[PATCH 0/2] powerpc prevents deadlock in the watchdog path

While doing LPM on large system (for instance a Brazos system with 1024
CPUs and 12TB of memory) with an heavy load (I ran 'stress-ng --futex 500
-vm 5'), watchdog hard lockup are seen when the hypervisor is taking
too much time handling the page tables to track page's changes.

When this happens, the system may hung with a deadlock between the watchdog
lock and the console owner lock.

The first patch of this series prevents that deadlock by not calling printk
while holding the watchdog lock, and also not sending IPI (and waiting for
CPU's answer during 1s) while holding the watchdog lock.

The second patch ensures that the watchdog's data are accessed under the
protection of the watchdog lock.

Laurent Dufour (2):
  powerpc/watchdog: prevent printk and send IPI while holding the wd
    lock
  powerpc/watchdog: ensure watchdog data accesses are protected

 arch/powerpc/kernel/watchdog.c | 45 +++++++++++++++++++---------------
 1 file changed, 25 insertions(+), 20 deletions(-)

-- 
2.33.1

The wd_smp_cpus_pending CPU mask should be accessed under the protection of
the __wd_smp_lock.

This prevents false alarm to be raised when the system is under an heavy
stress. This has been seen while doing LPM on large system with a big
workload.

Signed-off-by: Laurent Dufour <redacted>
---
 arch/powerpc/kernel/watchdog.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index bc7411327066..8d7a1a86187e 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c

@@ -203,12 +203,13 @@ static void watchdog_smp_panic(int cpu, u64 tb)
 
 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 {
+	unsigned long flags;
+
+	wd_smp_lock(&flags);
 	if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 		if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 			struct pt_regs *regs = get_irq_regs();
-			unsigned long flags;
 
-			wd_smp_lock(&flags);
 			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 			wd_smp_unlock(&flags);
 

@@ -219,22 +220,23 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 				show_regs(regs);
 			else
 				dump_stack();
+			return;
 		}
+
+		wd_smp_unlock(&flags);
 		return;
 	}
+
 	cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 	if (cpumask_empty(&wd_smp_cpus_pending)) {
-		unsigned long flags;
-
-		wd_smp_lock(&flags);
 		if (cpumask_empty(&wd_smp_cpus_pending)) {
 			wd_smp_last_reset_tb = tb;
 			cpumask_andnot(&wd_smp_cpus_pending,
 					&wd_cpus_enabled,
 					&wd_smp_cpus_stuck);
 		}
-		wd_smp_unlock(&flags);
 	}
+	wd_smp_unlock(&flags);
 }
 
 static void watchdog_timer_interrupt(int cpu)

-- 
2.33.1

When handling the Watchdog interrupt, long processing should not be done
while holding the __wd_smp_lock. This prevents the other CPUs to grab it
and to process Watchdog timer interrupts. Furhtermore, this could lead to
the following situation:

CPU x detect lockup on CPU y and grab the __wd_smp_lock
      in watchdog_smp_panic()
CPU y caught the watchdog interrupt and try to grab the __wd_smp_lock
      in soft_nmi_interrupt()
CPU x wait for CPU y to catch the IPI for 1s in __smp_send_nmi_ipi()
CPU x will timeout and so has spent 1s waiting while holding the
      __wd_smp_lock.

A deadlock may also happen between the __wd_smp_lock and the console_owner
'lock' this way:
CPU x grab the console_owner
CPU y grab the __wd_smp_lock
CPU x catch the watchdog timer interrupt and needs to grab __wd_smp_lock
CPU y wants to print something and wait for console_owner
-> deadlock

Doing all the long processing without holding the _wd_smp_lock prevents
these situations.

Signed-off-by: Laurent Dufour <redacted>
---
 arch/powerpc/kernel/watchdog.c | 31 +++++++++++++++++--------------
 1 file changed, 17 insertions(+), 14 deletions(-)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index f9ea0e5357f9..bc7411327066 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c

@@ -149,6 +149,8 @@ static void set_cpu_stuck(int cpu, u64 tb)
 
 static void watchdog_smp_panic(int cpu, u64 tb)
 {
+	cpumask_t cpus_pending_copy;
+	u64 last_reset_tb_copy;
 	unsigned long flags;
 	int c;
 

@@ -161,29 +163,32 @@ static void watchdog_smp_panic(int cpu, u64 tb)
 	if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 		goto out;
 
