powerpc/perf: Power8 PMU support
authorMichael Ellerman <michael@ellerman.id.au>
Thu, 25 Apr 2013 19:28:28 +0000 (19:28 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Fri, 26 Apr 2013 06:11:10 +0000 (16:11 +1000)
This patch adds support for the power8 PMU to perf.

Work is ongoing to add generic cache events.

Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Acked-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/perf/Makefile
arch/powerpc/perf/power8-pmu.c [new file with mode: 0644]

index af3fac23768c700791a92bfa87474f81894a57b1..472db183a6dd2f34e85b2a36c54273007a389cce 100644 (file)
@@ -4,7 +4,8 @@ obj-$(CONFIG_PERF_EVENTS)       += callchain.o
 
 obj-$(CONFIG_PPC_PERF_CTRS)    += core-book3s.o
 obj64-$(CONFIG_PPC_PERF_CTRS)  += power4-pmu.o ppc970-pmu.o power5-pmu.o \
-                                  power5+-pmu.o power6-pmu.o power7-pmu.o
+                                  power5+-pmu.o power6-pmu.o power7-pmu.o \
+                                  power8-pmu.o
 obj32-$(CONFIG_PPC_PERF_CTRS)  += mpc7450-pmu.o
 
 obj-$(CONFIG_FSL_EMB_PERF_EVENT) += core-fsl-emb.o
diff --git a/arch/powerpc/perf/power8-pmu.c b/arch/powerpc/perf/power8-pmu.c
new file mode 100644 (file)
index 0000000..c6aa713
--- /dev/null
@@ -0,0 +1,537 @@
+/*
+ * Performance counter support for POWER8 processors.
+ *
+ * Copyright 2009 Paul Mackerras, IBM Corporation.
+ * Copyright 2013 Michael Ellerman, IBM Corporation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#include <linux/kernel.h>
+#include <linux/perf_event.h>
+#include <asm/firmware.h>
+
+
+/*
+ * Some power8 event codes.
+ */
+#define PM_CYC                         0x0001e
+#define PM_GCT_NOSLOT_CYC              0x100f8
+#define PM_CMPLU_STALL                 0x4000a
+#define PM_INST_CMPL                   0x00002
+#define PM_BRU_FIN                     0x10068
+#define PM_BR_MPRED_CMPL               0x400f6
+
+
+/*
+ * Raw event encoding for POWER8:
+ *
+ *        60        56        52        48        44        40        36        32
+ * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
+ *                                     [      thresh_cmp     ]   [  thresh_ctl   ]
+ *                                                                       |
+ *                                       thresh start/stop OR FAB match -*
+ *
+ *        28        24        20        16        12         8         4         0
+ * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
+ *   [   ] [  sample ]   [cache]   [ pmc ]   [unit ]   c     m   [    pmcxsel    ]
+ *     |        |           |                          |     |
+ *     |        |           |                          |     *- mark
+ *     |        |           *- L1/L2/L3 cache_sel      |
+ *     |        |                                      |
+ *     |        *- sampling mode for marked events     *- combine
+ *     |
+ *     *- thresh_sel
+ *
+ * Below uses IBM bit numbering.
