[lkml]   [2007]   [Oct]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    Subject[RFC PATCH 0/4] Refcount Based Cpu-Hotplug Revisit.
    This patch series attempts to revisit the topic of
    cpu-hotplug locking model.

    Prior to this attempt, there were several different suggestions on
    how it should be implemented.

    The ones that were posted before were

    a) Refcount + Waitqueue model:
    Here the threads that want to avoid a cpu hotplug operation while
    they are operating in cpu-hotplug critical section, bump up the
    reference to the global online cpu state.
    The thread which wants to perform a cpu-hotplug,
    blocks until the reference to the global online state goes
    to zero. Any threads which want to enter the cpu-hotplug critical
    section during an ongoing cpu-hotplug operatoin, are blocked using
    a waitqueue.

    The advantange of this model was that it is along the lines of
    the well known get/put model. Only that it allows sleeping of readers
    and writers.

    The disadvantage, as Andrew pointed out was that there do exist
    a whole bunch of lock_cpu_hotplug()'s whose existance is undocumented,
    and an approach like this will not improve such a situation.

    b) Per Subsystem cpu-hotplug locks: Each subsystem which has cpu-hotplug
    critical data, uses a lock to protect that data. Such a subsystem
    needs to subscribe to the cpu-hotplug notification, especially the
    CPU_LOCK_ACQUIRE and CPU_LOCK_RELEASE events which are sent before
    and and after a cpu-hotplug operation. While handling these events
    respectively, the subsystem lock is taken or released.

    The advantage this model offered was that lock was associated with the
    data, which made easy to understand the purpose of locking.

    The disadvantage was that any cpu-hotplug aware function, could
    not be called from a cpu-hotplug callback path, since we would have
    acquired the subsystem lock during CPU_LOCK_ACQUIRE and attempting
    to reacquire it would result in a deadlock.
    The case which pointed this limitation out was the implementation of
    synchronize_sched in preemptible rcu.

    c) Freezer based cpu-hotplug:
    The idea here was to freeze the system using the process freezer
    technique which is being used for suspend/hibernate purpose, before
    performing a cpu-hotplug operation. This would ensure that none of
    the kernel threads are accessing any of the cpu-hotplug critical
    data, because they are frozen at well known points.

    This would have helped to remove all kinds of locks because when a
    thread is accessing a cpu-hotplug critical data, it meant that the
    system was not frozen and hence there would be no cpu-hotplug
    operation untill the thread either voluntarily calls try_to_freeze
    or returns out of the kernel.

    The disadvantage of this approach was that any kind of dependencies
    between threads might call the freezer to fail. For eg, thread A is
    waiting for thread B to accomplish something, but thread B is already
    frozen, leading to a freeze failure. There could be other subtle
    races which might be difficult to track.

    Some time in May 2007, Linus suggested using the refcount model, and
    this patch series simplifies and reimplements the Refcount + waitqueue
    model, based on the discussions in the past and inputs from
    Vatsa and Rusty.

    Patch 1/4: Implements the core refcount + waitqueue model.
    Patch 2/4: Replaces all the lock_cpu_hotplug/unlock_cpu_hotplug instances
    with get_online_cpus()/put_online_cpus()
    Patch 3/4: Replaces the subsystem mutexes (we do have three of them now,
    in sched.c, slab.c and workqueue.c) with get_online_cpus,
    Patch 4/4: Eliminates the CPU_DEAD and CPU_UP_CANCELLED event handling
    from workqueue.c

    The patch series has survived an overnight test with kernbench on i386.
    and has been tested with Paul Mckenney's latest preemptible rcu code.

    Awaiting thy feedback!

    Thanks and Regards
    Gautham R Shenoy
    Linux Technology Center
    IBM India.
    "Freedom comes with a price tag of responsibility, which is still a bargain,
    because Freedom is priceless!"
    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: 2007-10-16 12:35    [W:0.028 / U:83.232 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site