bcachefs: Plumb through subvolume id

To implement snapshots, we need every filesystem btree operation (every
btree operation without a subvolume) to start by looking up the
subvolume and getting the current snapshot ID, with
bch2_subvolume_get_snapshot() - then, that snapshot ID is used for doing
btree lookups in BTREE_ITER_FILTER_SNAPSHOTS mode.

This patch adds those bch2_subvolume_get_snapshot() calls, and also
switches to passing around a subvol_inum instead of just an inode
number.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

fs/bcachefs/acl.c

@@ -230,7 +230,7 @@ struct posix_acl *bch2_get_acl(struct mnt_idmap *idmap,
 	bch2_trans_begin(&trans);
 
 	ret = bch2_hash_lookup(&trans, &iter, bch2_xattr_hash_desc,
-			&hash, inode->v.i_ino,
+			&hash, inode_inum(inode),
 			&X_SEARCH(acl_to_xattr_type(type), "", 0),
 			0);
 	if (ret) {
@@ -260,11 +260,11 @@ struct posix_acl *bch2_get_acl(struct mnt_idmap *idmap,
 	return acl;
 }
 
-int bch2_set_acl_trans(struct btree_trans *trans,
+int bch2_set_acl_trans(struct btree_trans *trans, subvol_inum inum,
 		       struct bch_inode_unpacked *inode_u,
-		       const struct bch_hash_info *hash_info,
 		       struct posix_acl *acl, int type)
 {
+	struct bch_hash_info hash_info = bch2_hash_info_init(trans->c, inode_u);
 	int ret;
 
 	if (type == ACL_TYPE_DEFAULT &&
@@ -277,14 +277,14 @@ int bch2_set_acl_trans(struct btree_trans *trans,
 		if (IS_ERR(xattr))
 			return PTR_ERR(xattr);
 
-		ret = bch2_hash_set(trans, bch2_xattr_hash_desc, hash_info,
-				    inode_u->bi_inum, &xattr->k_i, 0);
+		ret = bch2_hash_set(trans, bch2_xattr_hash_desc, &hash_info,
+				    inum, &xattr->k_i, 0);
 	} else {
 		struct xattr_search_key search =
 			X_SEARCH(acl_to_xattr_type(type), "", 0);
 
-		ret = bch2_hash_delete(trans, bch2_xattr_hash_desc, hash_info,
-				       inode_u->bi_inum, &search);
+		ret = bch2_hash_delete(trans, bch2_xattr_hash_desc, &hash_info,
+				       inum, &search);
 	}
 
 	return ret == -ENOENT ? 0 : ret;
@@ -299,7 +299,6 @@ int bch2_set_acl(struct mnt_idmap *idmap,
 	struct btree_trans trans;
 	struct btree_iter inode_iter = { NULL };
 	struct bch_inode_unpacked inode_u;
-	struct bch_hash_info hash_info;
 	struct posix_acl *acl;
 	umode_t mode;
 	int ret;
@@ -310,7 +309,7 @@ int bch2_set_acl(struct mnt_idmap *idmap,
 	bch2_trans_begin(&trans);
 	acl = _acl;
 
-	ret = bch2_inode_peek(&trans, &inode_iter, &inode_u, inode->v.i_ino,
+	ret = bch2_inode_peek(&trans, &inode_iter, &inode_u, inode_inum(inode),
 			      BTREE_ITER_INTENT);
 	if (ret)
 		goto btree_err;
@@ -323,9 +322,7 @@ int bch2_set_acl(struct mnt_idmap *idmap,
 			goto btree_err;
 	}
 
-	hash_info = bch2_hash_info_init(c, &inode_u);
-
-	ret = bch2_set_acl_trans(&trans, &inode_u, &hash_info, acl, type);
+	ret = bch2_set_acl_trans(&trans, inode_inum(inode), &inode_u, acl, type);
 	if (ret)
 		goto btree_err;
 
@@ -354,7 +351,7 @@ int bch2_set_acl(struct mnt_idmap *idmap,
 	return ret;
 }
 
-int bch2_acl_chmod(struct btree_trans *trans,
+int bch2_acl_chmod(struct btree_trans *trans, subvol_inum inum,
 		   struct bch_inode_unpacked *inode,
 		   umode_t mode,
 		   struct posix_acl **new_acl)
@@ -368,7 +365,7 @@ int bch2_acl_chmod(struct btree_trans *trans,
 	int ret;
 
 	ret = bch2_hash_lookup(trans, &iter, bch2_xattr_hash_desc,
-			&hash_info, inode->bi_inum,
+			       &hash_info, inum,
 			&X_SEARCH(KEY_TYPE_XATTR_INDEX_POSIX_ACL_ACCESS, "", 0),
 			BTREE_ITER_INTENT);
 	if (ret)

fs/bcachefs/acl.h

@@ -28,25 +28,24 @@ typedef struct {
 
 struct posix_acl *bch2_get_acl(struct mnt_idmap *, struct dentry *, int);
 
-int bch2_set_acl_trans(struct btree_trans *,
+int bch2_set_acl_trans(struct btree_trans *, subvol_inum,
 		       struct bch_inode_unpacked *,
-		       const struct bch_hash_info *,
 		       struct posix_acl *, int);
 int bch2_set_acl(struct mnt_idmap *, struct dentry *, struct posix_acl *, int);
-int bch2_acl_chmod(struct btree_trans *, struct bch_inode_unpacked *,
+int bch2_acl_chmod(struct btree_trans *, subvol_inum,
+		   struct bch_inode_unpacked *,
 		   umode_t, struct posix_acl **);
 
 #else
 
-static inline int bch2_set_acl_trans(struct btree_trans *trans,
+static inline int bch2_set_acl_trans(struct btree_trans *trans, subvol_inum inum,
 				     struct bch_inode_unpacked *inode_u,
-				     const struct bch_hash_info *hash_info,
 				     struct posix_acl *acl, int type)
 {
 	return 0;
 }
 
-static inline int bch2_acl_chmod(struct btree_trans *trans,
+static inline int bch2_acl_chmod(struct btree_trans *trans, subvol_inum inum,
 				 struct bch_inode_unpacked *inode,
 				 umode_t mode,
 				 struct posix_acl **new_acl)

fs/bcachefs/dirent.c

@@ -8,6 +8,7 @@
 #include "fs.h"
 #include "keylist.h"
 #include "str_hash.h"
+#include "subvolume.h"
 
 #include <linux/dcache.h>
 
@@ -150,8 +151,8 @@ static struct bkey_i_dirent *dirent_create_key(struct btree_trans *trans,
 	return dirent;
 }
 
-int bch2_dirent_create(struct btree_trans *trans,
-		       u64 dir_inum, const struct bch_hash_info *hash_info,
+int bch2_dirent_create(struct btree_trans *trans, subvol_inum dir,
+		       const struct bch_hash_info *hash_info,
 		       u8 type, const struct qstr *name, u64 dst_inum,
 		       u64 *dir_offset, int flags)
 {
@@ -164,7 +165,7 @@ int bch2_dirent_create(struct btree_trans *trans,
 		return ret;
 
 	ret = bch2_hash_set(trans, bch2_dirent_hash_desc, hash_info,
-			    dir_inum, &dirent->k_i, flags);
+			    dir, &dirent->k_i, flags);
 	*dir_offset = dirent->k.p.offset;
 
 	return ret;
@@ -223,31 +224,40 @@ int __bch2_dirent_read_target(struct btree_trans *trans,
 	return ret;
 }
 
