Skip to contentSkip to conversation

kexec: disable cpu hotplug until the rebooting cpu is stable

PATCH Superseded Pingfan Liu 2 participants 5 messages Jan 25, 2022 – Jan 27, 2022

Subsystems (4)

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kexec: disable cpu hotplug until the rebooting cpu is stable

The following identical code piece appears in both migrate_to_reboot_cpu() and smp_shutdown_nonboot_cpus():

	if (!cpu_online(primary_cpu))
		primary_cpu = cpumask_first(cpu_online_mask);

Although the kexec-reboot task can get through a cpu_down() on its cpu, this code looks a little confusing.

Make things straight forward by keep cpu hotplug disabled until smp_shutdown_nonboot_cpus() holds cpu_add_remove_lock. By this way, the rebooting cpu can keep unchanged.

Cc Eric Biederman <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Vincent Donnefort <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Mark Rutland <[email protected]>
Cc YueHaibing <[email protected]>
Cc Baokun Li <[email protected]>
Cc Randy Dunlap <[email protected]>
Cc kexec at lists.infradead.org

To: [email protected]

---
 kernel/cpu.c        | 16 ++++++++++------
 kernel/kexec_core.c | 10 ++++------
 2 files changed, 14 insertions(+), 12 deletions(-)
kernel/cpu.c +10 -6
index 407a2568f35e..bc687d59ca90 100644
@@ -1227,20 +1227,24 @@ int remove_cpu(unsigned int cpu)
 }
 EXPORT_SYMBOL_GPL(remove_cpu);
 
+/* primary_cpu keeps unchanged after migrate_to_reboot_cpu() */
 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
 {
 	unsigned int cpu;
 	int error;
 
+	/*
+	 * Block other cpu hotplug event, so primary_cpu is always online if
+	 * it is not touched by us
+	 */
 	cpu_maps_update_begin();
-
 	/*
-	 * Make certain the cpu I'm about to reboot on is online.
-	 *
-	 * This is inline to what migrate_to_reboot_cpu() already do.
+	 * migrate_to_reboot_cpu() disables CPU hotplug assuming that
+	 * no further code needs to use CPU hotplug (which is true in
+	 * the reboot case). However, the kexec path depends on using
+	 * CPU hotplug again; so re-enable it here.
 	 */
-	if (!cpu_online(primary_cpu))
-		primary_cpu = cpumask_first(cpu_online_mask);
+	__cpu_hotplug_enable();
 
 	for_each_online_cpu(cpu) {
 		if (cpu == primary_cpu)
kernel/kexec_core.c +4 -6
index 68480f731192..db4fa6b174e3 100644
@@ -1168,14 +1168,12 @@ int kernel_kexec(void)
 		kexec_in_progress = true;
 		kernel_restart_prepare("kexec reboot");
 		migrate_to_reboot_cpu();
-
 		/*
-		 * migrate_to_reboot_cpu() disables CPU hotplug assuming that
-		 * no further code needs to use CPU hotplug (which is true in
-		 * the reboot case). However, the kexec path depends on using
-		 * CPU hotplug again; so re-enable it here.
+		 * migrate_to_reboot_cpu() disables CPU hotplug. If an arch
+		 * relies on the cpu teardown to achieve reboot, it needs to
+		 * re-enable CPU hotplug there.
 		 */
-		cpu_hotplug_enable();
+
 		pr_notice("Starting new kernel\n");
 		machine_shutdown();
 	}
Signature · 1 line
2.31.1
#1 lore raw
Quoted 6 lines · Pingfan Liu

The following identical code piece appears in both migrate_to_reboot_cpu() and smp_shutdown_nonboot_cpus():

	if (!cpu_online(primary_cpu))
		primary_cpu = cpumask_first(cpu_online_mask);
Quoted 7 lines

Although the kexec-reboot task can get through a cpu_down() on its cpu, this code looks a little confusing.

