CPU hotplug: Fix the long-standing "IPI to offline CPU" issue

Hi,

There is a long-standing problem related to CPU hotplug which causes IPIs to be delivered to offline CPUs, and the smp-call-function IPI handler code prints out a warning whenever this is detected. Every once in a while this (usually harmless) warning gets reported on LKML, but so far it has not been completely fixed. Usually the solution involves finding out the IPI sender and fixing it by adding appropriate synchronization with CPU hotplug.

However, while going through one such internal bug reports, I found that there is a significant bug in the receiver side itself (more specifically, in stop-machine) that can lead to this problem even when the sender code is perfectly fine. This patchset fixes that synchronization problem in the CPU hotplug stop-machine code.

Patch 1 adds some additional debug code to the smp-call-function framework, to help debug such issues easily.

Patch 2 modifies the stop-machine code to ensure that any IPIs that were sent while the target CPU was online, would be noticed and handled by that CPU without fail before it goes offline. Thus, this avoids scenarios where IPIs are received on offline CPUs (as long as the sender uses proper hotplug synchronization).

In fact, I debugged the problem by using Patch 1, and found that the payload of the IPI was always the block layer's trigger_softirq() function. But I was not able to find anything wrong with the block layer code. That's when I started looking at the stop-machine code and realized that there is a race-window which makes the IPI _receiver_ the culprit, not the sender. Patch 2 fixes that race and hence this should put an end to most of the hard-to-debug IPI-to-offline-CPU issues.

Srivatsa S. Bhat (2):
smp: Print more useful debug info upon receiving IPI on an offline CPU CPU hotplug, stop-machine: Plug race-window that leads to "IPI-to-offline-CPU"

 kernel/smp.c          |   15 ++++++++++++++-
 kernel/stop_machine.c |   22 +++++++++++++++++++---
 2 files changed, 33 insertions(+), 4 deletions(-)

Thanks,
Srivatsa S. Bhat
IBM Linux Technology Center

Today the smp-call-function code just prints a warning if we get an IPI on an offline CPU. This info is sufficient to let us know that something went wrong, but often it is very hard to debug exactly who sent the IPI and why, from this info alone.

In most cases, we get the warning about the IPI to an offline CPU, immediately after the CPU going offline comes out of the stop-machine phase and reenables interrupts. Since all online CPUs participate in stop-machine, the information regarding the sender of the IPI is already lost by the time we exit the stop-machine loop. So even if we dump the stack on each CPU at this point, we won't find anything useful since all of them will show the stack-trace of the stopper thread. So we need a better way to figure out who sent the IPI and why.

To achieve this, when we detect an IPI targeted to an offline CPU, loop through the call-single-data linked list and print out the payload (i.e., the name of the function which was supposed to be executed by the target CPU). This would give us an insight as to who might have sent the IPI and help us debug this further.

---

 kernel/smp.c |   15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

index 06d574e..6cb0e2e 100644
@@ -185,15 +185,28 @@ void generic_smp_call_function_single_interrupt(void)
 {
 	struct llist_node *entry;
 	struct call_single_data *csd, *csd_next;
+	int warn = 0;
 
 	/*
 	 * Shouldn't receive this interrupt on a cpu that is not yet online.
 	 */
-	WARN_ON_ONCE(!cpu_online(smp_processor_id()));
+	if (unlikely(!cpu_online(smp_processor_id()))) {
+		warn = 1;
+		WARN_ON_ONCE(1);
+	}
 
 	entry = llist_del_all(&__get_cpu_var(call_single_queue));
 	entry = llist_reverse_order(entry);
 
+	if (unlikely(warn)) {
+		/*
+		 * We don't have to use the _safe() variant here
+		 * because we are not invoking the IPI handlers yet.
+		 */
+		llist_for_each_entry(csd, entry, llist)
+			pr_warn("SMP IPI Payload: %pS \n", csd->func);
+	}
+
 	llist_for_each_entry_safe(csd, csd_next, entry, llist) {
 		csd->func(csd->info);
 		csd_unlock(csd);

Yeah, this version looks better. Sorry for missing this earlier. I'll incorporate this in my next version of the patchset.

Thanks a lot!

Regards,
Srivatsa S. Bhat

[...]

