Can we remove the logic of preventing MSI irq storms

Can we remove the logic of preventing MSI irq storms

[Haitao Shan]

Hi, Keir,

As you may remember (see c/s 17960), Xen implemented the logic of
preventing MSI irq storms. The reason of the IRQ storm at that time is
still unknown. But the logic is definitely needed at that time since
that NIC is the only device at my hand to test MSI.
The idea is simple: mask the second MSI interrupt when the first one
is still in processing. For HVM guests, we hooked at guest EOI write
to determime whether the first MSI is serviced already.

However, recently we find the logic has negative impact on 10G NIC
performance (assigned to guest). The logic lowers the interrupt
frequency that Xen can handle. It is a problem when the device is
generating too many interrupts as seen in this 10G NIC.

And now there is IRQ rate limit logic in Xenm which can also help to
prevent IRQ storms.

Given all the above, do you think it is time to remove the logic of
preventing MSI IRQ storm?

Shan Haitao

Re: Can we remove the logic of preventing MSI irq storms

[Keir Fraser]

On 14/07/2011 07:43, "Haitao Shan" <maillists.shan@gmail.com> wrote:

> Hi, Keir,
> 
> As you may remember (see c/s 17960), Xen implemented the logic of
> preventing MSI irq storms. The reason of the IRQ storm at that time is
> still unknown. But the logic is definitely needed at that time since
> that NIC is the only device at my hand to test MSI.
> The idea is simple: mask the second MSI interrupt when the first one
> is still in processing. For HVM guests, we hooked at guest EOI write
> to determime whether the first MSI is serviced already.
> 
> However, recently we find the logic has negative impact on 10G NIC
> performance (assigned to guest). The logic lowers the interrupt
> frequency that Xen can handle. It is a problem when the device is
> generating too many interrupts as seen in this 10G NIC.
> 
> And now there is IRQ rate limit logic in Xenm which can also help to
> prevent IRQ storms.
> 
> Given all the above, do you think it is time to remove the logic of
> preventing MSI IRQ storm?

I'd be happy to see it go.

 -- Keir

> Shan Haitao