x86/mce: Add CMCI poll mode

On Intel systems corrected machine check interrupts (CMCI) may be sent to
multiple logical processors; possibly to all processors on the affected
socket (SDM Volume 3B "15.5.1 CMCI Local APIC Interface").  This means
that a persistent error (such as a stuck bit in ECC memory) may cause
a storm of interrupts that greatly hinders or prevents forward progress
(probably on many processors).

To solve this we keep track of the rate at which each processor sees
CMCI. If we exceed a threshold, we disable CMCI delivery and switch to
polling the machine check banks. If the storm subsides (none of the
affected processors see any more errors for a complete poll interval) we
re-enable CMCI.

[Tony: Added console messages when storm begins/ends and increased storm
threshold from 5 to 15 so we have a few more logged entries before we
disable interrupts and start dropping reports]

Signed-off-by: Chen Gong <gong.chen@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Chen Gong <gong.chen@linux.intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>

3 files changed, 160 insertions, 7 deletions

diff --git a/arch/x86/kernel/cpu/mcheck/mce-internal.h b/arch/x86/kernel/cpu/mcheck/mce-internal.h
index ed44c8a6585..6a05c1d327a 100644
--- a/arch/x86/kernel/cpu/mcheck/mce-internal.h
+++ b/arch/x86/kernel/cpu/mcheck/mce-internal.h
@@ -28,6 +28,18 @@ extern int mce_ser;
 
 extern struct mce_bank *mce_banks;
 
+#ifdef CONFIG_X86_MCE_INTEL
+unsigned long mce_intel_adjust_timer(unsigned long interval);
+void mce_intel_cmci_poll(void);
+void mce_intel_hcpu_update(unsigned long cpu);
+#else
+# define mce_intel_adjust_timer mce_adjust_timer_default
+static inline void mce_intel_cmci_poll(void) { }
+static inline void mce_intel_hcpu_update(unsigned long cpu) { }
+#endif
+
+void mce_timer_kick(unsigned long interval);
+
 #ifdef CONFIG_ACPI_APEI
 int apei_write_mce(struct mce *m);
 ssize_t apei_read_mce(struct mce *m, u64 *record_id);
diff --git a/arch/x86/kernel/cpu/mcheck/mce.c b/arch/x86/kernel/cpu/mcheck/mce.c
index b4dde1527ed..8c1beea6cab 100644
--- a/arch/x86/kernel/cpu/mcheck/mce.c
+++ b/arch/x86/kernel/cpu/mcheck/mce.c
@@ -1260,6 +1260,14 @@ static unsigned long check_interval = 5 * 60; /* 5 minutes */
 static DEFINE_PER_CPU(unsigned long, mce_next_interval); /* in jiffies */
 static DEFINE_PER_CPU(struct timer_list, mce_timer);
 
+static unsigned long mce_adjust_timer_default(unsigned long interval)
+{
+	return interval;
+}
+
+static unsigned long (*mce_adjust_timer)(unsigned long interval) =
+	mce_adjust_timer_default;
+
 static void mce_timer_fn(unsigned long data)
 {
 	struct timer_list *t = &__get_cpu_var(mce_timer);
@@ -1270,6 +1278,7 @@ static void mce_timer_fn(unsigned long data)
 	if (mce_available(__this_cpu_ptr(&cpu_info))) {
 		machine_check_poll(MCP_TIMESTAMP,
 				&__get_cpu_var(mce_poll_banks));
+		mce_intel_cmci_poll();
 	}
 
 	/*
@@ -1277,14 +1286,38 @@ static void mce_timer_fn(unsigned long data)
 	 * polling interval, otherwise increase the polling interval.
 	 */
 	iv = __this_cpu_read(mce_next_interval);
-	if (mce_notify_irq())
+	if (mce_notify_irq()) {
 		iv = max(iv / 2, (unsigned long) HZ/100);
-	else
+	} else {
 		iv = min(iv * 2, round_jiffies_relative(check_interval * HZ));
+		iv = mce_adjust_timer(iv);
+	}
 	__this_cpu_write(mce_next_interval, iv);
+	/* Might have become 0 after CMCI storm subsided */
+	if (iv) {
+		t->expires = jiffies + iv;
+		add_timer_on(t, smp_processor_id());
+	}
+}
 
