Re: iotest 030 SIGSEGV

[Paolo Bonzini]

On 14/10/21 18:14, Vladimir Sementsov-Ogievskiy wrote:
> iotest 30 failing is a long story.. And as I understand the main source 
> of all these crashes is that we do diffreent graph modifications 
> simultaneously from parallel block jobs.
>
> In past I sent RFC series with global mutext, to fix a subset of the 
> problem: 
> https://patchew.org/QEMU/20201120161622.1537-1-vsementsov@virtuozzo.com/   
> [just look at patch 5: 
> https://patchew.org/QEMU/20201120161622.1537-1-vsementsov@virtuozzo.com/20201120161622.1537-6-vsementsov@virtuozzo.com/] 

Can you explain the way they interleave, and where the job callbacks are 
yielding in the middle of graph manipulations?

The problem with a CoMutex is that plenty of graph manipulations happen 
outside coroutines, and if coroutines such as stream_co_clean yield the 
monitor can do graph manipulations of its own.

So if the solution could be "no yielding in the middle of graph 
manipulations", that would be much better.  In fact, maybe the coroutine 
API should have a way to assert "no-yield regions" (similar to how Linux 
croaks if you call a sleeping function while preemption is disabled). 
More assertions = more bugs found early.

> Not sure was it good enough to try to recover it. I didn't look close at 
> Emanuele's "block layer: split block APIs in global state and I/O". 
> Wasn't there something on protecting graph operations?

In his series, graph operations are supposed to operate from the main 
thread (which they do) but he didn't cover the case of coroutines that 
yield.

Paolo