+	cpumask_copy(&cpus_pending_copy, &wd_smp_cpus_pending);
+	last_reset_tb_copy = wd_smp_last_reset_tb;
+
+	/* Take the stuck CPUs out of the watch group */
+	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
+
+	wd_smp_unlock(&flags);
+
 	pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
-		 cpu, cpumask_pr_args(&wd_smp_cpus_pending));
+		 cpu, cpumask_pr_args(&cpus_pending_copy));
 	pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
-		 cpu, tb, wd_smp_last_reset_tb,
-		 tb_to_ns(tb - wd_smp_last_reset_tb) / 1000000);
+		 cpu, tb, last_reset_tb_copy,
+		 tb_to_ns(tb - last_reset_tb_copy) / 1000000);
 
 	if (!sysctl_hardlockup_all_cpu_backtrace) {
 		/*
 		 * Try to trigger the stuck CPUs, unless we are going to
 		 * get a backtrace on all of them anyway.
 		 */
-		for_each_cpu(c, &wd_smp_cpus_pending) {
+		for_each_cpu(c, &cpus_pending_copy) {
 			if (c == cpu)
 				continue;
 			smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 		}
 	}
 
-	/* Take the stuck CPUs out of the watch group */
-	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
-
-	wd_smp_unlock(&flags);
-
 	if (sysctl_hardlockup_all_cpu_backtrace)
 		trigger_allbutself_cpu_backtrace();
 

@@ -204,6 +209,8 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 			unsigned long flags;
 
 			wd_smp_lock(&flags);
+			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
+			wd_smp_unlock(&flags);
 
 			pr_emerg("CPU %d became unstuck TB:%lld\n",
 				 cpu, tb);

@@ -212,9 +219,6 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 				show_regs(regs);
 			else
 				dump_stack();
-
-			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
-			wd_smp_unlock(&flags);
 		}
 		return;
 	}

@@ -267,6 +271,7 @@ DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
 			return 0;
 		}
 		set_cpu_stuck(cpu, tb);
+		wd_smp_unlock(&flags);
 
 		pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
 			 cpu, (void *)regs->nip);

@@ -277,8 +282,6 @@ DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
 		print_irqtrace_events(current);
 		show_regs(regs);
 
-		wd_smp_unlock(&flags);
-
 		if (sysctl_hardlockup_all_cpu_backtrace)
 			trigger_allbutself_cpu_backtrace();
 

-- 
2.33.1

Excerpts from Laurent Dufour's message of October 27, 2021 2:27 am:

When handling the Watchdog interrupt, long processing should not be done
while holding the __wd_smp_lock. This prevents the other CPUs to grab it
and to process Watchdog timer interrupts. Furhtermore, this could lead to
the following situation:

CPU x detect lockup on CPU y and grab the __wd_smp_lock
      in watchdog_smp_panic()
CPU y caught the watchdog interrupt and try to grab the __wd_smp_lock
      in soft_nmi_interrupt()
CPU x wait for CPU y to catch the IPI for 1s in __smp_send_nmi_ipi()

CPU y should get the IPI here if it's a NMI IPI (which will be true for 

= POWER9 64s).

That said, not all platforms support it and the console lock problem
seems real, so okay.

CPU x will timeout and so has spent 1s waiting while holding the
      __wd_smp_lock.

A deadlock may also happen between the __wd_smp_lock and the console_owner
'lock' this way:
CPU x grab the console_owner
CPU y grab the __wd_smp_lock
CPU x catch the watchdog timer interrupt and needs to grab __wd_smp_lock
CPU y wants to print something and wait for console_owner
-> deadlock

Doing all the long processing without holding the _wd_smp_lock prevents
these situations.

The intention was to avoid logs getting garbled e.g., if multiple 
different CPUs fire at once.

I wonder if instead we could deal with that by protecting the IPI 
sending and printing stuff with a trylock, and if you don't get the
trylock then just return, and you'll come back with the next timer
interrupt.

Thanks,
Nick

Excerpts from Laurent Dufour's message of October 27, 2021 2:27 am:

Hmm. I wanted to avoid the lock here because all CPUs will do it on 
every heartbeat timer.

Although maybe when you look at the cacheline transfers required it 
doesn't matter much (and the timer is only once every few seconds).

I guess it's always better to aovid lock free craziness unless it's
required... so where is the race coming from? I guess 2 CPUs can
clear wd_smp_cpus_pending but not see one another's update so they
both miss cpumask_empty == true! Good catch.