Make things straight forward by keep cpu hotplug disabled until smp_shutdown_nonboot_cpus() holds cpu_add_remove_lock. By this way, the rebooting cpu can keep unchanged.

So is this supposed to be a refactor with no change in behaviour? AFAICT it actually does change things (and isn't necessarily clearer).

Quoted 63 lines
Cc Eric Biederman <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Vincent Donnefort <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Mark Rutland <[email protected]>
Cc YueHaibing <[email protected]>
Cc Baokun Li <[email protected]>
Cc Randy Dunlap <[email protected]>
Cc kexec at lists.infradead.org

To: linux-kernel at vger.kernel.org

---
 kernel/cpu.c        | 16 ++++++++++------
 kernel/kexec_core.c | 10 ++++------
 2 files changed, 14 insertions(+), 12 deletions(-)
kernel/cpu.c +10 -6
index 407a2568f35e..bc687d59ca90 100644
@@ -1227,20 +1227,24 @@ int remove_cpu(unsigned int cpu)
 }
 EXPORT_SYMBOL_GPL(remove_cpu);
 
+/* primary_cpu keeps unchanged after migrate_to_reboot_cpu() */
 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
 {
 	unsigned int cpu;
 	int error;
 
+	/*
+	 * Block other cpu hotplug event, so primary_cpu is always online if
+	 * it is not touched by us
+	 */
 	cpu_maps_update_begin();
-
 	/*
-	 * Make certain the cpu I'm about to reboot on is online.
-	 *
-	 * This is inline to what migrate_to_reboot_cpu() already do.
+	 * migrate_to_reboot_cpu() disables CPU hotplug assuming that
+	 * no further code needs to use CPU hotplug (which is true in
+	 * the reboot case). However, the kexec path depends on using
+	 * CPU hotplug again; so re-enable it here.
 	 */
-	if (!cpu_online(primary_cpu))
-		primary_cpu = cpumask_first(cpu_online_mask);
+	__cpu_hotplug_enable();
 
 	for_each_online_cpu(cpu) {
 		if (cpu == primary_cpu)
kernel/kexec_core.c +0 -1
index 68480f731192..db4fa6b174e3 100644
@@ -1168,14 +1168,12 @@ int kernel_kexec(void)
 		kexec_in_progress = true;
 		kernel_restart_prepare("kexec reboot");
 		migrate_to_reboot_cpu();
-
 		/*

kernel/kexec_core.c int kernel_kexec(void) · line 1176

Quoted 10 lines
kernel/kexec_core.c +4 -5
-		 * migrate_to_reboot_cpu() disables CPU hotplug assuming that
-		 * no further code needs to use CPU hotplug (which is true in
-		 * the reboot case). However, the kexec path depends on using
-		 * CPU hotplug again; so re-enable it here.
+		 * migrate_to_reboot_cpu() disables CPU hotplug. If an arch
+		 * relies on the cpu teardown to achieve reboot, it needs to
+		 * re-enable CPU hotplug there.
 		 */
-		cpu_hotplug_enable();
+

Not all archs map machine_shutdown() to smp_shutdown_nonboot_cpus(), other archs will now be missing a cpu_hotplug_enable() prior to a kexec machine_shutdown(). That said, AFAICT none of those archs rely on the hotplug machinery in machine_shutdown(), so it might be OK, but that's not obvious at all.

Quoted 5 lines
kernel/kexec_core.c +0 -0
 		pr_notice("Starting new kernel\n");
 		machine_shutdown();
 	}
Signature · 1 line
2.31.1
#2 lore raw
Quoted 11 lines · Valentin Schneider
Quoted 9 lines · Pingfan Liu

The following identical code piece appears in both migrate_to_reboot_cpu() and smp_shutdown_nonboot_cpus():

      if (!cpu_online(primary_cpu))
              primary_cpu = cpumask_first(cpu_online_mask);

Although the kexec-reboot task can get through a cpu_down() on its cpu, this code looks a little confusing.

