On Fri, 09/11 14:22, Kevin Wolf wrote:
> Am 11.09.2015 um 13:46 hat Fam Zheng geschrieben:
> > On Fri, 09/11 12:39, Kevin Wolf wrote:
> > > Am 29.07.2015 um 06:42 hat Fam Zheng geschrieben:
> > > > v2: Switch to disable/enable model. [Paolo]
> > > >
> > > > Most existing nested aio_poll()'s in block layer are inconsiderate of
> > > > dispatching potential new r/w requests from ioeventfds and nbd exports,
> > > > which
> > > > might result in responsiveness issues (e.g. bdrv_drain_all will not
> > > > return when
> > > > new requests keep coming), or even wrong semantics (e.g.
> > > > qmp_transaction cannot
> > > > enforce atomicity due to aio_poll in bdrv_drain_all).
> > > >
> > > > Previous attampts to address this issue include new op blocker[1],
> > > > bdrv_lock[2]
> > > > and nested AioContext (patches not posted to qemu-devel).
> > > >
> > > > This approach is based on the idea proposed by Paolo Bonzini. The
> > > > original idea
> > > > is introducing "aio_context_disable_client / aio_context_enable_client
> > > > to
> > > > filter AioContext handlers according to the "client", e.g.
> > > > AIO_CLIENT_DATAPLANE (ioeventfd), AIO_CLIENT_PROTOCOL,
> > > > AIO_CLIENT_NBD_SERVER,
> > > > AIO_CLIENT_CONTEXT, ... Extend aio_set_{event_notifier,fd}_handler to
> > > > pass a
> > > > client (type)."
> > > >
> > > > After this series, block layer aio_poll() will only process those
> > > > "protocol"
> > > > fds that are used in block I/O, plus the ctx->notifier for
> > > > aio_notify(); other
> > > > aio_poll()'s keep unchanged.
> > > >
> > > > The biggest advantage over approaches [1] and [2] is, no change is
> > > > needed in
> > > > virtio-{blk,scsi}-dataplane code, also this doesn't depend on
> > > > converting QMP to
> > > > coroutines.
> > >
> > > It seems that I haven't replied on the mailing list yet, even though I
> > > think I already said this in person at KVM Forum: This series fixes only
> > > a special case of the real problem, which is that bdrv_drain/all at a
> > > single point doesn't make a lot of sense, but needs to be replaced by a
> > > whole section with exclusive access, like a bdrv_drained_begin/end pair.
> > >
> > > To be clear: Anything that works with types of users instead of
> > > individual users is bound to fall short of being a complete solution. I
> > > don't prefer partial solutions when we know there is a bigger problem.
> > >
> > > This series addresses your immediate need of protecting against new data
> > > plane requests, which it arguably achieves. The second case I always
> > > have in mind is Berto's case where he has multiple streaming block jobs
> > > in the same backing file chain [1].
> > >
> > > This involves a bdrv_reopen() of the target BDS to make it writable, and
> > > bdrv_reopen() uses bdrv_drain_all() so drivers don't have to cope with
> > > running requests while reopening themselves. It can however involve
> > > nested event loops for synchronous operations like bdrv_flush(), and if
> > > those can process completions of block jobs, which can respond by doing
> > > anything to the respective node, things can go wrong.
> >
> > Just to get a better idea of bdrv_drained_begin/end, could you explain how
> > to
> > use the pair to fix the above problem?
>
> How to use it is easy part: In bdrv_reopen_multiple(), you would replace
> the existing bdrv_drain_all() with begin and you would add the
> corresponding end right before the return statement.
OK, so that the completion of other jobs won't happen because we only complete
the relevant requests?
Does block_job_pause() work here?
>
> > > You don't solve this by adding client types (then problematic request
> > > would be PROTOCOL in your proposal and you can never exclude that), but
> > > you really need to have bdrv_drained_being/end pairs, where only
> > > requests issued in between are processed and everything else waits.
> >
> > What do you mean by "only requests issued in between are processed"? Where
> > are
> > the requests from?
>
> Generally speaking, you would have code that looks like this:
>
> bdrv_drain_begin()
> ...
> bdrv_something_synchronous()
> ...
> bdrv_drain_end()
>
> You want to process everything that is necessary for completing
> bdrv_something_synchronous(), but nothing else.
>
> The trickier question is how to implement this. I know that it's much
> easier to say that your series doesn't work than actually proposing
> something else that works...
I see the basic idea but that is still way too obscure. How would
bdrv_draind_begin/end know what is necessary in completing
bdrv_something_synchronous()?
>
> One relatively obvious answer we found when we discussed this a while
> back was some kind of a recursive CoRwLock (reader = in-flight request;
> writer = drained section), but that requires obviously that you're
> running in a coroutine if you want to do something with a drained
> request queue.
>
> I'm also not totally happy with the requirement of taking a reader lock
> more or less everywhere. But I'm not sure yet if there is a good
> alternative that can achieve the same.
We're basically trying to prevent new requests from being submitted, but
couldn't this blocking be a complication itself? A CoRwLock, if any, would be
implemented with something like a CoQueue, but considering the existing CoQueue
in BDS - throttled_reqs, aren't those what we want to keep *empty* between the
bdrv_drain_begin/end pair? Now we're talking about queueing requests to
another CoQueue, which feels contradictory to me.
Fam
