perf/x86: Add an MSR PMU driver
authorAndy Lutomirski <luto@kernel.org>
Mon, 20 Jul 2015 15:49:06 +0000 (11:49 -0400)
committerIngo Molnar <mingo@kernel.org>
Tue, 4 Aug 2015 08:17:00 +0000 (10:17 +0200)
This patch adds an MSR PMU to support free running MSR counters. Such
as time and freq related counters includes TSC, IA32_APERF, IA32_MPERF
and IA32_PPERF, but also SMI_COUNT.

The events are exposed in sysfs for use by perf stat and other tools.
The files are under /sys/devices/msr/events/

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Signed-off-by: Kan Liang <kan.liang@intel.com>
[ s/freq/msr/, added SMI_COUNT, fixed bugs. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@kernel.org
Cc: adrian.hunter@intel.com
Cc: dsahern@gmail.com
Cc: eranian@google.com
Cc: jolsa@kernel.org
Cc: mark.rutland@arm.com
Cc: namhyung@kernel.org
Link: http://lkml.kernel.org/r/1437407346-31186-1-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_msr.c [new file with mode: 0644]

index 9bff68798836232d51544e21ee3e15409171db11..4eb065c6bed26b38a2804ffcab57f9e09274a7ec 100644 (file)
@@ -46,6 +46,8 @@ obj-$(CONFIG_PERF_EVENTS_INTEL_UNCORE)        += perf_event_intel_uncore.o \
                                           perf_event_intel_uncore_snb.o \
                                           perf_event_intel_uncore_snbep.o \
                                           perf_event_intel_uncore_nhmex.o
+obj-$(CONFIG_CPU_SUP_INTEL)            += perf_event_msr.o
+obj-$(CONFIG_CPU_SUP_AMD)              += perf_event_msr.o
 endif
 
