Re: [PATCH 1/5] scsi: hisi_sas: Add host_tagset_enable module param for v3 hw

Re: [PATCH 1/5] scsi: hisi_sas: Add host_tagset_enable module param for v3 hw

[John Garry]

On 29/03/2025 09:49, yangxingui wrote:
> Hi，John
> 
> On 2025/3/29 16:50, John Garry wrote:
>> On 29/03/2025 07:32, Yihang Li wrote:
>>
>> +
>>
>>> From: Xingui Yang<yangxingui@huawei.com>
>>>
>>> After driver exposes all HW queues and application submits IO to 
>>> multiple
>>> queues in parallel, if NCQ and non-NCQ commands are mixed to sata disk,
>>> ata_qc_defer() causes non-NCQ commands to be requeued and possibly 
>>> repeated
>>> forever.
>>
>> I don't think that it is a good idea to mask out bugs with module 
>> parameters.
>>
>> Was this the same libata/libsas issue reported some time ago?
> 
> Yeah，related to this issue: https://lore.kernel.org/linux-block/ 
> eef1e927-c9b2-c61d-7f48-92e65d8b0418@huawei.com/
> 
> And, Niklas tried to help fix this problem: https://lore.kernel.org/ 
> linux-scsi/ZynmfyDA9R-lrW71@ryzen/
> 
> Considering that there is no formal solution yet. And our users rarely 
> use SATA disks and SAS disks together on a single machine. For this 
> reason, they can flexibly turn off the exposure of multiple queues in 
> the scenario of using only SATA disks. In addition, it is also 
> convenient to conduct performance comparison tests to expose multiple 
> hardware queues and single queues.
> 

The change in this series does not even solve the issues, as:
- you do not guarantee no SAS/SATA mix without that module param enabled
- the driver still uses managed interrupts in both cases, so with 
disabling host_tagset you are now exposed to CPU hotplug issue of IO 
being in-flight when HW queue interrupt is shutdown

And pm8001 driver will have the same issue, so we need to find a proper fix.

Let me consider this issue more.

Re: [PATCH 1/5] scsi: hisi_sas: Add host_tagset_enable module param for v3 hw

[yangxingui]

On 2025/3/31 15:43, John Garry wrote:
> On 29/03/2025 09:49, yangxingui wrote:
>> Hi，John
>>
>> On 2025/3/29 16:50, John Garry wrote:
>>> On 29/03/2025 07:32, Yihang Li wrote:
>>>
>>> +
>>>
>>>> From: Xingui Yang<yangxingui@huawei.com>
>>>>
>>>> After driver exposes all HW queues and application submits IO to 
>>>> multiple
>>>> queues in parallel, if NCQ and non-NCQ commands are mixed to sata disk,
>>>> ata_qc_defer() causes non-NCQ commands to be requeued and possibly 
>>>> repeated
>>>> forever.
>>>
>>> I don't think that it is a good idea to mask out bugs with module 
>>> parameters.
>>>
>>> Was this the same libata/libsas issue reported some time ago?
>>
>> Yeah，related to this issue: https://lore.kernel.org/linux-block/ 
>> eef1e927-c9b2-c61d-7f48-92e65d8b0418@huawei.com/
>>
>> And, Niklas tried to help fix this problem: https://lore.kernel.org/ 
>> linux-scsi/ZynmfyDA9R-lrW71@ryzen/
>>
>> Considering that there is no formal solution yet. And our users rarely 
>> use SATA disks and SAS disks together on a single machine. For this 
>> reason, they can flexibly turn off the exposure of multiple queues in 
>> the scenario of using only SATA disks. In addition, it is also 
>> convenient to conduct performance comparison tests to expose multiple 
>> hardware queues and single queues.
>>
> 
> The change in this series does not even solve the issues, as:
> - you do not guarantee no SAS/SATA mix without that module param enabled
> - the driver still uses managed interrupts in both cases, so with 
> disabling host_tagset you are now exposed to CPU hotplug issue of IO 
> being in-flight when HW queue interrupt is shutdown

Yes, there will be such problems.

> 
> And pm8001 driver will have the same issue, so we need to find a proper 
> fix.
> 
> Let me consider this issue more.

OK.

Thanks,
Xingui