lkml.org 
[lkml]   [2014]   [Jan]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Deadlock between cpu_hotplug_begin and cpu_add_remove_lock
On 01/24/2014 04:31 AM, Rusty Russell wrote:
> "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com> writes:
>> On 01/23/2014 07:59 AM, Rusty Russell wrote:
>>> "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com> writes:
>>>> On 01/22/2014 02:00 PM, Srivatsa S. Bhat wrote:
>>>>> Hi Paul,
>>>
>>> I find an old patch for register_allcpu_notifier(), but the "bool
>>> replay_history" should be eliminated (always true): it's too weird.
>>>
>>
>> Sorry, I didn't get this part. Why do you say that replay_history
>> will always be true?
>
> OK, let me start again and try to explain myself properly:
>
> register_cpu_notifier is a bad API. It's hard to get right because:
> 1) You need to loop over online (or present) cpus once before you call
> it.
> 2) You have to beware the race between the loop and registration, but
> much example code happens at boot time where it doesn't matter,
> so random author is likely to copy that and have a race.
> 3) You have two paths doing the same thing: the loop which is run on
> every machine (cpu hotplug or not), and the notifier callback which
> is run far less rarely.
>
> What we actually *want* is a routine which will reliably call for every
> current and future CPU, and then there are very few places which should
> use the current register_cpu_notifier().
>
> ie. halfway between register_cpu_notifier() (too racy) and
> register_allcpu_notifier() (too simplified).
>
> Let's call it register_cpu_callback / unregister_cpu_callback?
>

Thanks a lot for the detailed and profound explanation! It makes perfect
sense to me now.

>> By the way, I'm still tempted to try out the simpler-looking alternative
>> idea of exporting cpu_maps_update_begin() and cpu_maps_update_done()
>> and then mandating that the callers do:
>>
>> cpu_maps_update_begin();
>> for_each_online_cpu(cpu) {
>> ...
>> }
>>
>> __register_cpu_notifier(); // this doesn't take the add_remove_lock
>> cpu_maps_update_done();
>
> Sure, fix this one for -stable. But let's create an idiom we can be
> proud of for the longer term.
>

Ok, that sounds good, will work on that.

Thank you very much!

Regards,
Srivatsa S. Bhat



\
 
 \ /
  Last update: 2014-01-28 16:21    [W:0.047 / U:0.964 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site