-	t->expires = jiffies + iv;
-	add_timer_on(t, smp_processor_id());
+/*
+ * Ensure that the timer is firing in @interval from now.
+ */
+void mce_timer_kick(unsigned long interval)
+{
+	struct timer_list *t = &__get_cpu_var(mce_timer);
+	unsigned long when = jiffies + interval;
+	unsigned long iv = __this_cpu_read(mce_next_interval);
+
+	if (timer_pending(t)) {
+		if (time_before(when, t->expires))
+			mod_timer_pinned(t, when);
+	} else {
+		t->expires = round_jiffies(when);
+		add_timer_on(t, smp_processor_id());
+	}
+	if (interval < iv)
+		__this_cpu_write(mce_next_interval, interval);
 }
 
 /* Must not be called in IRQ context where del_timer_sync() can deadlock */
@@ -1548,6 +1581,7 @@ static void __mcheck_cpu_init_vendor(struct cpuinfo_x86 *c)
 	switch (c->x86_vendor) {
 	case X86_VENDOR_INTEL:
 		mce_intel_feature_init(c);
+		mce_adjust_timer = mce_intel_adjust_timer;
 		break;
 	case X86_VENDOR_AMD:
 		mce_amd_feature_init(c);
@@ -1559,7 +1593,7 @@ static void __mcheck_cpu_init_vendor(struct cpuinfo_x86 *c)
 
 static void mce_start_timer(unsigned int cpu, struct timer_list *t)
 {
-	unsigned long iv = check_interval * HZ;
+	unsigned long iv = mce_adjust_timer(check_interval * HZ);
 
 	__this_cpu_write(mce_next_interval, iv);
 
@@ -2272,10 +2306,11 @@ mce_cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 		if (threshold_cpu_callback)
 			threshold_cpu_callback(action, cpu);
 		mce_device_remove(cpu);
+		mce_intel_hcpu_update(cpu);
 		break;
 	case CPU_DOWN_PREPARE:
-		del_timer_sync(t);
 		smp_call_function_single(cpu, mce_disable_cpu, &action, 1);
+		del_timer_sync(t);
 		break;
 	case CPU_DOWN_FAILED:
 		smp_call_function_single(cpu, mce_reenable_cpu, &action, 1);
diff --git a/arch/x86/kernel/cpu/mcheck/mce_intel.c b/arch/x86/kernel/cpu/mcheck/mce_intel.c
index 59648e48a14..098386fed48 100644
--- a/arch/x86/kernel/cpu/mcheck/mce_intel.c
+++ b/arch/x86/kernel/cpu/mcheck/mce_intel.c
@@ -15,6 +15,8 @@
 #include <asm/msr.h>
 #include <asm/mce.h>
 
+#include "mce-internal.h"
+
 /*
  * Support for Intel Correct Machine Check Interrupts. This allows
  * the CPU to raise an interrupt when a corrected machine check happened.
@@ -30,7 +32,22 @@ static DEFINE_PER_CPU(mce_banks_t, mce_banks_owned);
  */
 static DEFINE_RAW_SPINLOCK(cmci_discover_lock);
 
-#define CMCI_THRESHOLD 1
+#define CMCI_THRESHOLD		1
+#define CMCI_POLL_INTERVAL	(30 * HZ)
+#define CMCI_STORM_INTERVAL	(1 * HZ)
+#define CMCI_STORM_THRESHOLD	15
+
+static DEFINE_PER_CPU(unsigned long, cmci_time_stamp);
+static DEFINE_PER_CPU(unsigned int, cmci_storm_cnt);
+static DEFINE_PER_CPU(unsigned int, cmci_storm_state);
+
+enum {
+	CMCI_STORM_NONE,
+	CMCI_STORM_ACTIVE,
+	CMCI_STORM_SUBSIDED,
+};
+
+static atomic_t cmci_storm_on_cpus;
 
 static int cmci_supported(int *banks)
 {
@@ -53,6 +70,93 @@ static int cmci_supported(int *banks)
 	return !!(cap & MCG_CMCI_P);
 }
 
