Commit 95158a89dd50 ("sched,rt: Use the full cpumask for balancing")
allow find_lock_lowest_rq to pick a task with migration disabled.
This commit is intended to push the current running task on this CPU
away.
There is a race scenario, which allows a migration disabled task to
be migrated to another CPU.
When there is a RT task with higher priority, rt sched class was
intended to migrate higher priority task to lowest rq via push_rt_tasks,
this BUG will happen here.
With the system running on PREEMPT_RT, rt_spin_lock will disable
migration, this will make the problem easier to reproduce.
I have seen this crash on PREEMPT_RT, from the logs, there is a race
when trying to migrate higher priority tasks to the lowest rq.
Please refer to the following scenarios.
CPU0 CPU1
------------------------------------------------------------------
push_rt_task
check is_migration_disabled(next_task)
task not running and
migration_disabled == 0
find_lock_lowest_rq(next_task, rq);
_double_lock_balance(this_rq, busiest);
raw_spin_rq_unlock(this_rq);
double_rq_lock(this_rq, busiest);
<<wait for busiest rq>>
<wakeup>
task become running
migrate_disable();
<context out>
deactivate_task(rq, next_task, 0);
set_task_cpu(next_task, lowest_rq->cpu);
WARN_ON_ONCE(is_migration_disabled(p));
---------OOPS-------------
Crash logs are as follows:
[123671.996430] WARNING: CPU: 2 PID: 13470 at kernel/sched/core.c:2485
set_task_cpu+0x8c/0x108
[123671.996800] pstate: 20400009 (nzCv daif +PAN -UAO -TCO BTYPE=--)
[123671.996811] pc : set_task_cpu+0x8c/0x108
[123671.996820] lr : set_task_cpu+0x7c/0x108
[123671.996828] sp : ffff80001268bd30
[123671.996832] pmr_save: 00000060
[123671.996835] x29: ffff80001268bd30 x28: ffff0001a3d68e80
[123671.996844] x27: ffff80001225f4a8 x26: ffff800010ab62cb
[123671.996854] x25: ffff80026d95e000 x24: 0000000000000005
[123671.996864] x23: ffff00019746c1b0 x22: 0000000000000000
[123671.996873] x21: ffff00027ee33a80 x20: 0000000000000000
[123671.996882] x19: ffff00019746ba00 x18: 0000000000000000
[123671.996890] x17: 0000000000000000 x16: 0000000000000000
[123671.996899] x15: 000000000000000a x14: 000000000000349e
[123671.996908] x13: ffff800012f4503d x12: 0000000000000001
[123671.996916] x11: 0000000000000000 x10: 0000000000000000
[123671.996925] x9 : 00000000000c0000 x8 : ffff00027ee58700
[123671.996933] x7 : ffff00027ee8da80 x6 : ffff00027ee8e580
[123671.996942] x5 : ffff00027ee8dcc0 x4 : 0000000000000005
[123671.996951] x3 : ffff00027ee8e338 x2 : 0000000000000000
[123671.996959] x1 : 00000000000000ff x0 : 0000000000000002
[123671.996969] Call trace:
[123671.996975] set_task_cpu+0x8c/0x108
[123671.996984] push_rt_task.part.0+0x144/0x184
[123671.996995] push_rt_tasks+0x28/0x3c
[123671.997002] task_woken_rt+0x58/0x68
[123671.997009] ttwu_do_wakeup+0x5c/0xd0
[123671.997019] ttwu_do_activate+0xc0/0xd4
[123671.997028] try_to_wake_up+0x244/0x288
[123671.997036] wake_up_process+0x18/0x24
[123671.997045] __irq_wake_thread+0x64/0x80
[123671.997056] __handle_irq_event_percpu+0x110/0x124
[123671.997064] handle_irq_event_percpu+0x50/0xac
[123671.997072] handle_irq_event+0x84/0xfc
To fix it, we need to check migration_disabled flag again to avoid
bad migration.
Fixes: 95158a89dd50 ("sched,rt: Use the full cpumask for balancing")
Signed-off-by: Schspa Shi <[email protected]>
--
Changelog:
v1 -> v2:
- Modify commit message to add fixed commit information.
