_recovery, &c->flags) ||
-	     fsck_err(trans, inode_i_size_dirty_but_clean,
-		      "filesystem marked clean, but inode %llu has i_size dirty",
-		      u.bi_inum))) {
-		bch_verbose(c, "truncating inode %llu", u.bi_inum);
-
-		/*
-		 * XXX: need to truncate partial blocks too here - or ideally
-		 * just switch units to bytes and that issue goes away
-		 */
-		ret = bch2_btree_delete_range_trans(trans, BTREE_ID_extents,
-				SPOS(u.bi_inum, round_up(u.bi_size, block_bytes(c)) >> 9,
-				     iter->pos.snapshot),
-				POS(u.bi_inum, U64_MAX),
-				0, NULL);
-		bch_err_msg(c, ret, "in fsck truncating inode");
-		if (ret)
-			return ret;
-
-		/*
-		 * We truncated without our normal sector accounting hook, just
-		 * make sure we recalculate it:
-		 */
-		u.bi_flags |= BCH_INODE_i_sectors_dirty;
-
-		u.bi_flags &= ~BCH_INODE_i_size_dirty;
-		do_update = true;
-	}
-
-	/* i_sectors_dirty is vestigal, i_sectors is always updated transactionally */
-	if (u.bi_flags & BCH_INODE_i_sectors_dirty &&
-	    (!test_bit(BCH_FS_clean_recovery, &c->flags) ||
-	     fsck_err(trans, inode_i_sectors_dirty_but_clean,
-		      "filesystem marked clean, but inode %llu has i_sectors dirty",
-		      u.bi_inum))) {
-		s64 sectors;
-
-		bch_verbose(c, "recounting sectors for inode %llu",
-			    u.bi_inum);
-
-		sectors = bch2_count_inode_sectors(trans, u.bi_inum, iter->pos.snapshot);
-		if (sectors < 0) {
-			bch_err_msg(c, sectors, "in fsck recounting inode sectors");
-			return sectors;
-		}
-
-		u.bi_sectors = sectors;
-		u.bi_flags &= ~BCH_INODE_i_sectors_dirty;
-		do_update = true;
-	}
-
-	if (u.bi_flags & BCH_INODE_backptr_untrusted) {
-		u.bi_dir = 0;
-		u.bi_dir_offset = 0;
-		u.bi_flags &= ~BCH_INODE_backptr_untrusted;
-		do_update = true;
 	}
 
 	if (fsck_err_on(u.bi_parent_subvol &&
···
 
 int bch2_check_inodes(struct bch_fs *c)
 {
-	bool full = c->opts.fsck;
 	struct bch_inode_unpacked prev = { 0 };
 	struct snapshots_seen s;
 
···
 				POS_MIN,
 				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k,
 				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
-			check_inode(trans, &iter, k, &prev, &s, full)));
+			check_inode(trans, &iter, k, &prev, &s)));
 
 	snapshots_seen_exit(&s);
+	bch_err_fn(c, ret);
+	return ret;
+}
+
+static int find_oldest_inode_needs_reattach(struct btree_trans *trans,
+					    struct bch_inode_unpacked *inode)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	/*
+	 * We look for inodes to reattach in natural key order, leaves first,
+	 * but we should do the reattach at the oldest version that needs to be
+	 * reattached:
+	 */
+	for_each_btree_key_norestart(trans, iter,
+				     BTREE_ID_inodes,
+				     SPOS(0, inode->bi_inum, inode->bi_snapshot + 1),
+				     BTREE_ITER_all_snapshots, k, ret) {
+		if (k.k->p.offset != inode->bi_inum)
+			break;
+
+		if (!bch2_snapshot_is_ancestor(c, inode->bi_snapshot, k.k->p.snapshot))
+			continue;
+
+		if (!bkey_is_inode(k.k))
+			break;
+
+		struct bch_inode_unpacked parent_inode;
+		bch2_inode_unpack(k, &parent_inode);
+
+		if (!inode_should_reattach(&parent_inode))
+			break;
+
+		*inode = parent_inode;
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
+
+static int check_unreachable_inode(struct btree_trans *trans,
+				   struct btree_iter *iter,
+				   struct bkey_s_c k)
+{
+	struct printbuf buf = PRINTBUF;
+	int ret = 0;
+
+	if (!bkey_is_inode(k.k))
+		return 0;
+
+	struct bch_inode_unpacked inode;
+	BUG_ON(bch2_inode_unpack(k, &inode));
+
+	if (!inode_should_reattach(&inode))
+		return 0;
+
+	ret = find_oldest_inode_needs_reattach(trans, &inode);
+	if (ret)
+		return ret;
+
+	if (fsck_err(trans, inode_unreachable,
+		     "unreachable inode:\n%s",
+		     (bch2_inode_unpacked_to_text(&buf, &inode),
+		      buf.buf)))
+		ret = reattach_inode(trans, &inode);
+fsck_err:
+	printbuf_exit(&buf);
+	return ret;
+}
+
+/*
+ * Reattach unreachable (but not unlinked) inodes
+ *
+ * Run after check_inodes() and check_dirents(), so we node that inode
+ * backpointer fields point to valid dirents, and every inode that has a dirent
+ * that points to it has its backpointer field set - so we're just looking for
+ * non-unlinked inodes without backpointers:
+ *
+ * XXX: this is racy w.r.t. hardlink removal in online fsck
+ */
+int bch2_check_unreachable_inodes(struct bch_fs *c)
+{
+	int ret = bch2_trans_run(c,
+		for_each_btree_key_commit(trans, iter, BTREE_ID_inodes,
+				POS_MIN,
+				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k,
+				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+			check_unreachable_inode(trans, &iter, k)));
 	bch_err_fn(c, ret);
 	return ret;
 }
···
 			    !key_visible_in_snapshot(c, s, i->snapshot, k.k->p.snapshot))
 				continue;
 
