Previously, seccomp filters were released in release_task(), which
required the process to exit and its zombie to be collected. However,
exited threads/processes can't trigger any seccomp events, making it
more logical to release filters upon task exits.

This adjustment simplifies scenarios where a parent is tracing its child
process. The parent process can now handle all events from a seccomp
listening descriptor and then call wait to collect a child zombie.

seccomp_filter_release takes the siglock to avoid races with
seccomp_sync_threads. There was an idea to bypass taking the lock by
checking PF_EXITING, but it can be set without holding siglock if
threads have SIGNAL_GROUP_EXIT. This means it can happen concurently
with seccomp_filter_release.

Signed-off-by: Andrei Vagin <avagin@google.com>
---
 kernel/exit.c    |  3 ++-
 kernel/seccomp.c | 22 ++++++++++++++++------
 2 files changed, 18 insertions(+), 7 deletions(-)

diff --git a/kernel/exit.c b/kernel/exit.c
index 41a12630cbbc..23439c021d8d 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -278,7 +278,6 @@ void release_task(struct task_struct *p)
 	}
 
 	write_unlock_irq(&tasklist_lock);
-	seccomp_filter_release(p);
 	proc_flush_pid(thread_pid);
 	put_pid(thread_pid);
 	release_thread(p);
@@ -836,6 +835,8 @@ void __noreturn do_exit(long code)
 	io_uring_files_cancel();
 	exit_signals(tsk);  /* sets PF_EXITING */
 
+	seccomp_filter_release(tsk);
+
 	acct_update_integrals(tsk);
 	group_dead = atomic_dec_and_test(&tsk->signal->live);
 	if (group_dead) {
diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 35435e8f1035..67305e776dd3 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -502,6 +502,9 @@ static inline pid_t seccomp_can_sync_threads(void)
 		/* Skip current, since it is initiating the sync. */
 		if (thread == caller)
 			continue;
+		/* Skip exited threads. */
+		if (thread->flags & PF_EXITING)
+			continue;
 
 		if (thread->seccomp.mode == SECCOMP_MODE_DISABLED ||
 		    (thread->seccomp.mode == SECCOMP_MODE_FILTER &&
@@ -563,18 +566,18 @@ static void __seccomp_filter_release(struct seccomp_filter *orig)
  * @tsk: task the filter should be released from.
  *
  * This function should only be called when the task is exiting as
- * it detaches it from its filter tree. As such, READ_ONCE() and
- * barriers are not needed here, as would normally be needed.
+ * it detaches it from its filter tree. PF_EXITING has to be set
+ * for the task.
  */
 void seccomp_filter_release(struct task_struct *tsk)
 {
-	struct seccomp_filter *orig = tsk->seccomp.filter;
-
-	/* We are effectively holding the siglock by not having any sighand. */
-	WARN_ON(tsk->sighand != NULL);
+	struct seccomp_filter *orig;
 
+	spin_lock_irq(&current->sighand->siglock);
+	orig = tsk->seccomp.filter;
 	/* Detach task from its filter tree. */
 	tsk->seccomp.filter = NULL;
+	spin_unlock_irq(&current->sighand->siglock);
 	__seccomp_filter_release(orig);
 }
 
@@ -602,6 +605,13 @@ static inline void seccomp_sync_threads(unsigned long flags)
 		if (thread == caller)
 			continue;
 
+		/*
+		 * Skip exited threads. seccomp_filter_release could have
+		 * been already called for this task.
+		 */
+		if (thread->flags & PF_EXITING)
+			continue;
+
 		/* Get a task reference for the new leaf node. */
 		get_seccomp_filter(caller);
 
