Re: [Qemu-devel] vhost-scsi port to v1.1.0 + MSI-X performance regression

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Jul 24, 2012 at 01:20:56PM -0700, Nicholas A. Bellinger wrote:
> On Tue, 2012-07-24 at 09:57 +0200, Jan Kiszka wrote:
> > On 2012-07-24 09:42, Nicholas A. Bellinger wrote:
> > > Hi Anthony, Stefan & QEMU folks,
> > > 
> 
> <SNIP>
> 
> > > However, thus far I've not been able to get virtio-scsi <-> tcm_vhost
> > > I/O to actually work against the latest qemu.git/master..  
> > > 
> > > So while doing a (manual) bisection w/ this series to track down the
> > > issue with qemu/master, I managed to run across something else..  With
> > > the vhost-scsi series applied, everything is working as expected up
> > > until the following commit:
> > > 
> > > commit 1523ed9e1d46b0b54540049d491475ccac7e6421
> > > Author: Jan Kiszka <jan.kiszka@siemens.com>
> > > Date:   Thu May 17 10:32:39 2012 -0300
> > > 
> > >     virtio/vhost: Add support for KVM in-kernel MSI injection
> > > 
> > > 
> > > This commit ends up triggering the following assert immediately after
> > > starting qemu with virtio-scsi <-> tcm_vhost:
> > > 
> > > qemu-system-x86_64: /usr/src/qemu.git/hw/msix.c:515:
> > >        msix_unset_vector_notifiers: Assertion `dev->msix_vector_use_notifier &&
> > >                                     dev->msix_vector_release_notifier' failed.
> > > 
> > > OK, so adding the following hack allows me to boot:
> > > 
> > > diff --git a/hw/msix.c b/hw/msix.c
> > > index 59c7a83..6036909 100644
> > > --- a/hw/msix.c
> > > +++ b/hw/msix.c
> > > @@ -511,6 +511,11 @@ void msix_unset_vector_notifiers(PCIDevice *dev)
> > >  {
> > >      int vector;
> > >  
> > > +    if (!dev->msix_vector_use_notifier && !dev->msix_vector_release_notifier) {
> > > +        printf("Hit NULL msix_unset_vector_notifiers for: %s\n", dev->name);
> > > +        return;
> > > +    }
> > > +
> > >      assert(dev->msix_vector_use_notifier &&
> > >             dev->msix_vector_release_notifier);
> > >  
> > > --
> > 
> > Can you post a backtrace from gdb?
> > 
> 
> Sure, w/o the above patch the backtrace with commit 1523ed9e1d looks
> like the following:
> 
> (gdb) run
> Starting program: /usr/src/qemu.git/x86_64-softmmu/qemu-system-x86_64 -enable-kvm -smp 2 -m 2048 -serial file:/tmp/vhost-serial.txt -hda /usr/src/qemu-vhost.git/debian_squeeze_amd64_standard-old.qcow2 -vhost-scsi id=vhost-scsi0,wwpn=naa.600140579ad21088,tpgt=1 -device virtio-scsi-pci,vhost-scsi=vhost-scsi0,event_idx=off
> [Thread debugging using libthread_db enabled]
> wwpn = "vhost-scsi0" tpgt = "1"
> [New Thread 0x7ffff45f8700 (LWP 26508)]
> [New Thread 0x7ffff3bf6700 (LWP 26509)]
> [New Thread 0x7ffff33f5700 (LWP 26510)]
> vhost_scsi_stop
> Failed to clear endpoint
> qemu-system-x86_64: /usr/src/qemu.git/hw/msix.c:515: msix_unset_vector_notifiers: Assertion `dev->msix_vector_use_notifier && dev->msix_vector_release_notifier' failed.
> 
> Program received signal SIGABRT, Aborted.
> 0x00007ffff5e8b165 in raise () from /lib/libc.so.6
> (gdb) bt
> #0  0x00007ffff5e8b165 in raise () from /lib/libc.so.6
> #1  0x00007ffff5e8df70 in abort () from /lib/libc.so.6
> #2  0x00007ffff5e842b1 in __assert_fail () from /lib/libc.so.6
> #3  0x00000000004a84a1 in msix_unset_vector_notifiers (dev=0x1463a70) at /usr/src/qemu.git/hw/msix.c:514
> #4  0x00000000004d2865 in virtio_pci_set_guest_notifiers (opaque=0x6788, assign=136)
>     at /usr/src/qemu.git/hw/virtio-pci.c:703
> #5  0x000000000062955f in vhost_dev_stop (hdev=0x126c8a8, vdev=0x1465220) at /usr/src/qemu.git/hw/vhost.c:954
> #6  0x0000000000628989 in vhost_scsi_stop (vs=0x126c890, vdev=0x1465220) at /usr/src/qemu.git/hw/vhost-scsi.c:115
> #7  0x000000000062f5c9 in virtio_scsi_set_status (vdev=0x1465220, val=<value optimized out>)
>     at /usr/src/qemu.git/hw/virtio-scsi.c:631
> #8  0x0000000000632082 in virtio_set_status (vdev=0x1465220, val=136 '\210') at /usr/src/qemu.git/hw/virtio.c:507
> #9  0x0000000000633410 in virtio_reset (opaque=0x6788) at /usr/src/qemu.git/hw/virtio.c:517


This is strange, code shows virtio_reset calls virtio_set_status
with value 0. Why is it 136 here?
Stack corruption?

