[LTP] [PATCH v2][STAGING] syscalls/fanotify26: Add test for nested permission events

[LTP] [PATCH v2][STAGING] syscalls/fanotify26: Add test for nested permission events

[Edvard Conrad via ltp]

The kernel creates the file descriptor it reports with a permission
event by an open that must not generate further events. When the
reported path is on a stacked filesystem, that open opens the real file
on the backing filesystem, and since commit ebe559609d78 ("fs: get rid
of __FMODE_NONOTIFY kludge") the backing file no longer inherits
FMODE_NONOTIFY. A listener that also watches the backing filesystem is
then sent a second permission event from inside its own read(2), which
only the listener itself can answer, so a single reader never returns.

Verify that a listener returns from read(2) and answers an open
permission event for a file on an overlayfs mount. The first test case
marks the overlay only and passes on any kernel, the second one also
marks the backing filesystem, which is the shape an anti-malware
scanner uses on an overlayfs root.

The second case fails on a kernel without the pending fix "fs: let
backing files inherit FMODE_NONOTIFY from the user file", so the test
is added to runtest/staging for now.

Link: https://lore.kernel.org/all/P4wV1J9--F-9@tuta.io/
Signed-off-by: Edvard Conrad <edvard@tuta.io>
---
v2:
- Add the copyright header block.
- Require FAN_MARK_FILESYSTEM support on the overlay and on the base
  mount in setup(). Both cases mark the overlay, so gating the check on
  the second case made the first one break instead of skip where the
  mark type is unsupported.
- Report TPASS from the child that answers the event. Inferring it from
  the child reaching 'Z' took any abnormal exit for a success.
- Use "!= -1" for file descriptor checks, and drop the redundant
  SAFE_CLOSE() before tst_brk() and the redundant fd_notify = -1.

Thanks to the LTP AI reviewer for the review.

 runtest/staging                               |   1 +
 testcases/kernel/syscalls/fanotify/.gitignore |   1 +
 .../kernel/syscalls/fanotify/fanotify26.c     | 185 ++++++++++++++++++
 3 files changed, 187 insertions(+)
 create mode 100644 testcases/kernel/syscalls/fanotify/fanotify26.c

