arch/x86/events/intel/knc.c

   1 /* Driver for Intel Xeon Phi "Knights Corner" PMU */
   2
   3 #include <linux/perf_event.h>
   4 #include <linux/types.h>
   5
   6 #include <asm/hardirq.h>
   7
   8 #include "../perf_event.h"
   9
  10 static const u64 knc_perfmon_event_map[] =
  11 {
  12   [PERF_COUNT_HW_CPU_CYCLES]            = 0x002a,
  13   [PERF_COUNT_HW_INSTRUCTIONS]          = 0x0016,
  14   [PERF_COUNT_HW_CACHE_REFERENCES]      = 0x0028,
  15   [PERF_COUNT_HW_CACHE_MISSES]          = 0x0029,
  16   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]   = 0x0012,
  17   [PERF_COUNT_HW_BRANCH_MISSES]         = 0x002b,
  18 };
  19
  20 static const u64 __initconst knc_hw_cache_event_ids
  21                                 [PERF_COUNT_HW_CACHE_MAX]
  22                                 [PERF_COUNT_HW_CACHE_OP_MAX]
  23                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
  24 {
  25  [ C(L1D) ] = {
  26         [ C(OP_READ) ] = {
  27                 /* On Xeon Phi event "0" is a valid DATA_READ          */
  28                 /*   (L1 Data Cache Reads) Instruction.                */
  29                 /* We code this as ARCH_PERFMON_EVENTSEL_INT as this   */
  30                 /* bit will always be set in x86_pmu_hw_config().      */
  31                 [ C(RESULT_ACCESS) ] = ARCH_PERFMON_EVENTSEL_INT,
  32                                                 /* DATA_READ           */
  33                 [ C(RESULT_MISS)   ] = 0x0003,  /* DATA_READ_MISS      */
  34         },
  35         [ C(OP_WRITE) ] = {
  36                 [ C(RESULT_ACCESS) ] = 0x0001,  /* DATA_WRITE          */
  37                 [ C(RESULT_MISS)   ] = 0x0004,  /* DATA_WRITE_MISS     */
  38         },
  39         [ C(OP_PREFETCH) ] = {
  40                 [ C(RESULT_ACCESS) ] = 0x0011,  /* L1_DATA_PF1         */
  41                 [ C(RESULT_MISS)   ] = 0x001c,  /* L1_DATA_PF1_MISS    */
  42         },
  43  },
  44  [ C(L1I ) ] = {
  45         [ C(OP_READ) ] = {
  46                 [ C(RESULT_ACCESS) ] = 0x000c,  /* CODE_READ          */
  47                 [ C(RESULT_MISS)   ] = 0x000e,  /* CODE_CACHE_MISS    */
  48         },
  49         [ C(OP_WRITE) ] = {
  50                 [ C(RESULT_ACCESS) ] = -1,
  51                 [ C(RESULT_MISS)   ] = -1,
  52         },
  53         [ C(OP_PREFETCH) ] = {
  54                 [ C(RESULT_ACCESS) ] = 0x0,
  55                 [ C(RESULT_MISS)   ] = 0x0,
  56         },
  57  },
  58  [ C(LL  ) ] = {
  59         [ C(OP_READ) ] = {
  60                 [ C(RESULT_ACCESS) ] = 0,
  61                 [ C(RESULT_MISS)   ] = 0x10cb,  /* L2_READ_MISS */
  62         },
  63         [ C(OP_WRITE) ] = {
  64                 [ C(RESULT_ACCESS) ] = 0x10cc,  /* L2_WRITE_HIT */
  65                 [ C(RESULT_MISS)   ] = 0,
  66         },
  67         [ C(OP_PREFETCH) ] = {
  68                 [ C(RESULT_ACCESS) ] = 0x10fc,  /* L2_DATA_PF2      */
  69                 [ C(RESULT_MISS)   ] = 0x10fe,  /* L2_DATA_PF2_MISS */
  70         },
  71  },
  72  [ C(DTLB) ] = {
  73         [ C(OP_READ) ] = {
