On 01/22/2015 03:44 AM, Ethan Zhao wrote: > On Wed, Jan 21, 2015 at 1:30 PM, Manfred Spraul > wrote: >> On 01/21/2015 04:53 AM, Ethan Zhao wrote: >>> On Tue, Jan 20, 2015 at 10:10 PM, Stephen Smalley >>> wrote: >>>> On 01/20/2015 04:18 AM, Ethan Zhao wrote: >>>>> sys_semget() >>>>> ->newary() >>>>> ->security_sem_alloc() >>>>> ->sem_alloc_security() >>>>> selinux_sem_alloc_security() >>>>> ->ipc_alloc_security() { >>>>> ->rc = avc_has_perm() >>>>> if (rc) { >>>>> >>>>> ipc_free_security(&sma->sem_perm); >>>>> return rc; >>>> We free the security structure here to avoid a memory leak on a >>>> failed/denied semaphore set creation. In this situation, we return an >>>> error to the caller (ultimately to newary), it does an >>>> ipc_rcu_putref(sma, ipc_rcu_free), and it returns an error to the >>>> caller. Thus, it never calls ipc_addid() and the semaphore set is not >>>> created. So how then can you call semtimedop() on it? >>> Seems it wouldn't happen after commit >>> e8577d1f0329d4842e8302e289fb2c22156abef4 ? >> That was my first guess when I read the bug report - but it can't be the >> fix, because security_sem_alloc() is before the ipc_addid(), with or without >> the patch. >> >> thread A: >> thread B: >> >> semtimedop() >> -> sem_obtain_object_check() >> semctl(IPC_RMID) >> -> freeary() >> -> ipc_rcu_putref() >> -> call_rcu() >> -> somehow a grace period >> -> sem_rcu_free() >> -> security_sem_free() >> >> Perhaps: modify ipc_free_security() to hexdump perm and a few more bytes if >> the pointer is NULL? > I tried to ask for vmcore and do more analysis, basically, the race condition > still exists and open a hole to be DoS. Is the issue reproducable? If yes, can you try something like the attached patch? > --- a/security/selinux/hooks.c > +++ b/security/selinux/hooks.c > @@ -5129,6 +5129,8 @@ static int ipc_has_perm(struct kern_ipc_perm > *ipc_perms, > u32 sid = current_sid(); > > isec = ipc_perms->security; > + if (!isec) > + return -EACCES; > > ad.type = LSM_AUDIT_DATA_IPC; > ad.u.ipc_id = ipc_perms->key; ipc_has_perm() runs without any spinlocks/semaphores, only rcu_read_lock(). Testing for ipc_perms->security!=NULL does solve the issue that ipc_perm->key could be an access to kfree'd memory: Nothing prevents that the kfree could happen just after the test. I.e.: The patch can't be the right solution. -- Manfred