Re: [PATCH v2 3/5] locking/qspinlock: Introduce CNA into the slow path of qspinlock

From: Alex Kogan
Date: Tue Jun 04 2019 - 19:26:41 EST

Hi, Peter, Longman,

> On Apr 3, 2019, at 12:01 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Wed, Apr 03, 2019 at 11:39:09AM -0400, Alex Kogan wrote:
>>>> The patch that I am looking for is to have a separate
>>>> numa_queued_spinlock_slowpath() that coexists with
>>>> native_queued_spinlock_slowpath() and
>>>> paravirt_queued_spinlock_slowpath(). At boot time, we select the most
>>>> appropriate one for the system at hand.
>> Is this how this selection works today for paravirt?
>> I see a PARAVIRT_SPINLOCKS config option, but IIUC you are talking about a different mechanism here.
>> Can you, please, elaborate or give me a link to a page that explains that?
> Oh man, you ask us to explain how paravirt patching works... that's
> magic :-)
> Basically, the compiler will emit a bunch of indirect calls to the
> various pv_ops.*.* functions.
> Then, at alternative_instructions() <- apply_paravirt() it will rewrite
> all these indirect calls to direct calls to the function pointers that
> are in the pv_ops structure at that time (+- more magic).
Trying to resume this work, I am looking for concrete steps required to integrate CNA with the paravirt patching.

Looking at alternative_instructions(), I wonder if I need to add another call, something like apply_numa() similar to apply_paravirt(), and do the patch work there.
Or perhaps I should âjust" initialize the pv_ops structure with the corresponding numa_queued_spinlock_slowpath() in paravirt.c?

Also, the paravirt code is under arch/x86, while CNA is generic (not x86-specific).
Do you still want to see CNA-related patching residing under arch/x86?

We still need a config option (something like NUMA_AWARE_SPINLOCKS) to enable CNA patching under this config only, correct?

Thanks in advance,
â Alex