[PATCH v2] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[PATCH v2] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Hui Peng]

In seccomp_notify_addfd(), when SECCOMP_ADDFD_FLAG_SEND is set,
knotif->state is transitioned from SECCOMP_NOTIFY_SENT to
SECCOMP_NOTIFY_REPLIED before dropping filter->notify_lock and waiting
in wait_for_completion_interruptible(&kaddfd.completion).

If wait_for_completion_interruptible() is interrupted by a signal before
the target task processes the kaddfd entry, seccomp_notify_addfd()
removes kaddfd.list from knotif->addfd via list_del(&kaddfd.list), but
leaves knotif->state set to SECCOMP_NOTIFY_REPLIED.

As a consequence:
1. When the target task wakes in seccomp_do_user_notification(), it
   observes knotif->state == SECCOMP_NOTIFY_REPLIED with an empty addfd
   list and zeroed knotif->val / knotif->error / knotif->flags, causing
   the supervised syscall to prematurely complete with return value 0
   without installing the file descriptor.
2. Any subsequent retry of SECCOMP_IOCTL_NOTIF_SEND or
   SECCOMP_IOCTL_NOTIF_ADDFD by the supervisor fails with -EINPROGRESS
   (or -ENOENT after the target exits) because knotif->state is no
   longer SECCOMP_NOTIFY_SENT.

Restore knotif->state = SECCOMP_NOTIFY_SENT when removing an unhandled
SECCOMP_ADDFD_FLAG_SEND request from knotif->addfd, and add a
kselftest (user_notification_addfd_send_interrupted) to cover this race.

Fixes: 7cf97b125455 ("seccomp: Introduce addfd ioctl to seccomp user notifier")
Reviewed-by: Bradley Morgan <brads@mainlining.org>
Assisted-by: LLM
Signed-off-by: Hui Peng <benquike@gmail.com>
---
Changes in v2:
- Added kselftest regression test (user_notification_addfd_send_interrupted)
  in tools/testing/selftests/seccomp/seccomp_bpf.c per Kees Cook.
- Collected Reviewed-by tag from Bradley Morgan.

 kernel/seccomp.c                              |  7 +-
 tools/testing/selftests/seccomp/seccomp_bpf.c | 98 +++++++++++++++++++
 2 files changed, 103 insertions(+), 2 deletions(-)

diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 86cf4460d69e..94c7ba80a8f7 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -1808,10 +1808,13 @@ static long seccomp_notify_addfd(struct seccomp_filter *filter,
 	 * We need to check again if the addfd request has been handled,
 	 * and if not, we will remove it from the queue.
 	 */
-	if (list_empty(&kaddfd.list))
+	if (list_empty(&kaddfd.list)) {
 		ret = kaddfd.ret;
-	else
+	} else {
 		list_del(&kaddfd.list);
+		if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND)
+			knotif->state = SECCOMP_NOTIFY_SENT;
+	}
 
 out_unlock:
 	mutex_unlock(&filter->notify_lock);
-- 
2.47.3

[PATCH v3] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Hui Peng]

In seccomp_notify_addfd(), when SECCOMP_ADDFD_FLAG_SEND is set in
addfd.flags, knotif->state is optimistically transitioned from
SECCOMP_NOTIFY_SENT to SECCOMP_NOTIFY_REPLIED before waking the tracee
and waiting on kaddfd.completion:

	if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND) {
		knotif->state = SECCOMP_NOTIFY_REPLIED;
		...
	}

If wait_for_completion_interruptible(&kaddfd.completion) is interrupted
by a signal before the tracee dequeues the kaddfd request
(!list_empty(&kaddfd.list)), seccomp_notify_addfd() removes kaddfd from
knotif->addfd and returns -EINTR to the supervisor without having
installed the file descriptor.

However, knotif->state is left as SECCOMP_NOTIFY_REPLIED (with
knotif->error == 0 and knotif->val == 0). This causes two problems:

1. When the tracee runs in do_user_notif(), it checks
   if (knotif.state != SECCOMP_NOTIFY_REPLIED) after calling
   seccomp_handle_addfd(). Because knotif->state is already
   SECCOMP_NOTIFY_REPLIED, the tracee exits the notification wait loop
   prematurely and returns 0 from the trapped syscall without the file
   descriptor ever having been installed.

2. If the supervisor retries SECCOMP_IOCTL_NOTIF_ADDFD or
   SECCOMP_IOCTL_NOTIF_SEND before the tracee runs, it fails with
   -EINPROGRESS because knotif->state is no longer SECCOMP_NOTIFY_SENT.

Fix this by restoring knotif->state back to SECCOMP_NOTIFY_SENT if
SECCOMP_ADDFD_FLAG_SEND was set and kaddfd was not consumed before the
interrupted wait. Also add a seccomp_bpf selftest
(user_notification_addfd_send_interrupted) covering this race.

Fixes: 0ae71c1990f3 ("seccomp: Add a flag to atomically addfd and perform send")
Signed-off-by: Hui Peng <benquike@gmail.com>
---
Changes in v3:
- Actually include the tools/testing/selftests/seccomp/seccomp_bpf.c
  regression test in the patch diff (v2 accidentally omitted the selftest
  hunk), with detailed comments explaining the race and test setup.

Changes in v2:
- Add user_notification_addfd_send_interrupted regression test to
  tools/testing/selftests/seccomp/seccomp_bpf.c as requested by Kees Cook.

 kernel/seccomp.c                              |   2 +
 tools/testing/selftests/seccomp/seccomp_bpf.c | 161 ++++++++++++++++++
 2 files changed, 163 insertions(+)

diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 86cf4460d69e..94c7ba80a8f7 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -1808,10 +1808,13 @@ static long seccomp_notify_addfd(struct seccomp_filter *filter,
 	 * We need to check again if the addfd request has been handled,
 	 * and if not, we will remove it from the queue.
 	 */
-	if (list_empty(&kaddfd.list))
+	if (list_empty(&kaddfd.list)) {
 		ret = kaddfd.ret;
-	else
+	} else {
 		list_del(&kaddfd.list);
+		if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND)
+			knotif->state = SECCOMP_NOTIFY_SENT;
+	}
 
 out_unlock:
 	mutex_unlock(&filter->notify_lock);
diff --git a/tools/testing/selftests/seccomp/seccomp_bpf.c b/tools/testing/selftests/seccomp/seccomp_bpf.c
index 0622bc2acad4..e565728f9eb2 100644
--- a/tools/testing/selftests/seccomp/seccomp_bpf.c
+++ b/tools/testing/selftests/seccomp/seccomp_bpf.c
@@ -4368,6 +4368,173 @@ TEST(user_notification_addfd_rlimit)
 	close(memfd);
 }
 
