Can RCU stall lead to hard lockups?

Hello, Paul.

So, I was looking at a machine which triggered crashdump from NMI hard lockup. The dmesg was filled up with backtraces - many were stuck in reclaim path, which seems to be the culprit, the others were stuck in RCU path. It looks like by the time crashdump was created, all CPUs got stuck in RCU path.

One of the CPUs looked like the following.

PID: 4149739 TASK: ffff881a46baa880 CPU: 13 COMMAND: "CPUThreadPool8"
#0 [ffff881fff945e48] crash_nmi_callback at ffffffff8103f7d0 #1 [ffff881fff945e58] nmi_handle at ffffffff81020653
#2 [ffff881fff945eb0] default_do_nmi at ffffffff81020c36
#3 [ffff881fff945ed0] do_nmi at ffffffff81020d32
#4 [ffff881fff945ef0] end_repeat_nmi at ffffffff81956a7e

      [exception RIP: io_serial_in+21]
      RIP: ffffffff81630e55  RSP: ffff881fff943b88  RFLAGS: 00000002
      RAX: 000000000000ca00  RBX: ffffffff8230e188  RCX: 0000000000000000
      RDX: 00000000000002fd  RSI: 0000000000000005  RDI: ffffffff8230e188
      RBP: ffff881fff943bb0   R8: 0000000000000000   R9: ffffffff820cb3c4
      R10: 0000000000000019  R11: 0000000000002000  R12: 00000000000026e1
      R13: 0000000000000020  R14: ffffffff820cd398  R15: 0000000000000035
      ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0000

--- <NMI exception stack> ---
#5 [ffff881fff943b88] io_serial_in at ffffffff81630e55
#6 [ffff881fff943b90] wait_for_xmitr at ffffffff8163175c #7 [ffff881fff943bb8] serial8250_console_putchar at ffffffff816317dc #8 [ffff881fff943bd8] uart_console_write at ffffffff8162ac00 #9 [ffff881fff943c08] serial8250_console_write at ffffffff81634691
#10 [ffff881fff943c80] univ8250_console_write at ffffffff8162f7c2 #11 [ffff881fff943c90] console_unlock at ffffffff810dfc55 #12 [ffff881fff943cf0] vprintk_emit at ffffffff810dffb5 #13 [ffff881fff943d50] vprintk_default at ffffffff810e01bf

#14 [ffff881fff943d60] vprintk_func at ffffffff810e1127 #15 [ffff881fff943d70] printk at ffffffff8119a8a4
#16 [ffff881fff943dd0] print_cpu_stall_info at ffffffff810eb78c #17 [ffff881fff943e88] rcu_check_callbacks at ffffffff810ef133 #18 [ffff881fff943ee8] update_process_times at ffffffff810f3497 #19 [ffff881fff943f10] tick_sched_timer at ffffffff81103037 #20 [ffff881fff943f38] __hrtimer_run_queues at ffffffff810f3f38 #21 [ffff881fff943f88] hrtimer_interrupt at ffffffff810f442b

The CPU which triggered hardlockup:

PID: 4149709 TASK: ffff88010f88c380 CPU: 26 COMMAND: "CPUThreadPool35"
#0 [ffff883fff1059d0] machine_kexec at ffffffff8104a874
#1 [ffff883fff105a30] __crash_kexec at ffffffff811116cc
#2 [ffff883fff105af0] __crash_kexec at ffffffff81111795
#3 [ffff883fff105b08] panic at ffffffff8119a6ae
#4 [ffff883fff105b98] watchdog_overflow_callback at ffffffff81135dbd #5 [ffff883fff105bb0] __perf_event_overflow at ffffffff81186866 #6 [ffff883fff105be8] perf_event_overflow at ffffffff81192bc4 #7 [ffff883fff105bf8] intel_pmu_handle_irq at ffffffff8100b265 #8 [ffff883fff105df8] perf_event_nmi_handler at ffffffff8100489f #9 [ffff883fff105e58] nmi_handle at ffffffff81020653
#10 [ffff883fff105eb0] default_do_nmi at ffffffff81020b94
#11 [ffff883fff105ed0] do_nmi at ffffffff81020d32
#12 [ffff883fff105ef0] end_repeat_nmi at ffffffff81956a7e

