Re: qemu:arm test failure due to commit 8053871d0f7f (smp: Fix smp_call_function_single_async() locking)

[Linus Torvalds <torvalds@linux-foundation.org>]

On Sun, Apr 19, 2015 at 10:39 PM, Ingo Molnar <mingo@kernel.org> wrote:
>
>>
>> So I _could_ imagine that somebody would want to do optimistic "prod
>> other cpu" calls that in all normal cases are for existing cpus, but
>> could be racy in theory.
>
> Yes, and I don't disagree with such optimizations in principle (it
> allows less references to be taken in the fast path), but is it really
> safe?
>
> If a CPU is going down and we potentially race against that, and send
> off an IPI, the IPI might be 'in flight' for an indeterminate amount
> of time, especially on wildly non-deterministic hardware like virtual
> platforms.

Well, it should be easy enough to handle that race in the cpu
offlining: after the cpu is marked "not present", just call
flush_smp_call_function_queue(), In fact, I thought we did exactly
that - it's the reason for the "warn_cpu_offline" argument, isn't it)?

So I don't think there should be any real race.  Sure, the HW IPI
itself might be in flight, but from a sw perspective isn't all done.

No, I was talking about something even more optimistic - the CPU
number we optimisitcally loaded and sent an IPI to might be completely
bogus just because we loaded it using some unlocked sequence, and
maybe the memory got re-assigned. So it might not even be a CPU number
that is "stale", it could be entirely invalid.

And no, I don't claim that we should do this, I'm just saying that I
could imagine this being a valid thing to do. But it might be a good
idea to add a WARN_ON_ONCE() for now to find the users that are not
being clever like this, they are just being stupid and wrong-headed,
and sending IPI's to bogus CPU's not because they are doing really
subtle smart stuff, but just because they never noticed how stupid
they are..

                      Linus