+static void sigusr1_handler(int signo)
+{
+}
+
+/*
+ * Verify that when SECCOMP_IOCTL_NOTIF_ADDFD with SECCOMP_ADDFD_FLAG_SEND is
+ * interrupted by a signal before the tracee dequeues the addfd request,
+ * knotif->state is restored from SECCOMP_NOTIFY_REPLIED back to
+ * SECCOMP_NOTIFY_SENT so that:
+ *   1. The woken tracee sees knotif->state == SECCOMP_NOTIFY_SENT in
+ *      do_user_notif() and goes back to sleep instead of prematurely
+ *      returning 0 from the trapped syscall without the FD installed.
+ *   2. The supervisor can retry SECCOMP_IOCTL_NOTIF_ADDFD (or
+ *      SECCOMP_IOCTL_NOTIF_SEND) instead of failing with -EINPROGRESS.
+ *
+ * To deterministically hit the race window where the supervisor sleeps in
+ * wait_for_completion_interruptible(&kaddfd.completion) after waking the
+ * tracee (complete(&knotif->ready)) but before the tracee runs
+ * seccomp_handle_addfd(), pin all processes to a single CPU and enforce a
+ * strict 3-tier scheduling priority hierarchy on that CPU:
+ *   - Supervisor:            SCHED_FIFO priority 99 (highest)
+ *   - Signal helper (sig_pid): SCHED_FIFO priority 50 (middle)
+ *   - Tracee (pid):          SCHED_IDLE             (lowest)
+ */
+TEST(user_notification_addfd_send_interrupted)
+{
+	/*
+	 * Save parent_pid before user_notif_syscall(__NR_getppid, ...) installs
+	 * the seccomp filter on the calling process; children inherit that
+	 * filter, so sig_pid must not call getppid().
+	 */
+	pid_t pid, sig_pid, parent_pid = getpid();
+	long ret;
+	int status, listener, memfd;
+	struct seccomp_notif_addfd addfd = {};
+	struct seccomp_notif req = {};
+	struct sigaction sa = {};
+	struct sched_param sp_tracee_idle = { .sched_priority = 0 };
+	struct sched_param sp_supervisor_fifo = { .sched_priority = 99 };
+	struct sched_param sp_sig_helper_fifo = { .sched_priority = 50 };
+	struct timespec delay = { .tv_nsec = 15000000 };
+	cpu_set_t cpuset;
+	int cpu;
+
+	/* Pin the supervisor (and its future child processes) to one CPU. */
+	cpu = sched_getcpu();
+	if (cpu >= 0) {
+		CPU_ZERO(&cpuset);
+		CPU_SET(cpu, &cpuset);
+		sched_setaffinity(0, sizeof(cpuset), &cpuset);
+	}
+
+	sa.sa_handler = sigusr1_handler;
+	ASSERT_EQ(sigaction(SIGUSR1, &sa, NULL), 0);
+
+	memfd = memfd_create("test", 0);
+	ASSERT_GE(memfd, 0);
+
+	ret = prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0);
+	ASSERT_EQ(0, ret) {
+		TH_LOG("Kernel does not support PR_SET_NO_NEW_PRIVS!");
+	}
+
+	/*
+	 * Follow the convention of other user_notification_* tests in this
+	 * file by trapping __NR_getppid: because the filter is installed on
+	 * the supervisor before fork(), the trapped syscall must be a
+	 * side-effect-free syscall that the supervisor itself never invokes.
+	 * Even though getppid() does not normally return an FD,
+	 * SECCOMP_ADDFD_FLAG_SEND replaces the trapped syscall's return value
+	 * with the newly installed FD number (42).
+	 */
+	listener = user_notif_syscall(__NR_getppid,
+				      SECCOMP_FILTER_FLAG_NEW_LISTENER);
+	ASSERT_GE(listener, 0);
+
+	pid = fork();
+	ASSERT_GE(pid, 0);
+
+	if (pid == 0) {
+		/*
+		 * Tracee: invoke __NR_getppid as a dummy trigger syscall to
+		 * trap into do_user_notif(). Verify that the syscall returns
+		 * the injected FD number (42) and that FD 42 is open.
+		 */
+		ret = syscall(__NR_getppid);
+		exit(ret != 42 || fcntl(42, F_GETFD) < 0);
+	}
+
+	/* Wait for the tracee to trap in do_user_notif(). */
+	ASSERT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_RECV, &req), 0);
+
+	/*
+	 * Demote the tracee to SCHED_IDLE and promote the supervisor to
+	 * SCHED_FIFO(99) on the same CPU.
+	 */
+	sched_setscheduler(pid, SCHED_IDLE, &sp_tracee_idle);
+	if (sched_setscheduler(0, SCHED_FIFO, &sp_supervisor_fifo) != 0) {
+		close(listener);
+		close(memfd);
+		kill(pid, SIGKILL);
+		waitpid(pid, NULL, 0);
+		SKIP(return, "SCHED_FIFO requires CAP_SYS_NICE");
+	}
+
+	addfd.srcfd = memfd;
+	addfd.newfd_flags = O_CLOEXEC;
+	addfd.newfd = 42;
+	addfd.id = req.id;
+	addfd.flags = SECCOMP_ADDFD_FLAG_SETFD | SECCOMP_ADDFD_FLAG_SEND;
+
+	/*
+	 * Fork a signal helper on the same CPU and set it to SCHED_FIFO(50).
+	 * Because the supervisor is currently running at SCHED_FIFO(99) on
+	 * this CPU, sig_pid is queued on the runqueue but cannot run until the
+	 * supervisor blocks inside the kernel.
+	 *
+	 * When the supervisor invokes ioctl(SECCOMP_IOCTL_NOTIF_ADDFD) below:
+	 *   1. seccomp_notify_addfd() sets knotif->state = SECCOMP_NOTIFY_REPLIED,
+	 *      wakes the tracee (SCHED_IDLE), and blocks in
+	 *      wait_for_completion_interruptible(&kaddfd.completion).
+	 *   2. The CPU scheduler immediately runs sig_pid (SCHED_FIFO 50)
+	 *      ahead of the woken tracee (SCHED_IDLE).
+	 *   3. sig_pid sends SIGUSR1 to parent_pid, waking the supervisor
+	 *      (SCHED_FIFO 99), which immediately preempts sig_pid, aborts the
+	 *      wait with -ERESTARTSYS (-EINTR), removes kaddfd from
+	 *      knotif->addfd, and restores knotif->state = SECCOMP_NOTIFY_SENT
+	 *      before the tracee has executed a single instruction.
+	 */
+	sig_pid = fork();
+	ASSERT_GE(sig_pid, 0);
+	if (sig_pid == 0) {
+		sched_setscheduler(0, SCHED_FIFO, &sp_sig_helper_fifo);
+		kill(parent_pid, SIGUSR1);
+		_exit(0);
+	}
+
+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), -1);
+	EXPECT_EQ(errno, EINTR);
+	EXPECT_EQ(waitpid(sig_pid, &status, 0), sig_pid);
+
+	/*
+	 * Restore normal scheduling and sleep briefly so the woken tracee
+	 * runs in do_user_notif(). With knotif->state restored to
+	 * SECCOMP_NOTIFY_SENT, the tracee must loop back to sleep waiting for
+	 * the notification reply rather than returning 0 from __NR_getppid.
+	 */
+	sched_setscheduler(0, SCHED_OTHER, &sp_tracee_idle);
+	sched_setscheduler(pid, SCHED_OTHER, &sp_tracee_idle);
+	nanosleep(&delay, NULL);
+
+	/*
+	 * Retry SECCOMP_IOCTL_NOTIF_ADDFD. Because knotif->state is
+	 * SECCOMP_NOTIFY_SENT, the retry succeeds (returns 42) instead of
+	 * failing with -EINPROGRESS, installs FD 42 into the tracee, and wakes
+	 * the tracee to complete the syscall with return value 42.
+	 */
+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), 42);
+
+	EXPECT_EQ(waitpid(pid, &status, 0), pid);
+	EXPECT_EQ(true, WIFEXITED(status));
+	EXPECT_EQ(0, WEXITSTATUS(status));
+
+	close(listener);
+	close(memfd);
+}
+
 #ifndef SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP
 #define SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP (1UL << 0)
 #define SECCOMP_IOCTL_NOTIF_SET_FLAGS  SECCOMP_IOW(4, __u64)
-- 
2.49.0

Re: [PATCH v3] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Bradley Morgan]