  74                 [ C(RESULT_ACCESS) ] = ARCH_PERFMON_EVENTSEL_INT,
  75                                                 /* DATA_READ */
  76                                                 /* see note on L1 OP_READ */
  77                 [ C(RESULT_MISS)   ] = 0x0002,  /* DATA_PAGE_WALK */
  78         },
  79         [ C(OP_WRITE) ] = {
  80                 [ C(RESULT_ACCESS) ] = 0x0001,  /* DATA_WRITE */
  81                 [ C(RESULT_MISS)   ] = 0x0002,  /* DATA_PAGE_WALK */
  82         },
  83         [ C(OP_PREFETCH) ] = {
  84                 [ C(RESULT_ACCESS) ] = 0x0,
  85                 [ C(RESULT_MISS)   ] = 0x0,
  86         },
  87  },
  88  [ C(ITLB) ] = {
  89         [ C(OP_READ) ] = {
  90                 [ C(RESULT_ACCESS) ] = 0x000c,  /* CODE_READ */
  91                 [ C(RESULT_MISS)   ] = 0x000d,  /* CODE_PAGE_WALK */
  92         },
  93         [ C(OP_WRITE) ] = {
  94                 [ C(RESULT_ACCESS) ] = -1,
  95                 [ C(RESULT_MISS)   ] = -1,
  96         },
  97         [ C(OP_PREFETCH) ] = {
  98                 [ C(RESULT_ACCESS) ] = -1,
  99                 [ C(RESULT_MISS)   ] = -1,
 100         },
 101  },
 102  [ C(BPU ) ] = {
 103         [ C(OP_READ) ] = {
 104                 [ C(RESULT_ACCESS) ] = 0x0012,  /* BRANCHES */
 105                 [ C(RESULT_MISS)   ] = 0x002b,  /* BRANCHES_MISPREDICTED */
 106         },
 107         [ C(OP_WRITE) ] = {
 108                 [ C(RESULT_ACCESS) ] = -1,
 109                 [ C(RESULT_MISS)   ] = -1,
 110         },
 111         [ C(OP_PREFETCH) ] = {
 112                 [ C(RESULT_ACCESS) ] = -1,
 113                 [ C(RESULT_MISS)   ] = -1,
 114         },
 115  },
 116 };
 117
 118
 119 static u64 knc_pmu_event_map(int hw_event)
 120 {
 121         return knc_perfmon_event_map[hw_event];
 122 }
 123
 124 static struct event_constraint knc_event_constraints[] =
 125 {
 126         INTEL_EVENT_CONSTRAINT(0xc3, 0x1),      /* HWP_L2HIT */
 127         INTEL_EVENT_CONSTRAINT(0xc4, 0x1),      /* HWP_L2MISS */
 128         INTEL_EVENT_CONSTRAINT(0xc8, 0x1),      /* L2_READ_HIT_E */
 129         INTEL_EVENT_CONSTRAINT(0xc9, 0x1),      /* L2_READ_HIT_M */
 130         INTEL_EVENT_CONSTRAINT(0xca, 0x1),      /* L2_READ_HIT_S */
 131         INTEL_EVENT_CONSTRAINT(0xcb, 0x1),      /* L2_READ_MISS */
 132         INTEL_EVENT_CONSTRAINT(0xcc, 0x1),      /* L2_WRITE_HIT */
 133         INTEL_EVENT_CONSTRAINT(0xce, 0x1),      /* L2_STRONGLY_ORDERED_STREAMING_VSTORES_MISS */
 134         INTEL_EVENT_CONSTRAINT(0xcf, 0x1),      /* L2_WEAKLY_ORDERED_STREAMING_VSTORE_MISS */
 135         INTEL_EVENT_CONSTRAINT(0xd7, 0x1),      /* L2_VICTIM_REQ_WITH_DATA */
 136         INTEL_EVENT_CONSTRAINT(0xe3, 0x1),      /* SNP_HITM_BUNIT */
 137         INTEL_EVENT_CONSTRAINT(0xe6, 0x1),      /* SNP_HIT_L2 */
 138         INTEL_EVENT_CONSTRAINT(0xe7, 0x1),      /* SNP_HITM_L2 */
 139         INTEL_EVENT_CONSTRAINT(0xf1, 0x1),      /* L2_DATA_READ_MISS_CACHE_FILL */
 140         INTEL_EVENT_CONSTRAINT(0xf2, 0x1),      /* L2_DATA_WRITE_MISS_CACHE_FILL */