We actually did have a memory corruptor with virtio scsi:

You sent patch
virtio-scsi: Fix incorrect VirtIOSCSI->cmd_vqs[0] definition

maybe the bug goes away if you apply that?

> #10 0x00000000004d30a9 in virtio_pci_reset (d=0x1463a70) at /usr/src/qemu.git/hw/virtio-pci.c:280
> #11 0x00000000004fc909 in qdev_reset_one (dev=0x6788, opaque=0x6788) at /usr/src/qemu.git/hw/qdev.c:207
> #12 0x00000000004fc670 in qdev_walk_children (dev=0x1463a70, devfn=0x4fc8f0 <qdev_reset_one>, 
>     busfn=0x4fc510 <qbus_reset_one>, opaque=0x0) at /usr/src/qemu.git/hw/qdev.c:372
> #13 0x00000000004ae43d in pci_device_reset (dev=0x6788) at /usr/src/qemu.git/hw/pci.c:163
> #14 0x00000000004ae64f in pci_bus_reset (bus=0x1415bd0) at /usr/src/qemu.git/hw/pci.c:206
> #15 0x00000000004ae699 in pcibus_reset (qbus=0x6788) at /usr/src/qemu.git/hw/pci.c:213
> #16 0x00000000004fc710 in qbus_walk_children (bus=0x1415bd0, devfn=0x4fc8f0 <qdev_reset_one>, busfn=0x6, opaque=0x0)
>     at /usr/src/qemu.git/hw/qdev.c:349
> #17 0x00000000004fc6a3 in qdev_walk_children (dev=<value optimized out>, devfn=0x4fc8f0 <qdev_reset_one>, 
>     busfn=0x4fc510 <qbus_reset_one>, opaque=0x0) at /usr/src/qemu.git/hw/qdev.c:379
> #18 0x00000000004fc745 in qbus_walk_children (bus=<value optimized out>, devfn=0x4fc8f0 <qdev_reset_one>, 
>     busfn=0x4fc510 <qbus_reset_one>, opaque=0x0) at /usr/src/qemu.git/hw/qdev.c:356
> #19 0x00000000004d5822 in qemu_system_reset (report=false) at /usr/src/qemu.git/vl.c:1412
> #20 0x00000000004d70bb in main (argc=<value optimized out>, argv=<value optimized out>, envp=<value optimized out>)
>     at /usr/src/qemu.git/vl.c:3647
> (gdb) 
> 
> 
> > Also, is there a git tree and a way to reproduce this without special
> > hardware needs?
> > 
> 
> I'll push this series + branches to demonstrate the issue into an public
> tree this afternoon.
> 
> Also, the particular backend is a Fusion-IO raw block flash device, but
> I'm pretty sure that using a TCM RAMDISK into tcm_vhost would exhibit
> the same type of behavior.  (Will double check on that shortly..)
> 
> > > 
> > > and virtio-scsi is then able to load + detect tcm_vhost LUNs as
> > > expected. 
> > > 
> > > However the random I/O performance with commit 1523ed9e1d46b is off by a
> > > couple of orders of magnitude, ~6K IOPs compared to ~60K IOPs on raw
> > > block flash using just the previous commit bdd00bdc64ba in Jan's series.
> > > 
> > > So AFAICT there appears to be a serious performance regression that is
> > > easily reproducible with that patch, which is about as far along as I've
> > > been able to diagnose yet.
> > > 
> > > Interestingly enough, virtio-scsi-raw performance does not seem to be
> > > effected AFAICT by this regression, and is still able to go ~20K IOPs
> > > with the same workload using commit 1523ed9e1d46b.  (Roughly the same as
> > > before)
> > > 
> > > Does anyone have any idea why commit 1523ed9e1d46b would be killing
> > > vhost / tcm_vhost performance so terribly, or is there something else
> > > that vhost / vhost-scsi should be doing with new code..?
> > 
> > No good idea yet, will have to look closer.
> > 
> 
> <nod>, thanks for your help Jan.  ;)
> 
> > Maybe you are somehow deassigning (via set_guest_notifiers) before
> > assigning. But that would not yet explain performance regressions. Your
> > target is exposing MSI-X, isn't it?
> > 
> 
> I believe that is correct.
> 
> --nab