We shouldn't strictly need the wd_smp_lock for the first test, but
that should be an uncommon case, so okay.

Can we clear wd_smp_cpus_pending with a non-atomic operation now?

Thanks,
Nick

Le 27/10/2021 à 05:29, Nicholas Piggin a écrit :

Excerpts from Laurent Dufour's message of October 27, 2021 2:27 am:

quoted

When handling the Watchdog interrupt, long processing should not be done
while holding the __wd_smp_lock. This prevents the other CPUs to grab it
and to process Watchdog timer interrupts. Furhtermore, this could lead to
the following situation:

CPU x detect lockup on CPU y and grab the __wd_smp_lock
       in watchdog_smp_panic()
CPU y caught the watchdog interrupt and try to grab the __wd_smp_lock
       in soft_nmi_interrupt()
CPU x wait for CPU y to catch the IPI for 1s in __smp_send_nmi_ipi()

CPU y should get the IPI here if it's a NMI IPI (which will be true for

quoted

= POWER9 64s).

That said, not all platforms support it and the console lock problem
seems real, so okay.

quoted

CPU x will timeout and so has spent 1s waiting while holding the
       __wd_smp_lock.

A deadlock may also happen between the __wd_smp_lock and the console_owner
'lock' this way:
CPU x grab the console_owner
CPU y grab the __wd_smp_lock
CPU x catch the watchdog timer interrupt and needs to grab __wd_smp_lock
CPU y wants to print something and wait for console_owner
-> deadlock

Doing all the long processing without holding the _wd_smp_lock prevents
these situations.

The intention was to avoid logs getting garbled e.g., if multiple
different CPUs fire at once.

I wonder if instead we could deal with that by protecting the IPI
sending and printing stuff with a trylock, and if you don't get the
trylock then just return, and you'll come back with the next timer
interrupt.

That sounds a bit risky to me, especially on large system when system goes 
wrong, all the CPU may try lock here.
Furthermore, now operation done under the lock protection are quite fast, there 
is no more spinning like the delay loop done when sending an IPI.

Protecting the IPI sending is a nightmare, if the target CPU is later play with 
the lock we are taking during the IPI processing, furthermore, if the target CPU 
is not responding the sending CPU is waiting for 1s, which slows all the system 
due to the lock held.
Since I do a copy of the pending CPU mask and clear it under the lock 
protection, the IPI sending is safe while done without holding the lock.

Regarding the interleaved traces, I don't think this has to be managed down 
here, but rather in the printk/console path.

Cheers,
Laurent.

Thanks,
Nick

quoted

Signed-off-by: Laurent Dufour <redacted>
---
  arch/powerpc/kernel/watchdog.c | 31 +++++++++++++++++--------------
  1 file changed, 17 insertions(+), 14 deletions(-)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index f9ea0e5357f9..bc7411327066 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c

@@ -149,6 +149,8 @@ static void set_cpu_stuck(int cpu, u64 tb)
  
  static void watchdog_smp_panic(int cpu, u64 tb)
  {
+	cpumask_t cpus_pending_copy;
+	u64 last_reset_tb_copy;
  	unsigned long flags;
  	int c;
  

@@ -161,29 +163,32 @@ static void watchdog_smp_panic(int cpu, u64 tb)
  	if (cpumask_weight(&wd_smp_cpus_pending) == 0)
  		goto out;
  
+	cpumask_copy(&cpus_pending_copy, &wd_smp_cpus_pending);
+	last_reset_tb_copy = wd_smp_last_reset_tb;
+
+	/* Take the stuck CPUs out of the watch group */
+	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
+
+	wd_smp_unlock(&flags);
+
  	pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
-		 cpu, cpumask_pr_args(&wd_smp_cpus_pending));
+		 cpu, cpumask_pr_args(&cpus_pending_copy));
  	pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
-		 cpu, tb, wd_smp_last_reset_tb,
-		 tb_to_ns(tb - wd_smp_last_reset_tb) / 1000000);
+		 cpu, tb, last_reset_tb_copy,
+		 tb_to_ns(tb - last_reset_tb_copy) / 1000000);
  
  	if (!sysctl_hardlockup_all_cpu_backtrace) {
  		/*
  		 * Try to trigger the stuck CPUs, unless we are going to
  		 * get a backtrace on all of them anyway.
  		 */
-		for_each_cpu(c, &wd_smp_cpus_pending) {
+		for_each_cpu(c, &cpus_pending_copy) {
  			if (c == cpu)
  				continue;
  			smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
  		}
  	}
  
