KEYS: find_keyring_by_name() can gain access to a freed keyring
authorToshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Fri, 30 Apr 2010 13:32:13 +0000 (14:32 +0100)
committerJames Morris <jmorris@namei.org>
Wed, 5 May 2010 13:49:10 +0000 (23:49 +1000)
find_keyring_by_name() can gain access to a keyring that has had its reference
count reduced to zero, and is thus ready to be freed.  This then allows the
dead keyring to be brought back into use whilst it is being destroyed.

The following timeline illustrates the process:

|(cleaner)                           (user)
|
| free_user(user)                    sys_keyctl()
|  |                                  |
|  key_put(user->session_keyring)     keyctl_get_keyring_ID()
|  || //=> keyring->usage = 0        |
|  |schedule_work(&key_cleanup_task)   lookup_user_key()
|  ||                                   |
|  kmem_cache_free(,user)               |
|  .                                    |[KEY_SPEC_USER_KEYRING]
|  .                                    install_user_keyrings()
|  .                                    ||
| key_cleanup() [<= worker_thread()]    ||
|  |                                    ||
|  [spin_lock(&key_serial_lock)]        |[mutex_lock(&key_user_keyr..mutex)]
|  |                                    ||
|  atomic_read() == 0                   ||
|  |{ rb_ease(&key->serial_node,) }     ||
|  |                                    ||
|  [spin_unlock(&key_serial_lock)]      |find_keyring_by_name()
|  |                                    |||
|  keyring_destroy(keyring)             ||[read_lock(&keyring_name_lock)]
|  ||                                   |||
|  |[write_lock(&keyring_name_lock)]    ||atomic_inc(&keyring->usage)
|  |.                                   ||| *** GET freeing keyring ***
|  |.                                   ||[read_unlock(&keyring_name_lock)]
|  ||                                   ||
|  |list_del()                          |[mutex_unlock(&key_user_k..mutex)]
|  ||                                   |
|  |[write_unlock(&keyring_name_lock)]  ** INVALID keyring is returned **
|  |                                    .
|  kmem_cache_free(,keyring)            .
|                                       .
|                                       atomic_dec(&keyring->usage)
v                                         *** DESTROYED ***
TIME

If CONFIG_SLUB_DEBUG=y then we may see the following message generated:

=============================================================================
BUG key_jar: Poison overwritten
-----------------------------------------------------------------------------

INFO: 0xffff880197a7e200-0xffff880197a7e200. First byte 0x6a instead of 0x6b
INFO: Allocated in key_alloc+0x10b/0x35f age=25 cpu=1 pid=5086
INFO: Freed in key_cleanup+0xd0/0xd5 age=12 cpu=1 pid=10
INFO: Slab 0xffffea000592cb90 objects=16 used=2 fp=0xffff880197a7e200 flags=0x200000000000c3
INFO: Object 0xffff880197a7e200 @offset=512 fp=0xffff880197a7e300

Bytes b4 0xffff880197a7e1f0:  5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZZZZZZZZZ
  Object 0xffff880197a7e200:  6a 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b jkkkkkkkkkkkkkkk

