[PATCH] PM / arch: x86: Rework the MSR_IA32_ENERGY_PERF_BIAS handling

* [PATCH] PM / arch: x86: Rework the MSR_IA32_ENERGY_PERF_BIAS handling
@ 2019-03-21  0:00 Rafael J. Wysocki
  2019-03-21  7:58 ` Peter Zijlstra
  0 siblings, 1 reply; 8+ messages in thread
From: Rafael J. Wysocki @ 2019-03-21  0:00 UTC (permalink / raw)
  To: x86
  Cc: LKML, Len Brown, Linux PM, Srinivas Pandruvada, Laura Abbott,
	Thomas Gleixner, Peter Zijlstra, Ingo Molnar, Simon Schricker,
	Borislav Petkov, Hannes Reinecke

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

The current handling of MSR_IA32_ENERGY_PERF_BIAS in the kernel is
problematic, because it may cause changes made by user space to that
MSR (with the help of the x86_energy_perf_policy tool, for example)
to be lost every time a CPU goes offline and then back online as well
as during system-wide power management transitions into sleep states
and back into the working state.

The first problem is that if the current EPB value for a CPU going
online is 0 ('performance'), the kernel will change it to 6 ('normal')
regardless of whether or not this is the first bring-up of that CPU.
That also happens during system-wide resume from sleep states
(including, but not limited to, hibernation).  However, the EPB may
have been adjusted by user space this way and the kernel should not
blindly override that setting.

The second problem is that if the platform firmware resets the EPB
values for any CPUs during system-wide resume from a sleep state,
the kernel will not restore their previous EPB values that may
have been set by user space before the preceding system-wide
suspend transition.  Again, that behavior may at least be confusing
from the user space perspective.

In order to address these issues, rework the handling of
MSR_IA32_ENERGY_PERF_BIAS so that the EPB value is saved on CPU
offline and restored on CPU online as well as (for the boot CPU)
during the syscore stages of system-wide suspend and resume
transitions, respectively.

However, retain the policy by which the EPB is set to 6 ('normal')
on the first bring-up of each CPU if its initial value is 0, based
on the observation that 0 may mean 'not initialized' just as well as
'performance' in that case.

While at it, move the MSR_IA32_ENERGY_PERF_BIAS handling code into
a separate file and document it in Documentation/admin-guide.

Fixes: abe48b108247 (x86, intel, power: Initialize MSR_IA32_ENERGY_PERF_BIAS)
Fixes: b51ef52df71c (x86/cpu: Restore MSR_IA32_ENERGY_PERF_BIAS after resume)
Reported-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---

I do realize that this may be regarded as overly conservative, but
I'm quite reluctant to change the policy choice made in 2011 and
present since then without rock solid evidence that it would not
introduce regressions anywhere.

---
 Documentation/admin-guide/pm/intel_epb.rst     |    6 +
 Documentation/admin-guide/pm/working-state.rst |    1 
 arch/x86/kernel/cpu/Makefile                   |    2 
 arch/x86/kernel/cpu/common.c                   |   17 ---
 arch/x86/kernel/cpu/cpu.h                      |    1 
 arch/x86/kernel/cpu/intel.c                    |   34 ------
 arch/x86/kernel/cpu/intel_epb.c                |  131 +++++++++++++++++++++++++
 include/linux/cpuhotplug.h                     |    1 
 8 files changed, 140 insertions(+), 53 deletions(-)

Index: linux-pm/include/linux/cpuhotplug.h
===================================================================

--- linux-pm.orig/include/linux/cpuhotplug.h
+++ linux-pm/include/linux/cpuhotplug.h
@@ -147,6 +147,7 @@ enum cpuhp_state {
 	CPUHP_AP_X86_VDSO_VMA_ONLINE,
 	CPUHP_AP_IRQ_AFFINITY_ONLINE,
 	CPUHP_AP_ARM_MVEBU_SYNC_CLOCKS,
+	CPUHP_AP_X86_INTEL_EPB_ONLINE,
 	CPUHP_AP_PERF_ONLINE,
 	CPUHP_AP_PERF_X86_ONLINE,
 	CPUHP_AP_PERF_X86_UNCORE_ONLINE,
Index: linux-pm/arch/x86/kernel/cpu/intel.c
===================================================================
--- linux-pm.orig/arch/x86/kernel/cpu/intel.c
+++ linux-pm/arch/x86/kernel/cpu/intel.c
@@ -596,36 +596,6 @@ detect_keyid_bits:
 	c->x86_phys_bits -= keyid_bits;
 }
 
