Re: [PATCH-tip] sched: Don't call kfree() in do_set_cpus_allowed()

[Waiman Long <longman@redhat.com>]

On 11/22/22 07:37, Peter Zijlstra wrote:
> On Mon, Nov 21, 2022 at 10:04:33AM -0500, Waiman Long wrote:
>> On 11/21/22 05:38, Peter Zijlstra wrote:
>>> On Fri, Nov 18, 2022 at 02:33:02PM -0500, Waiman Long wrote:
>>>> Commit 851a723e45d1 ("sched: Always clear user_cpus_ptr in
>>>> do_set_cpus_allowed()") may call kfree() if user_cpus_ptr was previously
>>>> set. Unfortunately, some of the callers of do_set_cpus_allowed()
>>> 'some' ? There's only 3 or so, which one triggers this?
>> It happenned at __kthread_bind_mask() where do_set_cpus_allowed() is called
>> with pi_lock held.
>>
>> [ 1084.820105]  <TASK>
>> [ 1084.820110]  dump_stack_lvl+0x57/0x81
>> [ 1084.820117]  check_noncircular+0x103/0x120
>> [ 10[ 1084.820160]  lock_acquire+0xba/0x230
>> [ 1084.820164]  ? kfree+0x10f/0x380
>> [ 1084.820172]  ? do_set_cpus_allowed+0x40/0x60
>> [ 1084.820181]  rt_spin_lock+0x27/0xe0
>> [ 1084.820184]  ? kfree+0x10f/0x380
>> [ 1084.820188]  kfree+0x10f/0x380
>> [ 1084.820195]  do_set_cpus_allowed+0x40/0x60
>> [ 1084.820203]  kthread_bind_mask+0x4a/0x70
>> [ 1084.820211]  create_worker+0xfb/0x1a0
>> [ 1084.820220]  worker_thread+0x2e3/0x3c0
>> [ 1084.820226]  ? process_one_work+0x450/0x450
>> [ 1084.820230]  kthread+0x111/0x130
>> [ 1084.820236]  ? kthread_complete_and_exit+0x20/0x20
>> [ 1084.820244]  ret_from_fork+0x22/0x30
>> [ 1084.820258]  </TASK>
>> [ 1084.820260] BUG: sleeping function called from invalid context at
>> kernel/locking/spinlock_rt.c:46
>>
>> It shows up with PREEMPT_RT kernel.
> Oh, I see ..
>
>> Maybe. One thing that I am not clear about is why user_cpus_ptr is set in
>> the first place.
> Perhaps someone set an affinity on kthreadd ?
>
> But I'm thinking this exact problem is also possible (rather more likely
> even) with select_fallback_rq() that too holds pi_lock (which account
> for both other users of this function).
>
> Bah.
>
> And the allocation is just the one long in size (for small configs)
> which is just enough space for a single linked list like you had.
That is exactly the reason why I use lockless list.
>
> Urgh.
>
> The below is yuck too, and I'm not sure Paul wants us to use
> kvfree_call_rcu() without its wrapper.
>
> ---
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 78b2d5cabcc5..0d0af0fc7fcf 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2606,7 +2606,12 @@ void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
>   	};
>   
>   	__do_set_cpus_allowed(p, &ac);
> -	kfree(ac.user_mask);
> +	/*
> +	 * Because this is called with p->pi_lock held, it is not possible
> +	 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
> +	 * kfree_rcu().
> +	 */
> +	kvfree_call_rcu((struct rcu_head *)ac.user_mask, (rcu_callback_t)0);
>   }

I guess you need to do a NULL check before calling kvfree_call_rcu() as 
I don't think kvfree_call_rcu() does that. Also it is unlikely that we 
need to call it.

>   
>   int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
> @@ -8196,7 +8201,7 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
>   	struct affinity_context ac;
>   	struct cpumask *user_mask;
>   	struct task_struct *p;
> -	int retval;
> +	int retval, size;
>   
>   	rcu_read_lock();
>   
> @@ -8229,7 +8234,11 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
>   	if (retval)
>   		goto out_put_task;
>   
> -	user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
> +	/*
> +	 * See do_set_cpus_allowed() for the rcu_head usage.
> +	 */
> +	size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
> +	user_mask = kmalloc(size, GFP_KERNEL);
>   	if (!user_mask) {
>   		retval = -ENOMEM;
>   		goto out_put_task;

I guess that will work too. Just like you, I am a bit uneasy to call 
into kvfree_call_rcu() directly as it may change in the future. How about

iff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 78b2d5cabcc5..5fac4aa6ac7f 100644
--- a/kernel/sched/core.c
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 78b2d5cabcc5..5fac4aa6ac7f 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2593,6 +2593,11 @@ __do_set_cpus_allowed(struct task_struct *p, 
struct affinity_context *ctx)
                 set_next_task(rq, p);
  }

+union cpumask_rcuhead {
+       void *cpumask;
+       struct rcu_head rcu;
+};
+
  /*
   * Used for kthread_bind() and select_fallback_rq(), in both cases the 
user
   * affinity (if any) should be destroyed too.
@@ -2606,7 +2611,12 @@ void do_set_cpus_allowed(struct task_struct *p, 
const struct cpumask *new_mask)
         };

         __do_set_cpus_allowed(p, &ac);
-       kfree(ac.user_mask);
+       /*
+        * Because this is called with p->pi_lock held, it is not possible
+        * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
+        * kfree_rcu().
+        */
+       kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
  }

  int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
@@ -8196,7 +8206,7 @@ long sched_setaffinity(pid_t pid, const struct 
cpumask *in_mask)
         struct affinity_context ac;
         struct cpumask *user_mask;
         struct task_struct *p;
-       int retval;
+       int retval, size;

         rcu_read_lock();

@@ -8229,7 +8239,11 @@ long sched_setaffinity(pid_t pid, const struct 
cpumask *in_mask)
         if (retval)
                 goto out_put_task;

-       user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
+       /*
+        * See do_set_cpus_allowed() for the rcu_head usage.
+        */
+       size = max_t(int, cpumask_size(), sizeof(union cpumask_rcuhead));
+       user_mask = kmalloc(size, GFP_KERNEL);
         if (!user_mask) {
                 retval = -ENOMEM;
                 goto out_put_task;

Cheers,
Longman