Re: [Qemu-devel] [big lock] Discussion about the convention of device's DMA each other after breaking down biglock

[Paolo Bonzini]

Il 19/09/2012 12:06, Avi Kivity ha scritto:
>> > (The hunch is that ts could be deleted exactly at the moment the
>> > callback is unlocked.  This can be solved with ref/unref on the opaque
>> > value, as you mention below).
> Are you saying that this works as is or not?  It does seem broken wrt
> deletion; after qemu_del_timer() completes the caller expects the
> callback not to be called.

Ouch, then I think the only solution is to remove this invariant if you
add fine-grained locking and re-check the enable conditions in the timer
callback.

If you allow calling del_timer to be called with the device lock held,
you cannot fixing without deadlocking.  If you don't, the caller of
del_timer cannot set any expectation because the timer could be run in
the window between dropping the device lock and calling del_timer

Paolo

> This isn't trivial to guarantee, we need something like
> 
>    dispatch_timer():
>       pending += 1
>       timer.ref()
>       drop lock
>       timer.cb()
>       take lock
>       timer.unref()
>       pending -= 1
>       notify
> 
>    del_timer():
>       take lock
>       timer.unlink()
>       while pending:
>           wait for notification
>       drop lock
> 
> but, if del_timer is called with the device lock held, we deadlock. ugh.

Wait for notification should be a cond_wait that drops the lock.

Alternatively:

    dispatch_timer():
       drop lock
       if timer has expired:
           timer.cb()
       take lock
       notify

    del_timer():
       take lock
       timer.unlink()
       drop lock