btrfs: fix race when picking most recent mod log operation for an old root

Commit dbcc7d57bf ("btrfs: fix race when cloning extent buffer during
rewind of an old root"), fixed a race when we need to rewind the extent
buffer of an old root. It was caused by picking a new mod log operation
for the extent buffer while getting a cloned extent buffer with an outdated
number of items (off by -1), because we cloned the extent buffer without
locking it first.

However there is still another similar race, but in the opposite direction.
The cloned extent buffer has a number of items that does not match the
number of tree mod log operations that are going to be replayed. This is
because right after we got the last (most recent) tree mod log operation to
replay and before locking and cloning the extent buffer, another task adds
a new pointer to the extent buffer, which results in adding a new tree mod
log operation and incrementing the number of items in the extent buffer.
So after cloning we have mismatch between the number of items in the extent
buffer and the number of mod log operations we are going to apply to it.
This results in hitting a BUG_ON() that produces the following stack trace:

   ------------[ cut here ]------------
   kernel BUG at fs/btrfs/tree-mod-log.c:675!
   invalid opcode: 0000 [#1] SMP KASAN PTI
   CPU: 3 PID: 4811 Comm: crawl_1215 Tainted: G        W         5.12.0-7d1efdf501f8-misc-next+ #99
   Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
   RIP: 0010:tree_mod_log_rewind+0x3b1/0x3c0
   Code: 05 48 8d 74 10 (...)
   RSP: 0018:ffffc90001027090 EFLAGS: 00010293
   RAX: 0000000000000000 RBX: ffff8880a8514600 RCX: ffffffffaa9e59b6
   RDX: 0000000000000007 RSI: dffffc0000000000 RDI: ffff8880a851462c
   RBP: ffffc900010270e0 R08: 00000000000000c0 R09: ffffed1004333417
   R10: ffff88802199a0b7 R11: ffffed1004333416 R12: 000000000000000e
   R13: ffff888135af8748 R14: ffff88818766ff00 R15: ffff8880a851462c
   FS:  00007f29acf62700(0000) GS:ffff8881f2200000(0000) knlGS:0000000000000000
   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
   CR2: 00007f0e6013f718 CR3: 000000010d42e003 CR4: 0000000000170ee0
   Call Trace:
    btrfs_get_old_root+0x16a/0x5c0
    ? lock_downgrade+0x400/0x400
    btrfs_search_old_slot+0x192/0x520
    ? btrfs_search_slot+0x1090/0x1090
    ? free_extent_buffer.part.61+0xd7/0x140
    ? free_extent_buffer+0x13/0x20
    resolve_indirect_refs+0x3e9/0xfc0
    ? lock_downgrade+0x400/0x400
    ? __kasan_check_read+0x11/0x20
    ? add_prelim_ref.part.11+0x150/0x150
    ? lock_downgrade+0x400/0x400
    ? __kasan_check_read+0x11/0x20
    ? lock_acquired+0xbb/0x620
    ? __kasan_check_write+0x14/0x20
    ? do_raw_spin_unlock+0xa8/0x140
    ? rb_insert_color+0x340/0x360
    ? prelim_ref_insert+0x12d/0x430
    find_parent_nodes+0x5c3/0x1830
    ? stack_trace_save+0x87/0xb0
    ? resolve_indirect_refs+0xfc0/0xfc0
    ? fs_reclaim_acquire+0x67/0xf0
    ? __kasan_check_read+0x11/0x20
    ? lockdep_hardirqs_on_prepare+0x210/0x210
    ? fs_reclaim_acquire+0x67/0xf0
    ? __kasan_check_read+0x11/0x20
    ? ___might_sleep+0x10f/0x1e0
    ? __kasan_kmalloc+0x9d/0xd0
    ? trace_hardirqs_on+0x55/0x120
    btrfs_find_all_roots_safe+0x142/0x1e0
    ? find_parent_nodes+0x1830/0x1830
    ? trace_hardirqs_on+0x55/0x120
    ? ulist_free+0x1f/0x30
    ? btrfs_inode_flags_to_xflags+0x50/0x50
