[PATCH] detect soft lockups
authorIngo Molnar <mingo@elte.hu>
Tue, 6 Sep 2005 22:16:27 +0000 (15:16 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 7 Sep 2005 23:57:17 +0000 (16:57 -0700)
This patch adds a new kernel debug feature: CONFIG_DETECT_SOFTLOCKUP.

When enabled then per-CPU watchdog threads are started, which try to run
once per second.  If they get delayed for more than 10 seconds then a
callback from the timer interrupt detects this condition and prints out a
warning message and a stack dump (once per lockup incident).  The feature
is otherwise non-intrusive, it doesnt try to unlock the box in any way, it
only gets the debug info out, automatically, and on all CPUs affected by
the lockup.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-Off-By: Matthias Urlichs <smurf@smurf.noris.de>
Signed-off-by: Richard Purdie <rpurdie@rpsys.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
12 files changed:
arch/i386/kernel/nmi.c
arch/i386/kernel/time.c
arch/x86_64/kernel/nmi.c
arch/x86_64/kernel/time.c
drivers/mtd/nand/nand_base.c
include/linux/sched.h
init/main.c
kernel/Makefile
kernel/power/swsusp.c
kernel/softlockup.c [new file with mode: 0644]
kernel/timer.c
lib/Kconfig.debug

index 8bbdbda07a2d6c0dd54d68ec06ea968b52d064fb..0178457db72130b54d3170dfd8150742b8addfe1 100644 (file)
@@ -478,6 +478,11 @@ void touch_nmi_watchdog (void)
         */
        for (i = 0; i < NR_CPUS; i++)
                alert_counter[i] = 0;
+
+       /*
+        * Tickle the softlockup detector too:
+        */
+       touch_softlockup_watchdog();
 }
 
 extern void die_nmi(struct pt_regs *, const char *msg);
index 6f794a78ee1e7be1f12074b7c0cfaad15aa8a9fd..b0c5ee2b3446d4f96520b4fd04458ec9a6d47cd1 100644 (file)
@@ -422,6 +422,7 @@ static int timer_resume(struct sys_device *dev)
                last_timer->resume();
        cur_timer = last_timer;
        last_timer = NULL;
+       touch_softlockup_watchdog();
        return 0;
 }
 
index 64a8e05d5811b82fd258365d539ea37c72ed0b2c..84cae81fff8b92277adccd928544cb0c19be8f92 100644 (file)
@@ -463,6 +463,8 @@ void touch_nmi_watchdog (void)
         */
        for (i = 0; i < NR_CPUS; i++)
                per_cpu(nmi_touch, i) = 1;
+
+       touch_softlockup_watchdog();
 }
 
 void nmi_watchdog_tick (struct pt_regs * regs, unsigned reason)
index 66bf6ddeb0c334a956c6bc2e75beeca9262b5f9e..2b5d9da912a26d2d7119563217d2b19dcf2c0e24 100644 (file)
@@ -1041,6 +1041,7 @@ static int timer_resume(struct sys_device *dev)
        write_sequnlock_irqrestore(&xtime_lock,flags);
        jiffies += sleep_length;
        wall_jiffies += sleep_length;
+       touch_softlockup_watchdog();
        return 0;
 }
 
index eee5115658c8c37bc3bf91136f48494e4221cb16..04e54318bc6a603f2c44b32e3b2f8b5b260a64dc 100644 (file)
@@ -526,6 +526,7 @@ static void nand_wait_ready(struct mtd_info *mtd)
        do {
                if (this->dev_ready(mtd))
                        return;
+               touch_softlockup_watchdog();
        } while (time_before(jiffies, timeo));  
 }
 
index dec5827c77422ebbd8e7f0238cdf7f34e08cb906..5fb31bede1036c378e759391573631e06c56e41c 100644 (file)
@@ -176,6 +176,23 @@ extern void trap_init(void);
 extern void update_process_times(int user);
 extern void scheduler_tick(void);
 