      [exception RIP: queued_spin_lock_slowpath+248]
      RIP: ffffffff810da958  RSP: ffff883fff103e68  RFLAGS: 00000046
      RAX: 0000000000000000  RBX: 0000000000000046  RCX: 00000000006d0000
      RDX: ffff883fff49a950  RSI: 0000000000d10101  RDI: ffffffff81e54300
      RBP: ffff883fff103e80   R8: ffff883fff11a950   R9: 0000000000000000
      R10: 000000000e5873ba  R11: 000000000000010f  R12: ffffffff81e54300
      R13: 0000000000000000  R14: ffff88010f88c380  R15: ffffffff81e54300
      ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018

--- <NMI exception stack> ---
#13 [ffff883fff103e68] queued_spin_lock_slowpath at ffffffff810da958 #14 [ffff883fff103e70] _raw_spin_lock_irqsave at ffffffff8195550b #15 [ffff883fff103e88] rcu_check_callbacks at ffffffff810eed18 #16 [ffff883fff103ee8] update_process_times at ffffffff810f3497 #17 [ffff883fff103f10] tick_sched_timer at ffffffff81103037 #18 [ffff883fff103f38] __hrtimer_run_queues at ffffffff810f3f38 #19 [ffff883fff103f88] hrtimer_interrupt at ffffffff810f442b --- <IRQ stack> ---
bt: read error: kernel virtual address: ffffffffffffffd8 type: "pt_regs"

All other CPUs have something similar to the following:

PID: 4150460 TASK: ffff88162c891b00 CPU: 8 COMMAND: "GlobalIOThreadP"
#0 [ffff881fff805e48] crash_nmi_callback at ffffffff8103f7d0 #1 [ffff881fff805e58] nmi_handle at ffffffff81020653 #2 [ffff881fff805eb0] default_do_nmi at ffffffff81020c36 #3 [ffff881fff805ed0] do_nmi at ffffffff81020d32 #4 [ffff881fff805ef0] end_repeat_nmi at ffffffff81956a7e

      [exception RIP: queued_spin_lock_slowpath+248]
      RIP: ffffffff810da958  RSP: ffff881fff803e68  RFLAGS: 00000046
      RAX: 0000000000000000  RBX: 0000000000000046  RCX: 0000000000250000
      RDX: ffff881fffb1a950  RSI: 00000000002d0101  RDI: ffffffff81e54300
      RBP: ffff881fff803e80   R8: ffff881fff81a950   R9: 0000000000000000
      R10: 000000000e5873ba  R11: ffff88207fffd000  R12: ffffffff81e54300
      R13: 0000000000000000  R14: ffff88162c891b00  R15: ffffffff81e54300
      ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0000

--- <NMI exception stack> ---
#5 [ffff881fff803e68] queued_spin_lock_slowpath at ffffffff810da958 #6 [ffff881fff803e70] _raw_spin_lock_irqsave at ffffffff8195550b #7 [ffff881fff803e88] rcu_check_callbacks at ffffffff810eed18 #8 [ffff881fff803ee8] update_process_times at ffffffff810f3497 #9 [ffff881fff803f10] tick_sched_timer at ffffffff81103037
#10 [ffff881fff803f38] __hrtimer_run_queues at ffffffff810f3f38 #11 [ffff881fff803f88] hrtimer_interrupt at ffffffff810f442b --- <IRQ stack> ---

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

kernel/rcu/tree_plugin.h +13 -1

index db85ca3..3c4c4d3 100644
@@ -561,8 +561,14 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 	}
 	t = list_entry(rnp->gp_tasks->prev,
 		       struct task_struct, rcu_node_entry);
-	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
+		touch_nmi_watchdog();
+		/*
+		 * We could be printing a lot of these messages while
+		 * holding a spinlock.  Avoid triggering hard lockup.
+		 */
 		sched_show_task(t);
+	}
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 }
 

@@ -1678,6 +1684,12 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 	char *ticks_title;
 	unsigned long ticks_value;
 
