Re: [PATCH 2/3] [BUGFIX] x86/x86_64: fix CPU offlining triggered inactive device IRQ interrruption

From: Yinghai Lu
Date: Thu Apr 09 2009 - 18:38:22 EST


On Thu, Apr 9, 2009 at 12:17 PM, Gary Hade <garyhade@xxxxxxxxxx> wrote:
> On Wed, Apr 08, 2009 at 04:59:35PM -0700, Yinghai Lu wrote:
>> On Wed, Apr 8, 2009 at 4:58 PM, Yinghai Lu <yhlu.kernel@xxxxxxxxx> wrote:
>> > On Wed, Apr 8, 2009 at 4:37 PM, Gary Hade <garyhade@xxxxxxxxxx> wrote:
>> >> On Wed, Apr 08, 2009 at 03:30:15PM -0700, Yinghai Lu wrote:
>> >>> On Wed, Apr 8, 2009 at 2:07 PM, Gary Hade <garyhade@xxxxxxxxxx> wrote:
>> >>> > Impact: Eliminates a race that can leave the system in an
>> >>> >        unusable state
>> >>> >
>> >>> > During rapid offlining of multiple CPUs there is a chance
>> >>> > that an IRQ affinity move destination CPU will be offlined
>> >>> > before the IRQ affinity move initiated during the offlining
>> >>> > of a previous CPU completes.  This can happen when the device
>> >>> > is not very active and thus fails to generate the IRQ that is
>> >>> > needed to complete the IRQ affinity move before the move
>> >>> > destination CPU is offlined.  When this happens there is an
>> >>> > -EBUSY return from __assign_irq_vector() during the offlining
>> >>> > of the IRQ move destination CPU which prevents initiation of
>> >>> > a new IRQ affinity move operation to an online CPU.  This
>> >>> > leaves the IRQ affinity set to an offlined CPU.
>> >>> >
>> >>> > I have been able to reproduce the problem on some of our
>> >>> > systems using the following script.  When the system is idle
>> >>> > the problem often reproduces during the first CPU offlining
>> >>> > sequence.
>> >>> >
>> >>> > #!/bin/sh
>> >>> >
>> >>> > SYS_CPU_DIR=/sys/devices/system/cpu
>> >>> > VICTIM_IRQ=25
>> >>> > IRQ_MASK=f0
>> >>> >
>> >>> > iteration=0
>> >>> > while true; do
>> >>> >  echo $iteration
>> >>> >  echo $IRQ_MASK > /proc/irq/$VICTIM_IRQ/smp_affinity
>> >>> >  for cpudir in $SYS_CPU_DIR/cpu[1-9] $SYS_CPU_DIR/cpu??; do
>> >>> >    echo 0 > $cpudir/online
>> >>> >  done
>> >>> >  for cpudir in $SYS_CPU_DIR/cpu[1-9] $SYS_CPU_DIR/cpu??; do
>> >>> >    echo 1 > $cpudir/online
>> >>> >  done
>> >>> >  iteration=`expr $iteration + 1`
>> >>> > done
>> >>> >
>> >>> > The proposed fix takes advantage of the fact that when all
>> >>> > CPUs in the old domain are offline there is nothing to be done
>> >>> > by send_cleanup_vector() during the affinity move completion.
>> >>> > So, we simply avoid setting cfg->move_in_progress preventing
>> >>> > the above mentioned -EBUSY return from __assign_irq_vector().
>> >>> > This allows initiation of a new IRQ affinity move to a CPU
>> >>> > that is not going offline.
>> >>> >
>> >>> > Signed-off-by: Gary Hade <garyhade@xxxxxxxxxx>
>> >>> >
>> >>> > ---
>> >>> >  arch/x86/kernel/apic/io_apic.c |   11 ++++++++---
>> >>> >  1 file changed, 8 insertions(+), 3 deletions(-)
>> >>> >
>> >>> > Index: linux-2.6.30-rc1/arch/x86/kernel/apic/io_apic.c
>> >>> > ===================================================================
>> >>> > --- linux-2.6.30-rc1.orig/arch/x86/kernel/apic/io_apic.c        2009-04-08 09:23:00.000000000 -0700
>> >>> > +++ linux-2.6.30-rc1/arch/x86/kernel/apic/io_apic.c     2009-04-08 09:23:16.000000000 -0700
>> >>> > @@ -363,7 +363,8 @@ set_extra_move_desc(struct irq_desc *des
>> >>> >        struct irq_cfg *cfg = desc->chip_data;
>> >>> >
>> >>> >        if (!cfg->move_in_progress) {
>> >>> > -               /* it means that domain is not changed */
>> >>> > +               /* it means that domain has not changed or all CPUs
>> >>> > +                * in old domain are offline */
>> >>> >                if (!cpumask_intersects(desc->affinity, mask))
>> >>> >                        cfg->move_desc_pending = 1;
>> >>> >        }
>> >>> > @@ -1262,8 +1263,11 @@ next:
>> >>> >                current_vector = vector;
