On 9/27/22 9:20 AM, Pankaj Raghav wrote:
>>> I guess the second patch should be enough to apply plugging when
>>> applicable for uring_cmd based nvme passthrough requests.
>>
>> Do we even need the 2nd patch? If we're just doing passthrough for the
>> blk_execute_nowait() API, then the condition should never trigger?
>
> I think this was the question I raised in your first version of the series.
>
> If we do a NVMe write using the passthrough interface, then we will be
> using REQ_OP_DRV_OUT op, which is:
>
> REQ_OP_DRV_OUT = (__force blk_opf_t)35, // write bit is set
>
> The condition in blk_mq_plug() will trigger as we only check if it is a
> _write_ (op & (__force blk_opf_t)1) to the device. Am I missing something?
Ah yes, good point. We used to have this notion of 'fs' request, don't
think we do anymore. Because it really should just be:
if (zoned && (op & REQ_OP_WRITE) && fs_request)
return NULL;
for that condition imho. I guess we could make it:
if (zoned && (op & REQ_OP_WRITE) && !(op & REQ_OP_DRV_OUT))
return NULL;
--
Jens Axboe