Alternatively, we may see a system panic happen, such as:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000001
IP: [<ffffffff810e61a3>] kmem_cache_alloc+0x5b/0xe9
PGD 6b2b4067 PUD 6a80d067 PMD 0
Oops: 0000 [#1] SMP
last sysfs file: /sys/kernel/kexec_crash_loaded
CPU 1
...
Pid: 31245, comm: su Not tainted 2.6.34-rc5-nofixed-nodebug #2 D2089/PRIMERGY
RIP: 0010:[<ffffffff810e61a3>]  [<ffffffff810e61a3>] kmem_cache_alloc+0x5b/0xe9
RSP: 0018:ffff88006af3bd98  EFLAGS: 00010002
RAX: 0000000000000000 RBX: 0000000000000001 RCX: ffff88007d19900b
RDX: 0000000100000000 RSI: 00000000000080d0 RDI: ffffffff81828430
RBP: ffffffff81828430 R08: ffff88000a293750 R09: 0000000000000000
R10: 0000000000000001 R11: 0000000000100000 R12: 00000000000080d0
R13: 00000000000080d0 R14: 0000000000000296 R15: ffffffff810f20ce
FS:  00007f97116bc700(0000) GS:ffff88000a280000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000001 CR3: 000000006a91c000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process su (pid: 31245, threadinfo ffff88006af3a000, task ffff8800374414c0)
Stack:
 0000000512e0958e 0000000000008000 ffff880037f8d180 0000000000000001
 0000000000000000 0000000000008001 ffff88007d199000 ffffffff810f20ce
 0000000000008000 ffff88006af3be48 0000000000000024 ffffffff810face3
Call Trace:
 [<ffffffff810f20ce>] ? get_empty_filp+0x70/0x12f
 [<ffffffff810face3>] ? do_filp_open+0x145/0x590
 [<ffffffff810ce208>] ? tlb_finish_mmu+0x2a/0x33
 [<ffffffff810ce43c>] ? unmap_region+0xd3/0xe2
 [<ffffffff810e4393>] ? virt_to_head_page+0x9/0x2d
 [<ffffffff81103916>] ? alloc_fd+0x69/0x10e
 [<ffffffff810ef4ed>] ? do_sys_open+0x56/0xfc
 [<ffffffff81008a02>] ? system_call_fastpath+0x16/0x1b
Code: 0f 1f 44 00 00 49 89 c6 fa 66 0f 1f 44 00 00 65 4c 8b 04 25 60 e8 00 00 48 8b 45 00 49 01 c0 49 8b 18 48 85 db 74 0d 48 63 45 18 <48> 8b 04 03 49 89 00 eb 14 4c 89 f9 83 ca ff 44 89 e6 48 89 ef
RIP  [<ffffffff810e61a3>] kmem_cache_alloc+0x5b/0xe9

This problem is that find_keyring_by_name does not confirm that the keyring is
valid before accepting it.

Skipping keyrings that have been reduced to a zero count seems the way to go.
To this end, use atomic_inc_not_zero() to increment the usage count and skip
the candidate keyring if that returns false.

The following script _may_ cause the bug to happen, but there's no guarantee
as the window of opportunity is small:

#!/bin/sh
LOOP=100000
USER=dummy_user
/bin/su -c "exit;" $USER || { /usr/sbin/adduser -m $USER; add=1; }
for ((i=0; i<LOOP; i++))
do
/bin/su -c "echo '$i' > /dev/null" $USER
done
(( add == 1 )) && /usr/sbin/userdel -r $USER
exit

Note that the nominated user must not be in use.

An alternative way of testing this may be:

for ((i=0; i<100000; i++))
do
keyctl session foo /bin/true || break
done >&/dev/null

as that uses a keyring named "foo" rather than relying on the user and
user-session named keyrings.

Reported-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Acked-by: Serge Hallyn <serue@us.ibm.com>
Signed-off-by: James Morris <jmorris@namei.org>
security/keys/keyring.c

index dd7cd0f8e13cffa6b561f707a6c508a38f21f7f5..0b27271c670c93df03f1de6c2e8da8983f9d0821 100644 (file)
@@ -526,9 +526,8 @@ struct key *find_keyring_by_name(const char *name, bool skip_perm_check)
        struct key *keyring;
        int bucket;
 
-       keyring = ERR_PTR(-EINVAL);
        if (!name)
-               goto error;
+               return ERR_PTR(-EINVAL);
 
        bucket = keyring_hash(name);
 
@@ -555,17 +554,18 @@ struct key *find_keyring_by_name(const char *name, bool skip_perm_check)
                                           KEY_SEARCH) < 0)
                                continue;
 
-                       /* we've got a match */
-                       atomic_inc(&keyring->usage);
-                       read_unlock(&keyring_name_lock);
-                       goto error;
+                       /* we've got a match but we might end up racing with
+                        * key_cleanup() if the keyring is currently 'dead'
+                        * (ie. it has a zero usage count) */
+                       if (!atomic_inc_not_zero(&keyring->usage))
+                               continue;
+                       goto out;
                }
        }
 
-       read_unlock(&keyring_name_lock);
        keyring = ERR_PTR(-ENOKEY);
-
- error:
+out:
+       read_unlock(&keyring_name_lock);
        return keyring;
 
 } /* end find_keyring_by_name() */