There is a fairly unlikely race condition in tree mod log rewind that
can result in a kernel panic which has the following trace:

  [530.569] BTRFS critical (device sda3): unable to find logical 0 length 4096
  [530.585] BTRFS critical (device sda3): unable to find logical 0 length 4096
  [530.602] BUG: kernel NULL pointer dereference, address: 0000000000000002
  [530.618] #PF: supervisor read access in kernel mode
  [530.629] #PF: error_code(0x0000) - not-present page
  [530.641] PGD 0 P4D 0
  [530.647] Oops: 0000 [#1] SMP
  [530.654] CPU: 30 PID: 398973 Comm: below Kdump: loaded Tainted: G S         O  K   5.12.0-0_fbk13_clang_7455_gb24de3bdb045 #1
  [530.680] Hardware name: Quanta Mono Lake-M.2 SATA 1HY9U9Z001G/Mono Lake-M.2 SATA, BIOS F20_3A15 08/16/2017
  [530.703] RIP: 0010:__btrfs_map_block+0xaa/0xd00
  [530.755] RSP: 0018:ffffc9002c2f7600 EFLAGS: 00010246
  [530.767] RAX: ffffffffffffffea RBX: ffff888292e41000 RCX: f2702d8b8be15100
  [530.784] RDX: ffff88885fda6fb8 RSI: ffff88885fd973c8 RDI: ffff88885fd973c8
  [530.800] RBP: ffff888292e410d0 R08: ffffffff82fd7fd0 R09: 00000000fffeffff
  [530.816] R10: ffffffff82e57fd0 R11: ffffffff82e57d70 R12: 0000000000000000
  [530.832] R13: 0000000000001000 R14: 0000000000001000 R15: ffffc9002c2f76f0
  [530.848] FS:  00007f38d64af000(0000) GS:ffff88885fd80000(0000) knlGS:0000000000000000
  [530.866] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  [530.880] CR2: 0000000000000002 CR3: 00000002b6770004 CR4: 00000000003706e0
  [530.896] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  [530.912] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  [530.928] Call Trace:
  [530.934]  ? btrfs_printk+0x13b/0x18c
  [530.943]  ? btrfs_bio_counter_inc_blocked+0x3d/0x130
  [530.955]  btrfs_map_bio+0x75/0x330
  [530.963]  ? kmem_cache_alloc+0x12a/0x2d0
  [530.973]  ? btrfs_submit_metadata_bio+0x63/0x100
  [530.984]  btrfs_submit_metadata_bio+0xa4/0x100
  [530.995]  submit_extent_page+0x30f/0x360
  [531.004]  read_extent_buffer_pages+0x49e/0x6d0
  [531.015]  ? submit_extent_page+0x360/0x360
  [531.025]  btree_read_extent_buffer_pages+0x5f/0x150
  [531.037]  read_tree_block+0x37/0x60
  [531.046]  read_block_for_search+0x18b/0x410
  [531.056]  btrfs_search_old_slot+0x198/0x2f0
  [531.066]  resolve_indirect_ref+0xfe/0x6f0
  [531.076]  ? ulist_alloc+0x31/0x60
  [531.084]  ? kmem_cache_alloc_trace+0x12e/0x2b0
  [531.095]  find_parent_nodes+0x720/0x1830
  [531.105]  ? ulist_alloc+0x10/0x60
  [531.113]  iterate_extent_inodes+0xea/0x370
  [531.123]  ? btrfs_previous_extent_item+0x8f/0x110
  [531.134]  ? btrfs_search_path_in_tree+0x240/0x240
  [531.146]  iterate_inodes_from_logical+0x98/0xd0
  [531.157]  ? btrfs_search_path_in_tree+0x240/0x240
  [531.168]  btrfs_ioctl_logical_to_ino+0xd9/0x180
  [531.179]  btrfs_ioctl+0xe2/0x2eb0

