Oleg Nesterov <oleg@xxxxxxxxxx> writes:
> On 05/04, Eric W. Biederman wrote:
>>
>> With the removal of the incomplete detection of the tracer going away
>> in ptrace_stop, ptrace_stop always sleeps in schedule after
>> ptrace_freeze_traced succeeds. Modify ptrace_check_attach to
>> warn if wait_task_inactive fails.
>
> Oh. Again, I don't understand the changelog. If we forget about RT,
> ptrace_stop() will always sleep if ptrace_freeze_traced() succeeds.
> may_ptrace_stop() has gone.
>
> IOW. Lets forget about RT
>
>> --- a/kernel/ptrace.c
>> +++ b/kernel/ptrace.c
>> @@ -266,17 +266,9 @@ static int ptrace_check_attach(struct task_struct *child, bool ignore_state)
>> }
>> read_unlock(&tasklist_lock);
>>
>> - if (!ret && !ignore_state) {
>> - if (!wait_task_inactive(child, __TASK_TRACED)) {
>> - /*
>> - * This can only happen if may_ptrace_stop() fails and
>> - * ptrace_stop() changes ->state back to TASK_RUNNING,
>> - * so we should not worry about leaking __TASK_TRACED.
>> - */
>> - WARN_ON(READ_ONCE(child->__state) == __TASK_TRACED);
>> - ret = -ESRCH;
>> - }
>> - }
>> + if (!ret && !ignore_state &&
>> + WARN_ON_ONCE(!wait_task_inactive(child, __TASK_TRACED)))
>> + ret = -ESRCH;
>>
>> return ret;
>> }
>
> Why do you think this change would be wrong without any other changes?
For purposes of this analysis ptrace_detach and ptrace_exit (when the
tracer exits) can't happen. So the bug you spotted in ptrace_stop does
not apply.
I was thinking that the test against !current->ptrace that replaced
the old may_ptrace_stop could trigger a failure here. If the
ptrace_freeze_traced happens before that test that branch clearly can
not happen.
*Looks twice* Both ptrace_check_attach and ptrace_stop taking a
read_lock on tasklist_lock does not protect against concurrency by each
other, but the write_lock on tasklist_lock in ptrace_attach does
protect against a ptrace_attach coming in after the test and before
__set_current_state(TASK_RUNNING).
So yes. I should really split that part out into it's own patch.
And yes that WARN_ON_ONCE can trigger on PREEMPT_RT but that is just
because PREMPT_RT is currently broken with respect to ptrace. Which
makes a WARN_ON_ONCE appropriate.
I will see how much of this analysis I can put in the changelog.
Thank you,
Eric
|