btrfs: fix false alert on bad tree level check

btrfs: fix false alert on bad tree level check

[BUG]
There is a bug report that on a RAID0 NVMe btrfs system, under heavy
write load the filesystem can flip RO randomly.

With extra debugging, it shows some tree blocks failed to pass their
level checks, and if that happens at critical path of a transaction, we
abort the transaction:

  BTRFS error (device nvme0n1p3): level verify failed on logical 5446121209856 mirror 1 wanted 0 found 1
  BTRFS error (device nvme0n1p3: state A): Transaction aborted (error -5)
  BTRFS: error (device nvme0n1p3: state A) in btrfs_finish_ordered_io:3343: errno=-5 IO failure
  BTRFS info (device nvme0n1p3: state EA): forced readonly

[CAUSE]
The reporter has already bisected to commit 947a629988f1 ("btrfs: move
tree block parentness check into validate_extent_buffer()").

And with extra debugging, it shows we can have btrfs_tree_parent_check
filled with all zeros in the following call trace:

  submit_one_bio+0xd4/0xe0
  submit_extent_page+0x142/0x550
  read_extent_buffer_pages+0x584/0x9c0
  ? __pfx_end_bio_extent_readpage+0x10/0x10
  ? folio_unlock+0x1d/0x50
  btrfs_read_extent_buffer+0x98/0x150
  read_tree_block+0x43/0xa0
  read_block_for_search+0x266/0x370
  btrfs_search_slot+0x351/0xd30
  ? lock_is_held_type+0xe8/0x140
  btrfs_lookup_csum+0x63/0x150
  btrfs_csum_file_blocks+0x197/0x6c0
  ? sched_clock_cpu+0x9f/0xc0
  ? lock_release+0x14b/0x440
  ? _raw_read_unlock+0x29/0x50
  btrfs_finish_ordered_io+0x441/0x860
  btrfs_work_helper+0xfe/0x400
  ? lock_is_held_type+0xe8/0x140
  process_one_work+0x294/0x5b0
  worker_thread+0x4f/0x3a0
  ? __pfx_worker_thread+0x10/0x10
  kthread+0xf5/0x120
  ? __pfx_kthread+0x10/0x10
  ret_from_fork+0x2c/0x50

Currently we only copy the btrfs_tree_parent_check structure into bbio
at read_extent_buffer_pages() after we have assembled the bbio.

But as shown above, submit_extent_page() itself can already submit the
bbio, leaving the bbio->parent_check uninitialized, and cause the false
alert.

[FIX]
Instead of copying @check into bbio after bbio is assembled, we pass
@check in btrfs_bio_ctrl::parent_check, and copy the content of
parent_check in submit_one_bio() for metadata read.

By this we should be able to pass the needed info for metadata endio
verification, and fix the false alert.

Reported-by: Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>
Link: https://lore.kernel.org/linux-btrfs/CABXGCsNzVxo4iq-tJSGm_kO1UggHXgq6CdcHDL=z5FL4njYXSQ@mail.gmail.com/
Fixes: 947a629988f1 ("btrfs: move tree block parentness check into validate_extent_buffer()")
Tested-by: Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>