+#ifdef CONFIG_DETECT_SOFTLOCKUP
+extern void softlockup_tick(struct pt_regs *regs);
+extern void spawn_softlockup_task(void);
+extern void touch_softlockup_watchdog(void);
+#else
+static inline void softlockup_tick(struct pt_regs *regs)
+{
+}
+static inline void spawn_softlockup_task(void)
+{
+}
+static inline void touch_softlockup_watchdog(void)
+{
+}
+#endif
+
+
 /* Attach to any functions which should be ignored in wchan output. */
 #define __sched                __attribute__((__section__(".sched.text")))
 /* Is this address in the __sched functions? */
index ff410063e4e13ca6a296a680a136a15297d079de..a29fb2ac724076433bf729b1f8c77a23aa673206 100644 (file)
@@ -614,6 +614,7 @@ static void do_pre_smp_initcalls(void)
        migration_init();
 #endif
        spawn_ksoftirqd();
+       spawn_softlockup_task();
 }
 
 static void run_init_process(char *init_filename)
index cb05cd05d2374b1ae316adb437ec14da370bd8a5..8d57a2f1226baa8034e8d93b7e4806d5a3f5fce1 100644 (file)
@@ -27,6 +27,7 @@ obj-$(CONFIG_AUDIT) += audit.o
 obj-$(CONFIG_AUDITSYSCALL) += auditsc.o
 obj-$(CONFIG_KPROBES) += kprobes.o
 obj-$(CONFIG_SYSFS) += ksysfs.o
+obj-$(CONFIG_DETECT_SOFTLOCKUP) += softlockup.o
 obj-$(CONFIG_GENERIC_HARDIRQS) += irq/
 obj-$(CONFIG_CRASH_DUMP) += crash_dump.o
 obj-$(CONFIG_SECCOMP) += seccomp.o
index eaacd5cb58895438ac28013968651f0d1752c4c6..d967e875ee82f974d9f4e8e365aa59270a6f5c61 100644 (file)
@@ -1059,6 +1059,7 @@ int swsusp_resume(void)
        BUG_ON(!error);
        restore_processor_state();
        restore_highmem();
+       touch_softlockup_watchdog();
        device_power_up();
        local_irq_enable();
        return error;
diff --git a/kernel/softlockup.c b/kernel/softlockup.c
new file mode 100644 (file)
index 0000000..7597620
--- /dev/null
@@ -0,0 +1,151 @@
+/*
+ * Detect Soft Lockups
+ *
+ * started by Ingo Molnar, (C) 2005, Red Hat
+ *
+ * this code detects soft lockups: incidents in where on a CPU
+ * the kernel does not reschedule for 10 seconds or more.
+ */
+
+#include <linux/mm.h>
+#include <linux/cpu.h>
+#include <linux/init.h>
+#include <linux/delay.h>
+#include <linux/kthread.h>
+#include <linux/notifier.h>
+#include <linux/module.h>
+
+static DEFINE_SPINLOCK(print_lock);
+
+static DEFINE_PER_CPU(unsigned long, timestamp) = 0;
+static DEFINE_PER_CPU(unsigned long, print_timestamp) = 0;
+static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
+
+static int did_panic = 0;
+static int softlock_panic(struct notifier_block *this, unsigned long event,
+                               void *ptr)
+{
+       did_panic = 1;
+
+       return NOTIFY_DONE;
+}
+
+static struct notifier_block panic_block = {
+       .notifier_call = softlock_panic,
+};
+
+void touch_softlockup_watchdog(void)
+{
+       per_cpu(timestamp, raw_smp_processor_id()) = jiffies;
+}
+EXPORT_SYMBOL(touch_softlockup_watchdog);
+
+/*
+ * This callback runs from the timer interrupt, and checks
+ * whether the watchdog thread has hung or not:
+ */
+void softlockup_tick(struct pt_regs *regs)
+{
+       int this_cpu = smp_processor_id();
+       unsigned long timestamp = per_cpu(timestamp, this_cpu);
+
+       if (per_cpu(print_timestamp, this_cpu) == timestamp)
+               return;
+
+       /* Do not cause a second panic when there already was one */
+       if (did_panic)
+               return;
+
+       if (time_after(jiffies, timestamp + 10*HZ)) {
+               per_cpu(print_timestamp, this_cpu) = timestamp;
+
+               spin_lock(&print_lock);
+               printk(KERN_ERR "BUG: soft lockup detected on CPU#%d!\n",
+                       this_cpu);
+               show_regs(regs);
+               spin_unlock(&print_lock);
+       }
+}
+
+/*
+ * The watchdog thread - runs every second and touches the timestamp.
+ */
