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

[Guenter Roeck <linux@roeck-us.net>]

On 04/18/2015 05:04 PM, Linus Torvalds wrote:
> On Sat, Apr 18, 2015 at 7:40 PM, Guenter Roeck <linux@roeck-us.net> wrote:
>> On Sat, Apr 18, 2015 at 04:23:25PM -0700, Guenter Roeck wrote:
>>>
>>> my qemu test for arm:vexpress fails with the latest upstream kernel. It fails
>>> hard - I don't get any output from the console. Bisect points to commit
>>> 8053871d0f7f ("smp: Fix smp_call_function_single_async() locking").
>>> Reverting this commit fixes the problem.
>
> Hmm. It being qemu, can you look at where it seems to lock?
>
I'll try. It must be very early in the boot process, prior to console
initialization - if I load qemu without -nographic I only get "Guest
has not initialized the display (yet)".

>> Additional observation: The system boots if I add "-smp cpus=4" to the qemu
>> options. It does still hang, however, with "-smp cpus=2" and "-smp cpus=3".
>
> Funky.
>
> That patch still looks obviously correct to me after looking at it
> again, but I guess we need to revert it if somebody can't see what's
> wrong.
>
> It does make async (wait=0) smp_call_function_single() possibly be
> *really* asynchronous, ie the 'csd' ends up being released and can be
> re-used even before the call-single function has completed. That
> should be a good thing, but I wonder if that triggers some ARM bug.
>
> Instead of doing a full revert, what happens if you replace this part:
>
> +               /* Do we wait until *after* callback? */
> +               if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
> +                       func(info);
> +                       csd_unlock(csd);
> +               } else {
> +                       csd_unlock(csd);
> +                       func(info);
> +               }
>
> with just
>
> +               func(info);
> +               csd_unlock(csd);
>
> ie keeping the csd locked until the function has actually completed? I
> guess for completeness, we should do the same thing for the cpu ==
> smp_processor_id() case (see the "We can unlock early" comment).
>
> Now, if that makes a difference, I think it implies a bug in the
> caller, so it's not the right fix, but it would be an interesting
> thing to test.
>
I applied the above. No difference. Applying the same change for the cpu ==
smp_processor_id() case does not make a difference either.

Guenter