On 10/25, Tetsuo Handa wrote: > > Oleg Nesterov wrote: > > On 10/22, Tetsuo Handa wrote: > > > > And again, I do not know how/if yama ensures that child is > > > > rcu-protected, perhaps > > > > task_is_descendant() needs to check pid_alive(child) right after > > > > rcu_read_lock() ? > > > > > > Since the caller (ptrace() path) called get_task_struct(child), child > > > itself can't be > > > released. Do we still need pid_alive(child) ? > > > > get_task_struct(child) can only ensure that this task_struct can't be freed. > > The report says that it is a use-after-free read at > > walker = rcu_dereference(walker->real_parent); > > which means that walker was already released.
quite possibly I missed something, but I am not sure I understand your concerns... So again, suppose that "child" is already dead. Its task_struct can't be freed, but child->real_parent can point to the already freed memory. This means that the 1st walker = rcu_dereference(walker->real_parent) is fine, this simply reads the child->real_parent pointer, but on the second iteration walker = rcu_dereference(walker->real_parent); reads the alredy freed memory. > I wonder whether pid_alive() test helps. > > We can get > > [ 40.620318] parent or walker is dead. > [ 40.624146] tracee is dead. > > messages using below patch and reproducer. again, I do not understand, this all looks correct... > ---------- > diff --git a/kernel/ptrace.c b/kernel/ptrace.c > index 99cfddd..0d9d786 100644 > --- a/kernel/ptrace.c > +++ b/kernel/ptrace.c > @@ -385,6 +385,7 @@ static int ptrace_attach(struct task_struct *task, long > request, > if (mutex_lock_interruptible(&task->signal->cred_guard_mutex)) > goto out; > > + schedule_timeout_killable(HZ); > task_lock(task); > retval = __ptrace_may_access(task, PTRACE_MODE_ATTACH_REALCREDS); > task_unlock(task); > diff --git a/security/yama/yama_lsm.c b/security/yama/yama_lsm.c > index ffda91a..a231ec6 100644 > --- a/security/yama/yama_lsm.c > +++ b/security/yama/yama_lsm.c > @@ -283,6 +283,11 @@ static int task_is_descendant(struct task_struct *parent, > return 0; > > rcu_read_lock(); > + if (!pid_alive(parent) || !pid_alive(walker)) { > + rcu_read_unlock(); > + printk("parent or walker is dead.\n"); This is what we need to do, except I think we should change yama_ptrace_access_check(). And iiuc parent == current, pid_alive(parent) looks unnecessary. Although we need to check ptracer_exception_found(), may be it needs some changes too. And yes, task_is_descendant() can hit the dead child, if nothing else it can be killed. This can explain the kasan report. > @@ -315,6 +320,10 @@ static int ptracer_exception_found(struct task_struct > *tracer, > bool found = false; > > rcu_read_lock(); > + if (!pid_alive(tracee)) { > + printk("tracee is dead.\n"); > + goto unlock; Sure, this is possible too. > But since "child" has at least one reference, reading "child->real_parent" > should > be safe. Therefore, I think that bailing out due to pid_is_alive(child) == > false > (like above patch does) cannot avoid this problem... Why? OK. Lets ignore ptracer_exception_found() for the moment. Why do you think the patch below can't help? Oleg. --- x/security/yama/yama_lsm.c +++ x/security/yama/yama_lsm.c @@ -368,7 +368,8 @@ static int yama_ptrace_access_check(stru break; case YAMA_SCOPE_RELATIONAL: rcu_read_lock(); - if (!task_is_descendant(current, child) && + if (!pid_alive(child) || + !task_is_descendant(current, child) && !ptracer_exception_found(current, child) && !ns_capable(__task_cred(child)->user_ns, CAP_SYS_PTRACE)) rc = -EPERM;