-static void init_intel_energy_perf(struct cpuinfo_x86 *c)
-{
-	u64 epb;
-
-	/*
-	 * Initialize MSR_IA32_ENERGY_PERF_BIAS if not already initialized.
-	 * (x86_energy_perf_policy(8) is available to change it at run-time.)
-	 */
-	if (!cpu_has(c, X86_FEATURE_EPB))
-		return;
-
-	rdmsrl(MSR_IA32_ENERGY_PERF_BIAS, epb);
-	if ((epb & 0xF) != ENERGY_PERF_BIAS_PERFORMANCE)
-		return;
-
-	pr_warn_once("ENERGY_PERF_BIAS: Set to 'normal', was 'performance'\n");
-	pr_warn_once("ENERGY_PERF_BIAS: View and update with x86_energy_perf_policy(8)\n");
-	epb = (epb & ~0xF) | ENERGY_PERF_BIAS_NORMAL;
-	wrmsrl(MSR_IA32_ENERGY_PERF_BIAS, epb);
-}
-
-static void intel_bsp_resume(struct cpuinfo_x86 *c)
-{
-	/*
-	 * MSR_IA32_ENERGY_PERF_BIAS is lost across suspend/resume,
-	 * so reinitialize it properly like during bootup:
-	 */
-	init_intel_energy_perf(c);
-}
-
 static void init_cpuid_fault(struct cpuinfo_x86 *c)
 {
 	u64 msr;
@@ -763,8 +733,6 @@ static void init_intel(struct cpuinfo_x8
 	if (cpu_has(c, X86_FEATURE_TME))
 		detect_tme(c);
 
-	init_intel_energy_perf(c);
-
 	init_intel_misc_features(c);
 }
 
@@ -1023,9 +991,7 @@ static const struct cpu_dev intel_cpu_de
 	.c_detect_tlb	= intel_detect_tlb,
 	.c_early_init   = early_init_intel,
 	.c_init		= init_intel,
-	.c_bsp_resume	= intel_bsp_resume,
 	.c_x86_vendor	= X86_VENDOR_INTEL,
 };
 
 cpu_dev_register(intel_cpu_dev);
-
Index: linux-pm/arch/x86/kernel/cpu/common.c
===================================================================
--- linux-pm.orig/arch/x86/kernel/cpu/common.c
+++ linux-pm/arch/x86/kernel/cpu/common.c
@@ -1864,23 +1864,6 @@ void cpu_init(void)
 }
 #endif
 
-static void bsp_resume(void)
-{
-	if (this_cpu->c_bsp_resume)
-		this_cpu->c_bsp_resume(&boot_cpu_data);
-}
-
-static struct syscore_ops cpu_syscore_ops = {
-	.resume		= bsp_resume,
-};
-
-static int __init init_cpu_syscore(void)
-{
-	register_syscore_ops(&cpu_syscore_ops);
-	return 0;
-}
-core_initcall(init_cpu_syscore);
-
 /*
  * The microcode loader calls this upon late microcode load to recheck features,
  * only when microcode has been updated. Caller holds microcode_mutex and CPU
Index: linux-pm/arch/x86/kernel/cpu/cpu.h
===================================================================
--- linux-pm.orig/arch/x86/kernel/cpu/cpu.h
+++ linux-pm/arch/x86/kernel/cpu/cpu.h
@@ -14,7 +14,6 @@ struct cpu_dev {
 	void		(*c_init)(struct cpuinfo_x86 *);
 	void		(*c_identify)(struct cpuinfo_x86 *);
 	void		(*c_detect_tlb)(struct cpuinfo_x86 *);
-	void		(*c_bsp_resume)(struct cpuinfo_x86 *);
 	int		c_x86_vendor;
 #ifdef CONFIG_X86_32
 	/* Optional vendor specific routine to obtain the cache size. */
Index: linux-pm/arch/x86/kernel/cpu/Makefile
===================================================================
--- linux-pm.orig/arch/x86/kernel/cpu/Makefile
+++ linux-pm/arch/x86/kernel/cpu/Makefile
@@ -28,7 +28,7 @@ obj-y			+= cpuid-deps.o
 obj-$(CONFIG_PROC_FS)	+= proc.o
 obj-$(CONFIG_X86_FEATURE_NAMES) += capflags.o powerflags.o
 