Here is the updated patch:

-------------------------------------------------------------------------

From: Srivatsa S. Bhat <[email protected]>
[PATCH v2 1/2] smp: Print more useful debug info upon receiving IPI on an offline CPU

Today the smp-call-function code just prints a warning if we get an IPI on an offline CPU. This info is sufficient to let us know that something went wrong, but often it is very hard to debug exactly who sent the IPI and why, from this info alone.

In most cases, we get the warning about the IPI to an offline CPU, immediately after the CPU going offline comes out of the stop-machine phase and reenables interrupts. Since all online CPUs participate in stop-machine, the information regarding the sender of the IPI is already lost by the time we exit the stop-machine loop. So even if we dump the stack on each CPU at this point, we won't find anything useful since all of them will show the stack-trace of the stopper thread. So we need a better way to figure out who sent the IPI and why.

To achieve this, when we detect an IPI targeted to an offline CPU, loop through the call-single-data linked list and print out the payload (i.e., the name of the function which was supposed to be executed by the target CPU). This would give us an insight as to who might have sent the IPI and help us debug this further.

---

 kernel/smp.c |   18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

index 06d574e..f864921 100644
@@ -185,14 +185,24 @@ void generic_smp_call_function_single_interrupt(void)
 {
 	struct llist_node *entry;
 	struct call_single_data *csd, *csd_next;
+	static bool warned;
+
+	entry = llist_del_all(&__get_cpu_var(call_single_queue));
+	entry = llist_reverse_order(entry);
 
 	/*
 	 * Shouldn't receive this interrupt on a cpu that is not yet online.
 	 */
-	WARN_ON_ONCE(!cpu_online(smp_processor_id()));
-
-	entry = llist_del_all(&__get_cpu_var(call_single_queue));
-	entry = llist_reverse_order(entry);
+	if (unlikely(!cpu_online(smp_processor_id()) && !warned)) {
+		warned = true;
+		WARN_ON(1);
+		/*
+		 * We don't have to use the _safe() variant here
+		 * because we are not invoking the IPI handlers yet.
+		 */
+		llist_for_each_entry(csd, entry, llist)
+			pr_warn("SMP IPI Payload: %pS \n", csd->func);
+	}
 
 	llist_for_each_entry_safe(csd, csd_next, entry, llist) {
 		csd->func(csd->info);

During CPU offline, stop-machine is used to take control over all the online CPUs (via the per-cpu stopper thread) and then run take_cpu_down() on the CPU that is to be taken offline.

But stop-machine itself has several stages: _PREPARE, _DISABLE_IRQ, _RUN etc. The important thing to note here is that the _DISABLE_IRQ stage comes much later after starting stop-machine, and hence there is a large window where other CPUs can send IPIs to the CPU going offline. As a result, we can encounter a scenario as depicted below, which causes IPIs to be sent to the CPU going offline, and that CPU notices them *after* it has gone offline, triggering the "IPI-to-offline-CPU" warning from the smp-call-function code.

CPU 1 CPU 2
(Online CPU) (CPU going offline)

Enter _PREPARE stage Enter _PREPARE stage

Enter _DISABLE_IRQ stage

=
Got a device interrupt, | Didn't notice the IPI
and the interrupt handler | since interrupts were
called smp_call_function() | disabled on this CPU.
and sent an IPI to CPU 2. |
=

Enter _DISABLE_IRQ stage

Enter _RUN stage Enter _RUN stage

=
Busy loop with interrupts | Invoke take_cpu_down()
disabled. | and take CPU 2 offline
=

Enter _EXIT stage Enter _EXIT stage

Re-enable interrupts Re-enable interrupts

The pending IPI is noted immediately, but alas, the CPU is offline at this point.

So, as we can observe from this scenario, the IPI was sent when CPU 2 was still online, and hence it was perfectly legal. But unfortunately it was noted only after CPU 2 went offline, resulting in the warning from the IPI handling code. In other words, the fault was not at the sender, but at the receiver side - and if we look closely, the real bug is in the stop-machine sequence itself.

The problem here is that the CPU going offline disabled its local interrupts (by entering _DISABLE_IRQ phase) *before* the other CPUs. And that's the reason why it was not able to respond to the IPI before going offline.