- Going to retry to push the current running task on this CPU
away, instead doing nothing for this migrate disabled task.
v2 -> v3:
- Change migration disabled check to the correct position
v3 -> v4:
- Check migrate disabled in find_lock_lowest_rq to avoid not
necessary check when task rq is not released as Steven advised.
v4 -> v5:
- Adjust the comment as Steve advised to make it clear.
v5 -> v6:
- Adjust the comment again as Steve advised.
v6 -> v7:
- Add missing put_task_struct && add this task migration
disable check to deadline scheduler too as Dietmar advised.
---
kernel/sched/deadline.c | 1 +
kernel/sched/rt.c | 4 ++++
2 files changed, 5 insertions(+)
diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index b5152961b743..cb3b886a081c 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -2238,6 +2238,7 @@ static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
!cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
task_running(rq, task) ||
!dl_task(task) ||
+ is_migration_disabled(task) ||
!task_on_rq_queued(task))) {
double_unlock_balance(rq, later_rq);
later_rq = NULL;
diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 8c9ed9664840..7bd3e6ecbe45 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -1998,11 +1998,15 @@ static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
* the mean time, task could have
* migrated already or had its affinity changed.
* Also make sure that it wasn't scheduled on its rq.
+ * It is possible the task was scheduled, set
+ * "migrate_disabled" and then got preempted, so we must
+ * check the task migration disable flag here too.
*/
if (unlikely(task_rq(task) != rq ||
!cpumask_test_cpu(lowest_rq->cpu, &task->cpus_mask) ||
task_running(rq, task) ||
!rt_task(task) ||
+ is_migration_disabled(task) ||
!task_on_rq_queued(task))) {
double_unlock_balance(rq, lowest_rq);
--
2.29.0
Hi Valentin, Steven, Dietmar:
Could you help to review this V7 patch again ?
I have been testing on our platform for more than a month, and it
seems that the system has no adverse reactions.
Thank you very much.
--
BRs
Schspa Shi
On 13/07/2022 15:48, Schspa Shi wrote:
> Commit 95158a89dd50 ("sched,rt: Use the full cpumask for balancing")
> allow find_lock_lowest_rq to pick a task with migration disabled.
> This commit is intended to push the current running task on this CPU
> away.
>
> There is a race scenario, which allows a migration disabled task to
> be migrated to another CPU.
>
> When there is a RT task with higher priority, rt sched class was
> intended to migrate higher priority task to lowest rq via push_rt_tasks,
> this BUG will happen here.
^^^
You mean the warning in set_task_cpu()?
> With the system running on PREEMPT_RT, rt_spin_lock will disable
> migration, this will make the problem easier to reproduce.
>
> I have seen this crash on PREEMPT_RT, from the logs, there is a race
^^^^^
?
We still talking about the set_task_cpu() warning, right?
[...]
Reviewed-by: Dietmar Eggemann <[email protected]>
You should carry the Reviewed-by:'s you got in previous versions.
Dietmar Eggemann <[email protected]> writes:
> On 13/07/2022 15:48, Schspa Shi wrote:
>> Commit 95158a89dd50 ("sched,rt: Use the full cpumask for balancing")
>> allow find_lock_lowest_rq to pick a task with migration disabled.
>> This commit is intended to push the current running task on this CPU
>> away.
>>
>> There is a race scenario, which allows a migration disabled task to
>> be migrated to another CPU.
>>
>> When there is a RT task with higher priority, rt sched class was
>> intended to migrate higher priority task to lowest rq via push_rt_tasks,
>> this BUG will happen here.
> ^^^
>
> You mean the warning in set_task_cpu()?
>
Yes, this is the warning in set_task_cpu()?
void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
{
#ifdef CONFIG_SCHED_DEBUG
...
WARN_ON_ONCE(is_migration_disabled(p));
#endif
...
}
>> With the system running on PREEMPT_RT, rt_spin_lock will disable
>> migration, this will make the problem easier to reproduce.
>>
>> I have seen this crash on PREEMPT_RT, from the logs, there is a race
> ^^^^^
> ?
>
> We still talking about the set_task_cpu() warning, right?
>
Yes.
> [...]
>
> Reviewed-by: Dietmar Eggemann <[email protected]>
>
> You should carry the Reviewed-by:'s you got in previous versions.
Thanks for the reminder, I'll pay attention to this next time.
I will upload a new patch version to carry it. and change the BUG in
comment message to WARN.
--
BRs
Schspa Shi