RE: [PATCH] cpuset: fix cpus_allowed mask for offline/online CPUs

[Chen, Yu C]

Hi Zefan,

> -----Original Message-----
> From: Zefan Li [mailto:lizefan@xxxxxxxxxx]
> Sent: Monday, January 04, 2016 9:52 AM
> To: Chen, Yu C; cgroups@xxxxxxxxxxxxxxx
> Cc: linux-kernel@xxxxxxxxxxxxxxx; Vlastimil Babka; Rik van Riel; Joonsoo Kim;
> David Rientjes; Vishnu Pratap Singh; Pintu Kumar; Michal Nazarewicz; Mel
> Gorman; Gortmaker, Paul (Wind River); Peter Zijlstra; Tim Chen; Hugh Dickins;
> Tejun Heo
> Subject: Re: [PATCH] cpuset: fix cpus_allowed mask for offline/online CPUs
> 
> On 2016/1/1 20:09, Chen Yu wrote:
> > Commit be4c9dd7aee5 ("cpuset: enable onlined cpu/node in effective
> > masks") leverages cpuset's cpus_allowed and its parent's
> > effective_cpus to calculate the new_cpus by:
> >
> > cpumask_and(&new_cpus, cs->cpus_allowed,
> > parent_cs(cs)->effective_cpus);
> >
> > However cpus_allowed will also be updated after the CPU is offline, in
> > hotplug_update_tasks_legacy, so when the CPU is online again, it will
> > use the old cpus_allowed mask to calculate the new_cpus, thus new_cpus
> > will get incorrect value after each round of offline/online.
> >
> > This problem is found on ubuntu 15.10 with cpuset mounted:
> >
> > 1. echo 0 > /sys/devices/system/cpu/cpu2/online
> > 2. echo 1 > /sys/devices/system/cpu/cpu2/online
> > 3. cat /sys/fs/cgroup/cpuset/cpuset.cpus
> >     0-3
> > 4. cat /sys/fs/cgroup/cpuset/user.slice/cpuset.cpus
> >     0-1,3
> > 5. taskset -c 2 ls
> >     taskset: failed to set pid 0's affinity: Invalid argument
> >
> 
> This is the expected behavior...In legacy hierachy onlining an offlined cpu
> won't restore cpuset configurations automatically.
[Yu] Ah, I see, we were just a little confused with it before.
 So if we want to online an offlined cpu in legacy hierarchy, we
should not only echo 1 to 'online' sysfs, but also restore the mask manually? 
It seems that we should not rely on /sys/devices/system/cpu/cpu2/online,
if cpuset has been mounted.
> 
> Commit be4c9dd7aee5 changed this behavior, but only for unified hierachy.

thanks,
yu
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/