Skip to contentSkip to conversation

inotify13: check rename cookie pairing after adjacent moves

PATCH In review Hungyu Lin 4 participants 7 messages Oct 29 – Nov 4 VFS

lore
Share & tools
lore

Hungyu Lin

author
#0 lore raw

PATCH v4 1/1 inotify13: check rename cookie pairing after adjacent moves

Queue an unpaired IN_MOVED_FROM event, followed by an IN_MOVED_FROM and IN_MOVED_TO pair with the same source watch descriptor and name as the first event. Watch only IN_MOVE so creation of the replacement entry does not separate the two IN_MOVED_FROM events in the queue.

Affected kernels coalesce the second IN_MOVED_FROM into the first, leaving the IN_MOVED_TO cookie without a matching IN_MOVED_FROM cookie. Check that all three events are retained, the two moves have distinct cookies, and the second IN_MOVED_FROM matches the IN_MOVED_TO cookie. Cover both regular files and directories.

---
Changes in v4:
- Put expected values immediately after the mismatched fields, following the existing per-field diagnostic format.
- Use the existing "got unnecessary event" and "didn't get event" wording.
- Keep event context for extra events and mask, watch descriptor and filename mismatches. Omit it from the two cookie-failure messages, since it is already present in TINFO, as confirmed by Amir.

Built with GCC 13.3.0 and checked with LTP checkpatch and sparse. On ARM64 QEMU, the test reports 0 pass / 2 fail on the baseline kernel and 2 pass / 0 fail with the kernel fix. Local diagnostic controls also exercise both cookie failures and verify the preceding TINFO and concise TFAIL output. The controls are kept outside this patch.

 runtest/syscalls                              |   1 +
 testcases/kernel/syscalls/inotify/.gitignore  |   1 +
 testcases/kernel/syscalls/inotify/inotify13.c | 211 ++++++++++++++++++++++++++
 3 files changed, 213 insertions(+)
 create mode 100644 testcases/kernel/syscalls/inotify/inotify13.c
runtest/syscalls +1 -0
index 2bdb4a9a2b4801d03d93efea686124530f5ce2d5..e578bd92d7e7ded5c0ca2f95dc87fce814d733cc 100644
@@ -677,6 +677,7 @@ inotify09 inotify09
 inotify10 inotify10
 inotify11 inotify11
 inotify12 inotify12
+inotify13 inotify13
 
 fanotify01 fanotify01
 fanotify02 fanotify02
testcases/kernel/syscalls/inotify/.gitignore +1 -0
index f6e5c546ac4032766cee9f16b3689be5d0302d37..b597ea63f493b2da2a58017d3eb7b16a7b033fc5 100644
@@ -10,3 +10,4 @@
 /inotify10
 /inotify11
 /inotify12
+/inotify13
testcases/kernel/syscalls/inotify/inotify13.c new file +211 -0

211 lines not on this page · raw message

Signature · 1 line
2.43.0
#1 lore raw

Hi Hungyu,

Quoted 7 lines · Hungyu Lin

inotify13: check rename cookie pairing after adjacent moves

+/*\
+ * Verify that two IN_MOVED_FROM events that differ only in their rename
+ * cookie are not coalesced.
+ *
+ * inotify(7) allows coalescing only of events with identical wd, mask,

Use :manpage:inotify(7) in the RST doc comment so that the test catalog renders a link to the man page.

testcases/kernel/syscalls/inotify/inotify13.c line 199

Quoted 5 lines
testcases/kernel/syscalls/inotify/inotify13.c new file +5 -0
+static void cleanup(void)
+{
+	if (fd_notify >= 0)
+		SAFE_CLOSE(fd_notify);
+}

Use "if (fd_notify != -1)" to check file descriptor validity in cleanup(), matching the -1 initialization convention.

testcases/kernel/syscalls/inotify/inotify13.c line 207

Quoted 7 lines
testcases/kernel/syscalls/inotify/inotify13.c new file +7 -0
+static struct tst_test test = {
+	.test = verify_inotify,
+	.tcnt = ARRAY_SIZE(tcases),
+	.setup = setup,
+	.cleanup = cleanup,
+	.needs_tmpdir = 1,
+};

Since this is a kernel regression test, add .tags with the "linux-git" fix commit hash in struct tst_test once the kernel fix is merged.

