Skip to contentSkip to conversation

cpuhp: fix some st->target issues

PATCH Superseded Phil Auld 2 participants 9 messages May 26, 2022 – Jun 1, 2022 CPU hotplug

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Phil Auld

author
#0 lore raw

v2 cover cpuhp: fix some st->target issues

Several small fixes that clean up some cpuhp inconsistencies. The first prevents target_store() from calling cpu_down() when target == state which prevents the cpu being incorrectly marked as dying. The second just makes the boot cpu have a valid cpuhp target rather than 0 (CPU_OFFLINE) while being in state CPU_ONLINE.

A further issue which these two patches don't address is that the cpuX/online file looks at the device->offline state and can thus get out of sync with the actual cpuhp state if the cpuhp target is used to change state.

Cheers,
Phil

Phil Auld (2):
cpuhp: make target_store() a nop when target == state
cpuhp: Set cpuhp target for boot cpu

 kernel/cpu.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
Signature · 1 line
2.18.0
author
#1 lore raw

PATCH v2 1/2 cpuhp: make target_store() a nop when target == state

writing the current state back in hotplug/target calls cpu_down() which will set cpu dying even when it isn't and then nothing will ever clear it. A stress test that reads values and writes them back for all cpu device files in sysfs will trigger the BUG() in select_fallback_rq once all cpus are marked as dying.

kernel/cpu.c::target_store()

	...
        if (st->state < target)
                ret = cpu_up(dev->id, target);
        else
                ret = cpu_down(dev->id, target);

cpu_down() -> cpu_set_state()

	 bool bringup = st->state < target;
	 ...
	 if (cpu_dying(cpu) != !bringup)
		set_cpu_dying(cpu, !bringup);

Fix this by letting state==target fall through in the target_store() conditional.

---
 kernel/cpu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/cpu.c +1 -1
index d0a9aa0b42e8..cdb6ac10ad94 100644
@@ -2315,7 +2315,7 @@ static ssize_t target_store(struct device *dev, struct device_attribute *attr,
 
 	if (st->state < target)
 		ret = cpu_up(dev->id, target);
-	else
+	else if (st->state > target)
 		ret = cpu_down(dev->id, target);
 out:
 	unlock_device_hotplug();
Signature · 1 line
2.18.0
Reviewed-by on v2 1/2
#2 lore raw

Re: [PATCH v2 1/2] cpuhp: make target_store() a nop when target == state

Quoted 13 lines · Phil Auld

writing the current state back in hotplug/target calls cpu_down() which will set cpu dying even when it isn't and then nothing will ever clear it. A stress test that reads values and writes them back for all cpu device files in sysfs will trigger the BUG() in select_fallback_rq once all cpus are marked as dying.

kernel/cpu.c::target_store()

	...
        if (st->state < target)
                ret = cpu_up(dev->id, target);
        else
                ret = cpu_down(dev->id, target);
Quoted 9 lines

cpu_down() -> cpu_set_state()

	 bool bringup = st->state < target;
	 ...
	 if (cpu_dying(cpu) != !bringup)
		set_cpu_dying(cpu, !bringup);

Fix this by letting state==target fall through in the target_store() conditional.

To go back on my data race paranoia: writes to both cpu$x/online and cpu$x/hotplug/target are serialized by device_hotplug_lock, and so are the exported kernel hotplug functions ({add, remove}_cpu()).

That's not cpu_add_remove_lock as I was looking for, but that's still all under one lock, so I think we're good. Sorry for that!

Quoted 1 line
Quoted 19 lines
---
 kernel/cpu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/cpu.c +1 -1
index d0a9aa0b42e8..cdb6ac10ad94 100644
@@ -2315,7 +2315,7 @@ static ssize_t target_store(struct device *dev, struct device_attribute *attr,
 
 	if (st->state < target)
 		ret = cpu_up(dev->id, target);
-	else
+	else if (st->state > target)
 		ret = cpu_down(dev->id, target);
 out:
 	unlock_device_hotplug();
Signature · 1 line
2.18.0
author
#3 lore raw

Re: [PATCH v2 1/2] cpuhp: make target_store() a nop when target == state

Quoted 24 lines · On Fri, May 27, 2022 at 10:38:24AM +0100 Valentin Schneider wrote:
Quoted 13 lines · Phil Auld

writing the current state back in hotplug/target calls cpu_down() which will set cpu dying even when it isn't and then nothing will ever clear it. A stress test that reads values and writes them back for all cpu device files in sysfs will trigger the BUG() in select_fallback_rq once all cpus are marked as dying.

kernel/cpu.c::target_store()

	...
        if (st->state < target)
                ret = cpu_up(dev->id, target);
        else
                ret = cpu_down(dev->id, target);
Quoted 9 lines

cpu_down() -> cpu_set_state()

	 bool bringup = st->state < target;
	 ...
	 if (cpu_dying(cpu) != !bringup)
		set_cpu_dying(cpu, !bringup);

Fix this by letting state==target fall through in the target_store() conditional.

