Re: [PATCH v2 16/21] virtio-net: Use replay_schedule_bh_event for bhs that affect machine state

[Nicholas Piggin]

On Thu Aug 15, 2024 at 3:25 AM AEST, Alex Bennée wrote:
> "Michael S. Tsirkin" <mst@redhat.com> writes:
>
> > On Wed, Aug 14, 2024 at 04:05:34PM +1000, Nicholas Piggin wrote:
> >> On Wed Aug 14, 2024 at 6:48 AM AEST, Michael S. Tsirkin wrote:
> >> > On Tue, Aug 13, 2024 at 09:23:24PM +0100, Alex Bennée wrote:
> >> > > From: Nicholas Piggin <npiggin@gmail.com>
> >> > > 
> >> > > The regular qemu_bh_schedule() calls result in non-deterministic
> >> > > execution of the bh in record-replay mode, which causes replay failure.
> >> > > 
> >> > > Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
> >> > > Reviewed-by: Pavel Dovgalyuk <Pavel.Dovgalyuk@ispras.ru>
> >> > > Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> >> > > Message-Id: <20240813050638.446172-9-npiggin@gmail.com>
> >> > > Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
> >> > > ---
> >> > >  hw/net/virtio-net.c | 11 ++++++-----
> >> > >  1 file changed, 6 insertions(+), 5 deletions(-)
> >> > > 
> >> > > diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
> >> > > index 08aa0b65e3..10ebaae5e2 100644
> >> > > --- a/hw/net/virtio-net.c
> >> > > +++ b/hw/net/virtio-net.c
> >> > > @@ -40,6 +40,7 @@
> >> > >  #include "migration/misc.h"
> >> > >  #include "standard-headers/linux/ethtool.h"
> >> > >  #include "sysemu/sysemu.h"
> >> > > +#include "sysemu/replay.h"
> >> > >  #include "trace.h"
> >> > >  #include "monitor/qdev.h"
> >> > >  #include "monitor/monitor.h"
> >> > > @@ -417,7 +418,7 @@ static void virtio_net_set_status(struct 
> >> > > VirtIODevice *vdev, uint8_t status)
> >> > >                  timer_mod(q->tx_timer,
> >> > >                                 qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) 
> >> > > + n->tx_timeout);
> >> > >              } else {
> >> > > -                qemu_bh_schedule(q->tx_bh);
> >> > > +                replay_bh_schedule_event(q->tx_bh);
> >> > >              }
> >> > >          } else {
> >> > >              if (q->tx_timer) {
> >> > > @@ -2672,7 +2673,7 @@ static void 
> >> > > virtio_net_tx_complete(NetClientState *nc, ssize_t len)
> >> > >           */
> >> > >          virtio_queue_set_notification(q->tx_vq, 0);
> >> > >          if (q->tx_bh) {
> >> > > -            qemu_bh_schedule(q->tx_bh);
> >> > > +            replay_bh_schedule_event(q->tx_bh);
> >> > >          } else {
> >> > >              timer_mod(q->tx_timer,
> >> > >                        qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 
> >> > > n->tx_timeout);
> >> > > @@ -2838,7 +2839,7 @@ static void virtio_net_handle_tx_bh(VirtIODevice 
> >> > > *vdev, VirtQueue *vq)
> >> > >          return;
> >> > >      }
> >> > >      virtio_queue_set_notification(vq, 0);
> >> > > -    qemu_bh_schedule(q->tx_bh);
> >> > > +    replay_bh_schedule_event(q->tx_bh);
> >> > >  }
> >> > >  
> >> > >  static void virtio_net_tx_timer(void *opaque)
> >> > > @@ -2921,7 +2922,7 @@ static void virtio_net_tx_bh(void *opaque)
> >> > >      /* If we flush a full burst of packets, assume there are
> >> > >       * more coming and immediately reschedule */
> >> > >      if (ret >= n->tx_burst) {
> >> > > -        qemu_bh_schedule(q->tx_bh);
> >> > > +        replay_bh_schedule_event(q->tx_bh);
> >> > >          q->tx_waiting = 1;
> >> > >          return;
> >> > >      }
> >> > > @@ -2935,7 +2936,7 @@ static void virtio_net_tx_bh(void *opaque)
> >> > >          return;
> >> > >      } else if (ret > 0) {
> >> > >          virtio_queue_set_notification(q->tx_vq, 0);
> >> > > -        qemu_bh_schedule(q->tx_bh);
> >> > > +        replay_bh_schedule_event(q->tx_bh);
> >> > >          q->tx_waiting = 1;
> >> > >      }
> >> > >  }
> >> > > -- 
> >> > > 2.39.2
> >> >
> >> >
> >> > Is this really the only way to fix this? I do not think
> >> > virtio has any business knowing about replay.
> >> > What does this API do, even? BH but not broken with replay?
> >> > Do we ever want replay broken? Why not fix qemu_bh_schedule?
> >> > And when we add another feature which we do not want to break
> >> > will we do foo_bar_replay_bh_schedule_event or what?
> >> 
> >> I agree with you. We need to do this (a couple of other hw
> >> subsystems already do and likely some are still broken vs
> >> replay and would need to be converted), but I think it's
> >> mostly a case of bad naming. You're right the caller should
> >> not know about replay at all, what it should be is whether
> >> the event is for the target machine or the host harness,
> >> same as timers are VIRTUAL / HOST.
> >> So I think we just need to make a qemu_bh_schedule_<type>,
> >> or qemu_bh_scheudle_event(... QEMU_EVENT_VIRTUAL/HOST/etc).
> >
> > Or just pass QEMUClockType?

Could be a good idea. Although I'm not sure what to do with
all types, maybe we can restrict what is supported.

> Is this wider re-factoring something that can wait for the next
> developer cycle?

I would say so. It's not quite trivial to do nicely since
things are a bit tangled between util/async and replay.

> >> I had started on a conversion once but not completed it.
> >> I could resurrect if there is agreement on the API?
>
> I would certainly welcome it being cleaned up. The supported replay
> devices are very piecemeal at the moment.

I'll tidy up and post an RFC for how the new API might look.

Thanks,
Nick