Verdict - Needs revision

---
Note:

The agent can sometimes produce false positives although often its findings are genuine. If you find issues with the review, please comment this email or ignore the suggestions.

Regards,
LTP AI Reviewer

Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Reviewed-by on v4 1/1
#2 lore raw
In reply to Hungyu Lin
Quoted 6 lines · Hungyu Lin

Queue an unpaired IN_MOVED_FROM event, followed by an IN_MOVED_FROM and IN_MOVED_TO pair with the same source watch descriptor and name as the first event. Watch only IN_MOVE so creation of the replacement entry does not separate the two IN_MOVED_FROM events in the queue.

Quoted 8 lines

Affected kernels coalesce the second IN_MOVED_FROM into the first, leaving the IN_MOVED_TO cookie without a matching IN_MOVED_FROM cookie. Check that all three events are retained, the two moves have distinct cookies, and the second IN_MOVED_FROM matches the IN_MOVED_TO cookie. Cover both regular files and directories.

Thank you for your patience with my review

Quoted 261 lines

---
Changes in v4:
- Put expected values immediately after the mismatched fields, following the existing per-field diagnostic format.
- Use the existing "got unnecessary event" and "didn't get event" wording.
- Keep event context for extra events and mask, watch descriptor and filename mismatches. Omit it from the two cookie-failure messages, since it is already present in TINFO, as confirmed by Amir.

Built with GCC 13.3.0 and checked with LTP checkpatch and sparse. On ARM64 QEMU, the test reports 0 pass / 2 fail on the baseline kernel and 2 pass / 0 fail with the kernel fix. Local diagnostic controls also exercise both cookie failures and verify the preceding TINFO and concise TFAIL output. The controls are kept outside this patch.

 runtest/syscalls                              |   1 +
 testcases/kernel/syscalls/inotify/.gitignore  |   1 +
 testcases/kernel/syscalls/inotify/inotify13.c | 211 ++++++++++++++++++++++++++
 3 files changed, 213 insertions(+)
 create mode 100644 testcases/kernel/syscalls/inotify/inotify13.c
runtest/syscalls +1 -0
index 2bdb4a9a2b4801d03d93efea686124530f5ce2d5..e578bd92d7e7ded5c0ca2f95dc87fce814d733cc 100644
@@ -677,6 +677,7 @@ inotify09 inotify09
 inotify10 inotify10
 inotify11 inotify11
 inotify12 inotify12
+inotify13 inotify13
 
