perf/x86: Add Intel cstate PMUs support
authorKan Liang <kan.liang@intel.com>
Mon, 28 Sep 2015 12:30:04 +0000 (08:30 -0400)
committerIngo Molnar <mingo@kernel.org>
Tue, 6 Oct 2015 15:31:51 +0000 (17:31 +0200)
This patch adds new PMUs to support cstate related free running
(read-only) counters. These counters may be used simultaneously by other
tools, such as turbostat. However, it still make sense to implement them
in perf. Because we can conveniently collect them together with other
events, and allow to use them from tools without special MSR access
code.

These counters include CORE_C*_RESIDENCY and PKG_C*_RESIDENCY.
According to counters' scope and category, two PMUs are registered with
the perf_event core subsystem.

 - 'cstate_core': The counter is available for each physical core. The
                  counters include CORE_C*_RESIDENCY.

 - 'cstate_pkg':  The counter is available for each physical package. The
                  counters include PKG_C*_RESIDENCY.

The events are exposed in sysfs for use by perf stat and other tools.
The files are:

  /sys/devices/cstate_core/events/c*-residency
  /sys/devices/cstate_pkg/events/c*-residency

These events only support system-wide mode counting.
The /sys/devices/cstate_*/cpumask file can be used by tools to figure
out which CPUs to monitor by default.

The PMU type (attr->type) is dynamically allocated and is available from
/sys/devices/core_misc/type and /sys/device/cstate_*/type.

Sampling is not supported.

Here is an example.

 - To caculate the fraction of time when the core is running in C6 state
   CORE_C6_time% = CORE_C6_RESIDENCY / TSC

 # perf stat -x, -e"cstate_core/c6-residency/,msr/tsc/" -C0 -- taskset -c 0 sleep 5

   11838820015,,cstate_core/c6-residency/,5175919658,100.00
   11877130740,,msr/tsc/,5175922010,100.00

 For sleep, 99.7% of time we ran in C6 state.

 # perf stat -x, -e"cstate_core/c6-residency/,msr/tsc/" -C0 -- taskset -c 0 busyloop

   1253316,,cstate_core/c6-residency/,4360969154,100.00
   10012635248,,msr/tsc/,4360972366,100.00

 For busyloop, 0.01% of time we ran in C6 state.

Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@kernel.org
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1443443404-8581-1-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/kernel/cpu/Makefile
arch/x86/kernel/cpu/perf_event_intel_cstate.c [new file with mode: 0644]

index 4eb065c6bed26b38a2804ffcab57f9e09274a7ec..58031303e30488c540d609f95d0693918c09fa64 100644 (file)
@@ -41,6 +41,7 @@ obj-$(CONFIG_CPU_SUP_INTEL)           += perf_event_p6.o perf_event_knc.o perf_event_p4.o
 obj-$(CONFIG_CPU_SUP_INTEL)            += perf_event_intel_lbr.o perf_event_intel_ds.o perf_event_intel.o
 obj-$(CONFIG_CPU_SUP_INTEL)            += perf_event_intel_rapl.o perf_event_intel_cqm.o
 obj-$(CONFIG_CPU_SUP_INTEL)            += perf_event_intel_pt.o perf_event_intel_bts.o
+obj-$(CONFIG_CPU_SUP_INTEL)            += perf_event_intel_cstate.o
 
