Re: [Qemu-devel] [PATCH 10/12] block: Drain throttling queue with BdrvChild callback

[Kevin Wolf <kwolf@redhat.com>]

Am 23.03.2016 um 22:29 hat Paolo Bonzini geschrieben:
> 
> 
> On 22/03/2016 16:33, Kevin Wolf wrote:
> > This removes the last part of I/O throttling from block/io.c and moves
> > it to the BlockBackend.
> > 
> > When draining the queue of a BlockDriverState, we must make sure that no
> > new requests can come in for it. Request sources from outside the block
> > layer are disabled with aio_disable_external(), but the throttling queue
> > must be handled separately.
> 
> I have looked at the strategy we talked about today to implement request
> cancellation (so that e.g. system reset doesn't take ages because of
> throttled requests).  While that may be a worthwhile addition anyway, I
> think throttling bdrv_drain() may impose an excessive cost for cases
> such as live migration.  The risk of the guest using bdrv_drain() to
> game throttling is low enough that we can keep on disabling throttling
> during bdrv_drain().

I think your cancellation series (allows to) gets rid of most if not all
blk_drain() callers in the device emulation, so it becomes harder for
guests to trigger one. Ideally only the monitor should allow triggering
a drain.

On the other hand, your other series introduces bdrv_drain() calls where
we have open-coded nested event loops waiting for a single request
today. I'm pretty sure that these can be triggered by the guest and that
throttling the drain would be desirable therefore. Maybe we need a
different function there, and maybe we can even retain the behaviour
that it doesn't unnecessarily flush everything instead of just waiting
for the completion of a single request.

> So for now I think we can merge the two series just fine.  The strategy
> I used in my patch, adding bdrv_no_throttling_begin and
> bdrv_no_throttling_end around the bdrv_drain loop, can be adapted just
> as use BdrvChildRole callbacks ->drained_begin and ->drained_end.

Okay. Actually, such a pair of callbacks - not only into the
BlockBackend, but from there into the guest device - was a thought
already when we introduced aio_disable_external(). Do you think it would
make sense to change things in the mid term so that the users of a
BlockBackend just get drain_begin/end callbacks?

> I will post v3 of my series tomorrow, adopting your patch 1/12 of this
> series and removing the recursion on bdrv_no_throttling_begin and
> bdrv_no_throttling_end, which is unnecessary.

Okay, I'll try to rebase then.

Kevin