Re: [Qemu-devel] [PATCH] KVM: Fix GSI number space limit

[Alexander Graf <agraf@suse.de>]

On 06.06.14 15:23, Cornelia Huck wrote:
> On Fri, 06 Jun 2014 15:15:54 +0200
> Alexander Graf <agraf@suse.de> wrote:
>
>> On 06.06.14 15:12, Cornelia Huck wrote:
>>> On Fri,  6 Jun 2014 14:46:05 +0200
>>> Alexander Graf <agraf@suse.de> wrote:
>>>
>>>> KVM tells us the number of GSIs it can handle inside the kernel. That value is
>>>> basically KVM_MAX_IRQ_ROUTES. However when we try to set the GSI mapping table,
>>>> it checks for
>>>>
>>>>       r = -EINVAL;
>>>>       if (routing.nr >= KVM_MAX_IRQ_ROUTES)
>>>>           goto out;
>>>>
>>>> erroring out even when we're only using all of the GSIs. To make sure we never
>>>> hit that limit, let's reduce the number of GSIs we get from KVM by one.
>>>>
>>>> Signed-off-by: Alexander Graf <agraf@suse.de>
>>>> ---
>>>>    kvm-all.c | 2 +-
>>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/kvm-all.c b/kvm-all.c
>>>> index 4e19eff..56a251b 100644
>>>> --- a/kvm-all.c
>>>> +++ b/kvm-all.c
>>>> @@ -938,7 +938,7 @@ void kvm_init_irq_routing(KVMState *s)
>>>>    {
>>>>        int gsi_count, i;
>>>>
>>>> -    gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
>>>> +    gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
>>>>        if (gsi_count > 0) {
>>>>            unsigned int gsi_bits, i;
>>>>
>>> But gsi_count is already marked as used further down in this function,
>>> isn't it? Confused.
>>     gsi_bits = ALIGN(gsi_count, 32);
>> [...]
>>           for (i = gsi_count; i < gsi_bits; i++) {
>>               set_gsi(s, i);
>>           }
>>
>> So if you take gsi_count = 1024, what happens?
>>
>>     gsi_count = 1024;
>>     gsi_bits = 1024;
>>     for (i = 1024; i < 1024; i++) {
>>               set_gsi(s, i);
>>     }
>>
>> At least in my world of C that loop never runs, no?
>>
> But then kvm_irqchip_get_virq() should never return 1024, shouldn't it?

Right, because it returns the virq number which starts at 0. However, to 
describe all virqs from [0..1023] we need 1024 entries which the kernel 
errors out on.

>
> And:
>
> void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
> {
> [...]
>      assert(pin < s->gsi_count);
>
> would trigger too early with your change, wouldn't it?

Not really - with my change we only support 1023 virqs. So the biggest 
virq number is 1022 which is < 1023 :).


Sorry for describing this with actual numbers - I find it easier to 
grasp when I think in concrete numbers here - this stuff is just really 
spinning my head :).

Alex