-			if (fsck_err_on(!(i->inode.bi_flags & BCH_INODE_i_size_dirty) &&
-					k.k->p.offset > round_up(i->inode.bi_size, block_bytes(c)) >> 9 &&
+			if (fsck_err_on(k.k->p.offset > round_up(i->inode.bi_size, block_bytes(c)) >> 9 &&
 					!bkey_extent_is_reservation(k),
 					trans, extent_past_end_of_inode,
 					"extent type past end of inode %llu:%u, i_size %llu\n  %s",
···
 		if (ret)
 			break;
 
-		/*
-		 * We've checked that inode backpointers point to valid dirents;
-		 * here, it's sufficient to check that the subvolume root has a
-		 * dirent:
-		 */
-		if (fsck_err_on(!subvol_root.bi_dir,
-				trans, subvol_unreachable,
-				"unreachable subvolume %s",
-				(bch2_bkey_val_to_text(&buf, c, s.s_c),
-				 prt_newline(&buf),
-				 bch2_inode_unpacked_to_text(&buf, &subvol_root),
-				 buf.buf))) {
-			ret = reattach_subvol(trans, s);
-			break;
-		}
-
 		u32 parent = le32_to_cpu(s.v->fs_path_parent);
 
 		if (darray_u32_has(&subvol_path, parent)) {
···
 	return false;
 }
 
-/*
- * Check that a given inode is reachable from its subvolume root - we already
- * verified subvolume connectivity:
- *
- * XXX: we should also be verifying that inodes are in the right subvolumes
- */
 static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c inode_k)
 {
 	struct bch_fs *c = trans->c;
···
 	p->nr = 0;
 
 	BUG_ON(bch2_inode_unpack(inode_k, &inode));
+
+	if (!S_ISDIR(inode.bi_mode))
+		return 0;
 
 	while (!inode.bi_subvol) {
 		struct btree_iter dirent_iter;
···
 			bch2_trans_iter_exit(trans, &dirent_iter);
 
 		if (bch2_err_matches(ret, ENOENT)) {
-			ret = 0;
-			if (fsck_err(trans, inode_unreachable,
-				     "unreachable inode\n%s",
-				     (printbuf_reset(&buf),
-				      bch2_bkey_val_to_text(&buf, c, inode_k),
-				      buf.buf)))
-				ret = reattach_inode(trans, &inode);
+			printbuf_reset(&buf);
+			bch2_bkey_val_to_text(&buf, c, inode_k);
+			bch_err(c, "unreachable inode in check_directory_structure: %s\n%s",
+				bch2_err_str(ret), buf.buf);
 			goto out;
 		}
 
 		bch2_trans_iter_exit(trans, &dirent_iter);
-
-		if (!S_ISDIR(inode.bi_mode))
-			break;
 
 		ret = darray_push(p, ((struct pathbuf_entry) {
 			.inum		= inode.bi_inum,
···
 }
 
 /*
- * Check for unreachable inodes, as well as loops in the directory structure:
- * After bch2_check_dirents(), if an inode backpointer doesn't exist that means it's
- * unreachable:
+ * Check for loops in the directory structure: all other connectivity issues
+ * have been fixed by prior passes
  */
 int bch2_check_directory_structure(struct bch_fs *c)
 {
···
 			if (S_ISDIR(u.bi_mode))
 				continue;
 
+			/*
+			 * Previous passes ensured that bi_nlink is nonzero if
+			 * it had multiple hardlinks:
+			 */
 			if (!u.bi_nlink)
 				continue;
 

fs/bcachefs/fsck.h

···
 int bch2_check_xattrs(struct bch_fs *);
 int bch2_check_root(struct bch_fs *);
 int bch2_check_subvolume_structure(struct bch_fs *);
+int bch2_check_unreachable_inodes(struct bch_fs *);
 int bch2_check_directory_structure(struct bch_fs *);
 int bch2_check_nlinks(struct bch_fs *);
 int bch2_fix_reflink_p(struct bch_fs *);

fs/bcachefs/inode.c

···
 #include "error.h"
 #include "extents.h"
 #include "extent_update.h"
+#include "fs.h"
 #include "inode.h"
 #include "str_hash.h"
 #include "snapshot.h"
···
 	NULL
 };
 #undef  x
+
+static int delete_ancestor_snapshot_inodes(struct btree_trans *, struct bpos);
 