+	/*
+	 * We could be printing a lot of these messages while holding a
+	 * spinlock.  Avoid triggering hard lockup.
+	 */
+	touch_nmi_watchdog();
+
 	if (rsp->gpnum == rdp->gpnum) {
 		ticks_title = "ticks this GP";
 		ticks_value = rdp->ticks_this_gp;

From 45f8e12038020e44ef7d589de9247ad4fcd62f44 Mon Sep 17 00:00:00 2001

From c1d1ae75ba4bde7c833560248b3f561085fa850e Mon Sep 17 00:00:00 2001

Quoted 1 line

From c1d1ae75ba4bde7c833560248b3f561085fa850e Mon Sep 17 00:00:00 2001

From: Tejun Heo <[email protected]>
Date: Tue, 9 Jan 2018 10:38:17 -0800
Subject: [PATCH] RCU: Call touch_nmi_watchdog() while printing stall warnings

When RCU stall warning triggers, it can print out a lot of messages while holding spinlocks. If the console device is slow (e.g. an actual or IPMI serial console), it may end up triggering NMI hard lockup watchdog like the following.

*** CPU printking while holding RCU spinlock

PID: 4149739 TASK: ffff881a46baa880 CPU: 13 COMMAND: "CPUThreadPool8"
#0 [ffff881fff945e48] crash_nmi_callback at ffffffff8103f7d0 #1 [ffff881fff945e58] nmi_handle at ffffffff81020653
#2 [ffff881fff945eb0] default_do_nmi at ffffffff81020c36
#3 [ffff881fff945ed0] do_nmi at ffffffff81020d32
#4 [ffff881fff945ef0] end_repeat_nmi at ffffffff81956a7e

      [exception RIP: io_serial_in+21]
      RIP: ffffffff81630e55  RSP: ffff881fff943b88  RFLAGS: 00000002
      RAX: 000000000000ca00  RBX: ffffffff8230e188  RCX: 0000000000000000
      RDX: 00000000000002fd  RSI: 0000000000000005  RDI: ffffffff8230e188
      RBP: ffff881fff943bb0   R8: 0000000000000000   R9: ffffffff820cb3c4
      R10: 0000000000000019  R11: 0000000000002000  R12: 00000000000026e1
      R13: 0000000000000020  R14: ffffffff820cd398  R15: 0000000000000035
      ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0000

--- <NMI exception stack> ---
#5 [ffff881fff943b88] io_serial_in at ffffffff81630e55
#6 [ffff881fff943b90] wait_for_xmitr at ffffffff8163175c #7 [ffff881fff943bb8] serial8250_console_putchar at ffffffff816317dc #8 [ffff881fff943bd8] uart_console_write at ffffffff8162ac00 #9 [ffff881fff943c08] serial8250_console_write at ffffffff81634691
#10 [ffff881fff943c80] univ8250_console_write at ffffffff8162f7c2 #11 [ffff881fff943c90] console_unlock at ffffffff810dfc55 #12 [ffff881fff943cf0] vprintk_emit at ffffffff810dffb5 #13 [ffff881fff943d50] vprintk_default at ffffffff810e01bf #14 [ffff881fff943d60] vprintk_func at ffffffff810e1127 #15 [ffff881fff943d70] printk at ffffffff8119a8a4
#16 [ffff881fff943dd0] print_cpu_stall_info at ffffffff810eb78c #17 [ffff881fff943e88] rcu_check_callbacks at ffffffff810ef133 #18 [ffff881fff943ee8] update_process_times at ffffffff810f3497 #19 [ffff881fff943f10] tick_sched_timer at ffffffff81103037 #20 [ffff881fff943f38] __hrtimer_run_queues at ffffffff810f3f38 #21 [ffff881fff943f88] hrtimer_interrupt at ffffffff810f442b