>> >>> >                current_offset = offset;
>> >>> >                if (old_vector) {
>> >>> > -                       cfg->move_in_progress = 1;
>> >>> >                        cpumask_copy(cfg->old_domain, cfg->domain);
>> >>> > +                       if (cpumask_intersects(cfg->old_domain,
>> >>> > +                                              cpu_online_mask)) {
>> >>> > +                               cfg->move_in_progress = 1;
>> >>> > +                       }
>> >>> >                }
>> >>> >                for_each_cpu_and(new_cpu, tmp_mask, cpu_online_mask)
>> >>> >                        per_cpu(vector_irq, new_cpu)[vector] = irq;
>> >>> > @@ -2492,7 +2496,8 @@ static void irq_complete_move(struct irq
>> >>> >                if (likely(!cfg->move_desc_pending))
>> >>> >                        return;
>> >>> >
>> >>> > -               /* domain has not changed, but affinity did */
>> >>> > +               /* domain has not changed or all CPUs in old domain
>> >>> > +                * are offline, but affinity changed */
>> >>> >                me = smp_processor_id();
>> >>> >                if (cpumask_test_cpu(me, desc->affinity)) {
>> >>> >                        *descp = desc = move_irq_desc(desc, me);
>> >>> > --
>> >>>
>> >>> so you mean during __assign_irq_vector(), cpu_online_mask get updated?
>> >>
>> >> No, the CPU being offlined is removed from cpu_online_mask
>> >> earlier via a call to remove_cpu_from_maps() from
>> >> cpu_disable_common().  This happens just before fixup_irqs()
>> >> is called.
>> >>
>> >>> with your patch, how about that it just happen right after you check
>> >>> that second time.
>> >>>
>> >>> it seems we are missing some lock_vector_lock() on the remove cpu from
>> >>> online mask.
>> >>
>> >> The remove_cpu_from_maps() call in cpu_disable_common() is vector
>> >> lock protected:
>> >> void cpu_disable_common(void)
>> >> {
>> >>               < snip >
>> >>        /* It's now safe to remove this processor from the online map */
>> >>        lock_vector_lock();
>> >>        remove_cpu_from_maps(cpu);
>> >>        unlock_vector_lock();
>> >>        fixup_irqs();
>> >> }
>> >
>> >
>> > __assign_irq_vector always has vector_lock locked...
>
> OK, I see the 'vector_lock' spin_lock_irqsave/spin_unlock_irqrestore
> surrounding the __assign_irq_vector call in assign_irq_vector.
>
>> > so cpu_online_mask will not changed during,
>
> I understand that this 'vector_lock' acquisition prevents
> multiple simultaneous executions of __assign_irq_vector but
> does that really prevent another thread executing outside
> __assign_irq_vector (or outside other 'vector_lock' serialized
> code) from modifying cpu_online_mask?
>
> Isn't it really 'cpu_add_remove_lock' (also held when
> __assign_irq_vector() is called in the context of a CPU add
> or remove) that is used for this purpose?
>
>> > why do you need to check that again in __assign_irq_vector ?
>
> Because that is where the cfg->move_in_progress flag was
> being set.
>
> Is there some reason that the content of cpu_online_mask
> cannot be trusted at this location?
>
> If all the CPUs in the old domain are offline doesn't
> that imply that we got to that location in response to
> a CPU offline request?
>
>> >
>> looks like you need to clear move_in_progress in fixup_irqs()
>
> This would be a difficult since I believe the code is
> currently partitioned in a manner that prevents access to
> irq_cfg records from functions defined in arch/x86/kernel/irq_32.c
> and arch/x86/kernel/irq_64.c.  It also doesn't feel right to
> allow cfg->move_in_progress to be set in __assign_irq_vector
> and then clear it in fixup_irqs().

it looks before fixup_irqs() cpu_online_mask get updated, and before
irq_complete_move get called.

so we could fixup_irqs to clear move_in_progress and cleanup percpu
vector_irq ...

YH
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/