[PATCH] x86_64: Collected NMI watchdog fixes.

Collected NMI watchdog fixes.

- Fix call of check_nmi_watchdog

- Remove earlier move of check_nmi_watchdog to later.  It does not fix the
  race it was supposed to fix fully.

- Remove unused P6 definitions

- Add support for performance counter based watchdog on P4 systems.

  This allows to run it only once per second, which saves some CPU time.
  Previously it would run at 1000Hz, which was too much.

  Code ported from i386

  Make this the default on Intel systems.

- Use check_nmi_watchdog with local APIC based nmi

- Fix race in touch_nmi_watchdog

- Fix bug that caused incorrect performance counters to be programmed in a
  few cases on K8.

- Remove useless check for local APIC

- Use local_t and per_cpu variables for per CPU data.

- Keep other CPUs busy during check_nmi_watchdog to make sure they really
  tick when in lapic mode.

- Only check CPUs that are actually online.

- Various other fixes.

- Fix fallback path when MSRs are unimplemented

Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

1 files changed, 172 insertions, 78 deletions

diff --git a/arch/x86_64/kernel/nmi.c b/arch/x86_64/kernel/nmi.c
index ec13eb97e8e6..31c0f2e6ac91 100644
--- a/arch/x86_64/kernel/nmi.c
+++ b/arch/x86_64/kernel/nmi.c
@@ -33,6 +33,7 @@
 #include <asm/msr.h>
 #include <asm/proto.h>
 #include <asm/kdebug.h>
+#include <asm/local.h>
 
 /*
  * lapic_nmi_owner tracks the ownership of the lapic NMI hardware:
@@ -59,7 +60,8 @@ int panic_on_timeout;
 
 unsigned int nmi_watchdog = NMI_DEFAULT;
 static unsigned int nmi_hz = HZ;
-unsigned int nmi_perfctr_msr;	/* the MSR to reset in NMI handler */
+static unsigned int nmi_perfctr_msr;	/* the MSR to reset in NMI handler */
+static unsigned int nmi_p4_cccr_val;
 
 /* Note that these events don't tick when the CPU idles. This means
    the frequency varies with CPU load. */
@@ -71,67 +73,87 @@ unsigned int nmi_perfctr_msr;	/* the MSR to reset in NMI handler */
 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING	0x76
 #define K7_NMI_EVENT		K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
 
-#define P6_EVNTSEL0_ENABLE	(1 << 22)
-#define P6_EVNTSEL_INT		(1 << 20)
-#define P6_EVNTSEL_OS		(1 << 17)
-#define P6_EVNTSEL_USR		(1 << 16)
-#define P6_EVENT_CPU_CLOCKS_NOT_HALTED	0x79
-#define P6_NMI_EVENT		P6_EVENT_CPU_CLOCKS_NOT_HALTED
+#define MSR_P4_MISC_ENABLE	0x1A0
+#define MSR_P4_MISC_ENABLE_PERF_AVAIL	(1<<7)
+#define MSR_P4_MISC_ENABLE_PEBS_UNAVAIL	(1<<12)
+#define MSR_P4_PERFCTR0		0x300
+#define MSR_P4_CCCR0		0x360
+#define P4_ESCR_EVENT_SELECT(N)	((N)<<25)
+#define P4_ESCR_OS		(1<<3)
+#define P4_ESCR_USR		(1<<2)
+#define P4_CCCR_OVF_PMI0	(1<<26)
+#define P4_CCCR_OVF_PMI1	(1<<27)
+#define P4_CCCR_THRESHOLD(N)	((N)<<20)
+#define P4_CCCR_COMPLEMENT	(1<<19)
+#define P4_CCCR_COMPARE		(1<<18)
+#define P4_CCCR_REQUIRED	(3<<16)
+#define P4_CCCR_ESCR_SELECT(N)	((N)<<13)
+#define P4_CCCR_ENABLE		(1<<12)
+/* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
+   CRU_ESCR0 (with any non-null event selector) through a complemented
+   max threshold. [IA32-Vol3, Section 14.9.9] */
+#define MSR_P4_IQ_COUNTER0	0x30C
+#define P4_NMI_CRU_ESCR0	(P4_ESCR_EVENT_SELECT(0x3F)|P4_ESCR_OS|P4_ESCR_USR)
+#define P4_NMI_IQ_CCCR0	\
+	(P4_CCCR_OVF_PMI0|P4_CCCR_THRESHOLD(15)|P4_CCCR_COMPLEMENT|	\
+	 P4_CCCR_COMPARE|P4_CCCR_REQUIRED|P4_CCCR_ESCR_SELECT(4)|P4_CCCR_ENABLE)
+
+static __init inline int nmi_known_cpu(void)
+{
+	switch (boot_cpu_data.x86_vendor) {
+	case X86_VENDOR_AMD:
+		return boot_cpu_data.x86 == 15;
+	case X86_VENDOR_INTEL:
+		return boot_cpu_data.x86 == 15;
+	}
+	return 0;
+}
 
