x86 idle: Repair large-server 50-watt idle-power regression
Len Brown [Wed, 18 Dec 2013 21:44:57 +0000 (16:44 -0500)]
commit 40e2d7f9b5dae048789c64672bf3027fbb663ffa upstream.

Linux 3.10 changed the timing of how thread_info->flags is touched:

x86: Use generic idle loop
(7d1a941731fabf27e5fb6edbebb79fe856edb4e5)

This caused Intel NHM-EX and WSM-EX servers to experience a large number
of immediate MONITOR/MWAIT break wakeups, which caused cpuidle to demote
from deep C-states to shallow C-states, which caused these platforms
to experience a significant increase in idle power.

Note that this issue was already present before the commit above,
however, it wasn't seen often enough to be noticed in power measurements.

Here we extend an errata workaround from the Core2 EX "Dunnington"
to extend to NHM-EX and WSM-EX, to prevent these immediate
returns from MWAIT, reducing idle power on these platforms.

While only acpi_idle ran on Dunnington, intel_idle
may also run on these two newer systems.
As of today, there are no other models that are known
to need this tweak.

Link: http://lkml.kernel.org/r/CAJvTdK=%2BaNN66mYpCGgbHGCHhYQAKx-vB0kJSWjVpsNb_hOAtQ@mail.gmail.com
Signed-off-by: Len Brown <len.brown@intel.com>
Link: http://lkml.kernel.org/r/baff264285f6e585df757d58b17788feabc68918.1387403066.git.len.brown@intel.com
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

arch/x86/kernel/cpu/intel.c
drivers/idle/intel_idle.c

index 9b0c441..32d37d9 100644 (file)
@@ -387,7 +387,8 @@ static void __cpuinit init_intel(struct cpuinfo_x86 *c)
                        set_cpu_cap(c, X86_FEATURE_PEBS);
        }
 
-       if (c->x86 == 6 && c->x86_model == 29 && cpu_has_clflush)
+       if (c->x86 == 6 && cpu_has_clflush &&
+           (c->x86_model == 29 || c->x86_model == 46 || c->x86_model == 47))
                set_cpu_cap(c, X86_FEATURE_CLFLUSH_MONITOR);
 
 #ifdef CONFIG_X86_64
index f116d66..d47bb0f 100644 (file)
@@ -361,6 +361,9 @@ static int intel_idle(struct cpuidle_device *dev,
 
        if (!current_set_polling_and_test()) {
 
+               if (this_cpu_has(X86_FEATURE_CLFLUSH_MONITOR))
+                       clflush((void *)&current_thread_info()->flags);
+
                __monitor((void *)&current_thread_info()->flags, 0, 0);
                smp_mb();
                if (!need_resched())