Re: [syzbot] KASAN: use-after-free Read in check_all_holdout_tasks_trace

On Tue, May 25, 2021 at 5:33 AM Paul E. McKenney [off-list ref] wrote:

On Tue, May 25, 2021 at 10:31:55AM +0800, Xu, Yanfei wrote:

quoted

On 5/25/21 6:46 AM, Paul E. McKenney wrote:

quoted

[Please note: This e-mail is from an EXTERNAL e-mail address]

On Sun, May 23, 2021 at 09:13:50PM -0700, Paul E. McKenney wrote:

quoted

On Sun, May 23, 2021 at 08:51:56AM +0200, Dmitry Vyukov wrote:

quoted

On Fri, May 21, 2021 at 7:29 PM syzbot
[off-list ref] wrote:

quoted

Hello,

syzbot found the following issue on:

HEAD commit:    f18ba26d libbpf: Add selftests for TC-BPF management API
git tree:       bpf-next
console output: https://syzkaller.appspot.com/x/log.txt?x=17f50d1ed00000
kernel config:  https://syzkaller.appspot.com/x/.config?x=8ff54addde0afb5d
dashboard link: https://syzkaller.appspot.com/bug?extid=7b2b13f4943374609532

Unfortunately, I don't have any reproducer for this issue yet.

IMPORTANT: if you fix the issue, please add the following tag to the commit:
Reported-by: syzbot+7b2b13f4943374609532@syzkaller.appspotmail.com

This looks rcu-related. +rcu mailing list

I think I see a possible cause for this, and will say more after some
testing and after becoming more awake Monday morning, Pacific time.

No joy.  From what I can see, within RCU Tasks Trace, the calls to
get_task_struct() are properly protected (either by RCU or by an earlier
get_task_struct()), and the calls to put_task_struct() are balanced by
those to get_task_struct().

I could of course have missed something, but at this point I am suspecting
an unbalanced put_task_struct() has been added elsewhere.

As always, extra eyes on this code would be a good thing.

If it were reproducible, I would of course suggest bisection.  :-/

                                                         Thanx, Paul

Hi Paul,

Could it be?

       CPU1                                        CPU2
trc_add_holdout(t, bhp)
//t->usage==2
                                      release_task
                                        put_task_struct_rcu_user
                                          delayed_put_task_struct
                                            ......
                                            put_task_struct(t)
                                            //t->usage==1

check_all_holdout_tasks_trace
  ->trc_wait_for_one_reader
    ->trc_del_holdout
      ->put_task_struct(t)
      //t->usage==0 and task_struct freed
  READ_ONCE(t->trc_reader_checked)
  //ops， t had been freed.

So, after excuting trc_wait_for_one_reader（）, task might had been removed
from holdout list and the corresponding task_struct was freed.
And we shouldn't do READ_ONCE(t->trc_reader_checked).

I was suspicious of that call to trc_del_holdout() from within
trc_wait_for_one_reader(), but the only time it executes is in the
context of the current running task, which means that CPU 2 had better
not be invoking release_task() on it just yet.

Or am I missing your point?

Of course, if you can reproduce it, the following patch might be
an interesting thing to try, my doubts notwithstanding.  But more
important, please check the patch to make sure that we are both
talking about the same call to trc_del_holdout()!

If we are talking about the same call to trc_del_holdout(), are you
actually seeing that code execute except when rcu_tasks_trace_pertask()
calls trc_wait_for_one_reader()?

quoted

I investigate the trc_wait_for_one_reader（） and found before we excute
trc_del_holdout, there is always set t->trc_reader_checked=true. How about
we just set the checked flag and unified excute trc_del_holdout()
in check_all_holdout_tasks_trace with checking the flag?

The problem is that we cannot execute trc_del_holdout() except in
the context of the RCU Tasks Trace grace-period kthread.  So it is
necessary to manipulate ->trc_reader_checked separately from the list
in order to safely synchronize with IPIs and with the exit code path
for any reader tasks, see for example trc_read_check_handler() and
exit_tasks_rcu_finish_trace().

Or are you thinking of some other approach?

This could be caused by a buggy extra put_pid somewhere else, right?
If so, I suspect that's what may be happening. We've 2 very similar
use-after-free reports on an internal kernel, but it also has a number
of other use-after-free reports in pid-related functions
(pid_task/pid_nr_ns/attach_pid). One of them is happening relatively
frequently (150 crashes) and is caused by something in the tty
subsystem. Presumably it may be causing one off use-after-free's in
other random places of the kernel as well. Unfortunately these crashes
don't happen on the upstream kernel (at least not yet).
So if you don't see any obvious smoking gun in rcu, I think we can
assume for now that it's due to tty.