lkml.org 
[lkml]   [2012]   [Mar]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: RFC [PATCH] SMP: Don't schedule tasks on inactive cpu(s)

    Thanks Peter for having a look. I will try to come up with a way to
    trigger this issue more easily.

    If I find a way, I will test with your original patch,
    https://lkml.org/lkml/2011/12/15/255, and tell you the result.

    Best regards,
    Jonas Aaberg

    On 02/29/2012 12:48 PM, Peter Zijlstra wrote:
    > On Wed, 2012-02-29 at 12:03 +0100, Peter Zijlstra wrote:
    >> On Wed, 2012-02-29 at 11:42 +0100, Jonas Aaberg wrote:
    >>> This patch removes the ability to schedule tasks on cpus that are online,
    >>> but not active. The reason for this patch is that during cpu hotplug
    >>> on ARM (atleast) there is a short window where cpuX (X > 0) is online, but
    >>> busy-waiting on cpu0 to put it active, meanwhile cpu0 can be interrupted
    >>> and try to schedule something on the cpu that is busy checking its active bit.
    >>
    >> https://lkml.org/lkml/2011/12/15/255
    >>
    >> that one?
    >>
    >> I _think_ its correct, but it would be so good if someone else could
    >> verify.
    >
    > Relevant patches to consider are: e761b772 and 3a101d05.
    >
    > Having looked at this again, I think we lost something in 3a101d05 since
    > it moves cpuset_update_active_cpus() from CPU_DEAD to CPU_DOWN_PREPARE
    > (and DOWN_FAILED) -- not that it matters that much. Also this patch does
    > leaves me somewhat puzzled as to what cpu_active_mask is for now..
    >
    > The suggested patch linked above moves setting active to CPU_STARTING
    > which is _before_ online. It looks like some parts of the scheduler
    > don't look at online at all anymore so that opens a 'window' where we
    > could select a cpu that isn't part of the sched_domain nor online
    > (select_fallback_rq and cpuset_cpus_allowed_fallback).
    >
    > Now this isn't really a problem because of stop-machine, by the time
    > anybody gets to run again both online and active are set and we should
    > be good to go. The bad part is of course us relying on this silly
    > stop-machine semantic.
    >
    > Bah, hotplug is such a pain..



    \
     
     \ /
      Last update: 2012-03-02 11:53    [W:3.470 / U:0.004 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site