-int bch2_dirent_read_target(struct btree_trans *trans,
-			    struct bkey_s_c_dirent d, u64 *target)
+static int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
+				   struct bkey_s_c_dirent d, subvol_inum *target)
 {
-	u32 subvol, snapshot;
+	u32 snapshot;
+	int ret = 0;
 
-	return __bch2_dirent_read_target(trans, d, &subvol,
-					 &snapshot, target, false);
+	ret = __bch2_dirent_read_target(trans, d, &target->subvol, &snapshot,
+					&target->inum, false);
+	if (!target->subvol)
+		target->subvol = dir.subvol;
+
+	return ret;
 }
 
 int bch2_dirent_rename(struct btree_trans *trans,
-		       u64 src_dir, struct bch_hash_info *src_hash,
-		       u64 dst_dir, struct bch_hash_info *dst_hash,
-		       const struct qstr *src_name, u64 *src_inum, u64 *src_offset,
-		       const struct qstr *dst_name, u64 *dst_inum, u64 *dst_offset,
-		       enum bch_rename_mode mode)
+		subvol_inum src_dir, struct bch_hash_info *src_hash,
+		subvol_inum dst_dir, struct bch_hash_info *dst_hash,
+		const struct qstr *src_name, subvol_inum *src_inum, u64 *src_offset,
+		const struct qstr *dst_name, subvol_inum *dst_inum, u64 *dst_offset,
+		enum bch_rename_mode mode)
 {
 	struct btree_iter src_iter = { NULL };
 	struct btree_iter dst_iter = { NULL };
 	struct bkey_s_c old_src, old_dst;
 	struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
 	struct bpos dst_pos =
-		POS(dst_dir, bch2_dirent_hash(dst_hash, dst_name));
+		POS(dst_dir.inum, bch2_dirent_hash(dst_hash, dst_name));
 	int ret = 0;
 
-	*src_inum = *dst_inum = 0;
+	if (src_dir.subvol != dst_dir.subvol)
+		return -EXDEV;
+
+	memset(src_inum, 0, sizeof(*src_inum));
+	memset(dst_inum, 0, sizeof(*dst_inum));
 
 	/*
 	 * Lookup dst:
@@ -270,8 +280,12 @@ int bch2_dirent_rename(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	if (mode != BCH_RENAME)
-		*dst_inum = le64_to_cpu(bkey_s_c_to_dirent(old_dst).v->d_inum);
+	if (mode != BCH_RENAME) {
+		ret = bch2_dirent_read_target(trans, dst_dir,
+				bkey_s_c_to_dirent(old_dst), dst_inum);
+		if (ret)
+			goto out;
+	}
 	if (mode != BCH_RENAME_EXCHANGE)
 		*src_offset = dst_iter.pos.offset;
 
@@ -287,7 +301,10 @@ int bch2_dirent_rename(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	*src_inum = le64_to_cpu(bkey_s_c_to_dirent(old_src).v->d_inum);
+	ret = bch2_dirent_read_target(trans, src_dir,
+			bkey_s_c_to_dirent(old_src), src_inum);
+	if (ret)
+		goto out;
 
 	/* Create new dst key: */
 	new_dst = dirent_create_key(trans, 0, dst_name, 0);
@@ -376,17 +393,22 @@ int bch2_dirent_delete_at(struct btree_trans *trans,
 
 int __bch2_dirent_lookup_trans(struct btree_trans *trans,
 			       struct btree_iter *iter,
-			       u64 dir_inum,
+			       subvol_inum dir,
 			       const struct bch_hash_info *hash_info,
-			       const struct qstr *name, u64 *inum,
+			       const struct qstr *name, subvol_inum *inum,
 			       unsigned flags)
 {
 	struct bkey_s_c k;
 	struct bkey_s_c_dirent d;
+	u32 snapshot;
 	int ret;
 
+	ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
+	if (ret)
+		return ret;
+
 	ret = bch2_hash_lookup(trans, iter, bch2_dirent_hash_desc,
-			       hash_info, dir_inum, name, flags);
+			       hash_info, dir, name, flags);
 	if (ret)
 		return ret;
 
@@ -399,44 +421,49 @@ int __bch2_dirent_lookup_trans(struct btree_trans *trans,
 
 	d = bkey_s_c_to_dirent(k);
 
-	ret = bch2_dirent_read_target(trans, d, inum);
+	ret = bch2_dirent_read_target(trans, dir, d, inum);
 	if (ret)
 		bch2_trans_iter_exit(trans, iter);
 
 	return ret;
 }
 
-u64 bch2_dirent_lookup(struct bch_fs *c, u64 dir_inum,
+u64 bch2_dirent_lookup(struct bch_fs *c, subvol_inum dir,
 		       const struct bch_hash_info *hash_info,
-		       const struct qstr *name)
+		       const struct qstr *name, subvol_inum *inum)
 {
 	struct btree_trans trans;
 	struct btree_iter iter;
-	u64 inum = 0;
-	int ret = 0;
+	int ret;
 
 	bch2_trans_init(&trans, c, 0, 0);
 retry:
 	bch2_trans_begin(&trans);
-	ret = __bch2_dirent_lookup_trans(&trans, &iter, dir_inum, hash_info,
-					 name, &inum, 0);
+
+	ret = __bch2_dirent_lookup_trans(&trans, &iter, dir, hash_info,
+					  name, inum, 0);
 
 	bch2_trans_iter_exit(&trans, &iter);
 	if (ret == -EINTR)
 		goto retry;
 	bch2_trans_exit(&trans);
-	return inum;
+	return ret;
 }
 
-int bch2_empty_dir_trans(struct btree_trans *trans, u64 dir_inum)
+int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
 {
 	struct btree_iter iter;
 	struct bkey_s_c k;
+	u32 snapshot;
 	int ret;
 
+	ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
+	if (ret)
+		return ret;
+
 	for_each_btree_key(trans, iter, BTREE_ID_dirents,
-			   POS(dir_inum, 0), 0, k, ret) {
-		if (k.k->p.inode > dir_inum)
+			   SPOS(dir.inum, 0, snapshot), 0, k, ret) {
+		if (k.k->p.inode > dir.inum)
 			break;
 
 		if (k.k->type == KEY_TYPE_dirent) {
@@ -449,19 +476,26 @@ int bch2_empty_dir_trans(struct btree_trans *trans, u64 dir_inum)
 	return ret;
 }
 
-int bch2_readdir(struct bch_fs *c, u64 inum, struct dir_context *ctx)
+int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
 {
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct bkey_s_c_dirent dirent;
+	u32 snapshot;
 	int ret;
 
 	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
 
 	for_each_btree_key(&trans, iter, BTREE_ID_dirents,
-			   POS(inum, ctx->pos), 0, k, ret) {
-		if (k.k->p.inode > inum)
+			   SPOS(inum.inum, ctx->pos, snapshot), 0, k, ret) {
+		if (k.k->p.inode > inum.inum)
 			break;
 
 		if (k.k->type != KEY_TYPE_dirent)
@@ -482,6 +516,9 @@ int bch2_readdir(struct bch_fs *c, u64 inum, struct dir_context *ctx)
 		ctx->pos = dirent.k->p.offset + 1;
 	}
 	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (ret == -EINTR)
+		goto retry;
 
 	ret = bch2_trans_exit(&trans) ?: ret;
 

fs/bcachefs/dirent.h

@@ -29,7 +29,7 @@ static inline unsigned dirent_val_u64s(unsigned len)
 			    sizeof(u64));
 }
 