 /* Run after command line and cpu_init init, but before all other checks */
 void __init nmi_watchdog_default(void)
 {
 	if (nmi_watchdog != NMI_DEFAULT)
 		return;
-
-	/* For some reason the IO APIC watchdog doesn't work on the AMD
-	   8111 chipset. For now switch to local APIC mode using
-	   perfctr0 there.  On Intel CPUs we don't have code to handle
-	   the perfctr and the IO-APIC seems to work, so use that.  */
-
-	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
-		nmi_watchdog = NMI_LOCAL_APIC; 
-		printk(KERN_INFO 
-              "Using local APIC NMI watchdog using perfctr0\n");
-	} else {
-		printk(KERN_INFO "Using IO APIC NMI watchdog\n");
+	if (nmi_known_cpu())
+		nmi_watchdog = NMI_LOCAL_APIC;
+	else
 		nmi_watchdog = NMI_IO_APIC;
-	}
 }
 
-/* Why is there no CPUID flag for this? */
-static __init int cpu_has_lapic(void)
+#ifdef CONFIG_SMP
+/* The performance counters used by NMI_LOCAL_APIC don't trigger when
+ * the CPU is idle. To make sure the NMI watchdog really ticks on all
+ * CPUs during the test make them busy.
+ */
+static __init void nmi_cpu_busy(void *data)
 {
-	switch (boot_cpu_data.x86_vendor) { 
-	case X86_VENDOR_INTEL:
-	case X86_VENDOR_AMD: 
-		return boot_cpu_data.x86 >= 6; 
-	/* .... add more cpus here or find a different way to figure this out. */	
-	default:
-		return 0;
-	} 	
+	volatile int *endflag = data;
+	local_irq_enable();
+	/* Intentionally don't use cpu_relax here. This is
+	   to make sure that the performance counter really ticks,
+	   even if there is a simulator or similar that catches the
+	   pause instruction. On a real HT machine this is fine because
+	   all other CPUs are busy with "useless" delay loops and don't
+	   care if they get somewhat less cycles. */
+	while (*endflag == 0)
+		barrier();
 }
+#endif
 
-static int __init check_nmi_watchdog (void)
+int __init check_nmi_watchdog (void)
 {
+	volatile int endflag = 0;
 	int *counts;
 	int cpu;
 
-	if (nmi_watchdog == NMI_NONE)
-		return 0;
+	counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
+	if (!counts)
+		return -1;
 
-	if (nmi_watchdog == NMI_LOCAL_APIC && !cpu_has_lapic())  {
-		nmi_watchdog = NMI_NONE;
-		return -1; 
-	}	
+	printk(KERN_INFO "testing NMI watchdog ... ");
 
-	counts = kmalloc(NR_CPUS * sizeof(int),GFP_KERNEL);
-	if (!counts) {
-		nmi_watchdog = NMI_NONE;
-		return 0;
-	}
-
-	printk(KERN_INFO "Testing NMI watchdog ... ");
+	if (nmi_watchdog == NMI_LOCAL_APIC)
+		smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 
 	for (cpu = 0; cpu < NR_CPUS; cpu++)
 		counts[cpu] = cpu_pda[cpu].__nmi_count; 
@@ -139,16 +161,22 @@ static int __init check_nmi_watchdog (void)
 	mdelay((10*1000)/nmi_hz); // wait 10 ticks
 
 	for (cpu = 0; cpu < NR_CPUS; cpu++) {
+		if (!cpu_online(cpu))
+			continue;
 		if (cpu_pda[cpu].__nmi_count - counts[cpu] <= 5) {
-			printk("CPU#%d: NMI appears to be stuck (%d)!\n", 
+			endflag = 1;
+			printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 			       cpu,
+			       counts[cpu],
 			       cpu_pda[cpu].__nmi_count);
 			nmi_active = 0;
 			lapic_nmi_owner &= ~LAPIC_NMI_WATCHDOG;
+			nmi_perfctr_msr = 0;
 			kfree(counts);
 			return -1;
 		}
 	}
+	endflag = 1;
 	printk("OK.\n");
 
