ipc,msg: shorten critical region in msgsnd
Hi Davidlohr,
I think the patch (3dd1f784ed6603d7ab1043e51e6371235edf2313) is still unsafe, i.e. my correction (bebcb928c820d0ee83aca4b192adc195e43e66a2) doesn't fix everything:
AFAICS, ipc_obtain_object_check:
- look up the id in the idr tree
- check if it is deleted
- return without taking any locks.
This means that the "is not deleted" information can be stale immediately.
Thus do_msgsnd() in ipc/msg.c contains a memory leak:
Quoted 6 lines
rcu_read_lock();
msq = msq_obtain_object_check(ns, msqid);
if (IS_ERR(msq)) {
err = PTR_ERR(msq);
goto out_unlock1;
}<<<< what if the code is preempted here and RMID is processed? The code below would queue the message into an already removed queue. The queue is freed by the rcu callback, but the message memory is leaked.
Quoted 2 lines
ipc_lock_object(&msq->q_perm);
Is this analysis correct?
And: What about the other users of obtain_object_check? exit_sem() is also quite long, but I didn't spot any obvious problems.
--
Manfred
Hi all,
Quoted 4 lines · Manfred Spraul
And: What about the other users of
obtain_object_check?exit_sem()is also quite long, but I didn't spot any obvious problems.
a) I think semtimed(), msgsnd() and msgrcv() must be fixed:
They either leak memory or tasks can sleep forever.
I haven't checked the shm code, I would expect that there are similar problems.
b) There are additional races at least with selinux:security/selinux/hooks.c
- selinux_sem_semop() accesses sma->sem_perm.security->sid.
- selinux_sem_free_security() does kfree() q_perm.security.
Right now, both operations can happen in parallel -> use after free.
I think the security_xx_yy() calls within ipc/*.c must only be called:
- after checking _perm.deleted
- with ipc_perm.lock acquired (to prevent parallel RMID calls).
Davidlohr:
What would be your proposal?
--
Manfred