 fanotify01 fanotify01
 fanotify02 fanotify02
testcases/kernel/syscalls/inotify/.gitignore +1 -0
index f6e5c546ac4032766cee9f16b3689be5d0302d37..b597ea63f493b2da2a58017d3eb7b16a7b033fc5 100644
@@ -10,3 +10,4 @@
 /inotify10
 /inotify11
 /inotify12
+/inotify13
testcases/kernel/syscalls/inotify/inotify13.c new file +211 -1
new file mode 100644
index 0000000000000000000000000000000000000000..9ee660e041435d16b63097ba8e2f86515d6bac57
@@ -0,0 +1,211 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2026 Hungyu Lin <[email protected]>
+ */
+
+/*\
+ * Verify that two IN_MOVED_FROM events that differ only in their rename
+ * cookie are not coalesced.
+ *
+ * inotify(7) allows coalescing only of events with identical wd, mask,
+ * cookie and name. A regression introduced in v2.6.31 dropped the cookie
+ * comparison. On affected kernels, the second IN_MOVED_FROM is coalesced
+ * into the first, leaving the following IN_MOVED_TO without a matching
+ * IN_MOVED_FROM cookie.
+ *
+ * With source/ and target/ watched for IN_MOVE, and without reading the queue:
+ *
+ * 1. Move source/entry to unwatched/ (IN_MOVED_FROM, cookie A).
+ * 2. Recreate source/entry.
+ * 3. Move source/entry to target/ (IN_MOVED_FROM and IN_MOVED_TO, cookie B).
+ *
+ * All three events must be queued, with A != B. IN_CREATE is not watched, so
+ * the two IN_MOVED_FROM events are adjacent; the queue is not read in between,
+ * because only the last unread event can be merged with a new one. Both a
+ * regular file and a directory are tested.
+ */
+
+#include "config.h"
+#include <fcntl.h>
+#include <string.h>
+#include "tst_test.h"
+#include "inotify.h"
+
+#ifdef HAVE_SYS_INOTIFY_H
+#include <sys/inotify.h>
+
+#define SOURCE "source/entry"
+#define UNWATCHED "unwatched/entry"
+#define TARGET "target/entry"
+
+struct event_t {
+       const char *name;
+       unsigned int mask;
+       int wd;
+};
+
+static struct tcase {
+       const char *type;
+       unsigned int isdir;
+} tcases[] = {
+       { "file", 0 },
+       { "directory", IN_ISDIR },
+};
+
+static int fd_notify = -1;
+
+static void create_entry(unsigned int isdir, const char *path)
+{
+       if (isdir)
+               SAFE_MKDIR(path, 0700);
+       else
+               SAFE_TOUCH(path, 0600, NULL);
+}
+
+static void delete_entry(unsigned int isdir, const char *path)
+{
+       if (isdir)
+               SAFE_RMDIR(path);
+       else
+               SAFE_UNLINK(path);
+}
+
+static void verify_inotify(unsigned int n)
+{
+       const char *type = tcases[n].type;
+       unsigned int isdir = tcases[n].isdir;
+       struct event_t event_set[] = {
+               { "entry", IN_MOVED_FROM | isdir, 0 },
+               { "entry", IN_MOVED_FROM | isdir, 0 },
+               { "entry", IN_MOVED_TO | isdir, 0 },
+       };
+       union {
+               struct inotify_event align;
+               char data[4096];
+       } buf;
+       struct inotify_event *event;
+       const struct event_t *expected;
+       unsigned int cookies[ARRAY_SIZE(event_set)] = { 0 };
+       unsigned int count = 0, i;
+       size_t pos = 0, len;
+       int wd_source, wd_target;
+
+       fd_notify = SAFE_MYINOTIFY_INIT1(O_NONBLOCK);
+       wd_source = SAFE_MYINOTIFY_ADD_WATCH(fd_notify, "source", IN_MOVE);
+       wd_target = SAFE_MYINOTIFY_ADD_WATCH(fd_notify, "target", IN_MOVE);
+
+       event_set[0].wd = wd_source;
+       event_set[1].wd = wd_source;
+       event_set[2].wd = wd_target;
+
+       create_entry(isdir, SOURCE);
+       SAFE_RENAME(SOURCE, UNWATCHED);
+       create_entry(isdir, SOURCE);
+       SAFE_RENAME(SOURCE, TARGET);
+
+       len = SAFE_READ(SAFE_READ_ANY_EAGAIN, fd_notify, buf.data, sizeof(buf.data));
+       SAFE_CLOSE(fd_notify);
+
+       /* Reuse these paths for the next case and for -i iterations. */
+       delete_entry(isdir, UNWATCHED);
+       delete_entry(isdir, TARGET);
+
+       while (pos < len) {
+               if (len - pos < sizeof(*event))
+                       tst_brk(TBROK, "%s: event %u has %zu header bytes (expected at least %zu)",