    iterate_extent_inodes+0x20e/0x580
    ? tree_backref_for_extent+0x230/0x230
    ? release_extent_buffer+0x225/0x280
    ? read_extent_buffer+0xdd/0x110
    ? lock_downgrade+0x400/0x400
    ? __kasan_check_read+0x11/0x20
    ? lock_acquired+0xbb/0x620
    ? __kasan_check_write+0x14/0x20
    ? do_raw_spin_unlock+0xa8/0x140
    ? _raw_spin_unlock+0x22/0x30
    ? release_extent_buffer+0x225/0x280
    iterate_inodes_from_logical+0x129/0x170
    ? iterate_inodes_from_logical+0x129/0x170
    ? btrfs_inode_flags_to_xflags+0x50/0x50
    ? iterate_extent_inodes+0x580/0x580
    ? __vmalloc_node+0x92/0xb0
    ? init_data_container+0x34/0xb0
    ? init_data_container+0x34/0xb0
    ? kvmalloc_node+0x60/0x80
    btrfs_ioctl_logical_to_ino+0x158/0x230
    btrfs_ioctl+0x2038/0x4360
    ? __kasan_check_write+0x14/0x20
    ? mmput+0x3b/0x220
    ? btrfs_ioctl_get_supported_features+0x30/0x30
    ? __kasan_check_read+0x11/0x20
    ? __kasan_check_read+0x11/0x20
    ? lock_release+0xc8/0x650
    ? __might_fault+0x64/0xd0
    ? __kasan_check_read+0x11/0x20
    ? lock_downgrade+0x400/0x400
    ? lockdep_hardirqs_on_prepare+0x210/0x210
    ? lockdep_hardirqs_on_prepare+0x13/0x210
    ? _raw_spin_unlock_irqrestore+0x51/0x63
    ? __kasan_check_read+0x11/0x20
    ? do_vfs_ioctl+0xfc/0x9d0
    ? ioctl_file_clone+0xe0/0xe0
    ? lock_downgrade+0x400/0x400
    ? lockdep_hardirqs_on_prepare+0x210/0x210
    ? __kasan_check_read+0x11/0x20
    ? lock_release+0xc8/0x650
    ? __task_pid_nr_ns+0xd3/0x250
    ? __kasan_check_read+0x11/0x20
    ? __fget_files+0x160/0x230
    ? __fget_light+0xf2/0x110
    __x64_sys_ioctl+0xc3/0x100
    do_syscall_64+0x37/0x80
    entry_SYSCALL_64_after_hwframe+0x44/0xae
   RIP: 0033:0x7f29ae85b427
   Code: 00 00 90 48 8b (...)
   RSP: 002b:00007f29acf5fcf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
   RAX: ffffffffffffffda RBX: 00007f29acf5ff40 RCX: 00007f29ae85b427
   RDX: 00007f29acf5ff48 RSI: 00000000c038943b RDI: 0000000000000003
   RBP: 0000000001000000 R08: 0000000000000000 R09: 00007f29acf60120
   R10: 00005640d5fc7b00 R11: 0000000000000246 R12: 0000000000000003
   R13: 00007f29acf5ff48 R14: 00007f29acf5ff40 R15: 00007f29acf5fef8
   Modules linked in:
   ---[ end trace 85e5fce078dfbe04 ]---

  (gdb) l *(tree_mod_log_rewind+0x3b1)
  0xffffffff819e5b21 is in tree_mod_log_rewind (fs/btrfs/tree-mod-log.c:675).
  670                      * the modification. As we're going backwards, we do the
  671                      * opposite of each operation here.
  672                      */
  673                     switch (tm->op) {
  674                     case BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING:
  675                             BUG_ON(tm->slot < n);
  676                             fallthrough;
  677                     case BTRFS_MOD_LOG_KEY_REMOVE_WHILE_MOVING:
  678                     case BTRFS_MOD_LOG_KEY_REMOVE:
  679                             btrfs_set_node_key(eb, &tm->key, tm->slot);
  (gdb) quit

The following steps explain in more detail how it happens:

1) We have one tree mod log user (through fiemap or the logical ino ioctl),
   with a sequence number of 1, so we have fs_info->tree_mod_seq == 1.
   This is task A;

2) Another task is at ctree.c:balance_level() and we have eb X currently as
   the root of the tree, and we promote its single child, eb Y, as the new
   root.