-- 
2.45.1.288.g0e0cd299f1-goog

On Thu, May 23, 2024 at 01:45:39AM +0000, Andrei Vagin wrote:
> Previously, seccomp filters were released in release_task(), which
> required the process to exit and its zombie to be collected. However,
> exited threads/processes can't trigger any seccomp events, making it
> more logical to release filters upon task exits.
> 
> This adjustment simplifies scenarios where a parent is tracing its child
> process. The parent process can now handle all events from a seccomp
> listening descriptor and then call wait to collect a child zombie.
> 
> seccomp_filter_release takes the siglock to avoid races with
> seccomp_sync_threads. There was an idea to bypass taking the lock by
> checking PF_EXITING, but it can be set without holding siglock if
> threads have SIGNAL_GROUP_EXIT. This means it can happen concurently
> with seccomp_filter_release.
> 
> Signed-off-by: Andrei Vagin <avagin@google.com>
> ---
>  kernel/exit.c    |  3 ++-
>  kernel/seccomp.c | 22 ++++++++++++++++------
>  2 files changed, 18 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/exit.c b/kernel/exit.c
> index 41a12630cbbc..23439c021d8d 100644
> --- a/kernel/exit.c
> +++ b/kernel/exit.c
> @@ -278,7 +278,6 @@ void release_task(struct task_struct *p)
>  	}
>  
>  	write_unlock_irq(&tasklist_lock);
> -	seccomp_filter_release(p);
>  	proc_flush_pid(thread_pid);
>  	put_pid(thread_pid);
>  	release_thread(p);
> @@ -836,6 +835,8 @@ void __noreturn do_exit(long code)
>  	io_uring_files_cancel();
>  	exit_signals(tsk);  /* sets PF_EXITING */
>  
> +	seccomp_filter_release(tsk);
> +
>  	acct_update_integrals(tsk);
>  	group_dead = atomic_dec_and_test(&tsk->signal->live);
>  	if (group_dead) {
> diff --git a/kernel/seccomp.c b/kernel/seccomp.c
> index 35435e8f1035..67305e776dd3 100644
> --- a/kernel/seccomp.c
> +++ b/kernel/seccomp.c
> @@ -502,6 +502,9 @@ static inline pid_t seccomp_can_sync_threads(void)
>  		/* Skip current, since it is initiating the sync. */
>  		if (thread == caller)
>  			continue;
> +		/* Skip exited threads. */
> +		if (thread->flags & PF_EXITING)
> +			continue;
>  
>  		if (thread->seccomp.mode == SECCOMP_MODE_DISABLED ||
>  		    (thread->seccomp.mode == SECCOMP_MODE_FILTER &&
> @@ -563,18 +566,18 @@ static void __seccomp_filter_release(struct seccomp_filter *orig)
>   * @tsk: task the filter should be released from.
>   *
>   * This function should only be called when the task is exiting as
> - * it detaches it from its filter tree. As such, READ_ONCE() and
> - * barriers are not needed here, as would normally be needed.
> + * it detaches it from its filter tree. PF_EXITING has to be set
> + * for the task.

Let's capture this requirement with a WARN_ON() (like was done for the
sighand case before). So before the spinlock, check for PF_EXITING and
fail safe (don't release):

	if (WARN_ON((tsk->flags & PF_EXITING) == 0))
		return;

>   */
>  void seccomp_filter_release(struct task_struct *tsk)
>  {
> -	struct seccomp_filter *orig = tsk->seccomp.filter;
> -
> -	/* We are effectively holding the siglock by not having any sighand. */
> -	WARN_ON(tsk->sighand != NULL);
> +	struct seccomp_filter *orig;
>  
> +	spin_lock_irq(&current->sighand->siglock);

Shouldn't this be "tsk" not "current"?

> +	orig = tsk->seccomp.filter;
>  	/* Detach task from its filter tree. */
>  	tsk->seccomp.filter = NULL;
> +	spin_unlock_irq(&current->sighand->siglock);

Same.

>  	__seccomp_filter_release(orig);
>  }
>  
> @@ -602,6 +605,13 @@ static inline void seccomp_sync_threads(unsigned long flags)
>  		if (thread == caller)
>  			continue;
>  
> +		/*
> +		 * Skip exited threads. seccomp_filter_release could have
> +		 * been already called for this task.
> +		 */
> +		if (thread->flags & PF_EXITING)
> +			continue;
> +
>  		/* Get a task reference for the new leaf node. */
>  		get_seccomp_filter(caller);
>  
> -- 
> 2.45.1.288.g0e0cd299f1-goog
> 

Otherwise, looks good!

-- 
Kees Cook

On 05/23, Andrei Vagin wrote:
>
> Previously, seccomp filters were released in release_task(), which
> required the process to exit and its zombie to be collected. However,
> exited threads/processes can't trigger any seccomp events, making it
> more logical to release filters upon task exits.
>
> This adjustment simplifies scenarios where a parent is tracing its child
> process. The parent process can now handle all events from a seccomp
> listening descriptor and then call wait to collect a child zombie.
>
> seccomp_filter_release takes the siglock to avoid races with
> seccomp_sync_threads. There was an idea to bypass taking the lock by
> checking PF_EXITING, but it can be set without holding siglock if
> threads have SIGNAL_GROUP_EXIT. This means it can happen concurently
> with seccomp_filter_release.
>
> Signed-off-by: Andrei Vagin <avagin@google.com>
> ---
>  kernel/exit.c    |  3 ++-
>  kernel/seccomp.c | 22 ++++++++++++++++------
>  2 files changed, 18 insertions(+), 7 deletions(-)

Reviewed-by: Oleg Nesterov <oleg@redhat.com>