Re: [Qemu-devel] [PATCH for-2.11 1/4] Revert "coroutine: abort if we try to schedule or enter a pending coroutine"

[Kevin Wolf <kwolf@redhat.com>]

Am 28.11.2017 um 17:18 hat Paolo Bonzini geschrieben:
> On 28/11/2017 16:43, Kevin Wolf wrote:
> > This reverts commit 6133b39f3c36623425a6ede9e89d93175fde15cd.
> > 
> > The commit checked conditions that would expose a bug, but there is no
> > real reason to forbid them apart from the bug, which we'll fix in a
> > minute.
> > 
> > In particular, reentering a coroutine during co_aio_sleep_ns() is fine;
> > the function is explicitly written to allow this.
> 
> This is true.
> 
> > aio_co_schedule() can indeed conflict with direct coroutine invocations,
> > but this is exactky what we want to fix, so remove that check again,
> > too.
> 
> I'm not sure this is a good idea, as I answered in patch 3.
> 
> It can also conflict badly with another aio_co_schedule().  Your patch
> here removes the assertion in this case, and patch 3 makes it easier to
> get into the situation where two aio_co_schedule()s conflict with each
> other.

I don't see how they conflict. If the second aio_co_schedule() comes
before the coroutine is actually entered, they are effectively simply
merged into a single one. Which is exactly what was intended.

> For example, say you have a coroutine that calls aio_co_schedule on
> itself, like
> 
> 	while (true) {
> 		aio_co_schedule(qemu_get_current_aio_context(),
> 				qemu_coroutine_self());
> 	}
> 
> If somebody else calls qemu_coroutine_enter on this coroutine, *that* is
> the bug.  These patches would just cause some random corruption or
> (perhaps worse) hang.

Obviously not every coroutine is made to be reentered from multiple
places, so for some cases it just might not make a whole lot of sense.
Coroutines that are made for it generally are one of the types I
explained in the commit message of patch 3.

But anyway, how would this cause corruption or a hang (apart from the
fact that this example doesn't have any state that could even be
corrupted)? The external qemu_coroutine_enter() would just replace the
scheduled coroutine call, so the coroutine wouldn't even notice that it
was called from qemu_coroutine_enter() rather than its own scheduled
call.

Kevin