 
diff --git a/arch/x86/kernel/cpu/perf_event_msr.c b/arch/x86/kernel/cpu/perf_event_msr.c
new file mode 100644 (file)
index 0000000..af216e9
--- /dev/null
@@ -0,0 +1,242 @@
+#include <linux/perf_event.h>
+
+enum perf_msr_id {
+       PERF_MSR_TSC                    = 0,
+       PERF_MSR_APERF                  = 1,
+       PERF_MSR_MPERF                  = 2,
+       PERF_MSR_PPERF                  = 3,
+       PERF_MSR_SMI                    = 4,
+
+       PERF_MSR_EVENT_MAX,
+};
+
+struct perf_msr {
+       int     id;
+       u64     msr;
+};
+
+static struct perf_msr msr[] = {
+       { PERF_MSR_TSC, 0 },
+       { PERF_MSR_APERF, MSR_IA32_APERF },
+       { PERF_MSR_MPERF, MSR_IA32_MPERF },
+       { PERF_MSR_PPERF, MSR_PPERF },
+       { PERF_MSR_SMI, MSR_SMI_COUNT },
+};
+
+PMU_EVENT_ATTR_STRING(tsc,   evattr_tsc,   "event=0x00");
+PMU_EVENT_ATTR_STRING(aperf, evattr_aperf, "event=0x01");
+PMU_EVENT_ATTR_STRING(mperf, evattr_mperf, "event=0x02");
+PMU_EVENT_ATTR_STRING(pperf, evattr_pperf, "event=0x03");
+PMU_EVENT_ATTR_STRING(smi,   evattr_smi,   "event=0x04");
+
+static struct attribute *events_attrs[PERF_MSR_EVENT_MAX + 1] = {
+       &evattr_tsc.attr.attr,
+};
+
+static struct attribute_group events_attr_group = {
+       .name = "events",
+       .attrs = events_attrs,
+};
+
+PMU_FORMAT_ATTR(event, "config:0-63");
+static struct attribute *format_attrs[] = {
+       &format_attr_event.attr,
+       NULL,
+};
+static struct attribute_group format_attr_group = {
+       .name = "format",
+       .attrs = format_attrs,
+};
+
+static const struct attribute_group *attr_groups[] = {
+       &events_attr_group,
+       &format_attr_group,
+       NULL,
+};
+
+static int msr_event_init(struct perf_event *event)
+{
+       u64 cfg = event->attr.config;
+
+       if (event->attr.type != event->pmu->type)
+               return -ENOENT;
+
+       if (cfg >= PERF_MSR_EVENT_MAX)
+               return -EINVAL;
+
+       /* 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;
+
+       event->hw.idx = -1;
+       event->hw.event_base = msr[cfg].msr;
+       event->hw.config = cfg;
+
+       return 0;
+}
+
+static inline u64 msr_read_counter(struct perf_event *event)
+{
+       u64 now;
+
+       if (event->hw.event_base)
+               rdmsrl(event->hw.event_base, now);
+       else
+               now = rdtsc();
+
+       return now;
+}
+static void msr_event_update(struct perf_event *event)
+{
+       u64 prev, now;
+       s64 delta;
+
+       /* Careful, an NMI might modify the previous event value. */
+again:
+       prev = local64_read(&event->hw.prev_count);
+       now = msr_read_counter(event);
+
+       if (local64_cmpxchg(&event->hw.prev_count, prev, now) != prev)
+               goto again;
+
+       delta = now - prev;
+       if (unlikely(event->hw.event_base == MSR_SMI_COUNT)) {
+               delta <<= 32;
+               delta >>= 32; /* sign extend */
+       }
+       local64_add(now - prev, &event->count);
+}
+
+static void msr_event_start(struct perf_event *event, int flags)
+{
+       u64 now;
+
+       now = msr_read_counter(event);
+       local64_set(&event->hw.prev_count, now);
+}
+
+static void msr_event_stop(struct perf_event *event, int flags)
+{
+       msr_event_update(event);
+}
+
+static void msr_event_del(struct perf_event *event, int flags)
+{
+       msr_event_stop(event, PERF_EF_UPDATE);
+}
+
+static int msr_event_add(struct perf_event *event, int flags)
+{
+       if (flags & PERF_EF_START)
+               msr_event_start(event, flags);
+
+       return 0;
+}
+
+static struct pmu pmu_msr = {
+       .task_ctx_nr    = perf_sw_context,
+       .attr_groups    = attr_groups,
+       .event_init     = msr_event_init,
+       .add            = msr_event_add,
+       .del            = msr_event_del,
+       .start          = msr_event_start,
+       .stop           = msr_event_stop,
+       .read           = msr_event_update,
+       .capabilities   = PERF_PMU_CAP_NO_INTERRUPT,
+};
+
+static int __init intel_msr_init(int idx)
+{
+       if (boot_cpu_data.x86 != 6)
+               return 0;
+
+       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 */
+
+       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 */
+               events_attrs[idx++] = &evattr_smi.attr.attr;
+               break;
+
+       case 78: /* 14nm Skylake Mobile */
+       case 94: /* 14nm Skylake Desktop */
+               events_attrs[idx++] = &evattr_pperf.attr.attr;
+               events_attrs[idx++] = &evattr_smi.attr.attr;
+               break;
+
+       case 55: /* 22nm Atom "Silvermont"                */
+       case 76: /* 14nm Atom "Airmont"                   */
+       case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */
+               events_attrs[idx++] = &evattr_smi.attr.attr;
+               break;
+       }
+
+       events_attrs[idx] = NULL;
+
+       return 0;
+}
+
+static int __init amd_msr_init(int idx)
+{
+       return 0;
+}
+
+static int __init msr_init(void)
+{
+       int err;
+       int idx = 1;
+
+       if (boot_cpu_has(X86_FEATURE_APERFMPERF)) {
+               events_attrs[idx++] = &evattr_aperf.attr.attr;
+               events_attrs[idx++] = &evattr_mperf.attr.attr;
+               events_attrs[idx] = NULL;
+       }
+
+       switch (boot_cpu_data.x86_vendor) {
+       case X86_VENDOR_INTEL:
+               err = intel_msr_init(idx);
+               break;
+
+       case X86_VENDOR_AMD:
+               err = amd_msr_init(idx);
+               break;
+
+       default:
+               err = -ENOTSUPP;
+       }
+
+       if (err != 0) {
+               pr_cont("no msr PMU driver.\n");
+               return 0;
+       }
+
+       perf_pmu_register(&pmu_msr, "msr", -1);
+
+       return 0;
+}
+device_initcall(msr_init);