-int bch2_dirent_create(struct btree_trans *, u64,
+int bch2_dirent_create(struct btree_trans *, subvol_inum,
 		       const struct bch_hash_info *, u8,
 		       const struct qstr *, u64, u64 *, int);
 
@@ -40,9 +40,6 @@ int bch2_dirent_delete_at(struct btree_trans *,
 int __bch2_dirent_read_target(struct btree_trans *, struct bkey_s_c_dirent,
 			      u32 *, u32 *, u64 *, bool);
 
-int bch2_dirent_read_target(struct btree_trans *,
-			    struct bkey_s_c_dirent, u64 *);
-
 static inline unsigned vfs_d_type(unsigned type)
 {
 	return type == DT_SUBVOL ? DT_DIR : type;
@@ -55,20 +52,20 @@ enum bch_rename_mode {
 };
 
 int bch2_dirent_rename(struct btree_trans *,
-		       u64, struct bch_hash_info *,
-		       u64, struct bch_hash_info *,
-		       const struct qstr *, u64 *, u64 *,
-		       const struct qstr *, u64 *, u64 *,
+		       subvol_inum, struct bch_hash_info *,
+		       subvol_inum, struct bch_hash_info *,
+		       const struct qstr *, subvol_inum *, u64 *,
+		       const struct qstr *, subvol_inum *, u64 *,
 		       enum bch_rename_mode);
 
-int __bch2_dirent_lookup_trans(struct btree_trans *, struct btree_iter *, u64,
-			   const struct bch_hash_info *,
-			   const struct qstr *, u64 *,
-			   unsigned);
-u64 bch2_dirent_lookup(struct bch_fs *, u64, const struct bch_hash_info *,
-		       const struct qstr *);
+int __bch2_dirent_lookup_trans(struct btree_trans *, struct btree_iter *,
+			       subvol_inum, const struct bch_hash_info *,
+			       const struct qstr *, subvol_inum *, unsigned);
+u64 bch2_dirent_lookup(struct bch_fs *, subvol_inum,
+		       const struct bch_hash_info *,
+		       const struct qstr *, subvol_inum *);
 
-int bch2_empty_dir_trans(struct btree_trans *, u64);
-int bch2_readdir(struct bch_fs *, u64, struct dir_context *);
+int bch2_empty_dir_trans(struct btree_trans *, subvol_inum);
+int bch2_readdir(struct bch_fs *, subvol_inum, struct dir_context *);
 
 #endif /* _BCACHEFS_DIRENT_H */

fs/bcachefs/extents.c

@@ -611,38 +611,6 @@ bool bch2_bkey_is_incompressible(struct bkey_s_c k)
 	return false;
 }
 
-bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
-				unsigned nr_replicas, bool compressed)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	struct bpos end = pos;
-	struct bkey_s_c k;
-	bool ret = true;
-	int err;
-
-	end.offset += size;
-
-	bch2_trans_init(&trans, c, 0, 0);
-
-	for_each_btree_key(&trans, iter, BTREE_ID_extents, pos,
-			   BTREE_ITER_SLOTS, k, err) {
-		if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
-			break;
-
-		if (nr_replicas > bch2_bkey_replicas(c, k) ||
-		    (!compressed && bch2_bkey_sectors_compressed(k))) {
-			ret = false;
-			break;
-		}
-	}
-	bch2_trans_iter_exit(&trans, &iter);
-
-	bch2_trans_exit(&trans);
-
-	return ret;
-}
-
 unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);

fs/bcachefs/extents.h

@@ -567,7 +567,6 @@ unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c);
 unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c);
 bool bch2_bkey_is_incompressible(struct bkey_s_c);
 unsigned bch2_bkey_sectors_compressed(struct bkey_s_c);
-bool bch2_check_range_allocated(struct bch_fs *, struct bpos, u64, unsigned, bool);
 
 unsigned bch2_bkey_replicas(struct bch_fs *, struct bkey_s_c);
 unsigned bch2_bkey_durability(struct bch_fs *, struct bkey_s_c);

fs/bcachefs/fs-common.h

@@ -4,27 +4,30 @@
 
 struct posix_acl;
 
-int bch2_create_trans(struct btree_trans *, u64,
+#define BCH_CREATE_TMPFILE		(1U << 0)
+
+int bch2_create_trans(struct btree_trans *, subvol_inum,
 		      struct bch_inode_unpacked *,
 		      struct bch_inode_unpacked *,
 		      const struct qstr *,
 		      uid_t, gid_t, umode_t, dev_t,
 		      struct posix_acl *,
-		      struct posix_acl *);
+		      struct posix_acl *,
+		      unsigned);
 
-int bch2_link_trans(struct btree_trans *, u64,
-		    u64, struct bch_inode_unpacked *,
-		    struct bch_inode_unpacked *,
+int bch2_link_trans(struct btree_trans *,
+		    subvol_inum, struct bch_inode_unpacked *,
+		    subvol_inum, struct bch_inode_unpacked *,
 		    const struct qstr *);
 
-int bch2_unlink_trans(struct btree_trans *,
-		      u64, struct bch_inode_unpacked *,
+int bch2_unlink_trans(struct btree_trans *, subvol_inum,
+		      struct bch_inode_unpacked *,
 		      struct bch_inode_unpacked *,
 		      const struct qstr *);
 
 int bch2_rename_trans(struct btree_trans *,
-		      u64, struct bch_inode_unpacked *,
-		      u64, struct bch_inode_unpacked *,
+		      subvol_inum, struct bch_inode_unpacked *,
+		      subvol_inum, struct bch_inode_unpacked *,
 		      struct bch_inode_unpacked *,
 		      struct bch_inode_unpacked *,
 		      const struct qstr *,

fs/bcachefs/fs-io.c

@@ -1790,6 +1790,49 @@ ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
 
 /* O_DIRECT writes */
 
+static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
+				       u64 offset, u64 size,
+				       unsigned nr_replicas, bool compressed)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u64 end = offset + size;
+	u32 snapshot;
+	bool ret = true;
+	int err;
+
+	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (err)
+		goto err;
+
+	for_each_btree_key(&trans, iter, BTREE_ID_extents,
+			   SPOS(inum.inum, offset, snapshot),
+			   BTREE_ITER_SLOTS, k, err) {
+		if (bkey_cmp(bkey_start_pos(k.k), POS(inum.inum, end)) >= 0)
+			break;
+
+		if (nr_replicas > bch2_bkey_replicas(c, k) ||
+		    (!compressed && bch2_bkey_sectors_compressed(k))) {
+			ret = false;
+			break;
+		}
+	}
+
+	offset = iter.pos.offset;
+	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (err == -EINTR)
+		goto retry;
+	bch2_trans_exit(&trans);
+
+	return err ? false : ret;
+}
+
 /*
  * We're going to return -EIOCBQUEUED, but we haven't finished consuming the
  * iov_iter yet, so we need to stash a copy of the iovec: it might be on the
@@ -1911,8 +1954,8 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 		ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
 						dio->op.opts.data_replicas, 0);
 		if (unlikely(ret) &&
-		    !bch2_check_range_allocated(c, dio->op.pos,
-				bio_sectors(bio),
+		    !bch2_check_range_allocated(c, inode_inum(inode),
+				dio->op.pos.offset, bio_sectors(bio),
 				dio->op.opts.data_replicas,
 				dio->op.opts.compression != 0))
 			goto err;
@@ -2141,9 +2184,9 @@ int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
 
