proc/sysctl: Don't grab i_lock under sysctl_lock.
authorEric W. Biederman <ebiederm@xmission.com>
Mon, 20 Feb 2017 05:17:03 +0000 (18:17 +1300)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 15 Aug 2018 16:14:43 +0000 (18:14 +0200)
commit631f93a6fe847d2d317010d5bbd7cb3bcc284336
tree5f3b6861fb3feb5405e58aca09fdef20eaf5ed63
parentb96e215e539509cae8bfe468689b70661cf511b4
proc/sysctl: Don't grab i_lock under sysctl_lock.

commit ace0c791e6c3cf5ef37cad2df69f0d90ccc40ffb upstream.

Konstantin Khlebnikov <khlebnikov@yandex-team.ru> writes:
> This patch has locking problem. I've got lockdep splat under LTP.
>
> [ 6633.115456] ======================================================
> [ 6633.115502] [ INFO: possible circular locking dependency detected ]
> [ 6633.115553] 4.9.10-debug+ #9 Tainted: G             L
> [ 6633.115584] -------------------------------------------------------
> [ 6633.115627] ksm02/284980 is trying to acquire lock:
> [ 6633.115659]  (&sb->s_type->i_lock_key#4){+.+...}, at: [<ffffffff816bc1ce>] igrab+0x1e/0x80
> [ 6633.115834] but task is already holding lock:
> [ 6633.115882]  (sysctl_lock){+.+...}, at: [<ffffffff817e379b>] unregister_sysctl_table+0x6b/0x110
> [ 6633.116026] which lock already depends on the new lock.
> [ 6633.116026]
> [ 6633.116080]
> [ 6633.116080] the existing dependency chain (in reverse order) is:
> [ 6633.116117]
> -> #2 (sysctl_lock){+.+...}:
> -> #1 (&(&dentry->d_lockref.lock)->rlock){+.+...}:
> -> #0 (&sb->s_type->i_lock_key#4){+.+...}:
>
> d_lock nests inside i_lock
> sysctl_lock nests inside d_lock in d_compare
>
> This patch adds i_lock nesting inside sysctl_lock.

Al Viro <viro@ZenIV.linux.org.uk> replied:
> Once ->unregistering is set, you can drop sysctl_lock just fine.  So I'd
> try something like this - use rcu_read_lock() in proc_sys_prune_dcache(),
> drop sysctl_lock() before it and regain after.  Make sure that no inodes
> are added to the list ones ->unregistering has been set and use RCU list
> primitives for modifying the inode list, with sysctl_lock still used to
> serialize its modifications.
>
> Freeing struct inode is RCU-delayed (see proc_destroy_inode()), so doing
> igrab() is safe there.  Since we don't drop inode reference until after we'd
> passed beyond it in the list, list_for_each_entry_rcu() should be fine.

I agree with Al Viro's analsysis of the situtation.

Fixes: d6cffbbe9a7e ("proc/sysctl: prune stale dentries during unregistering")
Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Tested-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Suggested-by: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/proc/proc_sysctl.c