Re: [RFC] sched: implement the exclusive wait queue as a LIFO queue

[Jamie Lokier <jamie@shareable.org>]

Changli Gao wrote:
> On Wed, Apr 28, 2010 at 9:21 PM, Jamie Lokier <jamie@shareable.org> wrote:
> > Changli Gao wrote:
> >>
> >> fs/eventpoll.c: 1443.
> >>                 wait.flags |= WQ_FLAG_EXCLUSIVE;
> >>                 __add_wait_queue(&ep->wq, &wait);
> >
> > The same thing about assumptions applies here.  The userspace process
> > may be waiting for an epoll condition to get access to a resource,
> > rather than being a worker thread interchangeable with others.
> 
> Oh, the lines above are the current ones. So the assumptions applies
> and works here.

No, because WQ_FLAG_EXCLUSIVE doesn't have your LIFO semantic at the moment.

Your patch changes the behaviour of epoll, though I don't know if it
matters.  Perhaps all programs which have multiple tasks waiting on
the same epoll fd are "interchangeable worker thread" types anyway :-)

> > For example, userspace might be using a pipe as a signal-safe lock, or
> > signal-safe multi-token semaphore, and epoll to wait for that pipe.
> >
> > WQ_FLAG_EXCLUSIVE means there is no point waking all tasks, to avoid a
> > pointless thundering herd.  It doesn't mean unfairness is ok.
> 
> The users should not make any assumption about the waking up sequence,
> neither LIFO nor FIFO.

Correct, but they should be able to assume non-starvation (eventual
progress) for all waiters.

It's one of those subtle things, possibly a unixy thing: Non-RT tasks
should always make progress when the competition is just other non-RT
tasks, even if the progress is slow.

Starvation can spread out beyond the starved process, to cause
priority inversions in other tasks that are waiting on a resource
locked by the starved process.  Among other things, that can cause
higher priority tasks, and RT priority tasks, to block permanently.
Very unpleasant.

> > The LIFO idea _might_ make sense for interchangeable worker-thread
> > situations - including userspace.  It would make sense for pipe
> > waiters, socket waiters (especially accept), etc.
> 
> Yea, and my following patches are for socket waiters.

Occasionally unix socketpairs are occasionally used in the above ways too.

I'm not against your patch, but I worry that starvation is a new
semantic, and it may have a significant effect on something - either
in the kernel, or in userspace which is harder to check.

> > Do you have any measurements which showing the LIFO mode performing
> > better than FIFO, and by how much?
> 
> I didn't do any test yet. But some work done by LSE project years ago
> showed that it is better.
> 
> http://lse.sourceforge.net/io/aionotes.txt
> 
> " Also in view of
> better cache utilization the wake queue mechanism is LIFO by default.
> (A new exclusive LIFO wakeup option has been introduced for this purpose)"

I suspect it's possible to combine LIFO-ish and FIFO-ish queuing to
prevent starvation while getting some of the locality benefit.
Something like add-LIFO and increment a small counter in the next wait
entry, but never add in front of an entry whose counter has reached
MAX_LIFO_WAITERS? :-)

-- Jamie