 141         INTEL_EVENT_CONSTRAINT(0xf6, 0x1),      /* L2_DATA_READ_MISS_MEM_FILL */
 142         INTEL_EVENT_CONSTRAINT(0xf7, 0x1),      /* L2_DATA_WRITE_MISS_MEM_FILL */
 143         INTEL_EVENT_CONSTRAINT(0xfc, 0x1),      /* L2_DATA_PF2 */
 144         INTEL_EVENT_CONSTRAINT(0xfd, 0x1),      /* L2_DATA_PF2_DROP */
 145         INTEL_EVENT_CONSTRAINT(0xfe, 0x1),      /* L2_DATA_PF2_MISS */
 146         INTEL_EVENT_CONSTRAINT(0xff, 0x1),      /* L2_DATA_HIT_INFLIGHT_PF2 */
 147         EVENT_CONSTRAINT_END
 148 };
 149
 150 #define MSR_KNC_IA32_PERF_GLOBAL_STATUS         0x0000002d
 151 #define MSR_KNC_IA32_PERF_GLOBAL_OVF_CONTROL    0x0000002e
 152 #define MSR_KNC_IA32_PERF_GLOBAL_CTRL           0x0000002f
 153
 154 #define KNC_ENABLE_COUNTER0                     0x00000001
 155 #define KNC_ENABLE_COUNTER1                     0x00000002
 156
 157 static void knc_pmu_disable_all(void)
 158 {
 159         u64 val;
 160
 161         rdmsrl(MSR_KNC_IA32_PERF_GLOBAL_CTRL, val);
 162         val &= ~(KNC_ENABLE_COUNTER0|KNC_ENABLE_COUNTER1);
 163         wrmsrl(MSR_KNC_IA32_PERF_GLOBAL_CTRL, val);
 164 }
 165
 166 static void knc_pmu_enable_all(int added)
 167 {
 168         u64 val;
 169
 170         rdmsrl(MSR_KNC_IA32_PERF_GLOBAL_CTRL, val);
 171         val |= (KNC_ENABLE_COUNTER0|KNC_ENABLE_COUNTER1);
 172         wrmsrl(MSR_KNC_IA32_PERF_GLOBAL_CTRL, val);
 173 }
 174
 175 static inline void
 176 knc_pmu_disable_event(struct perf_event *event)
 177 {
 178         struct hw_perf_event *hwc = &event->hw;
 179         u64 val;
 180
 181         val = hwc->config;
 182         val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
 183
 184         (void)wrmsrl_safe(hwc->config_base + hwc->idx, val);
 185 }
 186
 187 static void knc_pmu_enable_event(struct perf_event *event)
 188 {
 189         struct hw_perf_event *hwc = &event->hw;
 190         u64 val;
 191
 192         val = hwc->config;
 193         val |= ARCH_PERFMON_EVENTSEL_ENABLE;
 194
 195         (void)wrmsrl_safe(hwc->config_base + hwc->idx, val);
 196 }
 197
 198 static inline u64 knc_pmu_get_status(void)
 199 {
 200         u64 status;
 201
 202         rdmsrl(MSR_KNC_IA32_PERF_GLOBAL_STATUS, status);
 203
 204         return status;
 205 }
 206
 207 static inline void knc_pmu_ack_status(u64 ack)
 208 {
 209         wrmsrl(MSR_KNC_IA32_PERF_GLOBAL_OVF_CONTROL, ack);
 210 }
 211
 212 static int knc_pmu_handle_irq(struct pt_regs *regs)
 213 {
 214         struct perf_sample_data data;
 215         struct cpu_hw_events *cpuc;
 216         int handled = 0;
 217         int bit, loops;
 218         u64 status;
 219
 220         cpuc = this_cpu_ptr(&cpu_hw_events);
 221
 222         knc_pmu_disable_all();
 223
 224         status = knc_pmu_get_status();
 225         if (!status) {
 226                 knc_pmu_enable_all(0);
 227                 return handled;
 228         }
 229
 230         loops = 0;
 231 again:
 232         knc_pmu_ack_status(status);
 233         if (++loops > 100) {
 234                 WARN_ONCE(1, "perf: irq loop stuck!\n");
 235                 perf_event_print_debug();
 236                 goto done;