 	/* now that we know it works we can reduce NMI frequency to
@@ -159,8 +187,6 @@ static int __init check_nmi_watchdog (void)
 	kfree(counts);
 	return 0;
 }
-/* Have this called later during boot so counters are updating */
-late_initcall(check_nmi_watchdog);
 
 int __init setup_nmi_watchdog(char *str)
 {
@@ -178,7 +204,7 @@ int __init setup_nmi_watchdog(char *str)
 
 	if (nmi >= NMI_INVALID)
 		return 0;
-		nmi_watchdog = nmi;
+	nmi_watchdog = nmi;
 	return 1;
 }
 
@@ -193,7 +219,10 @@ static void disable_lapic_nmi_watchdog(void)
 		wrmsr(MSR_K7_EVNTSEL0, 0, 0);
 		break;
 	case X86_VENDOR_INTEL:
-		wrmsr(MSR_IA32_EVNTSEL0, 0, 0);
+		if (boot_cpu_data.x86 == 15) {
+			wrmsr(MSR_P4_IQ_CCCR0, 0, 0);
+			wrmsr(MSR_P4_CRU_ESCR0, 0, 0);
+		}
 		break;
 	}
 	nmi_active = -1;
@@ -261,7 +290,7 @@ void enable_timer_nmi_watchdog(void)
 
 static int nmi_pm_active; /* nmi_active before suspend */
 
-static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
+static int lapic_nmi_suspend(struct sys_device *dev, u32 state)
 {
 	nmi_pm_active = nmi_active;
 	disable_lapic_nmi_watchdog();
@@ -308,22 +337,27 @@ late_initcall(init_lapic_nmi_sysfs);
  * Original code written by Keith Owens.
  */
 
+static void clear_msr_range(unsigned int base, unsigned int n)
+{
+	unsigned int i;
+
+	for(i = 0; i < n; ++i)
+		wrmsr(base+i, 0, 0);
+}
+
 static void setup_k7_watchdog(void)
 {
 	int i;
 	unsigned int evntsel;
 
-	/* No check, so can start with slow frequency */
-	nmi_hz = 1; 
-
-	/* XXX should check these in EFER */
-
 	nmi_perfctr_msr = MSR_K7_PERFCTR0;
 
 	for(i = 0; i < 4; ++i) {
 		/* Simulator may not support it */
-		if (checking_wrmsrl(MSR_K7_EVNTSEL0+i, 0UL))
+		if (checking_wrmsrl(MSR_K7_EVNTSEL0+i, 0UL)) {
+			nmi_perfctr_msr = 0;
 			return;
+		}
 		wrmsrl(MSR_K7_PERFCTR0+i, 0UL);
 	}
 
@@ -333,12 +367,54 @@ static void setup_k7_watchdog(void)
 		| K7_NMI_EVENT;
 
 	wrmsr(MSR_K7_EVNTSEL0, evntsel, 0);
-	wrmsrl(MSR_K7_PERFCTR0, -((u64)cpu_khz*1000) / nmi_hz);
+	wrmsr(MSR_K7_PERFCTR0, -(cpu_khz/nmi_hz*1000), -1);
 	apic_write(APIC_LVTPC, APIC_DM_NMI);
 	evntsel |= K7_EVNTSEL_ENABLE;
 	wrmsr(MSR_K7_EVNTSEL0, evntsel, 0);
 }
 