 obj-$(CONFIG_PERF_EVENTS_INTEL_UNCORE) += perf_event_intel_uncore.o \
                                           perf_event_intel_uncore_snb.o \
diff --git a/arch/x86/kernel/cpu/perf_event_intel_cstate.c b/arch/x86/kernel/cpu/perf_event_intel_cstate.c
new file mode 100644 (file)
index 0000000..75a38b5
--- /dev/null
@@ -0,0 +1,694 @@
+/*
+ * perf_event_intel_cstate.c: support cstate residency counters
+ *
+ * Copyright (C) 2015, Intel Corp.
+ * Author: Kan Liang (kan.liang@intel.com)
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Library General Public License for more details.
+ *
+ */
+
+/*
+ * This file export cstate related free running (read-only) counters
+ * for perf. These counters may be use simultaneously by other tools,
+ * such as turbostat. However, it still make sense to implement them
+ * in perf. Because we can conveniently collect them together with
+ * other events, and allow to use them from tools without special MSR
+ * access code.
+ *
+ * The events only support system-wide mode counting. There is no
+ * sampling support because it is not supported by the hardware.
+ *
+ * According to counters' scope and category, two PMUs are registered
+ * with the perf_event core subsystem.
+ *  - 'cstate_core': The counter is available for each physical core.
+ *    The counters include CORE_C*_RESIDENCY.
+ *  - 'cstate_pkg': The counter is available for each physical package.
+ *    The counters include PKG_C*_RESIDENCY.
+ *
+ * All of these counters are specified in the IntelĀ® 64 and IA-32
+ * Architectures Software Developer.s Manual Vol3b.
+ *
+ * Model specific counters:
+ *     MSR_CORE_C1_RES: CORE C1 Residency Counter
+ *                      perf code: 0x00
+ *                      Available model: SLM,AMT
+ *                      Scope: Core (each processor core has a MSR)
+ *     MSR_CORE_C3_RESIDENCY: CORE C3 Residency Counter
+ *                            perf code: 0x01
+ *                            Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Core
+ *     MSR_CORE_C6_RESIDENCY: CORE C6 Residency Counter
+ *                            perf code: 0x02
+ *                            Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Core
+ *     MSR_CORE_C7_RESIDENCY: CORE C7 Residency Counter
+ *                            perf code: 0x03
+ *                            Available model: SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Core
+ *     MSR_PKG_C2_RESIDENCY:  Package C2 Residency Counter.
+ *                            perf code: 0x00
+ *                            Available model: SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Package (physical package)
+ *     MSR_PKG_C3_RESIDENCY:  Package C3 Residency Counter.
+ *                            perf code: 0x01
+ *                            Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Package (physical package)
+ *     MSR_PKG_C6_RESIDENCY:  Package C6 Residency Counter.
+ *                            perf code: 0x02
+ *                            Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Package (physical package)
+ *     MSR_PKG_C7_RESIDENCY:  Package C7 Residency Counter.
+ *                            perf code: 0x03
+ *                            Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL
+ *                            Scope: Package (physical package)
+ *     MSR_PKG_C8_RESIDENCY:  Package C8 Residency Counter.
+ *                            perf code: 0x04
+ *                            Available model: HSW ULT only
+ *                            Scope: Package (physical package)
+ *     MSR_PKG_C9_RESIDENCY:  Package C9 Residency Counter.
+ *                            perf code: 0x05
+ *                            Available model: HSW ULT only
+ *                            Scope: Package (physical package)
+ *     MSR_PKG_C10_RESIDENCY: Package C10 Residency Counter.
+ *                            perf code: 0x06
+ *                            Available model: HSW ULT only
+ *                            Scope: Package (physical package)