*** CPU triggering the hardlockup watchdog

PID: 4149709 TASK: ffff88010f88c380 CPU: 26 COMMAND: "CPUThreadPool35"
#0 [ffff883fff1059d0] machine_kexec at ffffffff8104a874
#1 [ffff883fff105a30] __crash_kexec at ffffffff811116cc
#2 [ffff883fff105af0] __crash_kexec at ffffffff81111795
#3 [ffff883fff105b08] panic at ffffffff8119a6ae
#4 [ffff883fff105b98] watchdog_overflow_callback at ffffffff81135dbd #5 [ffff883fff105bb0] __perf_event_overflow at ffffffff81186866 #6 [ffff883fff105be8] perf_event_overflow at ffffffff81192bc4 #7 [ffff883fff105bf8] intel_pmu_handle_irq at ffffffff8100b265 #8 [ffff883fff105df8] perf_event_nmi_handler at ffffffff8100489f #9 [ffff883fff105e58] nmi_handle at ffffffff81020653
#10 [ffff883fff105eb0] default_do_nmi at ffffffff81020b94
#11 [ffff883fff105ed0] do_nmi at ffffffff81020d32
#12 [ffff883fff105ef0] end_repeat_nmi at ffffffff81956a7e

      [exception RIP: queued_spin_lock_slowpath+248]
      RIP: ffffffff810da958  RSP: ffff883fff103e68  RFLAGS: 00000046
      RAX: 0000000000000000  RBX: 0000000000000046  RCX: 00000000006d0000
      RDX: ffff883fff49a950  RSI: 0000000000d10101  RDI: ffffffff81e54300
      RBP: ffff883fff103e80   R8: ffff883fff11a950   R9: 0000000000000000
      R10: 000000000e5873ba  R11: 000000000000010f  R12: ffffffff81e54300
      R13: 0000000000000000  R14: ffff88010f88c380  R15: ffffffff81e54300
      ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018

--- <NMI exception stack> ---
#13 [ffff883fff103e68] queued_spin_lock_slowpath at ffffffff810da958 #14 [ffff883fff103e70] _raw_spin_lock_irqsave at ffffffff8195550b #15 [ffff883fff103e88] rcu_check_callbacks at ffffffff810eed18 #16 [ffff883fff103ee8] update_process_times at ffffffff810f3497 #17 [ffff883fff103f10] tick_sched_timer at ffffffff81103037 #18 [ffff883fff103f38] __hrtimer_run_queues at ffffffff810f3f38 #19 [ffff883fff103f88] hrtimer_interrupt at ffffffff810f442b --- <IRQ stack> ---

Avoid spuriously triggering NMI hardlockup watchdog by touching it from the print functions. show_state_filter() shares the same problem and solution.

v2: Relocate the comment to where it belongs.

Signed-off-by Tejun Heo <[email protected]>

---
 kernel/rcu/tree_plugin.h | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

kernel/rcu/tree_plugin.h +13 -1

index db85ca3..bf05bf9 100644
@@ -561,8 +561,14 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 	}
 	t = list_entry(rnp->gp_tasks->prev,
 		       struct task_struct, rcu_node_entry);
-	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
+		/*
+		 * We could be printing a lot while holding a spinlock.
+		 * Avoid triggering hard lockup.
+		 */
+		touch_nmi_watchdog();
 		sched_show_task(t);
+	}
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 }
 

@@ -1678,6 +1684,12 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 	char *ticks_title;
 	unsigned long ticks_value;
 
+	/*
+	 * We could be printing a lot while holding a spinlock.  Avoid
+	 * triggering hard lockup.
+	 */
+	touch_nmi_watchdog();
+
 	if (rsp->gpnum == rdp->gpnum) {
 		ticks_title = "ticks this GP";
 		ticks_value = rdp->ticks_this_gp;

Signature · 1 line

2.9.5

When RCU stall warning triggers, it can print out a lot of messages while holding spinlocks. If the console device is slow (e.g. an actual or IPMI serial console), it may end up triggering NMI hard lockup watchdog like the following.