Quoted 8 lines
Quoted 4 lines

Make things straight forward by keep cpu hotplug disabled until smp_shutdown_nonboot_cpus() holds cpu_add_remove_lock. By this way, the rebooting cpu can keep unchanged.

So is this supposed to be a refactor with no change in behaviour? AFAICT it actually does change things (and isn't necessarily clearer).

Yes, as you have seen, it does change behavior. Before this patch, there is a breakage:
migrate_to_reboot_cpu();

  cpu_hotplug_enable();
                                     ----------> technical, here can

comes a cpu_down(this_cpu)
machine_shutdown();

And this patch squeezes out this breakage.

Quoted 71 lines
Quoted 71 lines
Cc Eric Biederman <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Vincent Donnefort <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Mark Rutland <[email protected]>
Cc YueHaibing <[email protected]>
Cc Baokun Li <[email protected]>
Cc Randy Dunlap <[email protected]>
Cc kexec at lists.infradead.org

To: linux-kernel at vger.kernel.org

---
 kernel/cpu.c        | 16 ++++++++++------
 kernel/kexec_core.c | 10 ++++------
 2 files changed, 14 insertions(+), 12 deletions(-)
kernel/cpu.c +10 -6
index 407a2568f35e..bc687d59ca90 100644
@@ -1227,20 +1227,24 @@ int remove_cpu(unsigned int cpu)
 }
 EXPORT_SYMBOL_GPL(remove_cpu);
 
+/* primary_cpu keeps unchanged after migrate_to_reboot_cpu() */
 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
 {
      unsigned int cpu;
      int error;
 
+     /*
+      * Block other cpu hotplug event, so primary_cpu is always online if
+      * it is not touched by us
+      */
      cpu_maps_update_begin();
-
      /*
-      * Make certain the cpu I'm about to reboot on is online.
-      *
-      * This is inline to what migrate_to_reboot_cpu() already do.
+      * migrate_to_reboot_cpu() disables CPU hotplug assuming that
+      * no further code needs to use CPU hotplug (which is true in
+      * the reboot case). However, the kexec path depends on using
+      * CPU hotplug again; so re-enable it here.
       */
-     if (!cpu_online(primary_cpu))
-             primary_cpu = cpumask_first(cpu_online_mask);
+     __cpu_hotplug_enable();
 
      for_each_online_cpu(cpu) {
              if (cpu == primary_cpu)
kernel/kexec_core.c +3 -5
index 68480f731192..db4fa6b174e3 100644
@@ -1168,14 +1168,12 @@ int kernel_kexec(void)
              kexec_in_progress = true;
              kernel_restart_prepare("kexec reboot");
              migrate_to_reboot_cpu();
-
              /*
-              * migrate_to_reboot_cpu() disables CPU hotplug assuming that
-              * no further code needs to use CPU hotplug (which is true in
-              * the reboot case). However, the kexec path depends on using
-              * CPU hotplug again; so re-enable it here.
+              * migrate_to_reboot_cpu() disables CPU hotplug. If an arch
+              * relies on the cpu teardown to achieve reboot, it needs to
+              * re-enable CPU hotplug there.
               */
Quoted 9 lines

kernel/kexec_core.c int kernel_kexec(void) · line 1176

Quoted 2 lines
kernel/kexec_core.c +1 -1
-             cpu_hotplug_enable();
+

Not all archs map machine_shutdown() to smp_shutdown_nonboot_cpus(), other archs will now be missing a cpu_hotplug_enable() prior to a kexec machine_shutdown(). That said, AFAICT none of those archs rely on the hotplug machinery in machine_shutdown(), so it might be OK, but that's not obvious at all.

At the first glance, it may be not obvious, but tracing down cpu_hotplug_enable() to the variable cpu_hotplug_disabled, you can find out the limited involved functions are all related to cpu_up/cpu_down.

IOW, if no code path connects with the interface of cpu_up/cpu_down, then kexec-reboot will not be affected.

