[PATCH v6] NVMe: conversion to blk-mq

[Ming Lei]

On Mon, Jun 9, 2014 at 10:08 PM, Jens Axboe <axboe at fb.com> wrote:
> On 2014-06-08 22:35, Ming Lei wrote:
>>
>> On Fri, Jun 6, 2014 at 8:20 PM, Matias Bjørling <m at bjorling.me> wrote:
>>>
>>> This converts the current NVMe driver to utilize the blk-mq layer.
>>
>>
>> Looks it can't be applied cleanly against 3.15-rc8 + Jens's for-linux
>> branch, when I fix the conflict manually, below failure is triggered:
>>
>> [  487.696057] nvme 0000:00:07.0: Cancelling I/O 202 QID 1
>> [  487.699005] nvme 0000:00:07.0: Aborting I/O 202 QID 1
>> [  487.704074] nvme 0000:00:07.0: Cancelling I/O 202 QID 1
>> [  487.717881] nvme 0000:00:07.0: Aborting I/O 202 QID 1
>> [  487.736093] end_request: I/O error, dev nvme0n1, sector 91532352
>> [  487.747378] nvme 0000:00:07.0: completed id 0 twice on queue 0
>>
>>
>> when running fio randread(libaio, iodepth:64) with more than 3 jobs.
>>
>> And looks no such failure when jobs is 1 or 2.
>
>
> It's a known issue, make sure you have this patch:
>
> http://git.kernel.dk/?p=linux-block.git;a=commit;h=f6be4fb4bcb396fc3b1c134b7863351972de081f

Looks assignment to 'rq->deadline' in blk_mq_start_request()
may be removed,  since blk_add_timer() already did that.

But that isn't the only issue.

Another one is that req->start_time isn't set without IO_STAT in
nvme, which may cause similar issue too, so req->start_time
can be set in nvme's queue_rq callback.

My real issue should be nvme specific: NVME_INTERNAL_DEV_ERROR(0x6) is
returned from device
with the conversion patch, but no such issue in current bio mode.


Thanks,
-- 
Ming Lei