Re: [PATCH] sched/fair: Prevent dead task groups from regaining cfs_rq's

[Benjamin Segall]

Mathias Krause <minipli@xxxxxxxxxxxxxx> writes:

> Am 04.11.21 um 09:50 schrieb Vincent Guittot:
>> On Wed, 3 Nov 2021 at 23:04, Benjamin Segall <bsegall@xxxxxxxxxx> wrote:
>>>
>>> Mathias Krause <minipli@xxxxxxxxxxxxxx> writes:
>>>
>>>> Kevin is reporting crashes which point to a use-after-free of a cfs_rq
>>>> in update_blocked_averages(). Initial debugging revealed that we've live
>>>> cfs_rq's (on_list=1) in an about to be kfree()'d task group in
>>>> free_fair_sched_group(). However, it was unclear how that can happen.
>>>> [...]
>>>> Fixes: a7b359fc6a37 ("sched/fair: Correctly insert cfs_rq's to list on unthrottle")
>>>> Cc: Odin Ugedal <odin@xxxxxxx>
>>>> Cc: Michal Koutný <mkoutny@xxxxxxxx>
>>>> Reported-by: Kevin Tanguy <kevin.tanguy@xxxxxxxxxxxx>
>>>> Suggested-by: Brad Spengler <spender@xxxxxxxxxxxxxx>
>>>> Signed-off-by: Mathias Krause <minipli@xxxxxxxxxxxxxx>
>>>> ---
>>>>  kernel/sched/core.c | 18 +++++++++++++++---
>>>>  1 file changed, 15 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>>>> index 978460f891a1..60125a6c9d1b 100644
>>>> --- a/kernel/sched/core.c
>>>> +++ b/kernel/sched/core.c
>>>> @@ -9506,13 +9506,25 @@ void sched_offline_group(struct task_group *tg)
>>>>  {
>>>>       unsigned long flags;
>>>>
>>>> -     /* End participation in shares distribution: */
>>>> -     unregister_fair_sched_group(tg);
>>>> -
>>>> +     /*
>>>> +      * Unlink first, to avoid walk_tg_tree_from() from finding us (via
>>>> +      * sched_cfs_period_timer()).
>>>> +      */
>>>>       spin_lock_irqsave(&task_group_lock, flags);
>>>>       list_del_rcu(&tg->list);
>>>>       list_del_rcu(&tg->siblings);
>>>>       spin_unlock_irqrestore(&task_group_lock, flags);
>>>> +
>>>> +     /*
>>>> +      * Wait for all pending users of this task group to leave their RCU
>>>> +      * critical section to ensure no new user will see our dying task
>>>> +      * group any more. Specifically ensure that tg_unthrottle_up() won't
>>>> +      * add decayed cfs_rq's to it.
>>>> +      */
>>>> +     synchronize_rcu();
>>>
>>> I was going to suggest that we could just clear all of avg.load_sum/etc, but
>>> that breaks the speculative on_list read. Currently the final avg update
>>> just races, but that's not good enough if we wanted to rely on it to
>>> prevent UAF. synchronize_rcu() doesn't look so bad if the alternative is
>>> taking every rqlock anyways.
>>>
>>> I do wonder if we can move the relevant part of
>>> unregister_fair_sched_group into sched_free_group_rcu. After all
>>> for_each_leaf_cfs_rq_safe is not _rcu and update_blocked_averages does
>>> in fact hold the rqlock (though print_cfs_stats thinks it is _rcu and
>>> should be updated).
>> 
>> I was wondering the same thing.
>> we would have to move unregister_fair_sched_group() completely in
>> sched_free_group_rcu() and probably in cpu_cgroup_css_free() too.
>
> Well, the point is, print_cfs_stats() pretty much relies on the list to
> be stable, i.e. safe to traverse. It doesn't take locks while walking it
> (beside the RCU lock), so if we would modify it concurrently, bad things
> would happen.


Yeah, my idea was that print_cfs_stats is just debug info so it wouldn't
be a disaster to hold the lock, but I forgot that we probably can't hold
it over all that writing.