Re: [PATCH 5/5] Fix kdump under KVM.

[Chris Lalancette <clalance@redhat.com>]

Marcelo Tosatti wrote:
> On Wed, Oct 28, 2009 at 11:21:42AM +0100, Chris Lalancette wrote:
>> Marcelo Tosatti wrote:
>>> On Tue, Oct 27, 2009 at 05:41:07PM +0100, Chris Lalancette wrote:
>>>> This patch is the main point of the series.  In order for
>>>> kdump to properly work inside a KVM guest, we need to make
>>>> sure that all VCPUs in virtual wire APIC mode get kicked
>>>> to try and pick up the timer interrupts.  To do this,
>>>> we iterate over the CPUs and deliver interrupts to the
>>>> proper VCPUs.
>>>>
>>>> I don't love the concept of doing kvm_irq_kick_vcpus() from
>>>> within pit_timer_fn().  A PIT is not connected to a CPU at all,
>>>> only to a PIC or APIC.  However, if a CPU enters idle, this is
>>>> the only way to wake it up to check for the interrupt.
>>> The reason the PIT interrupt was fixed to BSP is:
>>>
>>> http://www.mail-archive.com/kvm-devel@lists.sourceforge.net/msg13250.html
>>>
>>> Perhaps enhancing ioapic_deliver's (irq == 0) check to bypass the       
>>> destination overwrite in case its programmed by the guest to 
>>> a single CPU would fix it?
>> Ug, nasty.  I definitely don't want to re-introduce that bug.  What exactly do
>> you mean by "fix it"?  Do you mean fix the original RHEL-5.1 PAE issue, or fix
>> the kdump issue?
> 
> The kdump issue. Something like:
> 
> #ifdef CONFIG_X86
>         /* Always delivery PIT interrupt to vcpu 0 */
>         if (irq == 0 && dest_multiple_vcpus(entry)) {
>                 irqe.dest_mode = 0; /* Physical mode. */
>                 /* need to read apic_id from apic regiest since
>                  * it can be rewritten */
>                 irqe.dest_id = ioapic->kvm->bsp_vcpu->vcpu_id;
>         }
> #endif
> 
> The rest of the code should be ready to deal with it.

Do you have any more information about how to reproduce that original issue?  I
tried installing a RHEL-5.1 i386 guest with a PAE kernel, and I was able to
install and boot it just fine with or without the "always deliver irq 0 to vcpu
0" check above.  Either there's something in particular I need to do to trigger
the issue (hardware or software), or it's being solved another way.

In the meantime, I've gotten the "set_irq from IRQ context" that Avi suggested
working, and the fixing up of this IOAPIC check is the last bit to actually get
kdump working.

-- 
Chris Lalancette