On 20 September 2026 08:20:36 BST, Hui Peng <benquike@gmail.com> wrote:
>In seccomp_notify_addfd(), when SECCOMP_ADDFD_FLAG_SEND is set in
>addfd.flags, knotif->state is optimistically transitioned from
>SECCOMP_NOTIFY_SENT to SECCOMP_NOTIFY_REPLIED before waking the tracee
>and waiting on kaddfd.completion:
>
>	if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND) {
>		knotif->state = SECCOMP_NOTIFY_REPLIED;
>		...
>	}
>
>If wait_for_completion_interruptible(&kaddfd.completion) is interrupted
>by a signal before the tracee dequeues the kaddfd request
>(!list_empty(&kaddfd.list)), seccomp_notify_addfd() removes kaddfd from
>knotif->addfd and returns -EINTR to the supervisor without having
>installed the file descriptor.
>
>However, knotif->state is left as SECCOMP_NOTIFY_REPLIED (with
>knotif->error == 0 and knotif->val == 0). This causes two problems:
>
>1. When the tracee runs in do_user_notif(), it checks
>   if (knotif.state != SECCOMP_NOTIFY_REPLIED) after calling
>   seccomp_handle_addfd(). Because knotif->state is already
>   SECCOMP_NOTIFY_REPLIED, the tracee exits the notification wait loop
>   prematurely and returns 0 from the trapped syscall without the file
>   descriptor ever having been installed.
>
>2. If the supervisor retries SECCOMP_IOCTL_NOTIF_ADDFD or
>   SECCOMP_IOCTL_NOTIF_SEND before the tracee runs, it fails with
>   -EINPROGRESS because knotif->state is no longer SECCOMP_NOTIFY_SENT.
>
>Fix this by restoring knotif->state back to SECCOMP_NOTIFY_SENT if
>SECCOMP_ADDFD_FLAG_SEND was set and kaddfd was not consumed before the
>interrupted wait. Also add a seccomp_bpf selftest
>(user_notification_addfd_send_interrupted) covering this race.
>
>Fixes: 0ae71c1990f3 ("seccomp: Add a flag to atomically addfd and perform send")

Oh my godd you

1: removed me from CC
2: removed my R-B tag!

What is your AI doing!

NAK. I can't review something like this!

Add a assisted by tag and CC me and I'll look at it


>Signed-off-by: Hui Peng <benquike@gmail.com>
>---
>Changes in v3:
>- Actually include the tools/testing/selftests/seccomp/seccomp_bpf.c
>  regression test in the patch diff (v2 accidentally omitted the selftest
>  hunk), with detailed comments explaining the race and test setup.
>
>Changes in v2:
>- Add user_notification_addfd_send_interrupted regression test to
>  tools/testing/selftests/seccomp/seccomp_bpf.c as requested by Kees Cook.
>
> kernel/seccomp.c                              |   2 +
> tools/testing/selftests/seccomp/seccomp_bpf.c | 161 ++++++++++++++++++
> 2 files changed, 163 insertions(+)
>
>diff --git a/kernel/seccomp.c b/kernel/seccomp.c
>index 86cf4460d69e..94c7ba80a8f7 100644
>--- a/kernel/seccomp.c
>+++ b/kernel/seccomp.c
>@@ -1808,10 +1808,13 @@ static long seccomp_notify_addfd(struct seccomp_filter *filter,
> 	 * We need to check again if the addfd request has been handled,
> 	 * and if not, we will remove it from the queue.
> 	 */
>-	if (list_empty(&kaddfd.list))
>+	if (list_empty(&kaddfd.list)) {
> 		ret = kaddfd.ret;
>-	else
>+	} else {
> 		list_del(&kaddfd.list);
>+		if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND)
>+			knotif->state = SECCOMP_NOTIFY_SENT;
>+	}
> 
> out_unlock:
> 	mutex_unlock(&filter->notify_lock);
>diff --git a/tools/testing/selftests/seccomp/seccomp_bpf.c b/tools/testing/selftests/seccomp/seccomp_bpf.c
>index 0622bc2acad4..e565728f9eb2 100644
>--- a/tools/testing/selftests/seccomp/seccomp_bpf.c
>+++ b/tools/testing/selftests/seccomp/seccomp_bpf.c
>@@ -4368,6 +4368,173 @@ TEST(user_notification_addfd_rlimit)
> 	close(memfd);
> }
> 
>+static void sigusr1_handler(int signo)
>+{
>+}
>+
>+/*
>+ * Verify that when SECCOMP_IOCTL_NOTIF_ADDFD with SECCOMP_ADDFD_FLAG_SEND is
>+ * interrupted by a signal before the tracee dequeues the addfd request,
>+ * knotif->state is restored from SECCOMP_NOTIFY_REPLIED back to
>+ * SECCOMP_NOTIFY_SENT so that:
>+ *   1. The woken tracee sees knotif->state == SECCOMP_NOTIFY_SENT in
>+ *      do_user_notif() and goes back to sleep instead of prematurely
>+ *      returning 0 from the trapped syscall without the FD installed.
>+ *   2. The supervisor can retry SECCOMP_IOCTL_NOTIF_ADDFD (or
>+ *      SECCOMP_IOCTL_NOTIF_SEND) instead of failing with -EINPROGRESS.
>+ *
>+ * To deterministically hit the race window where the supervisor sleeps in
>+ * wait_for_completion_interruptible(&kaddfd.completion) after waking the
>+ * tracee (complete(&knotif->ready)) but before the tracee runs
>+ * seccomp_handle_addfd(), pin all processes to a single CPU and enforce a
>+ * strict 3-tier scheduling priority hierarchy on that CPU:
>+ *   - Supervisor:            SCHED_FIFO priority 99 (highest)
>+ *   - Signal helper (sig_pid): SCHED_FIFO priority 50 (middle)
>+ *   - Tracee (pid):          SCHED_IDLE             (lowest)
>+ */
>+TEST(user_notification_addfd_send_interrupted)
>+{
>+	/*
>+	 * Save parent_pid before user_notif_syscall(__NR_getppid, ...) installs
>+	 * the seccomp filter on the calling process; children inherit that
>+	 * filter, so sig_pid must not call getppid().
>+	 */
>+	pid_t pid, sig_pid, parent_pid = getpid();
>+	long ret;
>+	int status, listener, memfd;
>+	struct seccomp_notif_addfd addfd = {};
>+	struct seccomp_notif req = {};
>+	struct sigaction sa = {};
>+	struct sched_param sp_tracee_idle = { .sched_priority = 0 };
>+	struct sched_param sp_supervisor_fifo = { .sched_priority = 99 };
>+	struct sched_param sp_sig_helper_fifo = { .sched_priority = 50 };
>+	struct timespec delay = { .tv_nsec = 15000000 };
>+	cpu_set_t cpuset;
>+	int cpu;
>+
>+	/* Pin the supervisor (and its future child processes) to one CPU. */
>+	cpu = sched_getcpu();
>+	if (cpu >= 0) {
>+		CPU_ZERO(&cpuset);
>+		CPU_SET(cpu, &cpuset);
>+		sched_setaffinity(0, sizeof(cpuset), &cpuset);
>+	}
>+
>+	sa.sa_handler = sigusr1_handler;
>+	ASSERT_EQ(sigaction(SIGUSR1, &sa, NULL), 0);
>+
>+	memfd = memfd_create("test", 0);
>+	ASSERT_GE(memfd, 0);
>+
>+	ret = prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0);
>+	ASSERT_EQ(0, ret) {
>+		TH_LOG("Kernel does not support PR_SET_NO_NEW_PRIVS!");
>+	}
>+
>+	/*
>+	 * Follow the convention of other user_notification_* tests in this
>+	 * file by trapping __NR_getppid: because the filter is installed on
>+	 * the supervisor before fork(), the trapped syscall must be a
>+	 * side-effect-free syscall that the supervisor itself never invokes.
>+	 * Even though getppid() does not normally return an FD,
>+	 * SECCOMP_ADDFD_FLAG_SEND replaces the trapped syscall's return value
>+	 * with the newly installed FD number (42).
>+	 */
>+	listener = user_notif_syscall(__NR_getppid,
>+				      SECCOMP_FILTER_FLAG_NEW_LISTENER);
>+	ASSERT_GE(listener, 0);
>+
>+	pid = fork();
>+	ASSERT_GE(pid, 0);
>+
>+	if (pid == 0) {
>+		/*
>+		 * Tracee: invoke __NR_getppid as a dummy trigger syscall to
>+		 * trap into do_user_notif(). Verify that the syscall returns
>+		 * the injected FD number (42) and that FD 42 is open.
>+		 */
>+		ret = syscall(__NR_getppid);
>+		exit(ret != 42 || fcntl(42, F_GETFD) < 0);
>+	}
>+
>+	/* Wait for the tracee to trap in do_user_notif(). */
>+	ASSERT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_RECV, &req), 0);
>+
>+	/*
>+	 * Demote the tracee to SCHED_IDLE and promote the supervisor to
>+	 * SCHED_FIFO(99) on the same CPU.
>+	 */
>+	sched_setscheduler(pid, SCHED_IDLE, &sp_tracee_idle);
>+	if (sched_setscheduler(0, SCHED_FIFO, &sp_supervisor_fifo) != 0) {
>+		close(listener);
>+		close(memfd);
>+		kill(pid, SIGKILL);
>+		waitpid(pid, NULL, 0);
>+		SKIP(return, "SCHED_FIFO requires CAP_SYS_NICE");
>+	}
>+
>+	addfd.srcfd = memfd;
>+	addfd.newfd_flags = O_CLOEXEC;
>+	addfd.newfd = 42;
>+	addfd.id = req.id;
>+	addfd.flags = SECCOMP_ADDFD_FLAG_SETFD | SECCOMP_ADDFD_FLAG_SEND;
>+
>+	/*
>+	 * Fork a signal helper on the same CPU and set it to SCHED_FIFO(50).
>+	 * Because the supervisor is currently running at SCHED_FIFO(99) on
>+	 * this CPU, sig_pid is queued on the runqueue but cannot run until the
>+	 * supervisor blocks inside the kernel.
>+	 *
>+	 * When the supervisor invokes ioctl(SECCOMP_IOCTL_NOTIF_ADDFD) below:
>+	 *   1. seccomp_notify_addfd() sets knotif->state = SECCOMP_NOTIFY_REPLIED,
>+	 *      wakes the tracee (SCHED_IDLE), and blocks in
>+	 *      wait_for_completion_interruptible(&kaddfd.completion).
>+	 *   2. The CPU scheduler immediately runs sig_pid (SCHED_FIFO 50)
>+	 *      ahead of the woken tracee (SCHED_IDLE).
>+	 *   3. sig_pid sends SIGUSR1 to parent_pid, waking the supervisor
>+	 *      (SCHED_FIFO 99), which immediately preempts sig_pid, aborts the
>+	 *      wait with -ERESTARTSYS (-EINTR), removes kaddfd from
>+	 *      knotif->addfd, and restores knotif->state = SECCOMP_NOTIFY_SENT
>+	 *      before the tracee has executed a single instruction.
>+	 */
>+	sig_pid = fork();
>+	ASSERT_GE(sig_pid, 0);
>+	if (sig_pid == 0) {
>+		sched_setscheduler(0, SCHED_FIFO, &sp_sig_helper_fifo);
>+		kill(parent_pid, SIGUSR1);
>+		_exit(0);
>+	}
>+
>+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), -1);
>+	EXPECT_EQ(errno, EINTR);
>+	EXPECT_EQ(waitpid(sig_pid, &status, 0), sig_pid);
>+
>+	/*
>+	 * Restore normal scheduling and sleep briefly so the woken tracee
>+	 * runs in do_user_notif(). With knotif->state restored to
>+	 * SECCOMP_NOTIFY_SENT, the tracee must loop back to sleep waiting for
>+	 * the notification reply rather than returning 0 from __NR_getppid.
>+	 */
>+	sched_setscheduler(0, SCHED_OTHER, &sp_tracee_idle);
>+	sched_setscheduler(pid, SCHED_OTHER, &sp_tracee_idle);
>+	nanosleep(&delay, NULL);
>+
>+	/*
>+	 * Retry SECCOMP_IOCTL_NOTIF_ADDFD. Because knotif->state is
>+	 * SECCOMP_NOTIFY_SENT, the retry succeeds (returns 42) instead of
>+	 * failing with -EINPROGRESS, installs FD 42 into the tracee, and wakes
>+	 * the tracee to complete the syscall with return value 42.
>+	 */
>+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), 42);
>+
>+	EXPECT_EQ(waitpid(pid, &status, 0), pid);
>+	EXPECT_EQ(true, WIFEXITED(status));
>+	EXPECT_EQ(0, WEXITSTATUS(status));
>+
>+	close(listener);
>+	close(memfd);
>+}
>+
> #ifndef SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP
> #define SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP (1UL << 0)
> #define SECCOMP_IOCTL_NOTIF_SET_FLAGS  SECCOMP_IOW(4, __u64)
>

