Re: [ANNOUNCE] 3.12.5-rt6

From: Sebastian Andrzej Siewior
Date: Tue Dec 17 2013 - 07:42:57 EST


* Steven Rostedt | 2013-12-17 06:31:56 [-0500]:

>On Tue, 17 Dec 2013 08:16:31 +0100
>Mike Galbraith <bitbucket@xxxxxxxxx> wrote:
>
>> Hi Sebastian,
>>
>> Looks like there's a booboo here:
>>
>> On Mon, 2013-12-16 at 10:14 +0100, Sebastian Andrzej Siewior wrote:
>
>> "ptrace: fix ptrace vs tasklist_lock race" added..
>>
>> @@ -1068,8 +1082,11 @@ unsigned long wait_task_inactive(struct
>> * is actually now running somewhere else!
>> */
>> while (task_running(rq, p)) {
>> - if (match_state && unlikely(p->state != match_state))
>> + if (match_state) {
>> + if (!unlikely(check_task_state(p, match_state)))
>> + return 0;
>> return 0;
>> + }
>
>Ouch!

Not exactly sure how I managed this since I had this [0] to test this.
Which I used to come up with the patch. And now I see that the code as
it is here fails the testcase.

[0] http://breakpoint.cc/ptrace-test.c

>> cpu_relax();
>> }
>>
>> ..which is how it stays with the whole series applied.
>>
>> The patch contains hunk 2 from
>>
>> "sched/rt: Fix wait_task_interactive() to test rt_spin_lock state",
>>
>> which went away in -rt6, so it seems the busted hunk should be as below
>> if the two are to be merged.
>>
>> @@ -1068,8 +1082,10 @@ unsigned long wait_task_inactive(struct
>> * is actually now running somewhere else!
>> */
>> while (task_running(rq, p)) {
>> - if (match_state && unlikely(p->state != match_state))
>> + if (match_state && unlikely(p->state != match_state)
>> + && unlikely(p->saved_state != match_state))
>> return 0;
>> + }
>
>Yeah, it should just be:
>
> if (match_state && check_task_state(p, match_state))
> return 0;

Are you sure? If the state matches we should continue as long as it runs
therefore I would go for !check_task_state(). The problem here was that
I return 0 in both cases.

>Also, looking at check_task_state():
>
>+static bool check_task_state(struct task_struct *p, long match_state)
>+{
>+ bool match = false;
>+
>+ raw_spin_lock_irq(&p->pi_lock);
>+ if (p->state == match_state)
>+ match = true;
>+ else if (p->saved_state == match_state)
>+ match = true;
>
>Why the if () else if()? and not just:
>
> if (p->state == match_state || p->save_state == match_state)
> match = true;
>?
>
>The else if makes me think there's something missing.

Okay I can do this. But regarding the check_task_state part, I think I
should go with:

--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1076,9 +1076,7 @@ static bool check_task_state(struct task_struct *p, long match_state)
bool match = false;

raw_spin_lock_irq(&p->pi_lock);
- if (p->state == match_state)
- match = true;
- else if (p->saved_state == match_state)
+ if (p->state == match_state || p->saved_state == match_state)
match = true;
raw_spin_unlock_irq(&p->pi_lock);

@@ -1129,11 +1127,8 @@ unsigned long wait_task_inactive(struct task_struct *p, long match_state)
* is actually now running somewhere else!
*/
while (task_running(rq, p)) {
- if (match_state) {
- if (!unlikely(check_task_state(p, match_state)))
- return 0;
+ if (match_state && !check_task_state(p, match_state))
return 0;
- }
cpu_relax();
}


Any objections?

>
>-- Steve
>

Sebastian
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/