A simple solution to this problem is to ensure that the CPU going offline *follows* all other CPUs while entering each subsequent phase within stop-machine. In particular, all other CPUs will enter the _DISABLE_IRQ phase and disable their local interrupts, and only *then*, the CPU going offline will follow suit. Since the other CPUs are executing the stop-machine code with interrupts disabled, they won't send any IPIs at all, at that point. And by the time stop-machine ends, the CPU would have gone offline and disappeared from the cpu_online_mask, and hence future invocations of smp_call_function() and friends will automatically prune that CPU out. Thus, we can guarantee that no CPU will end up *inadvertently* sending IPIs to an offline CPU.

We can implement this by introducing a "holding area" for the CPUs marked as 'active_cpus', and use this infrastructure to let the other CPUs progress from one stage to the next, before allowing the active_cpus to do the same thing.

---

 kernel/stop_machine.c |   22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)

index 01fbae5..d65168e 100644
@@ -165,12 +165,21 @@ static void ack_state(struct multi_stop_data *msdata)
 		set_state(msdata, msdata->state + 1);
 }
 
+/* Holding area for active CPUs, to let all the non-active CPUs go first */
+static void hold_active_cpus(struct multi_stop_data *msdata,
+			     int num_active_cpus)
+{
+	/* Wait until all the non-active threads ack the state */
+	while (atomic_read(&msdata->thread_ack) > num_active_cpus)
+		cpu_relax();
+}
+
 /* This is the cpu_stop function which stops the CPU. */
 static int multi_cpu_stop(void *data)
 {
 	struct multi_stop_data *msdata = data;
 	enum multi_stop_state curstate = MULTI_STOP_NONE;
-	int cpu = smp_processor_id(), err = 0;
+	int cpu = smp_processor_id(), num_active_cpus, err = 0;
 	unsigned long flags;
 	bool is_active;
 

@@ -180,15 +189,22 @@ static int multi_cpu_stop(void *data)
 	 */
 	local_save_flags(flags);
 
-	if (!msdata->active_cpus)
+	if (!msdata->active_cpus) {
 		is_active = cpu == cpumask_first(cpu_online_mask);
-	else
+		num_active_cpus = 1;
+	} else {
 		is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
+		num_active_cpus = cpumask_weight(msdata->active_cpus);
+	}
 
 	/* Simple state machine */
 	do {
 		/* Chill out and ensure we re-read multi_stop_state. */
 		cpu_relax();
+
+		if (is_active)
+			hold_active_cpus(msdata, num_active_cpus);
+
 		if (msdata->state != curstate) {
 			curstate = msdata->state;
 			switch (curstate) {

Does it even need to be a separate function? I kinda really dislike trivial helpers which are used only once. It obfuscates more than helping anything. I think proper comment where the actual synchronization is happening along with open coded wait would be easier to follow.

Thanks.

tejun

Ok, I'll open code it and add an appropriate comment explaining the synchronization.

Thank you!

Regards,
Srivatsa S. Bhat

[...]

How about this?

---------------------------------------------------------------------------

From: Srivatsa S. Bhat <[email protected]>
[PATCH v2 2/2] CPU hotplug, stop-machine: Plug race-window that leads to "IPI-to-offline-CPU"

During CPU offline, stop-machine is used to take control over all the online CPUs (via the per-cpu stopper thread) and then run take_cpu_down() on the CPU that is to be taken offline.

But stop-machine itself has several stages: _PREPARE, _DISABLE_IRQ, _RUN etc. The important thing to note here is that the _DISABLE_IRQ stage comes much later after starting stop-machine, and hence there is a large window where other CPUs can send IPIs to the CPU going offline. As a result, we can encounter a scenario as depicted below, which causes IPIs to be sent to the CPU going offline, and that CPU notices them *after* it has gone offline, triggering the "IPI-to-offline-CPU" warning from the smp-call-function code.

CPU 1 CPU 2
(Online CPU) (CPU going offline)

Enter _PREPARE stage Enter _PREPARE stage

Enter _DISABLE_IRQ stage

=
Got a device interrupt, | Didn't notice the IPI
and the interrupt handler | since interrupts were
called smp_call_function() | disabled on this CPU.
and sent an IPI to CPU 2. |
=

Enter _DISABLE_IRQ stage

Enter _RUN stage Enter _RUN stage

=
Busy loop with interrupts | Invoke take_cpu_down()
disabled. | and take CPU 2 offline
=

Enter _EXIT stage Enter _EXIT stage

Re-enable interrupts Re-enable interrupts

The pending IPI is noted immediately, but alas, the CPU is offline at this point.