--- Thanks!
"I'm not a very positive person" - Linus torvalds

[PATCH v4] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Hui Peng]

In seccomp_notify_addfd(), when SECCOMP_ADDFD_FLAG_SEND is set in
addfd.flags, knotif->state is optimistically transitioned from
SECCOMP_NOTIFY_SENT to SECCOMP_NOTIFY_REPLIED before waking the tracee
and waiting on kaddfd.completion:

	if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND) {
		knotif->state = SECCOMP_NOTIFY_REPLIED;
		...
	}

If wait_for_completion_interruptible(&kaddfd.completion) is interrupted
by a signal before the tracee dequeues the kaddfd request
(!list_empty(&kaddfd.list)), seccomp_notify_addfd() removes kaddfd from
knotif->addfd and returns -EINTR to the supervisor without having
installed the file descriptor.

However, knotif->state is left as SECCOMP_NOTIFY_REPLIED (with
knotif->error == 0 and knotif->val == 0). This causes two problems:

1. When the tracee runs in do_user_notif(), it checks
   if (knotif.state != SECCOMP_NOTIFY_REPLIED) after calling
   seccomp_handle_addfd(). Because knotif->state is already
   SECCOMP_NOTIFY_REPLIED, the tracee exits the notification wait loop
   prematurely and returns 0 from the trapped syscall without the file
   descriptor ever having been installed.

2. If the supervisor retries SECCOMP_IOCTL_NOTIF_ADDFD or
   SECCOMP_IOCTL_NOTIF_SEND before the tracee runs, it fails with
   -EINPROGRESS because knotif->state is no longer SECCOMP_NOTIFY_SENT.

Fix this by restoring knotif->state back to SECCOMP_NOTIFY_SENT if
SECCOMP_ADDFD_FLAG_SEND was set and kaddfd was not consumed before the
interrupted wait. Also add a seccomp_bpf selftest
(user_notification_addfd_send_interrupted) covering this race.

Tested in QEMU against Linux 7.3.0-rc3 by exercising kernel/seccomp.c
and verifying the fix with KASAN enabled.