 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
 
···
 	}
 }
 
-static inline bool bkey_is_deleted_inode(struct bkey_s_c k)
+static inline void bkey_inode_flags_set(struct bkey_s k, u64 f)
 {
-	return bkey_inode_flags(k) & BCH_INODE_unlinked;
+	switch (k.k->type) {
+	case KEY_TYPE_inode:
+		bkey_s_to_inode(k).v->bi_flags = cpu_to_le32(f);
+		return;
+	case KEY_TYPE_inode_v2:
+		bkey_s_to_inode_v2(k).v->bi_flags = cpu_to_le64(f);
+		return;
+	case KEY_TYPE_inode_v3:
+		bkey_s_to_inode_v3(k).v->bi_flags = cpu_to_le64(f);
+		return;
+	default:
+		BUG();
+	}
+}
+
+static inline bool bkey_is_unlinked_inode(struct bkey_s_c k)
+{
+	unsigned f = bkey_inode_flags(k) & BCH_INODE_unlinked;
+
+	return (f & BCH_INODE_unlinked) && !(f & BCH_INODE_has_child_snapshot);
+}
+
+static struct bkey_s_c
+bch2_bkey_get_iter_snapshot_parent(struct btree_trans *trans, struct btree_iter *iter,
+				   enum btree_id btree, struct bpos pos,
+				   unsigned flags)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	for_each_btree_key_upto_norestart(trans, *iter, btree,
+					  bpos_successor(pos),
+					  SPOS(pos.inode, pos.offset, U32_MAX),
+					  flags|BTREE_ITER_all_snapshots, k, ret)
+		if (bch2_snapshot_is_ancestor(c, pos.snapshot, k.k->p.snapshot))
+			return k;
+
+	bch2_trans_iter_exit(trans, iter);
+	return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
+}
+
+static struct bkey_s_c
+bch2_inode_get_iter_snapshot_parent(struct btree_trans *trans, struct btree_iter *iter,
+				    struct bpos pos, unsigned flags)
+{
+	struct bkey_s_c k;
+again:
+	k = bch2_bkey_get_iter_snapshot_parent(trans, iter, BTREE_ID_inodes, pos, flags);
+	if (!k.k ||
+	    bkey_err(k) ||
+	    bkey_is_inode(k.k))
+		return k;
+
+	bch2_trans_iter_exit(trans, iter);
+	pos = k.k->p;
+	goto again;
+}
+
+int __bch2_inode_has_child_snapshots(struct btree_trans *trans, struct bpos pos)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	for_each_btree_key_upto_norestart(trans, iter,
+			BTREE_ID_inodes, POS(0, pos.offset), bpos_predecessor(pos),
+			BTREE_ITER_all_snapshots|
+			BTREE_ITER_with_updates, k, ret)
+		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot) &&
+		    bkey_is_inode(k.k)) {
+			ret = 1;
+			break;
+		}
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static int update_inode_has_children(struct btree_trans *trans,
+				     struct bkey_s k,
+				     bool have_child)
+{
+	if (!have_child) {
+		int ret = bch2_inode_has_child_snapshots(trans, k.k->p);
+		if (ret)
+			return ret < 0 ? ret : 0;
+	}
+
+	u64 f = bkey_inode_flags(k.s_c);
+	if (have_child != !!(f & BCH_INODE_has_child_snapshot))
+		bkey_inode_flags_set(k, f ^ BCH_INODE_has_child_snapshot);
+
+	return 0;
+}
+
+static int update_parent_inode_has_children(struct btree_trans *trans, struct bpos pos,
+					    bool have_child)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_inode_get_iter_snapshot_parent(trans,
+						&iter, pos, BTREE_ITER_with_updates);
+	int ret = bkey_err(k);
+	if (ret)
+		return ret;
+	if (!k.k)
+		return 0;
+
+	if (!have_child) {
+		ret = bch2_inode_has_child_snapshots(trans, k.k->p);
+		if (ret) {
+			ret = ret < 0 ? ret : 0;
+			goto err;
+		}
+	}
+
+	u64 f = bkey_inode_flags(k);
+	if (have_child != !!(f & BCH_INODE_has_child_snapshot)) {
+		struct bkey_i *update = bch2_bkey_make_mut(trans, &iter, &k,
+					     BTREE_UPDATE_internal_snapshot_node);
+		ret = PTR_ERR_OR_ZERO(update);
+		if (ret)
+			goto err;
+
+		bkey_inode_flags_set(bkey_i_to_s(update), f ^ BCH_INODE_has_child_snapshot);
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
 }
 
 int bch2_trigger_inode(struct btree_trans *trans,
···
 		       struct bkey_s new,
 		       enum btree_iter_update_trigger_flags flags)
 {
+	struct bch_fs *c = trans->c;
+
 	if ((flags & BTREE_TRIGGER_atomic) && (flags & BTREE_TRIGGER_insert)) {
 		BUG_ON(!trans->journal_res.seq);
 		bkey_s_to_inode_v3(new).v->bi_journal_seq = cpu_to_le64(trans->journal_res.seq);
···
 			return ret;
 	}
 