 /* truncate: */
 
-static inline int range_has_data(struct bch_fs *c,
-				  struct bpos start,
-				  struct bpos end)
+static inline int range_has_data(struct bch_fs *c, u32 subvol,
+				 struct bpos start,
+				 struct bpos end)
 {
 	struct btree_trans trans;
 	struct btree_iter iter;
@@ -2151,6 +2194,12 @@ static inline int range_has_data(struct bch_fs *c,
 	int ret = 0;
 
 	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
+	if (ret)
+		goto err;
 
 	for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
 		if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
@@ -2161,7 +2210,11 @@ static inline int range_has_data(struct bch_fs *c,
 			break;
 		}
 	}
+	start = iter.pos;
 	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (ret == -EINTR)
+		goto retry;
 
 	return bch2_trans_exit(&trans) ?: ret;
 }
@@ -2193,7 +2246,7 @@ static int __bch2_truncate_page(struct bch_inode_info *inode,
 		 * XXX: we're doing two index lookups when we end up reading the
 		 * page
 		 */
-		ret = range_has_data(c,
+		ret = range_has_data(c, inode->ei_subvol,
 				POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
 				POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
 		if (ret <= 0)
@@ -2327,7 +2380,7 @@ int bch2_truncate(struct mnt_idmap *idmap,
 	inode_dio_wait(&inode->v);
 	bch2_pagecache_block_get(&inode->ei_pagecache_lock);
 
-	ret = bch2_inode_find_by_inum(c, inode->v.i_ino, &inode_u);
+	ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
 	if (ret)
 		goto err;
 
@@ -2551,6 +2604,18 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 		struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
 		struct bpos atomic_end;
 		unsigned trigger_flags = 0;
+		u32 snapshot;
+
+		bch2_trans_begin(&trans);
+
+		ret = bch2_subvolume_get_snapshot(&trans,
+					inode->ei_subvol, &snapshot);
+		if (ret)
+			continue;
+
+		bch2_btree_iter_set_snapshot(&src, snapshot);
+		bch2_btree_iter_set_snapshot(&dst, snapshot);
+		bch2_btree_iter_set_snapshot(&del, snapshot);
 
 		bch2_trans_begin(&trans);
 
@@ -2671,9 +2736,17 @@ static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		struct bkey_i_reservation reservation;
 		struct bkey_s_c k;
 		unsigned sectors;
+		u32 snapshot;
 
 		bch2_trans_begin(&trans);
 
+		ret = bch2_subvolume_get_snapshot(&trans,
+					inode->ei_subvol, &snapshot);
+		if (ret)
+			goto bkey_err;
+
+		bch2_btree_iter_set_snapshot(&iter, snapshot);
+
 		k = bch2_btree_iter_peek_slot(&iter);
 		if ((ret = bkey_err(k)))
 			goto bkey_err;
@@ -2918,8 +2991,8 @@ loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 	mark_range_unallocated(src, pos_src, pos_src + aligned_len);
 
 	ret = bch2_remap_range(c,
-			       POS(dst->v.i_ino, pos_dst >> 9),
-			       POS(src->v.i_ino, pos_src >> 9),
+			       inode_inum(dst), pos_dst >> 9,
+			       inode_inum(src), pos_src >> 9,
 			       aligned_len >> 9,
 			       &dst->ei_journal_seq,
 			       pos_dst + len, &i_sectors_delta);
@@ -3012,7 +3085,9 @@ static loff_t bch2_seek_data(struct file *file, u64 offset)
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
+	subvol_inum inum = inode_inum(inode);
 	u64 isize, next_data = MAX_LFS_FILESIZE;
+	u32 snapshot;
 	int ret;
 
 	isize = i_size_read(&inode->v);
@@ -3020,9 +3095,15 @@ static loff_t bch2_seek_data(struct file *file, u64 offset)
 		return -ENXIO;
 
 	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
 
 	for_each_btree_key(&trans, iter, BTREE_ID_extents,
-			   POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
+			   SPOS(inode->v.i_ino, offset >> 9, snapshot), 0, k, ret) {
 		if (k.k->p.inode != inode->v.i_ino) {
 			break;
 		} else if (bkey_extent_is_data(k.k)) {
@@ -3032,6 +3113,9 @@ static loff_t bch2_seek_data(struct file *file, u64 offset)
 			break;
 	}
 	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (ret == -EINTR)
+		goto retry;
 
 	ret = bch2_trans_exit(&trans) ?: ret;
 	if (ret)
@@ -3108,7 +3192,9 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset)
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
+	subvol_inum inum = inode_inum(inode);
 	u64 isize, next_hole = MAX_LFS_FILESIZE;
+	u32 snapshot;
 	int ret;
 
 	isize = i_size_read(&inode->v);
@@ -3116,9 +3202,15 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset)
 		return -ENXIO;
 
 	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
 
 	for_each_btree_key(&trans, iter, BTREE_ID_extents,
-			   POS(inode->v.i_ino, offset >> 9),
+			   SPOS(inode->v.i_ino, offset >> 9, snapshot),
 			   BTREE_ITER_SLOTS, k, ret) {
 		if (k.k->p.inode != inode->v.i_ino) {
 			next_hole = bch2_seek_pagecache_hole(&inode->v,
@@ -3136,6 +3228,9 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset)
 		}
 	}
 	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (ret == -EINTR)
+		goto retry;
 
 	ret = bch2_trans_exit(&trans) ?: ret;
 	if (ret)

fs/bcachefs/fs-ioctl.c

@@ -192,7 +192,7 @@ static int bch2_ioc_reinherit_attrs(struct bch_fs *c,
 	char *kname = NULL;
 	struct qstr qstr;
 	int ret = 0;
-	subvol_inum inum = { .subvol = 1 };
+	subvol_inum inum;
 
 	kname = kmalloc(BCH_NAME_MAX + 1, GFP_KERNEL);
 	if (!kname)
@@ -205,10 +205,8 @@ static int bch2_ioc_reinherit_attrs(struct bch_fs *c,
 	qstr.len	= ret;
 	qstr.name	= kname;
 
-	ret = -ENOENT;
-	inum.inum = bch2_dirent_lookup(c, src->v.i_ino, &hash,
-				  &qstr);
-	if (!inum.inum)
+	ret = bch2_dirent_lookup(c, inode_inum(src), &hash, &qstr, &inum);
+	if (ret)
 		goto err1;
 
 	vinode = bch2_vfs_inode_get(c, inum);

fs/bcachefs/fs.c

@@ -150,7 +150,7 @@ int __must_check bch2_write_inode(struct bch_fs *c,
 retry:
 	bch2_trans_begin(&trans);
 
-	ret   = bch2_inode_peek(&trans, &iter, &inode_u, inode->v.i_ino,
+	ret   = bch2_inode_peek(&trans, &iter, &inode_u, inode_inum(inode),
 				BTREE_ITER_INTENT) ?:
 		(set ? set(inode, &inode_u, p) : 0) ?:
 		bch2_inode_write(&trans, &iter, &inode_u) ?:
@@ -256,7 +256,7 @@ struct inode *bch2_vfs_inode_get(struct bch_fs *c, subvol_inum inum)
 	if (!(inode->v.i_state & I_NEW))
 		return &inode->v;
 
-	ret = bch2_inode_find_by_inum(c, inum.inum, &inode_u);
+	ret = bch2_inode_find_by_inum(c, inum, &inode_u);
 	if (ret) {
 		iget_failed(&inode->v);
 		return ERR_PTR(ret);
@@ -271,10 +271,10 @@ struct inode *bch2_vfs_inode_get(struct bch_fs *c, subvol_inum inum)
 	return &inode->v;
 }
 