+ *
+ */
+
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/perf_event.h>
+#include <asm/cpu_device_id.h>
+#include "perf_event.h"
+
+#define DEFINE_CSTATE_FORMAT_ATTR(_var, _name, _format)                \
+static ssize_t __cstate_##_var##_show(struct kobject *kobj,    \
+                               struct kobj_attribute *attr,    \
+                               char *page)                     \
+{                                                              \
+       BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);             \
+       return sprintf(page, _format "\n");                     \
+}                                                              \
+static struct kobj_attribute format_attr_##_var =              \
+       __ATTR(_name, 0444, __cstate_##_var##_show, NULL)
+
+static ssize_t cstate_get_attr_cpumask(struct device *dev,
+                                      struct device_attribute *attr,
+                                      char *buf);
+
+struct perf_cstate_msr {
+       u64     msr;
+       struct  perf_pmu_events_attr *attr;
+       bool    (*test)(int idx);
+};
+
+
+/* cstate_core PMU */
+
+static struct pmu cstate_core_pmu;
+static bool has_cstate_core;
+
+enum perf_cstate_core_id {
+       /*
+        * cstate_core events
+        */
+       PERF_CSTATE_CORE_C1_RES = 0,
+       PERF_CSTATE_CORE_C3_RES,
+       PERF_CSTATE_CORE_C6_RES,
+       PERF_CSTATE_CORE_C7_RES,
+
+       PERF_CSTATE_CORE_EVENT_MAX,
+};
+
+bool test_core(int idx)
+{
+       if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
+           boot_cpu_data.x86 != 6)
+               return false;
+
+       switch (boot_cpu_data.x86_model) {
+       case 30: /* 45nm Nehalem    */
+       case 26: /* 45nm Nehalem-EP */
+       case 46: /* 45nm Nehalem-EX */
+
+       case 37: /* 32nm Westmere    */
+       case 44: /* 32nm Westmere-EP */
+       case 47: /* 32nm Westmere-EX */
+               if (idx == PERF_CSTATE_CORE_C3_RES ||
+                   idx == PERF_CSTATE_CORE_C6_RES)
+                       return true;
+               break;
+       case 42: /* 32nm SandyBridge         */
+       case 45: /* 32nm SandyBridge-E/EN/EP */
+
+       case 58: /* 22nm IvyBridge       */
+       case 62: /* 22nm IvyBridge-EP/EX */
+
+       case 60: /* 22nm Haswell Core */
+       case 63: /* 22nm Haswell Server */
+       case 69: /* 22nm Haswell ULT */
+       case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */
+
+       case 61: /* 14nm Broadwell Core-M */
+       case 86: /* 14nm Broadwell Xeon D */
+       case 71: /* 14nm Broadwell + GT3e (Intel Iris Pro graphics) */
+       case 79: /* 14nm Broadwell Server */
+
+       case 78: /* 14nm Skylake Mobile */
+       case 94: /* 14nm Skylake Desktop */
+               if (idx == PERF_CSTATE_CORE_C3_RES ||
+                   idx == PERF_CSTATE_CORE_C6_RES ||
+                   idx == PERF_CSTATE_CORE_C7_RES)
+                       return true;
+               break;
+       case 55: /* 22nm Atom "Silvermont"                */
+       case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */
+       case 76: /* 14nm Atom "Airmont"                   */
+               if (idx == PERF_CSTATE_CORE_C1_RES ||
+                   idx == PERF_CSTATE_CORE_C6_RES)
+                       return true;
+               break;
+       }
+
+       return false;
+}
+
+PMU_EVENT_ATTR_STRING(c1-residency, evattr_cstate_core_c1, "event=0x00");
+PMU_EVENT_ATTR_STRING(c3-residency, evattr_cstate_core_c3, "event=0x01");
+PMU_EVENT_ATTR_STRING(c6-residency, evattr_cstate_core_c6, "event=0x02");
+PMU_EVENT_ATTR_STRING(c7-residency, evattr_cstate_core_c7, "event=0x03");
+
+static struct perf_cstate_msr core_msr[] = {
+       [PERF_CSTATE_CORE_C1_RES] = { MSR_CORE_C1_RES,          &evattr_cstate_core_c1, test_core, },
+       [PERF_CSTATE_CORE_C3_RES] = { MSR_CORE_C3_RESIDENCY,    &evattr_cstate_core_c3, test_core, },
+       [PERF_CSTATE_CORE_C6_RES] = { MSR_CORE_C6_RESIDENCY,    &evattr_cstate_core_c6, test_core, },