-obj-$(CONFIG_CPU_SUP_INTEL)		+= intel.o intel_pconfig.o
+obj-$(CONFIG_CPU_SUP_INTEL)		+= intel.o intel_pconfig.o intel_epb.o
 obj-$(CONFIG_CPU_SUP_AMD)		+= amd.o
 obj-$(CONFIG_CPU_SUP_HYGON)		+= hygon.o
 obj-$(CONFIG_CPU_SUP_CYRIX_32)		+= cyrix.o
Index: linux-pm/arch/x86/kernel/cpu/intel_epb.c
===================================================================
--- /dev/null
+++ linux-pm/arch/x86/kernel/cpu/intel_epb.c
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Intel Performance and Energy Bias Hint support.
+ *
+ * Copyright (C) 2018 Intel Corporation
+ *
+ * Author:
+ *	Rafael J. Wysocki <rafael.j.wysocki@intel.com>
+ */
+
+#include <linux/cpuhotplug.h>
+#include <linux/kernel.h>
+#include <linux/syscore_ops.h>
+
+#include <asm/cpufeature.h>
+#include <asm/msr.h>
+
+/**
+ * DOC: overview
+ *
+ * The Performance and Energy Bias Hint (EPB) allows software to specify its
+ * preference with respect to the power-performance tradeoffs present in the
+ * processor.  Generally, the EPB is expected to be set by user space through
+ * the generic MSR interface (with the help of the x86_energy_perf_policy tool),
+ * but there are two reasons for the kernel to touch it.
+ *
+ * First, there are systems where the platform firmware resets the EPB during
+ * system-wide transitions from sleep states back into the working state
+ * effectively causing the previous EPB updates by user space to be lost.
+ * Thus the kernel needs to save the current EPB values for all CPUs during
+ * system-wide transitions to sleep states and restore them on the way back to
+ * the working state.  That can be achieved by saving EPB for secondary CPUs
+ * when they are taken offline during transitions into system sleep states and
+ * for the boot CPU in a syscore suspend operation, so that it can be restored
+ * for the boot CPU in a syscore resume operation and for the other CPUs when
+ * they are brought back online.  However, CPUs that are already offline when
+ * a system-wide PM transition is started are not taken offline again, but their
+ * EPB values may still be reset by the platform firmware during the transition,
+ * so in fact it is necessary to save the EPB of any CPU taken offline and to
+ * restore it when the given CPU goes back online at all times.
+ *
+ * Second, on many systems the initial EPB value coming from the platform
+ * firmware is 0 ('performance') and at least on some of them that is because
+ * the platform firmware does not initialize EPB at all with the assumption that
+ * the OS will do that anyway.  That sometimes is problematic, as it may cause
+ * the system battery to drain too fast, for example, so it is better to adjust
+ * it on CPU bring-up and if the initial EPB value for a given CPU is 0, the
+ * kernel changes it to 6 ('normal').
+ */
+
+static DEFINE_PER_CPU(u8, saved_epb);
+
+#define EPB_MASK	((u8)0x0f)
+#define EPB_SAVED	((u8)0x10)
+
+static int intel_epb_save(void)
+{
+	u64 epb;
+
+	rdmsrl(MSR_IA32_ENERGY_PERF_BIAS, epb);
+	/*
+	 * Ensure that saved_epb will always be nonzero after this write even if
+	 * the EPB value read from the MSR is 0.
+	 */
+	this_cpu_write(saved_epb, ((u8)epb & EPB_MASK) | EPB_SAVED);
+
+	return 0;
+}
+
+static void intel_epb_restore(void)
+{
+	u8 val = this_cpu_read(saved_epb);
+	u64 epb;
+
+	rdmsrl(MSR_IA32_ENERGY_PERF_BIAS, epb);
+	if (val) {
+		val &= EPB_MASK;
+	} else {
+		/*
+		 * Because intel_epb_save() has not run for the current CPU yet,
+		 * it is going online for the first time, so if its EPB value is
+		 * 0 ('performance') at this point, assume that it has not been
+		 * initialized by the platform firmware and set it to 6
+		 * ('normal').
+		 */
+		val = (u8)epb & EPB_MASK;
+		if (val == ENERGY_PERF_BIAS_PERFORMANCE) {
+			val = ENERGY_PERF_BIAS_NORMAL;
+			pr_warn_once("ENERGY_PERF_BIAS: Set to 'normal', was 'performance'\n");
+		}
+	}
+	wrmsrl(MSR_IA32_ENERGY_PERF_BIAS, (epb & ~(u64)EPB_MASK) | (u64)val);
+}
+
+static struct syscore_ops intel_epb_syscore_ops = {
+	.suspend = intel_epb_save,
+	.resume = intel_epb_restore,
+};
+
+static int intel_epb_online(unsigned int cpu)
+{
+	intel_epb_restore();
+	return 0;
+}
+
+static int intel_epb_offline(unsigned int cpu)
+{
+	return intel_epb_save();
+}
+
+static __init int intel_epb_init(void)
+{
+	int ret;
+
+	if (!boot_cpu_has(X86_FEATURE_EPB))
+		return -ENODEV;
+
+	ret = cpuhp_setup_state(CPUHP_AP_X86_INTEL_EPB_ONLINE,
+				"x86/intel/epb:online", intel_epb_online,
+				intel_epb_offline);
+	if (ret < 0)
+		goto err_out_online;
+
+	register_syscore_ops(&intel_epb_syscore_ops);
+	return 0;
+
+err_out_online:
+	cpuhp_remove_state(CPUHP_AP_X86_INTEL_EPB_ONLINE);
+	return ret;
+}
+subsys_initcall(intel_epb_init);
Index: linux-pm/Documentation/admin-guide/pm/intel_epb.rst
===================================================================
--- /dev/null
+++ linux-pm/Documentation/admin-guide/pm/intel_epb.rst
@@ -0,0 +1,6 @@
+======================================
+Intel Performance and Energy Bias Hint
+======================================
+
+.. kernel-doc:: arch/x86/kernel/cpu/intel_epb.c
+   :doc: overview
Index: linux-pm/Documentation/admin-guide/pm/working-state.rst
===================================================================
--- linux-pm.orig/Documentation/admin-guide/pm/working-state.rst
+++ linux-pm/Documentation/admin-guide/pm/working-state.rst
@@ -8,3 +8,4 @@ Working-State Power Management
    cpuidle
    cpufreq
    intel_pstate
+   intel_epb


^ permalink raw reply	[flat|nested] 8+ messages in thread