On Sun, Apr 21, 2019 at 10:50 PM Jason Wang <jasow...@redhat.com> wrote:
>
>
> On 2019/4/17 上午2:46, Dan Streetman wrote:
> > From: Dan Streetman <ddstr...@canonical.com>
> >
> > Buglink: https://launchpad.net/bugs/1823458
> >
> > There is a race condition when using the vhost-user driver, between a guest
> > shutdown and the vhost-user interface being closed. This is explained in
> > more detail at the bug link above; the short explanation is the vhost-user
> > device can be closed while the main thread is in the middle of stopping
> > the vhost_net. In this case, the main thread handling shutdown will
> > enter virtio_net_vhost_status() and move into the n->vhost_started (else)
> > block, and call vhost_net_stop(); while it is running that function,
> > another thread is notified that the vhost-user device has been closed,
> > and (indirectly) calls into virtio_net_vhost_status() also.
>
>
> I think we need figure out why there are multiple vhost_net_stop() calls
> simultaneously. E.g vhost-user register fd handlers like:
>
> qemu_chr_fe_set_handlers(&s->chr, NULL, NULL,
> net_vhost_user_event, NULL, nc0->name,
> NULL,
> true);
>
> which uses default main context, so it should only be called only in
> main thread.
net_vhost_user_event() schedules chr_closed_bh() to do its bottom half
work; does aio_bh_schedule_oneshot() execute its events from the main
thread?
For reference, the call chain is:
chr_closed_bh()
qmp_set_link()
nc->info->link_status_changed() -> virtio_net_set_link_status()
virtio_net_set_status()
virtio_net_vhost_status()
>
> Thanks
>
>
> > Since the
> > vhost_net status hasn't yet changed, the second thread also enters
> > the n->vhost_started block, and also calls vhost_net_stop(). This
> > causes problems for the second thread when it tries to stop the network
> > that's already been stopped.
> >
> > This adds a flag to the struct that's atomically set to prevent more than
> > one thread from calling vhost_net_stop(). The atomic_fetch_inc() is likely
> > overkill and probably could be done with a simple check-and-set, but
> > since it's a race condition there would still be a (very, very) small
> > window without using an atomic to set it.
> >
> > Signed-off-by: Dan Streetman <ddstr...@canonical.com>
> > ---
> > hw/net/virtio-net.c | 3 ++-
> > include/hw/virtio/virtio-net.h | 1 +
> > 2 files changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
> > index ffe0872fff..d36f50d5dd 100644
> > --- a/hw/net/virtio-net.c
> > +++ b/hw/net/virtio-net.c
> > @@ -13,6 +13,7 @@
> >
> > #include "qemu/osdep.h"
> > #include "qemu/iov.h"
> > +#include "qemu/atomic.h"
> > #include "hw/virtio/virtio.h"
> > #include "net/net.h"
> > #include "net/checksum.h"
> > @@ -240,7 +241,7 @@ static void virtio_net_vhost_status(VirtIONet *n,
> > uint8_t status)
> > "falling back on userspace virtio", -r);
> > n->vhost_started = 0;
> > }
> > - } else {
> > + } else if (atomic_fetch_inc(&n->vhost_stopped) == 0) {
> > vhost_net_stop(vdev, n->nic->ncs, queues);
> > n->vhost_started = 0;
> > }
> > diff --git a/include/hw/virtio/virtio-net.h b/include/hw/virtio/virtio-net.h
> > index b96f0c643f..d03fd933d0 100644
> > --- a/include/hw/virtio/virtio-net.h
> > +++ b/include/hw/virtio/virtio-net.h
> > @@ -164,6 +164,7 @@ struct VirtIONet {
> > uint8_t nouni;
> > uint8_t nobcast;
> > uint8_t vhost_started;
> > + int vhost_stopped;
> > struct {
> > uint32_t in_use;
> > uint32_t first_multi;
