Date:   Wed, 22 Apr 2020 11:44:30 -0700
From:   bsegall@...gle.com
To:     changhuaixin <changhuaixin@...ux.alibaba.com>
Cc:     bsegall@...gle.com, linux-kernel@...r.kernel.org,
        peterz@...radead.org, mingo@...hat.com, chiluk+linux@...eed.com,
        vincent.guittot@...aro.org, pauld@...head.com
Subject: Re: [PATCH 1/2] sched: Defend cfs and rt bandwidth quota against overflow

changhuaixin <changhuaixin@...ux.alibaba.com> writes:

>> 在 2020年4月21日，上午1:50，bsegall@...gle.com 写道：
>> 
>> Huaixin Chang <changhuaixin@...ux.alibaba.com> writes:
>> 
>>> Kernel limitation on cpu.cfs_quota_us is insufficient. Some large
>>> numbers might cause overflow in to_ratio() calculation and produce
>>> unexpected results.
>>> 
>>> For example, if we make two cpu cgroups and then write a reasonable
>>> value and a large value into child's and parent's cpu.cfs_quota_us. This
>>> will cause a write error.
>>> 
>>> 	cd /sys/fs/cgroup/cpu
>>> 	mkdir parent; mkdir parent/child
>>> 	echo 8000 > parent/child/cpu.cfs_quota_us
>>> 	# 17592186044416 is (1UL << 44)
>>> 	echo 17592186044416 > parent/cpu.cfs_quota_us
>>> 
>>> In this case, quota will overflow and thus fail the __cfs_schedulable
>>> check. Similar overflow also affects rt bandwidth.
>> 
>> More to the point is that I think doing
>> 
>> echo 17592186044416 > parent/cpu.cfs_quota_us
>> echo 8000 > parent/child/cpu.cfs_quota_us
>> 
>> will only fail on the second write, while with this patch it will fail
>> on the first, which should be more understandable.
>> 
>> 
>> to_ratio could be altered to avoid unnecessary internal overflow, but
>> min_cfs_quota_period is less than 1<<BW_SHIFT, so a cutoff would still
>> be needed.
>> 
>
> Yes, I will rewrite commit log in the following patch.
>
>> Also tg_rt_schedulable sums a bunch of to_ratio(), and doesn't check for
>> overflow on that sum, so if we consider preventing weirdness around
>> schedulable checks and max quotas relevant we should probably fix that too.
>> 
>
> It seems to me that check for overflow on sum of to_ratio(rt_period, rt_runtime)
> is not necessary. As to_ratio() of a rt group is bounded by global_rt_period()
> and global_rt_runtime() due to the checks in tg_rt_schedulable(). And
> global_rt_runtime() is not allowed to be greater than global_rt_period() thanks
> to sched_rt_global_validate(). Thus, to_ratio() of a rt group will not exceed
> BW_UNIT, sum of which is unlikely to overflow then. Checks against rt_runtime
> overflow during to_ratio is still needed.
>
> Is that correct?

Good point, that's probably not a problem then.

>
>>> 
>>> Signed-off-by: Huaixin Chang <changhuaixin@...ux.alibaba.com>
>>> ---
>>> kernel/sched/core.c  | 8 ++++++++
>>> kernel/sched/rt.c    | 9 +++++++++
>>> kernel/sched/sched.h | 2 ++
>>> 3 files changed, 19 insertions(+)
>>> 
>>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>>> index 3a61a3b8eaa9..f0a74e35c3f0 100644
>>> --- a/kernel/sched/core.c
>>> +++ b/kernel/sched/core.c
>>> @@ -7390,6 +7390,8 @@ static DEFINE_MUTEX(cfs_constraints_mutex);
>>> 
>>> const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
>>> static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
>>> +/* More than 203 days if BW_SHIFT equals 20. */
>>> +static const u64 max_cfs_runtime = MAX_BW_USEC * NSEC_PER_USEC;
>>> 
>>> static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
>>> 
>>> @@ -7417,6 +7419,12 @@ static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
>>> 	if (period > max_cfs_quota_period)
>>> 		return -EINVAL;
>>> 
>>> +	/*
>>> +	 * Bound quota to defend quota against overflow during bandwidth shift.
>>> +	 */
>>> +	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
>>> +		return -EINVAL;
>>> +
>>> 	/*
>>> 	 * Prevent race between setting of cfs_rq->runtime_enabled and
>>> 	 * unthrottle_offline_cfs_rqs().
>>> diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
>>> index df11d88c9895..f5eea19d68c4 100644
>>> --- a/kernel/sched/rt.c
>>> +++ b/kernel/sched/rt.c
>>> @@ -2569,6 +2569,9 @@ static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
>>> 	return ret;
>>> }
>>> 
>>> +/* More than 203 days if BW_SHIFT equals 20. */
>>> +static const u64 max_rt_runtime = MAX_BW_USEC * NSEC_PER_USEC;
>> 
>> It looks to me like __rt_schedulable doesn't divide by NSEC_PER_USEC, so
>> to_ratio is operating on nsec, and the limit is in nsec, and MAX_BW_USEC
>> should probably not be named USEC then as well.
>
> Yes, the limit for rt_runtime is in nsec. This should be changed.
>
>> 
>>> +
>>> static int tg_set_rt_bandwidth(struct task_group *tg,
>>> 		u64 rt_period, u64 rt_runtime)
>>> {
>>> @@ -2585,6 +2588,12 @@ static int tg_set_rt_bandwidth(struct task_group *tg,
>>> 	if (rt_period == 0)
>>> 		return -EINVAL;
>>> 
>>> +	/*
>>> +	 * Bound quota to defend quota against overflow during bandwidth shift.
>>> +	 */
>>> +	if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
>>> +		return -EINVAL;
>>> +
>>> 	mutex_lock(&rt_constraints_mutex);
>>> 	err = __rt_schedulable(tg, rt_period, rt_runtime);
>>> 	if (err)
>>> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
>>> index db3a57675ccf..6f6b7f545557 100644
>>> --- a/kernel/sched/sched.h
>>> +++ b/kernel/sched/sched.h
>>> @@ -1918,6 +1918,8 @@ extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se);
>>> #define BW_SHIFT		20
>>> #define BW_UNIT			(1 << BW_SHIFT)
>>> #define RATIO_SHIFT		8
>>> +#define MAX_BW_BITS		(64 - BW_SHIFT)
>>> +#define MAX_BW_USEC		((1UL << MAX_BW_BITS) - 1)
>>> unsigned long to_ratio(u64 period, u64 runtime);
>>> 
>>> extern void init_entity_runnable_average(struct sched_entity *se);

Powered by blists - more mailing lists