sched/dl: Fix race between dl_task_timer() and sched_setaffinity()
Subsystems (2)
PATCH 1/1 sched/dl: Fix race between dl_task_timer() and sched_setaffinity()
The race is in unlocked task_rq() access. In pair with parallel call of sched_setaffinity() it may be a reason of corruption of internal rq's data.
--- kernel/sched/deadline.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +8 -1
@@ -513,9 +513,16 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) struct sched_dl_entity, dl_timer);struct task_struct *p =dl_task_of (dl_se); -struct rq *rq =task_rq (p); +struct rq *rq; +again: + rq =task_rq (p);raw_spin_lock (&rq->lock); +if (unlikely (rq !=task_rq (p))) { +raw_spin_unlock (&rq->lock); + goto again; + } +/* * We need to take care of a possible races here. In fact, the * task might have changed its scheduling policy to something
Quoted 4 lines · On Sat, 17 May 2014 01:30:03 +0400 Kirill Tkhai <[email protected]> wrote:
The race is in unlocked
task_rq()access. In pair with parallel call ofsched_setaffinity()it may be a reason of corruption of internal rq's data.
Sure, the thing can happen!
Quoted 14 lines
Signed-off-by Kirill Tkhai <[email protected]>CC Juri Lelli <[email protected]>CC Peter Zijlstra <[email protected]>CC Ingo Molnar <[email protected]>Cc <[email protected]> # v3.14--- kernel/sched/deadline.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)kernel/sched/deadline.c +0
-0
kernel/sched/deadline.c static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) · line 516
Quoted 5 lines
kernel/sched/deadline.c +1
-1@@ -513,9 +513,16 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) struct sched_dl_entity, dl_timer);struct task_struct *p =dl_task_of (dl_se); -struct rq *rq =task_rq (p); +struct rq *rq;
We could maybe add a comment here, in line with what we have below, to document why we need this.
Thanks,
- Juri
Quoted 13 lines
kernel/sched/deadline.c +7
-0+again: + rq =task_rq (p);raw_spin_lock (&rq->lock); +if (unlikely (rq !=task_rq (p))) { +raw_spin_unlock (&rq->lock); + goto again; + } +/* * We need to take care of a possible races here. In fact, the * task might have changed its scheduling policy to something
19.05.2014, 17:11, "Juri Lelli" <[email protected]>:
Quoted 8 lines
Quoted 3 lines · On Sat, 17 May 2014 01:30:03 +0400 Kirill Tkhai <[email protected]> wrote:
О©╫The race is in unlocked task_rq() access. In pair with parallel О©╫call of sched_setaffinity() it may be a reason of corruption О©╫of internal rq's data.Sure, the thing can happen!
[snipped]
Quoted 9 lines
Quoted 6 lines
О©╫@@ -513,9 +513,16 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct sched_dl_entity, О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫dl_timer); О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct task_struct *p = dl_task_of(dl_se); О©╫- struct rq *rq = task_rq(p); О©╫+ struct rq *rq;We could maybe add a comment here, in line with what we have below, to document why we need this.
How about this? (I added comment and rewrote changelog).
[PATCH] sched/dl: Fix race between dl_task_timer() and sched_setaffinity()
Throttled task is still on rq, and it may be moved to other cpu if user is playing with sched_setaffinity(). Therefore, unlocked task_rq() access makes the race.
To fix that we do the same as made in __task_rq_lock(). We do not use __task_rq_lock() itself, because it has a useful lockdep check, which is not correct in case of dl_task_timer(). This case is an exception.
