[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [RFC PATCH] vhost_net: should not use max_queue_pairs for non-mq gue
|
From: |
Jason Wang |
|
Subject: |
Re: [RFC PATCH] vhost_net: should not use max_queue_pairs for non-mq guest |
|
Date: |
Fri, 25 Mar 2022 15:59:26 +0800 |
On Fri, Mar 25, 2022 at 3:02 PM Si-Wei Liu <si-wei.liu@oracle.com> wrote:
>
>
>
> On 3/21/2022 8:47 PM, Jason Wang wrote:
> > On Sat, Mar 19, 2022 at 12:14 PM Si-Wei Liu <si-wei.liu@oracle.com> wrote:
> >> With MQ enabled vdpa device and non-MQ supporting guest e.g.
> >> booting vdpa with mq=on over OVMF of single vqp, it's easy
> >> to hit assert failure as the following:
> >>
> >> ../hw/virtio/vhost-vdpa.c:560: vhost_vdpa_get_vq_index: Assertion `idx >=
> >> dev->vq_index && idx < dev->vq_index + dev->nvqs' failed.
> >>
> >> 0 0x00007f8ce3ff3387 in raise () at /lib64/libc.so.6
> >> 1 0x00007f8ce3ff4a78 in abort () at /lib64/libc.so.6
> >> 2 0x00007f8ce3fec1a6 in __assert_fail_base () at /lib64/libc.so.6
> >> 3 0x00007f8ce3fec252 in () at /lib64/libc.so.6
> >> 4 0x0000558f52d79421 in vhost_vdpa_get_vq_index (dev=<optimized out>,
> >> idx=<optimized out>) at ../hw/virtio/vhost-vdpa.c:563
> >> 5 0x0000558f52d79421 in vhost_vdpa_get_vq_index (dev=<optimized out>,
> >> idx=<optimized out>) at ../hw/virtio/vhost-vdpa.c:558
> >> 6 0x0000558f52d7329a in vhost_virtqueue_mask (hdev=0x558f55c01800,
> >> vdev=0x558f568f91f0, n=2, mask=<optimized out>) at
> >> ../hw/virtio/vhost.c:1557
> >> 7 0x0000558f52c6b89a in virtio_pci_set_guest_notifier
> >> (d=d@entry=0x558f568f0f60, n=n@entry=2, assign=assign@entry=true,
> >> with_irqfd=with_irqfd@entry=false)
> >> at ../hw/virtio/virtio-pci.c:974
> >> 8 0x0000558f52c6c0d8 in virtio_pci_set_guest_notifiers (d=0x558f568f0f60,
> >> nvqs=3, assign=true) at ../hw/virtio/virtio-pci.c:1019
> >> 9 0x0000558f52bf091d in vhost_net_start (dev=dev@entry=0x558f568f91f0,
> >> ncs=0x558f56937cd0, data_queue_pairs=data_queue_pairs@entry=1,
> >> cvq=cvq@entry=1)
> >> at ../hw/net/vhost_net.c:361
> >> 10 0x0000558f52d4e5e7 in virtio_net_set_status (status=<optimized out>,
> >> n=0x558f568f91f0) at ../hw/net/virtio-net.c:289
> >> 11 0x0000558f52d4e5e7 in virtio_net_set_status (vdev=0x558f568f91f0,
> >> status=15 '\017') at ../hw/net/virtio-net.c:370
> >> 12 0x0000558f52d6c4b2 in virtio_set_status
> >> (vdev=vdev@entry=0x558f568f91f0, val=val@entry=15 '\017') at
> >> ../hw/virtio/virtio.c:1945
> >> 13 0x0000558f52c69eff in virtio_pci_common_write (opaque=0x558f568f0f60,
> >> addr=<optimized out>, val=<optimized out>, size=<optimized out>) at
> >> ../hw/virtio/virtio-pci.c:1292
> >> 14 0x0000558f52d15d6e in memory_region_write_accessor (mr=0x558f568f19d0,
> >> addr=20, value=<optimized out>, size=1, shift=<optimized out>,
> >> mask=<optimized out>, attrs=...)
> >> at ../softmmu/memory.c:492
> >> 15 0x0000558f52d127de in access_with_adjusted_size (addr=addr@entry=20,
> >> value=value@entry=0x7f8cdbffe748, size=size@entry=1,
> >> access_size_min=<optimized out>, access_size_max=<optimized out>,
> >> access_fn=0x558f52d15cf0 <memory_region_write_accessor>,
> >> mr=0x558f568f19d0, attrs=...) at ../softmmu/memory.c:554
> >> 16 0x0000558f52d157ef in memory_region_dispatch_write
> >> (mr=mr@entry=0x558f568f19d0, addr=20, data=<optimized out>, op=<optimized
> >> out>, attrs=attrs@entry=...)