-	int deleted_delta =	(int) bkey_is_deleted_inode(new.s_c) -
-				(int) bkey_is_deleted_inode(old);
-	if ((flags & BTREE_TRIGGER_transactional) && deleted_delta) {
-		int ret = bch2_btree_bit_mod_buffered(trans, BTREE_ID_deleted_inodes,
-						      new.k->p, deleted_delta > 0);
-		if (ret)
-			return ret;
+	if (flags & BTREE_TRIGGER_transactional) {
+		int unlinked_delta =	(int) bkey_is_unlinked_inode(new.s_c) -
+					(int) bkey_is_unlinked_inode(old);
+		if (unlinked_delta) {
+			int ret = bch2_btree_bit_mod_buffered(trans, BTREE_ID_deleted_inodes,
+							      new.k->p, unlinked_delta > 0);
+			if (ret)
+				return ret;
+		}
+
+		/*
+		 * If we're creating or deleting an inode at this snapshot ID,
+		 * and there might be an inode in a parent snapshot ID, we might
+		 * need to set or clear the has_child_snapshot flag on the
+		 * parent.
+		 */
+		int deleted_delta = (int) bkey_is_inode(new.k) -
+				    (int) bkey_is_inode(old.k);
+		if (deleted_delta &&
+		    bch2_snapshot_parent(c, new.k->p.snapshot)) {
+			int ret = update_parent_inode_has_children(trans, new.k->p,
+								   deleted_delta > 0);
+			if (ret)
+				return ret;
+		}
+
+		/*
+		 * When an inode is first updated in a new snapshot, we may need
+		 * to clear has_child_snapshot
+		 */
+		if (deleted_delta > 0) {
+			int ret = update_inode_has_children(trans, new, false);
+			if (ret)
+				return ret;
+		}
 	}
 
 	return 0;
···
 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 		goto retry;
 
+	if (ret)
+		goto err2;
+
+	ret = delete_ancestor_snapshot_inodes(trans, SPOS(0, inum.inum, snapshot));
+err2:
 	bch2_trans_put(trans);
 	return ret;
 }
···
 	return 0;
 }
 
-int bch2_inode_rm_snapshot(struct btree_trans *trans, u64 inum, u32 snapshot)
+static noinline int __bch2_inode_rm_snapshot(struct btree_trans *trans, u64 inum, u32 snapshot)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter = { NULL };
···
 	return ret ?: -BCH_ERR_transaction_restart_nested;
 }
 
+/*
+ * After deleting an inode, there may be versions in older snapshots that should
+ * also be deleted - if they're not referenced by sibling snapshots and not open
+ * in other subvolumes:
+ */
+static int delete_ancestor_snapshot_inodes(struct btree_trans *trans, struct bpos pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+next_parent:
+	ret = lockrestart_do(trans,
+		bkey_err(k = bch2_inode_get_iter_snapshot_parent(trans, &iter, pos, 0)));
+	if (ret || !k.k)
+		return ret;
+
+	bool unlinked = bkey_is_unlinked_inode(k);
+	pos = k.k->p;
+	bch2_trans_iter_exit(trans, &iter);
+
+	if (!unlinked)
+		return 0;
+
+	ret = lockrestart_do(trans, bch2_inode_or_descendents_is_open(trans, pos));
+	if (ret)
+		return ret < 0 ? ret : 0;
+
+	ret = __bch2_inode_rm_snapshot(trans, pos.offset, pos.snapshot);
+	if (ret)
+		return ret;
+	goto next_parent;
+}
+
+int bch2_inode_rm_snapshot(struct btree_trans *trans, u64 inum, u32 snapshot)
+{
+	return __bch2_inode_rm_snapshot(trans, inum, snapshot) ?:
+		delete_ancestor_snapshot_inodes(trans, SPOS(0, inum, snapshot));
+}
+
 static int may_delete_deleted_inode(struct btree_trans *trans,
 				    struct btree_iter *iter,
 				    struct bpos pos,
···
 	struct btree_iter inode_iter;
 	struct bkey_s_c k;
 	struct bch_inode_unpacked inode;
+	struct printbuf buf = PRINTBUF;
 	int ret;
 
 	k = bch2_bkey_get_iter(trans, &inode_iter, BTREE_ID_inodes, pos, BTREE_ITER_cached);
···
 			pos.offset, pos.snapshot))
 		goto delete;
 
+	if (fsck_err_on(inode.bi_flags & BCH_INODE_has_child_snapshot,
+			trans, deleted_inode_has_child_snapshots,
+			"inode with child snapshots %llu:%u in deleted_inodes btree",
+			pos.offset, pos.snapshot))
+		goto delete;
+
+	ret = bch2_inode_has_child_snapshots(trans, k.k->p);
+	if (ret < 0)
+		goto out;
+
+	if (ret) {
+		if (fsck_err(trans, inode_has_child_snapshots_wrong,
+			     "inode has_child_snapshots flag wrong (should be set)\n%s",
+			     (printbuf_reset(&buf),
+			      bch2_inode_unpacked_to_text(&buf, &inode),
+			      buf.buf))) {
+			inode.bi_flags |= BCH_INODE_has_child_snapshot;
+			ret = __bch2_fsck_write_inode(trans, &inode);
+			if (ret)
+				goto out;
+		}
+		goto delete;
+
+	}
+
 	if (test_bit(BCH_FS_clean_recovery, &c->flags) &&
 	    !fsck_err(trans, deleted_inode_but_clean,
 		      "filesystem marked as clean but have deleted inode %llu:%u",
···
 		goto out;
 	}
 
