aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorAl Viro <viro@zeniv.linux.org.uk>2018-08-09 17:21:17 -0400
committerAl Viro <viro@zeniv.linux.org.uk>2018-08-09 17:21:17 -0400
commit9ea0a46ca2c318fcc449c1e6b62a7230a17888f1 (patch)
treea8871352d98abb3aae377a6eb079e030a27ae4ff
parent90bad5e05bcdb0308cfa3d3a60f5c0b9c8e2efb3 (diff)
downloadlinux-yocto-9ea0a46ca2c318fcc449c1e6b62a7230a17888f1.tar.gz
linux-yocto-9ea0a46ca2c318fcc449c1e6b62a7230a17888f1.tar.bz2
linux-yocto-9ea0a46ca2c318fcc449c1e6b62a7230a17888f1.zip
fix mntput/mntput race
mntput_no_expire() does the calculation of total refcount under mount_lock; unfortunately, the decrement (as well as all increments) are done outside of it, leading to false positives in the "are we dropping the last reference" test. Consider the following situation: * mnt is a lazy-umounted mount, kept alive by two opened files. One of those files gets closed. Total refcount of mnt is 2. On CPU 42 mntput(mnt) (called from __fput()) drops one reference, decrementing component * After it has looked at component #0, the process on CPU 0 does mntget(), incrementing component #0, gets preempted and gets to run again - on CPU 69. There it does mntput(), which drops the reference (component #69) and proceeds to spin on mount_lock. * On CPU 42 our first mntput() finishes counting. It observes the decrement of component #69, but not the increment of component #0. As the result, the total it gets is not 1 as it should've been - it's 0. At which point we decide that vfsmount needs to be killed and proceed to free it and shut the filesystem down. However, there's still another opened file on that filesystem, with reference to (now freed) vfsmount, etc. and we are screwed. It's not a wide race, but it can be reproduced with artificial slowdown of the mnt_get_count() loop, and it should be easier to hit on SMP KVM setups. Fix consists of moving the refcount decrement under mount_lock; the tricky part is that we want (and can) keep the fast case (i.e. mount that still has non-NULL ->mnt_ns) entirely out of mount_lock. All places that zero mnt->mnt_ns are dropping some reference to mnt and they call synchronize_rcu() before that mntput(). IOW, if mntput() observes (under rcu_read_lock()) a non-NULL ->mnt_ns, it is guaranteed that there is another reference yet to be dropped. Reported-by: Jann Horn <jannh@google.com> Tested-by: Jann Horn <jannh@google.com> Fixes: 48a066e72d97 ("RCU'd vsfmounts") Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
-rw-r--r--fs/namespace.c14
1 files changed, 12 insertions, 2 deletions
diff --git a/fs/namespace.c b/fs/namespace.c
index 8ddd14806799..d46a951bd541 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -1195,12 +1195,22 @@ static DECLARE_DELAYED_WORK(delayed_mntput_work, delayed_mntput);
static void mntput_no_expire(struct mount *mnt)
{
rcu_read_lock();
- mnt_add_count(mnt, -1);
- if (likely(mnt->mnt_ns)) { /* shouldn't be the last one */
+ if (likely(READ_ONCE(mnt->mnt_ns))) {
+ /*
+ * Since we don't do lock_mount_hash() here,
+ * ->mnt_ns can change under us. However, if it's
+ * non-NULL, then there's a reference that won't
+ * be dropped until after an RCU delay done after
+ * turning ->mnt_ns NULL. So if we observe it
+ * non-NULL under rcu_read_lock(), the reference
+ * we are dropping is not the final one.
+ */
+ mnt_add_count(mnt, -1);
rcu_read_unlock();
return;
}
lock_mount_hash();
+ mnt_add_count(mnt, -1);
if (mnt_get_count(mnt)) {
rcu_read_unlock();
unlock_mount_hash();