lkml.org 
[lkml]   [2012]   [Mar]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 0/4] arch/CPU hotplug: Add missing CPU Hotplug bits to fix nasty issues
From
Date
On Thu, 2012-03-22 at 13:13 +0100, Peter Zijlstra wrote:
> On Thu, 2012-03-22 at 16:58 +0530, Srivatsa S. Bhat wrote:
> > Commit 5fbd036b552f633abb394a319f7c62a5c86a9cd7 (sched: Cleanup cpu_active
> > madness) introduced some changes that made the scheduler rely on the
> > CPU_STARTING notifier. And hence those architectures which forgot to
> > send out the CPU_STARTING notification will almost surely get into trouble.
> > (Xen is one example[1]).
>
> The requirement is that CPU_STARTING is ran before set_cpu_online() or
> the open-coded equivalent -- eg. blackfin gets this wrong.
>
> However, it is also required that all this happens while local IRQs are
> still disabled, since the moment we enable IRQs interrupts can happen
> and interrupts can cause wakeups, and wakeups need to have this state
> set up -- blackfin, cris, m32r, mips, sh, sparc64, sparc32, um?, x86 get
> this wrong.
>
> Furthermore, all archs that use CONFIG_USE_GENERIC_SMP_HELPERS should
> hold ipi_call_lock() over setting the cpu online -- alpha, arm, m32r,
> mips, sh, sparc32 seem wrong.

So for parisc, anywhere in smp_callin() should be fine. Our last act on
CPU bringup before going into cpu_idle() is to enable interrupts.

> Furthermore, I was pondering the scenario where a 3rd cpu IPIs the newly
> booting cpu, I suspect we need a smp_wmb() after setting cpu_active and
> a rmb in select_fallback_rq() before reading active.
>
> All in all its a complete friggin trainwreck.

I haven't yet actually tried git head on parisc (and I won't be able to
until I actually get my machines back [currently being shipped across
the atlantic]) but if there's a problem it will show up quickly: parisc
actually uses cpu hotplug to boot its secondary cpus.

James




\
 
 \ /
  Last update: 2012-03-22 15:55    [W:0.192 / U:0.092 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site