Skip to contentSkip to conversation

powerpc/64s/interrupt: Fix decrementer storm

PATCH In review Nicholas Piggin 2 participants 5 messages Jan 24, 2022 – Mar 21, 2022 PowerPC

lore
Share & tools
lore

Nicholas Piggin

author reviewer
#0 lore raw

PATCH 1/3 powerpc/64s/interrupt: Fix decrementer storm

The decrementer exception can fail to be cleared when the interrupt returns in the case where the decrementer wraps with the next timer still beyond decrementer_max. This results in a decrementer interrupt storm. This is triggerable with small decrementer system with hard and soft watchdogs disabled.

Fix this by always programming the decrementer if there was no timer.

Fixes 0faf20a1ad16 ("powerpc/64s/interrupt: Don't enable MSR[EE] in irq handlers unless perf is in use")
---
 arch/powerpc/kernel/time.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
arch/powerpc/kernel/time.c +3 -2
index 62361cc7281c..cd0b8b71ecdd 100644
@@ -649,8 +649,9 @@ DEFINE_INTERRUPT_HANDLER_ASYNC(timer_interrupt)
 		__this_cpu_inc(irq_stat.timer_irqs_event);
 	} else {
 		now = *next_tb - now;
-		if (now <= decrementer_max)
-			set_dec_or_work(now);
+		if (now > decrementer_max)
+			now = decrementer_max;
+		set_dec_or_work(now);
 		__this_cpu_inc(irq_stat.timer_irqs_others);
 	}
 
Signature · 1 line
2.23.0
author reviewer
#1 lore raw

PATCH 2/3 powerpc/time: Fix KVM host re-arming a timer beyond decrementer range

If the next host timer is beyond decrementer range, timer_rearm_host_dec will leave decrementer not programmed. This will not cause a problem for the host it will just set the decrementer correctly when the decrementer interrupt hits, it seems safer not to leave the next host decrementer interrupt timing able to be influenced by a guest.

This code is only used in the P9 KVM paths so it's unlikely to be hit practically unless large decrementer is force disabled in the host.

Fixes 25aa145856cd ("powerpc/time: add API for KVM to re-arm the host timer/decrementer")
---
 arch/powerpc/kernel/time.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
arch/powerpc/kernel/time.c +3 -2
index cd0b8b71ecdd..384f58a3f373 100644
@@ -582,8 +582,9 @@ void timer_rearm_host_dec(u64 now)
 		local_paca->irq_happened |= PACA_IRQ_DEC;
 	} else {
 		now = *next_tb - now;
-		if (now <= decrementer_max)
-			set_dec_or_work(now);
+		if (now > decrementer_max)
+			now = decrementer_max;
+		set_dec_or_work(now);
 	}
 }
 EXPORT_SYMBOL_GPL(timer_rearm_host_dec);
Signature · 1 line
2.23.0
author reviewer
#2 lore raw

PATCH 3/3 powerpc/time: improve decrementer clockevent processing

In reply to Nicholas Piggin

The stop/shutdown op should not use decrementer_set_next_event because that sets decrementers_next_tb to now + decrementer_max, which means a decrementer interrupt that occurs after that time will call the clockevent event handler unexpectedly. Set next_tb to ~0 here to prevent any clock event call. Init all clockevents to stopped.

Then the decrementer clockevent device always has event_handler set and applicable because we know the clock event device was not stopped. So make this call unconditional to show that it is always called. next_tb need not be set to ~0 before the event handler is called because it will stop the clockevent device if there is no other timer.

Finally, the timer broadcast interrupt should not modify next_tb because it is not involved with the local decrementer clockevent on this CPU.

This doesn't fix a known bug, just tidies the code.

---
 arch/powerpc/kernel/time.c | 18 ++++++++++--------
 1 file changed, 10 insertions(+), 8 deletions(-)
arch/powerpc/kernel/time.c +10 -8
index 384f58a3f373..f3845601ab6a 100644
@@ -107,7 +107,12 @@ struct clock_event_device decrementer_clockevent = {
 };
 EXPORT_SYMBOL(decrementer_clockevent);
 
-DEFINE_PER_CPU(u64, decrementers_next_tb);
+/*
+ * This always puts next_tb beyond now, so the clock event will never fire
+ * with the usual comparison, no need for a separate test for stopped.
+ */
+#define DEC_CLOCKEVENT_STOPPED ~0ULL
+DEFINE_PER_CPU(u64, decrementers_next_tb) = DEC_CLOCKEVENT_STOPPED;
 EXPORT_SYMBOL_GPL(decrementers_next_tb);
 static DEFINE_PER_CPU(struct clock_event_device, decrementers);
 
@@ -644,9 +649,7 @@ DEFINE_INTERRUPT_HANDLER_ASYNC(timer_interrupt)
 
 	now = get_tb();
 	if (now >= *next_tb) {
-		*next_tb = ~(u64)0;
-		if (evt->event_handler)
-			evt->event_handler(evt);
+		evt->event_handler(evt);
 		__this_cpu_inc(irq_stat.timer_irqs_event);
 	} else {
 		now = *next_tb - now;
@@ -665,9 +668,6 @@ EXPORT_SYMBOL(timer_interrupt);
 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 void timer_broadcast_interrupt(void)
 {
-	u64 *next_tb = this_cpu_ptr(&decrementers_next_tb);
-
-	*next_tb = ~(u64)0;
 	tick_receive_broadcast();
 	__this_cpu_inc(irq_stat.broadcast_irqs_event);
 }
@@ -893,7 +893,9 @@ static int decrementer_set_next_event(unsigned long evt,
 
 static int decrementer_shutdown(struct clock_event_device *dev)
 {
-	decrementer_set_next_event(decrementer_max, dev);
+	__this_cpu_write(decrementers_next_tb, DEC_CLOCKEVENT_STOPPED);
+	set_dec_or_work(decrementer_max);
+
 	return 0;
 }
 
Signature · 1 line
2.23.0
#3 lore raw

Re: (subset) [PATCH 1/3] powerpc/64s/interrupt: Fix decrementer storm

In reply to Nicholas Piggin
Quoted 9 lines · Nicholas Piggin

The decrementer exception can fail to be cleared when the interrupt returns in the case where the decrementer wraps with the next timer still beyond decrementer_max. This results in a decrementer interrupt storm. This is triggerable with small decrementer system with hard and soft watchdogs disabled.

Fix this by always programming the decrementer if there was no timer.

[...]

Patch 1 applied to powerpc/fixes.

[1/3] powerpc/64s/interrupt: Fix decrementer storm

      https://git.kernel.org/powerpc/c/8defc2a5dd8f4c0cb19ecbaca8d3e89ab98524da

cheers

#4 lore raw
In reply to Nicholas Piggin
Quoted 9 lines · Nicholas Piggin

The decrementer exception can fail to be cleared when the interrupt returns in the case where the decrementer wraps with the next timer still beyond decrementer_max. This results in a decrementer interrupt storm. This is triggerable with small decrementer system with hard and soft watchdogs disabled.

Fix this by always programming the decrementer if there was no timer.

[...]

Patch 2 & 3 applied to powerpc/next.

[2/3] powerpc/time: Fix KVM host re-arming a timer beyond decrementer range

      https://git.kernel.org/powerpc/c/cf74ff52e352112be78c4c4c3637a37ec36a6608
[3/3] powerpc/time: improve decrementer clockevent processing
      https://git.kernel.org/powerpc/c/35de589cb8793573ed56a915af9cb4b5f15ad7d7

cheers