Re: Getting the new RxRPC patches upstream

[David Howells <dhowells@redhat.com>]

Oleg Nesterov <oleg@tv-sign.ru> wrote:

> > > We only care when del_timer() returns true. In that case, if the timer
> > > function still runs (possible for single-threaded wqs), it has already
> > > passed __queue_work().
> > 
> > Why do you assume that?

Sorry, I should have been more clear.  I meant the assumption that we only
care about a true return from del_timer().

> If del_timer() returns true, the timer was pending. This means it was
> started by work->func() (note that __run_timers() clears timer_pending()
> before calling timer->function). This in turn means that
> delayed_work_timer_fn() has already called __queue_work(dwork), otherwise
> work->func() has no chance to run.

But if del_timer() returns 0, then there may be a problem.  We can't tell the
difference between the following two cases:

 (1) The timer hadn't been started.

 (2) The timer had been started, has expired and is no longer pending, but
     another CPU is running its handler routine.

try_to_del_timer_sync() _does_, however, distinguish between these cases: the
first is the 0 return, the second is the -1 return, and the case where it
dequeued the timer is the 1 return.

BTW, can a timer handler be preempted?  I assume not...  But it can be delayed
by interrupt processing.

David