[PATCH RFC] possible atomicity issue in ptr_ring_resize_multiple_bh_noprof

[PATCH RFC] possible atomicity issue in ptr_ring_resize_multiple_bh_noprof

[Dheeraj Reddy Jonnalagadda]

Hi Maintainers,

While reviewing the ptr_ring_resize_multiple_bh_noprof function, I noticed 
a potential atomicity issue. The function appears to be callable from multiple
threads, based on the locking patterns and _bh suffix.

The current code frees queues[i] after releasing locks:

    for (i = 0; i < nrings; ++i) {
        spin_lock_bh(&(rings[i])->consumer_lock);
        spin_lock(&(rings[i])->producer_lock);
        queues[i] = __ptr_ring_swap_queue(rings[i], queues[i],
                                         size, gfp, destroy);
        spin_unlock(&(rings[i])->producer_lock);
        spin_unlock_bh(&(rings[i])->consumer_lock);
    }

    /* Free after releasing locks */
    for (i = 0; i < nrings; ++i)
        kvfree(queues[i]);

It seems that there could be a race condition where another thread modifies
queues[i] between the unlock and the kvfree. Would it be safer to do the
kvfree while still holding the locks and removing the kvfree loop later
as shown below?

    for (i = 0; i < nrings; ++i) {
        spin_lock_bh(&(rings[i])->consumer_lock);
        spin_lock(&(rings[i])->producer_lock);
        queues[i] = __ptr_ring_swap_queue(rings[i], queues[i],
                                         size, gfp, destroy);
        kvfree(queues[i]);
        spin_unlock(&(rings[i])->producer_lock);
        spin_unlock_bh(&(rings[i])->consumer_lock);
    }

    kfree(queues);

    return 0;

I've attached a potential fix, but would appreciate confirmation on whether
this is actually an issue that needs addressing.

-Dheeraj

[PATCH include] ptr_ring: fix potential race in ptr_ring_resize_multiple_bh_noprof

[Dheeraj Reddy Jonnalagadda]

The ptr_ring_resize_multiple_bh_noprof function may have a race condition
where queues[i] is freed after releasing the locks. Since this function
can be called from multiple threads, another thread could potentially modify
queues[i] between the unlock and kvfree operations.

Move the kvfree inside the critical section to ensure atomicity of the
queue swap and cleanup operations.

Fixes: 59e6ae53248a("ptr_ring: support resizing multiple queues")
Closes: https://scan7.scan.coverity.com/#/project-view/52337/11354?selectedIssue=1602644
Signed-off-by: Dheeraj Reddy Jonnalagadda <dheeraj.linuxdev@gmail.com>
---
 include/linux/ptr_ring.h | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h
index 551329220e4f..cb06c74fc791 100644
--- a/include/linux/ptr_ring.h
+++ b/include/linux/ptr_ring.h
@@ -641,13 +641,11 @@ static inline int ptr_ring_resize_multiple_bh_noprof(struct ptr_ring **rings,
 		spin_lock(&(rings[i])->producer_lock);
 		queues[i] = __ptr_ring_swap_queue(rings[i], queues[i],
 						  size, gfp, destroy);
+		kvfree(queues[i]);
 		spin_unlock(&(rings[i])->producer_lock);
 		spin_unlock_bh(&(rings[i])->consumer_lock);
 	}
 
-	for (i = 0; i < nrings; ++i)
-		kvfree(queues[i]);
-
 	kfree(queues);
 
 	return 0;
-- 
2.34.1

Re: [PATCH include] ptr_ring: fix potential race in ptr_ring_resize_multiple_bh_noprof

[Eric Dumazet]

On Sat, Dec 21, 2024 at 7:16 AM Dheeraj Reddy Jonnalagadda
<dheeraj.linuxdev@gmail.com> wrote:
>
> The ptr_ring_resize_multiple_bh_noprof function may have a race condition
> where queues[i] is freed after releasing the locks. Since this function
> can be called from multiple threads, another thread could potentially modify
> queues[i] between the unlock and kvfree operations.
>
> Move the kvfree inside the critical section to ensure atomicity of the
> queue swap and cleanup operations.
>
> Fixes: 59e6ae53248a("ptr_ring: support resizing multiple queues")
> Closes: https://scan7.scan.coverity.com/#/project-view/52337/11354?selectedIssue=1602644
> Signed-off-by: Dheeraj Reddy Jonnalagadda <dheeraj.linuxdev@gmail.com>
> ---
>  include/linux/ptr_ring.h | 4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)
>
> diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h
> index 551329220e4f..cb06c74fc791 100644
> --- a/include/linux/ptr_ring.h
> +++ b/include/linux/ptr_ring.h
> @@ -641,13 +641,11 @@ static inline int ptr_ring_resize_multiple_bh_noprof(struct ptr_ring **rings,
>                 spin_lock(&(rings[i])->producer_lock);
>                 queues[i] = __ptr_ring_swap_queue(rings[i], queues[i],
>                                                   size, gfp, destroy);
> +               kvfree(queues[i]);
>                 spin_unlock(&(rings[i])->producer_lock);
>                 spin_unlock_bh(&(rings[i])->consumer_lock);
>         }
>
> -       for (i = 0; i < nrings; ++i)
> -               kvfree(queues[i]);
> -
>         kfree(queues);

I do not think this patch makes sense.

queues[] is private to this thread, there is no possible race.