And after this patch, it is more clear how to handle the following cases:
arch/arm/kernel/reboot.c:94: smp_shutdown_nonboot_cpus(reboot_cpu);
arch/arm64/kernel/process.c:88: smp_shutdown_nonboot_cpus(reboot_cpu); arch/ia64/kernel/process.c:578: smp_shutdown_nonboot_cpus(reboot_cpu);

Thanks,
Pingfan

#3 lore raw
Quoted 24 lines · Pingfan Liu
Quoted 11 lines · Valentin Schneider
Quoted 9 lines · Pingfan Liu

The following identical code piece appears in both migrate_to_reboot_cpu() and smp_shutdown_nonboot_cpus():

      if (!cpu_online(primary_cpu))
              primary_cpu = cpumask_first(cpu_online_mask);

Although the kexec-reboot task can get through a cpu_down() on its cpu, this code looks a little confusing.

Quoted 8 lines
Quoted 4 lines

Make things straight forward by keep cpu hotplug disabled until smp_shutdown_nonboot_cpus() holds cpu_add_remove_lock. By this way, the rebooting cpu can keep unchanged.

So is this supposed to be a refactor with no change in behaviour? AFAICT it actually does change things (and isn't necessarily clearer).

Yes, as you have seen, it does change behavior. Before this patch, there is a breakage:
migrate_to_reboot_cpu();

Quoted 7 lines
  cpu_hotplug_enable();
                                     ----------> technical, here can

comes a cpu_down(this_cpu)
machine_shutdown();

And this patch squeezes out this breakage.

Ok, that's worth pointing out in the changelog.

Quoted 27 lines
Quoted 18 lines
Quoted 18 lines
kernel/kexec_core.c +3 -5
index 68480f731192..db4fa6b174e3 100644
@@ -1168,14 +1168,12 @@ int kernel_kexec(void)
              kexec_in_progress = true;
              kernel_restart_prepare("kexec reboot");
              migrate_to_reboot_cpu();
-
              /*
-              * migrate_to_reboot_cpu() disables CPU hotplug assuming that
-              * no further code needs to use CPU hotplug (which is true in
-              * the reboot case). However, the kexec path depends on using
-              * CPU hotplug again; so re-enable it here.
+              * migrate_to_reboot_cpu() disables CPU hotplug. If an arch
+              * relies on the cpu teardown to achieve reboot, it needs to
+              * re-enable CPU hotplug there.
               */
Quoted 9 lines

kernel/kexec_core.c int kernel_kexec(void) · line 1176

Quoted 2 lines
kernel/kexec_core.c +1 -1
-             cpu_hotplug_enable();
+

Not all archs map machine_shutdown() to smp_shutdown_nonboot_cpus(), other archs will now be missing a cpu_hotplug_enable() prior to a kexec machine_shutdown(). That said, AFAICT none of those archs rely on the hotplug machinery in machine_shutdown(), so it might be OK, but that's not obvious at all.

Quoted 8 lines

At the first glance, it may be not obvious, but tracing down cpu_hotplug_enable() to the variable cpu_hotplug_disabled, you can find out the limited involved functions are all related to cpu_up/cpu_down.

IOW, if no code path connects with the interface of cpu_up/cpu_down, then kexec-reboot will not be affected.

That's my point, this only works if the other archs truly do not rely on hotplug for machine_shutdown(), which seems to be the case but it wouldn't hurt for you to double-check that and explicitely call it out in the changelog.

Quoted 5 lines

And after this patch, it is more clear how to handle the following cases:
arch/arm/kernel/reboot.c:94: smp_shutdown_nonboot_cpus(reboot_cpu);
arch/arm64/kernel/process.c:88: smp_shutdown_nonboot_cpus(reboot_cpu); arch/ia64/kernel/process.c:578: smp_shutdown_nonboot_cpus(reboot_cpu);

FWIW riscv is also concerned.

