New "timeout" api, to replace callout

[John Baldwin]

On Friday 28 December 2007 05:30:14 am Robert Watson wrote:
> 
> On Fri, 28 Dec 2007, Hans Petter Selasky wrote:
> 
> >> The reason you need to do a drain is to allow for safe destroying of the
> >> lock. Specifically, drivers tend to do this:
> >>
> >> 	FOO_LOCK(sc);
> >> 	...
> >> 	callout_stop(...);
> >> 	FOO_UNLOCK(sc);
> >> 	...
> >> 	callout_drain(...);
> >> 	...
> >> 	mtx_destroy(&sc->foo_mtx);
> >>
> >> If you don't have the drain and softclock is trying to acquire the 
backing 
> >> mutex while you have it held (before the callout_stop) then Bad Things 
can 
> >> happen if you don't do the drain.  Having the lock just "give up" doesn't 
> >> work either because if the memory containing the lock is free'd and 
> >> reinitialized such that it looks enough like a valid lock then softclock 
> >> (or its equivalent) will still try to obtain it.  Also, you need to do a 
> >> drain so it is safe to free the callout structure to prevent it from 
being 
> >> recycled and having weird races where it gets recycled and rescheduled 
but 
> >> the timer code thinks it has a pending stop for that pointer and so it 
> >> aborts the wrong instance of the timer, etc.
> >
> > I completely agree to what John Baldwin is writing. You need two 
> > stop-functions:
> >
> > xxx_stop which is non-blocking and xxx_drain which can block i.e. sleep
> >
> > BTW: The USB code in P4 uses the same semantics, due to the same reasons:
> >
> > usbd_transfer_stop() and usbd_transfer_drain()
> >
> > The only difference is that I pass an error code to the callback which 
might 
> > happen after that usbd_transfer_stop is called.
> >
> > I think that xxx_stop() and xxx_drain() is a generic approach that should 
be 
> > applied to all callback systems. Whenever you have a callback you need to 
be 
> > able to stop it and drain it.
> 
> I think the argument that Poul-Henning is making is not that you don't need 
> something that behaves "like drain", but rather, we're like the wait for 
drain 
> to be a short, mutex-length wait rather than a long, msleep-length wait. 
> Remember that the bodies of callouts are expected to run in a very short 
> period of time in order to not stall the timer system, in fact, in such a 
way 
> that a mutex could be held over the entirely timeout call.  Given that this 
is 
> the case, one might reasonably expect callout_stop() to perform the drain 
> rather than having a separate call.  Such a model would be very advantageous 
> in TCP, where rather than having to defer GC'ing the inpcb/tcpcb to a GC 
> worker thread, which we don't do but should, we could use the stop call 
safely 
> and eliminate a whole class of races from the stack.

The problem is if softclock() (or similar replacement in future) was preempted 
and hasn't run yet but has already gotten to the point that callout_drain() 
has to block, then sitting in a spin loop waiting for softclock() to 
acknowledge the stop isn't very optimal.  The amount of time you are asleep 
in this case is actually very small, and you probably won't even block the 
vast majority of the time if you follow the 'lock / stop / unlock / drain' 
model.  You only sleep when you lose the race and softclock() has chosen to 
run your callout and is waiting for the driver/client/whatever's lock.

-- 
John Baldwin