*** CPU printking while holding RCU spinlock

PID: 4149739 TASK: ffff881a46baa880 CPU: 13 COMMAND: "CPUThreadPool8"
#0 [ffff881fff945e48] crash_nmi_callback at ffffffff8103f7d0 #1 [ffff881fff945e58] nmi_handle at ffffffff81020653
#2 [ffff881fff945eb0] default_do_nmi at ffffffff81020c36
#3 [ffff881fff945ed0] do_nmi at ffffffff81020d32
#4 [ffff881fff945ef0] end_repeat_nmi at ffffffff81956a7e

      [exception RIP: io_serial_in+21]
      RIP: ffffffff81630e55  RSP: ffff881fff943b88  RFLAGS: 00000002
      RAX: 000000000000ca00  RBX: ffffffff8230e188  RCX: 0000000000000000
      RDX: 00000000000002fd  RSI: 0000000000000005  RDI: ffffffff8230e188
      RBP: ffff881fff943bb0   R8: 0000000000000000   R9: ffffffff820cb3c4
      R10: 0000000000000019  R11: 0000000000002000  R12: 00000000000026e1
      R13: 0000000000000020  R14: ffffffff820cd398  R15: 0000000000000035
      ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0000

--- <NMI exception stack> ---
#5 [ffff881fff943b88] io_serial_in at ffffffff81630e55
#6 [ffff881fff943b90] wait_for_xmitr at ffffffff8163175c #7 [ffff881fff943bb8] serial8250_console_putchar at ffffffff816317dc #8 [ffff881fff943bd8] uart_console_write at ffffffff8162ac00 #9 [ffff881fff943c08] serial8250_console_write at ffffffff81634691
#10 [ffff881fff943c80] univ8250_console_write at ffffffff8162f7c2 #11 [ffff881fff943c90] console_unlock at ffffffff810dfc55 #12 [ffff881fff943cf0] vprintk_emit at ffffffff810dffb5 #13 [ffff881fff943d50] vprintk_default at ffffffff810e01bf

#14 [ffff881fff943d60] vprintk_func at ffffffff810e1127 #15 [ffff881fff943d70] printk at ffffffff8119a8a4
#16 [ffff881fff943dd0] print_cpu_stall_info at ffffffff810eb78c #17 [ffff881fff943e88] rcu_check_callbacks at ffffffff810ef133 #18 [ffff881fff943ee8] update_process_times at ffffffff810f3497 #19 [ffff881fff943f10] tick_sched_timer at ffffffff81103037 #20 [ffff881fff943f38] __hrtimer_run_queues at ffffffff810f3f38 #21 [ffff881fff943f88] hrtimer_interrupt at ffffffff810f442b

Tejun, while we are at it, shall we also touch NMI in show_workqueue_state()? I suppose that show_pwq() can be quite verbose on some systems, can't it?

for instance: sysrq-t [irq] -> show_workqueue_state()

kernel/workqueue.c +1 -0

index 9960f8842b31..649c00a9af12 100644
@@ -48,6 +48,7 @@
 #include <linux/moduleparam.h>
 #include <linux/uaccess.h>
 #include <linux/sched/isolation.h>
+#include <linux/nmi.h>
 
 #include "workqueue_internal.h"
 

Quoted 12 lines

kernel/workqueue.c +1 -0

index 9960f8842b31..649c00a9af12 100644
@@ -48,6 +48,7 @@
 #include <linux/moduleparam.h>
 #include <linux/uaccess.h>
 #include <linux/sched/isolation.h>
+#include <linux/nmi.h>
 
 #include "workqueue_internal.h"
 

kernel/workqueue.c void show_workqueue_state(void) · line 4478

Quoted 6 lines

kernel/workqueue.c +2 -0

@@ -4473,6 +4474,8 @@ void show_workqueue_state(void)
 			if (pwq->nr_active || !list_empty(&pwq->delayed_works))
 				show_pwq(pwq);
 			spin_unlock_irqrestore(&pwq->pool->lock, flags);
+
+			touch_nmi_watchdog();

Can you send a patch w/ signed-off-by adding touch_nmi_watchdog() to both outer loops in that function. Theoretically, we can have a lot of pools too.

Quoted 21 lines · Tejun Heo

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Quoted 4 lines

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

If this fixes things for you, I would welcome such a patch.