+                               type, count, len - pos, sizeof(*event));
+
+               event = (struct inotify_event *)(buf.data + pos);
+               if (event->len > len - pos - sizeof(*event))
+                       tst_brk(TBROK, "%s: event %u claims %u name bytes but only %zu are left in the buffer, wd=%d mask=%#x cookie=%u",
+                               type, count, event->len, len - pos - sizeof(*event),
+                               event->wd, event->mask, event->cookie);
+
+               if (count >= ARRAY_SIZE(event_set)) {
+                       tst_res(TFAIL, "%s: got unnecessary event %u: mask=%#x wd=%d cookie=%u len=%u name=\"%.*s\"",
+                               type, count, event->mask, event->wd, event->cookie,
+                               event->len, (int)event->len, event->name);
+                       return;
+               }
+
+               expected = &event_set[count];
+               if (event->mask != expected->mask) {
+                       tst_res(TFAIL, "%s: got event %u: mask=%#x (expected %#x) wd=%d cookie=%u name=\"%.*s\"",
+                               type, count, event->mask, expected->mask, event->wd,
+                               event->cookie, (int)event->len, event->name);
+                       return;
+               }
+
+               if (event->wd != expected->wd) {
+                       tst_res(TFAIL, "%s: got event %u: mask=%#x wd=%d (expected %d) cookie=%u name=\"%.*s\"",
+                               type, count, event->mask, event->wd, expected->wd,
+                               event->cookie, (int)event->len, event->name);
+                       return;
+               }
+
+               if (event->len < strlen(expected->name) + 1 ||
+                   strncmp(expected->name, event->name, event->len)) {
+                       tst_res(TFAIL, "%s: got event %u: mask=%#x wd=%d cookie=%u len=%u name=\"%.*s\" (expected \"%s\")",
+                               type, count, event->mask, event->wd, event->cookie,
+                               event->len, (int)event->len, event->name,
+                               expected->name);
+                       return;
+               }
+
+               cookies[count] = event->cookie;
+               tst_res(TINFO, "%s: event %u: wd=%d mask=%#x cookie=%u len=%u name=\"%.*s\"",
+                       type, count, event->wd, event->mask, event->cookie,
+                       event->len, (int)event->len, event->name);
+               count++;
+               pos += sizeof(*event) + event->len;
+       }
+
+       for (i = count; i < ARRAY_SIZE(event_set); i++) {
+               expected = &event_set[i];
+               tst_res(TFAIL, "%s: didn't get event %u: mask=%#x wd=%d name=\"%s\"",
+                       type, i, expected->mask, expected->wd, expected->name);
+       }
+
+       if (count < ARRAY_SIZE(event_set))
+               return;
+
+       if (cookies[0] == cookies[1]) {
+               tst_res(TFAIL, "%s: got second FROM cookie=%u (expected != %u)",
+                       type, cookies[1], cookies[0]);
+               return;
+       }
+
+       if (cookies[1] != cookies[2]) {
+               tst_res(TFAIL, "%s: got TO cookie=%u (expected %u from second FROM)",
+                       type, cookies[2], cookies[1]);
+               return;
+       }
+
+       tst_res(TPASS, "%s: FROM cookies %u and %u differ, TO cookie %u matches the second",
+               type, cookies[0], cookies[1], cookies[2]);
+}
+
+static void setup(void)
+{
+       SAFE_MKDIR("source", 0700);
+       SAFE_MKDIR("unwatched", 0700);
+       SAFE_MKDIR("target", 0700);
+}
+
+static void cleanup(void)
+{
+       if (fd_notify >= 0)
+               SAFE_CLOSE(fd_notify);
+}
+
+static struct tst_test test = {
+       .test = verify_inotify,
+       .tcnt = ARRAY_SIZE(tcases),
+       .setup = setup,
+       .cleanup = cleanup,
+       .needs_tmpdir = 1,
+};
+
+#else
+TST_TEST_TCONF("system doesn't have required inotify support");
+#endif
--

