Skip to contentSkip to conversation

kthread/smpboot: Serialize kthread parking against wakeup

PATCH In review Gaurav Kohli 2 participants 5 messages Apr 24, 2018

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kthread/smpboot: Serialize kthread parking against wakeup

The control cpu thread which initiates hotplug calls kthread_park() for hotplug thread and sets KTHREAD_SHOULD_PARK. After this control thread wakes up the hotplug thread. There is a chance that wakeup code sees the hotplug thread (running on AP core) in INTERRUPTIBLE state, but sets its state to RUNNING after hotplug thread has entered kthread_parkme() and changed its state to TASK_PARKED. This can result in panic later on in kthread_unpark(), as it sees KTHREAD_IS_PARKED flag set but fails to rebind the kthread, due to it being not in TASK_PARKED state. Fix this, by serializing wakeup state change, against state change before parking the kthread.

Below is the possible race:

Control thread Hotplug Thread

kthread_park()
set KTHREAD_SHOULD_PARK

					      smpboot_thread_fn
					      set_current_state(TASK_INTERRUPTIBLE);
					      kthread_parkme

wake_up_process()

raw_spin_lock_irqsave(&p->pi_lock, flags);
if (!(p->state & state)) -> this will fail

            goto out;

					      __kthread_parkme
					       __set_current_state(TASK_PARKED);

if (p->on_rq && ttwu_remote(p, wake_flags))

    ttwu_remote()
        p->state = TASK_RUNNING;
						schedule();

So to avoid this race, take pi_lock to serial state changes.

kernel/smpboot.c +2 -0
index 1650578..514b232 100644
@@ -121,7 +121,9 @@ static int smpboot_thread_fn(void *data)
 		}
 
 		if (kthread_should_park()) {
+			raw_spin_lock(&current->pi_lock);
 			__set_current_state(TASK_RUNNING);
+			raw_spin_unlock(&current->pi_lock);
 			preempt_enable();
 			if (ht->park && td->status == HP_THREAD_ACTIVE) {
 				BUG_ON(td->cpu != smp_processor_id());
Signature · 1 line
1.9.1
#1 lore raw

Hi ,

We can also fix below race by smpboot code as well:

@@ -109,7 +109,6 @@ static int smpboot_thread_fn(void *data)

        struct smp_hotplug_thread *ht = td->ht;

        while (1) {

-               set_current_state(TASK_INTERRUPTIBLE);

                preempt_disable();

                if (kthread_should_stop()) {

                        __set_current_state(TASK_RUNNING);

@@ -157,6 +156,7 @@ static int smpboot_thread_fn(void *data)

                if (!ht->thread_should_run(td->cpu)) {

                        preempt_enable_no_resched();

+                       set_current_state(TASK_INTERRUPTIBLE);

                        schedule();

                } else {

                        __set_current_state(TASK_RUNNING);

Please suggest if this approach is better.

Regards

Gaurav

Quoted 57 lines · Gaurav Kohli

The control cpu thread which initiates hotplug calls kthread_park() for hotplug thread and sets KTHREAD_SHOULD_PARK. After this control thread wakes up the hotplug thread. There is a chance that wakeup code sees the hotplug thread (running on AP core) in INTERRUPTIBLE state, but sets its state to RUNNING after hotplug thread has entered kthread_parkme() and changed its state to TASK_PARKED. This can result in panic later on in kthread_unpark(), as it sees KTHREAD_IS_PARKED flag set but fails to rebind the kthread, due to it being not in TASK_PARKED state. Fix this, by serializing wakeup state change, against state change before parking the kthread.

Below is the possible race:

Control thread Hotplug Thread

kthread_park()
set KTHREAD_SHOULD_PARK

					      smpboot_thread_fn
					      set_current_state(TASK_INTERRUPTIBLE);
					      kthread_parkme

wake_up_process()

raw_spin_lock_irqsave(&p->pi_lock, flags);
if (!(p->state & state)) -> this will fail

             goto out;

					      __kthread_parkme
					       __set_current_state(TASK_PARKED);

if (p->on_rq && ttwu_remote(p, wake_flags))

     ttwu_remote()
         p->state = TASK_RUNNING;
						schedule();

So to avoid this race, take pi_lock to serial state changes.

