Re: [PATCH v5 0/9] block device fixes for large block sizes, IOCB_NOWAIT, and direct I/O

[Tal Zussman]

On 10/7/26 10:05 AM, Jens Axboe wrote:
> On 9/23/26 5:56 PM, Tal Zussman wrote:
>> A set of independent fixes for the block device file operations. The
>> first two were found by Sashiko while reviewing the RWF_DONTCACHE series
>> [1]. The fourth and fifth were found by Sashiko's review of v1 of this
>> series, and the rest came from asking an LLM to find any similar or
>> related issues. Each issue has been reproduced, with the fixes resolving
>> the issues.
>> 
>> Patch 1 fixes silently lost mmap writes with CONFIG_BUFFER_HEAD=n.
>> 
>> Patches 2 and 3 take i_rwsem around the direct I/O write fallback and
>> the splice read path, which race set_blocksize() changing the mapping's
>> minimum folio order.
>> 
>> Patch 4 makes the buffered read path honor IOCB_NOWAIT instead of
>> blocking on i_rwsem.
>> 
>> Patch 5 makes IOCB_ATOMIC writes fail instead of tearing, and patch 6
>> stops the buffered fallback from retrying an atomic write. Block
>> devices can reject both paths into the fallback before submitting any
>> I/O, so we fail rather than issue a WARN() like ext4 does. Patch 7
>> makes iomap_file_buffered_write() reject IOCB_ATOMIC, so no other
>> buffered fallback can complete an atomic write either.
>> 
>> Patch 8 fixes leaked page pins in bio_iov_iter_align_down(), and patch
>> 9 removes dead metadata handling in the async direct I/O path.
>> 
>> These issues are currently unlikely to be hit in practice due to the
>> specific configurations required to trigger them.
>> 
>> The reproducer for patch 1 is in blktests as block/048, and tests for
>> patches 2, 3 and 8 are posted at [2].
> 
> This is based against the 7.3 series for some reason, can you please
> resend against the 7.4 based branch, for-7.4/block.
> 

Will do that shortly. I had assumed initially this would go in as fixes
for 7.3, hence basing it on the 7.3 branch.