+       [PERF_CSTATE_CORE_C7_RES] = { MSR_CORE_C7_RESIDENCY,    &evattr_cstate_core_c7, test_core, },
+};
+
+static struct attribute *core_events_attrs[PERF_CSTATE_CORE_EVENT_MAX + 1] = {
+       NULL,
+};
+
+static struct attribute_group core_events_attr_group = {
+       .name = "events",
+       .attrs = core_events_attrs,
+};
+
+DEFINE_CSTATE_FORMAT_ATTR(core_event, event, "config:0-63");
+static struct attribute *core_format_attrs[] = {
+       &format_attr_core_event.attr,
+       NULL,
+};
+
+static struct attribute_group core_format_attr_group = {
+       .name = "format",
+       .attrs = core_format_attrs,
+};
+
+static cpumask_t cstate_core_cpu_mask;
+static DEVICE_ATTR(cpumask, S_IRUGO, cstate_get_attr_cpumask, NULL);
+
+static struct attribute *cstate_cpumask_attrs[] = {
+       &dev_attr_cpumask.attr,
+       NULL,
+};
+
+static struct attribute_group cpumask_attr_group = {
+       .attrs = cstate_cpumask_attrs,
+};
+
+static const struct attribute_group *core_attr_groups[] = {
+       &core_events_attr_group,
+       &core_format_attr_group,
+       &cpumask_attr_group,
+       NULL,
+};
+
+/* cstate_core PMU end */
+
+
+/* cstate_pkg PMU */
+
+static struct pmu cstate_pkg_pmu;
+static bool has_cstate_pkg;
+
+enum perf_cstate_pkg_id {
+       /*
+        * cstate_pkg events
+        */
+       PERF_CSTATE_PKG_C2_RES = 0,
+       PERF_CSTATE_PKG_C3_RES,
+       PERF_CSTATE_PKG_C6_RES,
+       PERF_CSTATE_PKG_C7_RES,
+       PERF_CSTATE_PKG_C8_RES,
+       PERF_CSTATE_PKG_C9_RES,
+       PERF_CSTATE_PKG_C10_RES,
+
+       PERF_CSTATE_PKG_EVENT_MAX,
+};
+
+bool test_pkg(int idx)
+{
+       if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
+           boot_cpu_data.x86 != 6)
+               return false;
+
+       switch (boot_cpu_data.x86_model) {
+       case 30: /* 45nm Nehalem    */
+       case 26: /* 45nm Nehalem-EP */
+       case 46: /* 45nm Nehalem-EX */
+
+       case 37: /* 32nm Westmere    */
+       case 44: /* 32nm Westmere-EP */
+       case 47: /* 32nm Westmere-EX */
+               if (idx == PERF_CSTATE_CORE_C3_RES ||
+                   idx == PERF_CSTATE_CORE_C6_RES ||
+                   idx == PERF_CSTATE_CORE_C7_RES)
+                       return true;
+               break;
+       case 42: /* 32nm SandyBridge         */
+       case 45: /* 32nm SandyBridge-E/EN/EP */
+
+       case 58: /* 22nm IvyBridge       */
+       case 62: /* 22nm IvyBridge-EP/EX */
+
+       case 60: /* 22nm Haswell Core */
+       case 63: /* 22nm Haswell Server */
+       case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */
+
+       case 61: /* 14nm Broadwell Core-M */
+       case 86: /* 14nm Broadwell Xeon D */
+       case 71: /* 14nm Broadwell + GT3e (Intel Iris Pro graphics) */
+       case 79: /* 14nm Broadwell Server */
+
+       case 78: /* 14nm Skylake Mobile */
+       case 94: /* 14nm Skylake Desktop */
+               if (idx == PERF_CSTATE_PKG_C2_RES ||
+                   idx == PERF_CSTATE_PKG_C3_RES ||
+                   idx == PERF_CSTATE_PKG_C6_RES ||
+                   idx == PERF_CSTATE_PKG_C7_RES)
+                       return true;
+               break;
+       case 55: /* 22nm Atom "Silvermont"                */
+       case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */
+       case 76: /* 14nm Atom "Airmont"                   */
+               if (idx == PERF_CSTATE_CORE_C6_RES)
+                       return true;
+               break;
+       case 69: /* 22nm Haswell ULT */
+               if (idx == PERF_CSTATE_PKG_C2_RES ||
+                   idx == PERF_CSTATE_PKG_C3_RES ||
+                   idx == PERF_CSTATE_PKG_C6_RES ||
+                   idx == PERF_CSTATE_PKG_C7_RES ||
+                   idx == PERF_CSTATE_PKG_C8_RES ||
+                   idx == PERF_CSTATE_PKG_C9_RES ||
+                   idx == PERF_CSTATE_PKG_C10_RES)
+                       return true;
+               break;
+       }
+
+       return false;
+}
+
+PMU_EVENT_ATTR_STRING(c2-residency, evattr_cstate_pkg_c2, "event=0x00");