Quoted 7 lines

To go back on my data race paranoia: writes to both cpu$x/online and cpu$x/hotplug/target are serialized by device_hotplug_lock, and so are the exported kernel hotplug functions ({add, remove}_cpu()).

That's not cpu_add_remove_lock as I was looking for, but that's still all under one lock, so I think we're good. Sorry for that!

Right. This catches it up higher so that we don't get into the code that starts actually changing things. I wonder now in the state == target case if we should make sure st->target == target. With the second patch it's less likely to be needed. Thoughts?

Maybe I'll include that if/when I have code to keep cpux/online in sync with st->state and cpu_online_mask.

Quoted 3 lines
Quoted 1 line

Thanks!

Cheers,
Phil

Quoted 21 lines
Quoted 20 lines
---
 kernel/cpu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/cpu.c +1 -1
index d0a9aa0b42e8..cdb6ac10ad94 100644
@@ -2315,7 +2315,7 @@ static ssize_t target_store(struct device *dev, struct device_attribute *attr,
 
 	if (st->state < target)
 		ret = cpu_up(dev->id, target);
-	else
+	else if (st->state > target)
 		ret = cpu_down(dev->id, target);
 out:
 	unlock_device_hotplug();
Signature · 1 line
2.18.0
#4 lore raw

Re: [PATCH v2 1/2] cpuhp: make target_store() a nop when target == state

Quoted 29 lines · Phil Auld
Quoted 28 lines · On Fri, May 27, 2022 at 10:38:24AM +0100 Valentin Schneider wrote:
Quoted 13 lines · Phil Auld

writing the current state back in hotplug/target calls cpu_down() which will set cpu dying even when it isn't and then nothing will ever clear it. A stress test that reads values and writes them back for all cpu device files in sysfs will trigger the BUG() in select_fallback_rq once all cpus are marked as dying.

kernel/cpu.c::target_store()
...

        if (st->state < target)
                ret = cpu_up(dev->id, target);
        else
                ret = cpu_down(dev->id, target);
Quoted 9 lines

cpu_down() -> cpu_set_state()

    bool bringup = st->state < target;
    ...
    if (cpu_dying(cpu) != !bringup)
           set_cpu_dying(cpu, !bringup);

Fix this by letting state==target fall through in the target_store() conditional.

To go back on my data race paranoia: writes to both cpu$x/online and cpu$x/hotplug/target are serialized by device_hotplug_lock, and so are the exported kernel hotplug functions ({add, remove}_cpu()).

Quoted 9 lines
Quoted 3 lines

That's not cpu_add_remove_lock as I was looking for, but that's still all under one lock, so I think we're good. Sorry for that!

Right. This catches it up higher so that we don't get into the code that starts actually changing things. I wonder now in the state == target case if we should make sure st->target == target. With the second patch it's less likely to be needed. Thoughts?

Yeah, you could append a simple:

        else
                WARN_ON(st->state != target);
Quoted 2 lines

Maybe I'll include that if/when I have code to keep cpux/online in sync with st->state and cpu_online_mask.

author
#5 lore raw

Re: [PATCH v2 1/2] cpuhp: make target_store() a nop when target == state

Quoted 40 lines · On Mon, May 30, 2022 at 01:27:00PM +0100 Valentin Schneider wrote:
Quoted 29 lines · Phil Auld
Quoted 28 lines · On Fri, May 27, 2022 at 10:38:24AM +0100 Valentin Schneider wrote:
Quoted 13 lines · Phil Auld

writing the current state back in hotplug/target calls cpu_down() which will set cpu dying even when it isn't and then nothing will ever clear it. A stress test that reads values and writes them back for all cpu device files in sysfs will trigger the BUG() in select_fallback_rq once all cpus are marked as dying.

kernel/cpu.c::target_store()
...

        if (st->state < target)
                ret = cpu_up(dev->id, target);
        else
                ret = cpu_down(dev->id, target);
Quoted 9 lines

cpu_down() -> cpu_set_state()

    bool bringup = st->state < target;
    ...
    if (cpu_dying(cpu) != !bringup)
           set_cpu_dying(cpu, !bringup);

Fix this by letting state==target fall through in the target_store() conditional.