+ *
+ * MMCR1[x:y] = unit    (PMCxUNIT)
+ * MMCR1[x]   = combine (PMCxCOMB)
+ *
+ * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011
+ *     # PM_MRK_FAB_RSP_MATCH
+ *     MMCR1[20:27] = thresh_ctl   (FAB_CRESP_MATCH / FAB_TYPE_MATCH)
+ * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001
+ *     # PM_MRK_FAB_RSP_MATCH_CYC
+ *     MMCR1[20:27] = thresh_ctl   (FAB_CRESP_MATCH / FAB_TYPE_MATCH)
+ * else
+ *     MMCRA[48:55] = thresh_ctl   (THRESH START/END)
+ *
+ * if thresh_sel:
+ *     MMCRA[45:47] = thresh_sel
+ *
+ * if thresh_cmp:
+ *     MMCRA[22:24] = thresh_cmp[0:2]
+ *     MMCRA[25:31] = thresh_cmp[3:9]
+ *
+ * if unit == 6 or unit == 7
+ *     MMCRC[53:55] = cache_sel[1:3]      (L2EVENT_SEL)
+ * else if unit == 8 or unit == 9:
+ *     if cache_sel[0] == 0: # L3 bank
+ *             MMCRC[47:49] = cache_sel[1:3]  (L3EVENT_SEL0)
+ *     else if cache_sel[0] == 1:
+ *             MMCRC[50:51] = cache_sel[2:3]  (L3EVENT_SEL1)
+ * else if cache_sel[1]: # L1 event
+ *     MMCR1[16] = cache_sel[2]
+ *    MMCR1[17] = cache_sel[3]
+ *
+ * if mark:
+ *     MMCRA[63]    = 1                (SAMPLE_ENABLE)
+ *     MMCRA[57:59] = sample[0:2]      (RAND_SAMP_ELIG)
+ *    MMCRA[61:62] = sample[3:4]      (RAND_SAMP_MODE)
+ *
+ */
+
+#define EVENT_THR_CMP_SHIFT    40      /* Threshold CMP value */
+#define EVENT_THR_CMP_MASK     0x3ff
+#define EVENT_THR_CTL_SHIFT    32      /* Threshold control value (start/stop) */
+#define EVENT_THR_CTL_MASK     0xffull
+#define EVENT_THR_SEL_SHIFT    29      /* Threshold select value */
+#define EVENT_THR_SEL_MASK     0x7
+#define EVENT_THRESH_SHIFT     29      /* All threshold bits */
+#define EVENT_THRESH_MASK      0x1fffffull
+#define EVENT_SAMPLE_SHIFT     24      /* Sampling mode & eligibility */
+#define EVENT_SAMPLE_MASK      0x1f
+#define EVENT_CACHE_SEL_SHIFT  20      /* L2/L3 cache select */
+#define EVENT_CACHE_SEL_MASK   0xf
+#define EVENT_IS_L1            (4 << EVENT_CACHE_SEL_SHIFT)
+#define EVENT_PMC_SHIFT                16      /* PMC number (1-based) */
+#define EVENT_PMC_MASK         0xf
+#define EVENT_UNIT_SHIFT       12      /* Unit */
+#define EVENT_UNIT_MASK                0xf
+#define EVENT_COMBINE_SHIFT    11      /* Combine bit */
+#define EVENT_COMBINE_MASK     0x1
+#define EVENT_MARKED_SHIFT     8       /* Marked bit */
+#define EVENT_MARKED_MASK      0x1
+#define EVENT_IS_MARKED                (EVENT_MARKED_MASK << EVENT_MARKED_SHIFT)
+#define EVENT_PSEL_MASK                0xff    /* PMCxSEL value */
+
+/*
+ * Layout of constraint bits:
+ *
+ *        60        56        52        48        44        40        36        32
+ * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
+ *   [   fab_match   ]         [       thresh_cmp      ] [   thresh_ctl    ] [   ]
+ *                                                                             |
+ *                                                                 thresh_sel -*
+ *
+ *        28        24        20        16        12         8         4         0
+ * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
+ *                       [ ]   [  sample ]   [     ]   [6] [5]   [4] [3]   [2] [1]
+ *                        |                     |
+ *      L1 I/D qualifier -*                     |      Count of events for each PMC.
+ *                                              |        p1, p2, p3, p4, p5, p6.
+ *                     nc - number of counters -*
+ *
+ * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints
+ * we want the low bit of each field to be added to any existing value.
+ *
+ * Everything else is a value field.