+PMU_EVENT_ATTR_STRING(c3-residency, evattr_cstate_pkg_c3, "event=0x01");
+PMU_EVENT_ATTR_STRING(c6-residency, evattr_cstate_pkg_c6, "event=0x02");
+PMU_EVENT_ATTR_STRING(c7-residency, evattr_cstate_pkg_c7, "event=0x03");
+PMU_EVENT_ATTR_STRING(c8-residency, evattr_cstate_pkg_c8, "event=0x04");
+PMU_EVENT_ATTR_STRING(c9-residency, evattr_cstate_pkg_c9, "event=0x05");
+PMU_EVENT_ATTR_STRING(c10-residency, evattr_cstate_pkg_c10, "event=0x06");
+
+static struct perf_cstate_msr pkg_msr[] = {
+       [PERF_CSTATE_PKG_C2_RES] = { MSR_PKG_C2_RESIDENCY,      &evattr_cstate_pkg_c2,  test_pkg, },
+       [PERF_CSTATE_PKG_C3_RES] = { MSR_PKG_C3_RESIDENCY,      &evattr_cstate_pkg_c3,  test_pkg, },
+       [PERF_CSTATE_PKG_C6_RES] = { MSR_PKG_C6_RESIDENCY,      &evattr_cstate_pkg_c6,  test_pkg, },
+       [PERF_CSTATE_PKG_C7_RES] = { MSR_PKG_C7_RESIDENCY,      &evattr_cstate_pkg_c7,  test_pkg, },
+       [PERF_CSTATE_PKG_C8_RES] = { MSR_PKG_C8_RESIDENCY,      &evattr_cstate_pkg_c8,  test_pkg, },
+       [PERF_CSTATE_PKG_C9_RES] = { MSR_PKG_C9_RESIDENCY,      &evattr_cstate_pkg_c9,  test_pkg, },
+       [PERF_CSTATE_PKG_C10_RES] = { MSR_PKG_C10_RESIDENCY,    &evattr_cstate_pkg_c10, test_pkg, },
+};
+
+static struct attribute *pkg_events_attrs[PERF_CSTATE_PKG_EVENT_MAX + 1] = {
+       NULL,
+};
+
+static struct attribute_group pkg_events_attr_group = {
+       .name = "events",
+       .attrs = pkg_events_attrs,
+};
+
+DEFINE_CSTATE_FORMAT_ATTR(pkg_event, event, "config:0-63");
+static struct attribute *pkg_format_attrs[] = {
+       &format_attr_pkg_event.attr,
+       NULL,
+};
+static struct attribute_group pkg_format_attr_group = {
+       .name = "format",
+       .attrs = pkg_format_attrs,
+};
+
+static cpumask_t cstate_pkg_cpu_mask;
+
+static const struct attribute_group *pkg_attr_groups[] = {
+       &pkg_events_attr_group,
+       &pkg_format_attr_group,
+       &cpumask_attr_group,
+       NULL,
+};
+
+/* cstate_pkg PMU end*/
+
+static ssize_t cstate_get_attr_cpumask(struct device *dev,
+                                      struct device_attribute *attr,
+                                      char *buf)
+{
+       struct pmu *pmu = dev_get_drvdata(dev);
+
+       if (pmu == &cstate_core_pmu)
+               return cpumap_print_to_pagebuf(true, buf, &cstate_core_cpu_mask);
+       else if (pmu == &cstate_pkg_pmu)
+               return cpumap_print_to_pagebuf(true, buf, &cstate_pkg_cpu_mask);
+       else
+               return 0;
+}
+
+static int cstate_pmu_event_init(struct perf_event *event)
+{
+       u64 cfg = event->attr.config;
+       int ret = 0;
+
+       if (event->attr.type != event->pmu->type)
+               return -ENOENT;
+
+       /* unsupported modes and filters */
+       if (event->attr.exclude_user   ||
+           event->attr.exclude_kernel ||
+           event->attr.exclude_hv     ||
+           event->attr.exclude_idle   ||
+           event->attr.exclude_host   ||
+           event->attr.exclude_guest  ||
+           event->attr.sample_period) /* no sampling */
+               return -EINVAL;
+
+       if (event->pmu == &cstate_core_pmu) {
+               if (cfg >= PERF_CSTATE_CORE_EVENT_MAX)
+                       return -EINVAL;
+               if (!core_msr[cfg].attr)
+                       return -EINVAL;
+               event->hw.event_base = core_msr[cfg].msr;
+       } else if (event->pmu == &cstate_pkg_pmu) {
+               if (cfg >= PERF_CSTATE_PKG_EVENT_MAX)
+                       return -EINVAL;
+               if (!pkg_msr[cfg].attr)
+                       return -EINVAL;
+               event->hw.event_base = pkg_msr[cfg].msr;
+       } else
+               return -ENOENT;
+
+       /* must be done before validate_group */
+       event->hw.config = cfg;
+       event->hw.idx = -1;
+
+       return ret;
+}
+
+static inline u64 cstate_pmu_read_counter(struct perf_event *event)
+{
+       u64 val;
+
+       rdmsrl(event->hw.event_base, val);
+       return val;
+}
+
+static void cstate_pmu_event_update(struct perf_event *event)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       u64 prev_raw_count, new_raw_count;
