Re: [Qemu-block] [RFC PATCH] coroutines: generate wrapper code

[Kevin Wolf]

Am 12.02.2019 um 04:22 hat Stefan Hajnoczi geschrieben:
> On Mon, Feb 11, 2019 at 09:38:37AM +0000, Vladimir Sementsov-Ogievskiy wrote:
> > 11.02.2019 6:42, Stefan Hajnoczi wrote:
> > > On Fri, Feb 08, 2019 at 05:11:22PM +0300, Vladimir Sementsov-Ogievskiy 
> > > wrote:
> > >> Hi all!
> > >>
> > >> We have a very frequent pattern of wrapping a coroutine_fn function
> > >> to be called from non-coroutine context:
> > >>
> > >>    - create structure to pack parameters
> > >>    - create function to call original function taking parameters from
> > >>      struct
> > >>    - create wrapper, which in case of non-coroutine context will
> > >>      create a coroutine, enter it and start poll-loop.
> > >>
> > >> Here is a draft of template code + example how it can be used to drop a
> > >> lot of similar code.
> > >>
> > >> Hope someone like it except me)
> > > 
> > > My 2 cents.  Cons:
> > > 
> > >   * Synchronous poll loops are an anti-pattern.  They block all of QEMU
> > >     with the big mutex held.  Making them easier to write is
> > >     questionable because we should aim to have as few of these as
> > >     possible.
> > 
> > Understand. Do we have a concept or a kind of target for a future to get 
> > rid of
> > these a lot of poll-loops? What is the right way? At least for block-layer?
> 
> It's non-trivial.  The nested event loop could be flattened if there was
> a mechanism to stop further activity on a specific object only (e.g.
> BlockDriverState).  That way the event loop can continue processing
> events for other objects and device emulation could continue for other
> objects.

The mechanism to stop activity on BlockDriverStates is bdrv_drain(). But
I don't see how this is related. Nested event loops aren't for stopping
concurrent activity (events related to async operations started earlier
are still processed in nested event loops), but for making progress on
the operation we're waiting for. They happen when synchronous code calls
into asynchronous code.

The way to get rid of them is making their callers async. I think we
would come a long way if we ran QMP command handlers (at least the block
related ones) and qemu-img operations in coroutines instead of blocking
while we wait for the result.

> Unfortunately there are interactions between objects like in block jobs
> that act on multiple BDSes, so it becomes even tricky.
> 
> A simple way of imagining this is to make each object an "actor"
> coroutine.  The coroutine processes a single message (request) at a time
> and yields when it needs to wait.  Callers send messages and expect
> asynchronous responses.  This model is bad for efficiency (parallelism
> is necessary) but at least it offers a sane way of thinking about
> multiple asynchronous components coordinating together.  (It's another
> way of saying, let's put everything into coroutines.)
> 
> The advantage of a flat event loop is that a hang in one object (e.g.
> I/O getting stuck in one file) doesn't freeze the entire event loop.

I think this one is more theoretical because you'll still have
dependencies between the components. blk_drain_all() isn't hanging
because the code is designed suboptimally, but because its semantics is
to wait until all requests have completed. And it's called because this
semantics is required.

Kevin

signature.asc
Description: PGP signature