-	/* Take the stuck CPUs out of the watch group */
-	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
-
-	wd_smp_unlock(&flags);
-
  	if (sysctl_hardlockup_all_cpu_backtrace)
  		trigger_allbutself_cpu_backtrace();
  

@@ -204,6 +209,8 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
  			unsigned long flags;
  
  			wd_smp_lock(&flags);
+			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
+			wd_smp_unlock(&flags);
  
  			pr_emerg("CPU %d became unstuck TB:%lld\n",
  				 cpu, tb);

@@ -212,9 +219,6 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
  				show_regs(regs);
  			else
  				dump_stack();
-
-			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
-			wd_smp_unlock(&flags);
  		}
  		return;
  	}

@@ -267,6 +271,7 @@ DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
  			return 0;
  		}
  		set_cpu_stuck(cpu, tb);
+		wd_smp_unlock(&flags);
  
  		pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
  			 cpu, (void *)regs->nip);

@@ -277,8 +282,6 @@ DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
  		print_irqtrace_events(current);
  		show_regs(regs);
  
-		wd_smp_unlock(&flags);
-
  		if (sysctl_hardlockup_all_cpu_backtrace)
  			trigger_allbutself_cpu_backtrace();
  

-- 
2.33.1

Excerpts from Laurent Dufour's message of October 27, 2021 6:14 pm:

Le 27/10/2021 à 05:29, Nicholas Piggin a écrit :

quoted

Excerpts from Laurent Dufour's message of October 27, 2021 2:27 am:

quoted

When handling the Watchdog interrupt, long processing should not be done
while holding the __wd_smp_lock. This prevents the other CPUs to grab it
and to process Watchdog timer interrupts. Furhtermore, this could lead to
the following situation:

CPU x detect lockup on CPU y and grab the __wd_smp_lock
       in watchdog_smp_panic()
CPU y caught the watchdog interrupt and try to grab the __wd_smp_lock
       in soft_nmi_interrupt()
CPU x wait for CPU y to catch the IPI for 1s in __smp_send_nmi_ipi()

CPU y should get the IPI here if it's a NMI IPI (which will be true for

quoted

= POWER9 64s).

That said, not all platforms support it and the console lock problem
seems real, so okay.

quoted

CPU x will timeout and so has spent 1s waiting while holding the
       __wd_smp_lock.

A deadlock may also happen between the __wd_smp_lock and the console_owner
'lock' this way:
CPU x grab the console_owner
CPU y grab the __wd_smp_lock
CPU x catch the watchdog timer interrupt and needs to grab __wd_smp_lock
CPU y wants to print something and wait for console_owner
-> deadlock

Doing all the long processing without holding the _wd_smp_lock prevents
these situations.

The intention was to avoid logs getting garbled e.g., if multiple
different CPUs fire at once.

I wonder if instead we could deal with that by protecting the IPI
sending and printing stuff with a trylock, and if you don't get the
trylock then just return, and you'll come back with the next timer
interrupt.

That sounds a bit risky to me, especially on large system when system goes 
wrong, all the CPU may try lock here.

That should be okay though, one will get through and the others will 
skip.

Furthermore, now operation done under the lock protection are quite fast, there 
is no more spinning like the delay loop done when sending an IPI.

Protecting the IPI sending is a nightmare, if the target CPU is later play with 
the lock we are taking during the IPI processing, furthermore, if the target CPU 
is not responding the sending CPU is waiting for 1s, which slows all the system 
due to the lock held.
Since I do a copy of the pending CPU mask and clear it under the lock 
protection, the IPI sending is safe while done without holding the lock.

Protecting IPI sending basically has all the same issues in the NMI
IPI layer.

Regarding the interleaved traces, I don't think this has to be managed down 
here, but rather in the printk/console path.

It can't necessarily be because some of the problem is actually that a 
NMI handler can be interrupted by another NMI IPI because the caller
can return only after handlers start running rather than complete.

I don't think it would be an additional nightmare to trylock.