diff --git a/runtest/staging b/runtest/staging
index ef1cdea15..2ce76b119 100644
--- a/runtest/staging
+++ b/runtest/staging
@@ -1 +1,2 @@
 # Tests for features that are not yet in the stable kernel ABI
+fanotify26 fanotify26
diff --git a/testcases/kernel/syscalls/fanotify/.gitignore b/testcases/kernel/syscalls/fanotify/.gitignore
index d6d0599f1..c13d9b33c 100644
--- a/testcases/kernel/syscalls/fanotify/.gitignore
+++ b/testcases/kernel/syscalls/fanotify/.gitignore
@@ -23,4 +23,5 @@
 /fanotify23
 /fanotify24
 /fanotify25
+/fanotify26
 /fanotify_child
diff --git a/testcases/kernel/syscalls/fanotify/fanotify26.c b/testcases/kernel/syscalls/fanotify/fanotify26.c
new file mode 100644
index 000000000..4d7d6f1f5
--- /dev/null
+++ b/testcases/kernel/syscalls/fanotify/fanotify26.c
@@ -0,0 +1,185 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2026 Edvard Conrad <edvard@tuta.io>
+ */
+
+/*\
+ * Check that a permission event on a stacked filesystem does not deadlock the
+ * :manpage:`fanotify(7)` listener.
+ *
+ * When the kernel reports a permission event, it creates the event file
+ * descriptor with an open that must not generate further events. If the
+ * reported path is on a stacked filesystem, that open opens the real file on
+ * the backing filesystem. Should the backing file not inherit "generate no
+ * events", and the listener watch the backing filesystem as well, the kernel
+ * queues a second permission event to the same group while the listener is
+ * still inside :manpage:`read(2)`. Only the listener can answer it, so a
+ * listener that reads with a single thread never returns, and every task that
+ * touches a watched filesystem queues up behind it.
+ *
+ * The first test case marks the overlay only and must pass on any kernel. The
+ * second one marks the backing filesystem as well, which is the shape an
+ * anti-malware scanner uses on an overlayfs root.
+ *
+ * The test needs root to create permission events and to mount overlayfs.
+ */
+
+#define _GNU_SOURCE
+#include "config.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <errno.h>
+#include <string.h>
+#include <signal.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include "tst_test.h"
+
+#ifdef HAVE_SYS_FANOTIFY_H
+#include "fanotify.h"
+
+#define BACKING_MNT	OVL_BASE_MNTPOINT
+#define FNAME		"test_file"
+#define OVL_FILE	OVL_MNT "/" FNAME
+#define LOWER_FILE	OVL_LOWER "/" FNAME
+
+/*
+ * The listener either answers the event at once or is blocked for good, so
+ * this only has to outlast scheduling noise on a loaded machine.
+ */
+#define WAIT_MSEC	5000
+
+static int fd_notify = -1;
+
+static struct tcase {
+	const char *tname;
+	int mark_backing;
+} tcases[] = {
+	{"Mark on the overlay", 0},
+	{"Mark on the overlay and on the backing filesystem", 1},
+};
+
+static void answer_one_event(void)
+{
+	struct fanotify_event_metadata event;
+	struct fanotify_response resp;
+
+	/* Blocks for good if the event fd generates a nested event. */
+	SAFE_READ(1, fd_notify, &event, sizeof(event));
+
+	if (event.mask != FAN_OPEN_PERM) {
+		tst_brk(TBROK, "Got unexpected event mask %llx",
+			(unsigned long long)event.mask);
+	}
+
+	resp.fd = event.fd;
+	resp.response = FAN_ALLOW;
+	SAFE_WRITE(SAFE_WRITE_ALL, fd_notify, &resp, sizeof(resp));
+
+	if (event.fd != FAN_NOFD)
+		SAFE_CLOSE(event.fd);
+
+	tst_res(TPASS, "Listener answered the event");
+}
+
+static void run(unsigned int n)
+{
+	struct tcase *tc = &tcases[n];
+	pid_t opener, reader;
+
+	tst_res(TINFO, "Test #%d: %s", n, tc->tname);
+
+	fd_notify = SAFE_FANOTIFY_INIT(FAN_CLASS_CONTENT, O_RDONLY);
+
+	SAFE_FANOTIFY_MARK(fd_notify, FAN_MARK_ADD | FAN_MARK_FILESYSTEM,
+			   FAN_OPEN_PERM, AT_FDCWD, OVL_MNT);
+
+	if (tc->mark_backing) {
+		SAFE_FANOTIFY_MARK(fd_notify,
+				   FAN_MARK_ADD | FAN_MARK_FILESYSTEM,
+				   FAN_OPEN_PERM, AT_FDCWD, BACKING_MNT);
+	}
+
+	opener = SAFE_FORK();
+	if (!opener) {
+		int fd;
+
+		/* Must not hold the group, it has to be released on close. */
+		SAFE_CLOSE(fd_notify);
+		/*
+		 * The kernel denies the event when a listener that never
+		 * answered it goes away, so a failure here is the fallout of
+		 * the test tearing a blocked listener down, not a result.
+		 */
+		fd = open(OVL_FILE, O_RDONLY);
+		if (fd != -1)
+			SAFE_CLOSE(fd);
+		exit(0);
+	}
+
+	/* The opener sleeps in the kernel once its event is queued. */
+	if (TST_PROCESS_STATE_WAIT(opener, 'D', WAIT_MSEC))
+		tst_brk(TBROK, "Open generated no permission event");
+
+	reader = SAFE_FORK();
+	if (!reader) {
+		answer_one_event();
+		exit(0);
+	}
+
+	/*
+	 * The listener answers and exits, or is blocked for good. Only the
+	 * blocked case is handled here: the child reports its own success,
+	 * so a child that died some other way cannot be taken for one that
+	 * answered. The library reaps it on the normal path.
+	 */
+	if (TST_PROCESS_STATE_WAIT(reader, 'Z', WAIT_MSEC)) {
+		tst_res(TFAIL,
+			"Listener blocked in read() by a nested event");
+		SAFE_KILL(reader, SIGKILL);
+		SAFE_WAITPID(reader, NULL, 0);
+	}
+
+	/* Releases whatever is still queued, so the opener can finish. */
+	SAFE_CLOSE(fd_notify);
+
+	SAFE_WAITPID(opener, NULL, 0);
+}
+
+static void setup(void)
+{
+	/* The file has to exist in the lower layer to be opened through it. */
+	SAFE_UMOUNT(OVL_MNT);
+	SAFE_TOUCH(LOWER_FILE, 0644, NULL);
+	SAFE_MOUNT_OVERLAY();
+
+	require_fanotify_access_permissions_supported_on_fs(OVL_FILE);
+
+	/* Every case marks the overlay; the second also marks the base. */
+	REQUIRE_MARK_TYPE_SUPPORTED_ON_FS(FAN_MARK_FILESYSTEM, OVL_MNT);
+	REQUIRE_MARK_TYPE_SUPPORTED_ON_FS(FAN_MARK_FILESYSTEM, BACKING_MNT);
+}
+
+static void cleanup(void)
+{
+	if (fd_notify != -1)
+		SAFE_CLOSE(fd_notify);
+}
+
+static struct tst_test test = {
+	.test = run,
+	.tcnt = ARRAY_SIZE(tcases),
+	.setup = setup,
+	.cleanup = cleanup,
+	.forks_child = 1,
+	.needs_root = 1,
+	.mount_device = 1,
+	.needs_overlay = 1,
+	.mntpoint = OVL_BASE_MNTPOINT,
+};
+
+#else
+	TST_TEST_TCONF("system doesn't have required fanotify support");
+#endif
-- 
2.56.0


