Skip to contentSkip to conversation

libdm: fix races with udev

PATCH New Mikulas Patocka 2 participants 3 messages Sep 16, 2013 – Sep 23, 2013

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 libdm: fix races with udev

libdm: fix races with udev

On newer systems (Debian 6 and newer), udev manages nodes in /dev/mapper directory. It creates, deletes and renames the nodes according to the state of the kernel driver.

dmsetup tries to manage nodes in /dev/mapper too, so it can race with udev. dmsetup checks if the node was created/deleted/renamed with the stat syscall, and skips the operation if it was. However, if udev creates/deletes/renames the node after the stat syscall and before the mknod/unlink/rename syscall, dmsetup reports an error.

These races can be easily provoked by inserting sleep at appropriate places.

This patch avoids the error messages by ignoring errors that could happen as a result of the race.

---
 libdm/libdm-common.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
libdm/libdm-common.c +7 -3
Index: lvm2-copy/libdm/libdm-common.c
===================================================================
@@ -966,7 +966,9 @@ static int _add_dev_node(const char *dev
 
 	(void) dm_prepare_selinux_context(path, S_IFBLK);
 	old_mask = umask(0);
-	if (mknod(path, S_IFBLK | mode, dev) < 0) {
+
+	/* The node may already have been created by udev. So ignore EEXIST. */
+	if (mknod(path, S_IFBLK | mode, dev) < 0 && errno != EEXIST) {
 		log_error("%s: mknod for %s failed: %s", path, dev_name, strerror(errno));
 		umask(old_mask);
 		(void) dm_prepare_selinux_context(NULL, 0);
@@ -998,7 +1000,8 @@ static int _rm_dev_node(const char *dev_
 		log_warn("Node %s was not removed by udev. "
 			 "Falling back to direct node removal.", path);
 
-	if (unlink(path) < 0) {
+	/* udev may already have deleted the node. Ignore ENOENT. */
+	if (unlink(path) < 0 && errno != ENOENT) {
 		log_error("Unable to unlink device node for '%s'", dev_name);
 		return 0;
 	}
@@ -1054,7 +1057,8 @@ static int _rename_dev_node(const char *
 			 "Falling back to direct node rename.",
 			 oldpath, newpath);
 
-	if (rename(oldpath, newpath) < 0) {
+	/* udev may already have renamed the node. Ignore ENOENT. */
+	if (rename(oldpath, newpath) < 0 && errno != ENOENT) {
 		log_error("Unable to rename device node from '%s' to '%s'",
 			  old_name, new_name);
 		return 0;
#1 lore raw

PATCH 1/1 libdm: fix races with udev

Quoted 6 lines · Mikulas Patocka

libdm: fix races with udev

On newer systems (Debian 6 and newer), udev manages nodes in /dev/mapper directory. It creates, deletes and renames the nodes according to the state of the kernel driver.

Quoted 6 lines

dmsetup tries to manage nodes in /dev/mapper too, so it can race with udev. dmsetup checks if the node was created/deleted/renamed with the stat syscall, and skips the operation if it was. However, if udev creates/deletes/renames the node after the stat syscall and before the mknod/unlink/rename syscall, dmsetup reports an error.

These checks are performed after udev is synchronized (the dm_udev_wait call) and after the udev processing is complete, so we normally shouldn't get into this problem unless there is some bug in certain version of lvm2/udev rules.

Is this reproducible with upstream code as well or was that with some older version only that is currently in Debian?

Peter

#2 lore raw

PATCH 1/1 libdm: fix races with udev

Quoted 14 lines · Peter Rajnoha
Quoted 6 lines · Mikulas Patocka

libdm: fix races with udev

On newer systems (Debian 6 and newer), udev manages nodes in /dev/mapper directory. It creates, deletes and renames the nodes according to the state of the kernel driver.

Quoted 6 lines

dmsetup tries to manage nodes in /dev/mapper too, so it can race with udev. dmsetup checks if the node was created/deleted/renamed with the stat syscall, and skips the operation if it was. However, if udev creates/deletes/renames the node after the stat syscall and before the mknod/unlink/rename syscall, dmsetup reports an error.

Quoted 8 lines

These checks are performed after udev is synchronized (the dm_udev_wait call) and after the udev processing is complete, so we normally shouldn't get into this problem unless there is some bug in certain version of lvm2/udev rules.

Is this reproducible with upstream code as well or was that with some older version only that is currently in Debian?

Peter

This race condition can be easily provoked with this patch. Apply it and try dmsetup create/remove/rename.

One Debian 6, races in all three operations create/remove/rename can be provoked.

On Debian Sid and on RHEL 6, only races in create/rename can be provoked. Race in remove doesn't happen.

Mikulas

libdm/libdm-common.c +6 -0
Index: LVM2.2.02.102/libdm/libdm-common.c
===================================================================
@@ -964,6 +964,8 @@
 		log_warn("%s not set up by udev: Falling back to direct "
 			 "node creation.", path);
 
+	fprintf(stderr, "sleeping 1\n"); sleep(3);
+
 	(void) dm_prepare_selinux_context(path, S_IFBLK);
 	old_mask = umask(0);
 	if (mknod(path, S_IFBLK | mode, dev) < 0) {
@@ -998,6 +1000,8 @@
 		log_warn("Node %s was not removed by udev. "
 			 "Falling back to direct node removal.", path);
 
+	fprintf(stderr, "sleeping 2\n"); sleep(3);
+
 	if (unlink(path) < 0) {
 		log_error("Unable to unlink device node for '%s'", dev_name);
 		return 0;
@@ -1054,6 +1058,8 @@
 			 "Falling back to direct node rename.",
 			 oldpath, newpath);
 
+	fprintf(stderr, "sleeping 3\n"); sleep(3);
+
 	if (rename(oldpath, newpath) < 0) {
 		log_error("Unable to rename device node from '%s' to '%s'",
 			  old_name, new_name);