[PATCH v3 1/1] nvme: multipath: Implemented new iopolicy "queue-depth"

[Sagi Grimberg]

On 22/05/2024 13:48, Nilay Shroff wrote:
>
> On 5/21/24 20:14, John Meneghini wrote:
>> On 5/21/24 06:16, Sagi Grimberg wrote:
>>>>> Exactly, nvme_mpath_init_ctrl resets the counter.
>>>> Except you're right, the counter reset needs to move to nvme_mpath_init_identify()
>>>> or some place that is called on every controller reset.
>>> This however raises the question of how much failover/reset tests this patch has seen...
>> I has received quite a bit of testing with failover and controller resets.  I shared some of the testing that was done at LSFMM last week.
>>
>> It has received enough testing to make me confident that this code is safe.  That is: it won't panic, corrupt data, or otherwise do any harm.  We believe the error paths will not be affected by this change... but I agree that running the error paths could negatively impact the accuracy of the nr_active counters... which could lead to an inaccurate outcome with the queue-depth policy.
>>
>> I agree the nr_counter initialize should move to nvme_mpath_init_identify(), or maybe be done there in addition to in nvme_mpath_init_ctrl(). I'm will to make that change now... if that's what people want.  I don't think it would require any extensive retesting.
>>
>> /John
>>
>>
> I think with Keith's recent proposed patch for fixing io accounting on failover, the
> nvme_mpath_end_request() would be called even for cancelled IO and so the nr_active
> counter shall be adjusted correctly for cancelled IO requests. Having said that, IMO
> you shall consider moving initialization of nr_active counter to nvme_mpath_init_identify()
> as that's common function invoked from regular controller initialization code path as well
> the reset code path.

Yes, and preferably with a comment explaining why its there (despite 
having nothing to do with identify...)