On 3/4/21 3:09 AM, Mikulas Patocka wrote:
>
>
> On Thu, 4 Mar 2021, JeffleXu wrote:
>
>>> __split_and_process_non_flush records the poll cookie in ci.poll_cookie.
>>> When we processed all the bios, we poll for the last cookie here:
>>>
>>> if (ci.poll_cookie != BLK_QC_T_NONE) {
>>> while (atomic_read(&ci.io->io_count) > 1 &&
>>> blk_poll(ci.poll_queue, ci.poll_cookie, true)) ;
>>> }
>>
>> So what will happen if one bio submitted to dm device crosses the device
>> boundary among several target devices (e.g., dm-stripe)? Please refer
>> the following call graph.
>>
>> ```
>> submit_bio
>> __submit_bio_noacct
>> disk->fops->submit_bio(), calling into __split_and_process_bio(),
>> call __split_and_process_non_flush() once, submitting the *first* split bio
>> disk->fops->submit_bio(), calling into __split_and_process_bio(),
>> call __split_and_process_non_flush() once, submitting the *second* split bio
>> ...
>> ```
>>
>>
>> So the loop is in __submit_bio_noacct(), rather than
>> __split_and_process_bio(). Your design will send the first split bio,
>> and then poll on this split bio, then send the next split bio, polling
>> on this, go on and on...
>
> No. It will send all the bios and poll for the last one.
I took a quick look, and this seems very broken. You must not poll off
the submission path, polling should be invoked by the higher layer when
someone wants to reap events. IOW, dm should not be calling blk_poll()
by itself, only off mq_ops->poll(). Your patch seems to do it off
submission once you submit the last bio in that batch, effectively
implementing sync polling for that series. That's not right.
--
Jens Axboe
