sched_ext: Fix NULL pointer dereferences in put_prev_task_scx
PATCH 1/1 sched_ext: Fix NULL pointer dereferences in put_prev_task_scx
As put_prev_task can be used in other kernel modules which can lead to a NULL pointer. Fix this by checking for a valid next.
--- kernel/sched/ext.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/sched/ext.c +1 -1
@@ -3262,7 +3262,7 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,* ops.enqueue() that @p is the only one available for this cpu, * which should trigger an explicit follow-up scheduling event. */ -if (sched_class_above (&ext_sched_class, next->sched_class)) { +if (next &&sched_class_above (&ext_sched_class, next->sched_class)) {WARN_ON_ONCE (!static_branch_unlikely (&scx_ops_enq_last));do_enqueue_task (rq, p,SCX_ENQ_LAST ,-1 ); }else {
Signature · 1 line
2.17.1
Quoted 2 lines · liuwenfang
As
put_prev_taskcan be used in other kernel modules which can lead to a NULL pointer. Fix this by checking for a valid next.
Actually, put_prev_task() should be used only within kernel/sched/ and, in theory, you should have done a dequeue_task() before put_prev_task() in this scenario, so SCX_TASK_QUEUED shouldn't be set in p->scx.flags.
The change might still make sense, but can you clarify how you triggered the NULL pointer dereference?
Thanks,
-Andrea
Quoted 21 lines
Signed-off-by l00013971 <[email protected]>--- kernel/sched/ext.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)kernel/sched/ext.c +1
-1@@ -3262,7 +3262,7 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,* ops.enqueue() that @p is the only one available for this cpu, * which should trigger an explicit follow-up scheduling event. */ -if (sched_class_above (&ext_sched_class, next->sched_class)) { +if (next &&sched_class_above (&ext_sched_class, next->sched_class)) {WARN_ON_ONCE (!static_branch_unlikely (&scx_ops_enq_last));do_enqueue_task (rq, p,SCX_ENQ_LAST ,-1 ); }else {Signature · 1 line
2.17.1
Thanks for your feedback.
This is triggered in kernel modules developed In the mobile scenario:
tasks on rq are migrated while the current cpu should be halted for power saving policy. Its migration logic:drain_rq_cpu_stop -- migrate_all_tasks :
for (;;) {
if (rq->nr_running == 1)
break;
for_each_class(class) {
next = class->pick_next_task(rq);
if (next) {
next->sched_class->put_prev_task(rq, next, NULL);
break;
}
}
if (is_idle_task(next))
break;
dest_cpu = select_task_rq(next...);
move_queued_task(rq, rf, next, dest_cpu);
...
}put_prev_task in this function is selected to update util and statistics info for each runnable tasks, here they are not dequeued yet.
Best regards,
Quoted 37 lines · liuwenfang
Quoted 2 lines
As
put_prev_taskcan be used in other kernel modules which can lead to a NULL pointer. Fix this by checking for a valid next.Actually,
put_prev_task()should be used only withinkernel/sched/and, in theory, you should have done adequeue_task()beforeput_prev_task()in this scenario, soSCX_TASK_QUEUEDshouldn't be set inp->scx.flags.The change might still make sense, but can you clarify how you triggered the NULL pointer dereference?
Thanks,
-Andreakernel/sched/ext.c static void put_prev_task_scx(struct rq *rq, struct
Quoted 11 lines
Signed-off-by l00013971 <[email protected]>--- kernel/sched/ext.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)kernel/sched/ext.c index renamed from kernel/sched/ext.c +0
-0
f5133249f..6a579babd100644kernel/sched/ext.c +0
-0@@ -3262,7 +3262,7 @@ static void put_prev_task_scx(struct rq *rq, struct
task_struct*p,Quoted 11 lines
kernel/sched/ext.c +2
-2* ops.enqueue() that @p is the only one available for this cpu, * which should trigger an explicit follow-up scheduling event. */ -if (sched_class_above (&ext_sched_class, next->sched_class)) { +if (next &&sched_class_above (&ext_sched_class, next->sched_class)) +{WARN_ON_ONCE (!static_branch_unlikely (&scx_ops_enq_last));do_enqueue_task (rq, p,SCX_ENQ_LAST ,-1 ); }else { --2.17.1
Hi liuwenfang,
Quoted 5 lines · liuwenfang
Thanks for your feedback.
This is triggered in kernel modules developed In the mobile scenario:
tasks on rq are migrated while the current cpu should be halted for power saving policy. Its migration logic:drain_rq_cpu_stop--migrate_all_tasks:
Quoted 7 lines
for (;;) {
if (rq->nr_running == 1)
break;
for_each_class(class) {
next = class->pick_next_task(rq);
if (next) {
next->sched_class->put_prev_task(rq, next, NULL);Can you do something like this instead?
next->sched_class->put_prev_task(rq, next, next);
this should give you the same behavior you're relying on, without requiring the extra check in ext.c.
Quoted 10 lines
break; } } if (is_idle_task(next)) break; dest_cpu = select_task_rq(next...); move_queued_task(rq, rf, next, dest_cpu); ... }
Quoted 2 lines
put_prev_taskin this function is selected to update util and statistics info for each runnable tasks, here they are not dequeued yet.
I see, so it's an optimization/workaround to update utilization info without fully dequeuing the tasks. Is this out-of-tree code, I guess?
When you say the CPU is halted, we're not talking about CPU hotplugging, right? We're just migrating tasks off the CPU?