Thanx, Paul

Quoted 34 lines

kernel/rcu/tree_plugin.h +13 -1

index db85ca3..3c4c4d3 100644
@@ -561,8 +561,14 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 	}
 	t = list_entry(rnp->gp_tasks->prev,
 		       struct task_struct, rcu_node_entry);
-	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
+		touch_nmi_watchdog();
+		/*
+		 * We could be printing a lot of these messages while
+		 * holding a spinlock.  Avoid triggering hard lockup.
+		 */
 		sched_show_task(t);
+	}
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 }
 

@@ -1678,6 +1684,12 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 	char *ticks_title;
 	unsigned long ticks_value;
 
+	/*
+	 * We could be printing a lot of these messages while holding a
+	 * spinlock.  Avoid triggering hard lockup.
+	 */
+	touch_nmi_watchdog();
+
 	if (rsp->gpnum == rdp->gpnum) {
 		ticks_title = "ticks this GP";
 		ticks_value = rdp->ticks_this_gp;

Quoted 27 lines · Paul E. McKenney

Quoted 25 lines · Tejun Heo

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

Quoted 1 line

If this fixes things for you, I would welcome such a patch.

Hi - would this also be relevant to 4.9-stable and 4.4-stable, or has something elsewhere changed after 4.9 that actually triggers this?

thanks,
-serge

Quoted 36 lines

Thanx, Paul

Quoted 34 lines

kernel/rcu/tree_plugin.h +13 -1

index db85ca3..3c4c4d3 100644
@@ -561,8 +561,14 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 	}
 	t = list_entry(rnp->gp_tasks->prev,
 		       struct task_struct, rcu_node_entry);
-	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
+		touch_nmi_watchdog();
+		/*
+		 * We could be printing a lot of these messages while
+		 * holding a spinlock.  Avoid triggering hard lockup.
+		 */
 		sched_show_task(t);
+	}
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 }
 

@@ -1678,6 +1684,12 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 	char *ticks_title;
 	unsigned long ticks_value;
 
+	/*
+	 * We could be printing a lot of these messages while holding a
+	 * spinlock.  Avoid triggering hard lockup.
+	 */
+	touch_nmi_watchdog();
+
 	if (rsp->gpnum == rdp->gpnum) {
 		ticks_title = "ticks this GP";
 		ticks_value = rdp->ticks_this_gp;

Quoted 30 lines · Serge E. Hallyn

Quoted 28 lines · Paul E. McKenney

Quoted 21 lines · Tejun Heo

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Quoted 4 lines

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

If this fixes things for you, I would welcome such a patch.

Quoted 2 lines

Hi - would this also be relevant to 4.9-stable and 4.4-stable, or has something elsewhere changed after 4.9 that actually triggers this?

As far as I can tell, slow console lines have been prone to RCU CPU stall warnings for a very long time.

Thanx, Paul

Quoted 40 lines

thanks,
-serge

Quoted 36 lines

Thanx, Paul

Quoted 34 lines

kernel/rcu/tree_plugin.h +13 -1

index db85ca3..3c4c4d3 100644
@@ -561,8 +561,14 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 	}
 	t = list_entry(rnp->gp_tasks->prev,
 		       struct task_struct, rcu_node_entry);
-	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
+		touch_nmi_watchdog();
+		/*
+		 * We could be printing a lot of these messages while
+		 * holding a spinlock.  Avoid triggering hard lockup.
+		 */
 		sched_show_task(t);
+	}
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 }
 

@@ -1678,6 +1684,12 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 	char *ticks_title;
 	unsigned long ticks_value;
 
+	/*
+	 * We could be printing a lot of these messages while holding a
+	 * spinlock.  Avoid triggering hard lockup.
+	 */
+	touch_nmi_watchdog();
+
 	if (rsp->gpnum == rdp->gpnum) {
 		ticks_title = "ticks this GP";
 		ticks_value = rdp->ticks_this_gp;

Quoted 34 lines · Paul E. McKenney

Quoted 32 lines · Serge E. Hallyn

Quoted 22 lines · Paul E. McKenney

Quoted 21 lines · Tejun Heo

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Quoted 6 lines

Quoted 4 lines

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

If this fixes things for you, I would welcome such a patch.

