[PATCH 5/5] x86: mce: Add cmci poll mode

[Chen Gong <gong.chen@xxxxxxxxxxxxxxx>]

When CMCI is too many to handle, it should be disabled
to avoid hanging the whole system. In the meanwhile, CMCI poll
timer can be employed to receive CMCI periodically. When no
more CMCI happens CMCI handler can be switched from poll mode
to interrupt mode again.

By now, every CPU core owns one poll timer, but in fact, maybe
it should be enough that every package (or socket) owning one
poll timer. It is because CMCI gets broadcast to all threads on
the same socket. So if one cpu has a problem, all the cpus on
the same socket have a problem.

Signed-off-by: Chen Gong <gong.chen@xxxxxxxxxxxxxxx>
Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Tested-by: Chen Gong <gong.chen@xxxxxxxxxxxxxxx>
---
 arch/x86/kernel/cpu/mcheck/mce-internal.h |   12 ++++
 arch/x86/kernel/cpu/mcheck/mce.c          |   47 ++++++++++++--
 arch/x86/kernel/cpu/mcheck/mce_intel.c    |   99 ++++++++++++++++++++++++++++-
 3 files changed, 151 insertions(+), 7 deletions(-)

diff --git a/arch/x86/kernel/cpu/mcheck/mce-internal.h b/arch/x86/kernel/cpu/mcheck/mce-internal.h
index ed44c8a..6a05c1d 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 eff73e7..95738db0 100644
--- a/arch/x86/kernel/cpu/mcheck/mce.c
+++ b/arch/x86/kernel/cpu/mcheck/mce.c
@@ -1256,6 +1256,14 @@ 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);
@@ -1266,6 +1274,7 @@ 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();
 	}
 
 	/*
@@ -1273,14 +1282,38 @@ 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 */
@@ -1545,6 +1578,7 @@ 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);
@@ -1556,7 +1590,7 @@ 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);
 
@@ -2270,10 +2304,11 @@ 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 38e49bc..693bc7d 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_TRESHOLD	5
+
+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 +70,84 @@ 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)
+{
+	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);
+		atomic_dec(&cmci_storm_on_cpus);
+
+	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 wheather, 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;
+
+	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_TRESHOLD)
+		return false;
+
+	cmci_clear();
+	__this_cpu_write(cmci_storm_state, CMCI_STORM_ACTIVE);
+	atomic_inc(&cmci_storm_on_cpus);
+	mce_timer_kick(CMCI_POLL_INTERVAL);
+	return true;
+}
+
 /*
  * The interrupt handler. This is called on every event.
  * Just call the poller directly to log any events.
@@ -61,6 +156,8 @@ 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();
 }
-- 
1.7.10.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/