mnt: In propgate_umount handle visiting mounts in any order
authorEric W. Biederman <ebiederm@xmission.com>
Mon, 24 Oct 2016 21:16:13 +0000 (16:16 -0500)
committerEric W. Biederman <ebiederm@xmission.com>
Tue, 23 May 2017 13:41:16 +0000 (08:41 -0500)
While investigating some poor umount performance I realized that in
the case of overlapping mount trees where some of the mounts are locked
the code has been failing to unmount all of the mounts it should
have been unmounting.

This failure to unmount all of the necessary
mounts can be reproduced with:

$ cat locked_mounts_test.sh

mount -t tmpfs test-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 --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi'
sleep 1
umount -l /mnt/b
wait %%

$ unshare -Urm ./locked_mounts_test.sh

This failure is corrected by removing the prepass that marks mounts
that may be umounted.

A first pass is added that umounts mounts if possible and if not sets
mount mark if they could be unmounted if they weren't locked and adds
them to a list to umount possibilities.  This first pass reconsiders
the mounts parent if it is on the list of umount possibilities, ensuring
that information of umoutability will pass from child to mount parent.

A second pass then walks through all mounts that are umounted and processes
their children unmounting them or marking them for reparenting.

A last pass cleans up the state on the mounts that could not be umounted
and if applicable reparents them to their first parent that remained
mounted.

While a bit longer than the old code this code is much more robust
as it allows information to flow up from the leaves and down
from the trunk making the order in which mounts are encountered
in the umount propgation tree irrelevant.

Cc: stable@vger.kernel.org
Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts")
Reviewed-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
fs/mount.h
fs/namespace.c
fs/pnode.c

index ede5a1d5cf997060aee48558c09afc95874bfefc..de45d9e7674878db78d77d9aa4ac09e8da429dd9 100644 (file)
@@ -58,7 +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_reparent;  /* reparent list entry */
+       struct list_head mnt_umounting; /* list entry for umount propagation */
 #ifdef CONFIG_FSNOTIFY
        struct fsnotify_mark_connector __rcu *mnt_fsnotify_marks;
        __u32 mnt_fsnotify_mask;
index 51e49866e1fec34f1b1c4bae6b8f1d08d2f37305..5e3dcbeb1de576d0bd03a5cb87a0988c296255e0 100644 (file)
@@ -236,7 +236,7 @@ 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_reparent);
+               INIT_LIST_HEAD(&mnt->mnt_umounting);
                init_fs_pin(&mnt->mnt_umount, drop_mountpoint);
        }
        return mnt;
index 52aca0a118ff476f2ef116b14bb23154b32d9270..fbaca7df2eb08b5ddb5ae44a2a6e38e95c4ac999 100644 (file)
@@ -413,86 +413,95 @@ 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 void umount_one(struct mount *mnt, struct list_head *to_umount)
 {
-       struct mount *parent = mnt->mnt_parent;
-       struct mount *m;
-
-       BUG_ON(parent == mnt);
-
-       for (m = propagation_next(parent, parent); m;
-                       m = propagation_next(m, parent)) {
-               struct mount *child = __lookup_mnt(&m->mnt,
-                                               mnt->mnt_mountpoint);
-               if (!child || (child->mnt.mnt_flags & MNT_UMOUNT))
-                       continue;
-               if (!IS_MNT_LOCKED(child) || IS_MNT_MARKED(m)) {
-                       SET_MNT_MARK(child);
-               }
-       }
+       CLEAR_MNT_MARK(mnt);
+       mnt->mnt.mnt_flags |= MNT_UMOUNT;
+       list_del_init(&mnt->mnt_child);
+       list_del_init(&mnt->mnt_umounting);
+       list_move_tail(&mnt->mnt_list, to_umount);
 }
 
 /*
  * NOTE: unmounting 'mnt' naturally propagates to all other mounts its
  * parent propagates to.
  */
-static void __propagate_umount(struct mount *mnt, struct list_head *to_reparent)
+static bool __propagate_umount(struct mount *mnt,
+                              struct list_head *to_umount,
+                              struct list_head *to_restore)
 {
-       struct mount *parent = mnt->mnt_parent;
-       struct mount *m;
+       bool progress = false;
+       struct mount *child;
 
-       BUG_ON(parent == mnt);
+       /*
+        * The state of the parent won't change if this mount is
+        * already unmounted or marked as without children.
+        */
+       if (mnt->mnt.mnt_flags & (MNT_UMOUNT | MNT_MARKED))
+               goto out;
 