+
+again:
+       prev_raw_count = local64_read(&hwc->prev_count);
+       new_raw_count = cstate_pmu_read_counter(event);
+
+       if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
+                           new_raw_count) != prev_raw_count)
+               goto again;
+
+       local64_add(new_raw_count - prev_raw_count, &event->count);
+}
+
+static void cstate_pmu_event_start(struct perf_event *event, int mode)
+{
+       local64_set(&event->hw.prev_count, cstate_pmu_read_counter(event));
+}
+
+static void cstate_pmu_event_stop(struct perf_event *event, int mode)
+{
+       cstate_pmu_event_update(event);
+}
+
+static void cstate_pmu_event_del(struct perf_event *event, int mode)
+{
+       cstate_pmu_event_stop(event, PERF_EF_UPDATE);
+}
+
+static int cstate_pmu_event_add(struct perf_event *event, int mode)
+{
+       if (mode & PERF_EF_START)
+               cstate_pmu_event_start(event, mode);
+
+       return 0;
+}
+
+static void cstate_cpu_exit(int cpu)
+{
+       int i, id, target;
+
+       /* cpu exit for cstate core */
+       if (has_cstate_core) {
+               id = topology_core_id(cpu);
+               target = -1;
+
+               for_each_online_cpu(i) {
+                       if (i == cpu)
+                               continue;
+                       if (id == topology_core_id(i)) {
+                               target = i;
+                               break;
+                       }
+               }
+               if (cpumask_test_and_clear_cpu(cpu, &cstate_core_cpu_mask) && target >= 0)
+                       cpumask_set_cpu(target, &cstate_core_cpu_mask);
+               WARN_ON(cpumask_empty(&cstate_core_cpu_mask));
+               if (target >= 0)
+                       perf_pmu_migrate_context(&cstate_core_pmu, cpu, target);
+       }
+
+       /* cpu exit for cstate pkg */
+       if (has_cstate_pkg) {
+               id = topology_physical_package_id(cpu);
+               target = -1;
+
+               for_each_online_cpu(i) {
+                       if (i == cpu)
+                               continue;
+                       if (id == topology_physical_package_id(i)) {
+                               target = i;
+                               break;
+                       }
+               }
+               if (cpumask_test_and_clear_cpu(cpu, &cstate_pkg_cpu_mask) && target >= 0)
+                       cpumask_set_cpu(target, &cstate_pkg_cpu_mask);
+               WARN_ON(cpumask_empty(&cstate_pkg_cpu_mask));
+               if (target >= 0)
+                       perf_pmu_migrate_context(&cstate_pkg_pmu, cpu, target);
+       }
+}
+
+static void cstate_cpu_init(int cpu)
+{
+       int i, id;
+
+       /* cpu init for cstate core */
+       if (has_cstate_core) {
+               id = topology_core_id(cpu);
+               for_each_cpu(i, &cstate_core_cpu_mask) {
+                       if (id == topology_core_id(i))
+                               break;
+               }
+               if (i >= nr_cpu_ids)
+                       cpumask_set_cpu(cpu, &cstate_core_cpu_mask);
+       }
+
+       /* cpu init for cstate pkg */
+       if (has_cstate_pkg) {
+               id = topology_physical_package_id(cpu);
+               for_each_cpu(i, &cstate_pkg_cpu_mask) {
+                       if (id == topology_physical_package_id(i))
+                               break;
+               }
+               if (i >= nr_cpu_ids)
+                       cpumask_set_cpu(cpu, &cstate_pkg_cpu_mask);
+       }
+}
+
+static int cstate_cpu_notifier(struct notifier_block *self,
+                                 unsigned long action, void *hcpu)
+{
+       unsigned int cpu = (long)hcpu;
+
+       switch (action & ~CPU_TASKS_FROZEN) {
+       case CPU_UP_PREPARE:
+               break;
+       case CPU_STARTING:
+               cstate_cpu_init(cpu);
+               break;
+       case CPU_UP_CANCELED:
+       case CPU_DYING:
+               break;