Quoted 2 lines

Thanks,
Pingfan

#4 lore raw
Quoted 31 lines · Valentin Schneider
Quoted 29 lines · Pingfan Liu
Quoted 11 lines · Valentin Schneider
Quoted 9 lines · Pingfan Liu

The following identical code piece appears in both migrate_to_reboot_cpu() and smp_shutdown_nonboot_cpus():

      if (!cpu_online(primary_cpu))
              primary_cpu = cpumask_first(cpu_online_mask);

Although the kexec-reboot task can get through a cpu_down() on its cpu, this code looks a little confusing.

Quoted 8 lines
Quoted 4 lines

Make things straight forward by keep cpu hotplug disabled until smp_shutdown_nonboot_cpus() holds cpu_add_remove_lock. By this way, the rebooting cpu can keep unchanged.

So is this supposed to be a refactor with no change in behaviour? AFAICT it actually does change things (and isn't necessarily clearer).

Yes, as you have seen, it does change behavior. Before this patch, there is a breakage:
migrate_to_reboot_cpu();

  cpu_hotplug_enable();
                                     ----------> technical, here can

comes a cpu_down(this_cpu)
machine_shutdown();

Quoted 5 lines
Quoted 2 lines

And this patch squeezes out this breakage.

Ok, that's worth pointing out in the changelog.

Sure, I will update it.

Quoted 32 lines
Quoted 32 lines
Quoted 18 lines
Quoted 18 lines
kernel/kexec_core.c +3 -5
index 68480f731192..db4fa6b174e3 100644
@@ -1168,14 +1168,12 @@ int kernel_kexec(void)
              kexec_in_progress = true;
              kernel_restart_prepare("kexec reboot");
              migrate_to_reboot_cpu();
-
              /*
-              * migrate_to_reboot_cpu() disables CPU hotplug assuming that
-              * no further code needs to use CPU hotplug (which is true in
-              * the reboot case). However, the kexec path depends on using
-              * CPU hotplug again; so re-enable it here.
+              * migrate_to_reboot_cpu() disables CPU hotplug. If an arch
+              * relies on the cpu teardown to achieve reboot, it needs to
+              * re-enable CPU hotplug there.
               */
Quoted 9 lines

kernel/kexec_core.c int kernel_kexec(void) · line 1176

Quoted 2 lines
kernel/kexec_core.c +1 -1
-             cpu_hotplug_enable();
+

Not all archs map machine_shutdown() to smp_shutdown_nonboot_cpus(), other archs will now be missing a cpu_hotplug_enable() prior to a kexec machine_shutdown(). That said, AFAICT none of those archs rely on the hotplug machinery in machine_shutdown(), so it might be OK, but that's not obvious at all.

At the first glance, it may be not obvious, but tracing down cpu_hotplug_enable() to the variable cpu_hotplug_disabled, you can find out the limited involved functions are all related to cpu_up/cpu_down.

Quoted 9 lines
Quoted 3 lines

IOW, if no code path connects with the interface of cpu_up/cpu_down, then kexec-reboot will not be affected.

That's my point, this only works if the other archs truly do not rely on hotplug for machine_shutdown(), which seems to be the case but it wouldn't hurt for you to double-check that and explicitely call it out in the changelog.

OK, I will update the change log. BTW, besides x86, I have just finished the test on powerpc, both of them works fine with this patch

Quoted 7 lines
Quoted 5 lines

And after this patch, it is more clear how to handle the following cases:
arch/arm/kernel/reboot.c:94: smp_shutdown_nonboot_cpus(reboot_cpu);
arch/arm64/kernel/process.c:88: smp_shutdown_nonboot_cpus(reboot_cpu); arch/ia64/kernel/process.c:578: smp_shutdown_nonboot_cpus(reboot_cpu);

FWIW riscv is also concerned.

I think its current statement is right.
arch/riscv/kernel/machine_kexec.c:135:
smp_shutdown_nonboot_cpus(smp_processor_id());

Thanks,

Pingfan

Quoted 3 lines
Quoted 3 lines

Thanks,
Pingfan