-static struct bch_inode_info *
+struct bch_inode_info *
 __bch2_create(struct mnt_idmap *idmap,
 	      struct bch_inode_info *dir, struct dentry *dentry,
-	      umode_t mode, dev_t rdev, bool tmpfile)
+	      umode_t mode, dev_t rdev, unsigned flags)
 {
 	struct bch_fs *c = dir->v.i_sb->s_fs_info;
 	struct btree_trans trans;
@@ -303,20 +303,23 @@ __bch2_create(struct mnt_idmap *idmap,
 
 	bch2_inode_init_early(c, &inode_u);
 
-	if (!tmpfile)
+	if (!(flags & BCH_CREATE_TMPFILE))
 		mutex_lock(&dir->ei_update_lock);
 
 	bch2_trans_init(&trans, c, 8,
-			2048 + (!tmpfile ? dentry->d_name.len : 0));
+			2048 + (!(flags & BCH_CREATE_TMPFILE)
+				? dentry->d_name.len : 0));
 retry:
 	bch2_trans_begin(&trans);
 
-	ret   = bch2_create_trans(&trans, dir->v.i_ino, &dir_u, &inode_u,
-				  !tmpfile ? &dentry->d_name : NULL,
+	ret   = bch2_create_trans(&trans,
+				  inode_inum(dir), &dir_u, &inode_u,
+				  !(flags & BCH_CREATE_TMPFILE)
+				  ? &dentry->d_name : NULL,
 				  from_kuid(i_user_ns(&dir->v), current_fsuid()),
 				  from_kgid(i_user_ns(&dir->v), current_fsgid()),
 				  mode, rdev,
-				  default_acl, acl) ?:
+				  default_acl, acl, flags) ?:
 		bch2_quota_acct(c, bch_qid(&inode_u), Q_INO, 1,
 				KEY_TYPE_QUOTA_PREALLOC);
 	if (unlikely(ret))
@@ -332,7 +335,7 @@ __bch2_create(struct mnt_idmap *idmap,
 		goto err_trans;
 	}
 
-	if (!tmpfile) {
+	if (!(flags & BCH_CREATE_TMPFILE)) {
 		bch2_inode_update_after_write(c, dir, &dir_u,
 					      ATTR_MTIME|ATTR_CTIME);
 		journal_seq_copy(c, dir, journal_seq);
@@ -387,7 +390,7 @@ __bch2_create(struct mnt_idmap *idmap,
 	posix_acl_release(acl);
 	return inode;
 err_trans:
-	if (!tmpfile)
+	if (!(flags & BCH_CREATE_TMPFILE))
 		mutex_unlock(&dir->ei_update_lock);
 
 	bch2_trans_exit(&trans);
@@ -407,11 +410,12 @@ static struct dentry *bch2_lookup(struct inode *vdir, struct dentry *dentry,
 	struct bch_hash_info hash = bch2_hash_info_init(c, &dir->ei_inode);
 	struct inode *vinode = NULL;
 	subvol_inum inum = { .subvol = 1 };
+	int ret;
 
-	inum.inum = bch2_dirent_lookup(c, dir->v.i_ino, &hash,
-				  &dentry->d_name);
+	ret = bch2_dirent_lookup(c, inode_inum(dir), &hash,
+				 &dentry->d_name, &inum);
 
-	if (inum.inum)
+	if (!ret)
 		vinode = bch2_vfs_inode_get(c, inum);
 
 	return d_splice_alias(vinode, dentry);
@@ -422,7 +426,7 @@ static int bch2_mknod(struct mnt_idmap *idmap,
 		      umode_t mode, dev_t rdev)
 {
 	struct bch_inode_info *inode =
-		__bch2_create(idmap, to_bch_ei(vdir), dentry, mode, rdev, false);
+		__bch2_create(idmap, to_bch_ei(vdir), dentry, mode, rdev, 0);
 
 	if (IS_ERR(inode))
 		return PTR_ERR(inode);
@@ -452,8 +456,8 @@ static int __bch2_link(struct bch_fs *c,
 
 	ret = __bch2_trans_do(&trans, NULL, &inode->ei_journal_seq, 0,
 			bch2_link_trans(&trans,
-					dir->v.i_ino,
-					inode->v.i_ino, &dir_u, &inode_u,
+					inode_inum(dir),   &dir_u,
+					inode_inum(inode), &inode_u,
 					&dentry->d_name));
 
 	if (likely(!ret)) {
@@ -504,7 +508,7 @@ static int bch2_unlink(struct inode *vdir, struct dentry *dentry)
 	ret = __bch2_trans_do(&trans, NULL, &dir->ei_journal_seq,
 			      BTREE_INSERT_NOFAIL,
 			bch2_unlink_trans(&trans,
-					  dir->v.i_ino, &dir_u,
+					  inode_inum(dir), &dir_u,
 					  &inode_u, &dentry->d_name));
 
 	if (likely(!ret)) {
@@ -531,7 +535,8 @@ static int bch2_symlink(struct mnt_idmap *idmap,
 	struct bch_inode_info *dir = to_bch_ei(vdir), *inode;
 	int ret;
 
-	inode = __bch2_create(idmap, dir, dentry, S_IFLNK|S_IRWXUGO, 0, true);
+	inode = __bch2_create(idmap, dir, dentry, S_IFLNK|S_IRWXUGO, 0,
+			      BCH_CREATE_TMPFILE);
 	if (unlikely(IS_ERR(inode)))
 		return PTR_ERR(inode);
 
@@ -624,8 +629,8 @@ static int bch2_rename2(struct mnt_idmap *idmap,
 
 	ret = __bch2_trans_do(&trans, NULL, &journal_seq, 0,
 			bch2_rename_trans(&trans,
-					  src_dir->v.i_ino, &src_dir_u,
-					  dst_dir->v.i_ino, &dst_dir_u,
+					  inode_inum(src_dir), &src_dir_u,
+					  inode_inum(dst_dir), &dst_dir_u,
 					  &src_inode_u,
 					  &dst_inode_u,
 					  &src_dentry->d_name,
@@ -748,7 +753,7 @@ int bch2_setattr_nonsize(struct mnt_idmap *idmap,
 	kfree(acl);
 	acl = NULL;
 
-	ret = bch2_inode_peek(&trans, &inode_iter, &inode_u, inode->v.i_ino,
+	ret = bch2_inode_peek(&trans, &inode_iter, &inode_u, inode_inum(inode),
 			      BTREE_ITER_INTENT);
 	if (ret)
 		goto btree_err;
@@ -756,7 +761,8 @@ int bch2_setattr_nonsize(struct mnt_idmap *idmap,
 	bch2_setattr_copy(idmap, inode, &inode_u, attr);
 
 	if (attr->ia_valid & ATTR_MODE) {
-		ret = bch2_acl_chmod(&trans, &inode_u, inode_u.bi_mode, &acl);
+		ret = bch2_acl_chmod(&trans, inode_inum(inode), &inode_u,
+				     inode_u.bi_mode, &acl);
 		if (ret)
 			goto btree_err;
 	}
@@ -848,7 +854,8 @@ static int bch2_tmpfile(struct mnt_idmap *idmap,
 {
 	struct bch_inode_info *inode =
 		__bch2_create(idmap, to_bch_ei(vdir),
-			      file->f_path.dentry, mode, 0, true);
+			      file->f_path.dentry, mode, 0,
+			      BCH_CREATE_TMPFILE);
 
 	if (IS_ERR(inode))
 		return PTR_ERR(inode);
@@ -923,6 +930,7 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 	struct bpos end = POS(ei->v.i_ino, (start + len) >> 9);
 	unsigned offset_into_extent, sectors;
 	bool have_extent = false;
+	u32 snapshot;
 	int ret = 0;
 
 	ret = fiemap_prep(&ei->v, info, start, &len, FIEMAP_FLAG_SYNC);
@@ -932,15 +940,21 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 	if (start + len < start)
 		return -EINVAL;
 