+ */
+
+#define CNST_FAB_MATCH_VAL(v)  (((v) & EVENT_THR_CTL_MASK) << 56)
+#define CNST_FAB_MATCH_MASK    CNST_FAB_MATCH_VAL(EVENT_THR_CTL_MASK)
+
+/* We just throw all the threshold bits into the constraint */
+#define CNST_THRESH_VAL(v)     (((v) & EVENT_THRESH_MASK) << 32)
+#define CNST_THRESH_MASK       CNST_THRESH_VAL(EVENT_THRESH_MASK)
+
+#define CNST_L1_QUAL_VAL(v)    (((v) & 3) << 22)
+#define CNST_L1_QUAL_MASK      CNST_L1_QUAL_VAL(3)
+
+#define CNST_SAMPLE_VAL(v)     (((v) & EVENT_SAMPLE_MASK) << 16)
+#define CNST_SAMPLE_MASK       CNST_SAMPLE_VAL(EVENT_SAMPLE_MASK)
+
+/*
+ * For NC we are counting up to 4 events. This requires three bits, and we need
+ * the fifth event to overflow and set the 4th bit. To achieve that we bias the
+ * fields by 3 in test_adder.
+ */
+#define CNST_NC_SHIFT          12
+#define CNST_NC_VAL            (1 << CNST_NC_SHIFT)
+#define CNST_NC_MASK           (8 << CNST_NC_SHIFT)
+#define POWER8_TEST_ADDER      (3 << CNST_NC_SHIFT)
+
+/*
+ * For the per-PMC fields we have two bits. The low bit is added, so if two
+ * events ask for the same PMC the sum will overflow, setting the high bit,
+ * indicating an error. So our mask sets the high bit.
+ */
+#define CNST_PMC_SHIFT(pmc)    ((pmc - 1) * 2)
+#define CNST_PMC_VAL(pmc)      (1 << CNST_PMC_SHIFT(pmc))
+#define CNST_PMC_MASK(pmc)     (2 << CNST_PMC_SHIFT(pmc))
+
+/* Our add_fields is defined as: */
+#define POWER8_ADD_FIELDS      \
+       CNST_PMC_VAL(1) | CNST_PMC_VAL(2) | CNST_PMC_VAL(3) | \
+       CNST_PMC_VAL(4) | CNST_PMC_VAL(5) | CNST_PMC_VAL(6) | CNST_NC_VAL
+
+
+/* Bits in MMCR1 for POWER8 */
+#define MMCR1_UNIT_SHIFT(pmc)          (60 - (4 * ((pmc) - 1)))
+#define MMCR1_COMBINE_SHIFT(pmc)       (35 - ((pmc) - 1))
+#define MMCR1_PMCSEL_SHIFT(pmc)                (24 - (((pmc) - 1)) * 8)
+#define MMCR1_DC_QUAL_SHIFT            47
+#define MMCR1_IC_QUAL_SHIFT            46
+
+/* Bits in MMCRA for POWER8 */
+#define MMCRA_SAMP_MODE_SHIFT          1
+#define MMCRA_SAMP_ELIG_SHIFT          4
+#define MMCRA_THR_CTL_SHIFT            8
+#define MMCRA_THR_SEL_SHIFT            16
+#define MMCRA_THR_CMP_SHIFT            32
+#define MMCRA_SDAR_MODE_TLB            (1ull << 42)
+
+
+static inline bool event_is_fab_match(u64 event)
+{
+       /* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */
+       event &= 0xff0fe;
+
+       /* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */
+       return (event == 0x30056 || event == 0x4f052);
+}
+
+static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp)
+{
+       unsigned int unit, pmc, cache;
+       unsigned long mask, value;
+
+       mask = value = 0;
+
+       pmc   = (event >> EVENT_PMC_SHIFT)       & EVENT_PMC_MASK;
+       unit  = (event >> EVENT_UNIT_SHIFT)      & EVENT_UNIT_MASK;
+       cache = (event >> EVENT_CACHE_SEL_SHIFT) & EVENT_CACHE_SEL_MASK;
+
+       if (pmc) {
+               if (pmc > 6)
+                       return -1;
+
+               mask  |= CNST_PMC_MASK(pmc);
+               value |= CNST_PMC_VAL(pmc);
+
+               if (pmc >= 5 && event != 0x500fa && event != 0x600f4)
+                       return -1;
+       }
+
+       if (pmc <= 4) {
+               /*
+                * Add to number of counters in use. Note this includes events with
+                * a PMC of 0 - they still need a PMC, it's just assigned later.
