Re: [PATCH] ublk: skip blk_mq_tag_to_rq() bounds check

[Jens Axboe]

On 4/11/25 12:36 PM, Caleb Sander Mateos wrote:
> On Thu, Apr 10, 2025 at 6:13?AM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> On 4/10/25 3:24 AM, Christoph Hellwig wrote:
>>> On Tue, Apr 08, 2025 at 08:49:54PM -0600, Caleb Sander Mateos wrote:
>>>> The ublk driver calls blk_mq_tag_to_rq() in several places.
>>>> blk_mq_tag_to_rq() tolerates an invalid tag for the tagset, checking it
>>>> against the number of tags and returning NULL if it is out of bounds.
>>>> But all the calls from the ublk driver have already verified the tag
>>>> against the ublk queue's queue depth. In ublk_commit_completion(),
>>>> ublk_handle_need_get_data(), and case UBLK_IO_COMMIT_AND_FETCH_REQ, the
>>>> tag has already been checked in __ublk_ch_uring_cmd(). In
>>>> ublk_abort_queue(), the loop bounds the tag by the queue depth. In
>>>> __ublk_check_and_get_req(), the tag has already been checked in
>>>> __ublk_ch_uring_cmd(), in the case of ublk_register_io_buf(), or in
>>>> ublk_check_and_get_req().
>>>>
>>>> So just index the tagset's rqs array directly in the ublk driver.
>>>> Convert the tags to unsigned, as blk_mq_tag_to_rq() does.
>>>
>>> Poking directly into block layer internals feels like a really bad
>>> idea.  If this is important enough we'll need a non-checking helper
>>> in the core code, but as with all these kinds of micro-optimizations
>>> it better have a really good justification.
>>
>> FWIW, I agree, and I also have a hard time imagining this making much of
>> a measurable difference. Caleb, was this based "well this seems
>> pointless" or was it something you noticed in profiling/testing?
> 
> That's true, the nr_tags check doesn't show up super prominently in a
> CPU profile. The atomic reference counting in
> __ublk_check_and_get_req() or ublk_commit_completion() is
> significantly more expensive. Still, it seems like unnecessary work.

Matching atomics on either side is always going to be miserable, and I'd
wager a much bigger issue than the minor thing that this patch is trying
to address...

> nr_tags is in a different cache line from rqs, so there is the
> potential for a cache miss. And the prefetch() is another unnecessary
> cache miss in the cases where ublk doesn't access any of struct
> request's fields.
> I am happy to add a "blk_mq_tag_to_rq_unchecked()" helper to avoid
> accessing the blk-mq internals.

Or maybe go the route that Ming suggested? But if you go the other
route, I'd just add a __blk_mq_tag_to_rq() and have blk_mq_tag_to_rq()
call that with the validation happening before.

-- 
Jens Axboe