kernel/smpboot.c +2 -0
index 1650578..514b232 100644
@@ -121,7 +121,9 @@ static int smpboot_thread_fn(void *data)
  		}
 
  		if (kthread_should_park()) {
+			raw_spin_lock(&current->pi_lock);
  			__set_current_state(TASK_RUNNING);
+			raw_spin_unlock(&current->pi_lock);
  			preempt_enable();
  			if (ht->park && td->status == HP_THREAD_ACTIVE) {
  				BUG_ON(td->cpu != smp_processor_id());
Signature · 2 lines
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project.
maintainer
#2 lore raw
Quoted 10 lines · Gaurav
(quoted diff) +1 -0
@@ -157,6 +156,7 @@ static int smpboot_thread_fn(void *data)
 
                 if (!ht->thread_should_run(td->cpu)) {
 
                         preempt_enable_no_resched();
 
+                       set_current_state(TASK_INTERRUPTIBLE);
 

                        schedule();

Quoted 5 lines

                } else {

                        __set_current_state(TASK_RUNNING);

Please suggest if this approach is better.

Bah, my brain isn't working... see below for the 'correct' version of your second patch.

But this violates the normal pattern; see the comment near set_current_state(). That pattern ensures the thread either sees the wakeup condition or the actual wakeup.

I'm thinking that with this patch there is a scenario where we'll miss both the kthread_should_park() and the actual wakeup and end up not doing anything.

I do the like the end result, but I suspect it's buggy.

---
 kernel/smpboot.c | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)
kernel/smpboot.c +1 -6
index 5043e7433f4b..5bdf57f2ce68 100644
@@ -109,10 +109,8 @@ static int smpboot_thread_fn(void *data)
 	struct smp_hotplug_thread *ht = td->ht;
 
 	while (1) {
-		set_current_state(TASK_INTERRUPTIBLE);
 		preempt_disable();
 		if (kthread_should_stop()) {
-			__set_current_state(TASK_RUNNING);
 			preempt_enable();
 			/* cleanup must mirror setup */
 			if (ht->cleanup && td->status != HP_THREAD_NONE)
@@ -122,7 +120,6 @@ static int smpboot_thread_fn(void *data)
 		}
 
 		if (kthread_should_park()) {
-			__set_current_state(TASK_RUNNING);
 			preempt_enable();
 			if (ht->park && td->status == HP_THREAD_ACTIVE) {
 				BUG_ON(td->cpu != smp_processor_id());
@@ -139,7 +136,6 @@ static int smpboot_thread_fn(void *data)
 		/* Check for state change setup */
 		switch (td->status) {
 		case HP_THREAD_NONE:
-			__set_current_state(TASK_RUNNING);
 			preempt_enable();
 			if (ht->setup)
 				ht->setup(td->cpu);
@@ -147,7 +143,6 @@ static int smpboot_thread_fn(void *data)
 			continue;
 
 		case HP_THREAD_PARKED:
-			__set_current_state(TASK_RUNNING);
 			preempt_enable();
 			if (ht->unpark)
 				ht->unpark(td->cpu);
@@ -156,10 +151,10 @@ static int smpboot_thread_fn(void *data)
 		}
 
 		if (!ht->thread_should_run(td->cpu)) {
+			set_current_state(TASK_IDLE);
 			preempt_enable_no_resched();
 			schedule();
 		} else {
-			__set_current_state(TASK_RUNNING);
 			preempt_enable();
 			ht->thread_fn(td->cpu);
 		}
maintainer
#3 lore raw
In reply to Gaurav Kohli
Quoted 47 lines · Gaurav Kohli

The control cpu thread which initiates hotplug calls kthread_park() for hotplug thread and sets KTHREAD_SHOULD_PARK. After this control thread wakes up the hotplug thread. There is a chance that wakeup code sees the hotplug thread (running on AP core) in INTERRUPTIBLE state, but sets its state to RUNNING after hotplug thread has entered kthread_parkme() and changed its state to TASK_PARKED. This can result in panic later on in kthread_unpark(), as it sees KTHREAD_IS_PARKED flag set but fails to rebind the kthread, due to it being not in TASK_PARKED state. Fix this, by serializing wakeup state change, against state change before parking the kthread.