+	start >>= 9;
+
 	bch2_bkey_buf_init(&cur);
 	bch2_bkey_buf_init(&prev);
 	bch2_trans_init(&trans, c, 0, 0);
-
-	bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
-			     POS(ei->v.i_ino, start >> 9), 0);
 retry:
 	bch2_trans_begin(&trans);
 
+	ret = bch2_subvolume_get_snapshot(&trans, ei->ei_subvol, &snapshot);
+	if (ret)
+		goto err;
+
+	bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
+			     SPOS(ei->v.i_ino, start, snapshot), 0);
+
 	while ((k = bch2_btree_iter_peek(&iter)).k &&
 	       !(ret = bkey_err(k)) &&
 	       bkey_cmp(iter.pos, end) < 0) {
@@ -989,7 +1003,9 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 		bch2_btree_iter_set_pos(&iter,
 			POS(iter.pos.inode, iter.pos.offset + sectors));
 	}
-
+	start = iter.pos.offset;
+	bch2_trans_iter_exit(&trans, &iter);
+err:
 	if (ret == -EINTR)
 		goto retry;
 
@@ -997,7 +1013,6 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 		ret = bch2_fill_extent(c, info, bkey_i_to_s_c(prev.k),
 				       FIEMAP_EXTENT_LAST);
 
-	bch2_trans_iter_exit(&trans, &iter);
 	ret = bch2_trans_exit(&trans) ?: ret;
 	bch2_bkey_buf_exit(&cur, c);
 	bch2_bkey_buf_exit(&prev, c);
@@ -1034,7 +1049,7 @@ static int bch2_vfs_readdir(struct file *file, struct dir_context *ctx)
 	if (!dir_emit_dots(file, ctx))
 		return 0;
 
-	return bch2_readdir(c, inode->v.i_ino, ctx);
+	return bch2_readdir(c, inode_inum(inode), ctx);
 }
 
 static const struct file_operations bch_file_operations = {
@@ -1290,7 +1305,7 @@ static void bch2_evict_inode(struct inode *vinode)
 				KEY_TYPE_QUOTA_WARN);
 		bch2_quota_acct(c, inode->ei_qid, Q_INO, -1,
 				KEY_TYPE_QUOTA_WARN);
-		bch2_inode_rm(c, inode->v.i_ino, true);
+		bch2_inode_rm(c, inode_inum(inode), true);
 	}
 }
 

fs/bcachefs/fs.h

@@ -144,6 +144,10 @@ struct bch_inode_unpacked;
 
 #ifndef NO_BCACHEFS_FS
 
+struct bch_inode_info *
+__bch2_create(struct mnt_idmap *, struct bch_inode_info *,
+	      struct dentry *, umode_t, dev_t, unsigned);
+
 int bch2_fs_quota_transfer(struct bch_fs *,
 			   struct bch_inode_info *,
 			   struct bch_qid,

fs/bcachefs/fsck.c

@@ -858,7 +858,10 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 	d = bkey_s_c_to_dirent(k);
 	d_inum = le64_to_cpu(d.v->d_inum);
 
-	ret = bch2_dirent_read_target(trans, d, &d_inum);
+	ret = __bch2_dirent_read_target(&trans, d,
+					&target_subvol,
+					&target_snapshot,
+					&target_inum);
 	if (ret && ret != -ENOENT)
 		return ret;
 

fs/bcachefs/inode.c

@@ -6,6 +6,7 @@
 #include "btree_update.h"
 #include "error.h"
 #include "extents.h"
+#include "extent_update.h"
 #include "inode.h"
 #include "str_hash.h"
 #include "subvolume.h"
@@ -296,15 +297,21 @@ int bch2_inode_unpack(struct bkey_s_c_inode inode,
 int bch2_inode_peek(struct btree_trans *trans,
 		    struct btree_iter *iter,
 		    struct bch_inode_unpacked *inode,
-		    u64 inum, unsigned flags)
+		    subvol_inum inum, unsigned flags)
 {
 	struct bkey_s_c k;
+	u32 snapshot;
 	int ret;
 
 	if (trans->c->opts.inodes_use_key_cache)
 		flags |= BTREE_ITER_CACHED;
 
-	bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, inum), flags);
+	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+	if (ret)
+		return ret;
+
+	bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
+			     SPOS(0, inum.inum, snapshot), flags);
 	k = bch2_btree_iter_peek_slot(iter);
 	ret = bkey_err(k);
 	if (ret)
@@ -486,6 +493,9 @@ static inline u32 bkey_generation(struct bkey_s_c k)
 	}
 }
 
+/*
+ * This just finds an empty slot:
+ */
 int bch2_inode_create(struct btree_trans *trans,
 		      struct btree_iter *iter,
 		      struct bch_inode_unpacked *inode_u,
@@ -585,16 +595,74 @@ int bch2_inode_create(struct btree_trans *trans,
 	return 0;
 }
 
-int bch2_inode_rm(struct bch_fs *c, u64 inode_nr, bool cached)
+static int bch2_inode_delete_keys(struct btree_trans *trans,
+				  subvol_inum inum, enum btree_id id)
+{
+	u64 offset = 0;
+	int ret = 0;
+
+	while (!ret || ret == -EINTR) {
+		struct btree_iter iter;
+		struct bkey_s_c k;
+		struct bkey_i delete;
+		u32 snapshot;
+
+		bch2_trans_begin(trans);
+
+		ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+		if (ret)
+			continue;
+
+		bch2_trans_iter_init(trans, &iter, id,
+				     SPOS(inum.inum, offset, snapshot),
+				     BTREE_ITER_INTENT);
+		k = bch2_btree_iter_peek(&iter);
+
+		if (!k.k || iter.pos.inode != inum.inum) {
+			bch2_trans_iter_exit(trans, &iter);
+			break;
+		}
+
+		ret = bkey_err(k);
+		if (ret)
+			goto err;
+
+		bkey_init(&delete.k);
+		delete.k.p = iter.pos;
+
+		if (btree_node_type_is_extents(iter.btree_id)) {
+			unsigned max_sectors =
+				min_t(u64, U64_MAX - iter.pos.offset,
+				      KEY_SIZE_MAX & (~0 << trans->c->block_bits));
+
+			/* create the biggest key we can */
+			bch2_key_resize(&delete.k, max_sectors);
+
+			ret = bch2_extent_trim_atomic(trans, &iter, &delete);
+			if (ret)
+				goto err;
+		}
+
+		ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
+		      bch2_trans_commit(trans, NULL, NULL,
+					BTREE_INSERT_NOFAIL);
+err:
+		offset = iter.pos.offset;
+		bch2_trans_iter_exit(trans, &iter);
+	}
+
+	return ret;
+}
+
+int bch2_inode_rm(struct bch_fs *c, subvol_inum inum, bool cached)
 {
 	struct btree_trans trans;
 	struct btree_iter iter = { NULL };
 	struct bkey_i_inode_generation delete;
-	struct bpos start = POS(inode_nr, 0);
-	struct bpos end = POS(inode_nr + 1, 0);
 	struct bch_inode_unpacked inode_u;
 	struct bkey_s_c k;
 	unsigned iter_flags = BTREE_ITER_INTENT;
+	u32 snapshot;
 	int ret;
 
 	if (cached && c->opts.inodes_use_key_cache)
@@ -610,19 +678,20 @@ int bch2_inode_rm(struct bch_fs *c, u64 inode_nr, bool cached)
 	 * XXX: the dirent could ideally would delete whiteouts when they're no
 	 * longer needed
 	 */
-	ret   = bch2_btree_delete_range_trans(&trans, BTREE_ID_extents,
-					      start, end, NULL) ?:
-		bch2_btree_delete_range_trans(&trans, BTREE_ID_xattrs,
-					      start, end, NULL) ?:
-		bch2_btree_delete_range_trans(&trans, BTREE_ID_dirents,
-					      start, end, NULL);
+	ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
+		bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
+		bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 	if (ret)
 		goto err;
 retry:
 	bch2_trans_begin(&trans);
 