+
+static int setup_p4_watchdog(void)
+{
+	unsigned int misc_enable, dummy;
+
+	rdmsr(MSR_P4_MISC_ENABLE, misc_enable, dummy);
+	if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
+		return 0;
+
+	nmi_perfctr_msr = MSR_P4_IQ_COUNTER0;
+	nmi_p4_cccr_val = P4_NMI_IQ_CCCR0;
+#ifdef CONFIG_SMP
+	if (smp_num_siblings == 2)
+		nmi_p4_cccr_val |= P4_CCCR_OVF_PMI1;
+#endif
+
+	if (!(misc_enable & MSR_P4_MISC_ENABLE_PEBS_UNAVAIL))
+		clear_msr_range(0x3F1, 2);
+	/* MSR 0x3F0 seems to have a default value of 0xFC00, but current
+	   docs doesn't fully define it, so leave it alone for now. */
+	if (boot_cpu_data.x86_model >= 0x3) {
+		/* MSR_P4_IQ_ESCR0/1 (0x3ba/0x3bb) removed */
+		clear_msr_range(0x3A0, 26);
+		clear_msr_range(0x3BC, 3);
+	} else {
+		clear_msr_range(0x3A0, 31);
+	}
+	clear_msr_range(0x3C0, 6);
+	clear_msr_range(0x3C8, 6);
+	clear_msr_range(0x3E0, 2);
+	clear_msr_range(MSR_P4_CCCR0, 18);
+	clear_msr_range(MSR_P4_PERFCTR0, 18);
+
+	wrmsr(MSR_P4_CRU_ESCR0, P4_NMI_CRU_ESCR0, 0);
+	wrmsr(MSR_P4_IQ_CCCR0, P4_NMI_IQ_CCCR0 & ~P4_CCCR_ENABLE, 0);
+	Dprintk("setting P4_IQ_COUNTER0 to 0x%08lx\n", -(cpu_khz/nmi_hz*1000));
+	wrmsr(MSR_P4_IQ_COUNTER0, -(cpu_khz/nmi_hz*1000), -1);
+	apic_write(APIC_LVTPC, APIC_DM_NMI);
+	wrmsr(MSR_P4_IQ_CCCR0, nmi_p4_cccr_val, 0);
+	return 1;
+}
+
 void setup_apic_nmi_watchdog(void)
 {
 	switch (boot_cpu_data.x86_vendor) {
@@ -349,6 +425,13 @@ void setup_apic_nmi_watchdog(void)
 			return;
 		setup_k7_watchdog();
 		break;
+	case X86_VENDOR_INTEL:
+		if (boot_cpu_data.x86 != 15)
+			return;
+		if (!setup_p4_watchdog())
+			return;
+		break;
+
 	default:
 		return;
 	}
@@ -363,56 +446,67 @@ void setup_apic_nmi_watchdog(void)
  *
  * as these watchdog NMI IRQs are generated on every CPU, we only
  * have to check the current processor.
- *
- * since NMIs don't listen to _any_ locks, we have to be extremely
- * careful not to rely on unsafe variables. The printk might lock
- * up though, so we have to break up any console locks first ...
- * [when there will be more tty-related locks, break them up
- *  here too!]
  */
 
-static unsigned int
-	last_irq_sums [NR_CPUS],
-	alert_counter [NR_CPUS];
+static DEFINE_PER_CPU(unsigned, last_irq_sum);
+static DEFINE_PER_CPU(local_t, alert_counter);
+static DEFINE_PER_CPU(int, nmi_touch);
 
 void touch_nmi_watchdog (void)
 {
 	int i;
 
 	/*
-	 * Just reset the alert counters, (other CPUs might be
-	 * spinning on locks we hold):
+ 	 * Tell other CPUs to reset their alert counters. We cannot
+	 * do it ourselves because the alert count increase is not
+	 * atomic.
 	 */
 	for (i = 0; i < NR_CPUS; i++)
-		alert_counter[i] = 0;
+		per_cpu(nmi_touch, i) = 1;
 }
 
 void nmi_watchdog_tick (struct pt_regs * regs, unsigned reason)
 {
-	int sum, cpu;
+	int sum;
+	int touched = 0;
 
-	cpu = safe_smp_processor_id();
 	sum = read_pda(apic_timer_irqs);
-	if (last_irq_sums[cpu] == sum) {
+	if (__get_cpu_var(nmi_touch)) {
+		__get_cpu_var(nmi_touch) = 0;
+		touched = 1;
+	}
+	if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 		/*
 		 * Ayiee, looks like this CPU is stuck ...
 		 * wait a few IRQs (5 seconds) before doing the oops ...
 		 */
-		alert_counter[cpu]++;
-		if (alert_counter[cpu] == 5*nmi_hz) {
+		local_inc(&__get_cpu_var(alert_counter));
+		if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz) {
 			if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 							== NOTIFY_STOP) {
-				alert_counter[cpu] = 0; 
+				local_set(&__get_cpu_var(alert_counter), 0);
 				return;
 			} 
 			die_nmi("NMI Watchdog detected LOCKUP on CPU%d", regs);
 		}
 	} else {
-		last_irq_sums[cpu] = sum;
-		alert_counter[cpu] = 0;
+		__get_cpu_var(last_irq_sum) = sum;
+		local_set(&__get_cpu_var(alert_counter), 0);
 	}
-	if (nmi_perfctr_msr)
+	if (nmi_perfctr_msr) {
+ 		if (nmi_perfctr_msr == MSR_P4_IQ_COUNTER0) {
+ 			/*
+ 			 * P4 quirks:
+ 			 * - An overflown perfctr will assert its interrupt
+ 			 *   until the OVF flag in its CCCR is cleared.
+ 			 * - LVTPC is masked on interrupt and must be
+ 			 *   unmasked by the LVTPC handler.
+ 			 */
+ 			wrmsr(MSR_P4_IQ_CCCR0, nmi_p4_cccr_val, 0);
+ 			apic_write(APIC_LVTPC, APIC_DM_NMI);
+ 		}
 		wrmsr(nmi_perfctr_msr, -(cpu_khz/nmi_hz*1000), -1);
+	}
 }
 
 static int dummy_nmi_callback(struct pt_regs * regs, int cpu)