> >> at ../softmmu/memory.c:1504
> >> 17 0x0000558f52d078e7 in flatview_write_continue
> >> (fv=fv@entry=0x7f8accbc3b90, addr=addr@entry=103079215124, attrs=...,
> >> ptr=ptr@entry=0x7f8ce6300028, len=len@entry=1, addr1=<optimized out>,
> >> l=<optimized out>, mr=0x558f568f19d0) at
> >> /home/opc/qemu-upstream/include/qemu/host-utils.h:165
> >> 18 0x0000558f52d07b06 in flatview_write (fv=0x7f8accbc3b90,
> >> addr=103079215124, attrs=..., buf=0x7f8ce6300028, len=1) at
> >> ../softmmu/physmem.c:2822
> >> 19 0x0000558f52d0b36b in address_space_write (as=<optimized out>,
> >> addr=<optimized out>, attrs=..., buf=buf@entry=0x7f8ce6300028,
> >> len=<optimized out>)
> >> at ../softmmu/physmem.c:2914
> >> 20 0x0000558f52d0b3da in address_space_rw (as=<optimized out>,
> >> addr=<optimized out>, attrs=...,
> >> attrs@entry=..., buf=buf@entry=0x7f8ce6300028, len=<optimized out>,
> >> is_write=<optimized out>) at ../softmmu/physmem.c:2924
> >> 21 0x0000558f52dced09 in kvm_cpu_exec (cpu=cpu@entry=0x558f55c2da60) at
> >> ../accel/kvm/kvm-all.c:2903
> >> 22 0x0000558f52dcfabd in kvm_vcpu_thread_fn (arg=arg@entry=0x558f55c2da60)
> >> at ../accel/kvm/kvm-accel-ops.c:49
> >> 23 0x0000558f52f9f04a in qemu_thread_start (args=<optimized out>) at
> >> ../util/qemu-thread-posix.c:556
> >> 24 0x00007f8ce4392ea5 in start_thread () at /lib64/libpthread.so.0
> >> 25 0x00007f8ce40bb9fd in clone () at /lib64/libc.so.6
> >>
> >> The cause for the assert failure is due to that the vhost_dev index
> >> for the ctrl vq was not aligned with actual one in use by the guest.
> >> Upon multiqueue feature negotiation in virtio_net_set_multiqueue(),
> >> if guest doesn't support multiqueue, the guest vq layout would shrink
> >> to single queue pair of 3 vqs in total (rx, tx and ctrl). This results
> >> in ctrl_vq taking a different vhost_dev group index than the default
> >> n->max_queue_pairs, the latter of which is only valid for multiqueue
> >> guest. While on those additional vqs not exposed to the guest,
> >> vhost_net_set_vq_index() never populated vq_index properly, hence
> >> getting the assert failure.
> >>
> >> A possible fix is to pick the correct vhost_dev group for the control
> >> vq according to this table [*]:
> >>
> >> vdpa tool / QEMU arg / guest config / ctrl_vq group index
> >> ----------------------------------------------------------------
> >> max_vqp 8 / mq=on / mq=off (UEFI) => data_queue_pairs
> >> max_vqp 8 / mq=on / mq=on (Linux) => n->max_queue_pairs(>1)
> >> max_vqp 8 / mq=off / mq=on (Linux) => n->max_queue_pairs(=1)
> >>
> >> [*] Please see FIXME in the code for open question and discussion
> >>
> >> Signed-off-by: Si-Wei Liu <si-wei.liu@oracle.com>
> >> ---
> >> hw/net/vhost_net.c | 13 +++++++++----
> >> hw/virtio/vhost-vdpa.c | 25 ++++++++++++++++++++++++-
> >> 2 files changed, 33 insertions(+), 5 deletions(-)
> >>
> >> diff --git a/hw/net/vhost_net.c b/hw/net/vhost_net.c
> >> index 30379d2..9a4479b 100644
> >> --- a/hw/net/vhost_net.c
> >> +++ b/hw/net/vhost_net.c
> >> @@ -322,6 +322,7 @@ int vhost_net_start(VirtIODevice *dev, NetClientState
> >> *ncs,
> >> BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(dev)));
> >> VirtioBusState *vbus = VIRTIO_BUS(qbus);
> >> VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
> >> + bool mq = virtio_host_has_feature(dev, VIRTIO_NET_F_MQ);
> >> int total_notifiers = data_queue_pairs * 2 + cvq;
> >> VirtIONet *n = VIRTIO_NET(dev);
> >> int nvhosts = data_queue_pairs + cvq;
> >> @@ -343,7 +344,7 @@ int vhost_net_start(VirtIODevice *dev, NetClientState
> >> *ncs,
> >> if (i < data_queue_pairs) {
> >> peer = qemu_get_peer(ncs, i);