-       for (m = propagation_next(parent, parent); m;
-                       m = propagation_next(m, parent)) {
-               struct mount *topper;
-               struct mount *child = __lookup_mnt(&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))
+       /* Verify topper is the only grandchild that has not been
+        * speculatively unmounted.
+        */
+       list_for_each_entry(child, &mnt->mnt_mounts, mnt_child) {
+               if (child->mnt_mountpoint == mnt->mnt.mnt_root)
                        continue;
-               CLEAR_MNT_MARK(child);
+               if (!list_empty(&child->mnt_umounting) && IS_MNT_MARKED(child))
+                       continue;
+               /* Found a mounted child */
+               goto children;
+       }
 
-               /* If there is exactly one mount covering all of child
-                * replace child with that mount.
-                */
-               topper = find_topper(child);
-               if (topper)
-                       list_add_tail(&topper->mnt_reparent, to_reparent);
+       /* Mark mounts that can be unmounted if not locked */
+       SET_MNT_MARK(mnt);
+       progress = true;
 
-               if (topper || list_empty(&child->mnt_mounts)) {
-                       list_del_init(&child->mnt_child);
-                       list_del_init(&child->mnt_reparent);
-                       child->mnt.mnt_flags |= MNT_UMOUNT;
-                       list_move_tail(&child->mnt_list, &mnt->mnt_list);
+       /* If a mount is without children and not locked umount it. */
+       if (!IS_MNT_LOCKED(mnt)) {
+               umount_one(mnt, to_umount);
+       } else {
+children:
+               list_move_tail(&mnt->mnt_umounting, to_restore);
+       }
+out:
+       return progress;
+}
+
+static void umount_list(struct list_head *to_umount,
+                       struct list_head *to_restore)
+{
+       struct mount *mnt, *child, *tmp;
+       list_for_each_entry(mnt, to_umount, mnt_list) {
+               list_for_each_entry_safe(child, tmp, &mnt->mnt_mounts, mnt_child) {
+                       /* topper? */
+                       if (child->mnt_mountpoint == mnt->mnt.mnt_root)
+                               list_move_tail(&child->mnt_umounting, to_restore);
+                       else
+                               umount_one(child, to_umount);
                }
        }
 }
 
-static void reparent_mounts(struct list_head *to_reparent)
+static void restore_mounts(struct list_head *to_restore)
 {
-       while (!list_empty(to_reparent)) {
+       /* Restore mounts to a clean working state */
+       while (!list_empty(to_restore)) {
                struct mount *mnt, *parent;
                struct mountpoint *mp;
 
-               mnt = list_first_entry(to_reparent, struct mount, mnt_reparent);
-               list_del_init(&mnt->mnt_reparent);
+               mnt = list_first_entry(to_restore, struct mount, mnt_umounting);
+               CLEAR_MNT_MARK(mnt);
+               list_del_init(&mnt->mnt_umounting);
 
-               /* Where should this mount be reparented to? */
+               /* Should this mount be reparented? */
                mp = mnt->mnt_mp;
                parent = mnt->mnt_parent;
                while (parent->mnt.mnt_flags & MNT_UMOUNT) {
                        mp = parent->mnt_mp;
                        parent = parent->mnt_parent;
                }
-
-               mnt_change_mountpoint(parent, mp, mnt);
+               if (parent != mnt->mnt_parent)
+                       mnt_change_mountpoint(parent, mp, mnt);
        }
 }
 
@@ -506,15 +515,34 @@ static void reparent_mounts(struct list_head *to_reparent)
 int propagate_umount(struct list_head *list)
 {
        struct mount *mnt;
-       LIST_HEAD(to_reparent);
+       LIST_HEAD(to_restore);
+       LIST_HEAD(to_umount);
 
-       list_for_each_entry_reverse(mnt, list, mnt_list)
-               mark_umount_candidates(mnt);
+       list_for_each_entry(mnt, list, mnt_list) {
+               struct mount *parent = mnt->mnt_parent;
+               struct mount *m;
 
-       list_for_each_entry(mnt, list, mnt_list)
-               __propagate_umount(mnt, &to_reparent);
+               for (m = propagation_next(parent, parent); m;
+                    m = propagation_next(m, parent)) {
+                       struct mount *child = __lookup_mnt(&m->mnt,
+                                                          mnt->mnt_mountpoint);
+                       if (!child)
+                               continue;
+
+                       /* Check the child and parents while progress is made */
+                       while (__propagate_umount(child,
+                                                 &to_umount, &to_restore)) {
+                               /* Is the parent a umount candidate? */
+                               child = child->mnt_parent;
+                               if (list_empty(&child->mnt_umounting))
+                                       break;
+                       }
+               }
+       }
 
-       reparent_mounts(&to_reparent);
+       umount_list(&to_umount, &to_restore);
+       restore_mounts(&to_restore);
+       list_splice_tail(&to_umount, list);
 
        return 0;
 }