+                * Don't count events on PMC 5 & 6, there is only one valid event
+                * on each of those counters, and they are handled above.
+                */
+               mask  |= CNST_NC_MASK;
+               value |= CNST_NC_VAL;
+       }
+
+       if (unit >= 6 && unit <= 9) {
+               /*
+                * L2/L3 events contain a cache selector field, which is
+                * supposed to be programmed into MMCRC. However MMCRC is only
+                * HV writable, and there is no API for guest kernels to modify
+                * it. The solution is for the hypervisor to initialise the
+                * field to zeroes, and for us to only ever allow events that
+                * have a cache selector of zero.
+                */
+               if (cache)
+                       return -1;
+
+       } else if (event & EVENT_IS_L1) {
+               mask  |= CNST_L1_QUAL_MASK;
+               value |= CNST_L1_QUAL_VAL(cache);
+       }
+
+       if (event & EVENT_IS_MARKED) {
+               mask  |= CNST_SAMPLE_MASK;
+               value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT);
+       }
+
+       /*
+        * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
+        * the threshold control bits are used for the match value.
+        */
+       if (event_is_fab_match(event)) {
+               mask  |= CNST_FAB_MATCH_MASK;
+               value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT);
+       } else {
+               /*
+                * Check the mantissa upper two bits are not zero, unless the
+                * exponent is also zero. See the THRESH_CMP_MANTISSA doc.
+                */
+               unsigned int cmp, exp;
+
+               cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
+               exp = cmp >> 7;
+
+               if (exp && (cmp & 0x60) == 0)
+                       return -1;
+
+               mask  |= CNST_THRESH_MASK;
+               value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
+       }
+
+       *maskp = mask;
+       *valp = value;
+
+       return 0;
+}
+
+static int power8_compute_mmcr(u64 event[], int n_ev,
+                              unsigned int hwc[], unsigned long mmcr[])
+{
+       unsigned long mmcra, mmcr1, unit, combine, psel, cache, val;
+       unsigned int pmc, pmc_inuse;
+       int i;
+
+       pmc_inuse = 0;
+
+       /* First pass to count resource use */
+       for (i = 0; i < n_ev; ++i) {
+               pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
+               if (pmc)
+                       pmc_inuse |= 1 << pmc;
+       }
+
+       /* In continous sampling mode, update SDAR on TLB miss */
+       mmcra = MMCRA_SDAR_MODE_TLB;
+       mmcr1 = 0;
+
+       /* Second pass: assign PMCs, set all MMCR1 fields */
+       for (i = 0; i < n_ev; ++i) {
+               pmc     = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
+               unit    = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
+               combine = (event[i] >> EVENT_COMBINE_SHIFT) & EVENT_COMBINE_MASK;
+               psel    =  event[i] & EVENT_PSEL_MASK;
+
+               if (!pmc) {
+                       for (pmc = 1; pmc <= 4; ++pmc) {
+                               if (!(pmc_inuse & (1 << pmc)))
+                                       break;
+                       }
+
+                       pmc_inuse |= 1 << pmc;
+               }
+
+               if (pmc <= 4) {
+                       mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc);
+                       mmcr1 |= combine << MMCR1_COMBINE_SHIFT(pmc);
+                       mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc);
+               }
+
+               if (event[i] & EVENT_IS_L1) {
+                       cache = event[i] >> EVENT_CACHE_SEL_SHIFT;
+                       mmcr1 |= (cache & 1) << MMCR1_IC_QUAL_SHIFT;
+                       cache >>= 1;
+                       mmcr1 |= (cache & 1) << MMCR1_DC_QUAL_SHIFT;
+               }
+
+               if (event[i] & EVENT_IS_MARKED) {
+                       mmcra |= MMCRA_SAMPLE_ENABLE;
+
+                       val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
+                       if (val) {
+                               mmcra |= (val &  3) << MMCRA_SAMP_MODE_SHIFT;
+                               mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT;
+                       }
+               }
+
+               /*
+                * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
+                * the threshold bits are used for the match value.