kernel/sched/deadline.c +9 -1
@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) struct sched_dl_entity, dl_timer);struct task_struct *p =dl_task_of (dl_se); -struct rq *rq =task_rq (p); +struct rq *rq; +again: + rq =task_rq (p);raw_spin_lock (&rq->lock); +if (unlikely (rq !=task_rq (p))) { +/* Task was moved, retrying. */ +raw_spin_unlock (&rq->lock); + goto again; + } +/* * We need to take care of a possible races here. In fact, the * task might have changed its scheduling policy to something
Peter Zijlstra
maintainerQuoted 6 lines · Kirill Tkhai
kernel/sched/deadline.c +1
-1@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) struct sched_dl_entity, dl_timer);struct task_struct *p =dl_task_of (dl_se); -struct rq *rq =task_rq (p); +struct rq *rq;
kernel/sched/deadline.c static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) · line 526
Quoted 10 lines
kernel/sched/deadline.c +8
-0+again: + rq =task_rq (p);raw_spin_lock (&rq->lock); +if (unlikely (rq !=task_rq (p))) { +/* Task was moved, retrying. */ +raw_spin_unlock (&rq->lock); + goto again; + } +
That thing is called: rq = __task_rq_lock(p);
20.05.2014, 04:00, "Peter Zijlstra" <[email protected]>:
Quoted 12 lines · Kirill Tkhai
Quoted 11 lines
О©╫@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct sched_dl_entity, О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫dl_timer); О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct task_struct *p = dl_task_of(dl_se); О©╫- struct rq *rq = task_rq(p); О©╫+ struct rq *rq; О©╫+again: О©╫+ rq = task_rq(p); О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫raw_spin_lock(&rq->lock);
Quoted 8 lines
Quoted 6 lines
О©╫+ if (unlikely(rq != task_rq(p))) { О©╫+ /* Task was moved, retrying. */ О©╫+ raw_spin_unlock(&rq->lock); О©╫+ goto again; О©╫+ } О©╫+That thing is called: rq =
__task_rq_lock(p);
But p->pi_lock is not held. The problem is __task_rq_lock() has lockdep assert. Should we change it?
Kirill
20.05.2014, 09:08, "Kirill Tkhai" <[email protected]>:
Quoted 19 lines
20.05.2014, 04:00, "Peter Zijlstra" <[email protected]>:
Quoted 17 lines
О©╫On Mon, May 19, 2014 at 11:31:19PM +0400, Kirill Tkhai wrote: > О©╫О©╫@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) > О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct sched_dl_entity, > О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫dl_timer); > О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct task_struct *p = dl_task_of(dl_se); > О©╫О©╫- struct rq *rq = task_rq(p); > О©╫О©╫+ struct rq *rq; > О©╫О©╫+again: > О©╫О©╫+ rq = task_rq(p); > О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫raw_spin_lock(&rq->lock);Quoted 7 lines
О©╫О©╫+ if (unlikely(rq != task_rq(p))) { О©╫О©╫+ /* Task was moved, retrying. */ О©╫О©╫+ raw_spin_unlock(&rq->lock); О©╫О©╫+ goto again; О©╫О©╫+ } О©╫О©╫+
Quoted 4 lines
Quoted 1 line
О©╫That thing is called: rq =
__task_rq_lock(p);But
p->pi_lockis not held. The problem is__task_rq_lock()has lockdep assert. Should we change it?
Or make something like this?
static inline struct rq *_task_rq_lock(struct task_struct *p)
{
lockdep_assert_held(&p->pi_lock);
return __task_rq_lock(p);
}Thanks!
Peter Zijlstra
maintainerQuoted 22 lines · Kirill Tkhai
20.05.2014, 04:00, "Peter Zijlstra" <[email protected]>:
Quoted 19 lines · Kirill Tkhai
Quoted 17 lines
�@@ -513,9 +513,17 @@ static enum
hrtimer_restartdl_task_timer(struct hrtimer *timer)�������������������������������������������������������structsched_dl_entity, �������������������������������������������������������dl_timer); ����������structtask_struct*p =dl_task_of(dl_se);
�- struct rq *rq =task_rq(p);
�+ struct rq *rq;
�+again:
�+ rq =task_rq(p);
����������raw_spin_lock(&rq->lock);�+ if (
unlikely(rq != task_rq(p))) {
�+ /* Task was moved, retrying. */
�+raw_spin_unlock(&rq->lock);
�+ goto again;
�+ }
�+
Quoted 4 lines
Quoted 1 line
That thing is called: rq =
__task_rq_lock(p);But
p->pi_lockis not held. The problem is__task_rq_lock()has lockdep assert. Should we change it?
Ok, so now that I'm awake ;-)
So the trivial problem as described by your initial changelog isn't right, because we cannot call sched_setaffinity() on deadline tasks, or rather we can, but we can't actually change the affinity mask.
Now I suppose the problem can still actually happen when you change the root domain and trigger a effective affinity change that way.
That said, no leave it as you proposed, adding a *task_rq_lock() variant without lockdep assert in will only confuse things, as normally we really should be also taking ->pi_lock.
The only reason we don't strictly need ->pi_lock now is because we're guaranteed to have p->state == TASK_RUNNING here and are thus free of ttwu races.