2.43.0

Reviewed-by
#3 lore raw
Quoted 13 lines · Hungyu Lin
Quoted 11 lines

Queue an unpaired IN_MOVED_FROM event, followed by an IN_MOVED_FROM and IN_MOVED_TO pair with the same source watch descriptor and name as the first event. Watch only IN_MOVE so creation of the replacement entry does not separate the two IN_MOVED_FROM events in the queue.

Affected kernels coalesce the second IN_MOVED_FROM into the first, leaving the IN_MOVED_TO cookie without a matching IN_MOVED_FROM cookie. Check that all three events are retained, the two moves have distinct cookies, and the second IN_MOVED_FROM matches the IN_MOVED_TO cookie. Cover both regular files and directories.

Quoted 6 lines
Quoted 2 lines

Thank you for your patience with my review

Thank you!

I have no problem to fix very minor issues reported by agent before merge, but what bothers me there is no kernel fix for the failure. If I'm not mistaken, the fix in kernel [1] was reviewed [2] but not yet merged (I looked to next tree as well).
There are other fixes for 63c882a05416 ("inotify: reimplement inotify using fsnotify") in stable, e.g. e422777fdd474 ("fsnotify: Fix stale object mask after concurrent mark updates"), but I haven't found this one.

Merging test helps kernel developers to test if the problem was fixed but break for it for anybody who uses master for regular testing without even explaining fix is WIP. But I suppose it will be processed soon because stable was in Cc, right?

We have runtest/staging, but that's for new functionality, not for tests which reproduce problem which was not fixed.

@Jan @Amir is there a git kernel tree for fanotify and inotify? I found only VFS. If yes, please update MAINTAINERS file.

Kind regards,
Petr

[1] https://lore.kernel.org/lkml/[email protected]/
[2] https://lore.kernel.org/lkml/CAOQ4uxj3et_cyzD1iWHr3_z1dH1R3WQMCOm+kO2Q90TXjaV9pQ@mail.gmail.com/
[3] https://git.kernel.org/pub/scm/linux/kernel/git/overlayfs/vfs.git/

#4 lore raw
Quoted 38 lines · Petr Vorel
Quoted 14 lines · Hungyu Lin
Quoted 11 lines

Queue an unpaired IN_MOVED_FROM event, followed by an IN_MOVED_FROM and IN_MOVED_TO pair with the same source watch descriptor and name as the first event. Watch only IN_MOVE so creation of the replacement entry does not separate the two IN_MOVED_FROM events in the queue.

Affected kernels coalesce the second IN_MOVED_FROM into the first, leaving the IN_MOVED_TO cookie without a matching IN_MOVED_FROM cookie. Check that all three events are retained, the two moves have distinct cookies, and the second IN_MOVED_FROM matches the IN_MOVED_TO cookie. Cover both regular files and directories.

Quoted 6 lines
Quoted 2 lines

Thank you for your patience with my review

Thank you!

I have no problem to fix very minor issues reported by agent before merge, but what bothers me there is no kernel fix for the failure. If I'm not mistaken, the fix in kernel [1] was reviewed [2] but not yet merged (I looked to next tree as well).
There are other fixes for 63c882a05416 ("inotify: reimplement inotify using fsnotify") in stable, e.g. e422777fdd474 ("fsnotify: Fix stale object mask after concurrent mark updates"), but I haven't found this one.

Merging test helps kernel developers to test if the problem was fixed but break for it for anybody who uses master for regular testing without even explaining fix is WIP. But I suppose it will be processed soon because stable was in Cc, right?

Quoted 6 lines

We have runtest/staging, but that's for new functionality, not for tests which reproduce problem which was not fixed.

@Jan @Amir is there a git kernel tree for fanotify and inotify? I found only VFS. If yes, please update MAINTAINERS file.

FSNOTIFY: FILESYSTEM NOTIFICATION INFRASTRUCTURE
M:      Jan Kara <[email protected]>
R:      Amir Goldstein <[email protected]>
L:      [email protected]
S:      Maintained
T:      git git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs.git
fsnotify
F:      fs/notify/
F:      include/linux/fsnotify*.h

We could probably rename/move the fanotify/inotify entries to appear as sub entries mimicking the VFS entries:

FILESYSTEM NOTIFICATION (fsnotify infrastructure)
...
FILESYSTEM NOTIFICATION [INOTIFY]
...
FILESYSTEM NOTIFICATION [FANOTIFY]

Thanks,
Amir.