+                */
+               if (event_is_fab_match(event[i])) {
+                       mmcr1 |= (event[i] >> EVENT_THR_CTL_SHIFT) &
+                                 EVENT_THR_CTL_MASK;
+               } else {
+                       val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
+                       mmcra |= val << MMCRA_THR_CTL_SHIFT;
+                       val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
+                       mmcra |= val << MMCRA_THR_SEL_SHIFT;
+                       val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
+                       mmcra |= val << MMCRA_THR_CMP_SHIFT;
+               }
+
+               hwc[i] = pmc - 1;
+       }
+
+       /* Return MMCRx values */
+       mmcr[0] = 0;
+
+       /* pmc_inuse is 1-based */
+       if (pmc_inuse & 2)
+               mmcr[0] = MMCR0_PMC1CE;
+
+       if (pmc_inuse & 0x7c)
+               mmcr[0] |= MMCR0_PMCjCE;
+
+       mmcr[1] = mmcr1;
+       mmcr[2] = mmcra;
+
+       return 0;
+}
+
+#define MAX_ALT        2
+
+/* Table of alternatives, sorted by column 0 */
+static const unsigned int event_alternatives[][MAX_ALT] = {
+       { 0x10134, 0x301e2 },           /* PM_MRK_ST_CMPL */
+       { 0x10138, 0x40138 },           /* PM_BR_MRK_2PATH */
+       { 0x18082, 0x3e05e },           /* PM_L3_CO_MEPF */
+       { 0x1d14e, 0x401e8 },           /* PM_MRK_DATA_FROM_L2MISS */
+       { 0x1e054, 0x4000a },           /* PM_CMPLU_STALL */
+       { 0x20036, 0x40036 },           /* PM_BR_2PATH */
+       { 0x200f2, 0x300f2 },           /* PM_INST_DISP */
+       { 0x200f4, 0x600f4 },           /* PM_RUN_CYC */
+       { 0x2013c, 0x3012e },           /* PM_MRK_FILT_MATCH */
+       { 0x3e054, 0x400f0 },           /* PM_LD_MISS_L1 */
+       { 0x400fa, 0x500fa },           /* PM_RUN_INST_CMPL */
+};
+
+/*
+ * Scan the alternatives table for a match and return the
+ * index into the alternatives table if found, else -1.
+ */
+static int find_alternative(u64 event)
+{
+       int i, j;
+
+       for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
+               if (event < event_alternatives[i][0])
+                       break;
+
+               for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
+                       if (event == event_alternatives[i][j])
+                               return i;
+       }
+
+       return -1;
+}
+
+static int power8_get_alternatives(u64 event, unsigned int flags, u64 alt[])
+{
+       int i, j, num_alt = 0;
+       u64 alt_event;
+
+       alt[num_alt++] = event;
+
+       i = find_alternative(event);
+       if (i >= 0) {
+               /* Filter out the original event, it's already in alt[0] */
+               for (j = 0; j < MAX_ALT; ++j) {
+                       alt_event = event_alternatives[i][j];
+                       if (alt_event && alt_event != event)
+                               alt[num_alt++] = alt_event;
+               }
+       }
+
+       if (flags & PPMU_ONLY_COUNT_RUN) {
+               /*
+                * We're only counting in RUN state, so PM_CYC is equivalent to
+                * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL.