Below is the possible race:

Control thread Hotplug Thread

kthread_park()
set KTHREAD_SHOULD_PARK

					      smpboot_thread_fn
					      set_current_state(TASK_INTERRUPTIBLE);
					      kthread_parkme

wake_up_process()

raw_spin_lock_irqsave(&p->pi_lock, flags);
if (!(p->state & state)) -> this will fail

            goto out;

					      __kthread_parkme
					       __set_current_state(TASK_PARKED);

if (p->on_rq && ttwu_remote(p, wake_flags))

    ttwu_remote()
        p->state = TASK_RUNNING;
						schedule();

So to avoid this race, take pi_lock to serial state changes.

kernel/smpboot.c +0 -0
index 1650578..514b232 100644

kernel/smpboot.c static int smpboot_thread_fn(void *data) · line 129

Quoted 9 lines
kernel/smpboot.c +2 -0
@@ -121,7 +121,9 @@ static int smpboot_thread_fn(void *data)
 		}
 
 		if (kthread_should_park()) {
+			raw_spin_lock(&current->pi_lock);
 			__set_current_state(TASK_RUNNING);
+			raw_spin_unlock(&current->pi_lock);
 			preempt_enable();
 			if (ht->park && td->status == HP_THREAD_ACTIVE) {
 				BUG_ON(td->cpu != smp_processor_id());

Note how in your scenario above you didn't actually need the TASK_RUNNING state; so how is this change going to fix anything?

But yes, I suspect it is right, but it definitely needs a comment explaining wth we take that lock there.

Like I said earlier, my brain is entirely fried for the day; but I'll have a try tomorrow.

#4 lore raw
Quoted 50 lines · Peter Zijlstra
Quoted 49 lines · Gaurav Kohli

The control cpu thread which initiates hotplug calls kthread_park() for hotplug thread and sets KTHREAD_SHOULD_PARK. After this control thread wakes up the hotplug thread. There is a chance that wakeup code sees the hotplug thread (running on AP core) in INTERRUPTIBLE state, but sets its state to RUNNING after hotplug thread has entered kthread_parkme() and changed its state to TASK_PARKED. This can result in panic later on in kthread_unpark(), as it sees KTHREAD_IS_PARKED flag set but fails to rebind the kthread, due to it being not in TASK_PARKED state. Fix this, by serializing wakeup state change, against state change before parking the kthread.

Below is the possible race:

Control thread Hotplug Thread

kthread_park()
set KTHREAD_SHOULD_PARK

					      smpboot_thread_fn
					      set_current_state(TASK_INTERRUPTIBLE);
					      kthread_parkme

wake_up_process()

raw_spin_lock_irqsave(&p->pi_lock, flags);
if (!(p->state & state)) -> this will fail

             goto out;

					      __kthread_parkme
					       __set_current_state(TASK_PARKED);

if (p->on_rq && ttwu_remote(p, wake_flags))

     ttwu_remote()
         p->state = TASK_RUNNING;
						schedule();

So to avoid this race, take pi_lock to serial state changes.

kernel/smpboot.c +0 -0
index 1650578..514b232 100644
@@ -121,7 +121,9 @@ static int smpboot_thread_fn(void *data)
  		}
 
Quoted 9 lines

kernel/smpboot.c static int smpboot_thread_fn(void *data) · line 129

Quoted 7 lines
kernel/smpboot.c +2 -0
  		if (kthread_should_park()) {
+			raw_spin_lock(&current->pi_lock);
  			__set_current_state(TASK_RUNNING);
+			raw_spin_unlock(&current->pi_lock);
  			preempt_enable();
  			if (ht->park && td->status == HP_THREAD_ACTIVE) {
  				BUG_ON(td->cpu != smp_processor_id());

Note how in your scenario above you didn't actually need the TASK_RUNNING state; so how is this change going to fix anything?

Hi Peter,

As with help of this , if kthread_should_park run first so wake_up call of controller

get exited as task is already set as running, otherwise if controller runs first

then we will block here and set running and then sets TASK_PARKED .

So no chance of cpuhp set as running during  kthread_parkme call.

But as we discussed this can be fix by 2nd patch as well, So once you get time and able to

see , Please let us know or do you want me to try your 2nd patch

for testing first?

Quoted 7 lines

But yes, I suspect it is right, but it definitely needs a comment explaining wth we take that lock there.

Like I said earlier, my brain is entirely fried for the day; but I'll have a try tomorrow.

Signature · 2 lines
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project.