Re: [PATCH 0/9] pci-assign: 64bit MMIO + better MSI-X table support

[Shashidhar Patil <shashidhar.patil@gmail.com>]

HI Alex,
   I tried both FreeBSD stock and Intel official freebsd driver for 82599.
I tried both FreeBSD 9.0 and FreeBSD8.2.
No luck with traffic.

-Shashidhar

On Thu, Feb 9, 2012 at 10:53 PM, Alex Williamson
<alex.williamson@redhat.com> wrote:
> On Thu, 2012-02-09 at 21:53 +0530, Shashidhar Patil wrote:
>> Hi Alex,
>>     I tested your code changes for two problems
>> 1. MSIX vectors allocation for FreeBSD Guest
>>     Tested the allocation of MSI-X allocation happening for both
>> FreeBSD-9.0 (also tested 8.2) and Linux guest.
>>      Also tried multiple restart of the guests to see if the hosts
>> still maintains those allocations. With Linux
>>      as guest I could pass traffic. Though interrupt allocation worked
>> the FreeBSD could to pass traffic. The FreeBSD
>>      driver seems to be broken(atleast when running in guest).
>
> Which device/driver?  I was able to use an 82576, both physical device
> and sr-iov virtual function) with the freebsd 9.0 igb driver.
>
>> 2. Guest MSI-X smp_affinity fix.
>>      The smp_affinity configured in guest is honoured. The traffic
>> distribution is uniform. When smp affinity is configured
>>      in host and guest properly this fix provides multi queue and
>> hence multi core traffic distribution in both host and
>>      guest. The result is good packet throughput.
>>
>> Thanks for the the much needed fixes.
>
> Thanks for testing!
>
> Alex
>
>> -Shashidhar
>>
>> On Mon, Feb 6, 2012 at 10:59 PM, Alex Williamson
>> <alex.williamson@redhat.com> wrote:
>> > On Mon, 2012-02-06 at 21:25 +0530, Shashidhar Patil wrote:
>> >> HI Alex,
>> >>      I can give it a try. Please send me all patches as one file if possible.
>> >> I was about to report yet another problem with Guest MSI-X smp affinity
>> >> not being honoured by KVM while injecting interrupts. Its again in the context
>> >> of 82599 device (tried with Linux as guest). I had a hunch that the MSI-X vector
>> >> data which gets modified when SMP affinity is not tracked by the qemu_kvm.
>> >> So the interrupt injection was done in load balancing fashion. Because of this
>> >> interrupt LB in guest VCPUs I saw that packets for all the queues were processed
>> >> by one VCPU instead 2 or 4 available.
>> >> Anyway I will test the the MSI-X affinity fix also and get back with results.
>> >> And I hope 82599 send/receive works. Thanks for your help.
>> >
>> > I pushed the v2 version to github, so you can grab it from here for
>> > testing:
>> >
>> > git://github.com/awilliam/qemu-kvm.git
>> >
>> > Branch pci-assign-msix.  Thanks,
>> >
>> > Alex
>> >
>
>
>