Hi - would this also be relevant to 4.9-stable and 4.4-stable, or has something elsewhere changed after 4.9 that actually triggers this?

Quoted 2 lines

As far as I can tell, slow console lines have been prone to RCU CPU stall warnings for a very long time.

Ok, thanks Paul.

Tejun were you going to push this?

Quoted 40 lines

Quoted 40 lines

thanks,
-serge

Quoted 36 lines

Thanx, Paul

Quoted 34 lines

kernel/rcu/tree_plugin.h +13 -1

index db85ca3..3c4c4d3 100644
@@ -561,8 +561,14 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 	}
 	t = list_entry(rnp->gp_tasks->prev,
 		       struct task_struct, rcu_node_entry);
-	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
+		touch_nmi_watchdog();
+		/*
+		 * We could be printing a lot of these messages while
+		 * holding a spinlock.  Avoid triggering hard lockup.
+		 */
 		sched_show_task(t);
+	}
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 }
 

@@ -1678,6 +1684,12 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 	char *ticks_title;
 	unsigned long ticks_value;
 
+	/*
+	 * We could be printing a lot of these messages while holding a
+	 * spinlock.  Avoid triggering hard lockup.
+	 */
+	touch_nmi_watchdog();
+
 	if (rsp->gpnum == rdp->gpnum) {
 		ticks_title = "ticks this GP";
 		ticks_value = rdp->ticks_this_gp;

Quoted 38 lines · Serge E. Hallyn

Quoted 36 lines · Paul E. McKenney

Quoted 28 lines · Serge E. Hallyn

Quoted 27 lines · Paul E. McKenney

Quoted 25 lines · Tejun Heo

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

Quoted 4 lines

Quoted 1 line

If this fixes things for you, I would welcome such a patch.

Hi - would this also be relevant to 4.9-stable and 4.4-stable, or has something elsewhere changed after 4.9 that actually triggers this?

As far as I can tell, slow console lines have been prone to RCU CPU stall warnings for a very long time.

Quoted 3 lines

Ok, thanks Paul.

Tejun were you going to push this?

I have it queued for the next merge window. 3eea9623926f ("rcu: Call touch_nmi_watchdog() while printing stall warnings") in -rcu.

Quoted 39 lines · Paul E. McKenney

Quoted 38 lines · Serge E. Hallyn

Quoted 36 lines · Paul E. McKenney

Quoted 28 lines · Serge E. Hallyn

Quoted 27 lines · Paul E. McKenney

Quoted 25 lines · Tejun Heo

Hello, Paul.

Quoted 10 lines · Paul E. McKenney

Quoted 7 lines

I don't know the RCU code at all but it *looks* like the first CPU is taking a sweet while flushing printk buffer while holding a lock (the console is IPMI serial console, which faithfully emulates 115200 baud rate), and everyone else seems stuck waiting for that spinlock in rcu_check_callbacks().

Does this sound possible?

115200 baud? Ouch!!! That -will- result in trouble from console printing, and often also in RCU CPU stall warnings.

It could even be slower than 115200, and we occassionally see RCU stall warnings caused by printk storms, for example, while the kernel is trying to dump a lot of info after an OOM. That's an issue we probably want to improve from printk side; however, they don't usually lead to NMI hard lockup detector kicking in and crashing the machine, which is the peculiarity here.

Hmmm... show_state_filter(), the function which dumps all task backtraces, share a similar problem and it avoids it by explicitly calling touch_nmi_watchdog(). Maybe we can do something like the following from RCU too?

Quoted 4 lines

Quoted 1 line

If this fixes things for you, I would welcome such a patch.

Hi - would this also be relevant to 4.9-stable and 4.4-stable, or has something elsewhere changed after 4.9 that actually triggers this?

As far as I can tell, slow console lines have been prone to RCU CPU stall warnings for a very long time.

Quoted 6 lines

Quoted 3 lines

Ok, thanks Paul.

Tejun were you going to push this?

I have it queued for the next merge window. 3eea9623926f ("rcu: Call touch_nmi_watchdog() while printing stall warnings") in -rcu.

D'oh - thanks!