Hello Jing-Ting.
On Thu, Jun 30, 2022 at 06:52:10PM +0800, Jing-Ting Wu <[email protected]> wrote:
> Root cause:
> The rebind_subsystems() is no lock held when move css object from A
> list to B list,then let B's head be treated as css node at
> list_for_each_entry_rcu().
Nice outcome of the analysis.
> The call stack as following:
> kthread
> -> worker_thread
> -> process_one_work
> -> flush_memcg_stats_dwork
> -> __mem_cgroup_flush_stats
> -> cgroup_rstat_flush_irqsafe
> -> cgroup_rstat_flush_locked
>
> Detail:
> static void cgroup_rstat_flush_locked(struct cgroup *cgrp, bool
> may_sleep)
> {
> ...
> rcu_read_lock();
> list_for_each_entry_rcu(css, &pos->rstat_css_list,
> rstat_css_node)
> css->ss->css_rstat_flush(css, cpu);
> rcu_read_unlock();
> ...
> }
>
> Because the content of css->ss is not a function address,
> once the css_rstat_flush is called, kernel exception will happen.
I agree with your analysis. The list_for_each_entry_rcu() is not
accounting for the move from A to B and the head's rstat_css_list is
invalid in the cgroup_rstat_flush() context.
> When the issue happened, the list of pos->rstat_css_list at group A is
> empty.
> There must be racing conditions.
>
> From reviewing code, we find rebind_subsystems() is no lock held when
> move css object to other list.
>
> int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
> {
> ...
> if (ss->css_rstat_flush) {
> list_del_rcu(&css->rstat_css_node);
> list_add_rcu(&css->rstat_css_node,
> &dcgrp->rstat_css_list);
> }
> ...
> }
>
> If we held a css object from A group list, at same time this css object
> was moved to B group list.
> [...]
> Do you have any suggestion for this issue?
The css->rstat_css_node should not be modified if there are possible RCU
readers elsewhere.
One way to fix this would be to insert synchronize_rcu() after
list_del_rcu() and before list_add_rcu().
(A further alternative (I've heard about) would be to utilize 'nulls'
RCU lists [1] to make the move between lists detectable.)
But as I'm looking at it from distance, it may be simpler and sufficient
to just take cgroup_rstat_lock around the list migration (the nesting
under cgroup_mutex that's held with rebind_subsystems() is fine).
HTH,
Michal
[1] https://www.kernel.org/doc/html/latest/RCU/rculist_nulls.html
(resending, I messed up the message header, sorry)
Hello,
On Fri, Jul 15, 2022 at 01:59:38PM +0200, Michal Koutn? wrote:
> The css->rstat_css_node should not be modified if there are possible RCU
> readers elsewhere.
> One way to fix this would be to insert synchronize_rcu() after
> list_del_rcu() and before list_add_rcu().
> (A further alternative (I've heard about) would be to utilize 'nulls'
> RCU lists [1] to make the move between lists detectable.)
>
> But as I'm looking at it from distance, it may be simpler and sufficient
> to just take cgroup_rstat_lock around the list migration (the nesting
> under cgroup_mutex that's held with rebind_subsystems() is fine).
synchronize_rcu() prolly is the better fit here given how that list_node's
usage, but yeah, great find.
Thanks.
--
tejun