Quelle  generic-compat-pmu.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0+
//
// Copyright 2019 Madhavan Srinivasan, IBM Corporation.

#define pr_fmt(fmt) "generic-compat-pmu: " fmt

#include "isa207-common.h"

/*
 * Raw event encoding:
 *
 *        60        56        52        48        44        40        36        32
 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
 *
 *        28        24        20        16        12         8         4         0
 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
 *                                 [ pmc ]                       [    pmcxsel    ]
 */

/*
 * Event codes defined in ISA v3.0B
 */
#define EVENT(_name, _code) _name = _code,

enum {
 /* Cycles, alternate code */
 EVENT(PM_CYC_ALT,   0x100f0)
 /* One or more instructions completed in a cycle */
 EVENT(PM_CYC_INST_CMPL,   0x100f2)
 /* Floating-point instruction completed */
 EVENT(PM_FLOP_CMPL,   0x100f4)
 /* Instruction ERAT/L1-TLB miss */
 EVENT(PM_L1_ITLB_MISS,   0x100f6)
 /* All instructions completed and none available */
 EVENT(PM_NO_INST_AVAIL,   0x100f8)
 /* A load-type instruction completed (ISA v3.0+) */
 EVENT(PM_LD_CMPL,   0x100fc)
 /* Instruction completed, alternate code (ISA v3.0+) */
 EVENT(PM_INST_CMPL_ALT,   0x100fe)
 /* A store-type instruction completed */
 EVENT(PM_ST_CMPL,   0x200f0)
 /* Instruction Dispatched */
 EVENT(PM_INST_DISP,   0x200f2)
 /* Run_cycles */
 EVENT(PM_RUN_CYC,   0x200f4)
 /* Data ERAT/L1-TLB miss/reload */
 EVENT(PM_L1_DTLB_RELOAD,  0x200f6)
 /* Taken branch completed */
 EVENT(PM_BR_TAKEN_CMPL,   0x200fa)
 /* Demand iCache Miss */
 EVENT(PM_L1_ICACHE_MISS,  0x200fc)
 /* L1 Dcache reload from memory */
 EVENT(PM_L1_RELOAD_FROM_MEM,  0x200fe)
 /* L1 Dcache store miss */
 EVENT(PM_ST_MISS_L1,   0x300f0)
 /* Alternate code for PM_INST_DISP */
 EVENT(PM_INST_DISP_ALT,   0x300f2)
 /* Branch direction or target mispredicted */
 EVENT(PM_BR_MISPREDICT,   0x300f6)
 /* Data TLB miss/reload */
 EVENT(PM_DTLB_MISS,   0x300fc)
 /* Demand LD - L3 Miss (not L2 hit and not L3 hit) */
 EVENT(PM_DATA_FROM_L3MISS,  0x300fe)
 /* L1 Dcache load miss */
 EVENT(PM_LD_MISS_L1,   0x400f0)
 /* Cycle when instruction(s) dispatched */
 EVENT(PM_CYC_INST_DISP,   0x400f2)
 /* Branch or branch target mispredicted */
 EVENT(PM_BR_MPRED_CMPL,   0x400f6)
 /* Instructions completed with run latch set */
 EVENT(PM_RUN_INST_CMPL,   0x400fa)
 /* Instruction TLB miss/reload */
 EVENT(PM_ITLB_MISS,   0x400fc)
 /* Load data not cached */
 EVENT(PM_LD_NOT_CACHED,   0x400fe)
 /* Instructions */
 EVENT(PM_INST_CMPL,   0x500fa)
 /* Cycles */
 EVENT(PM_CYC,    0x600f4)
};

#undef EVENT

/* Table of alternatives, sorted in increasing order of column 0 */
/* Note that in each row, column 0 must be the smallest */
static const unsigned int generic_event_alternatives[][MAX_ALT] = {
 { PM_CYC_ALT,   PM_CYC },
 { PM_INST_CMPL_ALT,  PM_INST_CMPL },
 { PM_INST_DISP,   PM_INST_DISP_ALT },
};

static int generic_get_alternatives(u64 event, unsigned int flags, u64 alt[])
{
 int num_alt = 0;

 num_alt = isa207_get_alternatives(event, alt,
       ARRAY_SIZE(generic_event_alternatives), flags,
       generic_event_alternatives);

 return num_alt;
}

GENERIC_EVENT_ATTR(cpu-cycles,   PM_CYC);
GENERIC_EVENT_ATTR(instructions,  PM_INST_CMPL);
GENERIC_EVENT_ATTR(stalled-cycles-frontend, PM_NO_INST_AVAIL);
GENERIC_EVENT_ATTR(branch-misses,  PM_BR_MPRED_CMPL);
GENERIC_EVENT_ATTR(cache-misses,  PM_LD_MISS_L1);

