[lkml]   [2012]   [Jul]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [PATCHSET] workqueue: reimplement CPU hotplug to keep idle workers
    Hey, Peter.

    On Fri, Jul 20, 2012 at 05:48:31PM +0200, Peter Zijlstra wrote:
    > On Tue, 2012-07-17 at 10:12 -0700, Tejun Heo wrote:
    > > While this makes rebinding somewhat more complicated, as it has to be
    > > able to rebind idle workers too, it allows overall hotplug path to be
    > > much simpler.
    > I really don't see the point of re-binding.. at that point you've well
    > and proper violated any per-cpu expectation, so why not complete running
    > the works on the disassociated thing and let new works accrue on the
    > per-cpu things again?

    We've discussed this a couple times now, so the existing reasons were,

    * Local affinity is more often used as a form of affinity optimization
    since the beginning. This, mixed with queue_work() /
    queue_work_on(), does make things muddy.

    * With local affinity used for optimization, we better support
    detaching running workers - before cmwq, this used to be one of the
    sources of trouble during power state changes.

    * So, we have unbound workers which started as bound while a CPU is
    down. When the CPU comes back up again, we can do one of the
    followings - 1. migrate the unbound ones to WORK_CPU_UNBOUND (can
    also do this on CPU_DOWN), 2. leave them unbound and keep them
    running in parallel with bound ones, or 3. rebind them. #2 is the
    hariest - it contaminates the usual !hotplug code paths. #1 or #3,
    unsure, but given how global_cwq's don't usually interact with each
    other, I thought #3 would be lower impact on hot paths.

    So, the above was my rationale before this "we need to stop destroying
    and re-creating kthreads across CPU hotplug events because phones do
    it gazillion times". Now, I don't think we have any other way.



     \ /
      Last update: 2012-07-20 19:41    [W:0.022 / U:11.628 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site