Re: [PATCH 0/4] RFC/WIP: Fix scsi devices plug/unplug races w.r.t virtio-scsi iothread

[Paolo Bonzini <pbonzini@redhat.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1364 bytes --]

On 04/05/20 12:59, Stefan Hajnoczi wrote:
> Regarding drive_del, I guess the issue here is that this HMP command's
> semantics need to include not synchronize_rcu() but some kind of
> drain_call_rcu() operation as well that ensures deletion has completed?

Good idea, this would be Linux's rcu_barrier().

It would be a pity though that we have to do this instead of just having
the test rely on the DEVICE_DELETED event.

> drain_call_rcu() can be implemented by invoking call_rcu(temp,
> drain_call_rcu_cb, rcu) where drain_call_rcu_cb() sets a QemuEvent that
> the caller is waiting on. This way the caller can be sure that all
> previously queued call_rcu() callbacks have completed. call_rcu_thread()
> needs to be tweaked to avoid g_usleep() and instead use a timed wait so
> that drain_call_rcu() can immediately wake up the thread.

This was actually intentional in order to let some RCU callbacks pile up
(based on the observation, or the hope, that RCU data structures are
written rarely).  But the overall delay would be 50 ms so I don't think
it's a big deal to keep the unconditional sleep. The synchronize_rcu()
call could be on the order of 50 ms if --enable-membarrier is in use.

Another thing to care about is that call_rcu needs the iothread lock, so
you need to release it around the qemu_event_wait() call.

Paolo


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]