> >> } else { /* Control Virtqueue */
> >> - peer = qemu_get_peer(ncs, n->max_queue_pairs);
> >> + peer = qemu_get_peer(ncs, mq ? data_queue_pairs :
> >> n->max_queue_pairs);
> >> }
> >>
> >> net = get_vhost_net(peer);
> >> @@ -368,7 +369,7 @@ int vhost_net_start(VirtIODevice *dev, NetClientState
> >> *ncs,
> >> if (i < data_queue_pairs) {
> >> peer = qemu_get_peer(ncs, i);
> >> } else {
> >> - peer = qemu_get_peer(ncs, n->max_queue_pairs);
> >> + peer = qemu_get_peer(ncs, mq ? data_queue_pairs :
> >> n->max_queue_pairs);
> >> }
> >> r = vhost_net_start_one(get_vhost_net(peer), dev);
> >>
> >> @@ -390,7 +391,10 @@ int vhost_net_start(VirtIODevice *dev, NetClientState
> >> *ncs,
> >>
> >> err_start:
> >> while (--i >= 0) {
> >> - peer = qemu_get_peer(ncs , i);
> >> + if (mq)
> >> + peer = qemu_get_peer(ncs, i < data_queue_pairs ? i :
> >> data_queue_pairs);
> >> + else
> >> + peer = qemu_get_peer(ncs, i < data_queue_pairs ? i :
> >> n->max_queue_pairs);
> >> vhost_net_stop_one(get_vhost_net(peer), dev);
> >> }
> >> e = k->set_guest_notifiers(qbus->parent, total_notifiers, false);
> >> @@ -409,6 +413,7 @@ void vhost_net_stop(VirtIODevice *dev, NetClientState
> >> *ncs,
> >> VirtioBusState *vbus = VIRTIO_BUS(qbus);
> >> VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
> >> VirtIONet *n = VIRTIO_NET(dev);
> >> + bool mq = virtio_host_has_feature(dev, VIRTIO_NET_F_MQ);
> >> NetClientState *peer;
> >> int total_notifiers = data_queue_pairs * 2 + cvq;
> >> int nvhosts = data_queue_pairs + cvq;
> >> @@ -418,7 +423,7 @@ void vhost_net_stop(VirtIODevice *dev, NetClientState
> >> *ncs,
> >> if (i < data_queue_pairs) {
> >> peer = qemu_get_peer(ncs, i);
> >> } else {
> >> - peer = qemu_get_peer(ncs, n->max_queue_pairs);
> >> + peer = qemu_get_peer(ncs, mq ? data_queue_pairs :
> >> n->max_queue_pairs);
> >> }
> >> vhost_net_stop_one(get_vhost_net(peer), dev);
> >> }
> >> diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
> >> index 27ea706..623476e 100644
> >> --- a/hw/virtio/vhost-vdpa.c
> >> +++ b/hw/virtio/vhost-vdpa.c
> >> @@ -1097,7 +1097,30 @@ static int vhost_vdpa_dev_start(struct vhost_dev
> >> *dev, bool started)
> >> vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
> >> }
> >>
> >> - if (dev->vq_index + dev->nvqs != dev->vq_index_end) {
> >> + /* FIXME the vhost_dev group for the control vq may have bogus nvqs=2
> >> + * value rather than nvqs=1. This can happen in case the guest doesn't
> >> + * support multiqueue, as a result of
> >> virtio_net_change_num_queue_pairs()
> >> + * destroying and rebuilding all the vqs, the guest index for control
> >> vq
> >> + * will no longer align with the host's. Currently
> >> net_init_vhost_vdpa()
> >> + * only initializes all vhost_dev's and net_clients once during
> >> + * net_client_init1() time, way earlier before multiqueue feature
> >> + * negotiation can kick in.
> > See below, it looks like the code doesn't find the correct vhost_dev.
> >
> >> + *
> >> + * Discussion - some possible fixes so far I can think of:
> >> + *
> >> + * option 1: fix vhost_net->dev.nvqs and nc->is_datapath in place for
> >> + * vdpa's ctrl vq, or rebuild all vdpa's vhost_dev groups and the
> >> + * net_client array, in the virtio_net_set_multiqueue() path;
> >> + *
> >> + * option 2: fix vhost_dev->nvqs in place at vhost_vdpa_set_features()
> >> + * before coming down to vhost_vdpa_dev_start() (Q: nc->is_datapath
> >> + * seems only used in virtio_net_device_realize, is it relevant?);
> > Relevant but not directly related, for the vhost_dev where
> > nc->is_datapath is false, it will assume it is backed by a single
> > queue not a queue pair.