Thanks,
Nick

Excerpts from Nicholas Piggin's message of October 27, 2021 1:29 pm:

Excerpts from Laurent Dufour's message of October 27, 2021 2:27 am:

quoted

When handling the Watchdog interrupt, long processing should not be done
while holding the __wd_smp_lock. This prevents the other CPUs to grab it
and to process Watchdog timer interrupts. Furhtermore, this could lead to
the following situation:

CPU x detect lockup on CPU y and grab the __wd_smp_lock
      in watchdog_smp_panic()
CPU y caught the watchdog interrupt and try to grab the __wd_smp_lock
      in soft_nmi_interrupt()
CPU x wait for CPU y to catch the IPI for 1s in __smp_send_nmi_ipi()

CPU y should get the IPI here if it's a NMI IPI (which will be true for 

quoted

= POWER9 64s).

That said, not all platforms support it and the console lock problem
seems real, so okay.

quoted

CPU x will timeout and so has spent 1s waiting while holding the
      __wd_smp_lock.

A deadlock may also happen between the __wd_smp_lock and the console_owner
'lock' this way:
CPU x grab the console_owner
CPU y grab the __wd_smp_lock
CPU x catch the watchdog timer interrupt and needs to grab __wd_smp_lock
CPU y wants to print something and wait for console_owner
-> deadlock

Doing all the long processing without holding the _wd_smp_lock prevents
these situations.

The intention was to avoid logs getting garbled e.g., if multiple 
different CPUs fire at once.

I wonder if instead we could deal with that by protecting the IPI 
sending and printing stuff with a trylock, and if you don't get the
trylock then just return, and you'll come back with the next timer
interrupt.

Something like this (untested) should basically hold off concurrency on
watchdog panics. It does not serialize output from IPI targets but it
should prevent multiple CPUs trying to send IPIs at once, without 
holding the lock.

---

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index 2ffeb3f8b5a7..3a0db577da56 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c

@@ -85,6 +85,7 @@ static DEFINE_PER_CPU(u64, wd_timer_tb);
 
 /* SMP checker bits */
 static unsigned long __wd_smp_lock;
+static unsigned long __wd_printing;
 static cpumask_t wd_smp_cpus_pending;
 static cpumask_t wd_smp_cpus_stuck;
 static u64 wd_smp_last_reset_tb;

@@ -131,10 +132,10 @@ static void wd_lockup_ipi(struct pt_regs *regs)
 	/* Do not panic from here because that can recurse into NMI IPI layer */
 }
 
-static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
+static void set_cpu_stuck(int cpu, u64 tb)
 {
-	cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
-	cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
+	cpumask_set_cpu(cpu, &wd_smp_cpus_stuck);
+	cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 	if (cpumask_empty(&wd_smp_cpus_pending)) {
 		wd_smp_last_reset_tb = tb;
 		cpumask_andnot(&wd_smp_cpus_pending,

@@ -142,10 +143,6 @@ static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
 				&wd_smp_cpus_stuck);
 	}
 }
-static void set_cpu_stuck(int cpu, u64 tb)
-{
-	set_cpumask_stuck(cpumask_of(cpu), tb);
-}
 
 static void watchdog_smp_panic(int cpu, u64 tb)
 {

@@ -160,6 +157,10 @@ static void watchdog_smp_panic(int cpu, u64 tb)
 		goto out;
 	if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 		goto out;
+	if (__wd_printing)
+		goto out;
+	__wd_printing = 1;
+	wd_smp_unlock(&flags);
 
 	pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
 		 cpu, cpumask_pr_args(&wd_smp_cpus_pending));

@@ -172,24 +173,31 @@ static void watchdog_smp_panic(int cpu, u64 tb)
 		 * Try to trigger the stuck CPUs, unless we are going to
 		 * get a backtrace on all of them anyway.
 		 */
-		for_each_cpu(c, &wd_smp_cpus_pending) {
+		for_each_online_cpu(c) {
 			if (c == cpu)
 				continue;
+			if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
+				continue;
+			wd_smp_lock(&flags);
+			if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
+				wd_smp_unlock(&flags);
+				continue;
+			}
+			/* Take the stuck CPU out of the watch group */
+			set_cpu_stuck(cpu, tb);
+			wd_smp_unlock(&flags);
+
 			smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 		}
 	}
 
-	/* Take the stuck CPUs out of the watch group */
-	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
-
-	wd_smp_unlock(&flags);
-
 	if (sysctl_hardlockup_all_cpu_backtrace)
 		trigger_allbutself_cpu_backtrace();
 
 	if (hardlockup_panic)
 		nmi_panic(NULL, "Hard LOCKUP");
 
