dm: dm-crypt: Do not partially accept write BIOs with zoned targets

dm: dm-crypt: Do not partially accept write BIOs with zoned targets

Read and write operations issued to a dm-crypt target may be split
according to the dm-crypt internal limits defined by the max_read_size
and max_write_size module parameters (default is 128 KB). The intent is
to improve processing time of large BIOs by splitting them into smaller
operations that can be parallelized on different CPUs.

For zoned dm-crypt targets, this BIO splitting is still done but without
the parallel execution to ensure that the issuing order of write
operations to the underlying devices remains sequential. However, the
splitting itself causes other problems:

1) Since dm-crypt relies on the block layer zone write plugging to
   handle zone append emulation using regular write operations, the
   reminder of a split write BIO will always be plugged into the target
   zone write plugged. Once the on-going write BIO finishes, this
   reminder BIO is unplugged and issued from the zone write plug work.
   If this reminder BIO itself needs to be split, the reminder will be
   re-issued and plugged again, but that causes a call to a
   blk_queue_enter(), which may block if a queue freeze operation was
   initiated. This results in a deadlock as DM submission still holds
   BIOs that the queue freeze side is waiting for.

2) dm-crypt relies on the emulation done by the block layer using
   regular write operations for processing zone append operations. This
   still requires to properly return the written sector as the BIO
   sector of the original BIO. However, this can be done correctly only
   and only if there is a single clone BIO used for processing the
   original zone append operation issued by the user. If the size of a
   zone append operation is larger than dm-crypt max_write_size, then
   the orginal BIO will be split and processed as a chain of regular
   write operations. Such chaining result in an incorrect written sector
   being returned to the zone append issuer using the original BIO
   sector.  This in turn results in file system data corruptions using
   xfs or btrfs.

Fix this by modifying get_max_request_size() to always return the size
of the BIO to avoid it being split with dm_accpet_partial_bio() in
crypt_map(). get_max_request_size() is renamed to
get_max_request_sectors() to clarify the unit of the value returned
and its interface is changed to take a struct dm_target pointer and a
pointer to the struct bio being processed. In addition to this change,
to ensure that crypt_alloc_buffer() works correctly, set the dm-crypt
device max_hw_sectors limit to be at most
BIO_MAX_VECS << PAGE_SECTORS_SHIFT (1 MB with a 4KB page architecture).
This forces DM core to split write BIOs before passing them to
crypt_map(), and thus guaranteeing that dm-crypt can always accept an
entire write BIO without needing to split it.

This change does not have any effect on the read path of dm-crypt. Read
operations can still be split and the BIO fragments processed in
parallel. There is also no impact on the performance of the write path
given that all zone write BIOs were already processed inline instead of
in parallel.

This change also does not affect in any way regular dm-crypt block
devices.

Fixes: f211268ed1f9 ("dm: Use the block layer zone append emulation")
Cc: stable@vger.kernel.org
Signed-off-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Mikulas Patocka <mpatocka@redhat.com>
Link: https://lore.kernel.org/r/20250625093327.548866-5-dlemoal@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>