+void mce_intel_cmci_poll(void)
+{
+	if (__this_cpu_read(cmci_storm_state) == CMCI_STORM_NONE)
+		return;
+	machine_check_poll(MCP_TIMESTAMP, &__get_cpu_var(mce_banks_owned));
+}
+
+void mce_intel_hcpu_update(unsigned long cpu)
+{
+	if (per_cpu(cmci_storm_state, cpu) == CMCI_STORM_ACTIVE)
+		atomic_dec(&cmci_storm_on_cpus);
+
+	per_cpu(cmci_storm_state, cpu) = CMCI_STORM_NONE;
+}
+
+unsigned long mce_intel_adjust_timer(unsigned long interval)
+{
+	int r;
+
+	if (interval < CMCI_POLL_INTERVAL)
+		return interval;
+
+	switch (__this_cpu_read(cmci_storm_state)) {
+	case CMCI_STORM_ACTIVE:
+		/*
+		 * We switch back to interrupt mode once the poll timer has
+		 * silenced itself. That means no events recorded and the
+		 * timer interval is back to our poll interval.
+		 */
+		__this_cpu_write(cmci_storm_state, CMCI_STORM_SUBSIDED);
+		r = atomic_sub_return(1, &cmci_storm_on_cpus);
+		if (r == 0)
+			pr_notice("CMCI storm subsided: switching to interrupt mode\n");
+		/* FALLTHROUGH */
+
+	case CMCI_STORM_SUBSIDED:
+		/*
+		 * We wait for all cpus to go back to SUBSIDED
+		 * state. When that happens we switch back to
+		 * interrupt mode.
+		 */
+		if (!atomic_read(&cmci_storm_on_cpus)) {
+			__this_cpu_write(cmci_storm_state, CMCI_STORM_NONE);
+			cmci_reenable();
+			cmci_recheck();
+		}
+		return CMCI_POLL_INTERVAL;
+	default:
+		/*
+		 * We have shiny weather. Let the poll do whatever it
+		 * thinks.
+		 */
+		return interval;
+	}
+}
+
+static bool cmci_storm_detect(void)
+{
+	unsigned int cnt = __this_cpu_read(cmci_storm_cnt);
+	unsigned long ts = __this_cpu_read(cmci_time_stamp);
+	unsigned long now = jiffies;
+	int r;
+
+	if (__this_cpu_read(cmci_storm_state) != CMCI_STORM_NONE)
+		return true;
+
+	if (time_before_eq(now, ts + CMCI_STORM_INTERVAL)) {
+		cnt++;
+	} else {
+		cnt = 1;
+		__this_cpu_write(cmci_time_stamp, now);
+	}
+	__this_cpu_write(cmci_storm_cnt, cnt);
+
+	if (cnt <= CMCI_STORM_THRESHOLD)
+		return false;
+
+	cmci_clear();
+	__this_cpu_write(cmci_storm_state, CMCI_STORM_ACTIVE);
+	r = atomic_add_return(1, &cmci_storm_on_cpus);
+	mce_timer_kick(CMCI_POLL_INTERVAL);
+
+	if (r == 1)
+		pr_notice("CMCI storm detected: switching to poll mode\n");
+	return true;
+}
+
 /*
  * The interrupt handler. This is called on every event.
  * Just call the poller directly to log any events.
@@ -61,6 +165,8 @@ static int cmci_supported(int *banks)
  */
 static void intel_threshold_interrupt(void)
 {
+	if (cmci_storm_detect())
+		return;
 	machine_check_poll(MCP_TIMESTAMP, &__get_cpu_var(mce_banks_owned));
 	mce_notify_irq();
 }