   Then, at ctree.c:balance_level(), we call:

      ret = btrfs_tree_mod_log_insert_root(root->node, child, true);

3) At btrfs_tree_mod_log_insert_root() we create a tree mod log operation
   of type BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING, with a ->logical field
   pointing to ebX->start. We only have one item in eb X, so we create
   only one tree mod log operation, and store in the "tm_list" array;

4) Then, still at btrfs_tree_mod_log_insert_root(), we create a tree mod
   log element of operation type BTRFS_MOD_LOG_ROOT_REPLACE, ->logical set
   to ebY->start, ->old_root.logical set to ebX->start, ->old_root.level
   set to the level of eb X and ->generation set to the generation of eb X;

5) Then btrfs_tree_mod_log_insert_root() calls tree_mod_log_free_eb() with
   "tm_list" as argument. After that, tree_mod_log_free_eb() calls
   tree_mod_log_insert(). This inserts the mod log operation of type
   BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING from step 3 into the rbtree
   with a sequence number of 2 (and fs_info->tree_mod_seq set to 2);

6) Then, after inserting the "tm_list" single element into the tree mod
   log rbtree, the BTRFS_MOD_LOG_ROOT_REPLACE element is inserted, which
   gets the sequence number 3 (and fs_info->tree_mod_seq set to 3);

7) Back to ctree.c:balance_level(), we free eb X by calling
   btrfs_free_tree_block() on it. Because eb X was created in the current
   transaction, has no other references and writeback did not happen for
   it, we add it back to the free space cache/tree;

8) Later some other task B allocates the metadata extent from eb X, since
   it is marked as free space in the space cache/tree, and uses it as a
   node for some other btree;

9) The tree mod log user task calls btrfs_search_old_slot(), which calls
   btrfs_get_old_root(), and finally that calls tree_mod_log_oldest_root()
   with time_seq == 1 and eb_root == eb Y;

10) The first iteration of the while loop finds the tree mod log element
    with sequence number 3, for the logical address of eb Y and of type
    BTRFS_MOD_LOG_ROOT_REPLACE;

11) Because the operation type is BTRFS_MOD_LOG_ROOT_REPLACE, we don't
    break out of the loop, and set root_logical to point to
    tm->old_root.logical, which corresponds to the logical address of
    eb X;

12) On the next iteration of the while loop, the call to
    tree_mod_log_search_oldest() returns the smallest tree mod log element
    for the logical address of eb X, which has a sequence number of 2, an
    operation type of BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING and
    corresponds to the old slot 0 of eb X (eb X had only 1 item in it
    before being freed at step 7);

13) We then break out of the while loop and return the tree mod log
    operation of type BTRFS_MOD_LOG_ROOT_REPLACE (eb Y), and not the one
    for slot 0 of eb X, to btrfs_get_old_root();

14) At btrfs_get_old_root(), we process the BTRFS_MOD_LOG_ROOT_REPLACE
    operation and set "logical" to the logical address of eb X, which was
    the old root. We then call tree_mod_log_search() passing it the logical
    address of eb X and time_seq == 1;

15) But before calling tree_mod_log_search(), task B locks eb X, adds a
    key to eb X, which results in adding a tree mod log operation of type
    BTRFS_MOD_LOG_KEY_ADD, with a sequence number of 4, to the tree mod
    log, and increments the number of items in eb X from 0 to 1.
    Now fs_info->tree_mod_seq has a value of 4;

