[lkml]   [2010]   [May]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [PATCH 3/4] scheduler: replace migration_thread with cpu_stop
    On Wed, May 05, 2010 at 09:28:41AM +0200, Tejun Heo wrote:
    > Hello,
    > On 05/05/2010 03:33 AM, Paul E. McKenney wrote:
    > > o Therefore, when CPU 0 queues the work for CPU 1, CPU 1
    > > loops right around and processes it. There will be no
    > > context switch on CPU 1.
    > Yes, that can happen.
    > > At first glance, this looks safe because:
    > >
    > > 1. Although there is no context switch, there (presumably)
    > > can be no RCU read-side critical sections on CPU 1 that
    > > span this sequence of events. (As far as I can see,
    > > any such RCU read-side critical section would be due
    > > to abuse of rcu_read_lock() and friends.)
    > AFAICS, this must hold; otherwise, synchronize_sched_expedited()
    > wouldn't have worked in the first place. On entry to any cpu_stop
    > function, there can be no RCU read-side critical section in progress.

    Makes sense to me!

    The actual requirement is that, on each CPU, there must have been a
    context switch between the end of the last RCU read-side critical
    section and the end of a successful return from try_stop_cpus().

    For CONFIG_TREE_PREEMPT_RCU, the guarantee required is a bit different:
    on each CPU, either that CPU must not have been in an RCU read-side
    critical section, or, if it was, there must have been a context switch
    between the time that CPU entered its RCU read-side critical section
    and the memory barrier executed within a successful try_stop_cpus().

    As near as I can tell, the current implementation does meet these
    requirements (but I do like your suggested change below).

    > > 2. CPU 1 will acquire and release stopper->lock, and
    > > further more will do an atomic_dec_and_test() in
    > > cpu_stop_signal_done(). The former is a weak
    > > guarantee, but the latter guarantees a memory
    > > barrier, so that any subsequent code on CPU 1 will
    > > be guaranteed to see changes on CPU 0 prior to the
    > > call to synchronize_sched_expedited().
    > >
    > > The guarantee required is that there will be a
    > > full memory barrier on each affected CPU between
    > > the time that try_stop_cpus() is called and the
    > > time that it returns.
    > Ah, right. I think it would be dangerous to depend on the implicit
    > barriers there. It might work today but it can easily break with
    > later implementation detail changes which seem completely unrelated.
    > Adding smp_mb() in the cpu_stop function should suffice, right? It's
    > not like the cost of smp_mb() there would mean anything anyway.

    If I understand the code correctly, this would be very good!

    Thanx, Paul

     \ /
      Last update: 2010-05-05 19:51    [W:0.022 / U:12.124 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site