Fixes: 0ae71c7720e3 ("seccomp: Support atomic "addfd + send reply"")
Cc: stable@vger.kernel.org
Reviewed-by: Bradley Morgan <brads@mainlining.org>
Assisted-by: LLM
Signed-off-by: Hui Peng <benquike@gmail.com>
---
Changes in v4:
- Restore Reviewed-by: Bradley Morgan <brads@mainlining.org> and Cc:
  stable@vger.kernel.org tags from v1 (and restore Bradley Morgan to Cc),
  and add Assisted-by: LLM tag.
- Fix Fixes: tag SHA and subject (0ae71c7720e3 ("seccomp: Support atomic
  "addfd + send reply"")).
- Fix truncated word "interrupted" in the Subject line.

Changes in v3:
- Actually include the tools/testing/selftests/seccomp/seccomp_bpf.c
  regression test in the patch diff (v2 accidentally omitted the selftest
  hunk), with detailed comments explaining the race and test setup.

Changes in v2:
- Add user_notification_addfd_send_interrupted regression test to
  tools/testing/selftests/seccomp/seccomp_bpf.c as requested by Kees Cook.

 kernel/seccomp.c                              |   2 +
 tools/testing/selftests/seccomp/seccomp_bpf.c | 161 ++++++++++++++++++
 2 files changed, 163 insertions(+)

diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 86cf4460d69e..94c7ba80a8f7 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -1808,10 +1808,13 @@ static long seccomp_notify_addfd(struct seccomp_filter *filter,
 	 * We need to check again if the addfd request has been handled,
 	 * and if not, we will remove it from the queue.
 	 */
-	if (list_empty(&kaddfd.list))
+	if (list_empty(&kaddfd.list)) {
 		ret = kaddfd.ret;
-	else
+	} else {
 		list_del(&kaddfd.list);
+		if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND)
+			knotif->state = SECCOMP_NOTIFY_SENT;
+	}
 
 out_unlock:
 	mutex_unlock(&filter->notify_lock);
diff --git a/tools/testing/selftests/seccomp/seccomp_bpf.c b/tools/testing/selftests/seccomp/seccomp_bpf.c
index 0622bc2acad4..e565728f9eb2 100644
--- a/tools/testing/selftests/seccomp/seccomp_bpf.c
+++ b/tools/testing/selftests/seccomp/seccomp_bpf.c
@@ -4368,6 +4368,173 @@ TEST(user_notification_addfd_rlimit)
 	close(memfd);
 }
 
+static void sigusr1_handler(int signo)
+{
+}
+
+/*
+ * Verify that when SECCOMP_IOCTL_NOTIF_ADDFD with SECCOMP_ADDFD_FLAG_SEND is
+ * interrupted by a signal before the tracee dequeues the addfd request,
+ * knotif->state is restored from SECCOMP_NOTIFY_REPLIED back to
+ * SECCOMP_NOTIFY_SENT so that:
+ *   1. The woken tracee sees knotif->state == SECCOMP_NOTIFY_SENT in
+ *      do_user_notif() and goes back to sleep instead of prematurely
+ *      returning 0 from the trapped syscall without the FD installed.
+ *   2. The supervisor can retry SECCOMP_IOCTL_NOTIF_ADDFD (or
+ *      SECCOMP_IOCTL_NOTIF_SEND) instead of failing with -EINPROGRESS.
+ *
+ * To deterministically hit the race window where the supervisor sleeps in
+ * wait_for_completion_interruptible(&kaddfd.completion) after waking the
+ * tracee (complete(&knotif->ready)) but before the tracee runs
+ * seccomp_handle_addfd(), pin all processes to a single CPU and enforce a
+ * strict 3-tier scheduling priority hierarchy on that CPU:
+ *   - Supervisor:            SCHED_FIFO priority 99 (highest)
+ *   - Signal helper (sig_pid): SCHED_FIFO priority 50 (middle)
+ *   - Tracee (pid):          SCHED_IDLE             (lowest)
+ */
+TEST(user_notification_addfd_send_interrupted)
+{
+	/*
+	 * Save parent_pid before user_notif_syscall(__NR_getppid, ...) installs
+	 * the seccomp filter on the calling process; children inherit that
+	 * filter, so sig_pid must not call getppid().
+	 */
+	pid_t pid, sig_pid, parent_pid = getpid();
+	long ret;
+	int status, listener, memfd;
+	struct seccomp_notif_addfd addfd = {};
+	struct seccomp_notif req = {};
+	struct sigaction sa = {};
+	struct sched_param sp_tracee_idle = { .sched_priority = 0 };
+	struct sched_param sp_supervisor_fifo = { .sched_priority = 99 };
+	struct sched_param sp_sig_helper_fifo = { .sched_priority = 50 };
+	struct timespec delay = { .tv_nsec = 15000000 };
+	cpu_set_t cpuset;
+	int cpu;
+
+	/* Pin the supervisor (and its future child processes) to one CPU. */
+	cpu = sched_getcpu();
+	if (cpu >= 0) {
+		CPU_ZERO(&cpuset);
+		CPU_SET(cpu, &cpuset);
+		sched_setaffinity(0, sizeof(cpuset), &cpuset);
+	}
+
+	sa.sa_handler = sigusr1_handler;
+	ASSERT_EQ(sigaction(SIGUSR1, &sa, NULL), 0);
+
+	memfd = memfd_create("test", 0);
+	ASSERT_GE(memfd, 0);
+
+	ret = prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0);
+	ASSERT_EQ(0, ret) {
+		TH_LOG("Kernel does not support PR_SET_NO_NEW_PRIVS!");
+	}
+
+	/*
+	 * Follow the convention of other user_notification_* tests in this
+	 * file by trapping __NR_getppid: because the filter is installed on
+	 * the supervisor before fork(), the trapped syscall must be a
+	 * side-effect-free syscall that the supervisor itself never invokes.
+	 * Even though getppid() does not normally return an FD,
+	 * SECCOMP_ADDFD_FLAG_SEND replaces the trapped syscall's return value
+	 * with the newly installed FD number (42).
+	 */
+	listener = user_notif_syscall(__NR_getppid,
+				      SECCOMP_FILTER_FLAG_NEW_LISTENER);
+	ASSERT_GE(listener, 0);
+
+	pid = fork();
+	ASSERT_GE(pid, 0);
+
+	if (pid == 0) {
+		/*
+		 * Tracee: invoke __NR_getppid as a dummy trigger syscall to
+		 * trap into do_user_notif(). Verify that the syscall returns
+		 * the injected FD number (42) and that FD 42 is open.
+		 */
+		ret = syscall(__NR_getppid);
+		exit(ret != 42 || fcntl(42, F_GETFD) < 0);
+	}
+
+	/* Wait for the tracee to trap in do_user_notif(). */
+	ASSERT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_RECV, &req), 0);
+
+	/*
+	 * Demote the tracee to SCHED_IDLE and promote the supervisor to
+	 * SCHED_FIFO(99) on the same CPU.
+	 */
+	sched_setscheduler(pid, SCHED_IDLE, &sp_tracee_idle);
+	if (sched_setscheduler(0, SCHED_FIFO, &sp_supervisor_fifo) != 0) {
+		close(listener);
+		close(memfd);
+		kill(pid, SIGKILL);
+		waitpid(pid, NULL, 0);
+		SKIP(return, "SCHED_FIFO requires CAP_SYS_NICE");
+	}
+
+	addfd.srcfd = memfd;
+	addfd.newfd_flags = O_CLOEXEC;
+	addfd.newfd = 42;
+	addfd.id = req.id;
+	addfd.flags = SECCOMP_ADDFD_FLAG_SETFD | SECCOMP_ADDFD_FLAG_SEND;
+
+	/*
+	 * Fork a signal helper on the same CPU and set it to SCHED_FIFO(50).
+	 * Because the supervisor is currently running at SCHED_FIFO(99) on
+	 * this CPU, sig_pid is queued on the runqueue but cannot run until the
+	 * supervisor blocks inside the kernel.
+	 *
+	 * When the supervisor invokes ioctl(SECCOMP_IOCTL_NOTIF_ADDFD) below:
+	 *   1. seccomp_notify_addfd() sets knotif->state = SECCOMP_NOTIFY_REPLIED,
+	 *      wakes the tracee (SCHED_IDLE), and blocks in
+	 *      wait_for_completion_interruptible(&kaddfd.completion).
+	 *   2. The CPU scheduler immediately runs sig_pid (SCHED_FIFO 50)
+	 *      ahead of the woken tracee (SCHED_IDLE).
+	 *   3. sig_pid sends SIGUSR1 to parent_pid, waking the supervisor
+	 *      (SCHED_FIFO 99), which immediately preempts sig_pid, aborts the
+	 *      wait with -ERESTARTSYS (-EINTR), removes kaddfd from
+	 *      knotif->addfd, and restores knotif->state = SECCOMP_NOTIFY_SENT
+	 *      before the tracee has executed a single instruction.
+	 */
+	sig_pid = fork();
+	ASSERT_GE(sig_pid, 0);
+	if (sig_pid == 0) {
+		sched_setscheduler(0, SCHED_FIFO, &sp_sig_helper_fifo);
+		kill(parent_pid, SIGUSR1);
+		_exit(0);
+	}
+
+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), -1);
+	EXPECT_EQ(errno, EINTR);
+	EXPECT_EQ(waitpid(sig_pid, &status, 0), sig_pid);
+
+	/*
+	 * Restore normal scheduling and sleep briefly so the woken tracee
+	 * runs in do_user_notif(). With knotif->state restored to
+	 * SECCOMP_NOTIFY_SENT, the tracee must loop back to sleep waiting for
+	 * the notification reply rather than returning 0 from __NR_getppid.
+	 */
+	sched_setscheduler(0, SCHED_OTHER, &sp_tracee_idle);
+	sched_setscheduler(pid, SCHED_OTHER, &sp_tracee_idle);
+	nanosleep(&delay, NULL);
+
+	/*
+	 * Retry SECCOMP_IOCTL_NOTIF_ADDFD. Because knotif->state is
+	 * SECCOMP_NOTIFY_SENT, the retry succeeds (returns 42) instead of
+	 * failing with -EINPROGRESS, installs FD 42 into the tracee, and wakes
+	 * the tracee to complete the syscall with return value 42.
+	 */
+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), 42);
+
+	EXPECT_EQ(waitpid(pid, &status, 0), pid);
+	EXPECT_EQ(true, WIFEXITED(status));
+	EXPECT_EQ(0, WEXITSTATUS(status));
+
+	close(listener);
+	close(memfd);
+}
+
 #ifndef SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP
 #define SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP (1UL << 0)
 #define SECCOMP_IOCTL_NOTIF_SET_FLAGS  SECCOMP_IOW(4, __u64)
-- 
2.49.0

