linux-kernel - Re: [PATCH 1/2] sched: Defend cfs and rt bandwidth quota against overflow

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <xm261roim4hi.fsf@google.com>
Date:   Mon, 20 Apr 2020 10:50:01 -0700
From:   bsegall@...gle.com
To:     Huaixin Chang <changhuaixin@...ux.alibaba.com>
Cc:     linux-kernel@...r.kernel.org, peterz@...radead.org,
        mingo@...hat.com, bsegall@...gle.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

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.

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.

>
> 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.

> +
>  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);