+static int watchdog(void * __bind_cpu)
+{
+       struct sched_param param = { .sched_priority = 99 };
+       int this_cpu = (long) __bind_cpu;
+
+       printk("softlockup thread %d started up.\n", this_cpu);
+
+       sched_setscheduler(current, SCHED_FIFO, &param);
+       current->flags |= PF_NOFREEZE;
+
+       set_current_state(TASK_INTERRUPTIBLE);
+
+       /*
+        * Run briefly once per second - if this gets delayed for
+        * more than 10 seconds then the debug-printout triggers
+        * in softlockup_tick():
+        */
+       while (!kthread_should_stop()) {
+               msleep_interruptible(1000);
+               touch_softlockup_watchdog();
+       }
+       __set_current_state(TASK_RUNNING);
+
+       return 0;
+}
+
+/*
+ * Create/destroy watchdog threads as CPUs come and go:
+ */
+static int __devinit
+cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
+{
+       int hotcpu = (unsigned long)hcpu;
+       struct task_struct *p;
+
+       switch (action) {
+       case CPU_UP_PREPARE:
+               BUG_ON(per_cpu(watchdog_task, hotcpu));
+               p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
+               if (IS_ERR(p)) {
+                       printk("watchdog for %i failed\n", hotcpu);
+                       return NOTIFY_BAD;
+               }
+               per_cpu(watchdog_task, hotcpu) = p;
+               kthread_bind(p, hotcpu);
+               break;
+       case CPU_ONLINE:
+
+               wake_up_process(per_cpu(watchdog_task, hotcpu));
+               break;
+#ifdef CONFIG_HOTPLUG_CPU
+       case CPU_UP_CANCELED:
+               /* Unbind so it can run.  Fall thru. */
+               kthread_bind(per_cpu(watchdog_task, hotcpu), smp_processor_id());
+       case CPU_DEAD:
+               p = per_cpu(watchdog_task, hotcpu);
+               per_cpu(watchdog_task, hotcpu) = NULL;
+               kthread_stop(p);
+               break;
+#endif /* CONFIG_HOTPLUG_CPU */
+       }
+       return NOTIFY_OK;
+}
+
+static struct notifier_block __devinitdata cpu_nfb = {
+       .notifier_call = cpu_callback
+};
+
+__init void spawn_softlockup_task(void)
+{
+       void *cpu = (void *)(long)smp_processor_id();
+
+       cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
+       cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
+       register_cpu_notifier(&cpu_nfb);
+
+       notifier_chain_register(&panic_notifier_list, &panic_block);
+}
+
index 5377f40723ff0dc37f13897847ea040f8c317292..1433d87f46b3e520b7211f6a4f24efdd493cdd80 100644 (file)
@@ -950,6 +950,7 @@ void do_timer(struct pt_regs *regs)
 {
        jiffies_64++;
        update_times();
+       softlockup_tick(regs);
 }
 
 #ifdef __ARCH_WANT_SYS_ALARM
index 299f7f3b5b0871fdd332a647f83ea54d2ddf059b..3754c9a8f5c8e1e3fccebf2d0f282fd5eda6bcb4 100644 (file)
@@ -46,6 +46,25 @@ config LOG_BUF_SHIFT
                     13 =>  8 KB
                     12 =>  4 KB
 
+config DETECT_SOFTLOCKUP
+       bool "Detect Soft Lockups"
+       depends on DEBUG_KERNEL
+       default y
+       help
+         Say Y here to enable the kernel to detect "soft lockups",
+         which are bugs that cause the kernel to loop in kernel
+         mode for more than 10 seconds, without giving other tasks a
+         chance to run.
+
+         When a soft-lockup is detected, the kernel will print the
+         current stack trace (which you should report), but the
+         system will stay locked up. This feature has negligible
+         overhead.
+
+         (Note that "hard lockups" are separate type of bugs that
+          can be detected via the NMI-watchdog, on platforms that
+          support it.)
+
 config SCHEDSTATS
        bool "Collect scheduler statistics"
        depends on DEBUG_KERNEL && PROC_FS