Re: system hung up when offlining CPUs

From: Hannes Reinecke
Date: Wed Nov 01 2017 - 07:01:19 EST


On 11/01/2017 01:47 AM, Thomas Gleixner wrote:
> On Mon, 30 Oct 2017, Shivasharan Srikanteshwara wrote:
>
>> In managed-interrupts case, interrupts which were affine to the offlined
>> CPU is not getting migrated to another available CPU. But the
>> documentation at below link says that "all interrupts" are migrated to a
>> new CPU. So not all interrupts are getting migrated to a new CPU then.
>
> Correct.
>
>> https://www.kernel.org/doc/html/v4.11/core-api/cpu_hotplug.html#the-offlin
>> e-case
>> "- All interrupts targeted to this CPU are migrated to a new CPU"
>
> Well, documentation is not always up to date :)
>
>> Once the last CPU in the affinity mask is offlined and a particular IRQ
>> is shutdown, is there a way currently for the device driver to get
>> callback to complete all outstanding requests on that queue?
>
> No and I have no idea how the other drivers deal with that.
>
> The way you can do that is to have your own hotplug callback which is
> invoked when the cpu goes down, but way before the interrupt is shut down,
> which is one of the last steps. Ideally this would be a callback in the
> generic block code which then calls out to all instances like its done for
> the cpu dead state.
>
In principle, yes, that would be (and, in fact, might already) moved to
the block layer for blk-mq, as this has full control over the individual
queues and hence can ensure that the queues with dead/removed CPUs are
properly handled.

Here, OTOH, we are dealing with the legacy sq implementation (or, to be
precised, a blk-mq implementation utilizing only a single queue), so
that any of this handling need to be implemented in the driver.

So what would need to be done here is to implement a hotplug callback in
the driver, which would disable the CPU from the list/bitmap of valid
cpus. Then the driver could validate the CPU number with this bitmap
upon I/O submission (instead of just using raw_smp_cpu_number()), and
could set the queue ID to '0' if an invalid CPU was found.
With that the driver should be able to ensure that no new I/O will be
submitted which will hit the dead CPU, so with a bit of luck this might
already solve the problem.

Alternatively I could resurrect my patchset converting the driver to
blk-mq, which got vetoed the last time ...

Cheers,

Hannes
--
Dr. Hannes Reinecke Teamlead Storage & Networking
hare@xxxxxxx +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 NÃrnberg
GF: F. ImendÃrffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG NÃrnberg)