Re: [PATCH] cgroup/cpuset: Change nr_deadline_tasks to an atomic_t value

From: Waiman Long
Date: Thu Oct 12 2023 - 12:36:17 EST

On 10/11/23 08:54, Waiman Long wrote:

On 10/11/23 04:14, Juri Lelli wrote:
On 10/10/23 16:03, Waiman Long wrote:
On 10/10/23 15:44, Waiman Long wrote:
On 10/10/23 01:34, Juri Lelli wrote:
Hi,

On 09/10/23 15:15, Waiman Long wrote:
The nr_deadline_tasks field in cpuset structure was introduced by
commit 6c24849f5515 ("sched/cpuset: Keep track of SCHED_DEADLINE task
in cpusets"). Unlike nr_migrate_dl_tasks which is only modified under
cpuset_mutex, nr_deadline_tasks can be updated in various contexts
under different locks. As a result, data races may happen that cause
incorrect value to be stored in nr_deadline_tasks leading to incorrect
Could you please make an example of such data races?
Since update to cs->nr_deadline_tasks is not protected by a single lock,
it is possible that multiple CPUs may try to modify it at the same
time.  It is possible that nr_deadline_tasks++ and nr_deadline_tasks--
can be done in a single instruction like in x86 and hence atomic.
However, operation like "cs->nr_deadline_tasks +=
cs->nr_migrate_dl_tasks" is likely a RMW operation and so is subjected
to racing. It is mostly theoretical, but probably not impossible.
Sorry, even increment and decrement operators are not atomic.

inc_dl_tasks_cs() is only called from switched_to_dl() in deadline.c which
is protected by the rq_lock, but there are multiple rq's. dec_dl_tasks_cs()
is called from switched_from_dl() in deadline.c and cgroup_exit() in
cgroup.c. The later one is protected by css_set_lock. The other place where
nr_deadline_tasks can be changed is in cpuset_attach() protected by
cpuset_mutex.
So, let's see. :)

switched_to_dl(), switched_from_dl() and cpuset_attach() should all be
protected (for DEADLINE tasks) by cpuset_mutex, see [1] for the former
two.
Yes, I missed the cpuset_lock() call.
What leaves me perplexed is indeed cgroup_exit(), which seems to operate
under css_set_lock as you say. I however wonder why is that not racy
already wrt, say, cpuset_attach() which AFAIU uses css information w/o
holding css_set_lock?

The css_set_lock protects changes made to css_set. Looking at cgroup_migrate_execute(), css_set_lock is taken when the tasks are actually moving from one css_set to another one. cpuset_attach() is called just to update the CPU and node affinity and cpuset_mutex is taken to ensure stability of the CPU and node masks. There is no change to css_set and so css_set_lock isn't needed.

We can argue that there can be racing between cgroup_exit() and the iteration of tasks in cpuset_attach() or cpuset_can_attach(). An rcu_read_lock() is probably needed. I am stilling investigating that.

Cgroup has a rather complex task migration and iteration scheme. According to the following comments in include/linux/cgroup-defs.h:

        /*
         * Lists running through all tasks using this cgroup group.
         * mg_tasks lists tasks which belong to this cset but are in the
         * process of being migrated out or in.  Protected by
         * css_set_lock, but, during migration, once tasks are moved to
         * mg_tasks, it can be read safely while holding cgroup_mutex.
         */
        struct list_head tasks;
        struct list_head mg_tasks;
        struct list_head dying_tasks;

I haven't fully figured out how that protection works yet. Assuming that is the case, task iteration in cpuset_attach() should be fine since cgroup_mutex is indeed held when it is invoked. That protection, however, does not applied to nr_deadline_tasks. It may be too costly to acquire cpuset_mutex before updating nr_deadline_tasks in cgroup_exit(). So changing it to an atomic_t should be the easy way out of the potential racing problem.

I can update the commit log with these new analysis if you have no further objection to this change.

Cheers,
Longman