-- 
Mailing list info: https://lists.linux.it/listinfo/ltp

Re: [LTP] syscalls/fanotify26: Add test for nested permission events

[linuxtestproject.agent]

> Subject: [PATCH] syscalls/fanotify26: Add test for nested permission events
>
> A permission event reported on a stacked filesystem opens the real file
> on the backing filesystem to install the event file descriptor. If the
...
> The second case fails on a kernel without the pending fix "fs: let
> backing files inherit FMODE_NONOTIFY from the user file", so the test is
> added to runtest/staging for now.

Since the test is added to runtest/staging for an unreleased kernel fix,
the commit subject must include the [STAGING] prefix:

    [STAGING] syscalls/fanotify26: Add test for nested permission events

> +static void answer_one_event(void)
> +{
> +	struct fanotify_event_metadata event;
> +	struct fanotify_response resp;
> +
> +	/* Blocks for good if the event fd generates a nested event. */
> +	SAFE_READ(1, fd_notify, &event, sizeof(event));

Please use SAFE_READ_ALL instead of the literal 1, matching
SAFE_WRITE_ALL below:

	SAFE_READ(SAFE_READ_ALL, fd_notify, &event, sizeof(event));

> +	/* Releases whatever is still queued, so the opener can finish. */
> +	SAFE_CLOSE(fd_notify);
> +
> +	SAFE_WAITPID(opener, NULL, 0);
> +}

This SAFE_WAITPID() is redundant and discards the child's exit status.
The test library calls tst_reap_children() automatically after each
run() iteration, which waits for remaining children and verifies that
they exited cleanly.