+                */
+               j = num_alt;
+               for (i = 0; i < num_alt; ++i) {
+                       switch (alt[i]) {
+                       case 0x1e:      /* PM_CYC */
+                               alt[j++] = 0x600f4;     /* PM_RUN_CYC */
+                               break;
+                       case 0x600f4:   /* PM_RUN_CYC */
+                               alt[j++] = 0x1e;
+                               break;
+                       case 0x2:       /* PM_PPC_CMPL */
+                               alt[j++] = 0x500fa;     /* PM_RUN_INST_CMPL */
+                               break;
+                       case 0x500fa:   /* PM_RUN_INST_CMPL */
+                               alt[j++] = 0x2; /* PM_PPC_CMPL */
+                               break;
+                       }
+               }
+               num_alt = j;
+       }
+
+       return num_alt;
+}
+
+static void power8_disable_pmc(unsigned int pmc, unsigned long mmcr[])
+{
+       if (pmc <= 3)
+               mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1));
+}
+
+PMU_FORMAT_ATTR(event,         "config:0-49");
+PMU_FORMAT_ATTR(pmcxsel,       "config:0-7");
+PMU_FORMAT_ATTR(mark,          "config:8");
+PMU_FORMAT_ATTR(combine,       "config:11");
+PMU_FORMAT_ATTR(unit,          "config:12-15");
+PMU_FORMAT_ATTR(pmc,           "config:16-19");
+PMU_FORMAT_ATTR(cache_sel,     "config:20-23");
+PMU_FORMAT_ATTR(sample_mode,   "config:24-28");
+PMU_FORMAT_ATTR(thresh_sel,    "config:29-31");
+PMU_FORMAT_ATTR(thresh_stop,   "config:32-35");
+PMU_FORMAT_ATTR(thresh_start,  "config:36-39");
+PMU_FORMAT_ATTR(thresh_cmp,    "config:40-49");
+
+static struct attribute *power8_pmu_format_attr[] = {
+       &format_attr_event.attr,
+       &format_attr_pmcxsel.attr,
+       &format_attr_mark.attr,
+       &format_attr_combine.attr,
+       &format_attr_unit.attr,
+       &format_attr_pmc.attr,
+       &format_attr_cache_sel.attr,
+       &format_attr_sample_mode.attr,
+       &format_attr_thresh_sel.attr,
+       &format_attr_thresh_stop.attr,
+       &format_attr_thresh_start.attr,
+       &format_attr_thresh_cmp.attr,
+       NULL,
+};
+
+struct attribute_group power8_pmu_format_group = {
+       .name = "format",
+       .attrs = power8_pmu_format_attr,
+};
+
+static const struct attribute_group *power8_pmu_attr_groups[] = {
+       &power8_pmu_format_group,
+       NULL,
+};
+
+static int power8_generic_events[] = {
+       [PERF_COUNT_HW_CPU_CYCLES] =                    PM_CYC,
+       [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =       PM_GCT_NOSLOT_CYC,
+       [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =        PM_CMPLU_STALL,
+       [PERF_COUNT_HW_INSTRUCTIONS] =                  PM_INST_CMPL,
+       [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =           PM_BRU_FIN,
+       [PERF_COUNT_HW_BRANCH_MISSES] =                 PM_BR_MPRED_CMPL,
+};
+
+static struct power_pmu power8_pmu = {
+       .name                   = "POWER8",
+       .n_counter              = 6,
+       .max_alternatives       = MAX_ALT + 1,
+       .add_fields             = POWER8_ADD_FIELDS,
+       .test_adder             = POWER8_TEST_ADDER,
+       .compute_mmcr           = power8_compute_mmcr,
+       .get_constraint         = power8_get_constraint,
+       .get_alternatives       = power8_get_alternatives,
+       .disable_pmc            = power8_disable_pmc,
+       .flags                  = PPMU_HAS_SSLOT | PPMU_HAS_SIER,
+       .n_generic              = ARRAY_SIZE(power8_generic_events),
+       .generic_events         = power8_generic_events,
+       .attr_groups            = power8_pmu_attr_groups,
+};
+
+static int __init init_power8_pmu(void)
+{
+       if (!cur_cpu_spec->oprofile_cpu_type ||
+           strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power8"))
+               return -ENODEV;
+
+       return register_power_pmu(&power8_pmu);
+}
+early_initcall(init_power8_pmu);