So, as we can observe from this scenario, the IPI was sent when CPU 2 was still online, and hence it was perfectly legal. But unfortunately it was noted only after CPU 2 went offline, resulting in the warning from the IPI handling code. In other words, the fault was not at the sender, but at the receiver side - and if we look closely, the real bug is in the stop-machine sequence itself.

The problem here is that the CPU going offline disabled its local interrupts (by entering _DISABLE_IRQ phase) *before* the other CPUs. And that's the reason why it was not able to respond to the IPI before going offline.

A simple solution to this problem is to ensure that the CPU going offline *follows* all other CPUs while entering each subsequent phase within stop-machine. In particular, all other CPUs will enter the _DISABLE_IRQ phase and disable their local interrupts, and only *then*, the CPU going offline will follow suit. Since the other CPUs are executing the stop-machine code with interrupts disabled, they won't send any IPIs at all, at that point. And by the time stop-machine ends, the CPU would have gone offline and disappeared from the cpu_online_mask, and hence future invocations of smp_call_function() and friends will automatically prune that CPU out. Thus, we can guarantee that no CPU will end up *inadvertently* sending IPIs to an offline CPU.

We can implement this by introducing a "holding area" for the CPUs marked as 'active_cpus', and use this infrastructure to let the other CPUs progress from one stage to the next, before allowing the active_cpus to do the same thing.

---

 kernel/stop_machine.c |   30 +++++++++++++++++++++++++++---
 1 file changed, 27 insertions(+), 3 deletions(-)

index 01fbae5..7abb361 100644
@@ -165,12 +165,13 @@ static void ack_state(struct multi_stop_data *msdata)
 		set_state(msdata, msdata->state + 1);
 }
 
+
 /* This is the cpu_stop function which stops the CPU. */
 static int multi_cpu_stop(void *data)
 {
 	struct multi_stop_data *msdata = data;
 	enum multi_stop_state curstate = MULTI_STOP_NONE;
-	int cpu = smp_processor_id(), err = 0;
+	int cpu = smp_processor_id(), num_active_cpus, err = 0;
 	unsigned long flags;
 	bool is_active;
 

@@ -180,15 +181,38 @@ static int multi_cpu_stop(void *data)
 	 */
 	local_save_flags(flags);
 
-	if (!msdata->active_cpus)
+	if (!msdata->active_cpus) {
 		is_active = cpu == cpumask_first(cpu_online_mask);
-	else
+		num_active_cpus = 1;
+	} else {
 		is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
+		num_active_cpus = cpumask_weight(msdata->active_cpus);
+	}
 
 	/* Simple state machine */
 	do {
 		/* Chill out and ensure we re-read multi_stop_state. */
 		cpu_relax();
+
+		/*
+		 * In the case of CPU offline, we don't want the other CPUs to
+		 * send IPIs to the active_cpu (the one going offline) after it
+		 * has entered the _DISABLE_IRQ state (because, then it will
+		 * notice the IPIs only after it goes offline). So ensure that
+		 * the active_cpu always follows the others while entering
+		 * each subsequent state in this state-machine.
+		 *
+		 * msdata->thread_ack tracks the number of CPUs that are yet to
+		 * move to the next state, during each transition. So make the
+		 * active_cpu(s) wait until ->thread_ack indicates that the
+		 * active_cpus are the only ones left to complete the transition.
+		 */
+		if (is_active) {
+			/* Wait until all the non-active threads ack the state */
+			while (atomic_read(&msdata->thread_ack) > num_active_cpus)
+				cpu_relax();
+		}
+
 		if (msdata->state != curstate) {
 			curstate = msdata->state;
 			switch (curstate) {

Argh, a stray newline.. will remove it.

Yeah, now that you point out, it does look very odd. I don't remember why I wrote it that way in the first place! :-( I'll fix this in the next version. Thanks!

Sure, will use this convention.

Indeed, that looks like a much more elegant method! Thanks a lot for the suggestion Tejun, I'll use that in the next version of the patchset.

Thank you!

Regards,
Srivatsa S. Bhat