[lkml]   [2005]   [Mar]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [patch] Real-Time Preemption, -RT-2.6.12-rc1-V0.7.41-07
    On Wed, Mar 23, 2005 at 07:33:17AM +0100, Ingo Molnar wrote:
    > * Paul E. McKenney <> wrote:
    > > +#ifdef CONFIG_PREEMPT_RCU
    > > +
    > > +void rcu_read_lock(void)
    > > +{
    > > + if (current->rcu_read_lock_nesting++ == 0) {
    > > + current->rcu_data = &get_cpu_var(rcu_data);
    > > + atomic_inc(&current->rcu_data->active_readers);
    > > + put_cpu_var(rcu_data);
    > >
    > > Need an smp_mb() here for non-x86 CPUs. Otherwise, the CPU can
    > > re-order parts of the critical section to precede the rcu_read_lock().
    > > Could precede the put_cpu_var(), but why increase latency?
    > ok. It's enough to put a barrier into the else branch here, because the
    > atomic op in the main brain is a barrier by itself.

    For x86, the atomic op is a barrier, but not for other architectures.
    You don't need a barrier in the else branch, because in that case
    you are already in an enclosing RCU read-side critical section, so
    any bleeding of code will be into this enclosing section, thus still

    > > +void rcu_read_unlock(void)
    > > +{
    > [...]
    > > And need an smp_mb() here, again for non-x86 CPUs.
    > ok.
    > > Assuming that the memory barriers are added, I can see a bunch of ways
    > > for races to extend grace periods, but none so far that result in the
    > > fatal too-short grace period. Since rcu_qsctr_inc() refuses to
    > > increment the quiescent-state counter on any CPU that started an RCU
    > > read-side critical section that has not yet completed, any long
    > > critical section will have a corresponding CPU that will refuse to go
    > > through a quiescent state. And that will prevent the grace period
    > > from completing.
    > i'm worried about the following scenario: what happens when a task is
    > migrated from CPU#1 to CPU#2, while in an RCU read section that it
    > acquired on CPU#1, and queues a callback. E.g. d_free() does a
    > call_rcu(), to queue the freeing of the dentry.
    > That callback will be queued on CPU#2 - while the task still keeps
    > current->rcu_data of CPU#1. It also means that CPU#2's read counter did
    > _not_ get increased - and a too short grace period may occur.

    Let me make sure I understand the sequence of events:

    CPU#1 CPU#2

    task1: rcu_read_lock()

    task1: migrate to CPU#2

    task1: call_rcu()

    task1: rcu_read_unlock()

    This sequence will be safe because CPU#1's active_readers field will
    be non-zero throughout, so that CPU#1 will refuse to record any
    quiescent state from the time that task1 does the rcu_read_lock()
    on CPU#1 until the time that it does the rcu_read_unlock() on CPU#2.

    Now, it is true that CPU#2 might record a quiescent state during this
    time, but this will have no effect because -all- CPUs must pass through
    a quiescent state before any callbacks will be invoked. Since CPU#1
    is refusing to record a quiescent state, grace periods will be blocked
    for the full extent of task 1's RCU read-side critical section.

    Or am I misunderstanding your scenario? Or, for that matter, your code?

    > it seems to me that that only safe method is to pick an 'RCU CPU' when
    > first entering the read section, and then sticking to it, no matter
    > where the task gets migrated to. Or to 'migrate' the +1 read count from
    > one CPU to the other, within the scheduler.

    This would work too, but I don't believe it is necessary given what
    you are already doing.

    > the 'migrate read count' solution seems more promising, as it would keep
    > other parts of the RCU code unchanged. [ But it seems to break the nice
    > 'flip pointers' method you found to force a grace period. If a 'read
    > section' can migrate from one CPU to another then it can migrate back as
    > well, at which point it cannot have the 'old' pointer. Maybe it would
    > still work better than no flip pointers. ]

    Well, I do believe that suppressing migration of tasks during RCU read-side
    critical sections would simplify the flip-pointers/counters code. But
    that is easy to say in advance of actually producing this code. ;-)

    Thanx, Paul
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to
    More majordomo info at
    Please read the FAQ at

     \ /
      Last update: 2005-04-06 13:30    [W:0.027 / U:294.464 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site