CACHE_EVENT_ATTR(L1-dcache-load-misses,  PM_LD_MISS_L1);
CACHE_EVENT_ATTR(L1-dcache-store-misses, PM_ST_MISS_L1);
CACHE_EVENT_ATTR(L1-icache-load-misses,  PM_L1_ICACHE_MISS);
CACHE_EVENT_ATTR(LLC-load-misses,  PM_DATA_FROM_L3MISS);
CACHE_EVENT_ATTR(branch-load-misses,  PM_BR_MPRED_CMPL);
CACHE_EVENT_ATTR(dTLB-load-misses,  PM_DTLB_MISS);
CACHE_EVENT_ATTR(iTLB-load-misses,  PM_ITLB_MISS);

static struct attribute *generic_compat_events_attr[] = {
 GENERIC_EVENT_PTR(PM_CYC),
 GENERIC_EVENT_PTR(PM_INST_CMPL),
 GENERIC_EVENT_PTR(PM_NO_INST_AVAIL),
 GENERIC_EVENT_PTR(PM_BR_MPRED_CMPL),
 GENERIC_EVENT_PTR(PM_LD_MISS_L1),
 CACHE_EVENT_PTR(PM_LD_MISS_L1),
 CACHE_EVENT_PTR(PM_ST_MISS_L1),
 CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
 CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
 CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
 CACHE_EVENT_PTR(PM_DTLB_MISS),
 CACHE_EVENT_PTR(PM_ITLB_MISS),
 NULL
};

static const struct attribute_group generic_compat_pmu_events_group = {
 .name = "events",
 .attrs = generic_compat_events_attr,
};

PMU_FORMAT_ATTR(event,  "config:0-19");
PMU_FORMAT_ATTR(pmcxsel, "config:0-7");
PMU_FORMAT_ATTR(pmc,  "config:16-19");

static struct attribute *generic_compat_pmu_format_attr[] = {
 &format_attr_event.attr,
 &format_attr_pmcxsel.attr,
 &format_attr_pmc.attr,
 NULL,
};

static const struct attribute_group generic_compat_pmu_format_group = {
 .name = "format",
 .attrs = generic_compat_pmu_format_attr,
};

static struct attribute *generic_compat_pmu_caps_attrs[] = {
 NULL
};

static struct attribute_group generic_compat_pmu_caps_group = {
 .name  = "caps",
 .attrs = generic_compat_pmu_caps_attrs,
};

static const struct attribute_group *generic_compat_pmu_attr_groups[] = {
 &generic_compat_pmu_format_group,
 &generic_compat_pmu_events_group,
 &generic_compat_pmu_caps_group,
 NULL,
};

static int compat_generic_events[] = {
 [PERF_COUNT_HW_CPU_CYCLES] =   PM_CYC,
 [PERF_COUNT_HW_INSTRUCTIONS] =   PM_INST_CMPL,
 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = PM_NO_INST_AVAIL,
 [PERF_COUNT_HW_BRANCH_MISSES] =   PM_BR_MPRED_CMPL,
 [PERF_COUNT_HW_CACHE_MISSES] =   PM_LD_MISS_L1,
};

#define C(x) PERF_COUNT_HW_CACHE_##x

/*
 * Table of generalized cache-related events.
 * 0 means not supported, -1 means nonsensical, other values
 * are event codes.
 */
static u64 generic_compat_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 [ C(L1D) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_LD_MISS_L1,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_ST_MISS_L1,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = 0,
  },
 },
 [ C(L1I) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_L1_ICACHE_MISS,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = -1,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = 0,
  },
 },
 [ C(LL) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_DATA_FROM_L3MISS,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = 0,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = 0,
  },
 },
 [ C(DTLB) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_DTLB_MISS,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
 },
 [ C(ITLB) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_ITLB_MISS,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
 },
 [ C(BPU) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = 0,
   [ C(RESULT_MISS)   ] = PM_BR_MPRED_CMPL,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
 },
 [ C(NODE) ] = {
  [ C(OP_READ) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
  [ C(OP_WRITE) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
  [ C(OP_PREFETCH) ] = {
   [ C(RESULT_ACCESS) ] = -1,
   [ C(RESULT_MISS)   ] = -1,
  },
 },
};

#undef C

/*
 * We set MMCR0[CC5-6RUN] so we can use counters 5 and 6 for
 * PM_INST_CMPL and PM_CYC.
 */
static int generic_compute_mmcr(u64 event[], int n_ev,
    unsigned int hwc[], struct mmcr_regs *mmcr,
    struct perf_event *pevents[], u32 flags)
{
 int ret;

 ret = isa207_compute_mmcr(event, n_ev, hwc, mmcr, pevents, flags);
 if (!ret)
  mmcr->mmcr0 |= MMCR0_C56RUN;
 return ret;
}

static struct power_pmu generic_compat_pmu = {
 .name   = "ISAv3",
 .n_counter  = MAX_PMU_COUNTERS,
 .add_fields  = ISA207_ADD_FIELDS,
 .test_adder  = ISA207_TEST_ADDER,
 .compute_mmcr  = generic_compute_mmcr,
 .get_constraint  = isa207_get_constraint,
 .get_alternatives = generic_get_alternatives,
 .disable_pmc  = isa207_disable_pmc,
 .flags   = PPMU_HAS_SIER | PPMU_ARCH_207S,
 .n_generic  = ARRAY_SIZE(compat_generic_events),
 .generic_events  = compat_generic_events,
 .cache_events  = &generic_compat_cache_events,
 .attr_groups  = generic_compat_pmu_attr_groups,
};

int __init init_generic_compat_pmu(void)
{
 int rc = 0;

 /*
  * From ISA v2.07 on, PMU features are architected;
  * we require >= v3.0 because (a) that has PM_LD_CMPL and
  * PM_INST_CMPL_ALT, which v2.07 doesn't have, and
  * (b) we don't expect any non-IBM Power ISA
  * implementations that conform to v2.07 but not v3.0.
 */
 if (!cpu_has_feature(CPU_FTR_ARCH_300))
  return -ENODEV;

 rc = register_power_pmu(&generic_compat_pmu);
 if (rc)
  return rc;

 /* Tell userspace that EBB is supported */
 cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;

 return 0;
}