Re: [PATCH] scsi: bsg: copy uring_cmd payload to prevent double-fetch from shared SQE

[Jens Axboe <axboe@kernel.dk>]

On 5/27/26 10:27 AM, Caleb Sander Mateos wrote:
> On Wed, May 27, 2026 at 9:19 AM Rahul Chandelkar <rc@rexion.ai> wrote:
>>
>> On Wed, May 27, 2026 at 10:06:44AM -0600, Jens Axboe wrote:
>>> I don't think this is the right way to fix it, ->sqe should've been
>>> stable upfront if this ends up happening. Can you share your poc with
>>> me? Your trace has been trimmed down way too much to be useful.
>>
>> Agreed that a core-level copy before the inline callback would be the
>> right fix and would eliminate the entire class for every uring_cmd
>> driver. The per-driver copy was meant as a minimal backportable fix
>> for the immediate scsi_bsg path.
>>
>> PoC and full trace below.
>>
>> --- PoC (poc_bsg_toctou.c) ---
>>
>> Build:  gcc -O2 -pthread -static -o poc poc_bsg_toctou.c
>> Usage:  ./poc /dev/bsg/X
>> Needs:  2+ CPUs, io_uring, /dev/bsg/* access
>>
>> The racer thread flips request_len between 16 (passes the <=32 bounds
>> check) and 128 (used by copy_from_user, overflows scmd->cmnd[32]).
>> The overflow payload plants 0xdead000000001000 at the sense_buffer
>> pointer offset (+84 from cmnd[0]). When scsi_queue_rq() does
>> memset(scmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE) it faults on the
>> corrupted pointer.
> 
> Then the fix is to use READ_ONCE() to access the SQE fields, right?
> Copying the entire SQE seems like unnecessary overhead. See
> nvme_uring_cmd_io() for prior art.

That is indeed the correct fix.

-- 
Jens Axboe