-	if (bch2_snapshot_is_internal_node(c, pos.snapshot)) {
-		struct bpos new_min_pos;
-
-		ret = bch2_propagate_key_to_snapshot_leaves(trans, inode_iter.btree_id, k, &new_min_pos);
-		if (ret)
-			goto out;
-
-		inode.bi_flags &= ~BCH_INODE_unlinked;
-
-		ret = bch2_inode_write_flags(trans, &inode_iter, &inode,
-					     BTREE_UPDATE_internal_snapshot_node);
-		bch_err_msg(c, ret, "clearing inode unlinked flag");
-		if (ret)
-			goto out;
-
-		/*
-		 * We'll need another write buffer flush to pick up the new
-		 * unlinked inodes in the snapshot leaves:
-		 */
-		*need_another_pass = true;
-		goto out;
-	}
-
 	ret = 1;
 out:
 fsck_err:
 	bch2_trans_iter_exit(trans, &inode_iter);
+	printbuf_exit(&buf);
 	return ret;
 delete:
 	ret = bch2_btree_bit_mod_buffered(trans, BTREE_ID_deleted_inodes, pos, false);

fs/bcachefs/inode.h

···
 #include "bkey.h"
 #include "bkey_methods.h"
 #include "opts.h"
+#include "snapshot.h"
 
 enum bch_validate_flags;
 extern const char * const bch2_inode_opts[];
···
 int bch2_inode_v3_validate(struct bch_fs *, struct bkey_s_c,
 			  enum bch_validate_flags);
 void bch2_inode_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+
+int __bch2_inode_has_child_snapshots(struct btree_trans *, struct bpos);
+
+static inline int bch2_inode_has_child_snapshots(struct btree_trans *trans, struct bpos pos)
+{
+	return bch2_snapshot_is_leaf(trans->c, pos.snapshot) <= 0
+		? __bch2_inode_has_child_snapshots(trans, pos)
+		: 0;
+}
 
 int bch2_trigger_inode(struct btree_trans *, enum btree_id, unsigned,
 		       struct bkey_s_c, struct bkey_s,

fs/bcachefs/inode_format.h

···
 	x(i_size_dirty,			5)	\
 	x(i_sectors_dirty,		6)	\
 	x(unlinked,			7)	\
-	x(backptr_untrusted,		8)
+	x(backptr_untrusted,		8)	\
+	x(has_child_snapshot,		9)
 
 /* bits 20+ reserved for packed fields below: */
 

fs/bcachefs/journal.c

···
 {
 	int ret;
 
+	if (closure_wait_event_timeout(&j->async_wait,
+		   (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
+		   (flags & JOURNAL_RES_GET_NONBLOCK),
+		   HZ * 10))
+		return ret;
+
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+	struct printbuf buf = PRINTBUF;
+	bch2_journal_debug_to_text(&buf, j);
+	bch_err(c, "Journal stuck? Waited for 10 seconds...\n%s",
+		buf.buf);
+	printbuf_exit(&buf);
+
 	closure_wait_event(&j->async_wait,
 		   (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
 		   (flags & JOURNAL_RES_GET_NONBLOCK));

fs/bcachefs/opts.c

···
 			prt_printf(out, "%lli", v);
 		break;
 	case BCH_OPT_STR:
-		if (flags & OPT_SHOW_FULL_LIST)
+		if (v < opt->min || v >= opt->max - 1)
+			prt_printf(out, "(invalid option %lli)", v);
+		else if (flags & OPT_SHOW_FULL_LIST)
 			prt_string_option(out, opt->choices, v);
 		else
 			prt_str(out, opt->choices[v]);

fs/bcachefs/recovery.c

···
 				BCH_TRANS_COMMIT_no_enospc|
 				BCH_TRANS_COMMIT_journal_reclaim|
 				BCH_TRANS_COMMIT_skip_accounting_apply|
-				BCH_TRANS_COMMIT_no_journal_res,
+				BCH_TRANS_COMMIT_no_journal_res|
+				BCH_WATERMARK_reclaim,
 			     bch2_journal_replay_accounting_key(trans, k));
 		if (bch2_fs_fatal_err_on(ret, c, "error replaying accounting; %s", bch2_err_str(ret)))
 			goto err;

fs/bcachefs/recovery_passes_types.h

···
 	x(check_dirents,			27, PASS_FSCK)			\
 	x(check_xattrs,				28, PASS_FSCK)			\
 	x(check_root,				29, PASS_ONLINE|PASS_FSCK)	\
+	x(check_unreachable_inodes,		40, PASS_ONLINE|PASS_FSCK)	\
 	x(check_subvolume_structure,		36, PASS_ONLINE|PASS_FSCK)	\
 	x(check_directory_structure,		30, PASS_ONLINE|PASS_FSCK)	\
 	x(check_nlinks,				31, PASS_FSCK)			\