+	__wd_printing = 0;
 	return;
 
 out:

@@ -204,8 +212,6 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 		if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 			struct pt_regs *regs = get_irq_regs();
 
-			wd_smp_lock(&flags);
-
 			pr_emerg("CPU %d became unstuck TB:%lld\n",
 				 cpu, tb);
 			print_irqtrace_events(current);

@@ -214,6 +220,7 @@ static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 			else
 				dump_stack();
 
+			wd_smp_lock(&flags);
 			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 			wd_smp_unlock(&flags);
 		}

@@ -263,8 +270,16 @@ DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
 			wd_smp_unlock(&flags);
 			return 0;
 		}
+		if (__wd_printing) {
+			wd_smp_unlock(&flags);
+			return 0;
+		}
+		__wd_printing = 1;
+
 		set_cpu_stuck(cpu, tb);
 
+		wd_smp_unlock(&flags);
+
 		pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
 			 cpu, (void *)regs->nip);
 		pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",

@@ -274,13 +289,13 @@ DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
 		print_irqtrace_events(current);
 		show_regs(regs);
 
-		wd_smp_unlock(&flags);
-
 		if (sysctl_hardlockup_all_cpu_backtrace)
 			trigger_allbutself_cpu_backtrace();
 
 		if (hardlockup_panic)
 			nmi_panic(regs, "Hard LOCKUP");
+
+		__wd_printing = 0;
 	}
 	if (wd_panic_timeout_tb < 0x7fffffff)
 		mtspr(SPRN_DEC, wd_panic_timeout_tb);

Le 27/10/2021 à 11:49, Nicholas Piggin a écrit :

Excerpts from Nicholas Piggin's message of October 27, 2021 1:29 pm:

quoted

Excerpts from Laurent Dufour's message of October 27, 2021 2:27 am:

quoted

When handling the Watchdog interrupt, long processing should not be done
while holding the __wd_smp_lock. This prevents the other CPUs to grab it
and to process Watchdog timer interrupts. Furhtermore, this could lead to
the following situation:

CPU x detect lockup on CPU y and grab the __wd_smp_lock
       in watchdog_smp_panic()
CPU y caught the watchdog interrupt and try to grab the __wd_smp_lock
       in soft_nmi_interrupt()
CPU x wait for CPU y to catch the IPI for 1s in __smp_send_nmi_ipi()

CPU y should get the IPI here if it's a NMI IPI (which will be true for

quoted

= POWER9 64s).

That said, not all platforms support it and the console lock problem
seems real, so okay.

quoted

CPU x will timeout and so has spent 1s waiting while holding the
       __wd_smp_lock.

A deadlock may also happen between the __wd_smp_lock and the console_owner
'lock' this way:
CPU x grab the console_owner
CPU y grab the __wd_smp_lock
CPU x catch the watchdog timer interrupt and needs to grab __wd_smp_lock
CPU y wants to print something and wait for console_owner
-> deadlock

Doing all the long processing without holding the _wd_smp_lock prevents
these situations.

The intention was to avoid logs getting garbled e.g., if multiple
different CPUs fire at once.

I wonder if instead we could deal with that by protecting the IPI
sending and printing stuff with a trylock, and if you don't get the
trylock then just return, and you'll come back with the next timer
interrupt.

Something like this (untested) should basically hold off concurrency on
watchdog panics. It does not serialize output from IPI targets but it
should prevent multiple CPUs trying to send IPIs at once, without
holding the lock.

Got it, I'll work this way, despite the minor comments below.

I think WRITE_ONCE() is required here, to prevent any compiler trick.
Also, I think it might be safer (I don't have clear example in mind, just a bad 
feeling) to do so before the check of hardlockup_panic and the call to panic.

Here the CPU is not flagged as stuck and nothing is printed. We have to wait for 
the next soft_nmi_interrupt() to happen for this CPU to detect itself stuck and 
provide the backstack. Depending on the number of CPU and the stress on the 
system, this CPU may net see __wd_printing=0 when entering soft_nmi_interrupt() 
until a long period of time.
I think we should introduce a per CPU counter, when that counter reach a limit 
(let's say 10), the CPU is not taking care of __wd_printing and continue it's 
processing.

I think WRITE_ONCE() is required here, to prevent any compiler trick.

  	}
  	if (wd_panic_timeout_tb < 0x7fffffff)
  		mtspr(SPRN_DEC, wd_panic_timeout_tb);