This occurs when logical inode resolution takes a tree mod log sequence
number, and then while backref walking hits a rewind on a busy node
which has the following sequence of tree mod log operations (numbers
filled in from a specific example, but they are somewhat arbitrary)

  REMOVE_WHILE_FREEING slot 532
  REMOVE_WHILE_FREEING slot 531
  REMOVE_WHILE_FREEING slot 530
  ...
  REMOVE_WHILE_FREEING slot 0
  REMOVE slot 455
  REMOVE slot 454
  REMOVE slot 453
  ...
  REMOVE slot 0
  ADD slot 455
  ADD slot 454
  ADD slot 453
  ...
  ADD slot 0
  MOVE src slot 0 -> dst slot 456 nritems 533
  REMOVE slot 455
  REMOVE slot 454
  REMOVE slot 453
  ...
  REMOVE slot 0

When this sequence gets applied via btrfs_tree_mod_log_rewind, it
allocates a fresh rewind eb, and first inserts the correct key info for
the 533 elements, then overwrites the first 456 of them, then decrements
the count by 456 via the add ops, then rewinds the move by doing a
memmove from 456:988->0:532. We have never written anything past 532, so
that memmove writes garbage into the 0:532 range. In practice, this
results in a lot of fully 0 keys. The rewind then puts valid keys into
slots 0:455 with the last removes, but 456:532 are still invalid.

When search_old_slot uses this eb, if it uses one of those invalid
slots, it can then read the extent buffer and issue a bio for offset 0
which ultimately panics looking up extent mappings.

This bad tree mod log sequence gets generated when the node balancing
code happens to do a balance_node_right followed by a push_node_left
while logging in the tree mod log. Illustrated for ebs L and R (left and
right):

	L                 R
  start:
  [XXX|YYY|...]      [ZZZ|...|...]
  balance_node_right:
  [XXX|YYY|...]      [...|ZZZ|...] move Z to make room for Y
  [XXX|...|...]      [YYY|ZZZ|...] copy Y from L to R
  push_node_left:
  [XXX|YYY|...]      [...|ZZZ|...] copy Y from R to L
  [XXX|YYY|...]      [ZZZ|...|...] move Z into emptied space (NOT LOGGED!)

This is because balance_node_right logs a move, but push_node_left
explicitly doesn't. That is because logging the move would remove the
overwritten src < dst range in the right eb, which was already logged
when we called btrfs_tree_mod_log_eb_copy. The correct sequence would
include a move from 456:988 to 0:532 after remove 0:455 and before
removing 0:532. Reversing that sequence would entail creating keys for
0:532, then moving those keys out to 456:988, then creating more keys
for 0:455.

i.e.,

  REMOVE_WHILE_FREEING slot 532
  REMOVE_WHILE_FREEING slot 531
  REMOVE_WHILE_FREEING slot 530
  ...
  REMOVE_WHILE_FREEING slot 0
  MOVE src slot 456 -> dst slot 0 nritems 533
  REMOVE slot 455
  REMOVE slot 454
  REMOVE slot 453
  ...
  REMOVE slot 0
  ADD slot 455
  ADD slot 454
  ADD slot 453
  ...
  ADD slot 0
  MOVE src slot 0 -> dst slot 456 nritems 533
  REMOVE slot 455
  REMOVE slot 454
  REMOVE slot 453
  ...
  REMOVE slot 0

Fix this to log the move but avoid the double remove by putting all the
logging logic in btrfs_tree_mod_log_eb_copy which has enough information
to detect these cases and properly log moves, removes, and adds. Leave
btrfs_tree_mod_log_insert_move to handle insert_ptr and delete_ptr's
tree mod logging.

(Un)fortunately, this is quite difficult to reproduce, and I was only
able to reproduce it by adding sleeps in btrfs_search_old_slot that
would encourage more log rewinding during ino_to_logical ioctls. I was
able to hit the warning in the previous patch in the series without the
fix quite quickly, but not after this patch.

CC: stable@vger.kernel.org # 5.15+
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>