[RFC] fsnotify: move unlink/rmdir events to after d_delete

A regression was introduced by 116b9731ad76 ("fsnotify: add empty
fsnotify_{unlink,rmdir}() hooks"), which moved the fsnotify event for
unlink and rmdir to before the d_delete.  This was noticed by a tool we
have internally for validating a FUSE file system.  This tool watches
for IN_DELETE events and then stat's the file to make sure the file was
actually deleted.  This started failing on our newer kernels, and it was
traced to this patch.

The problem is there's a slight window where we emit the event and we
delete the dentry.  We can easily get the event before we've called
d_delete, and then stat the file before we're able to remove it.  This
is easily reproducible with the following reproducer

static int failed = 0;

static void *watch_inotify(void *arg)
{
	char name_buf[4096];
	char *buf = malloc(EVENT_BUF_LEN);
	char *directory = (char *)arg;
	int fd;
	int wd;

	if (!buf) {
		fprintf(stderr, "Failed to allocate inotify buf\n");
		return NULL;
	}

	fd = inotify_init();
	if (fd < 0) {
		fprintf(stderr, "Failed to init inotify %d %s\n",
			errno, strerror(errno));
		return NULL;
	}

	wd = inotify_add_watch(fd, directory, IN_DELETE);
	if (wd < 0) {
		fprintf(stderr, "inotify watch failed %d %s\n", errno,
			strerror(errno));
		close(fd);
		return NULL;
	}

	while (!failed) {
		ssize_t len = read(fd, buf, EVENT_BUF_LEN);
		size_t cur = 0;

		if (len < 0) {
			fprintf(stderr, "failed to read from inotify %d %s\n",
				errno, strerror(errno));
			break;
		}

		while (cur < len) {
			struct inotify_event *event =
				(struct inotify_event *)(buf + cur);
			struct stat st;
			int ret;

			cur += EVENT_SIZE + event->len;
			if (event->len == 0)
				continue;
			if (!(event->mask & IN_DELETE))
				continue;
			ret = snprintf(name_buf, 4096, "%s/%s", directory,
				       event->name);
			if (ret < 0) {
				fprintf(stderr, "Couldn't snprintf %d (%s)\n",
						errno, strerror(errno));
				break;
			}
			if (ret >= 4096) {
				fprintf(stderr, "Name was truncated, pick a shorter dir name\n");
				break;
			}
			ret = stat(name_buf, &st);
			if (!ret) {
				fprintf(stderr, "Found file %s after a delete event\n",
					name_buf);
				failed = 1;
				break;
			}
		}
	}
	inotify_rm_watch( fd, wd );
	close(fd);
	return NULL;
}

int main(int argc, char **argv)
{
	char buf[4096];
	char *dir;
	pthread_t inotify_thread;
	int ret, i;

	if (argc != 2) {
		fprintf(stderr, "Must specify a directory to use\n");
		return 1;
	}

	dir = strdup(argv[1]);
	if (!dir) {
		fprintf(stderr, "Couldn't dup directory\n");
		return 1;
	}

	printf("Creating files...\n");
	for (i = 0; i < NR_FILES; i++) {
		ret = snprintf(buf, 4096, "%s/file_%d", dir, i);
		if (ret < 0) {
			fprintf(stderr, "Couldn't snprintf %d (%s)\n",
				errno, strerror(errno));
			goto out;
		}
		if (ret >= 4096) {
			fprintf(stderr, "Name was truncated, pick a shorter dir name\n");
			goto out;
		}
		ret = creat(buf, 0600);
		if (ret < 0) {
			fprintf(stderr, "Failed to create %s %d %s\n",
				buf, errno, strerror(errno));
			goto out;
		}
		close(ret);
	}

	printf("Starting inotify thread\n");
	ret = pthread_create(&inotify_thread, NULL, watch_inotify, dir);
	if (ret) {
		fprintf(stderr, "Couldn't create inotify thread %d (%s)\n",
			errno, strerror(errno));
		goto out;
	}

	printf("Sleeping for a second\n");
	sleep(1);
	printf("Starting delete\n");
	for (i = 0; i < NR_FILES; i++) {
		ret = snprintf(buf, 4096, "%s/file_%d", dir, i);
		if (ret < 0) {
			fprintf(stderr, "Couldn't snprintf %d (%s)\n",
				errno, strerror(errno));
			goto out_pthread;
		}
		if (ret >= 4096) {
			fprintf(stderr, "Name was truncated, pick a shorter dir name\n");
			goto out_pthread;
		}
		ret = unlink(buf);
		if (ret)
			goto out_pthread;
	}
out_pthread:
	pthread_cancel(inotify_thread);
	pthread_join(inotify_thread, NULL);
out:
	free(dir);
	return ret ? 1 : failed;
}

Fix this by adjusting the fsnotify hooks to take the inode that we
unlink'ed/rmdir'ed, and simply hold onto the inode so that we can still
access it after the d_delete.  This patch returns us to the previous
behavior and fixes the problem we were seeing.

Fixes: 116b9731ad76 ("fsnotify: add empty fsnotify_{unlink,rmdir}() hooks")
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
---

I'm not married to this solution at all, I'm not sure how important it is to
have the original inode.  We could also push this back into d_delete().

 fs/btrfs/ioctl.c         |  2 +-
 fs/configfs/dir.c        | 16 ++++++++++++----
 fs/devpts/inode.c        |  8 ++++++--
 fs/libfs.c               | 10 ++++++++--
 fs/namei.c               | 17 +++++++++++------
 include/linux/fsnotify.h | 16 ++++++----------
 6 files changed, 44 insertions(+), 25 deletions(-)

Message ID	df48bd23ecacdc0b95a315d176d2cc03a8ffe81c.1620852240.git.josef@toxicpanda.com (mailing list archive)
State	New, archived
Headers	show Return-Path: <linux-fsdevel-owner@kernel.org> From: Josef Bacik <josef@toxicpanda.com> To: linux-btrfs@vger.kernel.org, kernel-team@fb.com, linux-fsdevel@vger.kernel.org, jack@suse.cz, amir73il@gmail.com, viro@zeniv.linux.org.uk Subject: [RFC][PATCH] fsnotify: move unlink/rmdir events to after d_delete Date: Wed, 12 May 2021 16:49:10 -0400 Message-Id: <df48bd23ecacdc0b95a315d176d2cc03a8ffe81c.1620852240.git.josef@toxicpanda.com> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Precedence: bulk
Series	[RFC] fsnotify: move unlink/rmdir events to after d_delete \| expand [RFC] fsnotify: move unlink/rmdir events to after d_delete

[RFC] fsnotify: move unlink/rmdir events to after d_delete

Commit Message

Patch