btrfs: split partial dio bios before submit

btrfs: split partial dio bios before submit

If an application is doing direct io to a btrfs file and experiences a
page fault reading from the write buffer, iomap will issue a partial
bio, and allow the fs to keep going. However, there was a subtle bug in
this code path in the btrfs dio iomap implementation that led to the
partial write ending up as a gap in the file's extents and to be read
back as zeros.

The sequence of events in a partial write, lightly summarized and
trimmed down for brevity is as follows:

==== WRITING TASK ====
 btrfs_direct_write
 __iomap_dio_write
 iomap_iter
 btrfs_dio_iomap_begin # create full ordered extent
 iomap_dio_bio_iter
 bio_iov_iter_get_pages # page fault; partial read
 submit_bio # partial bio
 iomap_iter
 btrfs_dio_iomap_end
 btrfs_mark_ordered_io_finished # sets BTRFS_ORDERED_IOERR;
				# submit to finish_ordered_fn wq
 fault_in_iov_iter_readable # btrfs_direct_write detects partial write
 __iomap_dio_write
 iomap_iter
 btrfs_dio_iomap_begin # create second partial ordered extent
 iomap_dio_bio_iter
 bio_iov_iter_get_pages # read all of remainder
 submit_bio # partial bio with all of remainder
 iomap_iter
 btrfs_dio_iomap_end # nothing exciting to do with ordered io

==== DIO ENDIO ====
== FIRST PARTIAL BIO ==
 btrfs_dio_end_io
 btrfs_mark_ordered_io_finished # bytes_left > 0
			        # don't submit to finish_ordered_fn wq
== SECOND PARTIAL BIO ==
 btrfs_dio_end_io
 btrfs_mark_ordered_io_finished # bytes_left == 0
			        # submit to finish_ordered_fn wq

==== BTRFS FINISH ORDERED WQ ====
== FIRST PARTIAL BIO ==
 btrfs_finish_ordered_io # called by dio_iomap_end_io, sees
		         # BTRFS_ORDERED_IOERR, just drops the
		         # ordered_extent
==SECOND PARTIAL BIO==
 btrfs_finish_ordered_io # called by btrfs_dio_end_io, writes out file
		         # extents, csums, etc...

The essence of the problem is that while btrfs_direct_write and iomap
properly interact to submit all the correct bios, there is insufficient
logic in the btrfs dio functions (btrfs_dio_iomap_begin,
btrfs_dio_submit_io, btrfs_dio_end_io, and btrfs_dio_iomap_end) to
ensure that every bio is at least a part of a completed ordered_extent.
And it is completing an ordered_extent that results in crucial
functionality like writing out a file extent for the range.

More specifically, btrfs_dio_end_io treats the ordered extent as
unfinished but btrfs_dio_iomap_end sets BTRFS_ORDERED_IOERR on it.
Thus, the finish io work doesn't result in file extents, csums, etc.
In the aftermath, such a file behaves as though it has a hole in it,
instead of the purportedly written data.

We considered a few options for fixing the bug:

  1. treat the partial bio as if we had truncated the file, which would
     result in properly finishing it.
  2. split the ordered extent when submitting a partial bio.
  3. cache the ordered extent across calls to __iomap_dio_rw in
     iter->private, so that we could reuse it and correctly apply
     several bios to it.

I had trouble with 1, and it felt the most like a hack, so I tried 2
and 3. Since 3 has the benefit of also not creating an extra file
extent, and avoids an ordered extent lookup during bio submission, it
felt like the best option. However, that turned out to re-introduce a
deadlock which this code discarding the ordered_extent between faults
was meant to fix in the first place. (Link to an explanation of the
deadlock below.)

Therefore, go with fix 2, which requires a bit more setup work but fixes
the corruption without introducing the deadlock, which is fundamentally
caused by the ordered extent existing when we attempt to fault in a
range that overlaps with it.

Put succinctly, what this patch does is: when we submit a dio bio, check
if it is partial against the ordered extent stored in dio_data, and if it
is, extract the ordered_extent that matches the bio exactly out of the
larger ordered_extent. Keep the remaining ordered_extent around in dio_data
for cancellation in iomap_end.

Thanks to Josef, Christoph, and Filipe with their help figuring out the
bug and the fix.

Fixes: 51bd9563b678 ("btrfs: fix deadlock due to page faults during direct IO reads and writes")
Link: https://bugzilla.redhat.com/show_bug.cgi?id=2169947
Link: https://lore.kernel.org/linux-btrfs/aa1fb69e-b613-47aa-a99e-a0a2c9ed273f@app.fastmail.com/
Link: https://pastebin.com/3SDaH8C6
Link: https://lore.kernel.org/linux-btrfs/20230315195231.GW10580@twin.jikos.cz/T/#t
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Tested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Boris Burkov <boris@bur.io>
[ hch: refactored the ordered_extent extraction ]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David Sterba <dsterba@suse.com>