Re: [Qemu-devel] MSI interrupt support with vioscsi.c miniport driver

[Vadim Rozenfeld <vrozenfe@redhat.com>]

On Tue, 2014-02-18 at 13:11 -0800, Nicholas A. Bellinger wrote:
> On Tue, 2014-02-18 at 13:00 -0800, Nicholas A. Bellinger wrote:
> > On Mon, 2014-02-10 at 11:05 -0800, Nicholas A. Bellinger wrote:
> > 
> > <SNIP>
> > 
> > > > > > Hi Yan,
> > > > > > 
> > > > > > So recently I've been doing some KVM guest performance comparisons
> > > > > > between the scsi-mq prototype using virtio-scsi + vhost-scsi, and
> > > > > > Windows Server 2012 with vioscsi.sys (virtio-win-0.1-74.iso) +
> > > > > > vhost-scsi using PCIe flash backend devices.
> > > > > > 
> > > > > > I've noticed that small block random performance for the MSFT guest is
> > > > > > at around ~80K IOPs with multiple vioscsi LUNs + adapters, which ends up
> > > > > > being well below what the Linux guest with scsi-mq + virtio-scsi is
> > > > > > capable of (~500K).
> > > > > > 
> > > > > > After searching through the various vioscsi registry settings, it
> > > > > > appears that MSIEnabled is being explicitly disabled (0x00000000), that
> > > > > > is different from what vioscsi.inx is currently defining:
> > > > > > 
> > > > > > [pnpsafe_pci_addreg_msix]
> > > > > > HKR, "Interrupt Management",, 0x00000010
> > > > > > HKR, "Interrupt Management\MessageSignaledInterruptProperties",, 0x00000010
> > > > > > HKR, "Interrupt Management\MessageSignaledInterruptProperties", MSISupported, 0x00010001, 0
> > > > > > HKR, "Interrupt Management\MessageSignaledInterruptProperties", MessageNumberLimit, 0x00010001, 4
> > > > > > 
> > > > > > Looking deeper at vioscsi.c code, I've noticed that MSI_SUPPORTED=0 is
> > > > > > explicitly disabled at build time in SOURCES + vioscsi.vcxproj, as well
> > > > > > as VioScsiFindAdapter() code always ends setting msix_enabled = FALSE
> > > > > > here, regardless of MSI_SUPPORTED:
> > > > > > 
> > > > > >  https://github.com/YanVugenfirer/kvm-guest-drivers-windows/blob/master/vioscsi/vioscsi.c#L340
> > > > > > 
> > > > > > Also looking at virtio_stor.c for the raw block driver, MSI_SUPPORTED=1
> > > > > > appears to be the default setting for the driver included in the offical
> > > > > > virtio-win iso builds, right..?
> > > > > > 
> > > > > > Sooo, I'd like to try enabling MSI_SUPPORTED=1 in a test vioscsi.sys
> > > > > > build of my own, but before going down the WDK development rabbit whole,
> > > > > > I'd like to better understand why you've explicitly disabled this logic
> > > > > > within vioscsi.c code to start..?
> > > > > > 
> > > > > > Is there anything that needs to be addressed / carried over from
> > > > > > virtio_stor.c in order to get MSI_SUPPORTED=1 to work with vioscsi.c
> > > > > > miniport code..?
> > > > 
> > > > Hi Nicholas,
> > > > 
> > > > I was thinking about enabling MSI in RHEL 6.6 (build 74) but for some
> > > > reasons decided to keep it disabled until adding mq support.
> > > > 
> > > > 
> > > > You definitely should be able to turn on MSI_SUPPORTED, rebuild the
> > > > driver, and switch MSISupported to 1 to make vioscsi driver working in
> > > > MSI mode.
> > > >    
> > > 
> > > Thanks for the quick response.  We'll give MSI_SUPPORTED=1 a shot over
> > > the next days with a test build on Server 2012 / Server 2008 R2 and see
> > > how things go..
> > > 
> > 
> > Just a quick update on progress.
> > 
> > I've been able to successfully build + load a unsigned vioscsi.sys
> > driver on Server 2012 with WDK 8.0.
> > 
> > Running with MSI_SUPPORTED=1 against vhost-scsi results in a significant
> > performance and efficiency gain, on the order of 100K to 225K IOPs for
> > 4K block random I/O workload, depending on read/write mix.
> > 
> 
> One other performance related question..
> 
> In vioscsi.c:VioScsiFindAdapter() code, the default setting for
> adaptExt->queue_depth ends up getting set to 32 (pageNum / 4) when
> indirect mode is enabled in the following bits:
> 
>     if(adaptExt->indirect) {
>         adaptExt->queue_depth = max(2, (pageNum / 4));
>     } else {
>         adaptExt->queue_depth = pageNum / ConfigInfo->NumberOfPhysicalBreaks - 1;
>     }
> 
> Looking at viostor/virtio_stor.c:VirtIoFindAdapter() code, the default
> setting for ->queue_depth appears to be 128 (pageNum):
> 
> #if (INDIRECT_SUPPORTED)
>     if(!adaptExt->dump_mode) {
>         adaptExt->indirect = CHECKBIT(adaptExt->features, VIRTIO_RING_F_INDIRECT_DESC);
>     }
>     if(adaptExt->indirect) {
>         adaptExt->queue_depth = pageNum;
>     }
> #else
>     adaptExt->indirect = 0;
> #endif
> 
> Is there a reason for the lower queue_depth for vioscsi vs. viostor..?

It's a horrible work around for the following bug:
https://bugzilla.redhat.com/show_bug.cgi?id=1013443

I'm going to remove it as soon as found better solution for it.

Best regards,
Vadim.


> 
> How about using min(adaptExt->scsi_config.cmd_per_lun, pageNum) instead..?
> 
> Thanks!
> 
> -nab
> 
>