+	ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
+
 	bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
-			     POS(0, inode_nr), iter_flags);
+			     SPOS(0, inum.inum, snapshot), iter_flags);
 	k = bch2_btree_iter_peek_slot(&iter);
 
 	ret = bkey_err(k);
@@ -632,7 +701,7 @@ int bch2_inode_rm(struct bch_fs *c, u64 inode_nr, bool cached)
 	if (k.k->type != KEY_TYPE_inode) {
 		bch2_fs_inconsistent(trans.c,
 				     "inode %llu not found when deleting",
-				     inode_nr);
+				     inum.inum);
 		ret = -EIO;
 		goto err;
 	}
@@ -662,20 +731,22 @@ int bch2_inode_rm(struct bch_fs *c, u64 inode_nr, bool cached)
 	return ret;
 }
 
-static int bch2_inode_find_by_inum_trans(struct btree_trans *trans, u64 inode_nr,
+static int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
+					 subvol_inum inum,
 					 struct bch_inode_unpacked *inode)
 {
-	struct btree_iter iter = { NULL };
+	struct btree_iter iter;
 	int ret;
 
-	ret = bch2_inode_peek(trans, &iter, inode, inode_nr, 0);
-	bch2_trans_iter_exit(trans, &iter);
+	ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
+	if (!ret)
+		bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
 
-int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
+int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 			    struct bch_inode_unpacked *inode)
 {
 	return bch2_trans_do(c, NULL, NULL, 0,
-		bch2_inode_find_by_inum_trans(&trans, inode_nr, inode));
+		bch2_inode_find_by_inum_trans(&trans, inum, inode));
 }

fs/bcachefs/inode.h

@@ -58,7 +58,7 @@ int bch2_inode_unpack(struct bkey_s_c_inode, struct bch_inode_unpacked *);
 void bch2_inode_unpacked_to_text(struct printbuf *, struct bch_inode_unpacked *);
 
 int bch2_inode_peek(struct btree_trans *, struct btree_iter *,
-		    struct bch_inode_unpacked *, u64, unsigned);
+		    struct bch_inode_unpacked *, subvol_inum, unsigned);
 int bch2_inode_write(struct btree_trans *, struct btree_iter *,
 		     struct bch_inode_unpacked *);
 
@@ -74,9 +74,10 @@ void bch2_inode_init(struct bch_fs *, struct bch_inode_unpacked *,
 int bch2_inode_create(struct btree_trans *, struct btree_iter *,
 		      struct bch_inode_unpacked *, u32, u64);
 
-int bch2_inode_rm(struct bch_fs *, u64, bool);
+int bch2_inode_rm(struct bch_fs *, subvol_inum, bool);
 
-int bch2_inode_find_by_inum(struct bch_fs *, u64, struct bch_inode_unpacked *);
+int bch2_inode_find_by_inum(struct bch_fs *, subvol_inum,
+			    struct bch_inode_unpacked *);
 
 static inline struct bch_io_opts bch2_inode_opts_get(struct bch_inode_unpacked *inode)
 {

fs/bcachefs/io.c

@@ -325,7 +325,10 @@ int bch2_extent_update(struct btree_trans *trans,
 		struct bch_inode_unpacked inode_u;
 
 		ret = bch2_inode_peek(trans, &inode_iter, &inode_u,
-				k->k.p.inode, BTREE_ITER_INTENT);
+				      (subvol_inum) {
+				      .subvol = BCACHEFS_ROOT_SUBVOL,
+				      .inum = k->k.p.inode,
+				      }, BTREE_ITER_INTENT);
 		if (ret)
 			return ret;
 

fs/bcachefs/move.c

@@ -581,7 +581,8 @@ static int __bch2_move_data(struct bch_fs *c,
 	stats->pos	= start;
 
 	bch2_trans_iter_init(&trans, &iter, btree_id, start,
-			     BTREE_ITER_PREFETCH);
+			     BTREE_ITER_PREFETCH|
+			     BTREE_ITER_ALL_SNAPSHOTS);
 
 	if (rate)
 		bch2_ratelimit_reset(rate);

fs/bcachefs/recovery.c

@@ -1480,11 +1480,12 @@ int bch2_fs_initialize(struct bch_fs *c)
 
 	err = "error creating lost+found";
 	ret = bch2_trans_do(c, NULL, NULL, 0,
-		bch2_create_trans(&trans, BCACHEFS_ROOT_INO,
+		bch2_create_trans(&trans,
+				  BCACHEFS_ROOT_SUBVOL_INUM,
 				  &root_inode, &lostfound_inode,
 				  &lostfound,
 				  0, 0, S_IFDIR|0700, 0,
-				  NULL, NULL));
+				  NULL, NULL, 0));
 	if (ret) {
 		bch_err(c, "error creating lost+found");
 		goto err;

fs/bcachefs/reflink.c

@@ -7,6 +7,7 @@
 #include "inode.h"
 #include "io.h"
 #include "reflink.h"
+#include "subvolume.h"
 
 #include <linux/sched/signal.h>
 
@@ -197,7 +198,8 @@ static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 }
 
 s64 bch2_remap_range(struct bch_fs *c,
-		     struct bpos dst_start, struct bpos src_start,
+		     subvol_inum dst_inum, u64 dst_offset,
+		     subvol_inum src_inum, u64 src_offset,
 		     u64 remap_sectors, u64 *journal_seq,
 		     u64 new_i_size, s64 *i_sectors_delta)
 {
@@ -205,6 +207,8 @@ s64 bch2_remap_range(struct bch_fs *c,
 	struct btree_iter dst_iter, src_iter;
 	struct bkey_s_c src_k;
 	struct bkey_buf new_dst, new_src;
+	struct bpos dst_start = POS(dst_inum.inum, dst_offset);
+	struct bpos src_start = POS(src_inum.inum, src_offset);
 	struct bpos dst_end = dst_start, src_end = src_start;
 	struct bpos src_want;
 	u64 dst_done;
@@ -238,6 +242,16 @@ s64 bch2_remap_range(struct bch_fs *c,
 			break;
 		}
 
+		ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
+						  &src_iter.snapshot);
+		if (ret)
+			continue;
+
+		ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
+						  &dst_iter.snapshot);
+		if (ret)
+			continue;
+
 		dst_done = dst_iter.pos.offset - dst_start.offset;
 		src_want = POS(src_start.inode, src_start.offset + dst_done);
 		bch2_btree_iter_set_pos(&src_iter, src_want);
@@ -311,7 +325,7 @@ s64 bch2_remap_range(struct bch_fs *c,
 		bch2_trans_begin(&trans);
 
