Re: [PATCH V3 15/23] xen/arm: Stick around in leave_hypervisor_to_guest until I/O has completed

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Wed, 9 Dec 2020, Julien Grall wrote:
> On 09/12/2020 23:35, Stefano Stabellini wrote:
> > On Wed, 9 Dec 2020, Stefano Stabellini wrote:
> > > On Mon, 30 Nov 2020, Oleksandr Tyshchenko wrote:
> > > > From: Oleksandr Tyshchenko <oleksandr_tyshchenko@xxxxxxxx>
> > > > 
> > > > This patch adds proper handling of return value of
> > > > vcpu_ioreq_handle_completion() which involves using a loop
> > > > in leave_hypervisor_to_guest().
> > > > 
> > > > The reason to use an unbounded loop here is the fact that vCPU
> > > > shouldn't continue until an I/O has completed. In Xen case, if an I/O
> > > > never completes then it most likely means that something went horribly
> > > > wrong with the Device Emulator. And it is most likely not safe to
> > > > continue. So letting the vCPU to spin forever if I/O never completes
> > > > is a safer action than letting it continue and leaving the guest in
> > > > unclear state and is the best what we can do for now.
> > > > 
> > > > This wouldn't be an issue for Xen as do_softirq() would be called at
> > > > every loop. In case of failure, the guest will crash and the vCPU
> > > > will be unscheduled.
> > > 
> > > Imagine that we have two guests: one that requires an ioreq server and
> > > one that doesn't. If I am not mistaken this loop could potentially spin
> > > forever on a pcpu, thus preventing any other guest being scheduled, even
> > > if the other guest doesn't need any ioreq servers.
> > > 
> > > 
> > > My other concern is that we are busy-looping. Could we call something
> > > like wfi() or do_idle() instead? The ioreq server event notification of
> > > completion should wake us up?
> > > 
> > > Following this line of thinking, I am wondering if instead of the
> > > busy-loop we should call vcpu_block_unless_event_pending(current) in
> > > try_handle_mmio if IO_RETRY. Then when the emulation is done, QEMU (or
> > > equivalent) calls xenevtchn_notify which ends up waking up the domU
> > > vcpu. Would that work?
> > 
> > I read now Julien's reply: we are already doing something similar to
> > what I suggested with the following call chain:
> > 
> > check_for_vcpu_work -> vcpu_ioreq_handle_completion -> wait_for_io ->
> > wait_on_xen_event_channel
> > 
> > So the busy-loop here is only a safety-belt in cause of a spurious
> > wake-up, in which case we are going to call again check_for_vcpu_work,
> > potentially causing a guest reschedule.
> > 
> > Then, this is fine and addresses both my concerns. Maybe let's add a note
> > in the commit message about it.
> 
> Damm, I hit the "sent" button just a second before seen your reply. :/ Oh
> well. I suggested the same because I have seen the same question multiple
> time.

:-)

 
> > I am also wondering if there is any benefit in calling wait_for_io()
> > earlier, maybe from try_handle_mmio if IO_RETRY?
> 
> wait_for_io() may end up to deschedule the vCPU. I would like to avoid this to
> happen in the middle of the I/O emulation because we need to happen it without
> lock held at all.
> 
> I don't think there are locks involved today, but the deeper in the call stack
> the scheduling happens, the more chance we may screw up in the future.
> 
> However...
> 
> > leave_hypervisor_to_guest is very late for that.
> 
> ... I am not sure what's the problem with that. The IOREQ will be notified of
> the pending I/O as soon as try_handle_mmio() put the I/O in the shared page.
> 
> If the IOREQ server is running on a different pCPU, then it might be possible
> that the I/O has completed before reached leave_hypervisor_to_guest(). In this
> case, we would not have to wait for the I/O.

Yeah, I was thinking about that too. Actually it could be faster
this way we end up being lucky.

The reason for moving it earlier would be that by the time
leave_hypervisor_to_guest is called "Xen" has already decided to
continue running this particular vcpu. If we called wait_for_io()
earlier, we would give important information to the scheduler before any
decision is made. This is more "philosophical" than practical though.
Let's leave it as is.