+       case CPU_ONLINE:
+       case CPU_DEAD:
+               break;
+       case CPU_DOWN_PREPARE:
+               cstate_cpu_exit(cpu);
+               break;
+       default:
+               break;
+       }
+
+       return NOTIFY_OK;
+}
+
+/*
+ * Probe the cstate events and insert the available one into sysfs attrs
+ * Return false if there is no available events.
+ */
+static bool cstate_probe_msr(struct perf_cstate_msr *msr,
+                            struct attribute   **events_attrs,
+                            int max_event_nr)
+{
+       int i, j = 0;
+       u64 val;
+
+       /* Probe the cstate events. */
+       for (i = 0; i < max_event_nr; i++) {
+               if (!msr[i].test(i) || rdmsrl_safe(msr[i].msr, &val))
+                       msr[i].attr = NULL;
+       }
+
+       /* List remaining events in the sysfs attrs. */
+       for (i = 0; i < max_event_nr; i++) {
+               if (msr[i].attr)
+                       events_attrs[j++] = &msr[i].attr->attr.attr;
+       }
+       events_attrs[j] = NULL;
+
+       return (j > 0) ? true : false;
+}
+
+static int __init cstate_init(void)
+{
+       /* SLM has different MSR for PKG C6 */
+       switch (boot_cpu_data.x86_model) {
+       case 55:
+       case 76:
+       case 77:
+               pkg_msr[PERF_CSTATE_PKG_C6_RES].msr = MSR_PKG_C7_RESIDENCY;
+       }
+
+       if (cstate_probe_msr(core_msr, core_events_attrs, PERF_CSTATE_CORE_EVENT_MAX))
+               has_cstate_core = true;
+
+       if (cstate_probe_msr(pkg_msr, pkg_events_attrs, PERF_CSTATE_PKG_EVENT_MAX))
+               has_cstate_pkg = true;
+
+       return (has_cstate_core || has_cstate_pkg) ? 0 : -ENODEV;
+}
+
+static void __init cstate_cpumask_init(void)
+{
+       int cpu;
+
+       cpu_notifier_register_begin();
+
+       for_each_online_cpu(cpu)
+               cstate_cpu_init(cpu);
+
+       __perf_cpu_notifier(cstate_cpu_notifier);
+
+       cpu_notifier_register_done();
+}
+
+static struct pmu cstate_core_pmu = {
+       .attr_groups    = core_attr_groups,
+       .name           = "cstate_core",
+       .task_ctx_nr    = perf_invalid_context,
+       .event_init     = cstate_pmu_event_init,
+       .add            = cstate_pmu_event_add, /* must have */
+       .del            = cstate_pmu_event_del, /* must have */
+       .start          = cstate_pmu_event_start,
+       .stop           = cstate_pmu_event_stop,
+       .read           = cstate_pmu_event_update,
+       .capabilities   = PERF_PMU_CAP_NO_INTERRUPT,
+};
+
+static struct pmu cstate_pkg_pmu = {
+       .attr_groups    = pkg_attr_groups,
+       .name           = "cstate_pkg",
+       .task_ctx_nr    = perf_invalid_context,
+       .event_init     = cstate_pmu_event_init,
+       .add            = cstate_pmu_event_add, /* must have */
+       .del            = cstate_pmu_event_del, /* must have */
+       .start          = cstate_pmu_event_start,
+       .stop           = cstate_pmu_event_stop,
+       .read           = cstate_pmu_event_update,
+       .capabilities   = PERF_PMU_CAP_NO_INTERRUPT,
+};
+
+static void __init cstate_pmus_register(void)
+{
+       int err;
+
+       if (has_cstate_core) {
+               err = perf_pmu_register(&cstate_core_pmu, cstate_core_pmu.name, -1);
+               if (WARN_ON(err))
+                       pr_info("Failed to register PMU %s error %d\n",
+                               cstate_core_pmu.name, err);
+       }
+
+       if (has_cstate_pkg) {
+               err = perf_pmu_register(&cstate_pkg_pmu, cstate_pkg_pmu.name, -1);
+               if (WARN_ON(err))
+                       pr_info("Failed to register PMU %s error %d\n",
+                               cstate_pkg_pmu.name, err);
+       }
+}
+
+static int __init cstate_pmu_init(void)
+{
+       int err;
+
+       if (cpu_has_hypervisor)
+               return -ENODEV;
+
+       err = cstate_init();
+       if (err)
+               return err;
+
+       cstate_cpumask_init();
+
+       cstate_pmus_register();
+
+       return 0;
+}
+
+device_initcall(cstate_pmu_init);