 		ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
-				dst_start.inode, BTREE_ITER_INTENT);
+				       dst_inum, BTREE_ITER_INTENT);
 
 		if (!ret2 &&
 		    inode_u.bi_size < new_i_size) {

fs/bcachefs/reflink.h

@@ -57,7 +57,7 @@ static inline __le64 *bkey_refcount(struct bkey_i *k)
 	}
 }
 
-s64 bch2_remap_range(struct bch_fs *, struct bpos, struct bpos,
-		     u64, u64 *, u64, s64 *);
+s64 bch2_remap_range(struct bch_fs *, subvol_inum, u64,
+		     subvol_inum, u64, u64, u64 *, u64, s64 *);
 
 #endif /* _BCACHEFS_REFLINK_H */

fs/bcachefs/str_hash.h

@@ -8,6 +8,7 @@
 #include "error.h"
 #include "inode.h"
 #include "siphash.h"
+#include "subvolume.h"
 #include "super.h"
 
 #include <linux/crc32c.h>
@@ -144,16 +145,21 @@ bch2_hash_lookup(struct btree_trans *trans,
 		 struct btree_iter *iter,
 		 const struct bch_hash_desc desc,
 		 const struct bch_hash_info *info,
-		 u64 inode, const void *key,
+		 subvol_inum inum, const void *key,
 		 unsigned flags)
 {
 	struct bkey_s_c k;
+	u32 snapshot;
 	int ret;
 
+	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+	if (ret)
+		return ret;
+
 	for_each_btree_key(trans, *iter, desc.btree_id,
-			   POS(inode, desc.hash_key(info, key)),
+			   SPOS(inum.inum, desc.hash_key(info, key), snapshot),
 			   BTREE_ITER_SLOTS|flags, k, ret) {
-		if (iter->pos.inode != inode)
+		if (iter->pos.inode != inum.inum)
 			break;
 
 		if (k.k->type == desc.key_type) {
@@ -176,15 +182,20 @@ bch2_hash_hole(struct btree_trans *trans,
 	       struct btree_iter *iter,
 	       const struct bch_hash_desc desc,
 	       const struct bch_hash_info *info,
-	       u64 inode, const void *key)
+	       subvol_inum inum, const void *key)
 {
 	struct bkey_s_c k;
+	u32 snapshot;
 	int ret;
 
+	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+	if (ret)
+		return ret;
+
 	for_each_btree_key(trans, *iter, desc.btree_id,
-			   POS(inode, desc.hash_key(info, key)),
+			   SPOS(inum.inum, desc.hash_key(info, key), snapshot),
 			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
-		if (iter->pos.inode != inode)
+		if (iter->pos.inode != inum.inum)
 			break;
 
 		if (k.k->type != desc.key_type)
@@ -229,17 +240,25 @@ static __always_inline
 int bch2_hash_set(struct btree_trans *trans,
 		  const struct bch_hash_desc desc,
 		  const struct bch_hash_info *info,
-		  u64 inode, struct bkey_i *insert, int flags)
+		  subvol_inum inum,
+		  struct bkey_i *insert, int flags)
 {
 	struct btree_iter iter, slot = { NULL };
 	struct bkey_s_c k;
 	bool found = false;
+	u32 snapshot;
 	int ret;
 
+	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+	if (ret)
+		return ret;
+
 	for_each_btree_key(trans, iter, desc.btree_id,
-			   POS(inode, desc.hash_bkey(info, bkey_i_to_s_c(insert))),
+			   SPOS(inum.inum,
+				desc.hash_bkey(info, bkey_i_to_s_c(insert)),
+				snapshot),
 			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
-		if (iter.pos.inode != inode)
+		if (iter.pos.inode != inum.inum)
 			break;
 
 		if (k.k->type == desc.key_type) {
@@ -313,12 +332,12 @@ static __always_inline
 int bch2_hash_delete(struct btree_trans *trans,
 		     const struct bch_hash_desc desc,
 		     const struct bch_hash_info *info,
-		     u64 inode, const void *key)
+		     subvol_inum inum, const void *key)
 {
 	struct btree_iter iter;
 	int ret;
 
-	ret = bch2_hash_lookup(trans, &iter, desc, info, inode, key,
+	ret = bch2_hash_lookup(trans, &iter, desc, info, inum, key,
 				BTREE_ITER_INTENT);
 	if (ret)
 		return ret;

fs/bcachefs/xattr.c

@@ -128,7 +128,7 @@ static int bch2_xattr_get_trans(struct btree_trans *trans, struct bch_inode_info
 	int ret;
 
 	ret = bch2_hash_lookup(trans, &iter, bch2_xattr_hash_desc, &hash,
-			       inode->v.i_ino,
+			       inode_inum(inode),
 			       &X_SEARCH(type, name, strlen(name)),
 			       0);
 	if (ret)
@@ -160,7 +160,7 @@ int bch2_xattr_get(struct bch_fs *c, struct bch_inode_info *inode,
 		bch2_xattr_get_trans(&trans, inode, name, buffer, size, type));
 }
 
-int bch2_xattr_set(struct btree_trans *trans, u64 inum,
+int bch2_xattr_set(struct btree_trans *trans, subvol_inum inum,
 		   const struct bch_hash_info *hash_info,
 		   const char *name, const void *value, size_t size,
 		   int type, int flags)
@@ -282,13 +282,21 @@ ssize_t bch2_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct xattr_buf buf = { .buf = buffer, .len = buffer_size };
-	u64 inum = dentry->d_inode->i_ino;
+	u64 offset = 0, inum = inode->ei_inode.bi_inum;
+	u32 snapshot;
 	int ret;
 
 	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+	iter = (struct btree_iter) { NULL };
+
+	ret = bch2_subvolume_get_snapshot(&trans, inode->ei_subvol, &snapshot);
+	if (ret)
+		goto err;
 
 	for_each_btree_key(&trans, iter, BTREE_ID_xattrs,
-			   POS(inum, 0), 0, k, ret) {
+			   SPOS(inum, offset, snapshot), 0, k, ret) {
 		BUG_ON(k.k->p.inode < inum);
 
 		if (k.k->p.inode > inum)
@@ -301,7 +309,12 @@ ssize_t bch2_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 		if (ret)
 			break;
 	}
+
+	offset = iter.pos.offset;
 	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (ret == -EINTR)
+		goto retry;
 
 	ret = bch2_trans_exit(&trans) ?: ret;
 
@@ -340,7 +353,7 @@ static int bch2_xattr_set_handler(const struct xattr_handler *handler,
 	struct bch_hash_info hash = bch2_hash_info_init(c, &inode->ei_inode);
 
 	return bch2_trans_do(c, NULL, &inode->ei_journal_seq, 0,
-			bch2_xattr_set(&trans, inode->v.i_ino, &hash,
+			bch2_xattr_set(&trans, inode_inum(inode), &hash,
 				       name, value, size,
 				       handler->flags, flags));
 }

fs/bcachefs/xattr.h

@@ -39,7 +39,8 @@ struct bch_inode_info;
 int bch2_xattr_get(struct bch_fs *, struct bch_inode_info *,
 		  const char *, void *, size_t, int);
 
-int bch2_xattr_set(struct btree_trans *, u64, const struct bch_hash_info *,
+int bch2_xattr_set(struct btree_trans *, subvol_inum,
+		   const struct bch_hash_info *,
 		   const char *, const void *, size_t, int, int);
 
 ssize_t bch2_xattr_list(struct dentry *, char *, size_t);