 237         }
 238
 239         inc_irq_stat(apic_perf_irqs);
 240
 241         for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
 242                 struct perf_event *event = cpuc->events[bit];
 243
 244                 handled++;
 245
 246                 if (!test_bit(bit, cpuc->active_mask))
 247                         continue;
 248
 249                 if (!intel_pmu_save_and_restart(event))
 250                         continue;
 251
 252                 perf_sample_data_init(&data, 0, event->hw.last_period);
 253
 254                 if (perf_event_overflow(event, &data, regs))
 255                         x86_pmu_stop(event, 0);
 256         }
 257
 258         /*
 259          * Repeat if there is more work to be done:
 260          */
 261         status = knc_pmu_get_status();
 262         if (status)
 263                 goto again;
 264
 265 done:
 266         /* Only restore PMU state when it's active. See x86_pmu_disable(). */
 267         if (cpuc->enabled)
 268                 knc_pmu_enable_all(0);
 269
 270         return handled;
 271 }
 272
 273
 274 PMU_FORMAT_ATTR(event,  "config:0-7"    );
 275 PMU_FORMAT_ATTR(umask,  "config:8-15"   );
 276 PMU_FORMAT_ATTR(edge,   "config:18"     );
 277 PMU_FORMAT_ATTR(inv,    "config:23"     );
 278 PMU_FORMAT_ATTR(cmask,  "config:24-31"  );
 279
 280 static struct attribute *intel_knc_formats_attr[] = {
 281         &format_attr_event.attr,
 282         &format_attr_umask.attr,
 283         &format_attr_edge.attr,
 284         &format_attr_inv.attr,
 285         &format_attr_cmask.attr,
 286         NULL,
 287 };
 288
 289 static const struct x86_pmu knc_pmu __initconst = {
 290         .name                   = "knc",
 291         .handle_irq             = knc_pmu_handle_irq,
 292         .disable_all            = knc_pmu_disable_all,
 293         .enable_all             = knc_pmu_enable_all,
 294         .enable                 = knc_pmu_enable_event,
 295         .disable                = knc_pmu_disable_event,
 296         .hw_config              = x86_pmu_hw_config,
 297         .schedule_events        = x86_schedule_events,
 298         .eventsel               = MSR_KNC_EVNTSEL0,
 299         .perfctr                = MSR_KNC_PERFCTR0,
 300         .event_map              = knc_pmu_event_map,
 301         .max_events             = ARRAY_SIZE(knc_perfmon_event_map),
 302         .apic                   = 1,
 303         .max_period             = (1ULL << 39) - 1,
 304         .version                = 0,
 305         .num_counters           = 2,
 306         .cntval_bits            = 40,
 307         .cntval_mask            = (1ULL << 40) - 1,
 308         .get_event_constraints  = x86_get_event_constraints,
 309         .event_constraints      = knc_event_constraints,
 310         .format_attrs           = intel_knc_formats_attr,
 311 };
 312
 313 __init int knc_pmu_init(void)
 314 {
 315         x86_pmu = knc_pmu;
 316
 317         memcpy(hw_cache_event_ids, knc_hw_cache_event_ids,
 318                 sizeof(hw_cache_event_ids));
 319
 320         return 0;
 321 }