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp

Re: [LTP] [PATCH v2][STAGING] syscalls/fanotify26: Add test for nested permission events

[Amir Goldstein]

On Wed, Oct 28, 2026 at 8:48 AM Edvard Conrad <edvard@tuta.io> wrote:
>
> The kernel creates the file descriptor it reports with a permission
> event by an open that must not generate further events. When the
> reported path is on a stacked filesystem, that open opens the real file
> on the backing filesystem, and since commit ebe559609d78 ("fs: get rid
> of __FMODE_NONOTIFY kludge") the backing file no longer inherits
> FMODE_NONOTIFY. A listener that also watches the backing filesystem is
> then sent a second permission event from inside its own read(2), which
> only the listener itself can answer, so a single reader never returns.
>
> Verify that a listener returns from read(2) and answers an open
> permission event for a file on an overlayfs mount. The first test case
> marks the overlay only and passes on any kernel, the second one also
> marks the backing filesystem, which is the shape an anti-malware
> scanner uses on an overlayfs root.
>
> The second case fails on a kernel without the pending fix "fs: let
> backing files inherit FMODE_NONOTIFY from the user file", so the test
> is added to runtest/staging for now.
>
> Link: https://lore.kernel.org/all/P4wV1J9--F-9@tuta.io/
> Signed-off-by: Edvard Conrad <edvard@tuta.io>
> ---

Very nice!

