lkml.org 
[lkml]   [2019]   [Mar]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: KASAN: use-after-free Read in get_mem_cgroup_from_mm
On 2019/3/7 2:29, Andrea Arcangeli wrote:
> Hello Zhong,
>
> On Wed, Mar 06, 2019 at 09:07:00PM +0800, zhong jiang wrote:
>> The patch use call_rcu to delay free the task_struct, but It is possible to free the task_struct
>> ahead of get_mem_cgroup_from_mm. is it right?
> Yes it is possible to free before get_mem_cgroup_from_mm, but if it's
> freed before get_mem_cgroup_from_mm rcu_read_lock,
> rcu_dereference(mm->owner) will return NULL in such case and there
> will be no problem.
Yes
> The simple fix also clears the mm->owner of the failed-fork-mm before
> doing the call_rcu. The call_rcu delays the freeing after no other CPU
> runs in between rcu_read_lock/unlock anymore. That guarantees that
> those critical section will see mm->owner == NULL if the freeing of
> the task strut already happened.
We has set the mm->owner to NULL when child process fails to fork ahead of freeing
the task struct.

Have those critical section chance to see the mm->owner, which is not NULL.

I has tested the patch. Not Oops and panic appear so far.

Thanks,
zhong jiang
> The solution Mike suggested for this and that we were wondering as
> ideal in the past for the signal issue too, is to move the uffd
> delivery at a point where fork is guaranteed to succeed. We should
> probably try that too to see how it looks like and if it can be done
> in a not intrusive way, but the simple fix that uses RCU should work
> too.
>
> Rolling back in case of errors inside fork itself isn't easily doable:
> the moment we push the uffd ctx to the other side of the uffd pipe
> there's no coming back as that information can reach the userland of
> the uffd monitor/reader thread immediately after. The rolling back is
> really the other thread failing at mmget_not_zero eventually. It's the
> userland that has to rollback in such case when it gets a -ESRCH
> retval.
>
> Note that this fork feature is only ever needed in the non-cooperative
> case, these things never need to happen when userfaultfd is used by an
> app (or a lib) that is aware that it is using userfaultfd.
>
> Thanks,
> Andrea
>
> .
>


\
 
 \ /
  Last update: 2019-03-07 08:59    [W:0.105 / U:0.268 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site