Re: [RFC PATCH 0/3] block-copy: lock tasks and calls list

[Paolo Bonzini]

On 21/04/21 10:53, Vladimir Sementsov-Ogievskiy wrote:
> > Good point. Emanuele, can you work on ProgressMeter and 
> > SharedResource? AioTaskPool can also be converted to just use CoQueue 
> > instead of manually waking up coroutines.
>
> That would be great.
>
> I have one more question in mind:
>
> Is it effective to use CoMutex here? We are protecting only some fast 
> manipulations with data, not io path or something like that. Will simple 
> QemuMutex work better? Even if CoMutex doesn't have any overhead, I 
> don't think than if thread A wants to modify task list, but mutex is 
> held by thread B (for similar thing), there is a reason for thread A to 
> yield and do some other things: it can just wait several moments on 
> mutex while B is modifying task list..

Indeed even CoQueue primitives count as simple manipulation of data, 
because they unlock/lock the mutex while the coroutine sleeps.  So 
you're right that it would be okay to use QemuMutex as well

The block copy code that Emanuele has touched so far is all coroutine 
based.  I like using CoMutex when that is the case, because it says 
implicitly "the monitor is not involved".  But we need to see what it 
will be like when the patches are complete.

Rate limiting ends up being called by the monitor, but it will have its 
own QemuMutex so it's fine.  What's left is cancellation and 
block_copy_kick; I think that we can make qemu_co_sleep thread-safe with 
an API similar to Linux's prepare_to_wait, so a QemuMutex wouldn't be 
needed there either.

Paolo