Re: [PATCH v4] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Kees Cook]

On Sun, Sep 20, 2026 at 06:59:10PM +0000, Hui Peng wrote:
> Fix this by restoring knotif->state back to SECCOMP_NOTIFY_SENT if
> SECCOMP_ADDFD_FLAG_SEND was set and kaddfd was not consumed before the
> interrupted wait. Also add a seccomp_bpf selftest
> (user_notification_addfd_send_interrupted) covering this race.

Yeah, this fix is the same restoration logic as seccomp_notify_recv()
uses (although that may actually need "if (state == SECCOMP_NOTIFY_SENT) ..."
added). It does make it clear there is a missing state in the state
machine. SECCOMP_NOTIFY_REPLIES means both "A reply is reserved" and
"A reply has been delivered". But the hidden state is maintained:
"REPLIED + queued ADDFD+SEND" == in flight, and "REPLIED + no queued
addfd" == delivered.

Thank you for the selftest addition, the sched trick is nice!

> Tested in QEMU against Linux 7.3.0-rc3 by exercising kernel/seccomp.c
> and verifying the fix with KASAN enabled.

I've verified this now too. It's kind of a ugly problem because unlucky
timing makes it look like the returned fd is fd 0. :(

> [...]
> +	/*
> +	 * Demote the tracee to SCHED_IDLE and promote the supervisor to
> +	 * SCHED_FIFO(99) on the same CPU.
> +	 */
> +	sched_setscheduler(pid, SCHED_IDLE, &sp_tracee_idle);

Unchecked return value?

> +	if (sched_setscheduler(0, SCHED_FIFO, &sp_supervisor_fifo) != 0) {
> +		close(listener);
> +		close(memfd);
> +		kill(pid, SIGKILL);
> +		waitpid(pid, NULL, 0);
> +		SKIP(return, "SCHED_FIFO requires CAP_SYS_NICE");
> +	}

Instead of this you may want to look at FIXTURE_TEARDOWN to clean up
(though it's not strictly needed since the harness is run in a
subprocess so all these go away on test exit). The one thing that might
be worth doing is making sure sched_setscheduler(0, SCHED_OTHER, ...)
happens ASAP or the test could block everything on a single CPU
machine/VM/CI.

> [...]
> +	sig_pid = fork();
> +	ASSERT_GE(sig_pid, 0);
> +	if (sig_pid == 0) {
> +		sched_setscheduler(0, SCHED_FIFO, &sp_sig_helper_fifo);

Missed return value check here too.

> +		kill(parent_pid, SIGUSR1);
> +		_exit(0);
> +	}
> +
> +	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), -1);
> +	EXPECT_EQ(errno, EINTR);
> +	EXPECT_EQ(waitpid(sig_pid, &status, 0), sig_pid);
> +
> +	/*
> +	 * Restore normal scheduling and sleep briefly so the woken tracee
> +	 * runs in do_user_notif(). With knotif->state restored to
> +	 * SECCOMP_NOTIFY_SENT, the tracee must loop back to sleep waiting for
> +	 * the notification reply rather than returning 0 from __NR_getppid.
> +	 */
> +	sched_setscheduler(0, SCHED_OTHER, &sp_tracee_idle);
> +	sched_setscheduler(pid, SCHED_OTHER, &sp_tracee_idle);

Need to check these too...

> +	nanosleep(&delay, NULL);
> +
> +	/*
> +	 * Retry SECCOMP_IOCTL_NOTIF_ADDFD. Because knotif->state is
> +	 * SECCOMP_NOTIFY_SENT, the retry succeeds (returns 42) instead of
> +	 * failing with -EINPROGRESS, installs FD 42 into the tracee, and wakes
> +	 * the tracee to complete the syscall with return value 42.
> +	 */
> +	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), 42);
> +
> +	EXPECT_EQ(waitpid(pid, &status, 0), pid);
> +	EXPECT_EQ(true, WIFEXITED(status));
> +	EXPECT_EQ(0, WEXITSTATUS(status));
> +
> +	close(listener);
> +	close(memfd);
> +}
> +
>  #ifndef SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP
>  #define SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP (1UL << 0)
>  #define SECCOMP_IOCTL_NOTIF_SET_FLAGS  SECCOMP_IOW(4, __u64)
> -- 
> 2.49.0

Thank you for the test, it really helps see the shape of the issue.

