Re: [PATCH 03/20] block: bdrv/blk_co_unref() for calls in coroutine context

[Kevin Wolf]

Am 27.04.2023 um 22:49 hat Paolo Bonzini geschrieben:
> Il gio 27 apr 2023, 19:00 Kevin Wolf <kwolf@redhat.com> ha scritto:
> 
> > By the way, and slightly unrelated, can vrc somehow help with finding
> > places that call coroutine wrappers without holding the AioContext lock?
> > (This results in an abort() when AIO_WAIT_WHILE() tries to unlock the
> > AioContext.) This is one of the classes of bugs we're seeing in 8.0.
> >
> 
> Seems more like a task for TSA.
> 
> Even though C TSA doesn't let you check that the *right* AioContext lock is
> taken, it can check statically that *one* such lock is taken, and in
> general I would guess it's rare for the wrong AioContext to be locked.

The problem with TSA is that you need to add annotations everywhere, it
doesn't look at the full call graph. It would be almost the same effort
as for adding the graph lock annotation. I don't think we want to do
this for a lock that we want to remove relatively soon anyway.

I'd really prefer a tool where I don't need to modify the source code,
at least not more than saying "aio_context_acquire() takes the lock",
"aio_context_release() releases the lock" and "you must have the lock in
AIO_WAIT_WHILE()".

I guess the other option is manual auditing.

Kevin