Re: [bug report] block: Non-NCQ commands will never be executed while fio is continuously running

[Damien Le Moal <dlemoal@kernel.org>]

On 9/10/24 20:27, Niklas Cassel wrote:
> On Tue, Sep 10, 2024 at 02:34:06PM +0800, yangxingui wrote:
>>
>>
>> On 2024/9/10 12:45, Damien Le Moal wrote:
>>> On 9/10/24 10:09 AM, yangxingui wrote:
>>>>
>>>>
>>>> On 2024/9/9 21:21, Damien Le Moal wrote:
>>>>> On 9/9/24 22:10, yangxingui wrote:
>>>>>> Hello axboe & John,
>>>>>>
>>>>>> After the driver exposes all HW queues to the block layer, non-NCQ
>>>>>> commands will never be executed while fio is continuously running, such
>>>>>> as a smartctl command.
>>>>>>
>>>>>> The cause of the problem is that other hctx used by the NCQ command is
>>>>>> still active and can continue to issue NCQ commands to the sata disk.
>>>>>> And the pio command keeps retrying in its corresponding hctx because
>>>>>> qc_defer() always returns true.
>>>>>>
>>>>>> hctx0: ncq, pio, ncq
>>>>>> hctx1：ncq, ncq, ...
>>>>>> ...
>>>>>> hctxn: ncq, ncq, ...
>>>>>>
>>>>>> Is there any good solution for this?
>>>>>
>>>>> SATA devices are single queue so how can you have multiple queues ?
>>>>> What adapter are you using ?
>>>>
>>>> In the following patch, we expose the host's 16 hardware queues to the block
>>>> layer. And when connecting to a sata disk, 16 hctx are used.
>>>>
>>>> 8d98416a55eb ("scsi: hisi_sas: Switch v3 hw to MQ")
>>>
>>> OK, so the HBA is a hisi one, using libsas...
>>> What is the device ? An SSD ? and HDD ?
>> Both SATA SSD and SATA HDD have this problem.
>>
>>>
>>> Do you set a block I/O scheduler for the drive, e.g. mq-deadline. If not, does
>>> setting a scheduler resolve the issue ?
>> Currently, the default configuration mq-deadline is used, and the same
>> phenomenon occurs when I try setting it to none. It seems to have nothing to
>> do with the scheduling strategy.
>>
>>>
>>> I do not have any hisi HBA. I use a lot of mpt3sas and mpi3mr HBAs which also
>>> have multiple queues with a shared tagset. Never seen the issue you are
>>> reporting though using HDDs with mq-deadline or bfq as the scheduler.
>> Unlike libsas, as these hosts don't use qc_defer()?
> 
> mpt3sas and mpi3mr do not use any libata code at all, the SCSI to ATA
> Translation (SAT) is done completely by the HBA, so from a Linux
> perspective, we are issuing SCSI commands to the HBA.

Yes, but we still can get requeue happening. Though for a SATA drive, that is
unlikely since the max queue depth is clearly defined, unlike for SAS drives

> We can see that libsas uses ata_std_qc_defer() as its .qc_defer callback:
> https://github.com/torvalds/linux/blob/v6.11-rc7/drivers/scsi/libsas/sas_ata.c#L566

And that may be the issue. More on this below.

> Without considering if it is a good idea or not, it should be possible to
> translate some commands to instead use the "NCQ encapsulated" variant of
> the ATA command that was used in the "ATA-16 passthrough" SCSI command.

That would be way too much work on the user side, and likely open up a can of
device bugs unseen until now.

> To be able to send a non-queued command, there has to be no NCQ commands queued
> on the device. I guess you could implement a scheduler that would be quiescing
> the queue, processes the non-queued command, and then thaw the queue, but that
> would essentially make non-queued commands high priority commands, and could
> thus be used to seriously limit throughput by just sending some non-queued
> commands every now and then :)

Passthrough commands do not go through the scheduler and are submitted directly
to the dispatch queue, generally at the head of it (see blk_mq_insert_request()).

So for a single queue device, even if ata_qc_defer causes a requeue, the
passthrough command ends up back at the top of the dispatch queue. After
repeating this a few times, all in-flight NCQ commands complete and the
passthrough command goes through.

But I feel this is very fragile given that the block layer requeue is done
through a work item, so in parallel to an application submitting IOs. So in
theory, I think that the requeue for the passthrough command could happen forever...

And for a multi-queue setup like with the hisi adapter, that is what is happening.

I do not have any good idea how to fix that yet. We need to find something.
scsi_queue_rq() and the budget/host or device blocked state management may help
with that, or we have a bug there... In any case, I do not think it is a block
layer issue as the block layer knows nothing about NCQ vs non-NCQ.

-- 
Damien Le Moal
Western Digital Research