Re: Intel i7/X 980 freezes with CONFIG_INTEL_IDLE and frequency scaling

[Carsten Emde]

On 04/02/2011 09:33 AM, Carsten Emde wrote:
> after upgrading a Fedora 14 based Intel i7/X 980 box to 2.6.39-rc1, the
> system freezes when frequency scaling is started. Looks like a complete
> CPU halt, since neither SysRq nor anything else can convince the box to
> provide any information.
> Loading acpi_cpufreq is save, but selecting the performance
> scaling_governor immediately freezes the system. Selecting the ondemand
> scaling_governor freezes the system when load is generated for the first
> time, presumably when the frequency is increased. Fortunately, when the
> kernel was built with another config file, the system did not freeze.
> After enabling and disabling some suspicious config items (before going
> through the hassle and bisect the config using ktest.pl), the culprit
> was found to be CONFIG_INTEL_IDLE. Fedora 14 enables it.
While digging deeper into this mystery, I found a detail that may 
connect this problem to the poweroff problem 
(https://lkml.org/lkml/2011/3/31/458).

The CPU (i7/X 980 box) apparently freezes when it executes

__monitor(&current_thread_info()->flags, 0, 0);

which is

static inline void __monitor(const void *eax, unsigned long ecx,
                             unsigned long edx)
{
        /* "monitor %eax, %ecx, %edx;" */
        asm volatile(".byte 0x0f, 0x01, 0xc8;"
                     :: "a" (eax), "c" (ecx), "d"(edx));
}

After removing this line from the execution path (see below patch), the 
problem is gone and the system no longer freezes when processor 
frequency scaling is enabled. I double-checked this, so I am pretty sure 
that this is, in fact, the case.

Interestingly, the __monitor() command is also used in mwait_play_dead() 
at arch/x86/kernel/smpboot.c where the poweroff problem is originating. 
Maybe, a misbehavior of this instruction also causes the poweroff problem.

But we should solve this problem first. Len or anybody else at Intel, I 
don't think that I can go any further without your help. Do you have any 
explanation what is going on here when the processor

vendor_id	: GenuineIntel
cpu family	: 6
model		: 44
model name	: Intel(R) Core(TM) i7 CPU       X 980  @ 3.33GHz
stepping	: 2

encounters the above instruction?

Thanks,
	-Carsten.


Index: linux-2.6.39-rc1/drivers/idle/intel_idle.c
===================================================================
--- linux-2.6.39-rc1.orig/drivers/idle/intel_idle.c
+++ linux-2.6.39-rc1/drivers/idle/intel_idle.c
@@ -99,11 +99,54 @@ static unsigned long long auto_demotion_
  */
 #define CPUIDLE_FLAG_TLB_FLUSHED	0x10000

+static int intel_idle_monitor(struct cpuidle_device *dev,
+	struct cpuidle_state *state, int monitor);
+
+static int intel_idle_westmere(struct cpuidle_device *dev,
+	struct cpuidle_state *state)
+{
+	return intel_idle_monitor(dev, state, 0);
+}
+
+static int intel_idle(struct cpuidle_device *dev,
+	struct cpuidle_state *state)
+{
+	return intel_idle_monitor(dev, state, 1);
+}
+
 /*
  * States are indexed by the cstate number,
  * which is also the index into the MWAIT hint array.
  * Thus C0 is a dummy.
  */
+static struct cpuidle_state westmere_cstates[MWAIT_MAX_NUM_CSTATES] = {
+	{ /* MWAIT C0 */ },
+	{ /* MWAIT C1 */
+		.name = "C1-NHM",
+		.desc = "MWAIT 0x00",
+		.driver_data = (void *) 0x00,
+		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.exit_latency = 3,
+		.target_residency = 6,
+		.enter = &intel_idle_westmere },
+	{ /* MWAIT C2 */
+		.name = "C3-NHM",
+		.desc = "MWAIT 0x10",
+		.driver_data = (void *) 0x10,
+		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
+		.exit_latency = 20,
+		.target_residency = 80,
+		.enter = &intel_idle_westmere },
+	{ /* MWAIT C3 */
+		.name = "C6-NHM",
+		.desc = "MWAIT 0x20",
+		.driver_data = (void *) 0x20,
+		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
+		.exit_latency = 200,
+		.target_residency = 800,
+		.enter = &intel_idle_westmere },
+};
+
 static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
 	{ /* MWAIT C0 */ },
 	{ /* MWAIT C1 */
@@ -212,7 +255,8 @@ static struct cpuidle_state atom_cstates
  * @state: cpuidle state
  *
  */
-static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state 
*state)
+static int intel_idle_monitor(struct cpuidle_device *dev,
+	struct cpuidle_state *state, int monitor)
 {
 	unsigned long ecx = 1; /* break on interrupt flag */
 	unsigned long eax = (unsigned long)cpuidle_get_statedata(state);
@@ -239,8 +283,8 @@ static int intel_idle(struct cpuidle_dev

 	stop_critical_timings();
 	if (!need_resched()) {
-
-		__monitor((void *)&current_thread_info()->flags, 0, 0);
+		if (monitor)
+			__monitor((void *)&current_thread_info()->flags, 0, 0);
 		smp_mb();
 		if (!need_resched())
 			__mwait(eax, ecx);
@@ -338,12 +382,17 @@ static int intel_idle_probe(void)
 	case 0x2E:	/* Nehalem-EX Xeon */
 	case 0x2F:	/* Westmere-EX Xeon */
 	case 0x25:	/* Westmere */
-	case 0x2C:	/* Westmere */
 		cpuidle_state_table = nehalem_cstates;
 		auto_demotion_disable_flags =
 			(NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE);
 		break;

+	case 0x2C:	/* Special Westmere? */
+		cpuidle_state_table = westmere_cstates;
+		auto_demotion_disable_flags =
+			(NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE);
+		break;
+
 	case 0x1C:	/* 28 - Atom Processor */
 		cpuidle_state_table = atom_cstates;
 		break;
--
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/