Hi,
Quoted 25 lines · On Tue, 20 May 2014 09:53:15 +0200 Peter Zijlstra <[email protected]> wrote:
Quoted 24 lines · Kirill Tkhai
20.05.2014, 04:00, "Peter Zijlstra" <[email protected]>:
Quoted 20 lines · Kirill Tkhai
Quoted 11 lines
@@ -513,9 +513,17 @@ static enum
hrtimer_restartdl_task_timer(struct hrtimer *timer)structsched_dl_entity,dl_timer);
structtask_struct*p =dl_task_of(dl_se);
- struct rq *rq =task_rq(p);
+ struct rq *rq;
+again:
+ rq =task_rq(p);
raw_spin_lock(&rq->lock);Quoted 6 lines
+ if (
unlikely(rq != task_rq(p))) {
+ /* Task was moved, retrying. */
+raw_spin_unlock(&rq->lock);
+ goto again;
+ }
+That thing is called: rq =
__task_rq_lock(p);
Quoted 9 lines
Quoted 2 lines
But
p->pi_lockis not held. The problem is__task_rq_lock()has lockdep assert. Should we change it?Ok, so now that I'm awake ;-)
So the trivial problem as described by your initial changelog isn't right, because we cannot call
sched_setaffinity()on deadline tasks, or rather we can, but we can't actually change the affinity mask.
Well, if we disable AC we can. And I was able to recreate that race in that case.
Quoted 3 lines
Now I suppose the problem can still actually happen when you change the root domain and trigger a effective affinity change that way.
Yeah, I think here too.
Quoted 7 lines
That said, no leave it as you proposed, adding a
*task_rq_lock()variant without lockdep assert in will only confuse things, as normally we really should be also taking ->pi_lock.The only reason we don't strictly need ->pi_lock now is because we're guaranteed to have
p->state==TASK_RUNNINGhere and are thus free of ttwu races.
Maybe we could add this as part of the comment.
Thanks,
- Juri
20.05.2014, 12:16, "Juri Lelli" <[email protected]>:
Quoted 45 lines
Hi,
Quoted 19 lines · On Tue, 20 May 2014 09:53:15 +0200 Peter Zijlstra <[email protected]> wrote:
О©╫On Tue, May 20, 2014 at 09:08:53AM +0400, Kirill Tkhai wrote: > О©╫20.05.2014, 04:00, "Peter Zijlstra" <[email protected]>: >> О©╫On Mon, May 19, 2014 at 11:31:19PM +0400, Kirill Tkhai wrote: >>> О©╫О©╫@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) >>> О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct sched_dl_entity, >>> О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫dl_timer); >>> О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫struct task_struct *p = dl_task_of(dl_se); >>> О©╫О©╫- struct rq *rq = task_rq(p); >>> О©╫О©╫+ struct rq *rq; >>> О©╫О©╫+again: >>> О©╫О©╫+ rq = task_rq(p); >>> О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫О©╫raw_spin_lock(&rq->lock);Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
О©╫О©╫+ if (unlikely(rq != task_rq(p))) { О©╫О©╫+ /* Task was moved, retrying. */ О©╫О©╫+ raw_spin_unlock(&rq->lock); О©╫О©╫+ goto again; О©╫О©╫+ } О©╫О©╫+Quoted 8 lines
Quoted 3 lines
Quoted 1 line
О©╫That thing is called: rq =
__task_rq_lock(p);О©╫But p->pi_lock is not held. The problem is __task_rq_lock() has lockdep assert. О©╫Should we change it?О©╫Ok, so now that I'm awake ;-)
О©╫So the trivial problem as described by your initial changelog isn't О©╫right, because we cannot call sched_setaffinity() on deadline tasks, or О©╫rather we can, but we can't actually change the affinity mask.Well, if we disable AC we can. And I was able to recreate that race in that case.
Quoted 2 lines
О©╫Now I suppose the problem can still actually happen when you change the О©╫root domain and trigger a effective affinity change that way.Yeah, I think here too.
Quoted 4 lines
О©╫That said, no leave it as you proposed, adding a *task_rq_lock() variant О©╫without lockdep assert in will only confuse things, as normally we О©╫really should be also taking ->pi_lock.
Quoted 5 lines
Quoted 3 lines
О©╫The only reason we don't strictly need ->pi_lock now is because we're О©╫guaranteed to have p->state == TASK_RUNNING here and are thus free of О©╫ttwu races.Maybe we could add this as part of the comment.