fs/bcachefs/replicas.c

···
 	prt_printf(out, "]");
 }
 
-static int bch2_replicas_entry_validate_locked(struct bch_replicas_entry_v1 *r,
-					       struct bch_sb *sb,
-					       struct printbuf *err)
+static int bch2_replicas_entry_sb_validate(struct bch_replicas_entry_v1 *r,
+					   struct bch_sb *sb,
+					   struct printbuf *err)
 {
 	if (!r->nr_devs) {
 		prt_printf(err, "no devices in entry ");
···
 				 struct bch_fs *c,
 				 struct printbuf *err)
 {
-	mutex_lock(&c->sb_lock);
-	int ret = bch2_replicas_entry_validate_locked(r, c->disk_sb.sb, err);
-	mutex_unlock(&c->sb_lock);
-	return ret;
+	if (!r->nr_devs) {
+		prt_printf(err, "no devices in entry ");
+		goto bad;
+	}
+
+	if (r->nr_required > 1 &&
+	    r->nr_required >= r->nr_devs) {
+		prt_printf(err, "bad nr_required in entry ");
+		goto bad;
+	}
+
+	for (unsigned i = 0; i < r->nr_devs; i++)
+		if (r->devs[i] != BCH_SB_MEMBER_INVALID &&
+		    !bch2_dev_exists(c, r->devs[i])) {
+			prt_printf(err, "invalid device %u in entry ", r->devs[i]);
+			goto bad;
+		}
+
+	return 0;
+bad:
+	bch2_replicas_entry_to_text(err, r);
+	return -BCH_ERR_invalid_replicas_entry;
 }
 
 void bch2_cpu_replicas_to_text(struct printbuf *out,
···
 		struct bch_replicas_entry_v1 *e =
 			cpu_replicas_entry(cpu_r, i);
 
-		int ret = bch2_replicas_entry_validate_locked(e, sb, err);
+		int ret = bch2_replicas_entry_sb_validate(e, sb, err);
 		if (ret)
 			return ret;
 
···
 
 		rcu_read_lock();
 		for (unsigned i = 0; i < e->nr_devs; i++) {
+			if (e->devs[i] == BCH_SB_MEMBER_INVALID) {
+				nr_failed++;
+				continue;
+			}
+
 			nr_online += test_bit(e->devs[i], devs.d);
 
 			struct bch_dev *ca = bch2_dev_rcu_noerror(c, e->devs[i]);

fs/bcachefs/sb-downgrade.c

···
 	  BCH_FSCK_ERR_accounting_mismatch)			\
 	x(rebalance_work_acct_fix,				\
 	  BIT_ULL(BCH_RECOVERY_PASS_check_allocations),		\
-	  BCH_FSCK_ERR_accounting_mismatch)
+	  BCH_FSCK_ERR_accounting_mismatch)			\
+	x(inode_has_child_snapshots,				\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_inodes),		\
+	  BCH_FSCK_ERR_inode_has_child_snapshots_wrong)
 
 #define DOWNGRADE_TABLE()					\
 	x(bucket_stripe_sectors,				\

fs/bcachefs/sb-errors_format.h

···
 	x(reflink_p_to_missing_reflink_v,			166,	0)		\
 	x(stripe_pos_bad,					167,	0)		\
 	x(stripe_val_size_bad,					168,	0)		\
+	x(stripe_csum_granularity_bad,				290,	0)		\
 	x(stripe_sector_count_wrong,				169,	0)		\
 	x(snapshot_tree_pos_bad,				170,	0)		\
 	x(snapshot_tree_to_missing_snapshot,			171,	0)		\
···
 	x(inode_multiple_links_but_nlink_0,			207,	FSCK_AUTOFIX)	\
 	x(inode_wrong_backpointer,				208,	FSCK_AUTOFIX)	\
 	x(inode_wrong_nlink,					209,	FSCK_AUTOFIX)	\
+	x(inode_has_child_snapshots_wrong,			287,	0)		\
 	x(inode_unreachable,					210,	FSCK_AUTOFIX)	\
 	x(deleted_inode_but_clean,				211,	FSCK_AUTOFIX)	\
 	x(deleted_inode_missing,				212,	FSCK_AUTOFIX)	\
 	x(deleted_inode_is_dir,					213,	FSCK_AUTOFIX)	\
 	x(deleted_inode_not_unlinked,				214,	FSCK_AUTOFIX)	\
+	x(deleted_inode_has_child_snapshots,			288,	FSCK_AUTOFIX)	\
 	x(extent_overlapping,					215,	0)		\
 	x(key_in_missing_inode,					216,	0)		\
 	x(key_in_wrong_inode_type,				217,	0)		\
···
 	x(alloc_key_stripe_sectors_wrong,			271,	FSCK_AUTOFIX)	\
 	x(accounting_mismatch,					272,	FSCK_AUTOFIX)	\
 	x(accounting_replicas_not_marked,			273,	0)		\
+	x(accounting_to_invalid_device,				289,	0)		\
 	x(invalid_btree_id,					274,	0)		\
 	x(alloc_key_io_time_bad,				275,	0)		\
 	x(alloc_key_fragmentation_lru_wrong,			276,	FSCK_AUTOFIX)	\
···
 	x(accounting_key_replicas_devs_unsorted,		280,	FSCK_AUTOFIX)	\
 	x(accounting_key_version_0,				282,	FSCK_AUTOFIX)	\
 	x(logged_op_but_clean,					283,	FSCK_AUTOFIX)	\
-	x(MAX,							287,	0)
+	x(MAX,							291,	0)
 
