Re: [PATCH] cgroup: cpuset: fix panic when offline a cpu

[Waiman Long <longman@redhat.com>]

On 09/22/2017 07:00 AM, shuwang@redhat.com wrote:
> From: Shu Wang <shuwang@redhat.com>
>
> cgroup_migrate assumes mgctx tset.csets is pointing to
> tset.src_csets at start, add tasks to tset.src_csets in
> func cgroup_migrate_add_task, change test.csets to
> tset.dst_csets in cgroup_migrate_execute.
>
> For offline a cpu in cgroup_transfer_tasks, it will first
> migrate a task and cause tset.csets pointing to dst_csets.
> Get a NULL pointer in cgroup_taskset_first.
>
> reproducer on my 2 cpus machine:
> mkdir /sys/fs/cgroup/cpuset/test
> cd /sys/fs/cgroup/cpuset/test
> echo 1 > cpuset.cpus
> echo 0 > cpuset.mems
> sleep 100 & echo $! > tasks
> sleep 100 & echo $! > tasks
> echo 0 > /sys/bus/cpu/devices/cpu1/online
>
> backtrace:
> BUG: unable to handle kernel NULL pointer dereference at 0000000000000cf8
> IP: cpuset_can_attach+0x2f/0x140
> Call Trace:
>  ? cpuset_attach+0x30f/0x3d0
>  cgroup_migrate_execute+0x71/0x3c0
>  cgroup_migrate+0x75/0x80
>  cgroup_transfer_tasks+0x1b2/0x230
>  cpuset_hotplug_workfn+0xa7d/0xce0
>  ? finish_task_switch+0x79/0x240
>  process_one_work+0x149/0x360
>  worker_thread+0x4d/0x3c0
>
> Signed-off-by: Shu Wang <shuwang@redhat.com>
> ---
>  kernel/cgroup/cgroup-v1.c | 6 ++++++
>  1 file changed, 6 insertions(+)
>
> diff --git a/kernel/cgroup/cgroup-v1.c b/kernel/cgroup/cgroup-v1.c
> index 024085daab1a..165734573b5e 100644
> --- a/kernel/cgroup/cgroup-v1.c
> +++ b/kernel/cgroup/cgroup-v1.c
> @@ -129,6 +129,12 @@ int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
>  		css_task_iter_end(&it);
>  
>  		if (task) {
> +			/*
> +			 * Reset csets to src_cets, as cgroup_migrate assumes
> +			 * csets is pointing to src_csets.
> +			 */
> +			mgctx.tset.csets = &mgctx.tset.src_csets;
> +
>  			ret = cgroup_migrate(task, false, &mgctx);
>  			if (!ret)
>  				trace_cgroup_transfer_tasks(to, task, false);

I had actually sent a patch to fix the same bug yesterday. See

https://lkml.org/lkml/2017/9/21/333

Cheers,
Longman