To go back on my data race paranoia: writes to both cpu$x/online and cpu$x/hotplug/target are serialized by device_hotplug_lock, and so are the exported kernel hotplug functions ({add, remove}_cpu()).

Quoted 9 lines
Quoted 3 lines

That's not cpu_add_remove_lock as I was looking for, but that's still all under one lock, so I think we're good. Sorry for that!

Right. This catches it up higher so that we don't get into the code that starts actually changing things. I wonder now in the state == target case if we should make sure st->target == target. With the second patch it's less likely to be needed. Thoughts?

Quoted 4 lines

Yeah, you could append a simple:

        else
                WARN_ON(st->state != target);

I was thinking more like:

      	 else 
                   if (st->target != target) st->target = target;

Since this is a write to the target field and we are not doing one of the operations that will set target because state == target we should make sure target == target. Although that could have its own issues, I suppose. But as I said fixing the boot cpu should make it much less likely that st->target != st->state once we have the hotplug lock.

I don't see how that WARN would ever fire. We're under the lock and nothing is re-reading the value of st->state anyway. Looks more like a compiler sanity check :)

Cheers,
Phil

Quoted 3 lines
Quoted 2 lines

Maybe I'll include that if/when I have code to keep cpux/online in sync with st->state and cpu_online_mask.

#6 lore raw

Re: [PATCH v2 1/2] cpuhp: make target_store() a nop when target == state

Quoted 19 lines · Phil Auld
Quoted 5 lines · On Mon, May 30, 2022 at 01:27:00PM +0100 Valentin Schneider wrote:

Yeah, you could append a simple:

        else
                WARN_ON(st->state != target);

I was thinking more like:

      	 else 
                   if (st->target != target) st->target = target;

Since this is a write to the target field and we are not doing one of the operations that will set target because state == target we should make sure target == target. Although that could have its own issues, I suppose. But as I said fixing the boot cpu should make it much less likely that st->target != st->state once we have the hotplug lock.

Quoted 4 lines

I don't see how that WARN would ever fire. We're under the lock and nothing is re-reading the value of st->state anyway. Looks more like a compiler sanity check :)

Right, having a warning in there would mostly be to catch unexpected/unintended scenarios like the one you've found for the boot CPU. I'd vote for

WARN("Huh, didn't expect that")
+
fix st->target

Quoted 10 lines

Cheers,
Phil

Quoted 3 lines
Quoted 2 lines

Maybe I'll include that if/when I have code to keep cpux/online in sync with st->state and cpu_online_mask.

author
#7 lore raw

PATCH v2 2/2 cpuhp: Set cpuhp target for boot cpu

In reply to Phil Auld

Since the boot cpu does not go through the hotplug process it ends up with state == CPUHP_ONLINE but target == CPUHP_OFFLINE. Set the target to match in boot_cpu_hotplug_init().

---
 kernel/cpu.c | 1 +
 1 file changed, 1 insertion(+)
kernel/cpu.c +1 -0
index cdb6ac10ad94..0bb8ad0fa2d9 100644
@@ -2677,6 +2677,7 @@ void __init boot_cpu_hotplug_init(void)
 	cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
 #endif
 	this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
+	this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
 }
 
 /*
Signature · 1 line
2.18.0
Reviewed-by on v2 2/2
#8 lore raw

Re: [PATCH v2 2/2] cpuhp: Set cpuhp target for boot cpu

Quoted 5 lines · Phil Auld

Since the boot cpu does not go through the hotplug process it ends up with state == CPUHP_ONLINE but target == CPUHP_OFFLINE. Set the target to match in boot_cpu_hotplug_init().

Quoted 18 lines
---
 kernel/cpu.c | 1 +
 1 file changed, 1 insertion(+)
kernel/cpu.c +1 -0
index cdb6ac10ad94..0bb8ad0fa2d9 100644
@@ -2677,6 +2677,7 @@ void __init boot_cpu_hotplug_init(void)
 	cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
 #endif
 	this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
+	this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
 }
 
 /*
Signature · 1 line
2.18.0