 enum bch_sb_error_id {
 #define x(t, n, ...) BCH_FSCK_ERR_##t = n,

fs/bcachefs/sb-members.c

···
 		return -BCH_ERR_invalid_sb_members;
 	}
 
+	if (m.btree_bitmap_shift >= 64) {
+		prt_printf(err, "device %u: invalid btree_bitmap_shift %u", i, m.btree_bitmap_shift);
+		return -BCH_ERR_invalid_sb_members;
+	}
+
 	return 0;
 }
 
···
 	prt_newline(out);
 
 	prt_printf(out, "Btree allocated bitmap blocksize:\t");
-	prt_units_u64(out, 1ULL << m.btree_bitmap_shift);
+	if (m.btree_bitmap_shift < 64)
+		prt_units_u64(out, 1ULL << m.btree_bitmap_shift);
+	else
+		prt_printf(out, "(invalid shift %u)", m.btree_bitmap_shift);
 	prt_newline(out);
 
 	prt_printf(out, "Btree allocated bitmap:\t");

fs/bcachefs/snapshot.c

···
 	if (bch2_snapshot_equiv(c, id))
 		return 0;
 
-	/* 0 is an invalid tree ID */
+	/* Do we need to reconstruct the snapshot_tree entry as well? */
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
 	u32 tree_id = 0;
-	int ret = bch2_snapshot_tree_create(trans, id, 0, &tree_id);
+
+	for_each_btree_key_norestart(trans, iter, BTREE_ID_snapshot_trees, POS_MIN,
+				     0, k, ret) {
+		if (le32_to_cpu(bkey_s_c_to_snapshot_tree(k).v->root_snapshot) == id) {
+			tree_id = k.k->p.offset;
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
 	if (ret)
 		return ret;
+
+	if (!tree_id) {
+		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);
···
 	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));
+
+	for_each_btree_key_norestart(trans, iter, BTREE_ID_subvolumes, POS_MIN,
+				     0, k, ret) {
+		if (le32_to_cpu(bkey_s_c_to_subvolume(k).v->snapshot) == id) {
+			snapshot->v.subvol = cpu_to_le32(k.k->p.offset);
+			SET_BCH_SNAPSHOT_SUBVOL(&snapshot->v, true);
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
 
 	return  bch2_btree_insert_trans(trans, BTREE_ID_snapshots, &snapshot->k_i, 0) ?:
 		bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0,
···
 	bch2_trans_iter_exit(trans, &iter);
 
 	return ret;
-}
-
-static u32 bch2_snapshot_smallest_child(struct bch_fs *c, u32 id)
-{
-	const struct snapshot_t *s = snapshot_t(c, id);
-
-	return s->children[1] ?: s->children[0];
-}
-
-static u32 bch2_snapshot_smallest_descendent(struct bch_fs *c, u32 id)
-{
-	u32 child;
-
-	while ((child = bch2_snapshot_smallest_child(c, id)))
-		id = child;
-	return id;
-}
-
-static int bch2_propagate_key_to_snapshot_leaf(struct btree_trans *trans,
-					       enum btree_id btree,
-					       struct bkey_s_c interior_k,
-					       u32 leaf_id, struct bpos *new_min_pos)
-{
-	struct btree_iter iter;
-	struct bpos pos = interior_k.k->p;
-	struct bkey_s_c k;
-	struct bkey_i *new;
-	int ret;
-
-	pos.snapshot = leaf_id;
-
-	bch2_trans_iter_init(trans, &iter, btree, pos, BTREE_ITER_intent);
-	k = bch2_btree_iter_peek_slot(&iter);
-	ret = bkey_err(k);
-	if (ret)
-		goto out;
-
-	/* key already overwritten in this snapshot? */
-	if (k.k->p.snapshot != interior_k.k->p.snapshot)
-		goto out;
-
-	if (bpos_eq(*new_min_pos, POS_MIN)) {
-		*new_min_pos = k.k->p;
-		new_min_pos->snapshot = leaf_id;
-	}
-
-	new = bch2_bkey_make_mut_noupdate(trans, interior_k);
-	ret = PTR_ERR_OR_ZERO(new);
-	if (ret)
-		goto out;
-
-	new->k.p.snapshot = leaf_id;
-	ret = bch2_trans_update(trans, &iter, new, 0);
-out:
-	bch2_set_btree_iter_dontneed(&iter);
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
-}
-
-int bch2_propagate_key_to_snapshot_leaves(struct btree_trans *trans,
-					  enum btree_id btree,
-					  struct bkey_s_c k,
-					  struct bpos *new_min_pos)
-{
-	struct bch_fs *c = trans->c;
-	struct bkey_buf sk;
-	u32 restart_count = trans->restart_count;
-	int ret = 0;
-
-	bch2_bkey_buf_init(&sk);
-	bch2_bkey_buf_reassemble(&sk, c, k);
-	k = bkey_i_to_s_c(sk.k);
-
-	*new_min_pos = POS_MIN;
-
-	for (u32 id = bch2_snapshot_smallest_descendent(c, k.k->p.snapshot);
-	     id < k.k->p.snapshot;
-	     id++) {
-		if (!bch2_snapshot_is_ancestor(c, id, k.k->p.snapshot) ||
-		    !bch2_snapshot_is_leaf(c, id))
-			continue;
-again:
-		ret =   btree_trans_too_many_iters(trans) ?:
-			bch2_propagate_key_to_snapshot_leaf(trans, btree, k, id, new_min_pos) ?:
-			bch2_trans_commit(trans, NULL, NULL, 0);
-		if (ret && bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
-			bch2_trans_begin(trans);
-			goto again;
-		}
-
-		if (ret)
-			break;
-	}
-
-	bch2_bkey_buf_exit(&sk, c);
-
-	return ret ?: trans_was_restarted(trans, restart_count);
 }
 