#5 lore raw
Quoted 46 lines · Petr Vorel
Quoted 38 lines
Quoted 14 lines · Hungyu Lin
Quoted 11 lines

Queue an unpaired IN_MOVED_FROM event, followed by an IN_MOVED_FROM and IN_MOVED_TO pair with the same source watch descriptor and name as the first event. Watch only IN_MOVE so creation of the replacement entry does not separate the two IN_MOVED_FROM events in the queue.

Affected kernels coalesce the second IN_MOVED_FROM into the first, leaving the IN_MOVED_TO cookie without a matching IN_MOVED_FROM cookie. Check that all three events are retained, the two moves have distinct cookies, and the second IN_MOVED_FROM matches the IN_MOVED_TO cookie. Cover both regular files and directories.

Quoted 6 lines
Quoted 2 lines

Thank you for your patience with my review

Thank you!

I have no problem to fix very minor issues reported by agent before merge, but what bothers me there is no kernel fix for the failure. If I'm not mistaken, the fix in kernel [1] was reviewed [2] but not yet merged (I looked to next tree as well).
There are other fixes for 63c882a05416 ("inotify: reimplement inotify using fsnotify") in stable, e.g. e422777fdd474 ("fsnotify: Fix stale object mask after concurrent mark updates"), but I haven't found this one.

Merging test helps kernel developers to test if the problem was fixed but break for it for anybody who uses master for regular testing without even explaining fix is WIP. But I suppose it will be processed soon because stable was in Cc, right?

Quoted 5 lines

We have runtest/staging, but that's for new functionality, not for tests which reproduce problem which was not fixed.

@Jan @Amir is there a git kernel tree for fanotify and inotify? I found only VFS. If yes, please update MAINTAINERS file.

Quoted 9 lines
FSNOTIFY: FILESYSTEM NOTIFICATION INFRASTRUCTURE
M:      Jan Kara <[email protected]>
R:      Amir Goldstein <[email protected]>
L:      [email protected]
S:      Maintained
T:      git git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs.git
fsnotify
F:      fs/notify/
F:      include/linux/fsnotify*.h

Thanks a lot for pointing this doc, I overlooked because I noticed the earlier:

FANOTIFY
M: Jan Kara <[email protected]>
R: Amir Goldstein <[email protected]>
R: Matt Bobrowski <[email protected]>
L: [email protected]
S: Maintained
F: fs/notify/fanotify/
F: include/linux/fanotify.h
F: include/uapi/linux/fanotify.h

INOTIFY
M: Jan Kara <[email protected]>
R: Amir Goldstein <[email protected]>
L: [email protected]
S: Maintained
F: Documentation/filesystems/inotify.rst
F: fs/notify/inotify/
F: include/linux/inotify.h
F: include/uapi/linux/inotify.h

Quoted 8 lines

We could probably rename/move the fanotify/inotify entries to appear as sub entries mimicking the VFS entries:

FILESYSTEM NOTIFICATION (fsnotify infrastructure)
...
FILESYSTEM NOTIFICATION [INOTIFY]
...
FILESYSTEM NOTIFICATION [FANOTIFY]

+1

Kind regards,
Petr

Quoted 2 lines

Thanks,
Amir.

#6 lore raw

Hi Hungyu Lin, Amir,

Given this is a test for bug from v2.6.31 I kept it in runtest/syscalls and merged, thank you!

I suppose the fix is [PATCH v2] inotify: do not coalesce rename events [1]. Although Jan stated "added to my tree" [2] I haven't found it in his for_next branch [3]. Please update the hash in the test or Cc me once it's merged.

Kind regards,
Petr

[1] https://lore.kernel.org/lkml/[email protected]/
[2] https://lore.kernel.org/lkml/u6fltejwlekpknwpxzy6pefif7ld3dhwvzqqalbzmtwtzjjaj2@hsulwhevxfds/
[3] https://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs.git/log/?h=for_next