CRED: Fix a race in creds_are_invalid() in credentials debugging
commitfd701336cefd200882c5c32dfea09690221067fb
authorDavid Howells <dhowells@redhat.com>
Wed, 21 Apr 2010 09:28:25 +0000 (21 10:28 +0100)
committerGreg Kroah-Hartman <gregkh@suse.de>
Wed, 12 May 2010 21:57:10 +0000 (12 14:57 -0700)
tree9db8c23285d9d056a4ed8e809fc423e125742112
parent85bf36e04883f422a10e6b95bd103e7b7f8ff064
CRED: Fix a race in creds_are_invalid() in credentials debugging

commit e134d200d57d43b171dcb0b55c178a1a0c7db14a upstream.

creds_are_invalid() reads both cred->usage and cred->subscribers and then
compares them to make sure the number of processes subscribed to a cred struct
never exceeds the refcount of that cred struct.

The problem is that this can cause a race with both copy_creds() and
exit_creds() as the two counters, whilst they are of atomic_t type, are only
atomic with respect to themselves, and not atomic with respect to each other.

This means that if creds_are_invalid() can read the values on one CPU whilst
they're being modified on another CPU, and so can observe an evolving state in
which the subscribers count now is greater than the usage count a moment
before.

Switching the order in which the counts are read cannot help, so the thing to
do is to remove that particular check.

I had considered rechecking the values to see if they're in flux if the test
fails, but I can't guarantee they won't appear the same, even if they've
changed several times in the meantime.

Note that this can only happen if CONFIG_DEBUG_CREDENTIALS is enabled.

The problem is only likely to occur with multithreaded programs, and can be
tested by the tst-eintr1 program from glibc's "make check".  The symptoms look
like:

CRED: Invalid credentials
CRED: At include/linux/cred.h:240
CRED: Specified credentials: ffff88003dda5878 [real][eff]
CRED: ->magic=43736564, put_addr=(null)
CRED: ->usage=766, subscr=766
CRED: ->*uid = { 0,0,0,0 }
CRED: ->*gid = { 0,0,0,0 }
CRED: ->security is ffff88003d72f538
CRED: ->security {359, 359}
------------[ cut here ]------------
kernel BUG at kernel/cred.c:850!
...
RIP: 0010:[<ffffffff81049889>]  [<ffffffff81049889>] __invalid_creds+0x4e/0x52
...
Call Trace:
 [<ffffffff8104a37b>] copy_creds+0x6b/0x23f

Note the ->usage=766 and subscr=766.  The values appear the same because
they've been re-read since the check was made.

Reported-by: Roland McGrath <roland@redhat.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: James Morris <jmorris@namei.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
kernel/cred.c