Re: [Qemu-devel] [PATCH] kvm: do not flush after deleting gsi

[Asias He <asias@redhat.com>]

Hello Jan,

On 12/13/2012 11:34 PM, Jan Kiszka wrote:
> On 2012-12-13 05:55, Asias He wrote:
>> Hello Michael,
>>
>> On 12/12/2012 06:48 PM, Michael S. Tsirkin wrote:
>>> Deleting a GSI isn't necessary: it is enough
>>> to stop using it. Delay flush until an entry is used.
>>>
>>> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
>>> ---
>>>  kvm-all.c | 2 --
>>>  1 file changed, 2 deletions(-)
>>>
>>> diff --git a/kvm-all.c b/kvm-all.c
>>> index 3bc3347..fb3180d 100644
>>> --- a/kvm-all.c
>>> +++ b/kvm-all.c
>>> @@ -991,8 +991,6 @@ void kvm_irqchip_release_virq(KVMState *s, int virq)
>>>          }
>>>      }
>>>      clear_gsi(s, virq);
>>> -
>>> -    kvm_irqchip_commit_routes(s);
>>>  }
>>>  
>>>  static unsigned int kvm_hash_msi(uint32_t data)
>>>
>>
>> I tried this patch with vhost-blk with qemu-1.3.0
>> 6d6c9f59ca1b1a76ade7ad868bef191818f58819.
>>
>> Without the drop of msix_fire_vector_notifier in msix_handle_mask_update
>> hack
>> Before: ~20K IOPS
>> After:  ~35K IOPS
>>
>> With the drop of msix_fire_vector_notifier in
>> msix_handle_mask_update hack
>> Before:  ~197K IOPS
>> After:   ~197K IOPS
>>
> 
> Is the guest balancing the IRQ(s) between its vCPUs all the time?

Yes. In all of the four the cases (w/o gsi patch, w/o drop of
msix_fire_vector_notifier hack), IRQs balanced evenly in all vCPUs.


-- 
Asias