-Kees

-- 
Kees Cook

[PATCH v5] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Hui Peng]

In seccomp_notify_addfd(), when SECCOMP_ADDFD_FLAG_SEND is set in
addfd.flags, knotif->state is optimistically transitioned from
SECCOMP_NOTIFY_SENT to SECCOMP_NOTIFY_REPLIED before waking the tracee
and waiting on kaddfd.completion:

	if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND) {
		knotif->state = SECCOMP_NOTIFY_REPLIED;
		...
	}

If wait_for_completion_interruptible(&kaddfd.completion) is interrupted
by a signal before the tracee dequeues the kaddfd request
(!list_empty(&kaddfd.list)), seccomp_notify_addfd() removes kaddfd from
knotif->addfd and returns -EINTR to the supervisor without having
installed the file descriptor.

However, knotif->state is left as SECCOMP_NOTIFY_REPLIED (with
knotif->error == 0 and knotif->val == 0). This causes two problems:

1. When the tracee runs in do_user_notif(), it checks
   if (knotif.state != SECCOMP_NOTIFY_REPLIED) after calling
   seccomp_handle_addfd(). Because knotif->state is already
   SECCOMP_NOTIFY_REPLIED, the tracee exits the notification wait loop
   prematurely and returns 0 from the trapped syscall without the file
   descriptor ever having been installed.

2. If the supervisor retries SECCOMP_IOCTL_NOTIF_ADDFD or
   SECCOMP_IOCTL_NOTIF_SEND before the tracee runs, it fails with
   -EINPROGRESS because knotif->state is no longer SECCOMP_NOTIFY_SENT.

Fix this by restoring knotif->state back to SECCOMP_NOTIFY_SENT if
SECCOMP_ADDFD_FLAG_SEND was set and kaddfd was not consumed before the
interrupted wait. Also add a seccomp_bpf selftest
(user_notification_addfd_send_interrupted) covering this race.

Tested in QEMU against Linux 7.3.0-rc3 using the
user_notification_addfd_send_interrupted selftest added in this patch
(./seccomp_bpf -t user_notification_addfd_send_interrupted): on the
unfixed kernel the test fails because the retry ioctl(listener,
SECCOMP_IOCTL_NOTIF_ADDFD, &addfd) returns -EINPROGRESS (-1) and the
tracee prematurely exits do_user_notif() with return value 0 instead of
FD 42, whereas with this patch applied the test passes.

Fixes: 0ae71c7720e3 ("seccomp: Support atomic "addfd + send reply"")
Cc: stable@vger.kernel.org
Reviewed-by: Bradley Morgan <brads@mainlining.org>
Assisted-by: LLM
Signed-off-by: Hui Peng <benquike@gmail.com>
---
Changes in v5:
- Check the return values of all sched_setscheduler() calls in
  user_notification_addfd_send_interrupted, as requested by Kees Cook.
- Restore SCHED_OTHER on the supervisor and tracee immediately after
  ioctl(SECCOMP_IOCTL_NOTIF_ADDFD) returns (and on fork() failure) so
  the test cannot remain SCHED_FIFO on a single-CPU VM/CI runner, as
  suggested by Kees Cook.
- Rename pid and struct sched_param variables (tracee_pid, sp_zero,
  sp_fifo_high, sp_fifo_mid) for clarity.

Changes in v4:
- Restore Reviewed-by: Bradley Morgan <brads@mainlining.org> and Cc:
  stable@vger.kernel.org tags from v1 (and restore Bradley Morgan to Cc),
  and add Assisted-by: LLM tag.
- Fix Fixes: tag SHA and subject (0ae71c7720e3 ("seccomp: Support atomic
  "addfd + send reply"")).
- Fix truncated word "interrupted" in the Subject line.

Changes in v3:
- Actually include the tools/testing/selftests/seccomp/seccomp_bpf.c
  regression test in the patch diff (v2 accidentally omitted the selftest
  hunk), with detailed comments explaining the race and test setup.

Changes in v2:
- Add user_notification_addfd_send_interrupted regression test to
  tools/testing/selftests/seccomp/seccomp_bpf.c as requested by Kees Cook.

 kernel/seccomp.c                              |   5 +-
 tools/testing/selftests/seccomp/seccomp_bpf.c | 184 ++++++++++++++++++
 2 files changed, 187 insertions(+), 2 deletions(-)

diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 86cf4460d69e..94c7ba80a8f7 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -1808,10 +1808,13 @@ static long seccomp_notify_addfd(struct seccomp_filter *filter,
 	 * We need to check again if the addfd request has been handled,
 	 * and if not, we will remove it from the queue.
 	 */
-	if (list_empty(&kaddfd.list))
+	if (list_empty(&kaddfd.list)) {
 		ret = kaddfd.ret;
-	else
+	} else {
 		list_del(&kaddfd.list);
+		if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND)
+			knotif->state = SECCOMP_NOTIFY_SENT;
+	}
 
 out_unlock:
 	mutex_unlock(&filter->notify_lock);
diff --git a/tools/testing/selftests/seccomp/seccomp_bpf.c b/tools/testing/selftests/seccomp/seccomp_bpf.c
index 0622bc2acad4..8cafac59ec7b 100644
--- a/tools/testing/selftests/seccomp/seccomp_bpf.c
+++ b/tools/testing/selftests/seccomp/seccomp_bpf.c
@@ -4368,6 +4368,184 @@ TEST(user_notification_addfd_rlimit)
 	close(memfd);
 }
 