 static int bch2_check_snapshot_needs_deletion(struct btree_trans *trans, struct bkey_s_c k)

fs/bcachefs/snapshot.h

···
 	return __bch2_key_has_snapshot_overwrites(trans, id, pos);
 }
 
-int bch2_propagate_key_to_snapshot_leaves(struct btree_trans *, enum btree_id,
-					  struct bkey_s_c, struct bpos *);
-
 int bch2_snapshots_read(struct bch_fs *);
 void bch2_fs_snapshots_exit(struct bch_fs *);
 

fs/bcachefs/super.c

···
 
 DECLARE_WAIT_QUEUE_HEAD(bch2_read_only_wait);
 
+static void bch2_dev_unlink(struct bch_dev *);
 static void bch2_dev_free(struct bch_dev *);
 static int bch2_dev_alloc(struct bch_fs *, unsigned);
 static int bch2_dev_sysfs_online(struct bch_fs *, struct bch_dev *);
···
 	up_write(&c->state_lock);
 
 	for_each_member_device(c, ca)
-		if (ca->kobj.state_in_sysfs &&
-		    ca->disk_sb.bdev)
-			sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
+		bch2_dev_unlink(ca);
 
 	if (c->kobj.state_in_sysfs)
 		kobject_del(&c->kobj);
···
 {
 	cancel_work_sync(&ca->io_error_work);
 
-	if (ca->kobj.state_in_sysfs &&
-	    ca->disk_sb.bdev)
-		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
+	bch2_dev_unlink(ca);
 
 	if (ca->kobj.state_in_sysfs)
 		kobject_del(&ca->kobj);
···
 	percpu_ref_kill(&ca->io_ref);
 	wait_for_completion(&ca->io_ref_completion);
 
-	if (ca->kobj.state_in_sysfs) {
-		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
-		sysfs_remove_link(&ca->kobj, "block");
-	}
+	bch2_dev_unlink(ca);
 
 	bch2_free_super(&ca->disk_sb);
 	bch2_dev_journal_exit(ca);
···
 	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
 
 	complete(&ca->io_ref_completion);
+}
+
+static void bch2_dev_unlink(struct bch_dev *ca)
+{
+	struct kobject *b;
+
+	/*
+	 * This is racy w.r.t. the underlying block device being hot-removed,
+	 * which removes it from sysfs.
+	 *
+	 * It'd be lovely if we had a way to handle this race, but the sysfs
+	 * code doesn't appear to provide a good method and block/holder.c is
+	 * susceptible as well:
+	 */
+	if (ca->kobj.state_in_sysfs &&
+	    ca->disk_sb.bdev &&
+	    (b = bdev_kobj(ca->disk_sb.bdev))->state_in_sysfs) {
+		sysfs_remove_link(b, "bcachefs");
+		sysfs_remove_link(&ca->kobj, "block");
+	}
 }
 
 static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)

include/linux/closure.h

···
 		__closure_wait_event(waitlist, _cond);			\
 } while (0)
 
+#define __closure_wait_event_timeout(waitlist, _cond, _until)		\
+({									\
+	struct closure cl;						\
+	long _t;							\
+									\
+	closure_init_stack(&cl);					\
+									\
+	while (1) {							\
+		closure_wait(waitlist, &cl);				\
+		if (_cond) {						\
+			_t = max_t(long, 1L, _until - jiffies);		\
+			break;						\
+		}							\
+		_t = max_t(long, 0L, _until - jiffies);			\
+		if (!_t)						\
+			break;						\
+		closure_sync_timeout(&cl, _t);				\
+	}								\
+	closure_wake_up(waitlist);					\
+	closure_sync(&cl);						\
+	_t;								\
+})
+
+/*
+ * Returns 0 if timeout expired, remaining time in jiffies (at least 1) if
+ * condition became true
+ */
+#define closure_wait_event_timeout(waitlist, _cond, _timeout)		\
+({									\
+	unsigned long _until = jiffies + _timeout;			\
+	(_cond)								\
+		? max_t(long, 1L, _until - jiffies)			\
+		: __closure_wait_event_timeout(waitlist, _cond, _until);\
+})
+
 #endif /* _LINUX_CLOSURE_H */