> >
> >> + *
> >> + * option 3: use host queue index all along in vhost-vdpa ioctls
> >> instead
> >> + * of using guest vq index, so that vhost_net_start/stop() can remain
> >> + * as-is today
> >> + */
> > Note that the vq_index of each vhost_dev is assigned during
> > vhost_net_start() according to whether or not the MQ or CVQ is
> > negotiated in vhost_net_start()
> >
> > for (i = 0; i < nvhosts; i++) {
> >
> > if (i < data_queue_pairs) {
> > peer = qemu_get_peer(ncs, i);
> > } else { /* Control Virtqueue */
> > peer = qemu_get_peer(ncs, n->max_queue_pairs);
> > }
> >
> > net = get_vhost_net(peer);
> > vhost_net_set_vq_index(net, i * 2, index_end);
> >
> > It means some of the peers won't be used when MQ is not negotiated. So
> > it looks to me the evil came from virtio_net_get_notifier_mask().
> Yes, there it is. Where the control virtqueue first ever needs a
> guest_notifier for vhost_dev.
> > Where it doesn't mask the correct vhost dev when the guest doesn't
> > support MQ but the host does. So we had option 4:
> >
> > diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
> > index 2087516253..5e9ac019cd 100644
> > --- a/hw/net/virtio-net.c
> > +++ b/hw/net/virtio-net.c
> > @@ -3179,7 +3179,13 @@ static void
> > virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,
> > bool mask)
> > {
> > VirtIONet *n = VIRTIO_NET(vdev);
> > - NetClientState *nc = qemu_get_subqueue(n->nic, vq2q(idx));
> > + NetClientState *nc;
> > +
> > + if (!virtio_vdev_has_feature(vdev, VIRTIO_NET_F_MQ) && idx == 2) {
> Hmmm, I thought it would be more natural to align the layout of
> vhost_dev's with that of virtqueue's, not the other way around.
The problem is that we need to make sure it works for vhost_net as
well where it doesn't support cvq.
> Not sure
> how this vhost_dev selection scheme may work with additional queues
> discovered through transport specific mechanism, such as the admin
> virtqueue, but I can live with it for now:
>
> --- a/hw/net/virtio-net.c
> +++ b/hw/net/virtio-net.c
> @@ -244,7 +244,8 @@ static void virtio_net_vhost_status(VirtIONet *n,
> uint8_t status)
> VirtIODevice *vdev = VIRTIO_DEVICE(n);
> NetClientState *nc = qemu_get_queue(n->nic);
> int queue_pairs = n->multiqueue ? n->max_queue_pairs : 1;
> - int cvq = n->max_ncs - n->max_queue_pairs;
> + int cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ?
> + n->max_ncs - n->max_queue_pairs : 0;
Any reason for this line?
Btw, would you mind to post a formal patch for this?
Thanks
>
> if (!get_vhost_net(nc->peer)) {
> return;
> @@ -3161,8 +3162,14 @@ static NetClientInfo net_virtio_info = {
> static bool virtio_net_guest_notifier_pending(VirtIODevice *vdev, int idx)
> {
> VirtIONet *n = VIRTIO_NET(vdev);
> - NetClientState *nc = qemu_get_subqueue(n->nic, vq2q(idx));
> + NetClientState *nc;
> assert(n->vhost_started);
> + if (!virtio_vdev_has_feature(vdev, VIRTIO_NET_F_MQ) && idx == 2) {
> + assert(virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ));
> + nc = qemu_get_subqueue(n->nic, n->max_queue_pairs);
> + } else {
> + nc = qemu_get_subqueue(n->nic, vq2q(idx));
> + }
> return vhost_net_virtqueue_pending(get_vhost_net(nc->peer), idx);
> }
>
> @@ -3170,8 +3177,14 @@ static void
> virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,
> bool mask)
> {
> VirtIONet *n = VIRTIO_NET(vdev);
> - NetClientState *nc = qemu_get_subqueue(n->nic, vq2q(idx));
> + NetClientState *nc;
> assert(n->vhost_started);
> + if (!virtio_vdev_has_feature(vdev, VIRTIO_NET_F_MQ) && idx == 2) {
> + assert(virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ));
> + nc = qemu_get_subqueue(n->nic, n->max_queue_pairs);
> + } else {
> + nc = qemu_get_subqueue(n->nic, vq2q(idx));
> + }
> vhost_net_virtqueue_mask(get_vhost_net(nc->peer),
> vdev, idx, mask);
> }
>
>
> Thanks,
> -Siwei
>
> > + nc = qemu_get_subqueue(n->nic, n->max_queue_pairs);
> > + } else {
> > + nc = qemu_get_subqueue(n->nic, vq2q(idx));
> > + }
> > assert(n->vhost_started);
> > vhost_net_virtqueue_mask(get_vhost_net(nc->peer),
> > vdev, idx, mask);
> >
> > Thanks
> >
> >> + if (dev->vq_index + dev->nvqs < dev->vq_index_end) {
> >> return 0;
> >> }
> >>
> >> --
> >> 1.8.3.1
> >>
>