Also, if you're running sched_ext there are ways to exclude certain CPUs at the scheduler's level via ops.select_cpu() / ops.dispatch(). Do you think this could be a viable solution for your specific use case?
Thanks,
-Andrea
PS https://subspace.kernel.org/etiquette.html#do-not-top-post-when-replying
Quoted 40 lines
Best regards,
Quoted 37 lines · liuwenfang
Quoted 2 lines
As
put_prev_taskcan be used in other kernel modules which can lead to a NULL pointer. Fix this by checking for a valid next.Actually,
put_prev_task()should be used only withinkernel/sched/and, in theory, you should have done adequeue_task()beforeput_prev_task()in this scenario, soSCX_TASK_QUEUEDshouldn't be set inp->scx.flags.The change might still make sense, but can you clarify how you triggered the NULL pointer dereference?
Thanks,
-Andreakernel/sched/ext.c static void put_prev_task_scx(struct rq *rq, struct
Quoted 11 lines
Signed-off-by l00013971 <[email protected]>--- kernel/sched/ext.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)kernel/sched/ext.c index renamed from kernel/sched/ext.c +0
-0
f5133249f..6a579babd100644kernel/sched/ext.c +0
-0@@ -3262,7 +3262,7 @@ static void put_prev_task_scx(struct rq *rq, struct
task_struct*p,Quoted 11 lines
kernel/sched/ext.c +2
-2* ops.enqueue() that @p is the only one available for this cpu, * which should trigger an explicit follow-up scheduling event. */ -if (sched_class_above (&ext_sched_class, next->sched_class)) { +if (next &&sched_class_above (&ext_sched_class, next->sched_class)) +{WARN_ON_ONCE (!static_branch_unlikely (&scx_ops_enq_last));do_enqueue_task (rq, p,SCX_ENQ_LAST ,-1 ); }else { --2.17.1
Quoted 18 lines
Hi liuwenfang,
Quoted 3 lines · liuwenfang
Thanks for your feedback.
This is triggered in kernel modules developed In the mobile scenario:
tasks on rq are migrated while the current cpu should be halted for powersaving policy.
Quoted 9 lines
Its migration logic:
drain_rq_cpu_stop--migrate_all_tasks:for (;;) { if (rq->nr_running == 1) break; for_each_class(class) { next = class->pick_next_task(rq); if (next) { next->sched_class->put_prev_task(rq, next, NULL);
Quoted 6 lines
Can you do something like this instead?
next->sched_class->put_prev_task(rq, next, next);this should give you the same behavior you're relying on, without requiring the extra check in
ext.c.
Yes, this can help me.
Quoted 11 lines
Quoted 11 lines
break; } } if (is_idle_task(next)) break; dest_cpu = select_task_rq(next...); move_queued_task(rq, rf, next, dest_cpu); ... }
Quoted 6 lines
Quoted 2 lines
put_prev_taskin this function is selected to update util and statistics info for each runnable tasks, here they are not dequeued yet.I see, so it's an optimization/workaround to update utilization info without fully dequeuing the tasks. Is this out-of-tree code, I guess?
Yes, this is out-of-tree code.
Quoted 3 lines
When you say the CPU is halted, we're not talking about CPU hotplugging, right? We're just migrating tasks off the CPU?
Yes, this is one power-saving method to control tasks placement by bypassing Certain CPUs in software, simply letting the halted CPU enter an idle state. It replaces CPU hotplugging to reduce costs in mobile gaming scenario.
Quoted 4 lines
Also, if you're running
sched_extthere are ways to exclude certain CPUs at the scheduler's level viaops.select_cpu()/ops.dispatch(). Do you think this could be a viable solution for your specific use case?
Yes, Thanks very much, This is exactly the approach we have implemented.
Best regards
Quoted 48 lines
Thanks,
-AndreaPS https://subspace.kernel.org/etiquette.html#do-not-top-post-when-replying
Quoted 10 lines
Best regards,
Quoted 7 lines · liuwenfang
Quoted 2 lines
As
put_prev_taskcan be used in other kernel modules which can lead to a NULL pointer. Fix this by checking for a valid next.Actually,
put_prev_task()should be used only withinkernel/sched/and, in theory, you should have done adequeue_task()beforeput_prev_task()in this scenario, soSCX_TASK_QUEUEDshouldn't be set in
p->scx.flags.Quoted 20 lines
Quoted 20 lines
The change might still make sense, but can you clarify how you triggered the NULL pointer dereference?
Thanks,
-AndreaQuoted 11 lines
Signed-off-by l00013971 <[email protected]>--- kernel/sched/ext.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)kernel/sched/ext.c index renamed from kernel/sched/ext.c +0
-0
f5133249f..6a579babd100644kernel/sched/ext.c +0
-0@@ -3262,7 +3262,7 @@ static void put_prev_task_scx(struct rq *rq,Quoted 1 line
struct
task_struct*p,Quoted 7 lines
Quoted 7 lines
kernel/sched/ext.c static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
Quoted 7 lines
kernel/sched/ext.c +2
-1* ops.enqueue() that @p is the only one available for this cpu, * which should trigger an explicit follow-up scheduling event. */ -if (sched_class_above (&ext_sched_class, next->sched_class)) { +if (next &&sched_class_above (&ext_sched_class, +next->sched_class)) {WARN_ON_ONCE(!static_branch_unlikely(&scx_ops_enq_last));Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
kernel/sched/ext.c +0
-1do_enqueue_task (rq, p,SCX_ENQ_LAST ,-1 ); }else { --2.17.1