[PATCH] sched: Fix NULL user_cpus_ptr check in dup_user_cpus_ptr()

[PATCH] sched: Fix NULL user_cpus_ptr check in dup_user_cpus_ptr()

[Waiman Long]

In general, a non-null user_cpus_ptr will remain set until the task dies.
A possible exception to this is the fact that do_set_cpus_allowed()
will clear a non-null user_cpus_ptr. To allow this possible racing
condition, we need to check for NULL user_cpus_ptr under the pi_lock
before duping the user mask.

Fixes: 851a723e45d1 ("sched: Always clear user_cpus_ptr in do_set_cpus_allowed()")
Signed-off-by: Waiman Long <longman@redhat.com>
---
 kernel/sched/core.c | 23 +++++++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 8df51b08bb38..f447a6285ea2 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2625,7 +2625,14 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
 		      int node)
 {
 	unsigned long flags;
+	cpumask_t *user_mask = NULL;
 
+	/*
+	 * If there is a concurrent sched_setaffinity(), we may miss the
+	 * newly updated user_cpus_ptr. However, a non-NULL user_cpus_ptr
+	 * is relatively unlikely and it is not worth the extra overhead
+	 * of taking the pi_lock on every fork/clone.
+	 */
 	if (!src->user_cpus_ptr)
 		return 0;
 
@@ -2633,10 +2640,22 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
 	if (!dst->user_cpus_ptr)
 		return -ENOMEM;
 
-	/* Use pi_lock to protect content of user_cpus_ptr */
+	/*
+	 * Use pi_lock to protect content of user_cpus_ptr
+	 *
+	 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
+	 * do_set_cpus_allowed().
+	 */
 	raw_spin_lock_irqsave(&src->pi_lock, flags);
-	cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
+	if (src->user_cpus_ptr)
+		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
+	else
+		swap(dst->user_cpus_ptr, user_mask);
 	raw_spin_unlock_irqrestore(&src->pi_lock, flags);
+
+	if (unlikely(user_mask))
+		kfree(user_mask);
+
 	return 0;
 }
 
-- 
2.31.1

Re: [PATCH] sched: Fix NULL user_cpus_ptr check in dup_user_cpus_ptr()

[Peter Zijlstra]

On Tue, Nov 22, 2022 at 02:06:53PM -0500, Waiman Long wrote:
> In general, a non-null user_cpus_ptr will remain set until the task dies.
> A possible exception to this is the fact that do_set_cpus_allowed()
> will clear a non-null user_cpus_ptr. To allow this possible racing
> condition, we need to check for NULL user_cpus_ptr under the pi_lock
> before duping the user mask.
> 
> Fixes: 851a723e45d1 ("sched: Always clear user_cpus_ptr in do_set_cpus_allowed()")
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  kernel/sched/core.c | 23 +++++++++++++++++++++--
>  1 file changed, 21 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 8df51b08bb38..f447a6285ea2 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2625,7 +2625,14 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
>  		      int node)
>  {
>  	unsigned long flags;
> +	cpumask_t *user_mask = NULL;

The inverse xmas tree is sad :-(

>  
> +	/*
> +	 * If there is a concurrent sched_setaffinity(), we may miss the
> +	 * newly updated user_cpus_ptr. However, a non-NULL user_cpus_ptr
> +	 * is relatively unlikely and it is not worth the extra overhead
> +	 * of taking the pi_lock on every fork/clone.
> +	 */

I think the correct argument is saying the thing is racy and loosing the
race is a valid situation. IOW, this is the same as if the concurrent
sched_setaffinity() happens after fork().

>  	if (!src->user_cpus_ptr)
>  		return 0;
>  
> @@ -2633,10 +2640,22 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
>  	if (!dst->user_cpus_ptr)
>  		return -ENOMEM;
>  
> -	/* Use pi_lock to protect content of user_cpus_ptr */
> +	/*
> +	 * Use pi_lock to protect content of user_cpus_ptr
> +	 *
> +	 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
> +	 * do_set_cpus_allowed().
> +	 */
>  	raw_spin_lock_irqsave(&src->pi_lock, flags);
> +	if (src->user_cpus_ptr)
> +		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
> +	else
> +		swap(dst->user_cpus_ptr, user_mask);

Uhhhh, did you mean to write:

	if (src->user_cpus_ptr) {
		cpumask_copy(user_mask, src->user_cpus_ptr);
		swap(dst->user_cpus_ptr, user_mask);
	}

?

>  	raw_spin_unlock_irqrestore(&src->pi_lock, flags);
> +
> +	if (unlikely(user_mask))
> +		kfree(user_mask);
> +
>  	return 0;
>  }
>  
> -- 
> 2.31.1
> 

Re: [PATCH] sched: Fix NULL user_cpus_ptr check in dup_user_cpus_ptr()

[Waiman Long]

On 11/23/22 06:20, Peter Zijlstra wrote:
> On Tue, Nov 22, 2022 at 02:06:53PM -0500, Waiman Long wrote:
>> In general, a non-null user_cpus_ptr will remain set until the task dies.
>> A possible exception to this is the fact that do_set_cpus_allowed()
>> will clear a non-null user_cpus_ptr. To allow this possible racing
>> condition, we need to check for NULL user_cpus_ptr under the pi_lock
>> before duping the user mask.
>>
>> Fixes: 851a723e45d1 ("sched: Always clear user_cpus_ptr in do_set_cpus_allowed()")
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>   kernel/sched/core.c | 23 +++++++++++++++++++++--
>>   1 file changed, 21 insertions(+), 2 deletions(-)
>>
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 8df51b08bb38..f447a6285ea2 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -2625,7 +2625,14 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
>>   		      int node)
>>   {
>>   	unsigned long flags;
>> +	cpumask_t *user_mask = NULL;
> The inverse xmas tree is sad :-(
Right. The inverse xmas tree rule. Will fix that.
>
>>   
>> +	/*
>> +	 * If there is a concurrent sched_setaffinity(), we may miss the
>> +	 * newly updated user_cpus_ptr. However, a non-NULL user_cpus_ptr
>> +	 * is relatively unlikely and it is not worth the extra overhead
>> +	 * of taking the pi_lock on every fork/clone.
>> +	 */
> I think the correct argument is saying the thing is racy and loosing the
> race is a valid situation. IOW, this is the same as if the concurrent
> sched_setaffinity() happens after fork().
Good point, will update the comment.
>>   	if (!src->user_cpus_ptr)
>>   		return 0;
>>   
>> @@ -2633,10 +2640,22 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
>>   	if (!dst->user_cpus_ptr)
>>   		return -ENOMEM;
>>   
>> -	/* Use pi_lock to protect content of user_cpus_ptr */
>> +	/*
>> +	 * Use pi_lock to protect content of user_cpus_ptr
>> +	 *
>> +	 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
>> +	 * do_set_cpus_allowed().
>> +	 */
>>   	raw_spin_lock_irqsave(&src->pi_lock, flags);
>> +	if (src->user_cpus_ptr)
>> +		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
>> +	else
>> +		swap(dst->user_cpus_ptr, user_mask);
> Uhhhh, did you mean to write:
>
> 	if (src->user_cpus_ptr) {
> 		cpumask_copy(user_mask, src->user_cpus_ptr);
> 		swap(dst->user_cpus_ptr, user_mask);
> 	}
>
> ?

Not really. The point is that dst->user_cpus_ptr has been allocated. If 
src->user_cpus_ptr turns out to be NULL, we need to clear 
dst->user_cpus_ptr which is what the swap() does and then free that 
memory after unlock. Will add a comment to make this point clear.

Cheers,
Longman