Peter, Juri, thanks for comment. Hope, I understood you right :)
[PATCH] sched/dl: Fix race in dl_task_timer()
Throttled task is still on rq, and it may be moved to other cpu if user is playing with sched_setaffinity(). Therefore, unlocked task_rq() access makes the race.
Juri Lelli reports he got this race when dl_bandwidth_enabled() was not set.
Other thing, pointed by Peter Zijlstra:
"Now I suppose the problem can still actually happen when
you change the root domain and trigger a effective affinity change that way".
To fix that we do the same as made in __task_rq_lock(). We do not use __task_rq_lock() itself, because it has a useful lockdep check, which is not correct in case of dl_task_timer(). We do not need pi_lock locked here. This case is an exception (PeterZ):
"The only reason we don't strictly need ->pi_lock now is because
we're guaranteed to have p->state == TASK_RUNNING here and are thus free of ttwu races".
--- kernel/sched/deadline.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +9 -1
@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) struct sched_dl_entity, dl_timer);struct task_struct *p =dl_task_of (dl_se); -struct rq *rq =task_rq (p); +struct rq *rq; +again: + rq =task_rq (p);raw_spin_lock (&rq->lock); +if (rq !=task_rq (p)) { +/* Task was moved, retrying. */ +raw_spin_unlock (&rq->lock); + goto again; + } +/* * We need to take care of a possible races here. In fact, the * task might have changed its scheduling policy to something
Peter Zijlstra
maintainerQuoted 24 lines · Kirill Tkhai
[PATCH] sched/dl: Fix race in
dl_task_timer()Throttled task is still on rq, and it may be moved to other cpu if user is playing with
sched_setaffinity(). Therefore, unlockedtask_rq()access makes the race.Juri Lelli reports he got this race when
dl_bandwidth_enabled()was not set.Other thing, pointed by Peter Zijlstra:
"Now I suppose the problem can still actually happen when
you change the root domain and trigger a effective affinity change that way".To fix that we do the same as made in
__task_rq_lock(). We do not use__task_rq_lock()itself, because it has a useful lockdep check, which is not correct in case ofdl_task_timer(). We do not needpi_locklocked here. This case is an exception (PeterZ):"The only reason we don't strictly need ->pi_lock now is because
we're guaranteed to havep->state==TASK_RUNNINGhere and are thus free of ttwu races".
Quoted 6 lines
Signed-off-by Kirill Tkhai <[email protected]>CC Juri Lelli <[email protected]>CC Peter Zijlstra <[email protected]>CC Ingo Molnar <[email protected]>Cc <[email protected]> # v3.14---
thanks Kirill!
tip-bot for Kirill Tkhai
merge bot[tip:sched/urgent] sched/dl: Fix race in dl_task_timer()
In reply to Kirill Tkhaisched/dl: Fix race in dl_task_timer()
Commit-ID: 0f397f2c90ce68821ee864c2c53baafe78de765d Gitweb: http://git.kernel.org/tip/0f397f2c90ce68821ee864c2c53baafe78de765d Author: Kirill Tkhai <[email protected]> AuthorDate: Tue, 20 May 2014 13:33:42 +0400 Committer: Ingo Molnar <[email protected]> CommitDate: Thu, 5 Jun 2014 11:51:12 +0200
sched/dl: Fix race in dl_task_timer()
Throttled task is still on rq, and it may be moved to other cpu if user is playing with sched_setaffinity(). Therefore, unlocked task_rq() access makes the race.
Juri Lelli reports he got this race when dl_bandwidth_enabled() was not set.
Other thing, pointed by Peter Zijlstra:
"Now I suppose the problem can still actually happen when
you change the root domain and trigger a effective affinity change that way".
To fix that we do the same as made in __task_rq_lock(). We do not use __task_rq_lock() itself, because it has a useful lockdep check, which is not correct in case of dl_task_timer(). We do not need pi_lock locked here. This case is an exception (PeterZ):
"The only reason we don't strictly need ->pi_lock now is because
we're guaranteed to have p->state == TASK_RUNNING here and are thus free of ttwu races".
--- kernel/sched/deadline.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +9 -1
@@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) struct sched_dl_entity, dl_timer);struct task_struct *p =dl_task_of (dl_se); -struct rq *rq =task_rq (p); +struct rq *rq; +again: + rq =task_rq (p);raw_spin_lock (&rq->lock); +if (rq !=task_rq (p)) { +/* Task was moved, retrying. */ +raw_spin_unlock (&rq->lock); + goto again; + } +/* * We need to take care of a possible races here. In fact, the * task might have changed its scheduling policy to something