+static void sigusr1_handler(int signo)
+{
+}
+
+/*
+ * Verify that when SECCOMP_IOCTL_NOTIF_ADDFD with SECCOMP_ADDFD_FLAG_SEND is
+ * interrupted by a signal before the tracee dequeues the addfd request,
+ * knotif->state is restored from SECCOMP_NOTIFY_REPLIED back to
+ * SECCOMP_NOTIFY_SENT so that:
+ *   1. The woken tracee sees knotif->state == SECCOMP_NOTIFY_SENT in
+ *      do_user_notif() and goes back to sleep instead of prematurely
+ *      returning 0 from the trapped syscall without the FD installed.
+ *   2. The supervisor can retry SECCOMP_IOCTL_NOTIF_ADDFD (or
+ *      SECCOMP_IOCTL_NOTIF_SEND) instead of failing with -EINPROGRESS.
+ *
+ * To deterministically hit the race window where the supervisor sleeps in
+ * wait_for_completion_interruptible(&kaddfd.completion) after waking the
+ * tracee (complete(&knotif->ready)) but before the tracee runs
+ * seccomp_handle_addfd(), pin all processes to a single CPU and enforce a
+ * strict 3-tier scheduling priority hierarchy on that CPU:
+ *   - Supervisor:            SCHED_FIFO priority 99 (highest)
+ *   - Signal helper (sig_pid): SCHED_FIFO priority 50 (middle)
+ *   - Tracee (pid):          SCHED_IDLE             (lowest)
+ */
+TEST(user_notification_addfd_send_interrupted)
+{
+	/*
+	 * Save parent_pid before user_notif_syscall(__NR_getppid, ...) installs
+	 * the seccomp filter on the calling process; children inherit that
+	 * filter, so sig_pid must not call getppid().
+	 */
+	pid_t tracee_pid, sig_pid, parent_pid = getpid();
+	long ret;
+	int status, listener, memfd, err;
+	struct seccomp_notif_addfd addfd = {};
+	struct seccomp_notif req = {};
+	struct sigaction sa = {};
+	struct sched_param sp_zero = { .sched_priority = 0 };
+	struct sched_param sp_fifo_high = { .sched_priority = 99 };
+	struct sched_param sp_fifo_mid = { .sched_priority = 50 };
+	struct timespec delay = { .tv_nsec = 15000000 };
+	cpu_set_t cpuset;
+	int cpu;
+
+	/* Pin the supervisor (and its future child processes) to one CPU. */
+	cpu = sched_getcpu();
+	if (cpu >= 0) {
+		CPU_ZERO(&cpuset);
+		CPU_SET(cpu, &cpuset);
+		sched_setaffinity(0, sizeof(cpuset), &cpuset);
+	}
+
+	sa.sa_handler = sigusr1_handler;
+	ASSERT_EQ(sigaction(SIGUSR1, &sa, NULL), 0);
+
+	memfd = memfd_create("test", 0);
+	ASSERT_GE(memfd, 0);
+
+	ret = prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0);
+	ASSERT_EQ(0, ret) {
+		TH_LOG("Kernel does not support PR_SET_NO_NEW_PRIVS!");
+	}
+
+	/*
+	 * Follow the convention of other user_notification_* tests in this
+	 * file by trapping __NR_getppid: because the filter is installed on
+	 * the supervisor before fork(), the trapped syscall must be a
+	 * side-effect-free syscall that the supervisor itself never invokes.
+	 * Even though getppid() does not normally return an FD,
+	 * SECCOMP_ADDFD_FLAG_SEND replaces the trapped syscall's return value
+	 * with the newly installed FD number (42).
+	 */
+	listener = user_notif_syscall(__NR_getppid,
+				      SECCOMP_FILTER_FLAG_NEW_LISTENER);
+	ASSERT_GE(listener, 0);
+
+	tracee_pid = fork();
+	ASSERT_GE(tracee_pid, 0);
+
+	if (tracee_pid == 0) {
+		/*
+		 * Tracee: invoke __NR_getppid as a dummy trigger syscall to
+		 * trap into do_user_notif(). Verify that the syscall returns
+		 * the injected FD number (42) and that FD 42 is open.
+		 */
+		ret = syscall(__NR_getppid);
+		exit(ret != 42 || fcntl(42, F_GETFD) < 0);
+	}
+
+	/* Wait for the tracee to trap in do_user_notif(). */
+	ASSERT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_RECV, &req), 0);
+
+	/*
+	 * Demote the tracee to SCHED_IDLE and promote the supervisor to
+	 * SCHED_FIFO(99) on the same CPU.
+	 */
+	ASSERT_EQ(sched_setscheduler(tracee_pid, SCHED_IDLE, &sp_zero), 0);
+	if (sched_setscheduler(0, SCHED_FIFO, &sp_fifo_high) != 0) {
+		kill(tracee_pid, SIGKILL);
+		waitpid(tracee_pid, NULL, 0);
+		SKIP(return, "SCHED_FIFO requires CAP_SYS_NICE");
+	}
+
+	addfd.srcfd = memfd;
+	addfd.newfd_flags = O_CLOEXEC;
+	addfd.newfd = 42;
+	addfd.id = req.id;
+	addfd.flags = SECCOMP_ADDFD_FLAG_SETFD | SECCOMP_ADDFD_FLAG_SEND;
+
+	/*
+	 * Fork a signal helper on the same CPU and set it to SCHED_FIFO(50).
+	 * Because the supervisor is currently running at SCHED_FIFO(99) on
+	 * this CPU, sig_pid is queued on the runqueue but cannot run until the
+	 * supervisor blocks inside the kernel.
+	 *
+	 * When the supervisor invokes ioctl(SECCOMP_IOCTL_NOTIF_ADDFD) below:
+	 *   1. seccomp_notify_addfd() sets knotif->state = SECCOMP_NOTIFY_REPLIED,
+	 *      wakes the tracee (SCHED_IDLE), and blocks in
+	 *      wait_for_completion_interruptible(&kaddfd.completion).
+	 *   2. The CPU scheduler immediately runs sig_pid (SCHED_FIFO 50)
+	 *      ahead of the woken tracee (SCHED_IDLE).
+	 *   3. sig_pid sends SIGUSR1 to parent_pid, waking the supervisor
+	 *      (SCHED_FIFO 99), which immediately preempts sig_pid, aborts the
+	 *      wait with -ERESTARTSYS (-EINTR), removes kaddfd from
+	 *      knotif->addfd, and restores knotif->state = SECCOMP_NOTIFY_SENT
+	 *      before the tracee has executed a single instruction.
+	 */
+	sig_pid = fork();
+	if (sig_pid < 0) {
+		sched_setscheduler(0, SCHED_OTHER, &sp_zero);
+		kill(tracee_pid, SIGKILL);
+		waitpid(tracee_pid, NULL, 0);
+	}
+	ASSERT_GE(sig_pid, 0);
+	if (sig_pid == 0) {
+		if (sched_setscheduler(0, SCHED_FIFO, &sp_fifo_mid) != 0)
+			_exit(1);
+		kill(parent_pid, SIGUSR1);
+		_exit(0);
+	}
+
+	ret = ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd);
+	err = errno;
+
+	/*
+	 * Restore normal scheduling ASAP so the supervisor does not remain
+	 * SCHED_FIFO on a single-CPU machine/VM/CI, then sleep briefly so the
+	 * woken tracee runs in do_user_notif(). With knotif->state restored to
+	 * SECCOMP_NOTIFY_SENT, the tracee must loop back to sleep waiting for
+	 * the notification reply rather than returning 0 from __NR_getppid.
+	 */
+	ASSERT_EQ(sched_setscheduler(0, SCHED_OTHER, &sp_zero), 0);
+	ASSERT_EQ(sched_setscheduler(tracee_pid, SCHED_OTHER, &sp_zero), 0);
+
+	EXPECT_EQ(ret, -1);
+	EXPECT_EQ(err, EINTR);
+	EXPECT_EQ(waitpid(sig_pid, &status, 0), sig_pid);
+	EXPECT_EQ(true, WIFEXITED(status));
+	EXPECT_EQ(0, WEXITSTATUS(status));
+
+	nanosleep(&delay, NULL);
+
+	/*
+	 * Retry SECCOMP_IOCTL_NOTIF_ADDFD. Because knotif->state is
+	 * SECCOMP_NOTIFY_SENT, the retry succeeds (returns 42) instead of
+	 * failing with -EINPROGRESS, installs FD 42 into the tracee, and wakes
+	 * the tracee to complete the syscall with return value 42.
+	 */
+	EXPECT_EQ(ioctl(listener, SECCOMP_IOCTL_NOTIF_ADDFD, &addfd), 42);
+
+	EXPECT_EQ(waitpid(tracee_pid, &status, 0), tracee_pid);
+	EXPECT_EQ(true, WIFEXITED(status));
+	EXPECT_EQ(0, WEXITSTATUS(status));
+
+	close(listener);
+	close(memfd);
+}
+
 #ifndef SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP
 #define SECCOMP_USER_NOTIF_FD_SYNC_WAKE_UP (1UL << 0)
 #define SECCOMP_IOCTL_NOTIF_SET_FLAGS  SECCOMP_IOW(4, __u64)
-- 
2.49.0

Re: [PATCH v5] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted

[Kees Cook]

On Tue, 22 Sep 2026 02:32:04 +0000, Hui Peng wrote:
> In seccomp_notify_addfd(), when SECCOMP_ADDFD_FLAG_SEND is set in
> addfd.flags, knotif->state is optimistically transitioned from
> SECCOMP_NOTIFY_SENT to SECCOMP_NOTIFY_REPLIED before waking the tracee
> and waiting on kaddfd.completion:
> 
> 	if (addfd.flags & SECCOMP_ADDFD_FLAG_SEND) {
> 		knotif->state = SECCOMP_NOTIFY_REPLIED;
> 		...
> 	}
> 
> [...]

I renamed the empty signal handler and added a comment, but otherwise,
applied to for-next/seccomp, thanks!

[1/1] seccomp: restore knotif->state when SECCOMP_ADDFD_FLAG_SEND is interrupted
      https://git.kernel.org/kees/c/832b9b176be0

Take care,

-- 
Kees Cook