Re: [Xen-devel] [PATCH] xen: Send spinlock IPI to all waiters

["Jan Beulich" <JBeulich@xxxxxxxx>]

(resending with proper list address)

>>> On 15.02.13 at 11:50, Stefan Bader <stefan.bader@xxxxxxxxxxxxx> wrote:
> Hopefully not mis-parsing Jan's last comments on the other thread,
> this would be the fix covering things until a better implementation
> is done.
> This also prevents the hang on older kernels, where it could be re-
> produced reliably.
> 
> -Stefan
> 
> From 7e042a253b06da96409a0e059744c217f396a17f Mon Sep 17 00:00:00 2001
> From: Stefan Bader <stefan.bader@xxxxxxxxxxxxx>
> Date: Fri, 15 Feb 2013 09:48:52 +0100
> Subject: [PATCH] xen: Send spinlock IPI to all waiters
> 
> There is a loophole between Xen's current implementation of
> pv-spinlocks and the scheduler. This was triggerable through
> a testcase until v3.6 changed the TLB flushing code. The
> problem potentially is still there just not observable in the
> same way.
> 
> What could happen was (is):
> 
> 1. CPU n tries to schedule task x away and goes into a slow
>    wait for the runq lock of CPU n-# (must be one with a lower
>    number).
> 2. CPU n-#, while processing softirqs, tries to balance domains
>    and goes into a slow wait for its own runq lock (for updating
>    some records). Since this is a spin_lock_irqsave in softirq
>    context, interrupts will be re-enabled for the duration of
>    the poll_irq hypercall used by Xen.
> 3. Before the runq lock of CPU n-# is unlocked, CPU n-1 receives
>    an interrupt (e.g. endio) and when processing the interrupt,
>    tries to wake up task x. But that is in schedule and still
>    on_cpu, so try_to_wake_up goes into a tight loop.
> 4. The runq lock of CPU n-# gets unlocked, but the message only
>    gets sent to the first waiter, which is CPU n-# and that is
>    busily stuck.
> 
> To avoid this and since the unlocking code has no real sense of
> which waiter is best suited to grab the lock, just send the IPI
> to all of them. This causes the waiters to return from the hyper-
> call (those not interrupted at least) and do active spinlocking.
> 
> BugLink: http://bugs.launchpad.net/bugs/1011792 
> 
> Signed-off-by: Stefan Bader <stefan.bader@xxxxxxxxxxxxx>

Acked-by: Jan Beulich <jbeulich@xxxxxxxx>

(but see below)

> Cc: stable@xxxxxxxxxxxxxxx 

(back to 3.0; earlier kernels shouldn't be affected)

> ---
>  arch/x86/xen/spinlock.c |    1 -
>  1 file changed, 1 deletion(-)
> 
> diff --git a/arch/x86/xen/spinlock.c b/arch/x86/xen/spinlock.c
> index 83e866d..f7a080e 100644
> --- a/arch/x86/xen/spinlock.c
> +++ b/arch/x86/xen/spinlock.c
> @@ -328,7 +328,6 @@ static noinline void xen_spin_unlock_slow(struct 
> xen_spinlock *xl)
>               if (per_cpu(lock_spinners, cpu) == xl) {
>                       ADD_STATS(released_slow_kicked, 1);
>                       xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
> -                     break;

Of course that's the minimal fix. Would be nice if this could be done
as a batch, since even if congestion on locks might be rare, it's
specifically that case where the biggest gain from doing things
quickly can be had.

But of course, as said earlier, switching to ticket locks should be the
real goal.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel