mount: dont execute propagate_umount() many times for same mounts

The reason of this optimization is that umount() can hold namespace_sem for a long time, this semaphore is global, so it affects all users. Recently Eric W. Biederman added a per mount namespace limit on the number of mounts. The default number of mounts allowed per mount namespace at 100,000. Currently this value is allowed to construct a tree which requires hours to be umounted.

In a worse case the current complexity of umount_tree() is O(n^3).
* Enumirate all mounts in a target tree (propagate_umount)
* Enumirate mounts to find where these changes have to be propagated (mark_umount_candidates)
* Enumirate mounts to find a requered mount by parent and dentry (__lookup_mnt_lat)

The worse case is when all mounts from the tree live in the same shared group. In this case we have to enumirate all mounts on each step.

Here we can optimize the second step. We don't need to make it for mounts which we already met when we did this step for previous mounts. It reduces the complexity of umount_tree() to O(n^2).

The reason of this optimization is that umount() can hold namespace_sem for a long time, this semaphore is global, so it affects all users. Recently Eric W. Biederman added a per mount namespace limit on the number of mounts. The default number of mounts allowed per mount namespace at 100,000. Currently this value is allowed to construct a tree which requires hours to be umounted.

In a worse case the current complexity of umount_tree() is O(n^3).
* Enumirate all mounts in a target tree (propagate_umount)
* Enumirate mounts to find where these changes have to be propagated (mark_umount_candidates)
* Enumirate mounts to find a requered mount by parent and dentry (__lookup_mnt_lat)

The worse case is when all mounts from the tree live in the same shared group. In this case we have to enumirate all mounts on each step.

Here we can optimize the second step. We don't need to make it for mounts which we already met when we did this step for previous mounts. It reduces the complexity of umount_tree() to O(n^2).

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

Solve this in the most straight forward way possible, by adding a new mount flag to mark parts of the mount propagation tree that have been visited, and use that mark to skip parts of the mount propagation tree that have already been visited during an unmount. This guarantees that each mountpoint in the possibly overlapping mount propagation trees will be visited exactly once.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate setting and clearling the visited mount mark.

Here is a script to generate such mount tree:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mounts | before | after (real sec)
-----------------------------

  1024 |  0.071 | 0.024
  2048 |  0.184 | 0.030
  4096 |  0.604 | 0.040
  8912 |  4.471 | 0.043
 16384 | 34.826 | 0.082
 32768 |        | 0.151
 65536 |        | 0.289
131072 |        | 0.659

Andrei Vagin fixing this performance problem is part of the work to fix CVE-2016-6213.

Cc [email protected]

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---

Andrei can you take a look at this patch and see if you can see any problems. My limited testing suggests this approach does a much better job of solving the problem you were seeing. With the time looking almost linear in the number of mounts now.

 fs/pnode.c            | 125 ++++++++++++++++++++++++++++++++++++++++++++++++--
 fs/pnode.h            |   4 ++
 include/linux/mount.h |   2 +
 3 files changed, 126 insertions(+), 5 deletions(-)

fs/pnode.c +120 -5

index 234a9ac49958..3acce0c75f94 100644
@@ -164,6 +164,120 @@ static struct mount *propagation_next(struct mount *m,
 	}
 }
 
+/*
+ * get the next mount in the propagation tree (that has not been visited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_visit_next(struct mount *m,
+					    struct mount *origin)
+{
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_VISITED(m))
+		return NULL;
+
+	SET_MNT_VISITED(m);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		struct mount *slave = first_slave(m);
+		while (1) {
+			if (!IS_MNT_VISITED(slave))
+				return slave;
+			if (slave->mnt_slave.next == &m->mnt_slave_list)
+				break;
+			slave = next_slave(slave);
+		}
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				if (!IS_MNT_VISITED(next))
+					return next;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+				if (!IS_MNT_VISITED(m))
+					return m;
+			}
+		}
+
+		/* back at master */
+		m = master;
+	}
+}
+
+/*
+ * get the next mount in the propagation tree (that has not been revisited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_revisit_next(struct mount *m,
+					      struct mount *origin)
+{
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_VISITED(m))
+		return NULL;
+
+	CLEAR_MNT_VISITED(m);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		struct mount *slave = first_slave(m);
+		while (1) {
+			if (IS_MNT_VISITED(slave))
+				return slave;
+			if (slave->mnt_slave.next == &m->mnt_slave_list)
+				break;
+			slave = next_slave(slave);
+		}
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				if (IS_MNT_VISITED(next))
+					return next;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+				if (IS_MNT_VISITED(m))
+					return m;
+			}
+		}
+
+		/* back at master */
+		m = master;
+	}
+}
+
 static struct mount *next_group(struct mount *m, struct mount *origin)
 {
 	while (1) {

@@ -399,11 +513,12 @@ static void mark_umount_candidates(struct mount *mnt)
 
 	BUG_ON(parent == mnt);
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	for (m = propagation_visit_next(parent, parent); m;
+			m = propagation_visit_next(m, parent)) {
 		struct mount *child = __lookup_mnt_last(&m->mnt,
 						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
+		if (child && (!IS_MNT_LOCKED(child) ||
+			      IS_MNT_MARKED(m))) {
 			SET_MNT_MARK(child);
 		}
 	}

@@ -420,8 +535,8 @@ static void __propagate_umount(struct mount *mnt)
 
 	BUG_ON(parent == mnt);
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	for (m = propagation_revisit_next(parent, parent); m;
+			m = propagation_revisit_next(m, parent)) {
 
 		struct mount *child = __lookup_mnt_last(&m->mnt,
 						mnt->mnt_mountpoint);

fs/pnode.h +4 -0

index 550f5a8b4fcf..988ea4945764 100644
@@ -21,6 +21,10 @@
 #define CLEAR_MNT_MARK(m) ((m)->mnt.mnt_flags &= ~MNT_MARKED)
 #define IS_MNT_LOCKED(m) ((m)->mnt.mnt_flags & MNT_LOCKED)
 
+#define IS_MNT_VISITED(m) ((m)->mnt.mnt_flags & MNT_VISITED)
+#define SET_MNT_VISITED(m) ((m)->mnt.mnt_flags |= MNT_VISITED)
+#define CLEAR_MNT_VISITED(m) ((m)->mnt.mnt_flags &= ~MNT_VISITED)
+
 #define CL_EXPIRE    		0x01
 #define CL_SLAVE     		0x02
 #define CL_COPY_UNBINDABLE	0x04

include/linux/mount.h +2 -0

index 9227b190fdf2..6048045b96c3 100644
@@ -52,6 +52,8 @@ struct mnt_namespace;
 
 #define MNT_INTERNAL	0x4000
 
+#define MNT_VISITED		0x010000
+
 #define MNT_LOCK_ATIME		0x040000
 #define MNT_LOCK_NOEXEC		0x080000
 #define MNT_LOCK_NOSUID		0x100000

Signature · 1 line

2.8.3

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

Solve this in the most straight forward way possible, by adding a new mount flag to mark parts of the mount propagation tree that have been visited, and use that mark to skip parts of the mount propagation tree that have already been visited during an unmount. This guarantees that each mountpoint in the possibly overlapping mount propagation trees will be visited exactly once.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate setting and clearling the visited mount mark.

Here is a script to generate such mount tree:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mounts | before | after (real sec)
-----------------------------

  1024 |  0.071 | 0.024
  2048 |  0.184 | 0.030
  4096 |  0.604 | 0.040
  8912 |  4.471 | 0.043
 16384 | 34.826 | 0.082
 32768 |        | 0.151
 65536 |        | 0.289
131072 |        | 0.659

Andrei Vagin fixing this performance problem is part of the work to fix CVE-2016-6213.

Cc [email protected]

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <[email protected]>

---

Andrei can you take a look at this patch and see if you can see any problems. My limited testing suggests this approach does a much better job of solving the problem you were seeing. With the time looking almost linear in the number of mounts now.

 fs/pnode.c            | 125 ++++++++++++++++++++++++++++++++++++++++++++++++--
 fs/pnode.h            |   4 ++
 include/linux/mount.h |   2 +
 3 files changed, 126 insertions(+), 5 deletions(-)

fs/pnode.c +120 -5

index 234a9ac49958..3acce0c75f94 100644
@@ -164,6 +164,120 @@ static struct mount *propagation_next(struct mount *m,
 	}
 }
 
+/*
+ * get the next mount in the propagation tree (that has not been visited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_visit_next(struct mount *m,
+					    struct mount *origin)
+{
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_VISITED(m))
+		return NULL;
+
+	SET_MNT_VISITED(m);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		struct mount *slave = first_slave(m);
+		while (1) {
+			if (!IS_MNT_VISITED(slave))
+				return slave;
+			if (slave->mnt_slave.next == &m->mnt_slave_list)
+				break;
+			slave = next_slave(slave);
+		}
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				if (!IS_MNT_VISITED(next))
+					return next;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+				if (!IS_MNT_VISITED(m))
+					return m;
+			}
+		}
+
+		/* back at master */
+		m = master;
+	}
+}
+
+/*
+ * get the next mount in the propagation tree (that has not been revisited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_revisit_next(struct mount *m,
+					      struct mount *origin)
+{
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_VISITED(m))
+		return NULL;
+
+	CLEAR_MNT_VISITED(m);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		struct mount *slave = first_slave(m);
+		while (1) {
+			if (IS_MNT_VISITED(slave))
+				return slave;
+			if (slave->mnt_slave.next == &m->mnt_slave_list)
+				break;
+			slave = next_slave(slave);
+		}
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				if (IS_MNT_VISITED(next))
+					return next;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+				if (IS_MNT_VISITED(m))
+					return m;
+			}
+		}
+
+		/* back at master */
+		m = master;
+	}
+}
+
 static struct mount *next_group(struct mount *m, struct mount *origin)
 {
 	while (1) {

@@ -399,11 +513,12 @@ static void mark_umount_candidates(struct mount *mnt)
 
 	BUG_ON(parent == mnt);
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	for (m = propagation_visit_next(parent, parent); m;
+			m = propagation_visit_next(m, parent)) {
 		struct mount *child = __lookup_mnt_last(&m->mnt,
 						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
+		if (child && (!IS_MNT_LOCKED(child) ||
+			      IS_MNT_MARKED(m))) {
 			SET_MNT_MARK(child);
 		}
 	}

@@ -420,8 +535,8 @@ static void __propagate_umount(struct mount *mnt)
 
 	BUG_ON(parent == mnt);
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	for (m = propagation_revisit_next(parent, parent); m;
+			m = propagation_revisit_next(m, parent)) {
 
 		struct mount *child = __lookup_mnt_last(&m->mnt,
 						mnt->mnt_mountpoint);

fs/pnode.h +4 -0

index 550f5a8b4fcf..988ea4945764 100644
@@ -21,6 +21,10 @@
 #define CLEAR_MNT_MARK(m) ((m)->mnt.mnt_flags &= ~MNT_MARKED)
 #define IS_MNT_LOCKED(m) ((m)->mnt.mnt_flags & MNT_LOCKED)
 
+#define IS_MNT_VISITED(m) ((m)->mnt.mnt_flags & MNT_VISITED)
+#define SET_MNT_VISITED(m) ((m)->mnt.mnt_flags |= MNT_VISITED)
+#define CLEAR_MNT_VISITED(m) ((m)->mnt.mnt_flags &= ~MNT_VISITED)
+
 #define CL_EXPIRE    		0x01
 #define CL_SLAVE     		0x02
 #define CL_COPY_UNBINDABLE	0x04

include/linux/mount.h +2 -0

index 9227b190fdf2..6048045b96c3 100644
@@ -52,6 +52,8 @@ struct mnt_namespace;
 
 #define MNT_INTERNAL	0x4000
 
+#define MNT_VISITED		0x010000
+
 #define MNT_LOCK_ATIME		0x040000
 #define MNT_LOCK_NOEXEC		0x080000
 #define MNT_LOCK_NOSUID		0x100000

Signature · 1 line

2.8.3

Quoted 47 lines · Eric W. Biederman

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

Solve this in the most straight forward way possible, by adding a new mount flag to mark parts of the mount propagation tree that have been visited, and use that mark to skip parts of the mount propagation tree that have already been visited during an unmount. This guarantees that each mountpoint in the possibly overlapping mount propagation trees will be visited exactly once.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate setting and clearling the visited mount mark.

Here is a script to generate such mount tree:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mounts | before | after (real sec)
-----------------------------

  1024 |  0.071 | 0.024
  2048 |  0.184 | 0.030
  4096 |  0.604 | 0.040
  8912 |  4.471 | 0.043
 16384 | 34.826 | 0.082
 32768 |        | 0.151
 65536 |        | 0.289
131072 |        | 0.659

Andrei Vagin fixing this performance problem is part of the work to fix CVE-2016-6213.

Cc [email protected]

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

Quoted 6 lines

---

Andrei can you take a look at this patch and see if you can see any problems. My limited testing suggests this approach does a much better job of solving the problem you were seeing. With the time looking almost linear in the number of mounts now.

I read this patch and I like the idea.

Then I run my tests and one of them doesn't work with this patch. I haven't found a reason yet.

Quoted 4 lines

fs/pnode.c +2 -0

+     for (m = propagation_visit_next(parent, parent); m;
+                     m = propagation_visit_next(m, parent)) {
              struct mount *child = __lookup_mnt_last(&m->mnt,
                                              mnt->mnt_mountpoint);

Here is the test:

[root@fc24 mounts]# cat run.sh
set -e
mount -t tmpfs zdtm /mnt
mkdir -p /mnt/1 /mnt/2
mount -t tmpfs zdtm /mnt/1
mount --make-shared /mnt/1
for i in seq $1; do

        mount --bind /mnt/1 `mktemp -d /mnt/1/test.XXXXXX`

done
mount --rbind /mnt/1 /mnt/2
cat /proc/self/mountinfo | grep zdtm | wc -l
time umount -l /mnt/1
cat /proc/self/mountinfo | grep zdtm | wc -l
umount /mnt/2

[root@fc24 mounts]# unshare -Urm ./run.sh 5
65

real    0m0.014s
user    0m0.000s
sys     0m0.004s

33
umount: /mnt/2: target is busy
(In some cases useful info about processes that
use the device is found by lsof(8) or fuser(1).)

Quoted 1 line

Quoted 55 lines · Andrei Vagin

Quoted 47 lines · Eric W. Biederman

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

Solve this in the most straight forward way possible, by adding a new mount flag to mark parts of the mount propagation tree that have been visited, and use that mark to skip parts of the mount propagation tree that have already been visited during an unmount. This guarantees that each mountpoint in the possibly overlapping mount propagation trees will be visited exactly once.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate setting and clearling the visited mount mark.

Here is a script to generate such mount tree:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mounts | before | after (real sec)
-----------------------------

  1024 |  0.071 | 0.024
  2048 |  0.184 | 0.030
  4096 |  0.604 | 0.040
  8912 |  4.471 | 0.043
 16384 | 34.826 | 0.082
 32768 |        | 0.151
 65536 |        | 0.289
131072 |        | 0.659

Andrei Vagin fixing this performance problem is part of the work to fix CVE-2016-6213.

Cc [email protected]

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

Quoted 6 lines

---

Andrei can you take a look at this patch and see if you can see any problems. My limited testing suggests this approach does a much better job of solving the problem you were seeing. With the time looking almost linear in the number of mounts now.

Quoted 9 lines

I read this patch and I like the idea.

Then I run my tests and one of them doesn't work with this patch. I haven't found a reason yet.

Quoted 4 lines

fs/pnode.c +2 -0

+     for (m = propagation_visit_next(parent, parent); m;
+                     m = propagation_visit_next(m, parent)) {
              struct mount *child = __lookup_mnt_last(&m->mnt,
                                              mnt->mnt_mountpoint);

The reason is that this loop is called for different "mnt", but it is executed only once with this optimization.

So I think the idea to mark parent will not work, because one parent can have a few children which have to be umounted.

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

Solve this in the most straight forward way possible, by adding a new mount flag to mark parts of the mount propagation tree that have been visited, and use that mark to skip parts of the mount propagation tree that have already been visited during an unmount. This guarantees that each mountpoint in the possibly overlapping mount propagation trees will be visited exactly once.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate setting and clearling the visited mount mark.

The skipping of already unmounted mounts has been moved from __lookup_mnt_last to mark_umount_candidates, so that the new propagation functions can notice record when the propagation tree passes through the initial set of unmounted mounts. Except in umount_tree as part of the unmounting process the only place where unmounted mounts should be found are in unmounted subtrees. All of the other callers of __lookup_mnt_last are from mounted subtrees so the not checking for unmounted mounts should not affect them.

Here is a script to generate such mount tree:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mhash  |  8192   |  8192  |  8192       | 131072 | 131072      | 104857 | 104857
mounts | before  | after  | after (sys) | after  | after (sys) |  after | after (sys)
-------------------------------------------------------------------------------------
  1024 |  0.071s | 0.023s | 0.008s      | 0.026s | 0.000s      | 0.024s | 0.008s
  2048 |  0.184s | 0.030s | 0.012s      | 0.035s | 0.008s      | 0.030s | 0.012s
  4096 |  0.604s | 0.047s | 0.012s      | 0.042s | 0.016s      | 0.032s | 0.016s
  8912 |  4.471s | 0.085s | 0.020s      | 0.059s | 0.059s      | 0.050s | 0.036s
 16384 | 34.826s | 0.105s | 0.092s      | 0.109s | 0.060s      | 0.087s | 0.068s
 32768 |         | 0.245s | 0.168s      | 0.192s | 0.144s      | 0.167s | 0.156s
 65536 |         | 0.833s | 0.716s      | 0.485s | 0.276s      | 0.468s | 0.316s
131072 |         | 4.628s | 4.108s      | 0.758s | 0.632s      | 0.736s | 0.612s

Andrei Vagin reports fixing this performance problem is part of the work to fix CVE-2016-6213.

Cc [email protected]

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---

I think this version is very close. I had to modify __lookup_mnt_last to not skip MOUNT_UMOUNT or we would never see when the mount propagation trees intersected.

This doesn't look as good as the previous buggy version but it looks good. When the hash table isn't getting full the times look pretty linear. So it may be necessary to do some hash table resizing.

That said there remains one issue I need to think about some more.

In mark_umount_candidates I don't mark mounts that are locked to their parent and their parent is not marked as a umount candidate. Given that we skip processing mounts multiple times this might result in a mount whose parent gets marked as unmountable after the first time we see a mount not getting marked as unmountable later.

Anyway Andrei if you could check this out and see if you can see anything I missed please let me know.

Eric

 fs/namespace.c        |   6 +--
 fs/pnode.c            | 147 ++++++++++++++++++++++++++++++++++++++++++++------
 fs/pnode.h            |   4 ++
 include/linux/mount.h |   2 +
 4 files changed, 138 insertions(+), 21 deletions(-)

fs/namespace.c +0 -0

index db1b5a38864e..1ca99fa2e0f4 100644
@@ -650,13 +650,11 @@ struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry)
 	p = __lookup_mnt(mnt, dentry);

fs/namespace.c +1 -0

+		res = p;
 	}
 out:
 	return res;

fs/pnode.c +130 -17

index 234a9ac49958..025e3d9339b0 100644
@@ -390,20 +390,137 @@ void propagate_mount_unlock(struct mount *mnt)
 }
 
 /*
+ * get the next mount in the propagation tree (that has not been visited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_visit_child(struct mount *last_child,
+					    struct mount *origin_child)
+{
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
+
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_VISITED(last_child))
+		return NULL;
+
+	SET_MNT_VISITED(last_child);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		m = first_slave(m);
+		goto check_slave;
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				child = __lookup_mnt_last(&next->mnt, mountpoint);
+				if (child && !IS_MNT_VISITED(child))
+					return child;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+			check_slave:
+				child = __lookup_mnt_last(&m->mnt, mountpoint);
+				if (child && !IS_MNT_VISITED(child))
+					return child;
+			}
+		}
+
+		/* back at master */
+		m = master;
+	}
+}
+
+/*
+ * get the next mount in the propagation tree (that has not been revisited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_revisit_child(struct mount *last_child,
+					       struct mount *origin_child)
+{
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
+
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_VISITED(last_child))
+		return NULL;
+
+	CLEAR_MNT_VISITED(last_child);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		m = first_slave(m);
+		goto check_slave;
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				child = __lookup_mnt_last(&next->mnt, mountpoint);
+				if (child && IS_MNT_VISITED(child))
+					return child;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+			check_slave:
+				child = __lookup_mnt_last(&m->mnt, mountpoint);
+				if (child && IS_MNT_VISITED(child))
+					return child;
+			}
+		}
+
+		/* back at master */
+		m = master;
+	}
+}
+
+/*
  * Mark all mounts that the MNT_LOCKED logic will allow to be unmounted.
  */
 static void mark_umount_candidates(struct mount *mnt)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	BUG_ON(mnt->mnt_parent == mnt);
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
+	for (child = propagation_visit_child(mnt, mnt); child;
+	     child = propagation_visit_child(child, mnt)) {
+		if (child->mnt.mnt_flags & MNT_UMOUNT)
+			continue;
+		if (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(child->mnt_parent)) {
 			SET_MNT_MARK(child);
 		}
 	}

@@ -415,21 +532,17 @@ static void mark_umount_candidates(struct mount *mnt)
  */
 static void __propagate_umount(struct mount *mnt)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
-
-	BUG_ON(parent == mnt);
+	struct mount *child;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	BUG_ON(mnt->mnt_parent == mnt);
 
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
+	for (child = propagation_revisit_child(mnt, mnt); child;
+	     child = propagation_revisit_child(child, mnt)) {
 		/*
 		 * umount the child only if the child has no children
 		 * and the child is marked safe to unmount.
 		 */
-		if (!child || !IS_MNT_MARKED(child))
+		if (!IS_MNT_MARKED(child))
 			continue;
 		CLEAR_MNT_MARK(child);
 		if (list_empty(&child->mnt_mounts)) {

fs/pnode.h +4 -0

index 550f5a8b4fcf..988ea4945764 100644
@@ -21,6 +21,10 @@
 #define CLEAR_MNT_MARK(m) ((m)->mnt.mnt_flags &= ~MNT_MARKED)
 #define IS_MNT_LOCKED(m) ((m)->mnt.mnt_flags & MNT_LOCKED)
 
+#define IS_MNT_VISITED(m) ((m)->mnt.mnt_flags & MNT_VISITED)
+#define SET_MNT_VISITED(m) ((m)->mnt.mnt_flags |= MNT_VISITED)
+#define CLEAR_MNT_VISITED(m) ((m)->mnt.mnt_flags &= ~MNT_VISITED)
+
 #define CL_EXPIRE    		0x01
 #define CL_SLAVE     		0x02
 #define CL_COPY_UNBINDABLE	0x04

include/linux/mount.h +2 -0

index 1172cce949a4..773464f85f93 100644
@@ -52,6 +52,8 @@ struct mnt_namespace;
 
 #define MNT_INTERNAL	0x4000
 
+#define MNT_VISITED		0x010000
+
 #define MNT_LOCK_ATIME		0x040000
 #define MNT_LOCK_NOEXEC		0x080000
 #define MNT_LOCK_NOSUID		0x100000

Signature · 1 line

2.8.3

Quoted 67 lines · Eric W. Biederman

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

Solve this in the most straight forward way possible, by adding a new mount flag to mark parts of the mount propagation tree that have been visited, and use that mark to skip parts of the mount propagation tree that have already been visited during an unmount. This guarantees that each mountpoint in the possibly overlapping mount propagation trees will be visited exactly once.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate setting and clearling the visited mount mark.

The skipping of already unmounted mounts has been moved from __lookup_mnt_last to mark_umount_candidates, so that the new propagation functions can notice record when the propagation tree passes through the initial set of unmounted mounts. Except in umount_tree as part of the unmounting process the only place where unmounted mounts should be found are in unmounted subtrees. All of the other callers of __lookup_mnt_last are from mounted subtrees so the not checking for unmounted mounts should not affect them.

Here is a script to generate such mount tree:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mhash  |  8192   |  8192  |  8192       | 131072 | 131072      | 104857 | 104857
mounts | before  | after  | after (sys) | after  | after (sys) |  after | after (sys)
-------------------------------------------------------------------------------------
  1024 |  0.071s | 0.023s | 0.008s      | 0.026s | 0.000s      | 0.024s | 0.008s
  2048 |  0.184s | 0.030s | 0.012s      | 0.035s | 0.008s      | 0.030s | 0.012s
  4096 |  0.604s | 0.047s | 0.012s      | 0.042s | 0.016s      | 0.032s | 0.016s
  8912 |  4.471s | 0.085s | 0.020s      | 0.059s | 0.059s      | 0.050s | 0.036s
 16384 | 34.826s | 0.105s | 0.092s      | 0.109s | 0.060s      | 0.087s | 0.068s
 32768 |         | 0.245s | 0.168s      | 0.192s | 0.144s      | 0.167s | 0.156s
 65536 |         | 0.833s | 0.716s      | 0.485s | 0.276s      | 0.468s | 0.316s
131072 |         | 4.628s | 4.108s      | 0.758s | 0.632s      | 0.736s | 0.612s

Andrei Vagin reports fixing this performance problem is part of the work to fix CVE-2016-6213.

Cc [email protected]

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---

I think this version is very close. I had to modify __lookup_mnt_last to not skip MOUNT_UMOUNT or we would never see when the mount propagation trees intersected.

This doesn't look as good as the previous buggy version but it looks good. When the hash table isn't getting full the times look pretty linear. So it may be necessary to do some hash table resizing.

Quoted 7 lines

That said there remains one issue I need to think about some more.

In mark_umount_candidates I don't mark mounts that are locked to their parent and their parent is not marked as a umount candidate. Given that we skip processing mounts multiple times this might result in a mount whose parent gets marked as unmountable after the first time we see a mount not getting marked as unmountable later.

Quoted 2 lines

Anyway Andrei if you could check this out and see if you can see anything I missed please let me know.

I've tested this patch today and it works to me. The idea of this patch looks good for me too. Thanks! There is one inline comment.

Quoted 18 lines

 fs/namespace.c        |   6 +--
 fs/pnode.c            | 147 ++++++++++++++++++++++++++++++++++++++++++++------
 fs/pnode.h            |   4 ++
 include/linux/mount.h |   2 +
 4 files changed, 138 insertions(+), 21 deletions(-)

fs/namespace.c +1 -2

index db1b5a38864e..1ca99fa2e0f4 100644
@@ -650,13 +650,11 @@ struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry)
 	p = __lookup_mnt(mnt, dentry);
 	if (!p)
 		goto out;
-	if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-		res = p;
+	res = p;
 	hlist_for_each_entry_continue(p, mnt_hash) {

fs/namespace.c struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry) · line 659

Quoted 4 lines

fs/namespace.c +0 -2

 		if (&p->mnt_parent->mnt != mnt || p->mnt_mountpoint != dentry)
 			break;
-		if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-			res = p;

__lookup_mnt_last is used in propagate_mount_busy and attach_recursive_mnt. Should we do smth to save old behaviour of these functions.

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

While investigating the horrible performance I realized that in the case overlapping mount trees since the addition of locked mount support the code has been failing to unmount all of the mounts it should have been unmounting.

Make the walk of the mount propagation trees nearly linear by using MNT_MARK to mark pieces of the mount propagation trees that have already been visited, allowing subsequent walks to skip over subtrees.

Make the processing of mounts order independent by adding a list of mount entries that need to be unmounted, and simply adding a mount to that list when it becomes apparent the mount can safely be unmounted. For mounts that are locked on other mounts but otherwise could be unmounted move them from their parnets mnt_mounts to mnt_umounts so that if and when their parent becomes unmounted these mounts can be added to the list of mounts to unmount.

Add a final pass to clear MNT_MARK and to restore mnt_mounts from mnt_umounts for anything that did not get unmounted.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate walking of the mount tree and setting and clearing the mount mark.

The skipping of already unmounted mounts has been moved from __lookup_mnt_last to mark_umount_candidates, so that the new propagation functions can notice when the propagation tree passes through the initial set of unmounted mounts. Except in umount_tree as part of the unmounting process the only place where unmounted mounts should be found are in unmounted subtrees. All of the other callers of __lookup_mnt_last are from mounted subtrees so the not checking for unmounted mounts should not affect them.

A script to generate overlapping mount propagation trees:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mhash  |  8192   |  8192  |  8192       | 131072 | 131072      | 104857 | 104857
mounts | before  | after  | after (sys) | after  | after (sys) |  after | after (sys)
-------------------------------------------------------------------------------------
  1024 |  0.071s | 0.020s | 0.000s      | 0.022s | 0.004s      | 0.020s | 0.004s
  2048 |  0.184s | 0.022s | 0.004s      | 0.023s | 0.004s      | 0.022s | 0.008s
  4096 |  0.604s | 0.025s | 0.020s      | 0.029s | 0.008s      | 0.026s | 0.004s
  8912 |  4.471s | 0.053s | 0.020s      | 0.051s | 0.024s      | 0.047s | 0.016s
 16384 | 34.826s | 0.088s | 0.060s      | 0.081s | 0.048s      | 0.082s | 0.052s
 32768 |         | 0.216s | 0.172s      | 0.160s | 0.124s      | 0.160s | 0.096s
 65536 |         | 0.819s | 0.726s      | 0.330s | 0.260s      | 0.338s | 0.256s
131072 |         | 4.502s | 4.168s      | 0.707s | 0.580s      | 0.709s | 0.592s

Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213.

A script for a pathlogical set of mounts:

$ cat pathological.sh

mount -t tmpfs base /mnt
mount --make-shared /mnt
mkdir -p /mnt/b

mount -t tmpfs test1 /mnt/b
mount --make-shared /mnt/b
mkdir -p /mnt/b/10

mount -t tmpfs test2 /mnt/b/10
mount --make-shared /mnt/b/10
mkdir -p /mnt/b/10/20

mount --rbind /mnt/b /mnt/b/10/20

unshare -Urm sleep 2
umount -l /mnt/b
wait %%

$ unsahre -Urm pathlogical.sh

Cc [email protected]

Fixes a05964f3917c ("[PATCH] shared mounts handling: umount")

Fixes 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts")

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---

Barring some stupid mistake this looks like it fixes both the performance and the correctness issues I was able to spot earlier. Andrei if you could give this version a look over I would appreciate it.

Unless we can find a problem I am going to call this the final version.

 fs/mount.h     |   1 +
 fs/namespace.c |   7 +-
 fs/pnode.c     | 198 ++++++++++++++++++++++++++++++++++++++++++++++-----------
 fs/pnode.h     |   2 +-
 4 files changed, 165 insertions(+), 43 deletions(-)

fs/mount.h +1 -0

index d2e25d7b64b3..00fe0d1d6ba7 100644
@@ -58,6 +58,7 @@ struct mount {
 	struct mnt_namespace *mnt_ns;	/* containing namespace */
 	struct mountpoint *mnt_mp;	/* where is it mounted */
 	struct hlist_node mnt_mp_list;	/* list mounts with the same mountpoint */
+	struct list_head mnt_umounts;	/* list of children that are being unmounted */
 #ifdef CONFIG_FSNOTIFY
 	struct hlist_head mnt_fsnotify_marks;
 	__u32 mnt_fsnotify_mask;

fs/namespace.c +3 -4

index e6c234b1a645..73801391bb00 100644
@@ -237,6 +237,7 @@ static struct mount *alloc_vfsmnt(const char *name)
 		INIT_LIST_HEAD(&mnt->mnt_slave_list);
 		INIT_LIST_HEAD(&mnt->mnt_slave);
 		INIT_HLIST_NODE(&mnt->mnt_mp_list);
+		INIT_LIST_HEAD(&mnt->mnt_umounts);
 #ifdef CONFIG_FSNOTIFY
 		INIT_HLIST_HEAD(&mnt->mnt_fsnotify_marks);
 #endif

@@ -650,13 +651,11 @@ struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry)
 	p = __lookup_mnt(mnt, dentry);
 	if (!p)
 		goto out;
-	if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-		res = p;
+	res = p;
 	hlist_for_each_entry_continue(p, mnt_hash) {
 		if (&p->mnt_parent->mnt != mnt || p->mnt_mountpoint != dentry)
 			break;
-		if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-			res = p;
+		res = p;
 	}
 out:
 	return res;

fs/pnode.c +160 -38

index 234a9ac49958..15e30e861a14 100644
@@ -390,56 +390,153 @@ void propagate_mount_unlock(struct mount *mnt)
 }
 
 /*
- * Mark all mounts that the MNT_LOCKED logic will allow to be unmounted.
+ * get the next mount in the propagation tree (that has not been visited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
  */
-static void mark_umount_candidates(struct mount *mnt)
+static struct mount *propagation_visit_child(struct mount *last_child,
+					    struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
-			SET_MNT_MARK(child);
+	SET_MNT_MARK(last_child);
+
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		m = first_slave(m);
+		goto check_slave;
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				child = __lookup_mnt_last(&next->mnt, mountpoint);
+				if (child && !IS_MNT_MARKED(child))
+					return child;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+			check_slave:
+				child = __lookup_mnt_last(&m->mnt, mountpoint);
+				if (child && !IS_MNT_MARKED(child))
+					return child;
+			}
 		}
+
+		/* back at master */
+		m = master;
 	}
 }
 
 /*
- * NOTE: unmounting 'mnt' naturally propagates to all other mounts its
- * parent propagates to.
+ * get the next mount in the propagation tree (that has not been revisited)
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
  */
-static void __propagate_umount(struct mount *mnt)
+static struct mount *propagation_revisit_child(struct mount *last_child,
+					       struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	CLEAR_MNT_MARK(last_child);
 
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		/*
-		 * umount the child only if the child has no children
-		 * and the child is marked safe to unmount.
-		 */
-		if (!child || !IS_MNT_MARKED(child))
-			continue;
-		CLEAR_MNT_MARK(child);
-		if (list_empty(&child->mnt_mounts)) {
-			list_del_init(&child->mnt_child);
-			child->mnt.mnt_flags |= MNT_UMOUNT;
-			list_move_tail(&child->mnt_list, &mnt->mnt_list);
+	/* are there any slaves of this mount? */
+	if (!list_empty(&m->mnt_slave_list)) {
+		m = first_slave(m);
+		goto check_slave;
+	}
+	while (1) {
+		struct mount *master = m->mnt_master;
+
+		if (master == origin->mnt_master) {
+			struct mount *next = next_peer(m);
+			while (1) {
+				if (next == origin)
+					return NULL;
+				child = __lookup_mnt_last(&next->mnt, mountpoint);
+				if (child && IS_MNT_MARKED(child))
+					return child;
+				next = next_peer(next);
+			}
+		} else {
+			while (1) {
+				if (m->mnt_slave.next == &master->mnt_slave_list)
+					break;
+				m = next_slave(m);
+			check_slave:
+				child = __lookup_mnt_last(&m->mnt, mountpoint);
+				if (child && IS_MNT_MARKED(child))
+					return child;
+			}
 		}
+
+		/* back at master */
+		m = master;
 	}
 }
 
+static void start_umount_propagation(struct mount *child,
+				     struct list_head *to_umount)
+{
+	do {
+		struct mount *parent = child->mnt_parent;
+
+		if ((child->mnt.mnt_flags & MNT_UMOUNT) ||
+		    !list_empty(&child->mnt_mounts))
+			return;
+
+		if (!IS_MNT_LOCKED(child))
+			list_move_tail(&child->mnt_child, to_umount);
+		else
+			list_move_tail(&child->mnt_child, &parent->mnt_umounts);
+
+		child = NULL;
+		if (IS_MNT_MARKED(parent))
+			child = parent;
+	} while (child);
+}
+
+static void end_umount_propagation(struct mount *child)
+{
+	struct mount *parent = child->mnt_parent;
+
+	if (!list_empty(&parent->mnt_umounts))
+		list_splice_tail_init(&parent->mnt_umounts, &parent->mnt_mounts);
+}
+
+
 /*
  * collect all mounts that receive propagation from the mount in @list,
  * and return these additional mounts in the same list.

@@ -447,14 +544,39 @@ static void __propagate_umount(struct mount *mnt)
  *
  * vfsmount lock must be held for write
  */
-int propagate_umount(struct list_head *list)
+void propagate_umount(struct list_head *list)
 {
 	struct mount *mnt;
+	LIST_HEAD(to_umount);
+	LIST_HEAD(tmp_list);
+
+	/* Find candidates for unmounting */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_visit_child(mnt, mnt); child;
+		     child = propagation_visit_child(child, mnt))
+			start_umount_propagation(child, &to_umount);
+	}
 
-	list_for_each_entry_reverse(mnt, list, mnt_list)
-		mark_umount_candidates(mnt);
+	/* Begin unmounting */
+	while (!list_empty(&to_umount)) {
+		mnt = list_first_entry(&to_umount, struct mount, mnt_child);
 
-	list_for_each_entry(mnt, list, mnt_list)
-		__propagate_umount(mnt);
-	return 0;
+		list_del_init(&mnt->mnt_child);
+		mnt->mnt.mnt_flags |= MNT_UMOUNT;
+		list_move_tail(&mnt->mnt_list, &tmp_list);
+
+		if (!list_empty(&mnt->mnt_umounts))
+			list_splice_tail_init(&mnt->mnt_umounts, &to_umount);
+	}
+
+	/* Cleanup the mount propagation tree */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_revisit_child(mnt, mnt); child;
+		     child = propagation_revisit_child(child, mnt))
+			end_umount_propagation(child);
+	}
+
+	list_splice_tail(&tmp_list, list);
 }

fs/pnode.h +1 -1

index 550f5a8b4fcf..38c6cdb96b34 100644
@@ -41,7 +41,7 @@ static inline void set_mnt_shared(struct mount *mnt)
 void change_mnt_propagation(struct mount *, int);
 int propagate_mnt(struct mount *, struct mountpoint *, struct mount *,
 		struct hlist_head *);
-int propagate_umount(struct list_head *);
+void propagate_umount(struct list_head *);
 int propagate_mount_busy(struct mount *, int);
 void propagate_mount_unlock(struct mount *);
 void mnt_release_group_id(struct mount *);

Signature · 1 line

2.10.1

Quoted 3 lines

Barring some stupid mistake this looks like it fixes both the performance and the correctness issues I was able to spot earlier. Andrei if you could give this version a look over I would appreciate it.

Eric, could you try out this script:

It hangs up on my host with this patch.

Andrei,

This fixes the issue you have reported and through a refactoring makes the code simpler and easier to verify. That said I find your
last test case very interesting. While looking at it in detail
I have realized I don't fully understand why we have both lookup_mnt and lookup_mnt_last, so I can't say that this change is fully correct.

Outside of propogate_umount I am don't have concerns but I am not 100% convinced that my change to lookup_mnt_last does the right thing in the case of propagate_umount.

I do see why your last test case scales badly. Long chains of shared mounts that we can't skip. At the same time I don't really understand that case. Part of it has to do with multiple child mounts of the same mount on the same mountpoint.

So I am working through my concerns. In the mean time I figured it would be useful to post this version. As this version is clearly better than the version of this change that have come before it.

Eric

From: "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>
Date: Thu, 13 Oct 2016 13:27:19 -0500

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

While investigating the horrible performance I realized that in the case overlapping mount trees since the addition of locked mount support the code has been failing to unmount all of the mounts it should have been unmounting.

Make the walk of the mount propagation trees nearly linear by using MNT_MARK to mark pieces of the mount propagation trees that have already been visited, allowing subsequent walks to skip over subtrees.

Make the processing of mounts order independent by adding a list of mount entries that need to be unmounted, and simply adding a mount to that list when it becomes apparent the mount can safely be unmounted. For mounts that are locked on other mounts but otherwise could be unmounted move them from their parnets mnt_mounts to mnt_umounts so that if and when their parent becomes unmounted these mounts can be added to the list of mounts to unmount.

Add a final pass to clear MNT_MARK and to restore mnt_mounts from mnt_umounts for anything that did not get unmounted.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate walking of the mount tree and setting and clearing the mount mark.

The skipping of already unmounted mounts has been moved from __lookup_mnt_last to mark_umount_candidates, so that the new propagation functions can notice when the propagation tree passes through the initial set of unmounted mounts. Except in umount_tree as part of the unmounting process the only place where unmounted mounts should be found are in unmounted subtrees. All of the other callers of __lookup_mnt_last are from mounted subtrees so the not checking for unmounted mounts should not affect them.

A script to generate overlapping mount propagation trees:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mhash  |  8192   |  8192  |  8192       | 131072 | 131072      | 104857 | 104857
mounts | before  | after  | after (sys) | after  | after (sys) |  after | after (sys)
-------------------------------------------------------------------------------------
  1024 |  0.071s | 0.020s | 0.000s      | 0.022s | 0.004s      | 0.020s | 0.004s
  2048 |  0.184s | 0.022s | 0.004s      | 0.023s | 0.004s      | 0.022s | 0.008s
  4096 |  0.604s | 0.025s | 0.020s      | 0.029s | 0.008s      | 0.026s | 0.004s
  8912 |  4.471s | 0.053s | 0.020s      | 0.051s | 0.024s      | 0.047s | 0.016s
 16384 | 34.826s | 0.088s | 0.060s      | 0.081s | 0.048s      | 0.082s | 0.052s
 32768 |         | 0.216s | 0.172s      | 0.160s | 0.124s      | 0.160s | 0.096s
 65536 |         | 0.819s | 0.726s      | 0.330s | 0.260s      | 0.338s | 0.256s
131072 |         | 4.502s | 4.168s      | 0.707s | 0.580s      | 0.709s | 0.592s

Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213.

A script for a pathlogical set of mounts:

$ cat pathological.sh

mount -t tmpfs base /mnt
mount --make-shared /mnt
mkdir -p /mnt/b

mount -t tmpfs test1 /mnt/b
mount --make-shared /mnt/b
mkdir -p /mnt/b/10

mount -t tmpfs test2 /mnt/b/10
mount --make-shared /mnt/b/10
mkdir -p /mnt/b/10/20

mount --rbind /mnt/b /mnt/b/10/20

unshare -Urm sleep 2
umount -l /mnt/b
wait %%

$ unshare -Urm pathlogical.sh

Cc [email protected]

Fixes a05964f3917c ("[PATCH] shared mounts handling: umount")

Fixes 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts")

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---
 fs/mount.h     |   1 +
 fs/namespace.c |   7 +--
 fs/pnode.c     | 179 +++++++++++++++++++++++++++++++++++++++++----------------
 fs/pnode.h     |   2 +-
 4 files changed, 133 insertions(+), 56 deletions(-)

fs/mount.h +1 -0

index d2e25d7b64b3..00fe0d1d6ba7 100644
@@ -58,6 +58,7 @@ struct mount {
 	struct mnt_namespace *mnt_ns;	/* containing namespace */
 	struct mountpoint *mnt_mp;	/* where is it mounted */
 	struct hlist_node mnt_mp_list;	/* list mounts with the same mountpoint */
+	struct list_head mnt_umounts;	/* list of children that are being unmounted */
 #ifdef CONFIG_FSNOTIFY
 	struct hlist_head mnt_fsnotify_marks;
 	__u32 mnt_fsnotify_mask;

fs/namespace.c +3 -4

index e6c234b1a645..73801391bb00 100644
@@ -237,6 +237,7 @@ static struct mount *alloc_vfsmnt(const char *name)
 		INIT_LIST_HEAD(&mnt->mnt_slave_list);
 		INIT_LIST_HEAD(&mnt->mnt_slave);
 		INIT_HLIST_NODE(&mnt->mnt_mp_list);
+		INIT_LIST_HEAD(&mnt->mnt_umounts);
 #ifdef CONFIG_FSNOTIFY
 		INIT_HLIST_HEAD(&mnt->mnt_fsnotify_marks);
 #endif

@@ -650,13 +651,11 @@ struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry)
 	p = __lookup_mnt(mnt, dentry);
 	if (!p)
 		goto out;
-	if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-		res = p;
+	res = p;
 	hlist_for_each_entry_continue(p, mnt_hash) {
 		if (&p->mnt_parent->mnt != mnt || p->mnt_mountpoint != dentry)
 			break;
-		if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-			res = p;
+		res = p;
 	}
 out:
 	return res;

fs/pnode.c +128 -51

index 234a9ac49958..8fd1a3fb420c 100644
@@ -134,7 +134,8 @@ void change_mnt_propagation(struct mount *mnt, int type)
 }
 
 /*
- * get the next mount in the propagation tree.
+ * get the next mount that is not a slave of the current mount in the
+ * propagation tree.
  * @m: the mount seen last
  * @origin: the original mount from where the tree walk initiated
  *

@@ -143,13 +144,9 @@ void change_mnt_propagation(struct mount *mnt, int type)
  * vfsmount found while iterating with propagation_next() is
  * a peer of one we'd found earlier.
  */
-static struct mount *propagation_next(struct mount *m,
-					 struct mount *origin)
+static struct mount *propagation_next_sib(struct mount *m,
+						struct mount *origin)
 {
-	/* are there any slaves of this mount? */
-	if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
-		return first_slave(m);
-
 	while (1) {
 		struct mount *master = m->mnt_master;
 

@@ -164,6 +161,26 @@ static struct mount *propagation_next(struct mount *m,
 	}
 }
 
+/*
+ * get the next mount in the propagation tree.
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_next(struct mount *m,
+					 struct mount *origin)
+{
+	/* are there any slaves of this mount? */
+	if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
+		return first_slave(m);
+
+	return propagation_next_sib(m, origin);
+}
+
 static struct mount *next_group(struct mount *m, struct mount *origin)
 {
 	while (1) {

@@ -389,57 +406,92 @@ void propagate_mount_unlock(struct mount *mnt)
 	}
 }
 
-/*
- * Mark all mounts that the MNT_LOCKED logic will allow to be unmounted.
- */
-static void mark_umount_candidates(struct mount *mnt)
+static struct mount *propagation_visit_child(struct mount *last_child,
+					    struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
-			SET_MNT_MARK(child);
-		}
+	SET_MNT_MARK(last_child);
+
+	m = propagation_next(m, origin);
+	while (m) {
+		child = __lookup_mnt_last(&m->mnt, mountpoint);
+		if (child && !IS_MNT_MARKED(child))
+			return child;
+
+		if (!child)
+			m = propagation_next(m, origin);
+		else
+			m = propagation_next_sib(m, origin);
 	}
+	return NULL;
 }
 
-/*
- * NOTE: unmounting 'mnt' naturally propagates to all other mounts its
- * parent propagates to.
- */
-static void __propagate_umount(struct mount *mnt)
+static struct mount *propagation_revisit_child(struct mount *last_child,
+					       struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	CLEAR_MNT_MARK(last_child);
 
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		/*
-		 * umount the child only if the child has no children
-		 * and the child is marked safe to unmount.
-		 */
-		if (!child || !IS_MNT_MARKED(child))
-			continue;
-		CLEAR_MNT_MARK(child);
-		if (list_empty(&child->mnt_mounts)) {
-			list_del_init(&child->mnt_child);
-			child->mnt.mnt_flags |= MNT_UMOUNT;
-			list_move_tail(&child->mnt_list, &mnt->mnt_list);
-		}
+	m = propagation_next(m, origin);
+	while (m) {
+		child = __lookup_mnt_last(&m->mnt, mountpoint);
+		if (child && IS_MNT_MARKED(child))
+			return child;
+
+		if (!child)
+			m = propagation_next(m, origin);
+		else
+			m = propagation_next_sib(m, origin);
 	}
+	return NULL;
 }
 
+static void start_umount_propagation(struct mount *child,
+				     struct list_head *to_umount)
+{
+	do {
+		struct mount *parent = child->mnt_parent;
+
+		if ((child->mnt.mnt_flags & MNT_UMOUNT) ||
+		    !list_empty(&child->mnt_mounts))
+			return;
+
+		if (!IS_MNT_LOCKED(child))
+			list_move_tail(&child->mnt_child, to_umount);
+		else
+			list_move_tail(&child->mnt_child, &parent->mnt_umounts);
+
+		child = NULL;
+		if (IS_MNT_MARKED(parent))
+			child = parent;
+	} while (child);
+}
+
+static void end_umount_propagation(struct mount *child)
+{
+	struct mount *parent = child->mnt_parent;
+
+	if (!list_empty(&parent->mnt_umounts))
+		list_splice_tail_init(&parent->mnt_umounts, &parent->mnt_mounts);
+}
+
+
 /*
  * collect all mounts that receive propagation from the mount in @list,
  * and return these additional mounts in the same list.

@@ -447,14 +499,39 @@ static void __propagate_umount(struct mount *mnt)
  *
  * vfsmount lock must be held for write
  */
-int propagate_umount(struct list_head *list)
+void propagate_umount(struct list_head *list)
 {
 	struct mount *mnt;
+	LIST_HEAD(to_umount);
+	LIST_HEAD(tmp_list);
+
+	/* Find candidates for unmounting */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_visit_child(mnt, mnt); child;
+		     child = propagation_visit_child(child, mnt))
+			start_umount_propagation(child, &to_umount);
+	}
 
-	list_for_each_entry_reverse(mnt, list, mnt_list)
-		mark_umount_candidates(mnt);
+	/* Begin unmounting */
+	while (!list_empty(&to_umount)) {
+		mnt = list_first_entry(&to_umount, struct mount, mnt_child);
 
-	list_for_each_entry(mnt, list, mnt_list)
-		__propagate_umount(mnt);
-	return 0;
+		list_del_init(&mnt->mnt_child);
+		mnt->mnt.mnt_flags |= MNT_UMOUNT;
+		list_move_tail(&mnt->mnt_list, &tmp_list);
+
+		if (!list_empty(&mnt->mnt_umounts))
+			list_splice_tail_init(&mnt->mnt_umounts, &to_umount);
+	}
+
+	/* Cleanup the mount propagation tree */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_revisit_child(mnt, mnt); child;
+		     child = propagation_revisit_child(child, mnt))
+			end_umount_propagation(child);
+	}
+
+	list_splice_tail(&tmp_list, list);
 }

fs/pnode.h +1 -1

index 550f5a8b4fcf..38c6cdb96b34 100644
@@ -41,7 +41,7 @@ static inline void set_mnt_shared(struct mount *mnt)
 void change_mnt_propagation(struct mount *, int);
 int propagate_mnt(struct mount *, struct mountpoint *, struct mount *,
 		struct hlist_head *);
-int propagate_umount(struct list_head *);
+void propagate_umount(struct list_head *);
 int propagate_mount_busy(struct mount *, int);
 void propagate_mount_unlock(struct mount *);
 void mnt_release_group_id(struct mount *);

Signature · 1 line

2.10.1

Quoted 13 lines · Eric W. Biederman

Andrei,

This fixes the issue you have reported and through a refactoring makes the code simpler and easier to verify. That said I find your
last test case very interesting. While looking at it in detail
I have realized I don't fully understand why we have both lookup_mnt and lookup_mnt_last, so I can't say that this change is fully correct.

Outside of propogate_umount I am don't have concerns but I am not 100% convinced that my change to lookup_mnt_last does the right thing in the case of propagate_umount.

Quoted 8 lines

I do see why your last test case scales badly. Long chains of shared mounts that we can't skip. At the same time I don't really understand that case. Part of it has to do with multiple child mounts of the same mount on the same mountpoint.

So I am working through my concerns. In the mean time I figured it would be useful to post this version. As this version is clearly better than the version of this change that have come before it.

Hi Eric,

I have tested this version and it works fine.

As for the the last test case, could you look at the attached patch? The idea is that we can skip all mounts from a shared group, if one of them already marked.

Quoted 400 lines

Eric

From: "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>
Date: Thu, 13 Oct 2016 13:27:19 -0500

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

While investigating the horrible performance I realized that in the case overlapping mount trees since the addition of locked mount support the code has been failing to unmount all of the mounts it should have been unmounting.

Make the walk of the mount propagation trees nearly linear by using MNT_MARK to mark pieces of the mount propagation trees that have already been visited, allowing subsequent walks to skip over subtrees.

Make the processing of mounts order independent by adding a list of mount entries that need to be unmounted, and simply adding a mount to that list when it becomes apparent the mount can safely be unmounted. For mounts that are locked on other mounts but otherwise could be unmounted move them from their parnets mnt_mounts to mnt_umounts so that if and when their parent becomes unmounted these mounts can be added to the list of mounts to unmount.

Add a final pass to clear MNT_MARK and to restore mnt_mounts from mnt_umounts for anything that did not get unmounted.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate walking of the mount tree and setting and clearing the mount mark.

The skipping of already unmounted mounts has been moved from __lookup_mnt_last to mark_umount_candidates, so that the new propagation functions can notice when the propagation tree passes through the initial set of unmounted mounts. Except in umount_tree as part of the unmounting process the only place where unmounted mounts should be found are in unmounted subtrees. All of the other callers of __lookup_mnt_last are from mounted subtrees so the not checking for unmounted mounts should not affect them.

A script to generate overlapping mount propagation trees:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mhash  |  8192   |  8192  |  8192       | 131072 | 131072      | 104857 | 104857
mounts | before  | after  | after (sys) | after  | after (sys) |  after | after (sys)
-------------------------------------------------------------------------------------
  1024 |  0.071s | 0.020s | 0.000s      | 0.022s | 0.004s      | 0.020s | 0.004s
  2048 |  0.184s | 0.022s | 0.004s      | 0.023s | 0.004s      | 0.022s | 0.008s
  4096 |  0.604s | 0.025s | 0.020s      | 0.029s | 0.008s      | 0.026s | 0.004s
  8912 |  4.471s | 0.053s | 0.020s      | 0.051s | 0.024s      | 0.047s | 0.016s
 16384 | 34.826s | 0.088s | 0.060s      | 0.081s | 0.048s      | 0.082s | 0.052s
 32768 |         | 0.216s | 0.172s      | 0.160s | 0.124s      | 0.160s | 0.096s
 65536 |         | 0.819s | 0.726s      | 0.330s | 0.260s      | 0.338s | 0.256s
131072 |         | 4.502s | 4.168s      | 0.707s | 0.580s      | 0.709s | 0.592s

Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213.

A script for a pathlogical set of mounts:

$ cat pathological.sh

mount -t tmpfs base /mnt
mount --make-shared /mnt
mkdir -p /mnt/b

mount -t tmpfs test1 /mnt/b
mount --make-shared /mnt/b
mkdir -p /mnt/b/10

mount -t tmpfs test2 /mnt/b/10
mount --make-shared /mnt/b/10
mkdir -p /mnt/b/10/20

mount --rbind /mnt/b /mnt/b/10/20

unshare -Urm sleep 2
umount -l /mnt/b
wait %%

$ unshare -Urm pathlogical.sh

Cc [email protected]

Fixes a05964f3917c ("[PATCH] shared mounts handling: umount")

Fixes 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts")

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---
 fs/mount.h     |   1 +
 fs/namespace.c |   7 +--
 fs/pnode.c     | 179 +++++++++++++++++++++++++++++++++++++++++----------------
 fs/pnode.h     |   2 +-
 4 files changed, 133 insertions(+), 56 deletions(-)

fs/mount.h +1 -0

index d2e25d7b64b3..00fe0d1d6ba7 100644
@@ -58,6 +58,7 @@ struct mount {
 	struct mnt_namespace *mnt_ns;	/* containing namespace */
 	struct mountpoint *mnt_mp;	/* where is it mounted */
 	struct hlist_node mnt_mp_list;	/* list mounts with the same mountpoint */
+	struct list_head mnt_umounts;	/* list of children that are being unmounted */
 #ifdef CONFIG_FSNOTIFY
 	struct hlist_head mnt_fsnotify_marks;
 	__u32 mnt_fsnotify_mask;

fs/namespace.c +3 -4

index e6c234b1a645..73801391bb00 100644
@@ -237,6 +237,7 @@ static struct mount *alloc_vfsmnt(const char *name)
 		INIT_LIST_HEAD(&mnt->mnt_slave_list);
 		INIT_LIST_HEAD(&mnt->mnt_slave);
 		INIT_HLIST_NODE(&mnt->mnt_mp_list);
+		INIT_LIST_HEAD(&mnt->mnt_umounts);
 #ifdef CONFIG_FSNOTIFY
 		INIT_HLIST_HEAD(&mnt->mnt_fsnotify_marks);
 #endif

@@ -650,13 +651,11 @@ struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry)
 	p = __lookup_mnt(mnt, dentry);
 	if (!p)
 		goto out;
-	if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-		res = p;
+	res = p;
 	hlist_for_each_entry_continue(p, mnt_hash) {
 		if (&p->mnt_parent->mnt != mnt || p->mnt_mountpoint != dentry)
 			break;
-		if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-			res = p;
+		res = p;
 	}
 out:
 	return res;

fs/pnode.c +128 -51

index 234a9ac49958..8fd1a3fb420c 100644
@@ -134,7 +134,8 @@ void change_mnt_propagation(struct mount *mnt, int type)
 }
 
 /*
- * get the next mount in the propagation tree.
+ * get the next mount that is not a slave of the current mount in the
+ * propagation tree.
  * @m: the mount seen last
  * @origin: the original mount from where the tree walk initiated
  *

@@ -143,13 +144,9 @@ void change_mnt_propagation(struct mount *mnt, int type)
  * vfsmount found while iterating with propagation_next() is
  * a peer of one we'd found earlier.
  */
-static struct mount *propagation_next(struct mount *m,
-					 struct mount *origin)
+static struct mount *propagation_next_sib(struct mount *m,
+						struct mount *origin)
 {
-	/* are there any slaves of this mount? */
-	if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
-		return first_slave(m);
-
 	while (1) {
 		struct mount *master = m->mnt_master;
 

@@ -164,6 +161,26 @@ static struct mount *propagation_next(struct mount *m,
 	}
 }
 
+/*
+ * get the next mount in the propagation tree.
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_next(struct mount *m,
+					 struct mount *origin)
+{
+	/* are there any slaves of this mount? */
+	if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
+		return first_slave(m);
+
+	return propagation_next_sib(m, origin);
+}
+
 static struct mount *next_group(struct mount *m, struct mount *origin)
 {
 	while (1) {

@@ -389,57 +406,92 @@ void propagate_mount_unlock(struct mount *mnt)
 	}
 }
 
-/*
- * Mark all mounts that the MNT_LOCKED logic will allow to be unmounted.
- */
-static void mark_umount_candidates(struct mount *mnt)
+static struct mount *propagation_visit_child(struct mount *last_child,
+					    struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
-			SET_MNT_MARK(child);
-		}
+	SET_MNT_MARK(last_child);
+
+	m = propagation_next(m, origin);
+	while (m) {
+		child = __lookup_mnt_last(&m->mnt, mountpoint);
+		if (child && !IS_MNT_MARKED(child))
+			return child;
+
+		if (!child)
+			m = propagation_next(m, origin);
+		else
+			m = propagation_next_sib(m, origin);
 	}
+	return NULL;
 }
 
-/*
- * NOTE: unmounting 'mnt' naturally propagates to all other mounts its
- * parent propagates to.
- */
-static void __propagate_umount(struct mount *mnt)
+static struct mount *propagation_revisit_child(struct mount *last_child,
+					       struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	CLEAR_MNT_MARK(last_child);
 
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		/*
-		 * umount the child only if the child has no children
-		 * and the child is marked safe to unmount.
-		 */
-		if (!child || !IS_MNT_MARKED(child))
-			continue;
-		CLEAR_MNT_MARK(child);
-		if (list_empty(&child->mnt_mounts)) {
-			list_del_init(&child->mnt_child);
-			child->mnt.mnt_flags |= MNT_UMOUNT;
-			list_move_tail(&child->mnt_list, &mnt->mnt_list);
-		}
+	m = propagation_next(m, origin);
+	while (m) {
+		child = __lookup_mnt_last(&m->mnt, mountpoint);
+		if (child && IS_MNT_MARKED(child))
+			return child;
+
+		if (!child)
+			m = propagation_next(m, origin);
+		else
+			m = propagation_next_sib(m, origin);
 	}
+	return NULL;
 }
 
+static void start_umount_propagation(struct mount *child,
+				     struct list_head *to_umount)
+{
+	do {
+		struct mount *parent = child->mnt_parent;
+
+		if ((child->mnt.mnt_flags & MNT_UMOUNT) ||
+		    !list_empty(&child->mnt_mounts))
+			return;
+
+		if (!IS_MNT_LOCKED(child))
+			list_move_tail(&child->mnt_child, to_umount);
+		else
+			list_move_tail(&child->mnt_child, &parent->mnt_umounts);
+
+		child = NULL;
+		if (IS_MNT_MARKED(parent))
+			child = parent;
+	} while (child);
+}
+
+static void end_umount_propagation(struct mount *child)
+{
+	struct mount *parent = child->mnt_parent;
+
+	if (!list_empty(&parent->mnt_umounts))
+		list_splice_tail_init(&parent->mnt_umounts, &parent->mnt_mounts);
+}
+
+
 /*
  * collect all mounts that receive propagation from the mount in @list,
  * and return these additional mounts in the same list.

@@ -447,14 +499,39 @@ static void __propagate_umount(struct mount *mnt)
  *
  * vfsmount lock must be held for write
  */
-int propagate_umount(struct list_head *list)
+void propagate_umount(struct list_head *list)
 {
 	struct mount *mnt;
+	LIST_HEAD(to_umount);
+	LIST_HEAD(tmp_list);
+
+	/* Find candidates for unmounting */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_visit_child(mnt, mnt); child;
+		     child = propagation_visit_child(child, mnt))
+			start_umount_propagation(child, &to_umount);
+	}
 
-	list_for_each_entry_reverse(mnt, list, mnt_list)
-		mark_umount_candidates(mnt);
+	/* Begin unmounting */
+	while (!list_empty(&to_umount)) {
+		mnt = list_first_entry(&to_umount, struct mount, mnt_child);
 
-	list_for_each_entry(mnt, list, mnt_list)
-		__propagate_umount(mnt);
-	return 0;
+		list_del_init(&mnt->mnt_child);
+		mnt->mnt.mnt_flags |= MNT_UMOUNT;
+		list_move_tail(&mnt->mnt_list, &tmp_list);
+
+		if (!list_empty(&mnt->mnt_umounts))
+			list_splice_tail_init(&mnt->mnt_umounts, &to_umount);
+	}
+
+	/* Cleanup the mount propagation tree */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_revisit_child(mnt, mnt); child;
+		     child = propagation_revisit_child(child, mnt))
+			end_umount_propagation(child);
+	}
+
+	list_splice_tail(&tmp_list, list);
 }

fs/pnode.h +1 -1

index 550f5a8b4fcf..38c6cdb96b34 100644
@@ -41,7 +41,7 @@ static inline void set_mnt_shared(struct mount *mnt)
 void change_mnt_propagation(struct mount *, int);
 int propagate_mnt(struct mount *, struct mountpoint *, struct mount *,
 		struct hlist_head *);
-int propagate_umount(struct list_head *);
+void propagate_umount(struct list_head *);
 int propagate_mount_busy(struct mount *, int);
 void propagate_mount_unlock(struct mount *);
 void mnt_release_group_id(struct mount *);

Signature · 1 line

2.10.1

From 8e0f45c0272aa1f789d1657a0acc98c58919dcc3 Mon Sep 17 00:00:00 2001
From: Andrei Vagin <[email protected]>
Date: Tue, 25 Oct 2016 13:57:31 -0700
Subject: [PATCH] mount: skip all mounts from a shared group if one is marked

If we meet a marked mount, it means that all mounts from its group have been already revised.

Signed-off-by Andrei Vagin <[email protected]>

---
 fs/pnode.c | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

fs/pnode.c +9 -3

index 8fd1a3f..ebb7134 100644
@@ -426,10 +426,16 @@ static struct mount *propagation_visit_child(struct mount *last_child,
 		if (child && !IS_MNT_MARKED(child))
 			return child;
 
-		if (!child)
+		if (!child) {
 			m = propagation_next(m, origin);
-		else
+		} else {
+			if (IS_MNT_MARKED(child)) {
+				if (m->mnt_group_id == origin->mnt_group_id)
+					return NULL;
+				m = m->mnt_master;
+			}
 			m = propagation_next_sib(m, origin);
+		}
 	}
 	return NULL;
 }

@@ -456,8 +462,14 @@ static struct mount *propagation_revisit_child(struct mount *last_child,
 
 		if (!child)
 			m = propagation_next(m, origin);
-		else
+		else {

Quoted 480 lines · Andrei Vagin

Quoted 13 lines · Eric W. Biederman

Andrei,

This fixes the issue you have reported and through a refactoring makes the code simpler and easier to verify. That said I find your
last test case very interesting. While looking at it in detail
I have realized I don't fully understand why we have both lookup_mnt and lookup_mnt_last, so I can't say that this change is fully correct.

Outside of propogate_umount I am don't have concerns but I am not 100% convinced that my change to lookup_mnt_last does the right thing in the case of propagate_umount.

Quoted 8 lines

I do see why your last test case scales badly. Long chains of shared mounts that we can't skip. At the same time I don't really understand that case. Part of it has to do with multiple child mounts of the same mount on the same mountpoint.

So I am working through my concerns. In the mean time I figured it would be useful to post this version. As this version is clearly better than the version of this change that have come before it.

Hi Eric,

I have tested this version and it works fine.

As for the the last test case, could you look at the attached patch? The idea is that we can skip all mounts from a shared group, if one of them already marked.

Quoted 400 lines

Eric

From: "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>
Date: Thu, 13 Oct 2016 13:27:19 -0500

Adrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap.

While investigating the horrible performance I realized that in the case overlapping mount trees since the addition of locked mount support the code has been failing to unmount all of the mounts it should have been unmounting.

Make the walk of the mount propagation trees nearly linear by using MNT_MARK to mark pieces of the mount propagation trees that have already been visited, allowing subsequent walks to skip over subtrees.

Make the processing of mounts order independent by adding a list of mount entries that need to be unmounted, and simply adding a mount to that list when it becomes apparent the mount can safely be unmounted. For mounts that are locked on other mounts but otherwise could be unmounted move them from their parnets mnt_mounts to mnt_umounts so that if and when their parent becomes unmounted these mounts can be added to the list of mounts to unmount.

Add a final pass to clear MNT_MARK and to restore mnt_mounts from mnt_umounts for anything that did not get unmounted.

Add the functions propagation_visit_next and propagation_revisit_next to coordinate walking of the mount tree and setting and clearing the mount mark.

The skipping of already unmounted mounts has been moved from __lookup_mnt_last to mark_umount_candidates, so that the new propagation functions can notice when the propagation tree passes through the initial set of unmounted mounts. Except in umount_tree as part of the unmounting process the only place where unmounted mounts should be found are in unmounted subtrees. All of the other callers of __lookup_mnt_last are from mounted subtrees so the not checking for unmounted mounts should not affect them.

A script to generate overlapping mount propagation trees:

$ cat run.sh
mount -t tmpfs test-mount /mnt
mount --make-shared /mnt
for i in `seq $1`; do
        mkdir /mnt/test.$i
        mount --bind /mnt /mnt/test.$i
done
cat /proc/mounts | grep test-mount | wc -l
time umount -l /mnt
$ for i in `seq 10 16`; do echo $i; unshare -Urm bash ./run.sh $i; done

Here are the performance numbers with and without the patch:

mhash  |  8192   |  8192  |  8192       | 131072 | 131072      | 104857 | 104857
mounts | before  | after  | after (sys) | after  | after (sys) |  after | after (sys)
-------------------------------------------------------------------------------------
  1024 |  0.071s | 0.020s | 0.000s      | 0.022s | 0.004s      | 0.020s | 0.004s
  2048 |  0.184s | 0.022s | 0.004s      | 0.023s | 0.004s      | 0.022s | 0.008s
  4096 |  0.604s | 0.025s | 0.020s      | 0.029s | 0.008s      | 0.026s | 0.004s
  8912 |  4.471s | 0.053s | 0.020s      | 0.051s | 0.024s      | 0.047s | 0.016s
 16384 | 34.826s | 0.088s | 0.060s      | 0.081s | 0.048s      | 0.082s | 0.052s
 32768 |         | 0.216s | 0.172s      | 0.160s | 0.124s      | 0.160s | 0.096s
 65536 |         | 0.819s | 0.726s      | 0.330s | 0.260s      | 0.338s | 0.256s
131072 |         | 4.502s | 4.168s      | 0.707s | 0.580s      | 0.709s | 0.592s

Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213.

A script for a pathlogical set of mounts:

$ cat pathological.sh

mount -t tmpfs base /mnt
mount --make-shared /mnt
mkdir -p /mnt/b

mount -t tmpfs test1 /mnt/b
mount --make-shared /mnt/b
mkdir -p /mnt/b/10

mount -t tmpfs test2 /mnt/b/10
mount --make-shared /mnt/b/10
mkdir -p /mnt/b/10/20

mount --rbind /mnt/b /mnt/b/10/20

unshare -Urm sleep 2
umount -l /mnt/b
wait %%

$ unshare -Urm pathlogical.sh

Cc [email protected]

Fixes a05964f3917c ("[PATCH] shared mounts handling: umount")

Fixes 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts")

Reported-by Andrei Vagin <[email protected]>

Signed-off-by "Eric W. Biederman" <ebiederm-aS9lmoZGLiVWk0Htik3J/[email protected]>

---
 fs/mount.h     |   1 +
 fs/namespace.c |   7 +--
 fs/pnode.c     | 179 +++++++++++++++++++++++++++++++++++++++++----------------
 fs/pnode.h     |   2 +-
 4 files changed, 133 insertions(+), 56 deletions(-)

fs/mount.h +1 -0

index d2e25d7b64b3..00fe0d1d6ba7 100644
@@ -58,6 +58,7 @@ struct mount {
 	struct mnt_namespace *mnt_ns;	/* containing namespace */
 	struct mountpoint *mnt_mp;	/* where is it mounted */
 	struct hlist_node mnt_mp_list;	/* list mounts with the same mountpoint */
+	struct list_head mnt_umounts;	/* list of children that are being unmounted */
 #ifdef CONFIG_FSNOTIFY
 	struct hlist_head mnt_fsnotify_marks;
 	__u32 mnt_fsnotify_mask;

fs/namespace.c +3 -4

index e6c234b1a645..73801391bb00 100644
@@ -237,6 +237,7 @@ static struct mount *alloc_vfsmnt(const char *name)
 		INIT_LIST_HEAD(&mnt->mnt_slave_list);
 		INIT_LIST_HEAD(&mnt->mnt_slave);
 		INIT_HLIST_NODE(&mnt->mnt_mp_list);
+		INIT_LIST_HEAD(&mnt->mnt_umounts);
 #ifdef CONFIG_FSNOTIFY
 		INIT_HLIST_HEAD(&mnt->mnt_fsnotify_marks);
 #endif

@@ -650,13 +651,11 @@ struct mount *__lookup_mnt_last(struct vfsmount *mnt, struct dentry *dentry)
 	p = __lookup_mnt(mnt, dentry);
 	if (!p)
 		goto out;
-	if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-		res = p;
+	res = p;
 	hlist_for_each_entry_continue(p, mnt_hash) {
 		if (&p->mnt_parent->mnt != mnt || p->mnt_mountpoint != dentry)
 			break;
-		if (!(p->mnt.mnt_flags & MNT_UMOUNT))
-			res = p;
+		res = p;
 	}
 out:
 	return res;

fs/pnode.c +128 -51

index 234a9ac49958..8fd1a3fb420c 100644
@@ -134,7 +134,8 @@ void change_mnt_propagation(struct mount *mnt, int type)
 }
 
 /*
- * get the next mount in the propagation tree.
+ * get the next mount that is not a slave of the current mount in the
+ * propagation tree.
  * @m: the mount seen last
  * @origin: the original mount from where the tree walk initiated
  *

@@ -143,13 +144,9 @@ void change_mnt_propagation(struct mount *mnt, int type)
  * vfsmount found while iterating with propagation_next() is
  * a peer of one we'd found earlier.
  */
-static struct mount *propagation_next(struct mount *m,
-					 struct mount *origin)
+static struct mount *propagation_next_sib(struct mount *m,
+						struct mount *origin)
 {
-	/* are there any slaves of this mount? */
-	if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
-		return first_slave(m);
-
 	while (1) {
 		struct mount *master = m->mnt_master;
 

@@ -164,6 +161,26 @@ static struct mount *propagation_next(struct mount *m,
 	}
 }
 
+/*
+ * get the next mount in the propagation tree.
+ * @m: the mount seen last
+ * @origin: the original mount from where the tree walk initiated
+ *
+ * Note that peer groups form contiguous segments of slave lists.
+ * We rely on that in get_source() to be able to find out if
+ * vfsmount found while iterating with propagation_next() is
+ * a peer of one we'd found earlier.
+ */
+static struct mount *propagation_next(struct mount *m,
+					 struct mount *origin)
+{
+	/* are there any slaves of this mount? */
+	if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
+		return first_slave(m);
+
+	return propagation_next_sib(m, origin);
+}
+
 static struct mount *next_group(struct mount *m, struct mount *origin)
 {
 	while (1) {

@@ -389,57 +406,92 @@ void propagate_mount_unlock(struct mount *mnt)
 	}
 }
 
-/*
- * Mark all mounts that the MNT_LOCKED logic will allow to be unmounted.
- */
-static void mark_umount_candidates(struct mount *mnt)
+static struct mount *propagation_visit_child(struct mount *last_child,
+					    struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been visited? */
+	if (IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		if (child && (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m))) {
-			SET_MNT_MARK(child);
-		}
+	SET_MNT_MARK(last_child);
+
+	m = propagation_next(m, origin);
+	while (m) {
+		child = __lookup_mnt_last(&m->mnt, mountpoint);
+		if (child && !IS_MNT_MARKED(child))
+			return child;
+
+		if (!child)
+			m = propagation_next(m, origin);
+		else
+			m = propagation_next_sib(m, origin);
 	}
+	return NULL;
 }
 
-/*
- * NOTE: unmounting 'mnt' naturally propagates to all other mounts its
- * parent propagates to.
- */
-static void __propagate_umount(struct mount *mnt)
+static struct mount *propagation_revisit_child(struct mount *last_child,
+					       struct mount *origin_child)
 {
-	struct mount *parent = mnt->mnt_parent;
-	struct mount *m;
+	struct mount *m = last_child->mnt_parent;
+	struct mount *origin = origin_child->mnt_parent;
+	struct dentry *mountpoint = origin_child->mnt_mountpoint;
+	struct mount *child;
 
-	BUG_ON(parent == mnt);
+	/* Has this part of the propgation tree already been revisited? */
+	if (!IS_MNT_MARKED(last_child))
+		return NULL;
 
-	for (m = propagation_next(parent, parent); m;
-			m = propagation_next(m, parent)) {
+	CLEAR_MNT_MARK(last_child);
 
-		struct mount *child = __lookup_mnt_last(&m->mnt,
-						mnt->mnt_mountpoint);
-		/*
-		 * umount the child only if the child has no children
-		 * and the child is marked safe to unmount.
-		 */
-		if (!child || !IS_MNT_MARKED(child))
-			continue;
-		CLEAR_MNT_MARK(child);
-		if (list_empty(&child->mnt_mounts)) {
-			list_del_init(&child->mnt_child);
-			child->mnt.mnt_flags |= MNT_UMOUNT;
-			list_move_tail(&child->mnt_list, &mnt->mnt_list);
-		}
+	m = propagation_next(m, origin);
+	while (m) {
+		child = __lookup_mnt_last(&m->mnt, mountpoint);
+		if (child && IS_MNT_MARKED(child))
+			return child;
+
+		if (!child)
+			m = propagation_next(m, origin);
+		else
+			m = propagation_next_sib(m, origin);
 	}
+	return NULL;
 }
 
+static void start_umount_propagation(struct mount *child,
+				     struct list_head *to_umount)
+{
+	do {
+		struct mount *parent = child->mnt_parent;
+
+		if ((child->mnt.mnt_flags & MNT_UMOUNT) ||
+		    !list_empty(&child->mnt_mounts))
+			return;
+
+		if (!IS_MNT_LOCKED(child))
+			list_move_tail(&child->mnt_child, to_umount);
+		else
+			list_move_tail(&child->mnt_child, &parent->mnt_umounts);
+
+		child = NULL;
+		if (IS_MNT_MARKED(parent))
+			child = parent;
+	} while (child);
+}
+
+static void end_umount_propagation(struct mount *child)
+{
+	struct mount *parent = child->mnt_parent;
+
+	if (!list_empty(&parent->mnt_umounts))
+		list_splice_tail_init(&parent->mnt_umounts, &parent->mnt_mounts);
+}
+
+
 /*
  * collect all mounts that receive propagation from the mount in @list,
  * and return these additional mounts in the same list.

@@ -447,14 +499,39 @@ static void __propagate_umount(struct mount *mnt)
  *
  * vfsmount lock must be held for write
  */
-int propagate_umount(struct list_head *list)
+void propagate_umount(struct list_head *list)
 {
 	struct mount *mnt;
+	LIST_HEAD(to_umount);
+	LIST_HEAD(tmp_list);
+
+	/* Find candidates for unmounting */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_visit_child(mnt, mnt); child;
+		     child = propagation_visit_child(child, mnt))
+			start_umount_propagation(child, &to_umount);
+	}
 
-	list_for_each_entry_reverse(mnt, list, mnt_list)
-		mark_umount_candidates(mnt);
+	/* Begin unmounting */
+	while (!list_empty(&to_umount)) {
+		mnt = list_first_entry(&to_umount, struct mount, mnt_child);
 
-	list_for_each_entry(mnt, list, mnt_list)
-		__propagate_umount(mnt);
-	return 0;
+		list_del_init(&mnt->mnt_child);
+		mnt->mnt.mnt_flags |= MNT_UMOUNT;
+		list_move_tail(&mnt->mnt_list, &tmp_list);
+
+		if (!list_empty(&mnt->mnt_umounts))
+			list_splice_tail_init(&mnt->mnt_umounts, &to_umount);
+	}
+
+	/* Cleanup the mount propagation tree */
+	list_for_each_entry(mnt, list, mnt_list) {
+		struct mount *child;
+		for (child = propagation_revisit_child(mnt, mnt); child;
+		     child = propagation_revisit_child(child, mnt))
+			end_umount_propagation(child);
+	}
+
+	list_splice_tail(&tmp_list, list);
 }

fs/pnode.h +1 -1

index 550f5a8b4fcf..38c6cdb96b34 100644
@@ -41,7 +41,7 @@ static inline void set_mnt_shared(struct mount *mnt)
 void change_mnt_propagation(struct mount *, int);
 int propagate_mnt(struct mount *, struct mountpoint *, struct mount *,
 		struct hlist_head *);
-int propagate_umount(struct list_head *);
+void propagate_umount(struct list_head *);
 int propagate_mount_busy(struct mount *, int);
 void propagate_mount_unlock(struct mount *);
 void mnt_release_group_id(struct mount *);

Signature · 1 line

2.10.1

From 8e0f45c0272aa1f789d1657a0acc98c58919dcc3 Mon Sep 17 00:00:00 2001
From: Andrei Vagin <[email protected]>
Date: Tue, 25 Oct 2016 13:57:31 -0700
Subject: [PATCH] mount: skip all mounts from a shared group if one is marked

If we meet a marked mount, it means that all mounts from its group have been already revised.

Signed-off-by Andrei Vagin <[email protected]>

---
 fs/pnode.c | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

fs/pnode.c +14 -3

index 8fd1a3f..ebb7134 100644
@@ -426,10 +426,16 @@ static struct mount *propagation_visit_child(struct mount *last_child,
 		if (child && !IS_MNT_MARKED(child))
 			return child;
 
-		if (!child)
+		if (!child) {
 			m = propagation_next(m, origin);
-		else
+		} else {
+			if (IS_MNT_MARKED(child)) {
+				if (m->mnt_group_id == origin->mnt_group_id)
+					return NULL;
+				m = m->mnt_master;
+			}
 			m = propagation_next_sib(m, origin);
+		}
 	}
 	return NULL;
 }

@@ -456,8 +462,14 @@ static struct mount *propagation_revisit_child(struct mount *last_child,
 
 		if (!child)
 			m = propagation_next(m, origin);
-		else
+		else {
+			if (!IS_MNT_MARKED(child)) {
+				if (m->mnt_group_id == origin->mnt_group_id)
+					return NULL;
+				m = m->mnt_master;
+			}
 			m = propagation_next_sib(m, origin);

fs/pnode.c static struct mount *propagation_revisit_child(struct mount *last_child, · line 475

Quoted 4 lines

fs/pnode.c +1 -0

+		}
 	}
 	return NULL;
 }

That is certainly interesting. The problem is that the reason we were going slow is that there were in fact mounts that had not been traversed in the share group.

And in fact the entire idea of visiting a vfsmount mountpoint pair exactly once is wrong in the face of shadow mounts. For a vfsmount mountpoint pair that has shadow mounts the number of shadow mounts needs to be descreased by one each time the propgation tree is traversed during unmount. Which means that as far as I can see we have to kill shadow mounts to correctly optimize this code. Once shadow mounts are gone I don't know of a case where need your optimization.

I am busily verifying my patch to kill shadow mounts but the following patch is the minimal version. As far as I can see propagate_one is the only place we create shadow mounts, and holding the namespace_lock over attach_recursive_mnt, propagate_mnt, and propgate_one is sufficient for that __lookup_mnt to be competely safe.

fs/pnode.c +3 -0

index 234a9ac49958..b14119b370d4 100644
@@ -217,6 +217,9 @@ static int propagate_one(struct mount *m)
        /* skip if mountpoint isn't covered by it */
        if (!is_subdir(mp->m_dentry, m->mnt.mnt_root))
                return 0;
+       /* skip if mountpoint already has a mount on it */
+       if (__lookup_mnt(&m->mnt, mp->m_dentry))
+               return 0;
        if (peers(m, last_dest)) {
                type = CL_MAKE_SHARED;
        } else {

If you run with that patch you will see that there are go faster stripes.

Eric

Quoted 3 lines · Eric W. Biederman

That is certainly interesting. The problem is that the reason we were going slow is that there were in fact mounts that had not been traversed in the share group.

You are right.

Quoted 8 lines

And in fact the entire idea of visiting a vfsmount mountpoint pair exactly once is wrong in the face of shadow mounts. For a vfsmount mountpoint pair that has shadow mounts the number of shadow mounts needs to be descreased by one each time the propgation tree is traversed during unmount. Which means that as far as I can see we have to kill shadow mounts to correctly optimize this code. Once shadow mounts are gone I don't know of a case where need your optimization.

Without shadow mounts, it will be hard to save predictable behaviour for cases like this:

$ unshare --propagation private -m sh test.sh
+ mount -t tmpfs --make-shared zzzz A
+ mkdir A/a
+ mount -t tmpfs zzzz A/a
+ mount --bind A B
+ mount -t tmpfs zzzz B/a
+ grep zzzz
+ cat /proc/self/mountinfo
155 123 0:44 / /root/tmp/A rw,relatime shared:70 - tmpfs zzzz rw
156 155 0:45 / /root/tmp/A/a rw,relatime shared:71 - tmpfs zzzz rw
157 123 0:44 / /root/tmp/B rw,relatime shared:70 - tmpfs zzzz rw
158 157 0:46 / /root/tmp/B/a rw,relatime shared:72 - tmpfs zzzz rw
159 155 0:46 / /root/tmp/A/a rw,relatime shared:72 - tmpfs zzzz rw
+ umount B/a
+ grep zzzz
+ cat /proc/self/mountinfo
155 123 0:44 / /root/tmp/A rw,relatime shared:70 - tmpfs zzzz rw
156 155 0:45 / /root/tmp/A/a rw,relatime shared:71 - tmpfs zzzz rw
157 123 0:44 / /root/tmp/B rw,relatime shared:70 - tmpfs zzzz rw

X + a - a = X

Maybe we need to add another ID for propagated mounts and when we do umount, we will detach only mounts with the same propagation id.

I support the idea to kill shadow mounts. I guess it will help us to simplify algorithm of dumping and restoring a mount tree in CRIU.

Currently it is a big pain for us.

Quoted 3 lines · Eric W. Biederman

That is certainly interesting. The problem is that the reason we were going slow is that there were in fact mounts that had not been traversed in the share group.

You are right.

Quoted 8 lines

And in fact the entire idea of visiting a vfsmount mountpoint pair exactly once is wrong in the face of shadow mounts. For a vfsmount mountpoint pair that has shadow mounts the number of shadow mounts needs to be descreased by one each time the propgation tree is traversed during unmount. Which means that as far as I can see we have to kill shadow mounts to correctly optimize this code. Once shadow mounts are gone I don't know of a case where need your optimization.

Without shadow mounts, it will be hard to save predictable behaviour for cases like this:

$ unshare --propagation private -m sh test.sh
+ mount -t tmpfs --make-shared zzzz A
+ mkdir A/a
+ mount -t tmpfs zzzz A/a
+ mount --bind A B
+ mount -t tmpfs zzzz B/a
+ grep zzzz
+ cat /proc/self/mountinfo
155 123 0:44 / /root/tmp/A rw,relatime shared:70 - tmpfs zzzz rw
156 155 0:45 / /root/tmp/A/a rw,relatime shared:71 - tmpfs zzzz rw
157 123 0:44 / /root/tmp/B rw,relatime shared:70 - tmpfs zzzz rw
158 157 0:46 / /root/tmp/B/a rw,relatime shared:72 - tmpfs zzzz rw
159 155 0:46 / /root/tmp/A/a rw,relatime shared:72 - tmpfs zzzz rw
+ umount B/a
+ grep zzzz
+ cat /proc/self/mountinfo
155 123 0:44 / /root/tmp/A rw,relatime shared:70 - tmpfs zzzz rw
156 155 0:45 / /root/tmp/A/a rw,relatime shared:71 - tmpfs zzzz rw
157 123 0:44 / /root/tmp/B rw,relatime shared:70 - tmpfs zzzz rw

X + a - a = X

Maybe we need to add another ID for propagated mounts and when we do umount, we will detach only mounts with the same propagation id.

I support the idea to kill shadow mounts. I guess it will help us to simplify algorithm of dumping and restoring a mount tree in CRIU.

Currently it is a big pain for us.

Quoted 43 lines · Andrei Vagin

From 8e0f45c0272aa1f789d1657a0acc98c58919dcc3 Mon Sep 17 00:00:00 2001
From: Andrei Vagin <[email protected]>
Date: Tue, 25 Oct 2016 13:57:31 -0700
Subject: [PATCH] mount: skip all mounts from a shared group if one is marked

If we meet a marked mount, it means that all mounts from its group have been already revised.

Signed-off-by Andrei Vagin <[email protected]>

---
 fs/pnode.c | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

fs/pnode.c +9 -3

index 8fd1a3f..ebb7134 100644
@@ -426,10 +426,16 @@ static struct mount *propagation_visit_child(struct mount *last_child,
 		if (child && !IS_MNT_MARKED(child))
 			return child;
 
-		if (!child)
+		if (!child) {
 			m = propagation_next(m, origin);
-		else
+		} else {
+			if (IS_MNT_MARKED(child)) {
+				if (m->mnt_group_id == origin->mnt_group_id)
+					return NULL;
+				m = m->mnt_master;
+			}
 			m = propagation_next_sib(m, origin);
+		}
 	}
 	return NULL;
 }

@@ -456,8 +462,14 @@ static struct mount *propagation_revisit_child(struct mount *last_child,
 
 		if (!child)
 			m = propagation_next(m, origin);
-		else
+		else {

fs/pnode.c static struct mount *propagation_revisit_child(struct mount *last_child, · line 475

Quoted 10 lines

fs/pnode.c +6 -0

+			if (!IS_MNT_MARKED(child)) {
+				if (m->mnt_group_id == origin->mnt_group_id)
+					return NULL;
+				m = m->mnt_master;
+			}
 			m = propagation_next_sib(m, origin);
+		}
 	}
 	return NULL;
 }

That is certainly interesting. The problem is that the reason we were going slow is that there were in fact mounts that had not been traversed in the share group.