intel_idle: disable NHM/WSM HW C-state auto-demotion
Hardware C-state auto-demotion is a mechanism where the HW overrides
the OS C-state request, instead demoting to a shallower state,
which is less expensive, but saves less power.
Modern Linux should generally get exactly the states it requests.
In particular, when a CPU is taken off-line, it must not be demoted, else
it can prevent the entire package from reaching deep C-states.
https://bugzilla.kernel.org/show_bug.cgi?id=25252
Signed-off-by: Len Brown <len.brown@intel.com>
diff --git a/drivers/idle/intel_idle.c b/drivers/idle/intel_idle.c
index 1fa091e..32b25bc 100644
--- a/drivers/idle/intel_idle.c
+++ b/drivers/idle/intel_idle.c
@@ -62,6 +62,7 @@
#include <linux/notifier.h>
#include <linux/cpu.h>
#include <asm/mwait.h>
+#include <asm/msr.h>
#define INTEL_IDLE_VERSION "0.4"
#define PREFIX "intel_idle: "
@@ -85,6 +86,12 @@
static struct cpuidle_state *cpuidle_state_table;
/*
+ * Hardware C-state auto-demotion may not always be optimal.
+ * Indicate which enable bits to clear here.
+ */
+static unsigned long long auto_demotion_disable_flags;
+
+/*
* Set this flag for states where the HW flushes the TLB for us
* and so we don't need cross-calls to keep it consistent.
* If this flag is set, SW flushes the TLB, so even if the
@@ -281,6 +288,15 @@
.notifier_call = setup_broadcast_cpuhp_notify,
};
+static void auto_demotion_disable(void *dummy)
+{
+ unsigned long long msr_bits;
+
+ rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
+ msr_bits &= ~auto_demotion_disable_flags;
+ wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
+}
+
/*
* intel_idle_probe()
*/
@@ -324,6 +340,8 @@
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 0x1C: /* 28 - Atom Processor */
@@ -436,6 +454,8 @@
return -EIO;
}
}
+ if (auto_demotion_disable_flags)
+ smp_call_function(auto_demotion_disable, NULL, 1);
return 0;
}