Re: [PATCH v2] x86/hotplug: Silence APIC only after all irq's are migrated
From: Raj, Ashok
Date: Tue Aug 25 2020 - 22:50:30 EST
Hi Thomas,
On Wed, Aug 26, 2020 at 02:40:45AM +0200, Thomas Gleixner wrote:
> Ashok,
>
> On Thu, Aug 20 2020 at 17:42, Ashok Raj wrote:
> > When offlining CPUs, fixup_irqs() migrates all interrupts away from the
> > outgoing CPU to an online CPU. It's always possible the device sent an
> > interrupt to the previous CPU destination. Pending interrupt bit in IRR in
> > LAPIC identifies such interrupts. apic_soft_disable() will not capture any
> > new interrupts in IRR. This causes interrupts from device to be lost during
> > CPU offline. The issue was found when explicitly setting MSI affinity to a
> > CPU and immediately offlining it. It was simple to recreate with a USB
> > ethernet device and doing I/O to it while the CPU is offlined. Lost
> > interrupts happen even when Interrupt Remapping is enabled.
>
> New lines exist for a reason. They help to structure information. For
> the content, please see below.
Will work on that :-)
>
> > Current code does apic_soft_disable() before migrating interrupts.
> >
> > native_cpu_disable()
> > {
> > ...
> > apic_soft_disable();
> > cpu_disable_common();
> > --> fixup_irqs(); // Too late to capture anything in IRR.
> > }
> >
> > Just flipping the above call sequence seems to hit the IRR checks
> > and the lost interrupt is fixed for both legacy MSI and when
> > interrupt remapping is enabled.
>
> Seems to hit? Come on, we really want changelogs which are based on
> facts and not on assumptions.
What I intended to convay was by placing a debug trace_printk() at
fixup_irqs(), it was *indeed* observed. Before the change I never noticed
that path being covered.
Just my Inglish (Indian English) tricking you :-).
Will make them sensible in the next update.
>
> Aside of that, yes that's a really subtle one and thanks for tracking it
> down! For some reason I never looked at that ordering, but now that you
> stick it in front of me, it's pretty clear that this is the root cause.
>
> > /*
> > * Disable the local APIC. Otherwise IPI broadcasts will reach
> > * it. It still responds normally to INIT, NMI, SMI, and SIPI
> > - * messages.
> > + * messages. It's important to do apic_soft_disable() after
> > + * fixup_irqs(), because fixup_irqs() called from cpu_disable_common()
> > + * depends on IRR being set.
>
> That sentence does not make sense to me.
Right, I was just stating the obvious. Since fixup_irqs() isn't called
right in that function, it was suggested to make that connection explicit.
Your writeup below is crystal.. so will replace with what you have below.
>
> > + .... After apic_soft_disable() CPU preserves
> > + * currently set IRR/ISR but new interrupts will not set IRR.
>
> I agree with the IRR part, but ISR is simply impossible to be set in
> this situation.
You are correct. I was trying to convey what the SDM said, but its probably
irrelavant for this discussion.
>
> > + * This causes interrupts sent to outgoing CPU before completion
> > + * of IRQ migration to be lost. Check SDM Vol 3 "10.4.7.2 Local
> > + * APIC State after It Has been Software Disabled" section for more
> > + * details.
>
> Please do not use the SDM chapter number of today. It's going to be a
> different one with the next version.
>
> Something like this perhaps?
>
> /*
> * Disable the local APIC. Otherwise IPI broadcasts will reach
> * it. It still responds normally to INIT, NMI, SMI, and SIPI
> * messages.
> *
> * Disabling the APIC must happen after cpu_disable_common()
> * which invokes fixup_irqs().
> *
> * Disabling the APIC preserves already set bits in IRR, but
> * an interrupt arriving after disabling the local APIC does not
> * set the corresponding IRR bit.
> *
> * fixup_irqs() scans IRR for set bits so it can raise a not
> * yet handled interrupt on the new destination CPU via an IPI
> * but obviously it can't do so for IRR bits which are not set.
> * IOW, interrupts arriving after disabling the local APIC will
> * be lost.
> */
>
> Hmm?
>
> The changelog wants to have a corresponding update.
Will do ...
Cheers,
Ashok