Thanks

Reviewed-by: Amir Goldstein <amir73il@gmail.com>

> v2:
> - Add the copyright header block.
> - Require FAN_MARK_FILESYSTEM support on the overlay and on the base
>   mount in setup(). Both cases mark the overlay, so gating the check on
>   the second case made the first one break instead of skip where the
>   mark type is unsupported.
> - Report TPASS from the child that answers the event. Inferring it from
>   the child reaching 'Z' took any abnormal exit for a success.
> - Use "!= -1" for file descriptor checks, and drop the redundant
>   SAFE_CLOSE() before tst_brk() and the redundant fd_notify = -1.
>
> Thanks to the LTP AI reviewer for the review.
>
>  runtest/staging                               |   1 +
>  testcases/kernel/syscalls/fanotify/.gitignore |   1 +
>  .../kernel/syscalls/fanotify/fanotify26.c     | 185 ++++++++++++++++++
>  3 files changed, 187 insertions(+)
>  create mode 100644 testcases/kernel/syscalls/fanotify/fanotify26.c
>
> diff --git a/runtest/staging b/runtest/staging
> index ef1cdea15..2ce76b119 100644
> --- a/runtest/staging
> +++ b/runtest/staging
> @@ -1 +1,2 @@
>  # Tests for features that are not yet in the stable kernel ABI
> +fanotify26 fanotify26
> diff --git a/testcases/kernel/syscalls/fanotify/.gitignore b/testcases/kernel/syscalls/fanotify/.gitignore
> index d6d0599f1..c13d9b33c 100644
> --- a/testcases/kernel/syscalls/fanotify/.gitignore
> +++ b/testcases/kernel/syscalls/fanotify/.gitignore
> @@ -23,4 +23,5 @@
>  /fanotify23
>  /fanotify24
>  /fanotify25
> +/fanotify26
>  /fanotify_child
> diff --git a/testcases/kernel/syscalls/fanotify/fanotify26.c b/testcases/kernel/syscalls/fanotify/fanotify26.c
> new file mode 100644
> index 000000000..4d7d6f1f5
> --- /dev/null
> +++ b/testcases/kernel/syscalls/fanotify/fanotify26.c
> @@ -0,0 +1,185 @@
> +// SPDX-License-Identifier: GPL-2.0-or-later
> +/*
> + * Copyright (c) 2026 Edvard Conrad <edvard@tuta.io>
> + */
> +
> +/*\
> + * Check that a permission event on a stacked filesystem does not deadlock the
> + * :manpage:`fanotify(7)` listener.
> + *
> + * When the kernel reports a permission event, it creates the event file
> + * descriptor with an open that must not generate further events. If the
> + * reported path is on a stacked filesystem, that open opens the real file on
> + * the backing filesystem. Should the backing file not inherit "generate no
> + * events", and the listener watch the backing filesystem as well, the kernel
> + * queues a second permission event to the same group while the listener is
> + * still inside :manpage:`read(2)`. Only the listener can answer it, so a
> + * listener that reads with a single thread never returns, and every task that
> + * touches a watched filesystem queues up behind it.
> + *
> + * The first test case marks the overlay only and must pass on any kernel. The
> + * second one marks the backing filesystem as well, which is the shape an
> + * anti-malware scanner uses on an overlayfs root.
> + *
> + * The test needs root to create permission events and to mount overlayfs.
> + */
> +
> +#define _GNU_SOURCE
> +#include "config.h"
> +
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <unistd.h>
> +#include <errno.h>
> +#include <string.h>
> +#include <signal.h>
> +#include <sys/types.h>
> +#include <sys/wait.h>
> +#include "tst_test.h"
> +
> +#ifdef HAVE_SYS_FANOTIFY_H
> +#include "fanotify.h"
> +
> +#define BACKING_MNT    OVL_BASE_MNTPOINT
> +#define FNAME          "test_file"
> +#define OVL_FILE       OVL_MNT "/" FNAME
> +#define LOWER_FILE     OVL_LOWER "/" FNAME
> +
> +/*
> + * The listener either answers the event at once or is blocked for good, so
> + * this only has to outlast scheduling noise on a loaded machine.
> + */
> +#define WAIT_MSEC      5000
> +
> +static int fd_notify = -1;
> +
> +static struct tcase {
> +       const char *tname;
> +       int mark_backing;
> +} tcases[] = {
> +       {"Mark on the overlay", 0},
> +       {"Mark on the overlay and on the backing filesystem", 1},
> +};
> +
> +static void answer_one_event(void)
> +{
> +       struct fanotify_event_metadata event;
> +       struct fanotify_response resp;
> +
> +       /* Blocks for good if the event fd generates a nested event. */
> +       SAFE_READ(1, fd_notify, &event, sizeof(event));
> +
> +       if (event.mask != FAN_OPEN_PERM) {
> +               tst_brk(TBROK, "Got unexpected event mask %llx",
> +                       (unsigned long long)event.mask);
> +       }
> +
> +       resp.fd = event.fd;
> +       resp.response = FAN_ALLOW;
> +       SAFE_WRITE(SAFE_WRITE_ALL, fd_notify, &resp, sizeof(resp));
> +
> +       if (event.fd != FAN_NOFD)
> +               SAFE_CLOSE(event.fd);
> +
> +       tst_res(TPASS, "Listener answered the event");
> +}
> +
> +static void run(unsigned int n)
> +{
> +       struct tcase *tc = &tcases[n];
> +       pid_t opener, reader;
> +
> +       tst_res(TINFO, "Test #%d: %s", n, tc->tname);
> +
> +       fd_notify = SAFE_FANOTIFY_INIT(FAN_CLASS_CONTENT, O_RDONLY);
> +
> +       SAFE_FANOTIFY_MARK(fd_notify, FAN_MARK_ADD | FAN_MARK_FILESYSTEM,
> +                          FAN_OPEN_PERM, AT_FDCWD, OVL_MNT);
> +
> +       if (tc->mark_backing) {
> +               SAFE_FANOTIFY_MARK(fd_notify,
> +                                  FAN_MARK_ADD | FAN_MARK_FILESYSTEM,
> +                                  FAN_OPEN_PERM, AT_FDCWD, BACKING_MNT);
> +       }
> +
> +       opener = SAFE_FORK();
> +       if (!opener) {
> +               int fd;
> +
> +               /* Must not hold the group, it has to be released on close. */
> +               SAFE_CLOSE(fd_notify);
> +               /*
> +                * The kernel denies the event when a listener that never
> +                * answered it goes away, so a failure here is the fallout of
> +                * the test tearing a blocked listener down, not a result.
> +                */
> +               fd = open(OVL_FILE, O_RDONLY);
> +               if (fd != -1)
> +                       SAFE_CLOSE(fd);
> +               exit(0);
> +       }
> +
> +       /* The opener sleeps in the kernel once its event is queued. */
> +       if (TST_PROCESS_STATE_WAIT(opener, 'D', WAIT_MSEC))
> +               tst_brk(TBROK, "Open generated no permission event");
> +
> +       reader = SAFE_FORK();
> +       if (!reader) {
> +               answer_one_event();
> +               exit(0);
> +       }
> +
> +       /*
> +        * The listener answers and exits, or is blocked for good. Only the
> +        * blocked case is handled here: the child reports its own success,
> +        * so a child that died some other way cannot be taken for one that
> +        * answered. The library reaps it on the normal path.
> +        */
> +       if (TST_PROCESS_STATE_WAIT(reader, 'Z', WAIT_MSEC)) {
> +               tst_res(TFAIL,
> +                       "Listener blocked in read() by a nested event");
> +               SAFE_KILL(reader, SIGKILL);
> +               SAFE_WAITPID(reader, NULL, 0);
> +       }
> +
> +       /* Releases whatever is still queued, so the opener can finish. */
> +       SAFE_CLOSE(fd_notify);
> +
> +       SAFE_WAITPID(opener, NULL, 0);
> +}
> +
> +static void setup(void)
> +{
> +       /* The file has to exist in the lower layer to be opened through it. */
> +       SAFE_UMOUNT(OVL_MNT);
> +       SAFE_TOUCH(LOWER_FILE, 0644, NULL);
> +       SAFE_MOUNT_OVERLAY();
> +
> +       require_fanotify_access_permissions_supported_on_fs(OVL_FILE);
> +
> +       /* Every case marks the overlay; the second also marks the base. */
> +       REQUIRE_MARK_TYPE_SUPPORTED_ON_FS(FAN_MARK_FILESYSTEM, OVL_MNT);
> +       REQUIRE_MARK_TYPE_SUPPORTED_ON_FS(FAN_MARK_FILESYSTEM, BACKING_MNT);
> +}
> +
> +static void cleanup(void)
> +{
> +       if (fd_notify != -1)
> +               SAFE_CLOSE(fd_notify);
> +}
> +
> +static struct tst_test test = {
> +       .test = run,
> +       .tcnt = ARRAY_SIZE(tcases),
> +       .setup = setup,
> +       .cleanup = cleanup,
> +       .forks_child = 1,
> +       .needs_root = 1,
> +       .mount_device = 1,
> +       .needs_overlay = 1,
> +       .mntpoint = OVL_BASE_MNTPOINT,
> +};
> +
> +#else
> +       TST_TEST_TCONF("system doesn't have required fanotify support");
> +#endif
> --
> 2.56.0
>

