lkml.org 
[lkml]   [2009]   [Oct]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC] Userspace RCU: (ab)using futexes to save cpu cycles and energy
On Sun, Oct 04, 2009 at 10:37:45AM -0400, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > On Wed, Sep 23, 2009 at 01:48:20PM -0400, Mathieu Desnoyers wrote:
> > > Hi,
> > >
> > > When implementing the call_rcu() "worker thread" in userspace, I ran
> > > into the problem that it had to be woken up periodically to check if
> > > there are any callbacks to execute. However, I easily imagine that this
> > > does not fit well with the "green computing" definition.
> > >
> > > Therefore, I've looked at ways to have the call_rcu() callers waking up
> > > this worker thread when callbacks are enqueued. However, I don't want to
> > > take any lock and the fast path (when no wake up is required) should not
> > > cause any cache-line exchange.
> > >
> > > Here are the primitives I've created. I'd like to have feedback on my
> > > futex use, just to make sure I did not do any incorrect assumptions.
> > >
> > > This could also be eventually used in the QSBR Userspace RCU quiescent
> > > state and in mb/signal userspace RCU when exiting RCU read-side C.S. to
> > > ensure synchronize_rcu() does not busy-wait for too long.
> > >
> > > /*
> > > * Wake-up any waiting defer thread. Called from many concurrent threads.
> > > */
> > > static void wake_up_defer(void)
> > > {
> > > if (unlikely(atomic_read(&defer_thread_futex) == -1))
> > > atomic_set(&defer_thread_futex, 0);
> > > futex(&defer_thread_futex, FUTEX_WAKE,
> > > 0, NULL, NULL, 0);
> > > }
> > >
> > > /*
> > > * Defer thread waiting. Single thread.
> > > */
> > > static void wait_defer(void)
> > > {
> > > atomic_dec(&defer_thread_futex);
> > > if (atomic_read(&defer_thread_futex) == -1)
> > > futex(&defer_thread_futex, FUTEX_WAIT, -1,
> > > NULL, NULL, 0);
> > > }
> >
> > The standard approach would be to use pthread_cond_wait() and
> > pthread_cond_broadcast(). Unfortunately, this would require holding a
> > pthread_mutex_lock across both operations, which would not necessarily
> > be so good for wake-up-side scalability.
>
> The pthread_cond_broadcast() mutex is really a bugger when it comes to
> execute it at each rcu_read_unlock(). We could as well use a mutex to
> protect the whole read-side.. :-(
>
> > That said, without this sort of heavy-locking approach, wakeup races
> > are quite difficult to avoid.
>
> I did a formal model of my futex-based wait/wakeup. The main idea is
> that the waiter:
>
> - Set itself to "waiting"
> - Checks the "real condition" for which it will wait (e.g. queues empty
> when used for rcu callbacks, no more ongoing old reader thread C.S.
> when used in synchronize_rcu())
> - Calls sys_futex if the variable have not changed.
>
> And the waker:
> - sets the "real condition" waking up the waiter (enqueuing, or
> rcu_read_unlock())
> - check if the waiter must be woken up, if so, wake it up by setting the
> state to "running" and calling sys_futex.
>
> But as you say, wakeup races are difficult (but not impossible!) to
> avoid. This is why I resorted to a formal model of the wait/wakeup
> scheme to ensure that we cannot end up in a situation where a waker
> races with the waiter and does not wake it up when it should. This is
> nothing fancy (does not model memory and instruction reordering
> automatically), but I figure that memory barriers are required between
> almost every steps of this algorithm, so by adding smp_mb() I end up
> ensure sequential behavior. I added test cases in the model to ensure
> that incorrect memory reordering _would_ cause errors by doing the
> reordering by hand in error-injection runs.

My question is whether pthread_cond_wait() and pthread_cond_broadcast()
can substitute for the raw call to futex. Unless I am missing something
(which I quite possibly am), the kernel will serialize on the futex
anyway, so serialization in user-mode code does not add much additional
pain.

> The model is available at:
> http://www.lttng.org/cgi-bin/gitweb.cgi?p=userspace-rcu.git;a=tree;f=futex-wakeup;h=4ddeaeb2784165cb0465d4ca9f7d27acb562eae3;hb=refs/heads/formal-model
>
> (this is in the formal-model branch of the urcu tree, futex-wakeup
> subdir)
>
> This is modeling this snippet of code :
>
> static int defer_thread_futex;
>
> /*
> * Wake-up any waiting defer thread. Called from many concurrent threads.
> */
> static void wake_up_defer(void)
> {
> if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
> uatomic_set(&defer_thread_futex, 0);
> futex(&defer_thread_futex, FUTEX_WAKE, 1,
> NULL, NULL, 0);
> }
> }
>
> static void enqueue(void *callback) /* not the actual types */
> {
> add_to_queue(callback);
> smp_mb();
> wake_up_defer();
> }
>
> /*
> * rcu_defer_num_callbacks() returns the total number of callbacks
> * enqueued.
> */
>
> /*
> * Defer thread waiting. Single thread.
> */
> static void wait_defer(void)
> {
> uatomic_dec(&defer_thread_futex);
> smp_mb(); /* Write futex before read queue */
> if (rcu_defer_num_callbacks()) {
> smp_mb(); /* Read queue before write futex */
> /* Callbacks are queued, don't wait. */
> uatomic_set(&defer_thread_futex, 0);
> } else {
> smp_rmb(); /* Read queue before read futex */
> if (uatomic_read(&defer_thread_futex) == -1)
> futex(&defer_thread_futex, FUTEX_WAIT, -1,
> NULL, NULL, 0);
> }
> }
>
>
> Comments are welcome,

I will take a look after further recovery from jetlag. Not yet competent
to review this kind of stuff. Give me a few days. ;-)

Thanx, Paul


\
 
 \ /
  Last update: 2009-10-04 22:43    [W:0.101 / U:0.132 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site