16) Task A then calls tree_mod_log_search(), which returns the most recent
    tree mod log operation for eb X, which is the one just added by task B
    at the previous step, with a sequence number of 4, a type of
    BTRFS_MOD_LOG_KEY_ADD and for slot 0;

17) Before task A locks and clones eb X, task A adds another key to eb X,
    which results in adding a new BTRFS_MOD_LOG_KEY_ADD mod log operation,
    with a sequence number of 5, for slot 1 of eb X, increments the
    number of items in eb X from 1 to 2, and unlocks eb X.
    Now fs_info->tree_mod_seq has a value of 5;

18) Task A then locks eb X and clones it. The clone has a value of 2 for
    the number of items and the pointer "tm" points to the tree mod log
    operation with sequence number 4, not the most recent one with a
    sequence number of 5, so there is mismatch between the number of
    mod log operations that are going to be applied to the cloned version
    of eb X and the number of items in the clone;

19) Task A then calls tree_mod_log_rewind() with the clone of eb X, the
    tree mod log operation with sequence number 4 and a type of
    BTRFS_MOD_LOG_KEY_ADD, and time_seq == 1;

20) At tree_mod_log_rewind(), we set the local variable "n" with a value
    of 2, which is the number of items in the clone of eb X.

    Then in the first iteration of the while loop, we process the mod log
    operation with sequence number 4, which is targeted at slot 0 and has
    a type of BTRFS_MOD_LOG_KEY_ADD. This results in decrementing "n" from
    2 to 1.

    Then we pick the next tree mod log operation for eb X, which is the
    tree mod log operation with a sequence number of 2, a type of
    BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING and for slot 0, it is the one
    added in step 5 to the tree mod log tree.

    We go back to the top of the loop to process this mod log operation,
    and because its slot is 0 and "n" has a value of 1, we hit the BUG_ON:

        (...)
        switch (tm->op) {
        case BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING:
                BUG_ON(tm->slot < n);
                fallthrough;
	(...)

Fix this by checking for a more recent tree mod log operation after locking
and cloning the extent buffer of the old root node, and use it as the first
operation to apply to the cloned extent buffer when rewinding it.

Stable backport notes: due to moved code and renames, in =< 5.11 the
change should be applied to ctree.c:get_old_root.

Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org>
Link: https://lore.kernel.org/linux-btrfs/20210404040732.GZ32440@hungrycats.org/
Fixes: 834328a849 ("Btrfs: tree mod log's old roots could still be part of the tree")
CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/tree-mod-log.c

@@ -830,10 +830,30 @@ struct extent_buffer *btrfs_get_old_root(struct btrfs_root *root, u64 time_seq)
 				   "failed to read tree block %llu from get_old_root",
 				   logical);
 		} else {
+			struct tree_mod_elem *tm2;
+
 			btrfs_tree_read_lock(old);
 			eb = btrfs_clone_extent_buffer(old);
+			/*
+			 * After the lookup for the most recent tree mod operation
+			 * above and before we locked and cloned the extent buffer
+			 * 'old', a new tree mod log operation may have been added.
+			 * So lookup for a more recent one to make sure the number
+			 * of mod log operations we replay is consistent with the
+			 * number of items we have in the cloned extent buffer,
+			 * otherwise we can hit a BUG_ON when rewinding the extent
+			 * buffer.
+			 */
+			tm2 = tree_mod_log_search(fs_info, logical, time_seq);
 			btrfs_tree_read_unlock(old);
 			free_extent_buffer(old);
+			ASSERT(tm2);
+			ASSERT(tm2 == tm || tm2->seq > tm->seq);
+			if (!tm2 || tm2->seq < tm->seq) {
+				free_extent_buffer(eb);
+				return NULL;
+			}
+			tm = tm2;
 		}
 	} else if (old_root) {
 		eb_root_owner = btrfs_header_owner(eb_root);