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp

Re: [LTP] syscalls/fanotify26: Add test for nested permission events

[Edvard Conrad via ltp]

Hi,

On Wed, Oct 28, 2026 at 09:15:42AM +0000, linuxtestproject.agent wrote:
> Since the test is added to runtest/staging for an unreleased kernel fix,
> the commit subject must include the [STAGING] prefix:
>
>     [STAGING] syscalls/fanotify26: Add test for nested permission events

ground_rules.rst asks for the keyword in the patch subject, and gives

  Subject: [PATCH v1][STAGING] fanotify: add test for <feature> (requires v6.19-rc3)

as the example. The mail went out as "[PATCH v2][STAGING] syscalls/fanotify26:
..." and the archive shows it arrived that way. fanotify23, the last test added
to runtest/staging in a8cd03037, likewise carries no [STAGING] in its commit
subject, so I have left the commit subject alone.

> Please use SAFE_READ_ALL instead of the literal 1, matching
> SAFE_WRITE_ALL below:

Done in v3.

> This SAFE_WAITPID() is redundant and discards the child's exit status.
> The test library calls tst_reap_children() automatically after each
> run() iteration [...]

Right, and dropped in v3 for the opener.

Worth noting it does not generalise to the other child: the reader is killed
with SIGKILL on the failure path, and check_child_status() reports TBROK for a
signal-killed child, so that one still has to be reaped by the test itself.

Thanks for the review.

Edvard

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp