arch/x86/events/intel/core.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Per core/cpu state
   4  *
   5  * Used to coordinate shared registers between HT threads or
   6  * among events on a single PMU.
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/stddef.h>
  12 #include <linux/types.h>
  13 #include <linux/init.h>
  14 #include <linux/slab.h>
  15 #include <linux/export.h>
  16 #include <linux/nmi.h>
  17
  18 #include <asm/cpufeature.h>
  19 #include <asm/hardirq.h>
  20 #include <asm/intel-family.h>
  21 #include <asm/intel_pt.h>
  22 #include <asm/apic.h>
  23 #include <asm/cpu_device_id.h>
  24
  25 #include "../perf_event.h"
  26
  27 /*
  28  * Intel PerfMon, used on Core and later.
  29  */
  30 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
  31 {
  32         [PERF_COUNT_HW_CPU_CYCLES]              = 0x003c,
  33         [PERF_COUNT_HW_INSTRUCTIONS]            = 0x00c0,
  34         [PERF_COUNT_HW_CACHE_REFERENCES]        = 0x4f2e,
  35         [PERF_COUNT_HW_CACHE_MISSES]            = 0x412e,
  36         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]     = 0x00c4,
  37         [PERF_COUNT_HW_BRANCH_MISSES]           = 0x00c5,
  38         [PERF_COUNT_HW_BUS_CYCLES]              = 0x013c,
  39         [PERF_COUNT_HW_REF_CPU_CYCLES]          = 0x0300, /* pseudo-encoding */
  40 };
  41
  42 static struct event_constraint intel_core_event_constraints[] __read_mostly =
  43 {
  44         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
  45         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
  46         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
  47         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
  48         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
  49         INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
  50         EVENT_CONSTRAINT_END
  51 };
  52
  53 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
  54 {
  55         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
  56         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
  57         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
  58         INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
  59         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
  60         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
  61         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
  62         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
  63         INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
  64         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
  65         INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
  66         INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
  67         INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
  68         EVENT_CONSTRAINT_END
  69 };
  70
  71 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
  72 {
  73         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
  74         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
  75         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
  76         INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
  77         INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
  78         INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
  79         INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
  80         INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
  81         INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
  82         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
  83         INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
  84         EVENT_CONSTRAINT_END
  85 };
  86
  87 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
  88 {
  89         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
  90         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
  91         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
  92         EVENT_EXTRA_END
  93 };
  94
  95 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
  96 {
  97         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
  98         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
  99         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
 100         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
 101         INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
 102         INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
 103         INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
 104         EVENT_CONSTRAINT_END
 105 };
 106
 107 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
 108 {
 109         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 110         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 111         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
 112         INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
 113         INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
 114         INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
 115         INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
 116         INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
 117         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
 118         INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
 119         INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
 120         INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
 121
 122         /*
 123          * When HT is off these events can only run on the bottom 4 counters
 124          * When HT is on, they are impacted by the HT bug and require EXCL access
 125          */
 126         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
 127         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
 128         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
 129         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
 130
 131         EVENT_CONSTRAINT_END
 132 };
 133
 134 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
 135 {
 136         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 137         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 138         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
 139         INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
 140         INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMPTY */
 141         INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
 142         INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */
 143         INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
 144         INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
 145         INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
 146         INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
 147         INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
 148         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
 149
 150         /*
 151          * When HT is off these events can only run on the bottom 4 counters
 152          * When HT is on, they are impacted by the HT bug and require EXCL access
 153          */
 154         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
 155         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
 156         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
 157         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
 158
 159         EVENT_CONSTRAINT_END
 160 };
 161
 162 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
 163 {
 164         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
 165         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
 166         INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
 167         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
 168         EVENT_EXTRA_END
 169 };
 170
 171 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
 172 {
 173         EVENT_CONSTRAINT_END
 174 };
 175
 176 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
 177 {
 178         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 179         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 180         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
 181         EVENT_CONSTRAINT_END
 182 };
 183
 184 static struct event_constraint intel_slm_event_constraints[] __read_mostly =
 185 {
 186         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 187         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 188         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
 189         EVENT_CONSTRAINT_END
 190 };
 191
 192 static struct event_constraint intel_skl_event_constraints[] = {
 193         FIXED_EVENT_CONSTRAINT(0x00c0, 0),      /* INST_RETIRED.ANY */
 194         FIXED_EVENT_CONSTRAINT(0x003c, 1),      /* CPU_CLK_UNHALTED.CORE */
 195         FIXED_EVENT_CONSTRAINT(0x0300, 2),      /* CPU_CLK_UNHALTED.REF */
 196         INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2),    /* INST_RETIRED.PREC_DIST */
 197
 198         /*
 199          * when HT is off, these can only run on the bottom 4 counters
 200          */
 201         INTEL_EVENT_CONSTRAINT(0xd0, 0xf),      /* MEM_INST_RETIRED.* */
 202         INTEL_EVENT_CONSTRAINT(0xd1, 0xf),      /* MEM_LOAD_RETIRED.* */
 203         INTEL_EVENT_CONSTRAINT(0xd2, 0xf),      /* MEM_LOAD_L3_HIT_RETIRED.* */
 204         INTEL_EVENT_CONSTRAINT(0xcd, 0xf),      /* MEM_TRANS_RETIRED.* */
 205         INTEL_EVENT_CONSTRAINT(0xc6, 0xf),      /* FRONTEND_RETIRED.* */
 206
 207         EVENT_CONSTRAINT_END
 208 };
 209
 210 static struct extra_reg intel_knl_extra_regs[] __read_mostly = {
 211         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0),
 212         INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1),
 213         EVENT_EXTRA_END
 214 };
 215
 216 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
 217         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
 218         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
 219         INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
 220         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
 221         EVENT_EXTRA_END
 222 };
 223
 224 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
 225         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
 226         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
 227         INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
 228         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
 229         EVENT_EXTRA_END
 230 };
 231
 232 static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
 233         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
 234         INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
 235         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
 236         /*
 237          * Note the low 8 bits eventsel code is not a continuous field, containing
 238          * some #GPing bits. These are masked out.
 239          */
 240         INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
 241         EVENT_EXTRA_END
 242 };
 243
 244 static struct event_constraint intel_icl_event_constraints[] = {
 245         FIXED_EVENT_CONSTRAINT(0x00c0, 0),      /* INST_RETIRED.ANY */
 246         FIXED_EVENT_CONSTRAINT(0x01c0, 0),      /* INST_RETIRED.PREC_DIST */
 247         FIXED_EVENT_CONSTRAINT(0x003c, 1),      /* CPU_CLK_UNHALTED.CORE */
 248         FIXED_EVENT_CONSTRAINT(0x0300, 2),      /* CPU_CLK_UNHALTED.REF */
 249         FIXED_EVENT_CONSTRAINT(0x0400, 3),      /* SLOTS */
 250         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
 251         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
 252         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
 253         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
 254         INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf),
 255         INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf),
 256         INTEL_EVENT_CONSTRAINT(0x32, 0xf),      /* SW_PREFETCH_ACCESS.* */
 257         INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x54, 0xf),
 258         INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf),
 259         INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff),  /* CYCLE_ACTIVITY.STALLS_TOTAL */
 260         INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff),  /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */
 261         INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff),  /* CYCLE_ACTIVITY.STALLS_MEM_ANY */
 262         INTEL_EVENT_CONSTRAINT(0xa3, 0xf),      /* CYCLE_ACTIVITY.* */
 263         INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf),
 264         INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf),
 265         INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf),
 266         INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf),
 267         EVENT_CONSTRAINT_END
 268 };
 269
 270 static struct extra_reg intel_icl_extra_regs[] __read_mostly = {
 271         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0),
 272         INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1),
 273         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
 274         INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
 275         EVENT_EXTRA_END
 276 };
 277
 278 static struct extra_reg intel_spr_extra_regs[] __read_mostly = {
 279         INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0),
 280         INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1),
 281         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
 282         INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
 283         INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE),
 284         INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE),
 285         EVENT_EXTRA_END
 286 };
 287
 288 static struct event_constraint intel_spr_event_constraints[] = {
 289         FIXED_EVENT_CONSTRAINT(0x00c0, 0),      /* INST_RETIRED.ANY */
 290         FIXED_EVENT_CONSTRAINT(0x01c0, 0),      /* INST_RETIRED.PREC_DIST */
 291         FIXED_EVENT_CONSTRAINT(0x003c, 1),      /* CPU_CLK_UNHALTED.CORE */
 292         FIXED_EVENT_CONSTRAINT(0x0300, 2),      /* CPU_CLK_UNHALTED.REF */
 293         FIXED_EVENT_CONSTRAINT(0x0400, 3),      /* SLOTS */
 294         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
 295         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
 296         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
 297         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
 298         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4),
 299         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5),
 300         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6),
 301         METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7),
 302
 303         INTEL_EVENT_CONSTRAINT(0x2e, 0xff),
 304         INTEL_EVENT_CONSTRAINT(0x3c, 0xff),
 305         /*
 306          * Generally event codes < 0x90 are restricted to counters 0-3.
 307          * The 0x2E and 0x3C are exception, which has no restriction.
 308          */
 309         INTEL_EVENT_CONSTRAINT_RANGE(0x01, 0x8f, 0xf),
 310
 311         INTEL_UEVENT_CONSTRAINT(0x01a3, 0xf),
 312         INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf),
 313         INTEL_UEVENT_CONSTRAINT(0x08a3, 0xf),
 314         INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1),
 315         INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1),
 316         INTEL_UEVENT_CONSTRAINT(0x02cd, 0x1),
 317         INTEL_EVENT_CONSTRAINT(0xce, 0x1),
 318         INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf),
 319         /*
 320          * Generally event codes >= 0x90 are likely to have no restrictions.
 321          * The exception are defined as above.
 322          */
 323         INTEL_EVENT_CONSTRAINT_RANGE(0x90, 0xfe, 0xff),
 324
 325         EVENT_CONSTRAINT_END
 326 };
 327
 328
 329 EVENT_ATTR_STR(mem-loads,       mem_ld_nhm,     "event=0x0b,umask=0x10,ldlat=3");
 330 EVENT_ATTR_STR(mem-loads,       mem_ld_snb,     "event=0xcd,umask=0x1,ldlat=3");
 331 EVENT_ATTR_STR(mem-stores,      mem_st_snb,     "event=0xcd,umask=0x2");
 332
 333 static struct attribute *nhm_mem_events_attrs[] = {
 334         EVENT_PTR(mem_ld_nhm),
 335         NULL,
 336 };
 337
 338 /*
 339  * topdown events for Intel Core CPUs.
 340  *
 341  * The events are all in slots, which is a free slot in a 4 wide
 342  * pipeline. Some events are already reported in slots, for cycle
 343  * events we multiply by the pipeline width (4).
 344  *
 345  * With Hyper Threading on, topdown metrics are either summed or averaged
 346  * between the threads of a core: (count_t0 + count_t1).
 347  *
 348  * For the average case the metric is always scaled to pipeline width,
 349  * so we use factor 2 ((count_t0 + count_t1) / 2 * 4)
 350  */
 351
 352 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots,
 353         "event=0x3c,umask=0x0",                 /* cpu_clk_unhalted.thread */
 354         "event=0x3c,umask=0x0,any=1");          /* cpu_clk_unhalted.thread_any */
 355 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2");
 356 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued,
 357         "event=0xe,umask=0x1");                 /* uops_issued.any */
 358 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired,
 359         "event=0xc2,umask=0x2");                /* uops_retired.retire_slots */
 360 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles,
 361         "event=0x9c,umask=0x1");                /* idq_uops_not_delivered_core */
 362 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles,
 363         "event=0xd,umask=0x3,cmask=1",          /* int_misc.recovery_cycles */
 364         "event=0xd,umask=0x3,cmask=1,any=1");   /* int_misc.recovery_cycles_any */
 365 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale,
 366         "4", "2");
 367
 368 EVENT_ATTR_STR(slots,                   slots,                  "event=0x00,umask=0x4");
 369 EVENT_ATTR_STR(topdown-retiring,        td_retiring,            "event=0x00,umask=0x80");
 370 EVENT_ATTR_STR(topdown-bad-spec,        td_bad_spec,            "event=0x00,umask=0x81");
 371 EVENT_ATTR_STR(topdown-fe-bound,        td_fe_bound,            "event=0x00,umask=0x82");
 372 EVENT_ATTR_STR(topdown-be-bound,        td_be_bound,            "event=0x00,umask=0x83");
 373 EVENT_ATTR_STR(topdown-heavy-ops,       td_heavy_ops,           "event=0x00,umask=0x84");
 374 EVENT_ATTR_STR(topdown-br-mispredict,   td_br_mispredict,       "event=0x00,umask=0x85");
 375 EVENT_ATTR_STR(topdown-fetch-lat,       td_fetch_lat,           "event=0x00,umask=0x86");
 376 EVENT_ATTR_STR(topdown-mem-bound,       td_mem_bound,           "event=0x00,umask=0x87");
 377
 378 static struct attribute *snb_events_attrs[] = {
 379         EVENT_PTR(td_slots_issued),
 380         EVENT_PTR(td_slots_retired),
 381         EVENT_PTR(td_fetch_bubbles),
 382         EVENT_PTR(td_total_slots),
 383         EVENT_PTR(td_total_slots_scale),
 384         EVENT_PTR(td_recovery_bubbles),
 385         EVENT_PTR(td_recovery_bubbles_scale),
 386         NULL,
 387 };
 388
 389 static struct attribute *snb_mem_events_attrs[] = {
 390         EVENT_PTR(mem_ld_snb),
 391         EVENT_PTR(mem_st_snb),
 392         NULL,
 393 };
 394
 395 static struct event_constraint intel_hsw_event_constraints[] = {
 396         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 397         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 398         FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
 399         INTEL_UEVENT_CONSTRAINT(0x148, 0x4),    /* L1D_PEND_MISS.PENDING */
 400         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
 401         INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
 402         /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
 403         INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
 404         /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
 405         INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
 406         /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
 407         INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
 408
 409         /*
 410          * When HT is off these events can only run on the bottom 4 counters
 411          * When HT is on, they are impacted by the HT bug and require EXCL access
 412          */
 413         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
 414         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
 415         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
 416         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
 417
 418         EVENT_CONSTRAINT_END
 419 };
 420
 421 static struct event_constraint intel_bdw_event_constraints[] = {
 422         FIXED_EVENT_CONSTRAINT(0x00c0, 0),      /* INST_RETIRED.ANY */
 423         FIXED_EVENT_CONSTRAINT(0x003c, 1),      /* CPU_CLK_UNHALTED.CORE */
 424         FIXED_EVENT_CONSTRAINT(0x0300, 2),      /* CPU_CLK_UNHALTED.REF */
 425         INTEL_UEVENT_CONSTRAINT(0x148, 0x4),    /* L1D_PEND_MISS.PENDING */
 426         INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4),        /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */
 427         /*
 428          * when HT is off, these can only run on the bottom 4 counters
 429          */
 430         INTEL_EVENT_CONSTRAINT(0xd0, 0xf),      /* MEM_INST_RETIRED.* */
 431         INTEL_EVENT_CONSTRAINT(0xd1, 0xf),      /* MEM_LOAD_RETIRED.* */
 432         INTEL_EVENT_CONSTRAINT(0xd2, 0xf),      /* MEM_LOAD_L3_HIT_RETIRED.* */
 433         INTEL_EVENT_CONSTRAINT(0xcd, 0xf),      /* MEM_TRANS_RETIRED.* */
 434         EVENT_CONSTRAINT_END
 435 };
 436
 437 static u64 intel_pmu_event_map(int hw_event)
 438 {
 439         return intel_perfmon_event_map[hw_event];
 440 }
 441
 442 static __initconst const u64 spr_hw_cache_event_ids
 443                                 [PERF_COUNT_HW_CACHE_MAX]
 444                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 445                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 446 {
 447  [ C(L1D ) ] = {
 448         [ C(OP_READ) ] = {
 449                 [ C(RESULT_ACCESS) ] = 0x81d0,
 450                 [ C(RESULT_MISS)   ] = 0xe124,
 451         },
 452         [ C(OP_WRITE) ] = {
 453                 [ C(RESULT_ACCESS) ] = 0x82d0,
 454         },
 455  },
 456  [ C(L1I ) ] = {
 457         [ C(OP_READ) ] = {
 458                 [ C(RESULT_MISS)   ] = 0xe424,
 459         },
 460         [ C(OP_WRITE) ] = {
 461                 [ C(RESULT_ACCESS) ] = -1,
 462                 [ C(RESULT_MISS)   ] = -1,
 463         },
 464  },
 465  [ C(LL  ) ] = {
 466         [ C(OP_READ) ] = {
 467                 [ C(RESULT_ACCESS) ] = 0x12a,
 468                 [ C(RESULT_MISS)   ] = 0x12a,
 469         },
 470         [ C(OP_WRITE) ] = {
 471                 [ C(RESULT_ACCESS) ] = 0x12a,
 472                 [ C(RESULT_MISS)   ] = 0x12a,
 473         },
 474  },
 475  [ C(DTLB) ] = {
 476         [ C(OP_READ) ] = {
 477                 [ C(RESULT_ACCESS) ] = 0x81d0,
 478                 [ C(RESULT_MISS)   ] = 0xe12,
 479         },
 480         [ C(OP_WRITE) ] = {
 481                 [ C(RESULT_ACCESS) ] = 0x82d0,
 482                 [ C(RESULT_MISS)   ] = 0xe13,
 483         },
 484  },
 485  [ C(ITLB) ] = {
 486         [ C(OP_READ) ] = {
 487                 [ C(RESULT_ACCESS) ] = -1,
 488                 [ C(RESULT_MISS)   ] = 0xe11,
 489         },
 490         [ C(OP_WRITE) ] = {
 491                 [ C(RESULT_ACCESS) ] = -1,
 492                 [ C(RESULT_MISS)   ] = -1,
 493         },
 494         [ C(OP_PREFETCH) ] = {
 495                 [ C(RESULT_ACCESS) ] = -1,
 496                 [ C(RESULT_MISS)   ] = -1,
 497         },
 498  },
 499  [ C(BPU ) ] = {
 500         [ C(OP_READ) ] = {
 501                 [ C(RESULT_ACCESS) ] = 0x4c4,
 502                 [ C(RESULT_MISS)   ] = 0x4c5,
 503         },
 504         [ C(OP_WRITE) ] = {
 505                 [ C(RESULT_ACCESS) ] = -1,
 506                 [ C(RESULT_MISS)   ] = -1,
 507         },
 508         [ C(OP_PREFETCH) ] = {
 509                 [ C(RESULT_ACCESS) ] = -1,
 510                 [ C(RESULT_MISS)   ] = -1,
 511         },
 512  },
 513  [ C(NODE) ] = {
 514         [ C(OP_READ) ] = {
 515                 [ C(RESULT_ACCESS) ] = 0x12a,
 516                 [ C(RESULT_MISS)   ] = 0x12a,
 517         },
 518  },
 519 };
 520
 521 static __initconst const u64 spr_hw_cache_extra_regs
 522                                 [PERF_COUNT_HW_CACHE_MAX]
 523                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 524                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 525 {
 526  [ C(LL  ) ] = {
 527         [ C(OP_READ) ] = {
 528                 [ C(RESULT_ACCESS) ] = 0x10001,
 529                 [ C(RESULT_MISS)   ] = 0x3fbfc00001,
 530         },
 531         [ C(OP_WRITE) ] = {
 532                 [ C(RESULT_ACCESS) ] = 0x3f3ffc0002,
 533                 [ C(RESULT_MISS)   ] = 0x3f3fc00002,
 534         },
 535  },
 536  [ C(NODE) ] = {
 537         [ C(OP_READ) ] = {
 538                 [ C(RESULT_ACCESS) ] = 0x10c000001,
 539                 [ C(RESULT_MISS)   ] = 0x3fb3000001,
 540         },
 541  },
 542 };
 543
 544 /*
 545  * Notes on the events:
 546  * - data reads do not include code reads (comparable to earlier tables)
 547  * - data counts include speculative execution (except L1 write, dtlb, bpu)
 548  * - remote node access includes remote memory, remote cache, remote mmio.
 549  * - prefetches are not included in the counts.
 550  * - icache miss does not include decoded icache
 551  */
 552
 553 #define SKL_DEMAND_DATA_RD              BIT_ULL(0)
 554 #define SKL_DEMAND_RFO                  BIT_ULL(1)
 555 #define SKL_ANY_RESPONSE                BIT_ULL(16)
 556 #define SKL_SUPPLIER_NONE               BIT_ULL(17)
 557 #define SKL_L3_MISS_LOCAL_DRAM          BIT_ULL(26)
 558 #define SKL_L3_MISS_REMOTE_HOP0_DRAM    BIT_ULL(27)
 559 #define SKL_L3_MISS_REMOTE_HOP1_DRAM    BIT_ULL(28)
 560 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM   BIT_ULL(29)
 561 #define SKL_L3_MISS                     (SKL_L3_MISS_LOCAL_DRAM| \
 562                                          SKL_L3_MISS_REMOTE_HOP0_DRAM| \
 563                                          SKL_L3_MISS_REMOTE_HOP1_DRAM| \
 564                                          SKL_L3_MISS_REMOTE_HOP2P_DRAM)
 565 #define SKL_SPL_HIT                     BIT_ULL(30)
 566 #define SKL_SNOOP_NONE                  BIT_ULL(31)
 567 #define SKL_SNOOP_NOT_NEEDED            BIT_ULL(32)
 568 #define SKL_SNOOP_MISS                  BIT_ULL(33)
 569 #define SKL_SNOOP_HIT_NO_FWD            BIT_ULL(34)
 570 #define SKL_SNOOP_HIT_WITH_FWD          BIT_ULL(35)
 571 #define SKL_SNOOP_HITM                  BIT_ULL(36)
 572 #define SKL_SNOOP_NON_DRAM              BIT_ULL(37)
 573 #define SKL_ANY_SNOOP                   (SKL_SPL_HIT|SKL_SNOOP_NONE| \
 574                                          SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
 575                                          SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
 576                                          SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM)
 577 #define SKL_DEMAND_READ                 SKL_DEMAND_DATA_RD
 578 #define SKL_SNOOP_DRAM                  (SKL_SNOOP_NONE| \
 579                                          SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
 580                                          SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
 581                                          SKL_SNOOP_HITM|SKL_SPL_HIT)
 582 #define SKL_DEMAND_WRITE                SKL_DEMAND_RFO
 583 #define SKL_LLC_ACCESS                  SKL_ANY_RESPONSE
 584 #define SKL_L3_MISS_REMOTE              (SKL_L3_MISS_REMOTE_HOP0_DRAM| \
 585                                          SKL_L3_MISS_REMOTE_HOP1_DRAM| \
 586                                          SKL_L3_MISS_REMOTE_HOP2P_DRAM)
 587
 588 static __initconst const u64 skl_hw_cache_event_ids
 589                                 [PERF_COUNT_HW_CACHE_MAX]
 590                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 591                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 592 {
 593  [ C(L1D ) ] = {
 594         [ C(OP_READ) ] = {
 595                 [ C(RESULT_ACCESS) ] = 0x81d0,  /* MEM_INST_RETIRED.ALL_LOADS */
 596                 [ C(RESULT_MISS)   ] = 0x151,   /* L1D.REPLACEMENT */
 597         },
 598         [ C(OP_WRITE) ] = {
 599                 [ C(RESULT_ACCESS) ] = 0x82d0,  /* MEM_INST_RETIRED.ALL_STORES */
 600                 [ C(RESULT_MISS)   ] = 0x0,
 601         },
 602         [ C(OP_PREFETCH) ] = {
 603                 [ C(RESULT_ACCESS) ] = 0x0,
 604                 [ C(RESULT_MISS)   ] = 0x0,
 605         },
 606  },
 607  [ C(L1I ) ] = {
 608         [ C(OP_READ) ] = {
 609                 [ C(RESULT_ACCESS) ] = 0x0,
 610                 [ C(RESULT_MISS)   ] = 0x283,   /* ICACHE_64B.MISS */
 611         },
 612         [ C(OP_WRITE) ] = {
 613                 [ C(RESULT_ACCESS) ] = -1,
 614                 [ C(RESULT_MISS)   ] = -1,
 615         },
 616         [ C(OP_PREFETCH) ] = {
 617                 [ C(RESULT_ACCESS) ] = 0x0,
 618                 [ C(RESULT_MISS)   ] = 0x0,
 619         },
 620  },
 621  [ C(LL  ) ] = {
 622         [ C(OP_READ) ] = {
 623                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
 624                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
 625         },
 626         [ C(OP_WRITE) ] = {
 627                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
 628                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
 629         },
 630         [ C(OP_PREFETCH) ] = {
 631                 [ C(RESULT_ACCESS) ] = 0x0,
 632                 [ C(RESULT_MISS)   ] = 0x0,
 633         },
 634  },
 635  [ C(DTLB) ] = {
 636         [ C(OP_READ) ] = {
 637                 [ C(RESULT_ACCESS) ] = 0x81d0,  /* MEM_INST_RETIRED.ALL_LOADS */
 638                 [ C(RESULT_MISS)   ] = 0xe08,   /* DTLB_LOAD_MISSES.WALK_COMPLETED */
 639         },
 640         [ C(OP_WRITE) ] = {
 641                 [ C(RESULT_ACCESS) ] = 0x82d0,  /* MEM_INST_RETIRED.ALL_STORES */
 642                 [ C(RESULT_MISS)   ] = 0xe49,   /* DTLB_STORE_MISSES.WALK_COMPLETED */
 643         },
 644         [ C(OP_PREFETCH) ] = {
 645                 [ C(RESULT_ACCESS) ] = 0x0,
 646                 [ C(RESULT_MISS)   ] = 0x0,
 647         },
 648  },
 649  [ C(ITLB) ] = {
 650         [ C(OP_READ) ] = {
 651                 [ C(RESULT_ACCESS) ] = 0x2085,  /* ITLB_MISSES.STLB_HIT */
 652                 [ C(RESULT_MISS)   ] = 0xe85,   /* ITLB_MISSES.WALK_COMPLETED */
 653         },
 654         [ C(OP_WRITE) ] = {
 655                 [ C(RESULT_ACCESS) ] = -1,
 656                 [ C(RESULT_MISS)   ] = -1,
 657         },
 658         [ C(OP_PREFETCH) ] = {
 659                 [ C(RESULT_ACCESS) ] = -1,
 660                 [ C(RESULT_MISS)   ] = -1,
 661         },
 662  },
 663  [ C(BPU ) ] = {
 664         [ C(OP_READ) ] = {
 665                 [ C(RESULT_ACCESS) ] = 0xc4,    /* BR_INST_RETIRED.ALL_BRANCHES */
 666                 [ C(RESULT_MISS)   ] = 0xc5,    /* BR_MISP_RETIRED.ALL_BRANCHES */
 667         },
 668         [ C(OP_WRITE) ] = {
 669                 [ C(RESULT_ACCESS) ] = -1,
 670                 [ C(RESULT_MISS)   ] = -1,
 671         },
 672         [ C(OP_PREFETCH) ] = {
 673                 [ C(RESULT_ACCESS) ] = -1,
 674                 [ C(RESULT_MISS)   ] = -1,
 675         },
 676  },
 677  [ C(NODE) ] = {
 678         [ C(OP_READ) ] = {
 679                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
 680                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
 681         },
 682         [ C(OP_WRITE) ] = {
 683                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
 684                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
 685         },
 686         [ C(OP_PREFETCH) ] = {
 687                 [ C(RESULT_ACCESS) ] = 0x0,
 688                 [ C(RESULT_MISS)   ] = 0x0,
 689         },
 690  },
 691 };
 692
 693 static __initconst const u64 skl_hw_cache_extra_regs
 694                                 [PERF_COUNT_HW_CACHE_MAX]
 695                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 696                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 697 {
 698  [ C(LL  ) ] = {
 699         [ C(OP_READ) ] = {
 700                 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
 701                                        SKL_LLC_ACCESS|SKL_ANY_SNOOP,
 702                 [ C(RESULT_MISS)   ] = SKL_DEMAND_READ|
 703                                        SKL_L3_MISS|SKL_ANY_SNOOP|
 704                                        SKL_SUPPLIER_NONE,
 705         },
 706         [ C(OP_WRITE) ] = {
 707                 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
 708                                        SKL_LLC_ACCESS|SKL_ANY_SNOOP,
 709                 [ C(RESULT_MISS)   ] = SKL_DEMAND_WRITE|
 710                                        SKL_L3_MISS|SKL_ANY_SNOOP|
 711                                        SKL_SUPPLIER_NONE,
 712         },
 713         [ C(OP_PREFETCH) ] = {
 714                 [ C(RESULT_ACCESS) ] = 0x0,
 715                 [ C(RESULT_MISS)   ] = 0x0,
 716         },
 717  },
 718  [ C(NODE) ] = {
 719         [ C(OP_READ) ] = {
 720                 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
 721                                        SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
 722                 [ C(RESULT_MISS)   ] = SKL_DEMAND_READ|
 723                                        SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
 724         },
 725         [ C(OP_WRITE) ] = {
 726                 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
 727                                        SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
 728                 [ C(RESULT_MISS)   ] = SKL_DEMAND_WRITE|
 729                                        SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
 730         },
 731         [ C(OP_PREFETCH) ] = {
 732                 [ C(RESULT_ACCESS) ] = 0x0,
 733                 [ C(RESULT_MISS)   ] = 0x0,
 734         },
 735  },
 736 };
 737
 738 #define SNB_DMND_DATA_RD        (1ULL << 0)
 739 #define SNB_DMND_RFO            (1ULL << 1)
 740 #define SNB_DMND_IFETCH         (1ULL << 2)
 741 #define SNB_DMND_WB             (1ULL << 3)
 742 #define SNB_PF_DATA_RD          (1ULL << 4)
 743 #define SNB_PF_RFO              (1ULL << 5)
 744 #define SNB_PF_IFETCH           (1ULL << 6)
 745 #define SNB_LLC_DATA_RD         (1ULL << 7)
 746 #define SNB_LLC_RFO             (1ULL << 8)
 747 #define SNB_LLC_IFETCH          (1ULL << 9)
 748 #define SNB_BUS_LOCKS           (1ULL << 10)
 749 #define SNB_STRM_ST             (1ULL << 11)
 750 #define SNB_OTHER               (1ULL << 15)
 751 #define SNB_RESP_ANY            (1ULL << 16)
 752 #define SNB_NO_SUPP             (1ULL << 17)
 753 #define SNB_LLC_HITM            (1ULL << 18)
 754 #define SNB_LLC_HITE            (1ULL << 19)
 755 #define SNB_LLC_HITS            (1ULL << 20)
 756 #define SNB_LLC_HITF            (1ULL << 21)
 757 #define SNB_LOCAL               (1ULL << 22)
 758 #define SNB_REMOTE              (0xffULL << 23)
 759 #define SNB_SNP_NONE            (1ULL << 31)
 760 #define SNB_SNP_NOT_NEEDED      (1ULL << 32)
 761 #define SNB_SNP_MISS            (1ULL << 33)
 762 #define SNB_NO_FWD              (1ULL << 34)
 763 #define SNB_SNP_FWD             (1ULL << 35)
 764 #define SNB_HITM                (1ULL << 36)
 765 #define SNB_NON_DRAM            (1ULL << 37)
 766
 767 #define SNB_DMND_READ           (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
 768 #define SNB_DMND_WRITE          (SNB_DMND_RFO|SNB_LLC_RFO)
 769 #define SNB_DMND_PREFETCH       (SNB_PF_DATA_RD|SNB_PF_RFO)
 770
 771 #define SNB_SNP_ANY             (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
 772                                  SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
 773                                  SNB_HITM)
 774
 775 #define SNB_DRAM_ANY            (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
 776 #define SNB_DRAM_REMOTE         (SNB_REMOTE|SNB_SNP_ANY)
 777
 778 #define SNB_L3_ACCESS           SNB_RESP_ANY
 779 #define SNB_L3_MISS             (SNB_DRAM_ANY|SNB_NON_DRAM)
 780
 781 static __initconst const u64 snb_hw_cache_extra_regs
 782                                 [PERF_COUNT_HW_CACHE_MAX]
 783                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 784                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 785 {
 786  [ C(LL  ) ] = {
 787         [ C(OP_READ) ] = {
 788                 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
 789                 [ C(RESULT_MISS)   ] = SNB_DMND_READ|SNB_L3_MISS,
 790         },
 791         [ C(OP_WRITE) ] = {
 792                 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
 793                 [ C(RESULT_MISS)   ] = SNB_DMND_WRITE|SNB_L3_MISS,
 794         },
 795         [ C(OP_PREFETCH) ] = {
 796                 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
 797                 [ C(RESULT_MISS)   ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
 798         },
 799  },
 800  [ C(NODE) ] = {
 801         [ C(OP_READ) ] = {
 802                 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
 803                 [ C(RESULT_MISS)   ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
 804         },
 805         [ C(OP_WRITE) ] = {
 806                 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
 807                 [ C(RESULT_MISS)   ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
 808         },
 809         [ C(OP_PREFETCH) ] = {
 810                 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
 811                 [ C(RESULT_MISS)   ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
 812         },
 813  },
 814 };
 815
 816 static __initconst const u64 snb_hw_cache_event_ids
 817                                 [PERF_COUNT_HW_CACHE_MAX]
 818                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 819                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 820 {
 821  [ C(L1D) ] = {
 822         [ C(OP_READ) ] = {
 823                 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS        */
 824                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPLACEMENT              */
 825         },
 826         [ C(OP_WRITE) ] = {
 827                 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES       */
 828                 [ C(RESULT_MISS)   ] = 0x0851, /* L1D.ALL_M_REPLACEMENT        */
 829         },
 830         [ C(OP_PREFETCH) ] = {
 831                 [ C(RESULT_ACCESS) ] = 0x0,
 832                 [ C(RESULT_MISS)   ] = 0x024e, /* HW_PRE_REQ.DL1_MISS          */
 833         },
 834  },
 835  [ C(L1I ) ] = {
 836         [ C(OP_READ) ] = {
 837                 [ C(RESULT_ACCESS) ] = 0x0,
 838                 [ C(RESULT_MISS)   ] = 0x0280, /* ICACHE.MISSES */
 839         },
 840         [ C(OP_WRITE) ] = {
 841                 [ C(RESULT_ACCESS) ] = -1,
 842                 [ C(RESULT_MISS)   ] = -1,
 843         },
 844         [ C(OP_PREFETCH) ] = {
 845                 [ C(RESULT_ACCESS) ] = 0x0,
 846                 [ C(RESULT_MISS)   ] = 0x0,
 847         },
 848  },
 849  [ C(LL  ) ] = {
 850         [ C(OP_READ) ] = {
 851                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
 852                 [ C(RESULT_ACCESS) ] = 0x01b7,
 853                 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
 854                 [ C(RESULT_MISS)   ] = 0x01b7,
 855         },
 856         [ C(OP_WRITE) ] = {
 857                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
 858                 [ C(RESULT_ACCESS) ] = 0x01b7,
 859                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
 860                 [ C(RESULT_MISS)   ] = 0x01b7,
 861         },
 862         [ C(OP_PREFETCH) ] = {
 863                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
 864                 [ C(RESULT_ACCESS) ] = 0x01b7,
 865                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
 866                 [ C(RESULT_MISS)   ] = 0x01b7,
 867         },
 868  },
 869  [ C(DTLB) ] = {
 870         [ C(OP_READ) ] = {
 871                 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
 872                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
 873         },
 874         [ C(OP_WRITE) ] = {
 875                 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
 876                 [ C(RESULT_MISS)   ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
 877         },
 878         [ C(OP_PREFETCH) ] = {
 879                 [ C(RESULT_ACCESS) ] = 0x0,
 880                 [ C(RESULT_MISS)   ] = 0x0,
 881         },
 882  },
 883  [ C(ITLB) ] = {
 884         [ C(OP_READ) ] = {
 885                 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT         */
 886                 [ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK    */
 887         },
 888         [ C(OP_WRITE) ] = {
 889                 [ C(RESULT_ACCESS) ] = -1,
 890                 [ C(RESULT_MISS)   ] = -1,
 891         },
 892         [ C(OP_PREFETCH) ] = {
 893                 [ C(RESULT_ACCESS) ] = -1,
 894                 [ C(RESULT_MISS)   ] = -1,
 895         },
 896  },
 897  [ C(BPU ) ] = {
 898         [ C(OP_READ) ] = {
 899                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
 900                 [ C(RESULT_MISS)   ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
 901         },
 902         [ C(OP_WRITE) ] = {
 903                 [ C(RESULT_ACCESS) ] = -1,
 904                 [ C(RESULT_MISS)   ] = -1,
 905         },
 906         [ C(OP_PREFETCH) ] = {
 907                 [ C(RESULT_ACCESS) ] = -1,
 908                 [ C(RESULT_MISS)   ] = -1,
 909         },
 910  },
 911  [ C(NODE) ] = {
 912         [ C(OP_READ) ] = {
 913                 [ C(RESULT_ACCESS) ] = 0x01b7,
 914                 [ C(RESULT_MISS)   ] = 0x01b7,
 915         },
 916         [ C(OP_WRITE) ] = {
 917                 [ C(RESULT_ACCESS) ] = 0x01b7,
 918                 [ C(RESULT_MISS)   ] = 0x01b7,
 919         },
 920         [ C(OP_PREFETCH) ] = {
 921                 [ C(RESULT_ACCESS) ] = 0x01b7,
 922                 [ C(RESULT_MISS)   ] = 0x01b7,
 923         },
 924  },
 925
 926 };
 927
 928 /*
 929  * Notes on the events:
 930  * - data reads do not include code reads (comparable to earlier tables)
 931  * - data counts include speculative execution (except L1 write, dtlb, bpu)
 932  * - remote node access includes remote memory, remote cache, remote mmio.
 933  * - prefetches are not included in the counts because they are not
 934  *   reliably counted.
 935  */
 936
 937 #define HSW_DEMAND_DATA_RD              BIT_ULL(0)
 938 #define HSW_DEMAND_RFO                  BIT_ULL(1)
 939 #define HSW_ANY_RESPONSE                BIT_ULL(16)
 940 #define HSW_SUPPLIER_NONE               BIT_ULL(17)
 941 #define HSW_L3_MISS_LOCAL_DRAM          BIT_ULL(22)
 942 #define HSW_L3_MISS_REMOTE_HOP0         BIT_ULL(27)
 943 #define HSW_L3_MISS_REMOTE_HOP1         BIT_ULL(28)
 944 #define HSW_L3_MISS_REMOTE_HOP2P        BIT_ULL(29)
 945 #define HSW_L3_MISS                     (HSW_L3_MISS_LOCAL_DRAM| \
 946                                          HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
 947                                          HSW_L3_MISS_REMOTE_HOP2P)
 948 #define HSW_SNOOP_NONE                  BIT_ULL(31)
 949 #define HSW_SNOOP_NOT_NEEDED            BIT_ULL(32)
 950 #define HSW_SNOOP_MISS                  BIT_ULL(33)
 951 #define HSW_SNOOP_HIT_NO_FWD            BIT_ULL(34)
 952 #define HSW_SNOOP_HIT_WITH_FWD          BIT_ULL(35)
 953 #define HSW_SNOOP_HITM                  BIT_ULL(36)
 954 #define HSW_SNOOP_NON_DRAM              BIT_ULL(37)
 955 #define HSW_ANY_SNOOP                   (HSW_SNOOP_NONE| \
 956                                          HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
 957                                          HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
 958                                          HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
 959 #define HSW_SNOOP_DRAM                  (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
 960 #define HSW_DEMAND_READ                 HSW_DEMAND_DATA_RD
 961 #define HSW_DEMAND_WRITE                HSW_DEMAND_RFO
 962 #define HSW_L3_MISS_REMOTE              (HSW_L3_MISS_REMOTE_HOP0|\
 963                                          HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
 964 #define HSW_LLC_ACCESS                  HSW_ANY_RESPONSE
 965
 966 #define BDW_L3_MISS_LOCAL               BIT(26)
 967 #define BDW_L3_MISS                     (BDW_L3_MISS_LOCAL| \
 968                                          HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
 969                                          HSW_L3_MISS_REMOTE_HOP2P)
 970
 971
 972 static __initconst const u64 hsw_hw_cache_event_ids
 973                                 [PERF_COUNT_HW_CACHE_MAX]
 974                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 975                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 976 {
 977  [ C(L1D ) ] = {
 978         [ C(OP_READ) ] = {
 979                 [ C(RESULT_ACCESS) ] = 0x81d0,  /* MEM_UOPS_RETIRED.ALL_LOADS */
 980                 [ C(RESULT_MISS)   ] = 0x151,   /* L1D.REPLACEMENT */
 981         },
 982         [ C(OP_WRITE) ] = {
 983                 [ C(RESULT_ACCESS) ] = 0x82d0,  /* MEM_UOPS_RETIRED.ALL_STORES */
 984                 [ C(RESULT_MISS)   ] = 0x0,
 985         },
 986         [ C(OP_PREFETCH) ] = {
 987                 [ C(RESULT_ACCESS) ] = 0x0,
 988                 [ C(RESULT_MISS)   ] = 0x0,
 989         },
 990  },
 991  [ C(L1I ) ] = {
 992         [ C(OP_READ) ] = {
 993                 [ C(RESULT_ACCESS) ] = 0x0,
 994                 [ C(RESULT_MISS)   ] = 0x280,   /* ICACHE.MISSES */
 995         },
 996         [ C(OP_WRITE) ] = {
 997                 [ C(RESULT_ACCESS) ] = -1,
 998                 [ C(RESULT_MISS)   ] = -1,
 999         },
1000         [ C(OP_PREFETCH) ] = {
1001                 [ C(RESULT_ACCESS) ] = 0x0,
1002                 [ C(RESULT_MISS)   ] = 0x0,
1003         },
1004  },
1005  [ C(LL  ) ] = {
1006         [ C(OP_READ) ] = {
1007                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
1008                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
1009         },
1010         [ C(OP_WRITE) ] = {
1011                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
1012                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
1013         },
1014         [ C(OP_PREFETCH) ] = {
1015                 [ C(RESULT_ACCESS) ] = 0x0,
1016                 [ C(RESULT_MISS)   ] = 0x0,
1017         },
1018  },
1019  [ C(DTLB) ] = {
1020         [ C(OP_READ) ] = {
1021                 [ C(RESULT_ACCESS) ] = 0x81d0,  /* MEM_UOPS_RETIRED.ALL_LOADS */
1022                 [ C(RESULT_MISS)   ] = 0x108,   /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
1023         },
1024         [ C(OP_WRITE) ] = {
1025                 [ C(RESULT_ACCESS) ] = 0x82d0,  /* MEM_UOPS_RETIRED.ALL_STORES */
1026                 [ C(RESULT_MISS)   ] = 0x149,   /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
1027         },
1028         [ C(OP_PREFETCH) ] = {
1029                 [ C(RESULT_ACCESS) ] = 0x0,
1030                 [ C(RESULT_MISS)   ] = 0x0,
1031         },
1032  },
1033  [ C(ITLB) ] = {
1034         [ C(OP_READ) ] = {
1035                 [ C(RESULT_ACCESS) ] = 0x6085,  /* ITLB_MISSES.STLB_HIT */
1036                 [ C(RESULT_MISS)   ] = 0x185,   /* ITLB_MISSES.MISS_CAUSES_A_WALK */
1037         },
1038         [ C(OP_WRITE) ] = {
1039                 [ C(RESULT_ACCESS) ] = -1,
1040                 [ C(RESULT_MISS)   ] = -1,
1041         },
1042         [ C(OP_PREFETCH) ] = {
1043                 [ C(RESULT_ACCESS) ] = -1,
1044                 [ C(RESULT_MISS)   ] = -1,
1045         },
1046  },
1047  [ C(BPU ) ] = {
1048         [ C(OP_READ) ] = {
1049                 [ C(RESULT_ACCESS) ] = 0xc4,    /* BR_INST_RETIRED.ALL_BRANCHES */
1050                 [ C(RESULT_MISS)   ] = 0xc5,    /* BR_MISP_RETIRED.ALL_BRANCHES */
1051         },
1052         [ C(OP_WRITE) ] = {
1053                 [ C(RESULT_ACCESS) ] = -1,
1054                 [ C(RESULT_MISS)   ] = -1,
1055         },
1056         [ C(OP_PREFETCH) ] = {
1057                 [ C(RESULT_ACCESS) ] = -1,
1058                 [ C(RESULT_MISS)   ] = -1,
1059         },
1060  },
1061  [ C(NODE) ] = {
1062         [ C(OP_READ) ] = {
1063                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
1064                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
1065         },
1066         [ C(OP_WRITE) ] = {
1067                 [ C(RESULT_ACCESS) ] = 0x1b7,   /* OFFCORE_RESPONSE */
1068                 [ C(RESULT_MISS)   ] = 0x1b7,   /* OFFCORE_RESPONSE */
1069         },
1070         [ C(OP_PREFETCH) ] = {
1071                 [ C(RESULT_ACCESS) ] = 0x0,
1072                 [ C(RESULT_MISS)   ] = 0x0,
1073         },
1074  },
1075 };
1076
1077 static __initconst const u64 hsw_hw_cache_extra_regs
1078                                 [PERF_COUNT_HW_CACHE_MAX]
1079                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1080                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1081 {
1082  [ C(LL  ) ] = {
1083         [ C(OP_READ) ] = {
1084                 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
1085                                        HSW_LLC_ACCESS,
1086                 [ C(RESULT_MISS)   ] = HSW_DEMAND_READ|
1087                                        HSW_L3_MISS|HSW_ANY_SNOOP,
1088         },
1089         [ C(OP_WRITE) ] = {
1090                 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
1091                                        HSW_LLC_ACCESS,
1092                 [ C(RESULT_MISS)   ] = HSW_DEMAND_WRITE|
1093                                        HSW_L3_MISS|HSW_ANY_SNOOP,
1094         },
1095         [ C(OP_PREFETCH) ] = {
1096                 [ C(RESULT_ACCESS) ] = 0x0,
1097                 [ C(RESULT_MISS)   ] = 0x0,
1098         },
1099  },
1100  [ C(NODE) ] = {
1101         [ C(OP_READ) ] = {
1102                 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
1103                                        HSW_L3_MISS_LOCAL_DRAM|
1104                                        HSW_SNOOP_DRAM,
1105                 [ C(RESULT_MISS)   ] = HSW_DEMAND_READ|
1106                                        HSW_L3_MISS_REMOTE|
1107                                        HSW_SNOOP_DRAM,
1108         },
1109         [ C(OP_WRITE) ] = {
1110                 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
1111                                        HSW_L3_MISS_LOCAL_DRAM|
1112                                        HSW_SNOOP_DRAM,
1113                 [ C(RESULT_MISS)   ] = HSW_DEMAND_WRITE|
1114                                        HSW_L3_MISS_REMOTE|
1115                                        HSW_SNOOP_DRAM,
1116         },
1117         [ C(OP_PREFETCH) ] = {
1118                 [ C(RESULT_ACCESS) ] = 0x0,
1119                 [ C(RESULT_MISS)   ] = 0x0,
1120         },
1121  },
1122 };
1123
1124 static __initconst const u64 westmere_hw_cache_event_ids
1125                                 [PERF_COUNT_HW_CACHE_MAX]
1126                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1127                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1128 {
1129  [ C(L1D) ] = {
1130         [ C(OP_READ) ] = {
1131                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
1132                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
1133         },
1134         [ C(OP_WRITE) ] = {
1135                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
1136                 [ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
1137         },
1138         [ C(OP_PREFETCH) ] = {
1139                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
1140                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
1141         },
1142  },
1143  [ C(L1I ) ] = {
1144         [ C(OP_READ) ] = {
1145                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
1146                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
1147         },
1148         [ C(OP_WRITE) ] = {
1149                 [ C(RESULT_ACCESS) ] = -1,
1150                 [ C(RESULT_MISS)   ] = -1,
1151         },
1152         [ C(OP_PREFETCH) ] = {
1153                 [ C(RESULT_ACCESS) ] = 0x0,
1154                 [ C(RESULT_MISS)   ] = 0x0,
1155         },
1156  },
1157  [ C(LL  ) ] = {
1158         [ C(OP_READ) ] = {
1159                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1160                 [ C(RESULT_ACCESS) ] = 0x01b7,
1161                 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1162                 [ C(RESULT_MISS)   ] = 0x01b7,
1163         },
1164         /*
1165          * Use RFO, not WRITEBACK, because a write miss would typically occur
1166          * on RFO.
1167          */
1168         [ C(OP_WRITE) ] = {
1169                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1170                 [ C(RESULT_ACCESS) ] = 0x01b7,
1171                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1172                 [ C(RESULT_MISS)   ] = 0x01b7,
1173         },
1174         [ C(OP_PREFETCH) ] = {
1175                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1176                 [ C(RESULT_ACCESS) ] = 0x01b7,
1177                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1178                 [ C(RESULT_MISS)   ] = 0x01b7,
1179         },
1180  },
1181  [ C(DTLB) ] = {
1182         [ C(OP_READ) ] = {
1183                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
1184                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
1185         },
1186         [ C(OP_WRITE) ] = {
1187                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
1188                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
1189         },
1190         [ C(OP_PREFETCH) ] = {
1191                 [ C(RESULT_ACCESS) ] = 0x0,
1192                 [ C(RESULT_MISS)   ] = 0x0,
1193         },
1194  },
1195  [ C(ITLB) ] = {
1196         [ C(OP_READ) ] = {
1197                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
1198                 [ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.ANY              */
1199         },
1200         [ C(OP_WRITE) ] = {
1201                 [ C(RESULT_ACCESS) ] = -1,
1202                 [ C(RESULT_MISS)   ] = -1,
1203         },
1204         [ C(OP_PREFETCH) ] = {
1205                 [ C(RESULT_ACCESS) ] = -1,
1206                 [ C(RESULT_MISS)   ] = -1,
1207         },
1208  },
1209  [ C(BPU ) ] = {
1210         [ C(OP_READ) ] = {
1211                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1212                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
1213         },
1214         [ C(OP_WRITE) ] = {
1215                 [ C(RESULT_ACCESS) ] = -1,
1216                 [ C(RESULT_MISS)   ] = -1,
1217         },
1218         [ C(OP_PREFETCH) ] = {
1219                 [ C(RESULT_ACCESS) ] = -1,
1220                 [ C(RESULT_MISS)   ] = -1,
1221         },
1222  },
1223  [ C(NODE) ] = {
1224         [ C(OP_READ) ] = {
1225                 [ C(RESULT_ACCESS) ] = 0x01b7,
1226                 [ C(RESULT_MISS)   ] = 0x01b7,
1227         },
1228         [ C(OP_WRITE) ] = {
1229                 [ C(RESULT_ACCESS) ] = 0x01b7,
1230                 [ C(RESULT_MISS)   ] = 0x01b7,
1231         },
1232         [ C(OP_PREFETCH) ] = {
1233                 [ C(RESULT_ACCESS) ] = 0x01b7,
1234                 [ C(RESULT_MISS)   ] = 0x01b7,
1235         },
1236  },
1237 };
1238
1239 /*
1240  * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
1241  * See IA32 SDM Vol 3B 30.6.1.3
1242  */
1243
1244 #define NHM_DMND_DATA_RD        (1 << 0)
1245 #define NHM_DMND_RFO            (1 << 1)
1246 #define NHM_DMND_IFETCH         (1 << 2)
1247 #define NHM_DMND_WB             (1 << 3)
1248 #define NHM_PF_DATA_RD          (1 << 4)
1249 #define NHM_PF_DATA_RFO         (1 << 5)
1250 #define NHM_PF_IFETCH           (1 << 6)
1251 #define NHM_OFFCORE_OTHER       (1 << 7)
1252 #define NHM_UNCORE_HIT          (1 << 8)
1253 #define NHM_OTHER_CORE_HIT_SNP  (1 << 9)
1254 #define NHM_OTHER_CORE_HITM     (1 << 10)
1255                                 /* reserved */
1256 #define NHM_REMOTE_CACHE_FWD    (1 << 12)
1257 #define NHM_REMOTE_DRAM         (1 << 13)
1258 #define NHM_LOCAL_DRAM          (1 << 14)
1259 #define NHM_NON_DRAM            (1 << 15)
1260
1261 #define NHM_LOCAL               (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
1262 #define NHM_REMOTE              (NHM_REMOTE_DRAM)
1263
1264 #define NHM_DMND_READ           (NHM_DMND_DATA_RD)
1265 #define NHM_DMND_WRITE          (NHM_DMND_RFO|NHM_DMND_WB)
1266 #define NHM_DMND_PREFETCH       (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
1267
1268 #define NHM_L3_HIT      (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
1269 #define NHM_L3_MISS     (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
1270 #define NHM_L3_ACCESS   (NHM_L3_HIT|NHM_L3_MISS)
1271
1272 static __initconst const u64 nehalem_hw_cache_extra_regs
1273                                 [PERF_COUNT_HW_CACHE_MAX]
1274                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1275                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1276 {
1277  [ C(LL  ) ] = {
1278         [ C(OP_READ) ] = {
1279                 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
1280                 [ C(RESULT_MISS)   ] = NHM_DMND_READ|NHM_L3_MISS,
1281         },
1282         [ C(OP_WRITE) ] = {
1283                 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
1284                 [ C(RESULT_MISS)   ] = NHM_DMND_WRITE|NHM_L3_MISS,
1285         },
1286         [ C(OP_PREFETCH) ] = {
1287                 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
1288                 [ C(RESULT_MISS)   ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
1289         },
1290  },
1291  [ C(NODE) ] = {
1292         [ C(OP_READ) ] = {
1293                 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
1294                 [ C(RESULT_MISS)   ] = NHM_DMND_READ|NHM_REMOTE,
1295         },
1296         [ C(OP_WRITE) ] = {
1297                 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
1298                 [ C(RESULT_MISS)   ] = NHM_DMND_WRITE|NHM_REMOTE,
1299         },
1300         [ C(OP_PREFETCH) ] = {
1301                 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
1302                 [ C(RESULT_MISS)   ] = NHM_DMND_PREFETCH|NHM_REMOTE,
1303         },
1304  },
1305 };
1306
1307 static __initconst const u64 nehalem_hw_cache_event_ids
1308                                 [PERF_COUNT_HW_CACHE_MAX]
1309                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1310                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1311 {
1312  [ C(L1D) ] = {
1313         [ C(OP_READ) ] = {
1314                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
1315                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
1316         },
1317         [ C(OP_WRITE) ] = {
1318                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
1319                 [ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
1320         },
1321         [ C(OP_PREFETCH) ] = {
1322                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
1323                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
1324         },
1325  },
1326  [ C(L1I ) ] = {
1327         [ C(OP_READ) ] = {
1328                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
1329                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
1330         },
1331         [ C(OP_WRITE) ] = {
1332                 [ C(RESULT_ACCESS) ] = -1,
1333                 [ C(RESULT_MISS)   ] = -1,
1334         },
1335         [ C(OP_PREFETCH) ] = {
1336                 [ C(RESULT_ACCESS) ] = 0x0,
1337                 [ C(RESULT_MISS)   ] = 0x0,
1338         },
1339  },
1340  [ C(LL  ) ] = {
1341         [ C(OP_READ) ] = {
1342                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1343                 [ C(RESULT_ACCESS) ] = 0x01b7,
1344                 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1345                 [ C(RESULT_MISS)   ] = 0x01b7,
1346         },
1347         /*
1348          * Use RFO, not WRITEBACK, because a write miss would typically occur
1349          * on RFO.
1350          */
1351         [ C(OP_WRITE) ] = {
1352                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1353                 [ C(RESULT_ACCESS) ] = 0x01b7,
1354                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1355                 [ C(RESULT_MISS)   ] = 0x01b7,
1356         },
1357         [ C(OP_PREFETCH) ] = {
1358                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1359                 [ C(RESULT_ACCESS) ] = 0x01b7,
1360                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1361                 [ C(RESULT_MISS)   ] = 0x01b7,
1362         },
1363  },
1364  [ C(DTLB) ] = {
1365         [ C(OP_READ) ] = {
1366                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI   (alias)  */
1367                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
1368         },
1369         [ C(OP_WRITE) ] = {
1370                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI   (alias)  */
1371                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
1372         },
1373         [ C(OP_PREFETCH) ] = {
1374                 [ C(RESULT_ACCESS) ] = 0x0,
1375                 [ C(RESULT_MISS)   ] = 0x0,
1376         },
1377  },
1378  [ C(ITLB) ] = {
1379         [ C(OP_READ) ] = {
1380                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
1381                 [ C(RESULT_MISS)   ] = 0x20c8, /* ITLB_MISS_RETIRED            */
1382         },
1383         [ C(OP_WRITE) ] = {
1384                 [ C(RESULT_ACCESS) ] = -1,
1385                 [ C(RESULT_MISS)   ] = -1,
1386         },
1387         [ C(OP_PREFETCH) ] = {
1388                 [ C(RESULT_ACCESS) ] = -1,
1389                 [ C(RESULT_MISS)   ] = -1,
1390         },
1391  },
1392  [ C(BPU ) ] = {
1393         [ C(OP_READ) ] = {
1394                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1395                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
1396         },
1397         [ C(OP_WRITE) ] = {
1398                 [ C(RESULT_ACCESS) ] = -1,
1399                 [ C(RESULT_MISS)   ] = -1,
1400         },
1401         [ C(OP_PREFETCH) ] = {
1402                 [ C(RESULT_ACCESS) ] = -1,
1403                 [ C(RESULT_MISS)   ] = -1,
1404         },
1405  },
1406  [ C(NODE) ] = {
1407         [ C(OP_READ) ] = {
1408                 [ C(RESULT_ACCESS) ] = 0x01b7,
1409                 [ C(RESULT_MISS)   ] = 0x01b7,
1410         },
1411         [ C(OP_WRITE) ] = {
1412                 [ C(RESULT_ACCESS) ] = 0x01b7,
1413                 [ C(RESULT_MISS)   ] = 0x01b7,
1414         },
1415         [ C(OP_PREFETCH) ] = {
1416                 [ C(RESULT_ACCESS) ] = 0x01b7,
1417                 [ C(RESULT_MISS)   ] = 0x01b7,
1418         },
1419  },
1420 };
1421
1422 static __initconst const u64 core2_hw_cache_event_ids
1423                                 [PERF_COUNT_HW_CACHE_MAX]
1424                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1425                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1426 {
1427  [ C(L1D) ] = {
1428         [ C(OP_READ) ] = {
1429                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI          */
1430                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE       */
1431         },
1432         [ C(OP_WRITE) ] = {
1433                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI          */
1434                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE       */
1435         },
1436         [ C(OP_PREFETCH) ] = {
1437                 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS      */
1438                 [ C(RESULT_MISS)   ] = 0,
1439         },
1440  },
1441  [ C(L1I ) ] = {
1442         [ C(OP_READ) ] = {
1443                 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS                  */
1444                 [ C(RESULT_MISS)   ] = 0x0081, /* L1I.MISSES                 */
1445         },
1446         [ C(OP_WRITE) ] = {
1447                 [ C(RESULT_ACCESS) ] = -1,
1448                 [ C(RESULT_MISS)   ] = -1,
1449         },
1450         [ C(OP_PREFETCH) ] = {
1451                 [ C(RESULT_ACCESS) ] = 0,
1452                 [ C(RESULT_MISS)   ] = 0,
1453         },
1454  },
1455  [ C(LL  ) ] = {
1456         [ C(OP_READ) ] = {
1457                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
1458                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
1459         },
1460         [ C(OP_WRITE) ] = {
1461                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
1462                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
1463         },
1464         [ C(OP_PREFETCH) ] = {
1465                 [ C(RESULT_ACCESS) ] = 0,
1466                 [ C(RESULT_MISS)   ] = 0,
1467         },
1468  },
1469  [ C(DTLB) ] = {
1470         [ C(OP_READ) ] = {
1471                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI  (alias) */
1472                 [ C(RESULT_MISS)   ] = 0x0208, /* DTLB_MISSES.MISS_LD        */
1473         },
1474         [ C(OP_WRITE) ] = {
1475                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI  (alias) */
1476                 [ C(RESULT_MISS)   ] = 0x0808, /* DTLB_MISSES.MISS_ST        */
1477         },
1478         [ C(OP_PREFETCH) ] = {
1479                 [ C(RESULT_ACCESS) ] = 0,
1480                 [ C(RESULT_MISS)   ] = 0,
1481         },
1482  },
1483  [ C(ITLB) ] = {
1484         [ C(OP_READ) ] = {
1485                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
1486                 [ C(RESULT_MISS)   ] = 0x1282, /* ITLBMISSES                 */
1487         },
1488         [ C(OP_WRITE) ] = {
1489                 [ C(RESULT_ACCESS) ] = -1,
1490                 [ C(RESULT_MISS)   ] = -1,
1491         },
1492         [ C(OP_PREFETCH) ] = {
1493                 [ C(RESULT_ACCESS) ] = -1,
1494                 [ C(RESULT_MISS)   ] = -1,
1495         },
1496  },
1497  [ C(BPU ) ] = {
1498         [ C(OP_READ) ] = {
1499                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
1500                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
1501         },
1502         [ C(OP_WRITE) ] = {
1503                 [ C(RESULT_ACCESS) ] = -1,
1504                 [ C(RESULT_MISS)   ] = -1,
1505         },
1506         [ C(OP_PREFETCH) ] = {
1507                 [ C(RESULT_ACCESS) ] = -1,
1508                 [ C(RESULT_MISS)   ] = -1,
1509         },
1510  },
1511 };
1512
1513 static __initconst const u64 atom_hw_cache_event_ids
1514                                 [PERF_COUNT_HW_CACHE_MAX]
1515                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1516                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1517 {
1518  [ C(L1D) ] = {
1519         [ C(OP_READ) ] = {
1520                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD               */
1521                 [ C(RESULT_MISS)   ] = 0,
1522         },
1523         [ C(OP_WRITE) ] = {
1524                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST               */
1525                 [ C(RESULT_MISS)   ] = 0,
1526         },
1527         [ C(OP_PREFETCH) ] = {
1528                 [ C(RESULT_ACCESS) ] = 0x0,
1529                 [ C(RESULT_MISS)   ] = 0,
1530         },
1531  },
1532  [ C(L1I ) ] = {
1533         [ C(OP_READ) ] = {
1534                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                  */
1535                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                 */
1536         },
1537         [ C(OP_WRITE) ] = {
1538                 [ C(RESULT_ACCESS) ] = -1,
1539                 [ C(RESULT_MISS)   ] = -1,
1540         },
1541         [ C(OP_PREFETCH) ] = {
1542                 [ C(RESULT_ACCESS) ] = 0,
1543                 [ C(RESULT_MISS)   ] = 0,
1544         },
1545  },
1546  [ C(LL  ) ] = {
1547         [ C(OP_READ) ] = {
1548                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
1549                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
1550         },
1551         [ C(OP_WRITE) ] = {
1552                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
1553                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
1554         },
1555         [ C(OP_PREFETCH) ] = {
1556                 [ C(RESULT_ACCESS) ] = 0,
1557                 [ C(RESULT_MISS)   ] = 0,
1558         },
1559  },
1560  [ C(DTLB) ] = {
1561         [ C(OP_READ) ] = {
1562                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI  (alias) */
1563                 [ C(RESULT_MISS)   ] = 0x0508, /* DTLB_MISSES.MISS_LD        */
1564         },
1565         [ C(OP_WRITE) ] = {
1566                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI  (alias) */
1567                 [ C(RESULT_MISS)   ] = 0x0608, /* DTLB_MISSES.MISS_ST        */
1568         },
1569         [ C(OP_PREFETCH) ] = {
1570                 [ C(RESULT_ACCESS) ] = 0,
1571                 [ C(RESULT_MISS)   ] = 0,
1572         },
1573  },
1574  [ C(ITLB) ] = {
1575         [ C(OP_READ) ] = {
1576                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
1577                 [ C(RESULT_MISS)   ] = 0x0282, /* ITLB.MISSES                */
1578         },
1579         [ C(OP_WRITE) ] = {
1580                 [ C(RESULT_ACCESS) ] = -1,
1581                 [ C(RESULT_MISS)   ] = -1,
1582         },
1583         [ C(OP_PREFETCH) ] = {
1584                 [ C(RESULT_ACCESS) ] = -1,
1585                 [ C(RESULT_MISS)   ] = -1,
1586         },
1587  },
1588  [ C(BPU ) ] = {
1589         [ C(OP_READ) ] = {
1590                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
1591                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
1592         },
1593         [ C(OP_WRITE) ] = {
1594                 [ C(RESULT_ACCESS) ] = -1,
1595                 [ C(RESULT_MISS)   ] = -1,
1596         },
1597         [ C(OP_PREFETCH) ] = {
1598                 [ C(RESULT_ACCESS) ] = -1,
1599                 [ C(RESULT_MISS)   ] = -1,
1600         },
1601  },
1602 };
1603
1604 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c");
1605 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2");
1606 /* no_alloc_cycles.not_delivered */
1607 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm,
1608                "event=0xca,umask=0x50");
1609 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2");
1610 /* uops_retired.all */
1611 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm,
1612                "event=0xc2,umask=0x10");
1613 /* uops_retired.all */
1614 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm,
1615                "event=0xc2,umask=0x10");
1616
1617 static struct attribute *slm_events_attrs[] = {
1618         EVENT_PTR(td_total_slots_slm),
1619         EVENT_PTR(td_total_slots_scale_slm),
1620         EVENT_PTR(td_fetch_bubbles_slm),
1621         EVENT_PTR(td_fetch_bubbles_scale_slm),
1622         EVENT_PTR(td_slots_issued_slm),
1623         EVENT_PTR(td_slots_retired_slm),
1624         NULL
1625 };
1626
1627 static struct extra_reg intel_slm_extra_regs[] __read_mostly =
1628 {
1629         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1630         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0),
1631         INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1),
1632         EVENT_EXTRA_END
1633 };
1634
1635 #define SLM_DMND_READ           SNB_DMND_DATA_RD
1636 #define SLM_DMND_WRITE          SNB_DMND_RFO
1637 #define SLM_DMND_PREFETCH       (SNB_PF_DATA_RD|SNB_PF_RFO)
1638
1639 #define SLM_SNP_ANY             (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1640 #define SLM_LLC_ACCESS          SNB_RESP_ANY
1641 #define SLM_LLC_MISS            (SLM_SNP_ANY|SNB_NON_DRAM)
1642
1643 static __initconst const u64 slm_hw_cache_extra_regs
1644                                 [PERF_COUNT_HW_CACHE_MAX]
1645                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1646                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1647 {
1648  [ C(LL  ) ] = {
1649         [ C(OP_READ) ] = {
1650                 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS,
1651                 [ C(RESULT_MISS)   ] = 0,
1652         },
1653         [ C(OP_WRITE) ] = {
1654                 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS,
1655                 [ C(RESULT_MISS)   ] = SLM_DMND_WRITE|SLM_LLC_MISS,
1656         },
1657         [ C(OP_PREFETCH) ] = {
1658                 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS,
1659                 [ C(RESULT_MISS)   ] = SLM_DMND_PREFETCH|SLM_LLC_MISS,
1660         },
1661  },
1662 };
1663
1664 static __initconst const u64 slm_hw_cache_event_ids
1665                                 [PERF_COUNT_HW_CACHE_MAX]
1666                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1667                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1668 {
1669  [ C(L1D) ] = {
1670         [ C(OP_READ) ] = {
1671                 [ C(RESULT_ACCESS) ] = 0,
1672                 [ C(RESULT_MISS)   ] = 0x0104, /* LD_DCU_MISS */
1673         },
1674         [ C(OP_WRITE) ] = {
1675                 [ C(RESULT_ACCESS) ] = 0,
1676                 [ C(RESULT_MISS)   ] = 0,
1677         },
1678         [ C(OP_PREFETCH) ] = {
1679                 [ C(RESULT_ACCESS) ] = 0,
1680                 [ C(RESULT_MISS)   ] = 0,
1681         },
1682  },
1683  [ C(L1I ) ] = {
1684         [ C(OP_READ) ] = {
1685                 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */
1686                 [ C(RESULT_MISS)   ] = 0x0280, /* ICACGE.MISSES */
1687         },
1688         [ C(OP_WRITE) ] = {
1689                 [ C(RESULT_ACCESS) ] = -1,
1690                 [ C(RESULT_MISS)   ] = -1,
1691         },
1692         [ C(OP_PREFETCH) ] = {
1693                 [ C(RESULT_ACCESS) ] = 0,
1694                 [ C(RESULT_MISS)   ] = 0,
1695         },
1696  },
1697  [ C(LL  ) ] = {
1698         [ C(OP_READ) ] = {
1699                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1700                 [ C(RESULT_ACCESS) ] = 0x01b7,
1701                 [ C(RESULT_MISS)   ] = 0,
1702         },
1703         [ C(OP_WRITE) ] = {
1704                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1705                 [ C(RESULT_ACCESS) ] = 0x01b7,
1706                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1707                 [ C(RESULT_MISS)   ] = 0x01b7,
1708         },
1709         [ C(OP_PREFETCH) ] = {
1710                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1711                 [ C(RESULT_ACCESS) ] = 0x01b7,
1712                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1713                 [ C(RESULT_MISS)   ] = 0x01b7,
1714         },
1715  },
1716  [ C(DTLB) ] = {
1717         [ C(OP_READ) ] = {
1718                 [ C(RESULT_ACCESS) ] = 0,
1719                 [ C(RESULT_MISS)   ] = 0x0804, /* LD_DTLB_MISS */
1720         },
1721         [ C(OP_WRITE) ] = {
1722                 [ C(RESULT_ACCESS) ] = 0,
1723                 [ C(RESULT_MISS)   ] = 0,
1724         },
1725         [ C(OP_PREFETCH) ] = {
1726                 [ C(RESULT_ACCESS) ] = 0,
1727                 [ C(RESULT_MISS)   ] = 0,
1728         },
1729  },
1730  [ C(ITLB) ] = {
1731         [ C(OP_READ) ] = {
1732                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1733                 [ C(RESULT_MISS)   ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1734         },
1735         [ C(OP_WRITE) ] = {
1736                 [ C(RESULT_ACCESS) ] = -1,
1737                 [ C(RESULT_MISS)   ] = -1,
1738         },
1739         [ C(OP_PREFETCH) ] = {
1740                 [ C(RESULT_ACCESS) ] = -1,
1741                 [ C(RESULT_MISS)   ] = -1,
1742         },
1743  },
1744  [ C(BPU ) ] = {
1745         [ C(OP_READ) ] = {
1746                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1747                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1748         },
1749         [ C(OP_WRITE) ] = {
1750                 [ C(RESULT_ACCESS) ] = -1,
1751                 [ C(RESULT_MISS)   ] = -1,
1752         },
1753         [ C(OP_PREFETCH) ] = {
1754                 [ C(RESULT_ACCESS) ] = -1,
1755                 [ C(RESULT_MISS)   ] = -1,
1756         },
1757  },
1758 };
1759
1760 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c");
1761 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3");
1762 /* UOPS_NOT_DELIVERED.ANY */
1763 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c");
1764 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */
1765 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02");
1766 /* UOPS_RETIRED.ANY */
1767 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2");
1768 /* UOPS_ISSUED.ANY */
1769 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e");
1770
1771 static struct attribute *glm_events_attrs[] = {
1772         EVENT_PTR(td_total_slots_glm),
1773         EVENT_PTR(td_total_slots_scale_glm),
1774         EVENT_PTR(td_fetch_bubbles_glm),
1775         EVENT_PTR(td_recovery_bubbles_glm),
1776         EVENT_PTR(td_slots_issued_glm),
1777         EVENT_PTR(td_slots_retired_glm),
1778         NULL
1779 };
1780
1781 static struct extra_reg intel_glm_extra_regs[] __read_mostly = {
1782         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1783         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0),
1784         INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1),
1785         EVENT_EXTRA_END
1786 };
1787
1788 #define GLM_DEMAND_DATA_RD              BIT_ULL(0)
1789 #define GLM_DEMAND_RFO                  BIT_ULL(1)
1790 #define GLM_ANY_RESPONSE                BIT_ULL(16)
1791 #define GLM_SNP_NONE_OR_MISS            BIT_ULL(33)
1792 #define GLM_DEMAND_READ                 GLM_DEMAND_DATA_RD
1793 #define GLM_DEMAND_WRITE                GLM_DEMAND_RFO
1794 #define GLM_DEMAND_PREFETCH             (SNB_PF_DATA_RD|SNB_PF_RFO)
1795 #define GLM_LLC_ACCESS                  GLM_ANY_RESPONSE
1796 #define GLM_SNP_ANY                     (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM)
1797 #define GLM_LLC_MISS                    (GLM_SNP_ANY|SNB_NON_DRAM)
1798
1799 static __initconst const u64 glm_hw_cache_event_ids
1800                                 [PERF_COUNT_HW_CACHE_MAX]
1801                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1802                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1803         [C(L1D)] = {
1804                 [C(OP_READ)] = {
1805                         [C(RESULT_ACCESS)]      = 0x81d0,       /* MEM_UOPS_RETIRED.ALL_LOADS */
1806                         [C(RESULT_MISS)]        = 0x0,
1807                 },
1808                 [C(OP_WRITE)] = {
1809                         [C(RESULT_ACCESS)]      = 0x82d0,       /* MEM_UOPS_RETIRED.ALL_STORES */
1810                         [C(RESULT_MISS)]        = 0x0,
1811                 },
1812                 [C(OP_PREFETCH)] = {
1813                         [C(RESULT_ACCESS)]      = 0x0,
1814                         [C(RESULT_MISS)]        = 0x0,
1815                 },
1816         },
1817         [C(L1I)] = {
1818                 [C(OP_READ)] = {
1819                         [C(RESULT_ACCESS)]      = 0x0380,       /* ICACHE.ACCESSES */
1820                         [C(RESULT_MISS)]        = 0x0280,       /* ICACHE.MISSES */
1821                 },
1822                 [C(OP_WRITE)] = {
1823                         [C(RESULT_ACCESS)]      = -1,
1824                         [C(RESULT_MISS)]        = -1,
1825                 },
1826                 [C(OP_PREFETCH)] = {
1827                         [C(RESULT_ACCESS)]      = 0x0,
1828                         [C(RESULT_MISS)]        = 0x0,
1829                 },
1830         },
1831         [C(LL)] = {
1832                 [C(OP_READ)] = {
1833                         [C(RESULT_ACCESS)]      = 0x1b7,        /* OFFCORE_RESPONSE */
1834                         [C(RESULT_MISS)]        = 0x1b7,        /* OFFCORE_RESPONSE */
1835                 },
1836                 [C(OP_WRITE)] = {
1837                         [C(RESULT_ACCESS)]      = 0x1b7,        /* OFFCORE_RESPONSE */
1838                         [C(RESULT_MISS)]        = 0x1b7,        /* OFFCORE_RESPONSE */
1839                 },
1840                 [C(OP_PREFETCH)] = {
1841                         [C(RESULT_ACCESS)]      = 0x1b7,        /* OFFCORE_RESPONSE */
1842                         [C(RESULT_MISS)]        = 0x1b7,        /* OFFCORE_RESPONSE */
1843                 },
1844         },
1845         [C(DTLB)] = {
1846                 [C(OP_READ)] = {
1847                         [C(RESULT_ACCESS)]      = 0x81d0,       /* MEM_UOPS_RETIRED.ALL_LOADS */
1848                         [C(RESULT_MISS)]        = 0x0,
1849                 },
1850                 [C(OP_WRITE)] = {
1851                         [C(RESULT_ACCESS)]      = 0x82d0,       /* MEM_UOPS_RETIRED.ALL_STORES */
1852                         [C(RESULT_MISS)]        = 0x0,
1853                 },
1854                 [C(OP_PREFETCH)] = {
1855                         [C(RESULT_ACCESS)]      = 0x0,
1856                         [C(RESULT_MISS)]        = 0x0,
1857                 },
1858         },
1859         [C(ITLB)] = {
1860                 [C(OP_READ)] = {
1861                         [C(RESULT_ACCESS)]      = 0x00c0,       /* INST_RETIRED.ANY_P */
1862                         [C(RESULT_MISS)]        = 0x0481,       /* ITLB.MISS */
1863                 },
1864                 [C(OP_WRITE)] = {
1865                         [C(RESULT_ACCESS)]      = -1,
1866                         [C(RESULT_MISS)]        = -1,
1867                 },
1868                 [C(OP_PREFETCH)] = {
1869                         [C(RESULT_ACCESS)]      = -1,
1870                         [C(RESULT_MISS)]        = -1,
1871                 },
1872         },
1873         [C(BPU)] = {
1874                 [C(OP_READ)] = {
1875                         [C(RESULT_ACCESS)]      = 0x00c4,       /* BR_INST_RETIRED.ALL_BRANCHES */
1876                         [C(RESULT_MISS)]        = 0x00c5,       /* BR_MISP_RETIRED.ALL_BRANCHES */
1877                 },
1878                 [C(OP_WRITE)] = {
1879                         [C(RESULT_ACCESS)]      = -1,
1880                         [C(RESULT_MISS)]        = -1,
1881                 },
1882                 [C(OP_PREFETCH)] = {
1883                         [C(RESULT_ACCESS)]      = -1,
1884                         [C(RESULT_MISS)]        = -1,
1885                 },
1886         },
1887 };
1888
1889 static __initconst const u64 glm_hw_cache_extra_regs
1890                                 [PERF_COUNT_HW_CACHE_MAX]
1891                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1892                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1893         [C(LL)] = {
1894                 [C(OP_READ)] = {
1895                         [C(RESULT_ACCESS)]      = GLM_DEMAND_READ|
1896                                                   GLM_LLC_ACCESS,
1897                         [C(RESULT_MISS)]        = GLM_DEMAND_READ|
1898                                                   GLM_LLC_MISS,
1899                 },
1900                 [C(OP_WRITE)] = {
1901                         [C(RESULT_ACCESS)]      = GLM_DEMAND_WRITE|
1902                                                   GLM_LLC_ACCESS,
1903                         [C(RESULT_MISS)]        = GLM_DEMAND_WRITE|
1904                                                   GLM_LLC_MISS,
1905                 },
1906                 [C(OP_PREFETCH)] = {
1907                         [C(RESULT_ACCESS)]      = GLM_DEMAND_PREFETCH|
1908                                                   GLM_LLC_ACCESS,
1909                         [C(RESULT_MISS)]        = GLM_DEMAND_PREFETCH|
1910                                                   GLM_LLC_MISS,
1911                 },
1912         },
1913 };
1914
1915 static __initconst const u64 glp_hw_cache_event_ids
1916                                 [PERF_COUNT_HW_CACHE_MAX]
1917                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1918                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1919         [C(L1D)] = {
1920                 [C(OP_READ)] = {
1921                         [C(RESULT_ACCESS)]      = 0x81d0,       /* MEM_UOPS_RETIRED.ALL_LOADS */
1922                         [C(RESULT_MISS)]        = 0x0,
1923                 },
1924                 [C(OP_WRITE)] = {
1925                         [C(RESULT_ACCESS)]      = 0x82d0,       /* MEM_UOPS_RETIRED.ALL_STORES */
1926                         [C(RESULT_MISS)]        = 0x0,
1927                 },
1928                 [C(OP_PREFETCH)] = {
1929                         [C(RESULT_ACCESS)]      = 0x0,
1930                         [C(RESULT_MISS)]        = 0x0,
1931                 },
1932         },
1933         [C(L1I)] = {
1934                 [C(OP_READ)] = {
1935                         [C(RESULT_ACCESS)]      = 0x0380,       /* ICACHE.ACCESSES */
1936                         [C(RESULT_MISS)]        = 0x0280,       /* ICACHE.MISSES */
1937                 },
1938                 [C(OP_WRITE)] = {
1939                         [C(RESULT_ACCESS)]      = -1,
1940                         [C(RESULT_MISS)]        = -1,
1941                 },
1942                 [C(OP_PREFETCH)] = {
1943                         [C(RESULT_ACCESS)]      = 0x0,
1944                         [C(RESULT_MISS)]        = 0x0,
1945                 },
1946         },
1947         [C(LL)] = {
1948                 [C(OP_READ)] = {
1949                         [C(RESULT_ACCESS)]      = 0x1b7,        /* OFFCORE_RESPONSE */
1950                         [C(RESULT_MISS)]        = 0x1b7,        /* OFFCORE_RESPONSE */
1951                 },
1952                 [C(OP_WRITE)] = {
1953                         [C(RESULT_ACCESS)]      = 0x1b7,        /* OFFCORE_RESPONSE */
1954                         [C(RESULT_MISS)]        = 0x1b7,        /* OFFCORE_RESPONSE */
1955                 },
1956                 [C(OP_PREFETCH)] = {
1957                         [C(RESULT_ACCESS)]      = 0x0,
1958                         [C(RESULT_MISS)]        = 0x0,
1959                 },
1960         },
1961         [C(DTLB)] = {
1962                 [C(OP_READ)] = {
1963                         [C(RESULT_ACCESS)]      = 0x81d0,       /* MEM_UOPS_RETIRED.ALL_LOADS */
1964                         [C(RESULT_MISS)]        = 0xe08,        /* DTLB_LOAD_MISSES.WALK_COMPLETED */
1965                 },
1966                 [C(OP_WRITE)] = {
1967                         [C(RESULT_ACCESS)]      = 0x82d0,       /* MEM_UOPS_RETIRED.ALL_STORES */
1968                         [C(RESULT_MISS)]        = 0xe49,        /* DTLB_STORE_MISSES.WALK_COMPLETED */
1969                 },
1970                 [C(OP_PREFETCH)] = {
1971                         [C(RESULT_ACCESS)]      = 0x0,
1972                         [C(RESULT_MISS)]        = 0x0,
1973                 },
1974         },
1975         [C(ITLB)] = {
1976                 [C(OP_READ)] = {
1977                         [C(RESULT_ACCESS)]      = 0x00c0,       /* INST_RETIRED.ANY_P */
1978                         [C(RESULT_MISS)]        = 0x0481,       /* ITLB.MISS */
1979                 },
1980                 [C(OP_WRITE)] = {
1981                         [C(RESULT_ACCESS)]      = -1,
1982                         [C(RESULT_MISS)]        = -1,
1983                 },
1984                 [C(OP_PREFETCH)] = {
1985                         [C(RESULT_ACCESS)]      = -1,
1986                         [C(RESULT_MISS)]        = -1,
1987                 },
1988         },
1989         [C(BPU)] = {
1990                 [C(OP_READ)] = {
1991                         [C(RESULT_ACCESS)]      = 0x00c4,       /* BR_INST_RETIRED.ALL_BRANCHES */
1992                         [C(RESULT_MISS)]        = 0x00c5,       /* BR_MISP_RETIRED.ALL_BRANCHES */
1993                 },
1994                 [C(OP_WRITE)] = {
1995                         [C(RESULT_ACCESS)]      = -1,
1996                         [C(RESULT_MISS)]        = -1,
1997                 },
1998                 [C(OP_PREFETCH)] = {
1999                         [C(RESULT_ACCESS)]      = -1,
2000                         [C(RESULT_MISS)]        = -1,
2001                 },
2002         },
2003 };
2004
2005 static __initconst const u64 glp_hw_cache_extra_regs
2006                                 [PERF_COUNT_HW_CACHE_MAX]
2007                                 [PERF_COUNT_HW_CACHE_OP_MAX]
2008                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2009         [C(LL)] = {
2010                 [C(OP_READ)] = {
2011                         [C(RESULT_ACCESS)]      = GLM_DEMAND_READ|
2012                                                   GLM_LLC_ACCESS,
2013                         [C(RESULT_MISS)]        = GLM_DEMAND_READ|
2014                                                   GLM_LLC_MISS,
2015                 },
2016                 [C(OP_WRITE)] = {
2017                         [C(RESULT_ACCESS)]      = GLM_DEMAND_WRITE|
2018                                                   GLM_LLC_ACCESS,
2019                         [C(RESULT_MISS)]        = GLM_DEMAND_WRITE|
2020                                                   GLM_LLC_MISS,
2021                 },
2022                 [C(OP_PREFETCH)] = {
2023                         [C(RESULT_ACCESS)]      = 0x0,
2024                         [C(RESULT_MISS)]        = 0x0,
2025                 },
2026         },
2027 };
2028
2029 #define TNT_LOCAL_DRAM                  BIT_ULL(26)
2030 #define TNT_DEMAND_READ                 GLM_DEMAND_DATA_RD
2031 #define TNT_DEMAND_WRITE                GLM_DEMAND_RFO
2032 #define TNT_LLC_ACCESS                  GLM_ANY_RESPONSE
2033 #define TNT_SNP_ANY                     (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \
2034                                          SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM)
2035 #define TNT_LLC_MISS                    (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM)
2036
2037 static __initconst const u64 tnt_hw_cache_extra_regs
2038                                 [PERF_COUNT_HW_CACHE_MAX]
2039                                 [PERF_COUNT_HW_CACHE_OP_MAX]
2040                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2041         [C(LL)] = {
2042                 [C(OP_READ)] = {
2043                         [C(RESULT_ACCESS)]      = TNT_DEMAND_READ|
2044                                                   TNT_LLC_ACCESS,
2045                         [C(RESULT_MISS)]        = TNT_DEMAND_READ|
2046                                                   TNT_LLC_MISS,
2047                 },
2048                 [C(OP_WRITE)] = {
2049                         [C(RESULT_ACCESS)]      = TNT_DEMAND_WRITE|
2050                                                   TNT_LLC_ACCESS,
2051                         [C(RESULT_MISS)]        = TNT_DEMAND_WRITE|
2052                                                   TNT_LLC_MISS,
2053                 },
2054                 [C(OP_PREFETCH)] = {
2055                         [C(RESULT_ACCESS)]      = 0x0,
2056                         [C(RESULT_MISS)]        = 0x0,
2057                 },
2058         },
2059 };
2060
2061 EVENT_ATTR_STR(topdown-fe-bound,       td_fe_bound_tnt,        "event=0x71,umask=0x0");
2062 EVENT_ATTR_STR(topdown-retiring,       td_retiring_tnt,        "event=0xc2,umask=0x0");
2063 EVENT_ATTR_STR(topdown-bad-spec,       td_bad_spec_tnt,        "event=0x73,umask=0x6");
2064 EVENT_ATTR_STR(topdown-be-bound,       td_be_bound_tnt,        "event=0x74,umask=0x0");
2065
2066 static struct attribute *tnt_events_attrs[] = {
2067         EVENT_PTR(td_fe_bound_tnt),
2068         EVENT_PTR(td_retiring_tnt),
2069         EVENT_PTR(td_bad_spec_tnt),
2070         EVENT_PTR(td_be_bound_tnt),
2071         NULL,
2072 };
2073
2074 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = {
2075         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
2076         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0),
2077         INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1),
2078         EVENT_EXTRA_END
2079 };
2080
2081 static struct extra_reg intel_grt_extra_regs[] __read_mostly = {
2082         /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
2083         INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0),
2084         INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1),
2085         INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0),
2086         EVENT_EXTRA_END
2087 };
2088
2089 #define KNL_OT_L2_HITE          BIT_ULL(19) /* Other Tile L2 Hit */
2090 #define KNL_OT_L2_HITF          BIT_ULL(20) /* Other Tile L2 Hit */
2091 #define KNL_MCDRAM_LOCAL        BIT_ULL(21)
2092 #define KNL_MCDRAM_FAR          BIT_ULL(22)
2093 #define KNL_DDR_LOCAL           BIT_ULL(23)
2094 #define KNL_DDR_FAR             BIT_ULL(24)
2095 #define KNL_DRAM_ANY            (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \
2096                                     KNL_DDR_LOCAL | KNL_DDR_FAR)
2097 #define KNL_L2_READ             SLM_DMND_READ
2098 #define KNL_L2_WRITE            SLM_DMND_WRITE
2099 #define KNL_L2_PREFETCH         SLM_DMND_PREFETCH
2100 #define KNL_L2_ACCESS           SLM_LLC_ACCESS
2101 #define KNL_L2_MISS             (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \
2102                                    KNL_DRAM_ANY | SNB_SNP_ANY | \
2103                                                   SNB_NON_DRAM)
2104
2105 static __initconst const u64 knl_hw_cache_extra_regs
2106                                 [PERF_COUNT_HW_CACHE_MAX]
2107                                 [PERF_COUNT_HW_CACHE_OP_MAX]
2108                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2109         [C(LL)] = {
2110                 [C(OP_READ)] = {
2111                         [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS,
2112                         [C(RESULT_MISS)]   = 0,
2113                 },
2114                 [C(OP_WRITE)] = {
2115                         [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS,
2116                         [C(RESULT_MISS)]   = KNL_L2_WRITE | KNL_L2_MISS,
2117                 },
2118                 [C(OP_PREFETCH)] = {
2119                         [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS,
2120                         [C(RESULT_MISS)]   = KNL_L2_PREFETCH | KNL_L2_MISS,
2121                 },
2122         },
2123 };
2124
2125 /*
2126  * Used from PMIs where the LBRs are already disabled.
2127  *
2128  * This function could be called consecutively. It is required to remain in
2129  * disabled state if called consecutively.
2130  *
2131  * During consecutive calls, the same disable value will be written to related
2132  * registers, so the PMU state remains unchanged.
2133  *
2134  * intel_bts events don't coexist with intel PMU's BTS events because of
2135  * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them
2136  * disabled around intel PMU's event batching etc, only inside the PMI handler.
2137  *
2138  * Avoid PEBS_ENABLE MSR access in PMIs.
2139  * The GLOBAL_CTRL has been disabled. All the counters do not count anymore.
2140  * It doesn't matter if the PEBS is enabled or not.
2141  * Usually, the PEBS status are not changed in PMIs. It's unnecessary to
2142  * access PEBS_ENABLE MSR in disable_all()/enable_all().
2143  * However, there are some cases which may change PEBS status, e.g. PMI
2144  * throttle. The PEBS_ENABLE should be updated where the status changes.
2145  */
2146 static void __intel_pmu_disable_all(void)
2147 {
2148         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2149
2150         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2151
2152         if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
2153                 intel_pmu_disable_bts();
2154 }
2155
2156 static void intel_pmu_disable_all(void)
2157 {
2158         __intel_pmu_disable_all();
2159         intel_pmu_pebs_disable_all();
2160         intel_pmu_lbr_disable_all();
2161 }
2162
2163 static void __intel_pmu_enable_all(int added, bool pmi)
2164 {
2165         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2166         u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
2167
2168         intel_pmu_lbr_enable_all(pmi);
2169         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
2170                intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
2171
2172         if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
2173                 struct perf_event *event =
2174                         cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
2175
2176                 if (WARN_ON_ONCE(!event))
2177                         return;
2178
2179                 intel_pmu_enable_bts(event->hw.config);
2180         }
2181 }
2182
2183 static void intel_pmu_enable_all(int added)
2184 {
2185         intel_pmu_pebs_enable_all();
2186         __intel_pmu_enable_all(added, false);
2187 }
2188
2189 /*
2190  * Workaround for:
2191  *   Intel Errata AAK100 (model 26)
2192  *   Intel Errata AAP53  (model 30)
2193  *   Intel Errata BD53   (model 44)
2194  *
2195  * The official story:
2196  *   These chips need to be 'reset' when adding counters by programming the
2197  *   magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
2198  *   in sequence on the same PMC or on different PMCs.
2199  *
2200  * In practice it appears some of these events do in fact count, and
2201  * we need to program all 4 events.
2202  */
2203 static void intel_pmu_nhm_workaround(void)
2204 {
2205         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2206         static const unsigned long nhm_magic[4] = {
2207                 0x4300B5,
2208                 0x4300D2,
2209                 0x4300B1,
2210                 0x4300B1
2211         };
2212         struct perf_event *event;
2213         int i;
2214
2215         /*
2216          * The Errata requires below steps:
2217          * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
2218          * 2) Configure 4 PERFEVTSELx with the magic events and clear
2219          *    the corresponding PMCx;
2220          * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
2221          * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
2222          * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
2223          */
2224
2225         /*
2226          * The real steps we choose are a little different from above.
2227          * A) To reduce MSR operations, we don't run step 1) as they
2228          *    are already cleared before this function is called;
2229          * B) Call x86_perf_event_update to save PMCx before configuring
2230          *    PERFEVTSELx with magic number;
2231          * C) With step 5), we do clear only when the PERFEVTSELx is
2232          *    not used currently.
2233          * D) Call x86_perf_event_set_period to restore PMCx;
2234          */
2235
2236         /* We always operate 4 pairs of PERF Counters */
2237         for (i = 0; i < 4; i++) {
2238                 event = cpuc->events[i];
2239                 if (event)
2240                         x86_perf_event_update(event);
2241         }
2242
2243         for (i = 0; i < 4; i++) {
2244                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
2245                 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
2246         }
2247
2248         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
2249         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
2250
2251         for (i = 0; i < 4; i++) {
2252                 event = cpuc->events[i];
2253
2254                 if (event) {
2255                         x86_perf_event_set_period(event);
2256                         __x86_pmu_enable_event(&event->hw,
2257                                         ARCH_PERFMON_EVENTSEL_ENABLE);
2258                 } else
2259                         wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
2260         }
2261 }
2262
2263 static void intel_pmu_nhm_enable_all(int added)
2264 {
2265         if (added)
2266                 intel_pmu_nhm_workaround();
2267         intel_pmu_enable_all(added);
2268 }
2269
2270 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on)
2271 {
2272         u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0;
2273
2274         if (cpuc->tfa_shadow != val) {
2275                 cpuc->tfa_shadow = val;
2276                 wrmsrl(MSR_TSX_FORCE_ABORT, val);
2277         }
2278 }
2279
2280 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2281 {
2282         /*
2283          * We're going to use PMC3, make sure TFA is set before we touch it.
2284          */
2285         if (cntr == 3)
2286                 intel_set_tfa(cpuc, true);
2287 }
2288
2289 static void intel_tfa_pmu_enable_all(int added)
2290 {
2291         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2292
2293         /*
2294          * If we find PMC3 is no longer used when we enable the PMU, we can
2295          * clear TFA.
2296          */
2297         if (!test_bit(3, cpuc->active_mask))
2298                 intel_set_tfa(cpuc, false);
2299
2300         intel_pmu_enable_all(added);
2301 }
2302
2303 static inline u64 intel_pmu_get_status(void)
2304 {
2305         u64 status;
2306
2307         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
2308
2309         return status;
2310 }
2311
2312 static inline void intel_pmu_ack_status(u64 ack)
2313 {
2314         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
2315 }
2316
2317 static inline bool event_is_checkpointed(struct perf_event *event)
2318 {
2319         return unlikely(event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0;
2320 }
2321
2322 static inline void intel_set_masks(struct perf_event *event, int idx)
2323 {
2324         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2325
2326         if (event->attr.exclude_host)
2327                 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask);
2328         if (event->attr.exclude_guest)
2329                 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask);
2330         if (event_is_checkpointed(event))
2331                 __set_bit(idx, (unsigned long *)&cpuc->intel_cp_status);
2332 }
2333
2334 static inline void intel_clear_masks(struct perf_event *event, int idx)
2335 {
2336         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2337
2338         __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask);
2339         __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask);
2340         __clear_bit(idx, (unsigned long *)&cpuc->intel_cp_status);
2341 }
2342
2343 static void intel_pmu_disable_fixed(struct perf_event *event)
2344 {
2345         struct hw_perf_event *hwc = &event->hw;
2346         u64 ctrl_val, mask;
2347         int idx = hwc->idx;
2348
2349         if (is_topdown_idx(idx)) {
2350                 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2351
2352                 /*
2353                  * When there are other active TopDown events,
2354                  * don't disable the fixed counter 3.
2355                  */
2356                 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx))
2357                         return;
2358                 idx = INTEL_PMC_IDX_FIXED_SLOTS;
2359         }
2360
2361         intel_clear_masks(event, idx);
2362
2363         mask = 0xfULL << ((idx - INTEL_PMC_IDX_FIXED) * 4);
2364         rdmsrl(hwc->config_base, ctrl_val);
2365         ctrl_val &= ~mask;
2366         wrmsrl(hwc->config_base, ctrl_val);
2367 }
2368
2369 static void intel_pmu_disable_event(struct perf_event *event)
2370 {
2371         struct hw_perf_event *hwc = &event->hw;
2372         int idx = hwc->idx;
2373
2374         switch (idx) {
2375         case 0 ... INTEL_PMC_IDX_FIXED - 1:
2376                 intel_clear_masks(event, idx);
2377                 x86_pmu_disable_event(event);
2378                 break;
2379         case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1:
2380         case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
2381                 intel_pmu_disable_fixed(event);
2382                 break;
2383         case INTEL_PMC_IDX_FIXED_BTS:
2384                 intel_pmu_disable_bts();
2385                 intel_pmu_drain_bts_buffer();
2386                 return;
2387         case INTEL_PMC_IDX_FIXED_VLBR:
2388                 intel_clear_masks(event, idx);
2389                 break;
2390         default:
2391                 intel_clear_masks(event, idx);
2392                 pr_warn("Failed to disable the event with invalid index %d\n",
2393                         idx);
2394                 return;
2395         }
2396
2397         /*
2398          * Needs to be called after x86_pmu_disable_event,
2399          * so we don't trigger the event without PEBS bit set.
2400          */
2401         if (unlikely(event->attr.precise_ip))
2402                 intel_pmu_pebs_disable(event);
2403 }
2404
2405 static void intel_pmu_del_event(struct perf_event *event)
2406 {
2407         if (needs_branch_stack(event))
2408                 intel_pmu_lbr_del(event);
2409         if (event->attr.precise_ip)
2410                 intel_pmu_pebs_del(event);
2411 }
2412
2413 static int icl_set_topdown_event_period(struct perf_event *event)
2414 {
2415         struct hw_perf_event *hwc = &event->hw;
2416         s64 left = local64_read(&hwc->period_left);
2417
2418         /*
2419          * The values in PERF_METRICS MSR are derived from fixed counter 3.
2420          * Software should start both registers, PERF_METRICS and fixed
2421          * counter 3, from zero.
2422          * Clear PERF_METRICS and Fixed counter 3 in initialization.
2423          * After that, both MSRs will be cleared for each read.
2424          * Don't need to clear them again.
2425          */
2426         if (left == x86_pmu.max_period) {
2427                 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
2428                 wrmsrl(MSR_PERF_METRICS, 0);
2429                 hwc->saved_slots = 0;
2430                 hwc->saved_metric = 0;
2431         }
2432
2433         if ((hwc->saved_slots) && is_slots_event(event)) {
2434                 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, hwc->saved_slots);
2435                 wrmsrl(MSR_PERF_METRICS, hwc->saved_metric);
2436         }
2437
2438         perf_event_update_userpage(event);
2439
2440         return 0;
2441 }
2442
2443 static int adl_set_topdown_event_period(struct perf_event *event)
2444 {
2445         struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
2446
2447         if (pmu->cpu_type != hybrid_big)
2448                 return 0;
2449
2450         return icl_set_topdown_event_period(event);
2451 }
2452
2453 static inline u64 icl_get_metrics_event_value(u64 metric, u64 slots, int idx)
2454 {
2455         u32 val;
2456
2457         /*
2458          * The metric is reported as an 8bit integer fraction
2459          * summing up to 0xff.
2460          * slots-in-metric = (Metric / 0xff) * slots
2461          */
2462         val = (metric >> ((idx - INTEL_PMC_IDX_METRIC_BASE) * 8)) & 0xff;
2463         return  mul_u64_u32_div(slots, val, 0xff);
2464 }
2465
2466 static u64 icl_get_topdown_value(struct perf_event *event,
2467                                        u64 slots, u64 metrics)
2468 {
2469         int idx = event->hw.idx;
2470         u64 delta;
2471
2472         if (is_metric_idx(idx))
2473                 delta = icl_get_metrics_event_value(metrics, slots, idx);
2474         else
2475                 delta = slots;
2476
2477         return delta;
2478 }
2479
2480 static void __icl_update_topdown_event(struct perf_event *event,
2481                                        u64 slots, u64 metrics,
2482                                        u64 last_slots, u64 last_metrics)
2483 {
2484         u64 delta, last = 0;
2485
2486         delta = icl_get_topdown_value(event, slots, metrics);
2487         if (last_slots)
2488                 last = icl_get_topdown_value(event, last_slots, last_metrics);
2489
2490         /*
2491          * The 8bit integer fraction of metric may be not accurate,
2492          * especially when the changes is very small.
2493          * For example, if only a few bad_spec happens, the fraction
2494          * may be reduced from 1 to 0. If so, the bad_spec event value
2495          * will be 0 which is definitely less than the last value.
2496          * Avoid update event->count for this case.
2497          */
2498         if (delta > last) {
2499                 delta -= last;
2500                 local64_add(delta, &event->count);
2501         }
2502 }
2503
2504 static void update_saved_topdown_regs(struct perf_event *event, u64 slots,
2505                                       u64 metrics, int metric_end)
2506 {
2507         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2508         struct perf_event *other;
2509         int idx;
2510
2511         event->hw.saved_slots = slots;
2512         event->hw.saved_metric = metrics;
2513
2514         for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
2515                 if (!is_topdown_idx(idx))
2516                         continue;
2517                 other = cpuc->events[idx];
2518                 other->hw.saved_slots = slots;
2519                 other->hw.saved_metric = metrics;
2520         }
2521 }
2522
2523 /*
2524  * Update all active Topdown events.
2525  *
2526  * The PERF_METRICS and Fixed counter 3 are read separately. The values may be
2527  * modify by a NMI. PMU has to be disabled before calling this function.
2528  */
2529
2530 static u64 intel_update_topdown_event(struct perf_event *event, int metric_end)
2531 {
2532         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2533         struct perf_event *other;
2534         u64 slots, metrics;
2535         bool reset = true;
2536         int idx;
2537
2538         /* read Fixed counter 3 */
2539         rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots);
2540         if (!slots)
2541                 return 0;
2542
2543         /* read PERF_METRICS */
2544         rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics);
2545
2546         for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
2547                 if (!is_topdown_idx(idx))
2548                         continue;
2549                 other = cpuc->events[idx];
2550                 __icl_update_topdown_event(other, slots, metrics,
2551                                            event ? event->hw.saved_slots : 0,
2552                                            event ? event->hw.saved_metric : 0);
2553         }
2554
2555         /*
2556          * Check and update this event, which may have been cleared
2557          * in active_mask e.g. x86_pmu_stop()
2558          */
2559         if (event && !test_bit(event->hw.idx, cpuc->active_mask)) {
2560                 __icl_update_topdown_event(event, slots, metrics,
2561                                            event->hw.saved_slots,
2562                                            event->hw.saved_metric);
2563
2564                 /*
2565                  * In x86_pmu_stop(), the event is cleared in active_mask first,
2566                  * then drain the delta, which indicates context switch for
2567                  * counting.
2568                  * Save metric and slots for context switch.
2569                  * Don't need to reset the PERF_METRICS and Fixed counter 3.
2570                  * Because the values will be restored in next schedule in.
2571                  */
2572                 update_saved_topdown_regs(event, slots, metrics, metric_end);
2573                 reset = false;
2574         }
2575
2576         if (reset) {
2577                 /* The fixed counter 3 has to be written before the PERF_METRICS. */
2578                 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
2579                 wrmsrl(MSR_PERF_METRICS, 0);
2580                 if (event)
2581                         update_saved_topdown_regs(event, 0, 0, metric_end);
2582         }
2583
2584         return slots;
2585 }
2586
2587 static u64 icl_update_topdown_event(struct perf_event *event)
2588 {
2589         return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE +
2590                                                  x86_pmu.num_topdown_events - 1);
2591 }
2592
2593 static u64 adl_update_topdown_event(struct perf_event *event)
2594 {
2595         struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
2596
2597         if (pmu->cpu_type != hybrid_big)
2598                 return 0;
2599
2600         return icl_update_topdown_event(event);
2601 }
2602
2603
2604 static void intel_pmu_read_topdown_event(struct perf_event *event)
2605 {
2606         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2607
2608         /* Only need to call update_topdown_event() once for group read. */
2609         if ((cpuc->txn_flags & PERF_PMU_TXN_READ) &&
2610             !is_slots_event(event))
2611                 return;
2612
2613         perf_pmu_disable(event->pmu);
2614         x86_pmu.update_topdown_event(event);
2615         perf_pmu_enable(event->pmu);
2616 }
2617
2618 static void intel_pmu_read_event(struct perf_event *event)
2619 {
2620         if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2621                 intel_pmu_auto_reload_read(event);
2622         else if (is_topdown_count(event) && x86_pmu.update_topdown_event)
2623                 intel_pmu_read_topdown_event(event);
2624         else
2625                 x86_perf_event_update(event);
2626 }
2627
2628 static void intel_pmu_enable_fixed(struct perf_event *event)
2629 {
2630         struct hw_perf_event *hwc = &event->hw;
2631         u64 ctrl_val, mask, bits = 0;
2632         int idx = hwc->idx;
2633
2634         if (is_topdown_idx(idx)) {
2635                 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2636                 /*
2637                  * When there are other active TopDown events,
2638                  * don't enable the fixed counter 3 again.
2639                  */
2640                 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx))
2641                         return;
2642
2643                 idx = INTEL_PMC_IDX_FIXED_SLOTS;
2644         }
2645
2646         intel_set_masks(event, idx);
2647
2648         /*
2649          * Enable IRQ generation (0x8), if not PEBS,
2650          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
2651          * if requested:
2652          */
2653         if (!event->attr.precise_ip)
2654                 bits |= 0x8;
2655         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
2656                 bits |= 0x2;
2657         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
2658                 bits |= 0x1;
2659
2660         /*
2661          * ANY bit is supported in v3 and up
2662          */
2663         if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
2664                 bits |= 0x4;
2665
2666         idx -= INTEL_PMC_IDX_FIXED;
2667         bits <<= (idx * 4);
2668         mask = 0xfULL << (idx * 4);
2669
2670         if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) {
2671                 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4);
2672                 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4);
2673         }
2674
2675         rdmsrl(hwc->config_base, ctrl_val);
2676         ctrl_val &= ~mask;
2677         ctrl_val |= bits;
2678         wrmsrl(hwc->config_base, ctrl_val);
2679 }
2680
2681 static void intel_pmu_enable_event(struct perf_event *event)
2682 {
2683         struct hw_perf_event *hwc = &event->hw;
2684         int idx = hwc->idx;
2685
2686         if (unlikely(event->attr.precise_ip))
2687                 intel_pmu_pebs_enable(event);
2688
2689         switch (idx) {
2690         case 0 ... INTEL_PMC_IDX_FIXED - 1:
2691                 intel_set_masks(event, idx);
2692                 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
2693                 break;
2694         case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1:
2695         case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
2696                 intel_pmu_enable_fixed(event);
2697                 break;
2698         case INTEL_PMC_IDX_FIXED_BTS:
2699                 if (!__this_cpu_read(cpu_hw_events.enabled))
2700                         return;
2701                 intel_pmu_enable_bts(hwc->config);
2702                 break;
2703         case INTEL_PMC_IDX_FIXED_VLBR:
2704                 intel_set_masks(event, idx);
2705                 break;
2706         default:
2707                 pr_warn("Failed to enable the event with invalid index %d\n",
2708                         idx);
2709         }
2710 }
2711
2712 static void intel_pmu_add_event(struct perf_event *event)
2713 {
2714         if (event->attr.precise_ip)
2715                 intel_pmu_pebs_add(event);
2716         if (needs_branch_stack(event))
2717                 intel_pmu_lbr_add(event);
2718 }
2719
2720 /*
2721  * Save and restart an expired event. Called by NMI contexts,
2722  * so it has to be careful about preempting normal event ops:
2723  */
2724 int intel_pmu_save_and_restart(struct perf_event *event)
2725 {
2726         x86_perf_event_update(event);
2727         /*
2728          * For a checkpointed counter always reset back to 0.  This
2729          * avoids a situation where the counter overflows, aborts the
2730          * transaction and is then set back to shortly before the
2731          * overflow, and overflows and aborts again.
2732          */
2733         if (unlikely(event_is_checkpointed(event))) {
2734                 /* No race with NMIs because the counter should not be armed */
2735                 wrmsrl(event->hw.event_base, 0);
2736                 local64_set(&event->hw.prev_count, 0);
2737         }
2738         return x86_perf_event_set_period(event);
2739 }
2740
2741 static void intel_pmu_reset(void)
2742 {
2743         struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2744         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2745         int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
2746         int num_counters = hybrid(cpuc->pmu, num_counters);
2747         unsigned long flags;
2748         int idx;
2749
2750         if (!num_counters)
2751                 return;
2752
2753         local_irq_save(flags);
2754
2755         pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
2756
2757         for (idx = 0; idx < num_counters; idx++) {
2758                 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
2759                 wrmsrl_safe(x86_pmu_event_addr(idx),  0ull);
2760         }
2761         for (idx = 0; idx < num_counters_fixed; idx++) {
2762                 if (fixed_counter_disabled(idx, cpuc->pmu))
2763                         continue;
2764                 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
2765         }
2766
2767         if (ds)
2768                 ds->bts_index = ds->bts_buffer_base;
2769
2770         /* Ack all overflows and disable fixed counters */
2771         if (x86_pmu.version >= 2) {
2772                 intel_pmu_ack_status(intel_pmu_get_status());
2773                 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2774         }
2775
2776         /* Reset LBRs and LBR freezing */
2777         if (x86_pmu.lbr_nr) {
2778                 update_debugctlmsr(get_debugctlmsr() &
2779                         ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR));
2780         }
2781
2782         local_irq_restore(flags);
2783 }
2784
2785 static int handle_pmi_common(struct pt_regs *regs, u64 status)
2786 {
2787         struct perf_sample_data data;
2788         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2789         int bit;
2790         int handled = 0;
2791         u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
2792
2793         inc_irq_stat(apic_perf_irqs);
2794
2795         /*
2796          * Ignore a range of extra bits in status that do not indicate
2797          * overflow by themselves.
2798          */
2799         status &= ~(GLOBAL_STATUS_COND_CHG |
2800                     GLOBAL_STATUS_ASIF |
2801                     GLOBAL_STATUS_LBRS_FROZEN);
2802         if (!status)
2803                 return 0;
2804         /*
2805          * In case multiple PEBS events are sampled at the same time,
2806          * it is possible to have GLOBAL_STATUS bit 62 set indicating
2807          * PEBS buffer overflow and also seeing at most 3 PEBS counters
2808          * having their bits set in the status register. This is a sign
2809          * that there was at least one PEBS record pending at the time
2810          * of the PMU interrupt. PEBS counters must only be processed
2811          * via the drain_pebs() calls and not via the regular sample
2812          * processing loop coming after that the function, otherwise
2813          * phony regular samples may be generated in the sampling buffer
2814          * not marked with the EXACT tag. Another possibility is to have
2815          * one PEBS event and at least one non-PEBS event which overflows
2816          * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will
2817          * not be set, yet the overflow status bit for the PEBS counter will
2818          * be on Skylake.
2819          *
2820          * To avoid this problem, we systematically ignore the PEBS-enabled
2821          * counters from the GLOBAL_STATUS mask and we always process PEBS
2822          * events via drain_pebs().
2823          */
2824         if (x86_pmu.flags & PMU_FL_PEBS_ALL)
2825                 status &= ~cpuc->pebs_enabled;
2826         else
2827                 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
2828
2829         /*
2830          * PEBS overflow sets bit 62 in the global status register
2831          */
2832         if (__test_and_clear_bit(GLOBAL_STATUS_BUFFER_OVF_BIT, (unsigned long *)&status)) {
2833                 u64 pebs_enabled = cpuc->pebs_enabled;
2834
2835                 handled++;
2836                 x86_pmu.drain_pebs(regs, &data);
2837                 status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
2838
2839                 /*
2840                  * PMI throttle may be triggered, which stops the PEBS event.
2841                  * Although cpuc->pebs_enabled is updated accordingly, the
2842                  * MSR_IA32_PEBS_ENABLE is not updated. Because the
2843                  * cpuc->enabled has been forced to 0 in PMI.
2844                  * Update the MSR if pebs_enabled is changed.
2845                  */
2846                 if (pebs_enabled != cpuc->pebs_enabled)
2847                         wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
2848         }
2849
2850         /*
2851          * Intel PT
2852          */
2853         if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) {
2854                 handled++;
2855                 if (unlikely(perf_guest_cbs && perf_guest_cbs->is_in_guest() &&
2856                         perf_guest_cbs->handle_intel_pt_intr))
2857                         perf_guest_cbs->handle_intel_pt_intr();
2858                 else
2859                         intel_pt_interrupt();
2860         }
2861
2862         /*
2863          * Intel Perf metrics
2864          */
2865         if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) {
2866                 handled++;
2867                 if (x86_pmu.update_topdown_event)
2868                         x86_pmu.update_topdown_event(NULL);
2869         }
2870
2871         /*
2872          * Checkpointed counters can lead to 'spurious' PMIs because the
2873          * rollback caused by the PMI will have cleared the overflow status
2874          * bit. Therefore always force probe these counters.
2875          */
2876         status |= cpuc->intel_cp_status;
2877
2878         for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
2879                 struct perf_event *event = cpuc->events[bit];
2880
2881                 handled++;
2882
2883                 if (!test_bit(bit, cpuc->active_mask))
2884                         continue;
2885
2886                 if (!intel_pmu_save_and_restart(event))
2887                         continue;
2888
2889                 perf_sample_data_init(&data, 0, event->hw.last_period);
2890
2891                 if (has_branch_stack(event))
2892                         data.br_stack = &cpuc->lbr_stack;
2893
2894                 if (perf_event_overflow(event, &data, regs))
2895                         x86_pmu_stop(event, 0);
2896         }
2897
2898         return handled;
2899 }
2900
2901 /*
2902  * This handler is triggered by the local APIC, so the APIC IRQ handling
2903  * rules apply:
2904  */
2905 static int intel_pmu_handle_irq(struct pt_regs *regs)
2906 {
2907         struct cpu_hw_events *cpuc;
2908         int loops;
2909         u64 status;
2910         int handled;
2911         int pmu_enabled;
2912
2913         cpuc = this_cpu_ptr(&cpu_hw_events);
2914
2915         /*
2916          * Save the PMU state.
2917          * It needs to be restored when leaving the handler.
2918          */
2919         pmu_enabled = cpuc->enabled;
2920         /*
2921          * No known reason to not always do late ACK,
2922          * but just in case do it opt-in.
2923          */
2924         if (!x86_pmu.late_ack)
2925                 apic_write(APIC_LVTPC, APIC_DM_NMI);
2926         intel_bts_disable_local();
2927         cpuc->enabled = 0;
2928         __intel_pmu_disable_all();
2929         handled = intel_pmu_drain_bts_buffer();
2930         handled += intel_bts_interrupt();
2931         status = intel_pmu_get_status();
2932         if (!status)
2933                 goto done;
2934
2935         loops = 0;
2936 again:
2937         intel_pmu_lbr_read();
2938         intel_pmu_ack_status(status);
2939         if (++loops > 100) {
2940                 static bool warned;
2941
2942                 if (!warned) {
2943                         WARN(1, "perfevents: irq loop stuck!\n");
2944                         perf_event_print_debug();
2945                         warned = true;
2946                 }
2947                 intel_pmu_reset();
2948                 goto done;
2949         }
2950
2951         handled += handle_pmi_common(regs, status);
2952
2953         /*
2954          * Repeat if there is more work to be done:
2955          */
2956         status = intel_pmu_get_status();
2957         if (status)
2958                 goto again;
2959
2960 done:
2961         /* Only restore PMU state when it's active. See x86_pmu_disable(). */
2962         cpuc->enabled = pmu_enabled;
2963         if (pmu_enabled)
2964                 __intel_pmu_enable_all(0, true);
2965         intel_bts_enable_local();
2966
2967         /*
2968          * Only unmask the NMI after the overflow counters
2969          * have been reset. This avoids spurious NMIs on
2970          * Haswell CPUs.
2971          */
2972         if (x86_pmu.late_ack)
2973                 apic_write(APIC_LVTPC, APIC_DM_NMI);
2974         return handled;
2975 }
2976
2977 static struct event_constraint *
2978 intel_bts_constraints(struct perf_event *event)
2979 {
2980         if (unlikely(intel_pmu_has_bts(event)))
2981                 return &bts_constraint;
2982
2983         return NULL;
2984 }
2985
2986 /*
2987  * Note: matches a fake event, like Fixed2.
2988  */
2989 static struct event_constraint *
2990 intel_vlbr_constraints(struct perf_event *event)
2991 {
2992         struct event_constraint *c = &vlbr_constraint;
2993
2994         if (unlikely(constraint_match(c, event->hw.config)))
2995                 return c;
2996
2997         return NULL;
2998 }
2999
3000 static int intel_alt_er(struct cpu_hw_events *cpuc,
3001                         int idx, u64 config)
3002 {
3003         struct extra_reg *extra_regs = hybrid(cpuc->pmu, extra_regs);
3004         int alt_idx = idx;
3005
3006         if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1))
3007                 return idx;
3008
3009         if (idx == EXTRA_REG_RSP_0)
3010                 alt_idx = EXTRA_REG_RSP_1;
3011
3012         if (idx == EXTRA_REG_RSP_1)
3013                 alt_idx = EXTRA_REG_RSP_0;
3014
3015         if (config & ~extra_regs[alt_idx].valid_mask)
3016                 return idx;
3017
3018         return alt_idx;
3019 }
3020
3021 static void intel_fixup_er(struct perf_event *event, int idx)
3022 {
3023         struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs);
3024         event->hw.extra_reg.idx = idx;
3025
3026         if (idx == EXTRA_REG_RSP_0) {
3027                 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
3028                 event->hw.config |= extra_regs[EXTRA_REG_RSP_0].event;
3029                 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
3030         } else if (idx == EXTRA_REG_RSP_1) {
3031                 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
3032                 event->hw.config |= extra_regs[EXTRA_REG_RSP_1].event;
3033                 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
3034         }
3035 }
3036
3037 /*
3038  * manage allocation of shared extra msr for certain events
3039  *
3040  * sharing can be:
3041  * per-cpu: to be shared between the various events on a single PMU
3042  * per-core: per-cpu + shared by HT threads
3043  */
3044 static struct event_constraint *
3045 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
3046                                    struct perf_event *event,
3047                                    struct hw_perf_event_extra *reg)
3048 {
3049         struct event_constraint *c = &emptyconstraint;
3050         struct er_account *era;
3051         unsigned long flags;
3052         int idx = reg->idx;
3053
3054         /*
3055          * reg->alloc can be set due to existing state, so for fake cpuc we
3056          * need to ignore this, otherwise we might fail to allocate proper fake
3057          * state for this extra reg constraint. Also see the comment below.
3058          */
3059         if (reg->alloc && !cpuc->is_fake)
3060                 return NULL; /* call x86_get_event_constraint() */
3061
3062 again:
3063         era = &cpuc->shared_regs->regs[idx];
3064         /*
3065          * we use spin_lock_irqsave() to avoid lockdep issues when
3066          * passing a fake cpuc
3067          */
3068         raw_spin_lock_irqsave(&era->lock, flags);
3069
3070         if (!atomic_read(&era->ref) || era->config == reg->config) {
3071
3072                 /*
3073                  * If its a fake cpuc -- as per validate_{group,event}() we
3074                  * shouldn't touch event state and we can avoid doing so
3075                  * since both will only call get_event_constraints() once
3076                  * on each event, this avoids the need for reg->alloc.
3077                  *
3078                  * Not doing the ER fixup will only result in era->reg being
3079                  * wrong, but since we won't actually try and program hardware
3080                  * this isn't a problem either.
3081                  */
3082                 if (!cpuc->is_fake) {
3083                         if (idx != reg->idx)
3084                                 intel_fixup_er(event, idx);
3085
3086                         /*
3087                          * x86_schedule_events() can call get_event_constraints()
3088                          * multiple times on events in the case of incremental
3089                          * scheduling(). reg->alloc ensures we only do the ER
3090                          * allocation once.
3091                          */
3092                         reg->alloc = 1;
3093                 }
3094
3095                 /* lock in msr value */
3096                 era->config = reg->config;
3097                 era->reg = reg->reg;
3098
3099                 /* one more user */
3100                 atomic_inc(&era->ref);
3101
3102                 /*
3103                  * need to call x86_get_event_constraint()
3104                  * to check if associated event has constraints
3105                  */
3106                 c = NULL;
3107         } else {
3108                 idx = intel_alt_er(cpuc, idx, reg->config);
3109                 if (idx != reg->idx) {
3110                         raw_spin_unlock_irqrestore(&era->lock, flags);
3111                         goto again;
3112                 }
3113         }
3114         raw_spin_unlock_irqrestore(&era->lock, flags);
3115
3116         return c;
3117 }
3118
3119 static void
3120 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
3121                                    struct hw_perf_event_extra *reg)
3122 {
3123         struct er_account *era;
3124
3125         /*
3126          * Only put constraint if extra reg was actually allocated. Also takes
3127          * care of event which do not use an extra shared reg.
3128          *
3129          * Also, if this is a fake cpuc we shouldn't touch any event state
3130          * (reg->alloc) and we don't care about leaving inconsistent cpuc state
3131          * either since it'll be thrown out.
3132          */
3133         if (!reg->alloc || cpuc->is_fake)
3134                 return;
3135
3136         era = &cpuc->shared_regs->regs[reg->idx];
3137
3138         /* one fewer user */
3139         atomic_dec(&era->ref);
3140
3141         /* allocate again next time */
3142         reg->alloc = 0;
3143 }
3144
3145 static struct event_constraint *
3146 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
3147                               struct perf_event *event)
3148 {
3149         struct event_constraint *c = NULL, *d;
3150         struct hw_perf_event_extra *xreg, *breg;
3151
3152         xreg = &event->hw.extra_reg;
3153         if (xreg->idx != EXTRA_REG_NONE) {
3154                 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
3155                 if (c == &emptyconstraint)
3156                         return c;
3157         }
3158         breg = &event->hw.branch_reg;
3159         if (breg->idx != EXTRA_REG_NONE) {
3160                 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
3161                 if (d == &emptyconstraint) {
3162                         __intel_shared_reg_put_constraints(cpuc, xreg);
3163                         c = d;
3164                 }
3165         }
3166         return c;
3167 }
3168
3169 struct event_constraint *
3170 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3171                           struct perf_event *event)
3172 {
3173         struct event_constraint *event_constraints = hybrid(cpuc->pmu, event_constraints);
3174         struct event_constraint *c;
3175
3176         if (event_constraints) {
3177                 for_each_event_constraint(c, event_constraints) {
3178                         if (constraint_match(c, event->hw.config)) {
3179                                 event->hw.flags |= c->flags;
3180                                 return c;
3181                         }
3182                 }
3183         }
3184
3185         return &hybrid_var(cpuc->pmu, unconstrained);
3186 }
3187
3188 static struct event_constraint *
3189 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3190                             struct perf_event *event)
3191 {
3192         struct event_constraint *c;
3193
3194         c = intel_vlbr_constraints(event);
3195         if (c)
3196                 return c;
3197
3198         c = intel_bts_constraints(event);
3199         if (c)
3200                 return c;
3201
3202         c = intel_shared_regs_constraints(cpuc, event);
3203         if (c)
3204                 return c;
3205
3206         c = intel_pebs_constraints(event);
3207         if (c)
3208                 return c;
3209
3210         return x86_get_event_constraints(cpuc, idx, event);
3211 }
3212
3213 static void
3214 intel_start_scheduling(struct cpu_hw_events *cpuc)
3215 {
3216         struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3217         struct intel_excl_states *xl;
3218         int tid = cpuc->excl_thread_id;
3219
3220         /*
3221          * nothing needed if in group validation mode
3222          */
3223         if (cpuc->is_fake || !is_ht_workaround_enabled())
3224                 return;
3225
3226         /*
3227          * no exclusion needed
3228          */
3229         if (WARN_ON_ONCE(!excl_cntrs))
3230                 return;
3231
3232         xl = &excl_cntrs->states[tid];
3233
3234         xl->sched_started = true;
3235         /*
3236          * lock shared state until we are done scheduling
3237          * in stop_event_scheduling()
3238          * makes scheduling appear as a transaction
3239          */
3240         raw_spin_lock(&excl_cntrs->lock);
3241 }
3242
3243 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
3244 {
3245         struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3246         struct event_constraint *c = cpuc->event_constraint[idx];
3247         struct intel_excl_states *xl;
3248         int tid = cpuc->excl_thread_id;
3249
3250         if (cpuc->is_fake || !is_ht_workaround_enabled())
3251                 return;
3252
3253         if (WARN_ON_ONCE(!excl_cntrs))
3254                 return;
3255
3256         if (!(c->flags & PERF_X86_EVENT_DYNAMIC))
3257                 return;
3258
3259         xl = &excl_cntrs->states[tid];
3260
3261         lockdep_assert_held(&excl_cntrs->lock);
3262
3263         if (c->flags & PERF_X86_EVENT_EXCL)
3264                 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE;
3265         else
3266                 xl->state[cntr] = INTEL_EXCL_SHARED;
3267 }
3268
3269 static void
3270 intel_stop_scheduling(struct cpu_hw_events *cpuc)
3271 {
3272         struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3273         struct intel_excl_states *xl;
3274         int tid = cpuc->excl_thread_id;
3275
3276         /*
3277          * nothing needed if in group validation mode
3278          */
3279         if (cpuc->is_fake || !is_ht_workaround_enabled())
3280                 return;
3281         /*
3282          * no exclusion needed
3283          */
3284         if (WARN_ON_ONCE(!excl_cntrs))
3285                 return;
3286
3287         xl = &excl_cntrs->states[tid];
3288
3289         xl->sched_started = false;
3290         /*
3291          * release shared state lock (acquired in intel_start_scheduling())
3292          */
3293         raw_spin_unlock(&excl_cntrs->lock);
3294 }
3295
3296 static struct event_constraint *
3297 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx)
3298 {
3299         WARN_ON_ONCE(!cpuc->constraint_list);
3300
3301         if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
3302                 struct event_constraint *cx;
3303
3304                 /*
3305                  * grab pre-allocated constraint entry
3306                  */
3307                 cx = &cpuc->constraint_list[idx];
3308
3309                 /*
3310                  * initialize dynamic constraint
3311                  * with static constraint
3312                  */
3313                 *cx = *c;
3314
3315                 /*
3316                  * mark constraint as dynamic
3317                  */
3318                 cx->flags |= PERF_X86_EVENT_DYNAMIC;
3319                 c = cx;
3320         }
3321
3322         return c;
3323 }
3324
3325 static struct event_constraint *
3326 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
3327                            int idx, struct event_constraint *c)
3328 {
3329         struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3330         struct intel_excl_states *xlo;
3331         int tid = cpuc->excl_thread_id;
3332         int is_excl, i, w;
3333
3334         /*
3335          * validating a group does not require
3336          * enforcing cross-thread  exclusion
3337          */
3338         if (cpuc->is_fake || !is_ht_workaround_enabled())
3339                 return c;
3340
3341         /*
3342          * no exclusion needed
3343          */
3344         if (WARN_ON_ONCE(!excl_cntrs))
3345                 return c;
3346
3347         /*
3348          * because we modify the constraint, we need
3349          * to make a copy. Static constraints come
3350          * from static const tables.
3351          *
3352          * only needed when constraint has not yet
3353          * been cloned (marked dynamic)
3354          */
3355         c = dyn_constraint(cpuc, c, idx);
3356
3357         /*
3358          * From here on, the constraint is dynamic.
3359          * Either it was just allocated above, or it
3360          * was allocated during a earlier invocation
3361          * of this function
3362          */
3363
3364         /*
3365          * state of sibling HT
3366          */
3367         xlo = &excl_cntrs->states[tid ^ 1];
3368
3369         /*
3370          * event requires exclusive counter access
3371          * across HT threads
3372          */
3373         is_excl = c->flags & PERF_X86_EVENT_EXCL;
3374         if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) {
3375                 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT;
3376                 if (!cpuc->n_excl++)
3377                         WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
3378         }
3379
3380         /*
3381          * Modify static constraint with current dynamic
3382          * state of thread
3383          *
3384          * EXCLUSIVE: sibling counter measuring exclusive event
3385          * SHARED   : sibling counter measuring non-exclusive event
3386          * UNUSED   : sibling counter unused
3387          */
3388         w = c->weight;
3389         for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) {
3390                 /*
3391                  * exclusive event in sibling counter
3392                  * our corresponding counter cannot be used
3393                  * regardless of our event
3394                  */
3395                 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) {
3396                         __clear_bit(i, c->idxmsk);
3397                         w--;
3398                         continue;
3399                 }
3400                 /*
3401                  * if measuring an exclusive event, sibling
3402                  * measuring non-exclusive, then counter cannot
3403                  * be used
3404                  */
3405                 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) {
3406                         __clear_bit(i, c->idxmsk);
3407                         w--;
3408                         continue;
3409                 }
3410         }
3411
3412         /*
3413          * if we return an empty mask, then switch
3414          * back to static empty constraint to avoid
3415          * the cost of freeing later on
3416          */
3417         if (!w)
3418                 c = &emptyconstraint;
3419
3420         c->weight = w;
3421
3422         return c;
3423 }
3424
3425 static struct event_constraint *
3426 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3427                             struct perf_event *event)
3428 {
3429         struct event_constraint *c1, *c2;
3430
3431         c1 = cpuc->event_constraint[idx];
3432
3433         /*
3434          * first time only
3435          * - static constraint: no change across incremental scheduling calls
3436          * - dynamic constraint: handled by intel_get_excl_constraints()
3437          */
3438         c2 = __intel_get_event_constraints(cpuc, idx, event);
3439         if (c1) {
3440                 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC));
3441                 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX);
3442                 c1->weight = c2->weight;
3443                 c2 = c1;
3444         }
3445
3446         if (cpuc->excl_cntrs)
3447                 return intel_get_excl_constraints(cpuc, event, idx, c2);
3448
3449         return c2;
3450 }
3451
3452 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc,
3453                 struct perf_event *event)
3454 {
3455         struct hw_perf_event *hwc = &event->hw;
3456         struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3457         int tid = cpuc->excl_thread_id;
3458         struct intel_excl_states *xl;
3459
3460         /*
3461          * nothing needed if in group validation mode
3462          */
3463         if (cpuc->is_fake)
3464                 return;
3465
3466         if (WARN_ON_ONCE(!excl_cntrs))
3467                 return;
3468
3469         if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) {
3470                 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT;
3471                 if (!--cpuc->n_excl)
3472                         WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
3473         }
3474
3475         /*
3476          * If event was actually assigned, then mark the counter state as
3477          * unused now.
3478          */
3479         if (hwc->idx >= 0) {
3480                 xl = &excl_cntrs->states[tid];
3481
3482                 /*
3483                  * put_constraint may be called from x86_schedule_events()
3484                  * which already has the lock held so here make locking
3485                  * conditional.
3486                  */
3487                 if (!xl->sched_started)
3488                         raw_spin_lock(&excl_cntrs->lock);
3489
3490                 xl->state[hwc->idx] = INTEL_EXCL_UNUSED;
3491
3492                 if (!xl->sched_started)
3493                         raw_spin_unlock(&excl_cntrs->lock);
3494         }
3495 }
3496
3497 static void
3498 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
3499                                         struct perf_event *event)
3500 {
3501         struct hw_perf_event_extra *reg;
3502
3503         reg = &event->hw.extra_reg;
3504         if (reg->idx != EXTRA_REG_NONE)
3505                 __intel_shared_reg_put_constraints(cpuc, reg);
3506
3507         reg = &event->hw.branch_reg;
3508         if (reg->idx != EXTRA_REG_NONE)
3509                 __intel_shared_reg_put_constraints(cpuc, reg);
3510 }
3511
3512 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
3513                                         struct perf_event *event)
3514 {
3515         intel_put_shared_regs_event_constraints(cpuc, event);
3516
3517         /*
3518          * is PMU has exclusive counter restrictions, then
3519          * all events are subject to and must call the
3520          * put_excl_constraints() routine
3521          */
3522         if (cpuc->excl_cntrs)
3523                 intel_put_excl_constraints(cpuc, event);
3524 }
3525
3526 static void intel_pebs_aliases_core2(struct perf_event *event)
3527 {
3528         if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3529                 /*
3530                  * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3531                  * (0x003c) so that we can use it with PEBS.
3532                  *
3533                  * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3534                  * PEBS capable. However we can use INST_RETIRED.ANY_P
3535                  * (0x00c0), which is a PEBS capable event, to get the same
3536                  * count.
3537                  *
3538                  * INST_RETIRED.ANY_P counts the number of cycles that retires
3539                  * CNTMASK instructions. By setting CNTMASK to a value (16)
3540                  * larger than the maximum number of instructions that can be
3541                  * retired per cycle (4) and then inverting the condition, we
3542                  * count all cycles that retire 16 or less instructions, which
3543                  * is every cycle.
3544                  *
3545                  * Thereby we gain a PEBS capable cycle counter.
3546                  */
3547                 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
3548
3549                 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3550                 event->hw.config = alt_config;
3551         }
3552 }
3553
3554 static void intel_pebs_aliases_snb(struct perf_event *event)
3555 {
3556         if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3557                 /*
3558                  * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3559                  * (0x003c) so that we can use it with PEBS.
3560                  *
3561                  * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3562                  * PEBS capable. However we can use UOPS_RETIRED.ALL
3563                  * (0x01c2), which is a PEBS capable event, to get the same
3564                  * count.
3565                  *
3566                  * UOPS_RETIRED.ALL counts the number of cycles that retires
3567                  * CNTMASK micro-ops. By setting CNTMASK to a value (16)
3568                  * larger than the maximum number of micro-ops that can be
3569                  * retired per cycle (4) and then inverting the condition, we
3570                  * count all cycles that retire 16 or less micro-ops, which
3571                  * is every cycle.
3572                  *
3573                  * Thereby we gain a PEBS capable cycle counter.
3574                  */
3575                 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
3576
3577                 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3578                 event->hw.config = alt_config;
3579         }
3580 }
3581
3582 static void intel_pebs_aliases_precdist(struct perf_event *event)
3583 {
3584         if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3585                 /*
3586                  * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3587                  * (0x003c) so that we can use it with PEBS.
3588                  *
3589                  * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3590                  * PEBS capable. However we can use INST_RETIRED.PREC_DIST
3591                  * (0x01c0), which is a PEBS capable event, to get the same
3592                  * count.
3593                  *
3594                  * The PREC_DIST event has special support to minimize sample
3595                  * shadowing effects. One drawback is that it can be
3596                  * only programmed on counter 1, but that seems like an
3597                  * acceptable trade off.
3598                  */
3599                 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16);
3600
3601                 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3602                 event->hw.config = alt_config;
3603         }
3604 }
3605
3606 static void intel_pebs_aliases_ivb(struct perf_event *event)
3607 {
3608         if (event->attr.precise_ip < 3)
3609                 return intel_pebs_aliases_snb(event);
3610         return intel_pebs_aliases_precdist(event);
3611 }
3612
3613 static void intel_pebs_aliases_skl(struct perf_event *event)
3614 {
3615         if (event->attr.precise_ip < 3)
3616                 return intel_pebs_aliases_core2(event);
3617         return intel_pebs_aliases_precdist(event);
3618 }
3619
3620 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event)
3621 {
3622         unsigned long flags = x86_pmu.large_pebs_flags;
3623
3624         if (event->attr.use_clockid)
3625                 flags &= ~PERF_SAMPLE_TIME;
3626         if (!event->attr.exclude_kernel)
3627                 flags &= ~PERF_SAMPLE_REGS_USER;
3628         if (event->attr.sample_regs_user & ~PEBS_GP_REGS)
3629                 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR);
3630         return flags;
3631 }
3632
3633 static int intel_pmu_bts_config(struct perf_event *event)
3634 {
3635         struct perf_event_attr *attr = &event->attr;
3636
3637         if (unlikely(intel_pmu_has_bts(event))) {
3638                 /* BTS is not supported by this architecture. */
3639                 if (!x86_pmu.bts_active)
3640                         return -EOPNOTSUPP;
3641
3642                 /* BTS is currently only allowed for user-mode. */
3643                 if (!attr->exclude_kernel)
3644                         return -EOPNOTSUPP;
3645
3646                 /* BTS is not allowed for precise events. */
3647                 if (attr->precise_ip)
3648                         return -EOPNOTSUPP;
3649
3650                 /* disallow bts if conflicting events are present */
3651                 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3652                         return -EBUSY;
3653
3654                 event->destroy = hw_perf_lbr_event_destroy;
3655         }
3656
3657         return 0;
3658 }
3659
3660 static int core_pmu_hw_config(struct perf_event *event)
3661 {
3662         int ret = x86_pmu_hw_config(event);
3663
3664         if (ret)
3665                 return ret;
3666
3667         return intel_pmu_bts_config(event);
3668 }
3669
3670 #define INTEL_TD_METRIC_AVAILABLE_MAX   (INTEL_TD_METRIC_RETIRING + \
3671                                          ((x86_pmu.num_topdown_events - 1) << 8))
3672
3673 static bool is_available_metric_event(struct perf_event *event)
3674 {
3675         return is_metric_event(event) &&
3676                 event->attr.config <= INTEL_TD_METRIC_AVAILABLE_MAX;
3677 }
3678
3679 static inline bool is_mem_loads_event(struct perf_event *event)
3680 {
3681         return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0xcd, .umask=0x01);
3682 }
3683
3684 static inline bool is_mem_loads_aux_event(struct perf_event *event)
3685 {
3686         return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0x03, .umask=0x82);
3687 }
3688
3689 static inline bool require_mem_loads_aux_event(struct perf_event *event)
3690 {
3691         if (!(x86_pmu.flags & PMU_FL_MEM_LOADS_AUX))
3692                 return false;
3693
3694         if (is_hybrid())
3695                 return hybrid_pmu(event->pmu)->cpu_type == hybrid_big;
3696
3697         return true;
3698 }
3699
3700 static inline bool intel_pmu_has_cap(struct perf_event *event, int idx)
3701 {
3702         union perf_capabilities *intel_cap = &hybrid(event->pmu, intel_cap);
3703
3704         return test_bit(idx, (unsigned long *)&intel_cap->capabilities);
3705 }
3706
3707 static int intel_pmu_hw_config(struct perf_event *event)
3708 {
3709         int ret = x86_pmu_hw_config(event);
3710
3711         if (ret)
3712                 return ret;
3713
3714         ret = intel_pmu_bts_config(event);
3715         if (ret)
3716                 return ret;
3717
3718         if (event->attr.precise_ip) {
3719                 if ((event->attr.config & INTEL_ARCH_EVENT_MASK) == INTEL_FIXED_VLBR_EVENT)
3720                         return -EINVAL;
3721
3722                 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) {
3723                         event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD;
3724                         if (!(event->attr.sample_type &
3725                               ~intel_pmu_large_pebs_flags(event))) {
3726                                 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS;
3727                                 event->attach_state |= PERF_ATTACH_SCHED_CB;
3728                         }
3729                 }
3730                 if (x86_pmu.pebs_aliases)
3731                         x86_pmu.pebs_aliases(event);
3732
3733                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3734                         event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY;
3735         }
3736
3737         if (needs_branch_stack(event)) {
3738                 ret = intel_pmu_setup_lbr_filter(event);
3739                 if (ret)
3740                         return ret;
3741                 event->attach_state |= PERF_ATTACH_SCHED_CB;
3742
3743                 /*
3744                  * BTS is set up earlier in this path, so don't account twice
3745                  */
3746                 if (!unlikely(intel_pmu_has_bts(event))) {
3747                         /* disallow lbr if conflicting events are present */
3748                         if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3749                                 return -EBUSY;
3750
3751                         event->destroy = hw_perf_lbr_event_destroy;
3752                 }
3753         }
3754
3755         if (event->attr.aux_output) {
3756                 if (!event->attr.precise_ip)
3757                         return -EINVAL;
3758
3759                 event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT;
3760         }
3761
3762         if ((event->attr.type == PERF_TYPE_HARDWARE) ||
3763             (event->attr.type == PERF_TYPE_HW_CACHE))
3764                 return 0;
3765
3766         /*
3767          * Config Topdown slots and metric events
3768          *
3769          * The slots event on Fixed Counter 3 can support sampling,
3770          * which will be handled normally in x86_perf_event_update().
3771          *
3772          * Metric events don't support sampling and require being paired
3773          * with a slots event as group leader. When the slots event
3774          * is used in a metrics group, it too cannot support sampling.
3775          */
3776         if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) {
3777                 if (event->attr.config1 || event->attr.config2)
3778                         return -EINVAL;
3779
3780                 /*
3781                  * The TopDown metrics events and slots event don't
3782                  * support any filters.
3783                  */
3784                 if (event->attr.config & X86_ALL_EVENT_FLAGS)
3785                         return -EINVAL;
3786
3787                 if (is_available_metric_event(event)) {
3788                         struct perf_event *leader = event->group_leader;
3789
3790                         /* The metric events don't support sampling. */
3791                         if (is_sampling_event(event))
3792                                 return -EINVAL;
3793
3794                         /* The metric events require a slots group leader. */
3795                         if (!is_slots_event(leader))
3796                                 return -EINVAL;
3797
3798                         /*
3799                          * The leader/SLOTS must not be a sampling event for
3800                          * metric use; hardware requires it starts at 0 when used
3801                          * in conjunction with MSR_PERF_METRICS.
3802                          */
3803                         if (is_sampling_event(leader))
3804                                 return -EINVAL;
3805
3806                         event->event_caps |= PERF_EV_CAP_SIBLING;
3807                         /*
3808                          * Only once we have a METRICs sibling do we
3809                          * need TopDown magic.
3810                          */
3811                         leader->hw.flags |= PERF_X86_EVENT_TOPDOWN;
3812                         event->hw.flags  |= PERF_X86_EVENT_TOPDOWN;
3813                 }
3814         }
3815
3816         /*
3817          * The load latency event X86_CONFIG(.event=0xcd, .umask=0x01) on SPR
3818          * doesn't function quite right. As a work-around it needs to always be
3819          * co-scheduled with a auxiliary event X86_CONFIG(.event=0x03, .umask=0x82).
3820          * The actual count of this second event is irrelevant it just needs
3821          * to be active to make the first event function correctly.
3822          *
3823          * In a group, the auxiliary event must be in front of the load latency
3824          * event. The rule is to simplify the implementation of the check.
3825          * That's because perf cannot have a complete group at the moment.
3826          */
3827         if (require_mem_loads_aux_event(event) &&
3828             (event->attr.sample_type & PERF_SAMPLE_DATA_SRC) &&
3829             is_mem_loads_event(event)) {
3830                 struct perf_event *leader = event->group_leader;
3831                 struct perf_event *sibling = NULL;
3832
3833                 if (!is_mem_loads_aux_event(leader)) {
3834                         for_each_sibling_event(sibling, leader) {
3835                                 if (is_mem_loads_aux_event(sibling))
3836                                         break;
3837                         }
3838                         if (list_entry_is_head(sibling, &leader->sibling_list, sibling_list))
3839                                 return -ENODATA;
3840                 }
3841         }
3842
3843         if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
3844                 return 0;
3845
3846         if (x86_pmu.version < 3)
3847                 return -EINVAL;
3848
3849         ret = perf_allow_cpu(&event->attr);
3850         if (ret)
3851                 return ret;
3852
3853         event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
3854
3855         return 0;
3856 }
3857
3858 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
3859 {
3860         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3861         struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3862         u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
3863
3864         arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
3865         arr[0].host = intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
3866         arr[0].guest = intel_ctrl & ~cpuc->intel_ctrl_host_mask;
3867         if (x86_pmu.flags & PMU_FL_PEBS_ALL)
3868                 arr[0].guest &= ~cpuc->pebs_enabled;
3869         else
3870                 arr[0].guest &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
3871         *nr = 1;
3872
3873         if (x86_pmu.pebs && x86_pmu.pebs_no_isolation) {
3874                 /*
3875                  * If PMU counter has PEBS enabled it is not enough to
3876                  * disable counter on a guest entry since PEBS memory
3877                  * write can overshoot guest entry and corrupt guest
3878                  * memory. Disabling PEBS solves the problem.
3879                  *
3880                  * Don't do this if the CPU already enforces it.
3881                  */
3882                 arr[1].msr = MSR_IA32_PEBS_ENABLE;
3883                 arr[1].host = cpuc->pebs_enabled;
3884                 arr[1].guest = 0;
3885                 *nr = 2;
3886         }
3887
3888         return arr;
3889 }
3890
3891 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
3892 {
3893         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3894         struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3895         int idx;
3896
3897         for (idx = 0; idx < x86_pmu.num_counters; idx++)  {
3898                 struct perf_event *event = cpuc->events[idx];
3899
3900                 arr[idx].msr = x86_pmu_config_addr(idx);
3901                 arr[idx].host = arr[idx].guest = 0;
3902
3903                 if (!test_bit(idx, cpuc->active_mask))
3904                         continue;
3905
3906                 arr[idx].host = arr[idx].guest =
3907                         event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
3908
3909                 if (event->attr.exclude_host)
3910                         arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3911                 else if (event->attr.exclude_guest)
3912                         arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3913         }
3914
3915         *nr = x86_pmu.num_counters;
3916         return arr;
3917 }
3918
3919 static void core_pmu_enable_event(struct perf_event *event)
3920 {
3921         if (!event->attr.exclude_host)
3922                 x86_pmu_enable_event(event);
3923 }
3924
3925 static void core_pmu_enable_all(int added)
3926 {
3927         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3928         int idx;
3929
3930         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3931                 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
3932
3933                 if (!test_bit(idx, cpuc->active_mask) ||
3934                                 cpuc->events[idx]->attr.exclude_host)
3935                         continue;
3936
3937                 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
3938         }
3939 }
3940
3941 static int hsw_hw_config(struct perf_event *event)
3942 {
3943         int ret = intel_pmu_hw_config(event);
3944
3945         if (ret)
3946                 return ret;
3947         if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
3948                 return 0;
3949         event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
3950
3951         /*
3952          * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
3953          * PEBS or in ANY thread mode. Since the results are non-sensical forbid
3954          * this combination.
3955          */
3956         if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
3957              ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
3958               event->attr.precise_ip > 0))
3959                 return -EOPNOTSUPP;
3960
3961         if (event_is_checkpointed(event)) {
3962                 /*
3963                  * Sampling of checkpointed events can cause situations where
3964                  * the CPU constantly aborts because of a overflow, which is
3965                  * then checkpointed back and ignored. Forbid checkpointing
3966                  * for sampling.
3967                  *
3968                  * But still allow a long sampling period, so that perf stat
3969                  * from KVM works.
3970                  */
3971                 if (event->attr.sample_period > 0 &&
3972                     event->attr.sample_period < 0x7fffffff)
3973                         return -EOPNOTSUPP;
3974         }
3975         return 0;
3976 }
3977
3978 static struct event_constraint counter0_constraint =
3979                         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1);
3980
3981 static struct event_constraint counter2_constraint =
3982                         EVENT_CONSTRAINT(0, 0x4, 0);
3983
3984 static struct event_constraint fixed0_constraint =
3985                         FIXED_EVENT_CONSTRAINT(0x00c0, 0);
3986
3987 static struct event_constraint fixed0_counter0_constraint =
3988                         INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL);
3989
3990 static struct event_constraint *
3991 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3992                           struct perf_event *event)
3993 {
3994         struct event_constraint *c;
3995
3996         c = intel_get_event_constraints(cpuc, idx, event);
3997
3998         /* Handle special quirk on in_tx_checkpointed only in counter 2 */
3999         if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
4000                 if (c->idxmsk64 & (1U << 2))
4001                         return &counter2_constraint;
4002                 return &emptyconstraint;
4003         }
4004
4005         return c;
4006 }
4007
4008 static struct event_constraint *
4009 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4010                           struct perf_event *event)
4011 {
4012         /*
4013          * Fixed counter 0 has less skid.
4014          * Force instruction:ppp in Fixed counter 0
4015          */
4016         if ((event->attr.precise_ip == 3) &&
4017             constraint_match(&fixed0_constraint, event->hw.config))
4018                 return &fixed0_constraint;
4019
4020         return hsw_get_event_constraints(cpuc, idx, event);
4021 }
4022
4023 static struct event_constraint *
4024 spr_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4025                           struct perf_event *event)
4026 {
4027         struct event_constraint *c;
4028
4029         c = icl_get_event_constraints(cpuc, idx, event);
4030
4031         /*
4032          * The :ppp indicates the Precise Distribution (PDist) facility, which
4033          * is only supported on the GP counter 0. If a :ppp event which is not
4034          * available on the GP counter 0, error out.
4035          * Exception: Instruction PDIR is only available on the fixed counter 0.
4036          */
4037         if ((event->attr.precise_ip == 3) &&
4038             !constraint_match(&fixed0_constraint, event->hw.config)) {
4039                 if (c->idxmsk64 & BIT_ULL(0))
4040                         return &counter0_constraint;
4041
4042                 return &emptyconstraint;
4043         }
4044
4045         return c;
4046 }
4047
4048 static struct event_constraint *
4049 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4050                           struct perf_event *event)
4051 {
4052         struct event_constraint *c;
4053
4054         /* :ppp means to do reduced skid PEBS which is PMC0 only. */
4055         if (event->attr.precise_ip == 3)
4056                 return &counter0_constraint;
4057
4058         c = intel_get_event_constraints(cpuc, idx, event);
4059
4060         return c;
4061 }
4062
4063 static struct event_constraint *
4064 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4065                           struct perf_event *event)
4066 {
4067         struct event_constraint *c;
4068
4069         /*
4070          * :ppp means to do reduced skid PEBS,
4071          * which is available on PMC0 and fixed counter 0.
4072          */
4073         if (event->attr.precise_ip == 3) {
4074                 /* Force instruction:ppp on PMC0 and Fixed counter 0 */
4075                 if (constraint_match(&fixed0_constraint, event->hw.config))
4076                         return &fixed0_counter0_constraint;
4077
4078                 return &counter0_constraint;
4079         }
4080
4081         c = intel_get_event_constraints(cpuc, idx, event);
4082
4083         return c;
4084 }
4085
4086 static bool allow_tsx_force_abort = true;
4087
4088 static struct event_constraint *
4089 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4090                           struct perf_event *event)
4091 {
4092         struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event);
4093
4094         /*
4095          * Without TFA we must not use PMC3.
4096          */
4097         if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) {
4098                 c = dyn_constraint(cpuc, c, idx);
4099                 c->idxmsk64 &= ~(1ULL << 3);
4100                 c->weight--;
4101         }
4102
4103         return c;
4104 }
4105
4106 static struct event_constraint *
4107 adl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4108                           struct perf_event *event)
4109 {
4110         struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4111
4112         if (pmu->cpu_type == hybrid_big)
4113                 return spr_get_event_constraints(cpuc, idx, event);
4114         else if (pmu->cpu_type == hybrid_small)
4115                 return tnt_get_event_constraints(cpuc, idx, event);
4116
4117         WARN_ON(1);
4118         return &emptyconstraint;
4119 }
4120
4121 static int adl_hw_config(struct perf_event *event)
4122 {
4123         struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4124
4125         if (pmu->cpu_type == hybrid_big)
4126                 return hsw_hw_config(event);
4127         else if (pmu->cpu_type == hybrid_small)
4128                 return intel_pmu_hw_config(event);
4129
4130         WARN_ON(1);
4131         return -EOPNOTSUPP;
4132 }
4133
4134 static u8 adl_get_hybrid_cpu_type(void)
4135 {
4136         return hybrid_big;
4137 }
4138
4139 /*
4140  * Broadwell:
4141  *
4142  * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
4143  * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
4144  * the two to enforce a minimum period of 128 (the smallest value that has bits
4145  * 0-5 cleared and >= 100).
4146  *
4147  * Because of how the code in x86_perf_event_set_period() works, the truncation
4148  * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
4149  * to make up for the 'lost' events due to carrying the 'error' in period_left.
4150  *
4151  * Therefore the effective (average) period matches the requested period,
4152  * despite coarser hardware granularity.
4153  */
4154 static u64 bdw_limit_period(struct perf_event *event, u64 left)
4155 {
4156         if ((event->hw.config & INTEL_ARCH_EVENT_MASK) ==
4157                         X86_CONFIG(.event=0xc0, .umask=0x01)) {
4158                 if (left < 128)
4159                         left = 128;
4160                 left &= ~0x3fULL;
4161         }
4162         return left;
4163 }
4164
4165 static u64 nhm_limit_period(struct perf_event *event, u64 left)
4166 {
4167         return max(left, 32ULL);
4168 }
4169
4170 static u64 spr_limit_period(struct perf_event *event, u64 left)
4171 {
4172         if (event->attr.precise_ip == 3)
4173                 return max(left, 128ULL);
4174
4175         return left;
4176 }
4177
4178 PMU_FORMAT_ATTR(event,  "config:0-7"    );
4179 PMU_FORMAT_ATTR(umask,  "config:8-15"   );
4180 PMU_FORMAT_ATTR(edge,   "config:18"     );
4181 PMU_FORMAT_ATTR(pc,     "config:19"     );
4182 PMU_FORMAT_ATTR(any,    "config:21"     ); /* v3 + */
4183 PMU_FORMAT_ATTR(inv,    "config:23"     );
4184 PMU_FORMAT_ATTR(cmask,  "config:24-31"  );
4185 PMU_FORMAT_ATTR(in_tx,  "config:32");
4186 PMU_FORMAT_ATTR(in_tx_cp, "config:33");
4187
4188 static struct attribute *intel_arch_formats_attr[] = {
4189         &format_attr_event.attr,
4190         &format_attr_umask.attr,
4191         &format_attr_edge.attr,
4192         &format_attr_pc.attr,
4193         &format_attr_inv.attr,
4194         &format_attr_cmask.attr,
4195         NULL,
4196 };
4197
4198 ssize_t intel_event_sysfs_show(char *page, u64 config)
4199 {
4200         u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
4201
4202         return x86_event_sysfs_show(page, config, event);
4203 }
4204
4205 static struct intel_shared_regs *allocate_shared_regs(int cpu)
4206 {
4207         struct intel_shared_regs *regs;
4208         int i;
4209
4210         regs = kzalloc_node(sizeof(struct intel_shared_regs),
4211                             GFP_KERNEL, cpu_to_node(cpu));
4212         if (regs) {
4213                 /*
4214                  * initialize the locks to keep lockdep happy
4215                  */
4216                 for (i = 0; i < EXTRA_REG_MAX; i++)
4217                         raw_spin_lock_init(&regs->regs[i].lock);
4218
4219                 regs->core_id = -1;
4220         }
4221         return regs;
4222 }
4223
4224 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu)
4225 {
4226         struct intel_excl_cntrs *c;
4227
4228         c = kzalloc_node(sizeof(struct intel_excl_cntrs),
4229                          GFP_KERNEL, cpu_to_node(cpu));
4230         if (c) {
4231                 raw_spin_lock_init(&c->lock);
4232                 c->core_id = -1;
4233         }
4234         return c;
4235 }
4236
4237
4238 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu)
4239 {
4240         cpuc->pebs_record_size = x86_pmu.pebs_record_size;
4241
4242         if (is_hybrid() || x86_pmu.extra_regs || x86_pmu.lbr_sel_map) {
4243                 cpuc->shared_regs = allocate_shared_regs(cpu);
4244                 if (!cpuc->shared_regs)
4245                         goto err;
4246         }
4247
4248         if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) {
4249                 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint);
4250
4251                 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
4252                 if (!cpuc->constraint_list)
4253                         goto err_shared_regs;
4254         }
4255
4256         if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
4257                 cpuc->excl_cntrs = allocate_excl_cntrs(cpu);
4258                 if (!cpuc->excl_cntrs)
4259                         goto err_constraint_list;
4260
4261                 cpuc->excl_thread_id = 0;
4262         }
4263
4264         return 0;
4265
4266 err_constraint_list:
4267         kfree(cpuc->constraint_list);
4268         cpuc->constraint_list = NULL;
4269
4270 err_shared_regs:
4271         kfree(cpuc->shared_regs);
4272         cpuc->shared_regs = NULL;
4273
4274 err:
4275         return -ENOMEM;
4276 }
4277
4278 static int intel_pmu_cpu_prepare(int cpu)
4279 {
4280         return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu);
4281 }
4282
4283 static void flip_smm_bit(void *data)
4284 {
4285         unsigned long set = *(unsigned long *)data;
4286
4287         if (set > 0) {
4288                 msr_set_bit(MSR_IA32_DEBUGCTLMSR,
4289                             DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
4290         } else {
4291                 msr_clear_bit(MSR_IA32_DEBUGCTLMSR,
4292                               DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
4293         }
4294 }
4295
4296 static bool init_hybrid_pmu(int cpu)
4297 {
4298         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
4299         u8 cpu_type = get_this_hybrid_cpu_type();
4300         struct x86_hybrid_pmu *pmu = NULL;
4301         int i;
4302
4303         if (!cpu_type && x86_pmu.get_hybrid_cpu_type)
4304                 cpu_type = x86_pmu.get_hybrid_cpu_type();
4305
4306         for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) {
4307                 if (x86_pmu.hybrid_pmu[i].cpu_type == cpu_type) {
4308                         pmu = &x86_pmu.hybrid_pmu[i];
4309                         break;
4310                 }
4311         }
4312         if (WARN_ON_ONCE(!pmu || (pmu->pmu.type == -1))) {
4313                 cpuc->pmu = NULL;
4314                 return false;
4315         }
4316
4317         /* Only check and dump the PMU information for the first CPU */
4318         if (!cpumask_empty(&pmu->supported_cpus))
4319                 goto end;
4320
4321         if (!check_hw_exists(&pmu->pmu, pmu->num_counters, pmu->num_counters_fixed))
4322                 return false;
4323
4324         pr_info("%s PMU driver: ", pmu->name);
4325
4326         if (pmu->intel_cap.pebs_output_pt_available)
4327                 pr_cont("PEBS-via-PT ");
4328
4329         pr_cont("\n");
4330
4331         x86_pmu_show_pmu_cap(pmu->num_counters, pmu->num_counters_fixed,
4332                              pmu->intel_ctrl);
4333
4334 end:
4335         cpumask_set_cpu(cpu, &pmu->supported_cpus);
4336         cpuc->pmu = &pmu->pmu;
4337
4338         x86_pmu_update_cpu_context(&pmu->pmu, cpu);
4339
4340         return true;
4341 }
4342
4343 static void intel_pmu_cpu_starting(int cpu)
4344 {
4345         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
4346         int core_id = topology_core_id(cpu);
4347         int i;
4348
4349         if (is_hybrid() && !init_hybrid_pmu(cpu))
4350                 return;
4351
4352         init_debug_store_on_cpu(cpu);
4353         /*
4354          * Deal with CPUs that don't clear their LBRs on power-up.
4355          */
4356         intel_pmu_lbr_reset();
4357
4358         cpuc->lbr_sel = NULL;
4359
4360         if (x86_pmu.flags & PMU_FL_TFA) {
4361                 WARN_ON_ONCE(cpuc->tfa_shadow);
4362                 cpuc->tfa_shadow = ~0ULL;
4363                 intel_set_tfa(cpuc, false);
4364         }
4365
4366         if (x86_pmu.version > 1)
4367                 flip_smm_bit(&x86_pmu.attr_freeze_on_smi);
4368
4369         /*
4370          * Disable perf metrics if any added CPU doesn't support it.
4371          *
4372          * Turn off the check for a hybrid architecture, because the
4373          * architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicate
4374          * the architecture features. The perf metrics is a model-specific
4375          * feature for now. The corresponding bit should always be 0 on
4376          * a hybrid platform, e.g., Alder Lake.
4377          */
4378         if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) {
4379                 union perf_capabilities perf_cap;
4380
4381                 rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap.capabilities);
4382                 if (!perf_cap.perf_metrics) {
4383                         x86_pmu.intel_cap.perf_metrics = 0;
4384                         x86_pmu.intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS);
4385                 }
4386         }
4387
4388         if (!cpuc->shared_regs)
4389                 return;
4390
4391         if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) {
4392                 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
4393                         struct intel_shared_regs *pc;
4394
4395                         pc = per_cpu(cpu_hw_events, i).shared_regs;
4396                         if (pc && pc->core_id == core_id) {
4397                                 cpuc->kfree_on_online[0] = cpuc->shared_regs;
4398                                 cpuc->shared_regs = pc;
4399                                 break;
4400                         }
4401                 }
4402                 cpuc->shared_regs->core_id = core_id;
4403                 cpuc->shared_regs->refcnt++;
4404         }
4405
4406         if (x86_pmu.lbr_sel_map)
4407                 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
4408
4409         if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
4410                 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
4411                         struct cpu_hw_events *sibling;
4412                         struct intel_excl_cntrs *c;
4413
4414                         sibling = &per_cpu(cpu_hw_events, i);
4415                         c = sibling->excl_cntrs;
4416                         if (c && c->core_id == core_id) {
4417                                 cpuc->kfree_on_online[1] = cpuc->excl_cntrs;
4418                                 cpuc->excl_cntrs = c;
4419                                 if (!sibling->excl_thread_id)
4420                                         cpuc->excl_thread_id = 1;
4421                                 break;
4422                         }
4423                 }
4424                 cpuc->excl_cntrs->core_id = core_id;
4425                 cpuc->excl_cntrs->refcnt++;
4426         }
4427 }
4428
4429 static void free_excl_cntrs(struct cpu_hw_events *cpuc)
4430 {
4431         struct intel_excl_cntrs *c;
4432
4433         c = cpuc->excl_cntrs;
4434         if (c) {
4435                 if (c->core_id == -1 || --c->refcnt == 0)
4436                         kfree(c);
4437                 cpuc->excl_cntrs = NULL;
4438         }
4439
4440         kfree(cpuc->constraint_list);
4441         cpuc->constraint_list = NULL;
4442 }
4443
4444 static void intel_pmu_cpu_dying(int cpu)
4445 {
4446         fini_debug_store_on_cpu(cpu);
4447 }
4448
4449 void intel_cpuc_finish(struct cpu_hw_events *cpuc)
4450 {
4451         struct intel_shared_regs *pc;
4452
4453         pc = cpuc->shared_regs;
4454         if (pc) {
4455                 if (pc->core_id == -1 || --pc->refcnt == 0)
4456                         kfree(pc);
4457                 cpuc->shared_regs = NULL;
4458         }
4459
4460         free_excl_cntrs(cpuc);
4461 }
4462
4463 static void intel_pmu_cpu_dead(int cpu)
4464 {
4465         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
4466
4467         intel_cpuc_finish(cpuc);
4468
4469         if (is_hybrid() && cpuc->pmu)
4470                 cpumask_clear_cpu(cpu, &hybrid_pmu(cpuc->pmu)->supported_cpus);
4471 }
4472
4473 static void intel_pmu_sched_task(struct perf_event_context *ctx,
4474                                  bool sched_in)
4475 {
4476         intel_pmu_pebs_sched_task(ctx, sched_in);
4477         intel_pmu_lbr_sched_task(ctx, sched_in);
4478 }
4479
4480 static void intel_pmu_swap_task_ctx(struct perf_event_context *prev,
4481                                     struct perf_event_context *next)
4482 {
4483         intel_pmu_lbr_swap_task_ctx(prev, next);
4484 }
4485
4486 static int intel_pmu_check_period(struct perf_event *event, u64 value)
4487 {
4488         return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0;
4489 }
4490
4491 static int intel_pmu_aux_output_match(struct perf_event *event)
4492 {
4493         if (!x86_pmu.intel_cap.pebs_output_pt_available)
4494                 return 0;
4495
4496         return is_intel_pt_event(event);
4497 }
4498
4499 static int intel_pmu_filter_match(struct perf_event *event)
4500 {
4501         struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4502         unsigned int cpu = smp_processor_id();
4503
4504         return cpumask_test_cpu(cpu, &pmu->supported_cpus);
4505 }
4506
4507 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
4508
4509 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
4510
4511 PMU_FORMAT_ATTR(frontend, "config1:0-23");
4512
4513 static struct attribute *intel_arch3_formats_attr[] = {
4514         &format_attr_event.attr,
4515         &format_attr_umask.attr,
4516         &format_attr_edge.attr,
4517         &format_attr_pc.attr,
4518         &format_attr_any.attr,
4519         &format_attr_inv.attr,
4520         &format_attr_cmask.attr,
4521         NULL,
4522 };
4523
4524 static struct attribute *hsw_format_attr[] = {
4525         &format_attr_in_tx.attr,
4526         &format_attr_in_tx_cp.attr,
4527         &format_attr_offcore_rsp.attr,
4528         &format_attr_ldlat.attr,
4529         NULL
4530 };
4531
4532 static struct attribute *nhm_format_attr[] = {
4533         &format_attr_offcore_rsp.attr,
4534         &format_attr_ldlat.attr,
4535         NULL
4536 };
4537
4538 static struct attribute *slm_format_attr[] = {
4539         &format_attr_offcore_rsp.attr,
4540         NULL
4541 };
4542
4543 static struct attribute *skl_format_attr[] = {
4544         &format_attr_frontend.attr,
4545         NULL,
4546 };
4547
4548 static __initconst const struct x86_pmu core_pmu = {
4549         .name                   = "core",
4550         .handle_irq             = x86_pmu_handle_irq,
4551         .disable_all            = x86_pmu_disable_all,
4552         .enable_all             = core_pmu_enable_all,
4553         .enable                 = core_pmu_enable_event,
4554         .disable                = x86_pmu_disable_event,
4555         .hw_config              = core_pmu_hw_config,
4556         .schedule_events        = x86_schedule_events,
4557         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
4558         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
4559         .event_map              = intel_pmu_event_map,
4560         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
4561         .apic                   = 1,
4562         .large_pebs_flags       = LARGE_PEBS_FLAGS,
4563
4564         /*
4565          * Intel PMCs cannot be accessed sanely above 32-bit width,
4566          * so we install an artificial 1<<31 period regardless of
4567          * the generic event period:
4568          */
4569         .max_period             = (1ULL<<31) - 1,
4570         .get_event_constraints  = intel_get_event_constraints,
4571         .put_event_constraints  = intel_put_event_constraints,
4572         .event_constraints      = intel_core_event_constraints,
4573         .guest_get_msrs         = core_guest_get_msrs,
4574         .format_attrs           = intel_arch_formats_attr,
4575         .events_sysfs_show      = intel_event_sysfs_show,
4576
4577         /*
4578          * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
4579          * together with PMU version 1 and thus be using core_pmu with
4580          * shared_regs. We need following callbacks here to allocate
4581          * it properly.
4582          */
4583         .cpu_prepare            = intel_pmu_cpu_prepare,
4584         .cpu_starting           = intel_pmu_cpu_starting,
4585         .cpu_dying              = intel_pmu_cpu_dying,
4586         .cpu_dead               = intel_pmu_cpu_dead,
4587
4588         .check_period           = intel_pmu_check_period,
4589
4590         .lbr_reset              = intel_pmu_lbr_reset_64,
4591         .lbr_read               = intel_pmu_lbr_read_64,
4592         .lbr_save               = intel_pmu_lbr_save,
4593         .lbr_restore            = intel_pmu_lbr_restore,
4594 };
4595
4596 static __initconst const struct x86_pmu intel_pmu = {
4597         .name                   = "Intel",
4598         .handle_irq             = intel_pmu_handle_irq,
4599         .disable_all            = intel_pmu_disable_all,
4600         .enable_all             = intel_pmu_enable_all,
4601         .enable                 = intel_pmu_enable_event,
4602         .disable                = intel_pmu_disable_event,
4603         .add                    = intel_pmu_add_event,
4604         .del                    = intel_pmu_del_event,
4605         .read                   = intel_pmu_read_event,
4606         .hw_config              = intel_pmu_hw_config,
4607         .schedule_events        = x86_schedule_events,
4608         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
4609         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
4610         .event_map              = intel_pmu_event_map,
4611         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
4612         .apic                   = 1,
4613         .large_pebs_flags       = LARGE_PEBS_FLAGS,
4614         /*
4615          * Intel PMCs cannot be accessed sanely above 32 bit width,
4616          * so we install an artificial 1<<31 period regardless of
4617          * the generic event period:
4618          */
4619         .max_period             = (1ULL << 31) - 1,
4620         .get_event_constraints  = intel_get_event_constraints,
4621         .put_event_constraints  = intel_put_event_constraints,
4622         .pebs_aliases           = intel_pebs_aliases_core2,
4623
4624         .format_attrs           = intel_arch3_formats_attr,
4625         .events_sysfs_show      = intel_event_sysfs_show,
4626
4627         .cpu_prepare            = intel_pmu_cpu_prepare,
4628         .cpu_starting           = intel_pmu_cpu_starting,
4629         .cpu_dying              = intel_pmu_cpu_dying,
4630         .cpu_dead               = intel_pmu_cpu_dead,
4631
4632         .guest_get_msrs         = intel_guest_get_msrs,
4633         .sched_task             = intel_pmu_sched_task,
4634         .swap_task_ctx          = intel_pmu_swap_task_ctx,
4635
4636         .check_period           = intel_pmu_check_period,
4637
4638         .aux_output_match       = intel_pmu_aux_output_match,
4639
4640         .lbr_reset              = intel_pmu_lbr_reset_64,
4641         .lbr_read               = intel_pmu_lbr_read_64,
4642         .lbr_save               = intel_pmu_lbr_save,
4643         .lbr_restore            = intel_pmu_lbr_restore,
4644 };
4645
4646 static __init void intel_clovertown_quirk(void)
4647 {
4648         /*
4649          * PEBS is unreliable due to:
4650          *
4651          *   AJ67  - PEBS may experience CPL leaks
4652          *   AJ68  - PEBS PMI may be delayed by one event
4653          *   AJ69  - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
4654          *   AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
4655          *
4656          * AJ67 could be worked around by restricting the OS/USR flags.
4657          * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
4658          *
4659          * AJ106 could possibly be worked around by not allowing LBR
4660          *       usage from PEBS, including the fixup.
4661          * AJ68  could possibly be worked around by always programming
4662          *       a pebs_event_reset[0] value and coping with the lost events.
4663          *
4664          * But taken together it might just make sense to not enable PEBS on
4665          * these chips.
4666          */
4667         pr_warn("PEBS disabled due to CPU errata\n");
4668         x86_pmu.pebs = 0;
4669         x86_pmu.pebs_constraints = NULL;
4670 }
4671
4672 static const struct x86_cpu_desc isolation_ucodes[] = {
4673         INTEL_CPU_DESC(INTEL_FAM6_HASWELL,               3, 0x0000001f),
4674         INTEL_CPU_DESC(INTEL_FAM6_HASWELL_L,             1, 0x0000001e),
4675         INTEL_CPU_DESC(INTEL_FAM6_HASWELL_G,             1, 0x00000015),
4676         INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X,             2, 0x00000037),
4677         INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X,             4, 0x0000000a),
4678         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL,             4, 0x00000023),
4679         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_G,           1, 0x00000014),
4680         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D,           2, 0x00000010),
4681         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D,           3, 0x07000009),
4682         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D,           4, 0x0f000009),
4683         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D,           5, 0x0e000002),
4684         INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X,           1, 0x0b000014),
4685         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X,             3, 0x00000021),
4686         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X,             4, 0x00000000),
4687         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X,             5, 0x00000000),
4688         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X,             6, 0x00000000),
4689         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X,             7, 0x00000000),
4690         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_L,             3, 0x0000007c),
4691         INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE,               3, 0x0000007c),
4692         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE,              9, 0x0000004e),
4693         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L,            9, 0x0000004e),
4694         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L,           10, 0x0000004e),
4695         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L,           11, 0x0000004e),
4696         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L,           12, 0x0000004e),
4697         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE,             10, 0x0000004e),
4698         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE,             11, 0x0000004e),
4699         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE,             12, 0x0000004e),
4700         INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE,             13, 0x0000004e),
4701         {}
4702 };
4703
4704 static void intel_check_pebs_isolation(void)
4705 {
4706         x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes);
4707 }
4708
4709 static __init void intel_pebs_isolation_quirk(void)
4710 {
4711         WARN_ON_ONCE(x86_pmu.check_microcode);
4712         x86_pmu.check_microcode = intel_check_pebs_isolation;
4713         intel_check_pebs_isolation();
4714 }
4715
4716 static const struct x86_cpu_desc pebs_ucodes[] = {
4717         INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE,          7, 0x00000028),
4718         INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X,        6, 0x00000618),
4719         INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X,        7, 0x0000070c),
4720         {}
4721 };
4722
4723 static bool intel_snb_pebs_broken(void)
4724 {
4725         return !x86_cpu_has_min_microcode_rev(pebs_ucodes);
4726 }
4727
4728 static void intel_snb_check_microcode(void)
4729 {
4730         if (intel_snb_pebs_broken() == x86_pmu.pebs_broken)
4731                 return;
4732
4733         /*
4734          * Serialized by the microcode lock..
4735          */
4736         if (x86_pmu.pebs_broken) {
4737                 pr_info("PEBS enabled due to microcode update\n");
4738                 x86_pmu.pebs_broken = 0;
4739         } else {
4740                 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
4741                 x86_pmu.pebs_broken = 1;
4742         }
4743 }
4744
4745 static bool is_lbr_from(unsigned long msr)
4746 {
4747         unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr;
4748
4749         return x86_pmu.lbr_from <= msr && msr < lbr_from_nr;
4750 }
4751
4752 /*
4753  * Under certain circumstances, access certain MSR may cause #GP.
4754  * The function tests if the input MSR can be safely accessed.
4755  */
4756 static bool check_msr(unsigned long msr, u64 mask)
4757 {
4758         u64 val_old, val_new, val_tmp;
4759
4760         /*
4761          * Disable the check for real HW, so we don't
4762          * mess with potentially enabled registers:
4763          */
4764         if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
4765                 return true;
4766
4767         /*
4768          * Read the current value, change it and read it back to see if it
4769          * matches, this is needed to detect certain hardware emulators
4770          * (qemu/kvm) that don't trap on the MSR access and always return 0s.
4771          */
4772         if (rdmsrl_safe(msr, &val_old))
4773                 return false;
4774
4775         /*
4776          * Only change the bits which can be updated by wrmsrl.
4777          */
4778         val_tmp = val_old ^ mask;
4779
4780         if (is_lbr_from(msr))
4781                 val_tmp = lbr_from_signext_quirk_wr(val_tmp);
4782
4783         if (wrmsrl_safe(msr, val_tmp) ||
4784             rdmsrl_safe(msr, &val_new))
4785                 return false;
4786
4787         /*
4788          * Quirk only affects validation in wrmsr(), so wrmsrl()'s value
4789          * should equal rdmsrl()'s even with the quirk.
4790          */
4791         if (val_new != val_tmp)
4792                 return false;
4793
4794         if (is_lbr_from(msr))
4795                 val_old = lbr_from_signext_quirk_wr(val_old);
4796
4797         /* Here it's sure that the MSR can be safely accessed.
4798          * Restore the old value and return.
4799          */
4800         wrmsrl(msr, val_old);
4801
4802         return true;
4803 }
4804
4805 static __init void intel_sandybridge_quirk(void)
4806 {
4807         x86_pmu.check_microcode = intel_snb_check_microcode;
4808         cpus_read_lock();
4809         intel_snb_check_microcode();
4810         cpus_read_unlock();
4811 }
4812
4813 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
4814         { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
4815         { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
4816         { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
4817         { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
4818         { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
4819         { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
4820         { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
4821 };
4822
4823 static __init void intel_arch_events_quirk(void)
4824 {
4825         int bit;
4826
4827         /* disable event that reported as not present by cpuid */
4828         for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
4829                 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
4830                 pr_warn("CPUID marked event: \'%s\' unavailable\n",
4831                         intel_arch_events_map[bit].name);
4832         }
4833 }
4834
4835 static __init void intel_nehalem_quirk(void)
4836 {
4837         union cpuid10_ebx ebx;
4838
4839         ebx.full = x86_pmu.events_maskl;
4840         if (ebx.split.no_branch_misses_retired) {
4841                 /*
4842                  * Erratum AAJ80 detected, we work it around by using
4843                  * the BR_MISP_EXEC.ANY event. This will over-count
4844                  * branch-misses, but it's still much better than the
4845                  * architectural event which is often completely bogus:
4846                  */
4847                 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
4848                 ebx.split.no_branch_misses_retired = 0;
4849                 x86_pmu.events_maskl = ebx.full;
4850                 pr_info("CPU erratum AAJ80 worked around\n");
4851         }
4852 }
4853
4854 /*
4855  * enable software workaround for errata:
4856  * SNB: BJ122
4857  * IVB: BV98
4858  * HSW: HSD29
4859  *
4860  * Only needed when HT is enabled. However detecting
4861  * if HT is enabled is difficult (model specific). So instead,
4862  * we enable the workaround in the early boot, and verify if
4863  * it is needed in a later initcall phase once we have valid
4864  * topology information to check if HT is actually enabled
4865  */
4866 static __init void intel_ht_bug(void)
4867 {
4868         x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED;
4869
4870         x86_pmu.start_scheduling = intel_start_scheduling;
4871         x86_pmu.commit_scheduling = intel_commit_scheduling;
4872         x86_pmu.stop_scheduling = intel_stop_scheduling;
4873 }
4874
4875 EVENT_ATTR_STR(mem-loads,       mem_ld_hsw,     "event=0xcd,umask=0x1,ldlat=3");
4876 EVENT_ATTR_STR(mem-stores,      mem_st_hsw,     "event=0xd0,umask=0x82")
4877
4878 /* Haswell special events */
4879 EVENT_ATTR_STR(tx-start,        tx_start,       "event=0xc9,umask=0x1");
4880 EVENT_ATTR_STR(tx-commit,       tx_commit,      "event=0xc9,umask=0x2");
4881 EVENT_ATTR_STR(tx-abort,        tx_abort,       "event=0xc9,umask=0x4");
4882 EVENT_ATTR_STR(tx-capacity,     tx_capacity,    "event=0x54,umask=0x2");
4883 EVENT_ATTR_STR(tx-conflict,     tx_conflict,    "event=0x54,umask=0x1");
4884 EVENT_ATTR_STR(el-start,        el_start,       "event=0xc8,umask=0x1");
4885 EVENT_ATTR_STR(el-commit,       el_commit,      "event=0xc8,umask=0x2");
4886 EVENT_ATTR_STR(el-abort,        el_abort,       "event=0xc8,umask=0x4");
4887 EVENT_ATTR_STR(el-capacity,     el_capacity,    "event=0x54,umask=0x2");
4888 EVENT_ATTR_STR(el-conflict,     el_conflict,    "event=0x54,umask=0x1");
4889 EVENT_ATTR_STR(cycles-t,        cycles_t,       "event=0x3c,in_tx=1");
4890 EVENT_ATTR_STR(cycles-ct,       cycles_ct,      "event=0x3c,in_tx=1,in_tx_cp=1");
4891
4892 static struct attribute *hsw_events_attrs[] = {
4893         EVENT_PTR(td_slots_issued),
4894         EVENT_PTR(td_slots_retired),
4895         EVENT_PTR(td_fetch_bubbles),
4896         EVENT_PTR(td_total_slots),
4897         EVENT_PTR(td_total_slots_scale),
4898         EVENT_PTR(td_recovery_bubbles),
4899         EVENT_PTR(td_recovery_bubbles_scale),
4900         NULL
4901 };
4902
4903 static struct attribute *hsw_mem_events_attrs[] = {
4904         EVENT_PTR(mem_ld_hsw),
4905         EVENT_PTR(mem_st_hsw),
4906         NULL,
4907 };
4908
4909 static struct attribute *hsw_tsx_events_attrs[] = {
4910         EVENT_PTR(tx_start),
4911         EVENT_PTR(tx_commit),
4912         EVENT_PTR(tx_abort),
4913         EVENT_PTR(tx_capacity),
4914         EVENT_PTR(tx_conflict),
4915         EVENT_PTR(el_start),
4916         EVENT_PTR(el_commit),
4917         EVENT_PTR(el_abort),
4918         EVENT_PTR(el_capacity),
4919         EVENT_PTR(el_conflict),
4920         EVENT_PTR(cycles_t),
4921         EVENT_PTR(cycles_ct),
4922         NULL
4923 };
4924
4925 EVENT_ATTR_STR(tx-capacity-read,  tx_capacity_read,  "event=0x54,umask=0x80");
4926 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2");
4927 EVENT_ATTR_STR(el-capacity-read,  el_capacity_read,  "event=0x54,umask=0x80");
4928 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2");
4929
4930 static struct attribute *icl_events_attrs[] = {
4931         EVENT_PTR(mem_ld_hsw),
4932         EVENT_PTR(mem_st_hsw),
4933         NULL,
4934 };
4935
4936 static struct attribute *icl_td_events_attrs[] = {
4937         EVENT_PTR(slots),
4938         EVENT_PTR(td_retiring),
4939         EVENT_PTR(td_bad_spec),
4940         EVENT_PTR(td_fe_bound),
4941         EVENT_PTR(td_be_bound),
4942         NULL,
4943 };
4944
4945 static struct attribute *icl_tsx_events_attrs[] = {
4946         EVENT_PTR(tx_start),
4947         EVENT_PTR(tx_abort),
4948         EVENT_PTR(tx_commit),
4949         EVENT_PTR(tx_capacity_read),
4950         EVENT_PTR(tx_capacity_write),
4951         EVENT_PTR(tx_conflict),
4952         EVENT_PTR(el_start),
4953         EVENT_PTR(el_abort),
4954         EVENT_PTR(el_commit),
4955         EVENT_PTR(el_capacity_read),
4956         EVENT_PTR(el_capacity_write),
4957         EVENT_PTR(el_conflict),
4958         EVENT_PTR(cycles_t),
4959         EVENT_PTR(cycles_ct),
4960         NULL,
4961 };
4962
4963
4964 EVENT_ATTR_STR(mem-stores,      mem_st_spr,     "event=0xcd,umask=0x2");
4965 EVENT_ATTR_STR(mem-loads-aux,   mem_ld_aux,     "event=0x03,umask=0x82");
4966
4967 static struct attribute *spr_events_attrs[] = {
4968         EVENT_PTR(mem_ld_hsw),
4969         EVENT_PTR(mem_st_spr),
4970         EVENT_PTR(mem_ld_aux),
4971         NULL,
4972 };
4973
4974 static struct attribute *spr_td_events_attrs[] = {
4975         EVENT_PTR(slots),
4976         EVENT_PTR(td_retiring),
4977         EVENT_PTR(td_bad_spec),
4978         EVENT_PTR(td_fe_bound),
4979         EVENT_PTR(td_be_bound),
4980         EVENT_PTR(td_heavy_ops),
4981         EVENT_PTR(td_br_mispredict),
4982         EVENT_PTR(td_fetch_lat),
4983         EVENT_PTR(td_mem_bound),
4984         NULL,
4985 };
4986
4987 static struct attribute *spr_tsx_events_attrs[] = {
4988         EVENT_PTR(tx_start),
4989         EVENT_PTR(tx_abort),
4990         EVENT_PTR(tx_commit),
4991         EVENT_PTR(tx_capacity_read),
4992         EVENT_PTR(tx_capacity_write),
4993         EVENT_PTR(tx_conflict),
4994         EVENT_PTR(cycles_t),
4995         EVENT_PTR(cycles_ct),
4996         NULL,
4997 };
4998
4999 static ssize_t freeze_on_smi_show(struct device *cdev,
5000                                   struct device_attribute *attr,
5001                                   char *buf)
5002 {
5003         return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi);
5004 }
5005
5006 static DEFINE_MUTEX(freeze_on_smi_mutex);
5007
5008 static ssize_t freeze_on_smi_store(struct device *cdev,
5009                                    struct device_attribute *attr,
5010                                    const char *buf, size_t count)
5011 {
5012         unsigned long val;
5013         ssize_t ret;
5014
5015         ret = kstrtoul(buf, 0, &val);
5016         if (ret)
5017                 return ret;
5018
5019         if (val > 1)
5020                 return -EINVAL;
5021
5022         mutex_lock(&freeze_on_smi_mutex);
5023
5024         if (x86_pmu.attr_freeze_on_smi == val)
5025                 goto done;
5026
5027         x86_pmu.attr_freeze_on_smi = val;
5028
5029         get_online_cpus();
5030         on_each_cpu(flip_smm_bit, &val, 1);
5031         put_online_cpus();
5032 done:
5033         mutex_unlock(&freeze_on_smi_mutex);
5034
5035         return count;
5036 }
5037
5038 static void update_tfa_sched(void *ignored)
5039 {
5040         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
5041
5042         /*
5043          * check if PMC3 is used
5044          * and if so force schedule out for all event types all contexts
5045          */
5046         if (test_bit(3, cpuc->active_mask))
5047                 perf_pmu_resched(x86_get_pmu(smp_processor_id()));
5048 }
5049
5050 static ssize_t show_sysctl_tfa(struct device *cdev,
5051                               struct device_attribute *attr,
5052                               char *buf)
5053 {
5054         return snprintf(buf, 40, "%d\n", allow_tsx_force_abort);
5055 }
5056
5057 static ssize_t set_sysctl_tfa(struct device *cdev,
5058                               struct device_attribute *attr,
5059                               const char *buf, size_t count)
5060 {
5061         bool val;
5062         ssize_t ret;
5063
5064         ret = kstrtobool(buf, &val);
5065         if (ret)
5066                 return ret;
5067
5068         /* no change */
5069         if (val == allow_tsx_force_abort)
5070                 return count;
5071
5072         allow_tsx_force_abort = val;
5073
5074         get_online_cpus();
5075         on_each_cpu(update_tfa_sched, NULL, 1);
5076         put_online_cpus();
5077
5078         return count;
5079 }
5080
5081
5082 static DEVICE_ATTR_RW(freeze_on_smi);
5083
5084 static ssize_t branches_show(struct device *cdev,
5085                              struct device_attribute *attr,
5086                              char *buf)
5087 {
5088         return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
5089 }
5090
5091 static DEVICE_ATTR_RO(branches);
5092
5093 static struct attribute *lbr_attrs[] = {
5094         &dev_attr_branches.attr,
5095         NULL
5096 };
5097
5098 static char pmu_name_str[30];
5099
5100 static ssize_t pmu_name_show(struct device *cdev,
5101                              struct device_attribute *attr,
5102                              char *buf)
5103 {
5104         return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str);
5105 }
5106
5107 static DEVICE_ATTR_RO(pmu_name);
5108
5109 static struct attribute *intel_pmu_caps_attrs[] = {
5110        &dev_attr_pmu_name.attr,
5111        NULL
5112 };
5113
5114 static DEVICE_ATTR(allow_tsx_force_abort, 0644,
5115                    show_sysctl_tfa,
5116                    set_sysctl_tfa);
5117
5118 static struct attribute *intel_pmu_attrs[] = {
5119         &dev_attr_freeze_on_smi.attr,
5120         &dev_attr_allow_tsx_force_abort.attr,
5121         NULL,
5122 };
5123
5124 static umode_t
5125 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5126 {
5127         return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0;
5128 }
5129
5130 static umode_t
5131 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5132 {
5133         return x86_pmu.pebs ? attr->mode : 0;
5134 }
5135
5136 static umode_t
5137 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5138 {
5139         return x86_pmu.lbr_nr ? attr->mode : 0;
5140 }
5141
5142 static umode_t
5143 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5144 {
5145         return x86_pmu.version >= 2 ? attr->mode : 0;
5146 }
5147
5148 static umode_t
5149 default_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5150 {
5151         if (attr == &dev_attr_allow_tsx_force_abort.attr)
5152                 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0;
5153
5154         return attr->mode;
5155 }
5156
5157 static struct attribute_group group_events_td  = {
5158         .name = "events",
5159 };
5160
5161 static struct attribute_group group_events_mem = {
5162         .name       = "events",
5163         .is_visible = pebs_is_visible,
5164 };
5165
5166 static struct attribute_group group_events_tsx = {
5167         .name       = "events",
5168         .is_visible = tsx_is_visible,
5169 };
5170
5171 static struct attribute_group group_caps_gen = {
5172         .name  = "caps",
5173         .attrs = intel_pmu_caps_attrs,
5174 };
5175
5176 static struct attribute_group group_caps_lbr = {
5177         .name       = "caps",
5178         .attrs      = lbr_attrs,
5179         .is_visible = lbr_is_visible,
5180 };
5181
5182 static struct attribute_group group_format_extra = {
5183         .name       = "format",
5184         .is_visible = exra_is_visible,
5185 };
5186
5187 static struct attribute_group group_format_extra_skl = {
5188         .name       = "format",
5189         .is_visible = exra_is_visible,
5190 };
5191
5192 static struct attribute_group group_default = {
5193         .attrs      = intel_pmu_attrs,
5194         .is_visible = default_is_visible,
5195 };
5196
5197 static const struct attribute_group *attr_update[] = {
5198         &group_events_td,
5199         &group_events_mem,
5200         &group_events_tsx,
5201         &group_caps_gen,
5202         &group_caps_lbr,
5203         &group_format_extra,
5204         &group_format_extra_skl,
5205         &group_default,
5206         NULL,
5207 };
5208
5209 EVENT_ATTR_STR_HYBRID(slots,                 slots_adl,        "event=0x00,umask=0x4",                       hybrid_big);
5210 EVENT_ATTR_STR_HYBRID(topdown-retiring,      td_retiring_adl,  "event=0xc2,umask=0x0;event=0x00,umask=0x80", hybrid_big_small);
5211 EVENT_ATTR_STR_HYBRID(topdown-bad-spec,      td_bad_spec_adl,  "event=0x73,umask=0x0;event=0x00,umask=0x81", hybrid_big_small);
5212 EVENT_ATTR_STR_HYBRID(topdown-fe-bound,      td_fe_bound_adl,  "event=0x71,umask=0x0;event=0x00,umask=0x82", hybrid_big_small);
5213 EVENT_ATTR_STR_HYBRID(topdown-be-bound,      td_be_bound_adl,  "event=0x74,umask=0x0;event=0x00,umask=0x83", hybrid_big_small);
5214 EVENT_ATTR_STR_HYBRID(topdown-heavy-ops,     td_heavy_ops_adl, "event=0x00,umask=0x84",                      hybrid_big);
5215 EVENT_ATTR_STR_HYBRID(topdown-br-mispredict, td_br_mis_adl,    "event=0x00,umask=0x85",                      hybrid_big);
5216 EVENT_ATTR_STR_HYBRID(topdown-fetch-lat,     td_fetch_lat_adl, "event=0x00,umask=0x86",                      hybrid_big);
5217 EVENT_ATTR_STR_HYBRID(topdown-mem-bound,     td_mem_bound_adl, "event=0x00,umask=0x87",                      hybrid_big);
5218
5219 static struct attribute *adl_hybrid_events_attrs[] = {
5220         EVENT_PTR(slots_adl),
5221         EVENT_PTR(td_retiring_adl),
5222         EVENT_PTR(td_bad_spec_adl),
5223         EVENT_PTR(td_fe_bound_adl),
5224         EVENT_PTR(td_be_bound_adl),
5225         EVENT_PTR(td_heavy_ops_adl),
5226         EVENT_PTR(td_br_mis_adl),
5227         EVENT_PTR(td_fetch_lat_adl),
5228         EVENT_PTR(td_mem_bound_adl),
5229         NULL,
5230 };
5231
5232 /* Must be in IDX order */
5233 EVENT_ATTR_STR_HYBRID(mem-loads,     mem_ld_adl,     "event=0xd0,umask=0x5,ldlat=3;event=0xcd,umask=0x1,ldlat=3", hybrid_big_small);
5234 EVENT_ATTR_STR_HYBRID(mem-stores,    mem_st_adl,     "event=0xd0,umask=0x6;event=0xcd,umask=0x2",                 hybrid_big_small);
5235 EVENT_ATTR_STR_HYBRID(mem-loads-aux, mem_ld_aux_adl, "event=0x03,umask=0x82",                                     hybrid_big);
5236
5237 static struct attribute *adl_hybrid_mem_attrs[] = {
5238         EVENT_PTR(mem_ld_adl),
5239         EVENT_PTR(mem_st_adl),
5240         EVENT_PTR(mem_ld_aux_adl),
5241         NULL,
5242 };
5243
5244 EVENT_ATTR_STR_HYBRID(tx-start,          tx_start_adl,          "event=0xc9,umask=0x1",          hybrid_big);
5245 EVENT_ATTR_STR_HYBRID(tx-commit,         tx_commit_adl,         "event=0xc9,umask=0x2",          hybrid_big);
5246 EVENT_ATTR_STR_HYBRID(tx-abort,          tx_abort_adl,          "event=0xc9,umask=0x4",          hybrid_big);
5247 EVENT_ATTR_STR_HYBRID(tx-conflict,       tx_conflict_adl,       "event=0x54,umask=0x1",          hybrid_big);
5248 EVENT_ATTR_STR_HYBRID(cycles-t,          cycles_t_adl,          "event=0x3c,in_tx=1",            hybrid_big);
5249 EVENT_ATTR_STR_HYBRID(cycles-ct,         cycles_ct_adl,         "event=0x3c,in_tx=1,in_tx_cp=1", hybrid_big);
5250 EVENT_ATTR_STR_HYBRID(tx-capacity-read,  tx_capacity_read_adl,  "event=0x54,umask=0x80",         hybrid_big);
5251 EVENT_ATTR_STR_HYBRID(tx-capacity-write, tx_capacity_write_adl, "event=0x54,umask=0x2",          hybrid_big);
5252
5253 static struct attribute *adl_hybrid_tsx_attrs[] = {
5254         EVENT_PTR(tx_start_adl),
5255         EVENT_PTR(tx_abort_adl),
5256         EVENT_PTR(tx_commit_adl),
5257         EVENT_PTR(tx_capacity_read_adl),
5258         EVENT_PTR(tx_capacity_write_adl),
5259         EVENT_PTR(tx_conflict_adl),
5260         EVENT_PTR(cycles_t_adl),
5261         EVENT_PTR(cycles_ct_adl),
5262         NULL,
5263 };
5264
5265 FORMAT_ATTR_HYBRID(in_tx,       hybrid_big);
5266 FORMAT_ATTR_HYBRID(in_tx_cp,    hybrid_big);
5267 FORMAT_ATTR_HYBRID(offcore_rsp, hybrid_big_small);
5268 FORMAT_ATTR_HYBRID(ldlat,       hybrid_big_small);
5269 FORMAT_ATTR_HYBRID(frontend,    hybrid_big);
5270
5271 static struct attribute *adl_hybrid_extra_attr_rtm[] = {
5272         FORMAT_HYBRID_PTR(in_tx),
5273         FORMAT_HYBRID_PTR(in_tx_cp),
5274         FORMAT_HYBRID_PTR(offcore_rsp),
5275         FORMAT_HYBRID_PTR(ldlat),
5276         FORMAT_HYBRID_PTR(frontend),
5277         NULL,
5278 };
5279
5280 static struct attribute *adl_hybrid_extra_attr[] = {
5281         FORMAT_HYBRID_PTR(offcore_rsp),
5282         FORMAT_HYBRID_PTR(ldlat),
5283         FORMAT_HYBRID_PTR(frontend),
5284         NULL,
5285 };
5286
5287 static bool is_attr_for_this_pmu(struct kobject *kobj, struct attribute *attr)
5288 {
5289         struct device *dev = kobj_to_dev(kobj);
5290         struct x86_hybrid_pmu *pmu =
5291                 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
5292         struct perf_pmu_events_hybrid_attr *pmu_attr =
5293                 container_of(attr, struct perf_pmu_events_hybrid_attr, attr.attr);
5294
5295         return pmu->cpu_type & pmu_attr->pmu_type;
5296 }
5297
5298 static umode_t hybrid_events_is_visible(struct kobject *kobj,
5299                                         struct attribute *attr, int i)
5300 {
5301         return is_attr_for_this_pmu(kobj, attr) ? attr->mode : 0;
5302 }
5303
5304 static inline int hybrid_find_supported_cpu(struct x86_hybrid_pmu *pmu)
5305 {
5306         int cpu = cpumask_first(&pmu->supported_cpus);
5307
5308         return (cpu >= nr_cpu_ids) ? -1 : cpu;
5309 }
5310
5311 static umode_t hybrid_tsx_is_visible(struct kobject *kobj,
5312                                      struct attribute *attr, int i)
5313 {
5314         struct device *dev = kobj_to_dev(kobj);
5315         struct x86_hybrid_pmu *pmu =
5316                  container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
5317         int cpu = hybrid_find_supported_cpu(pmu);
5318
5319         return (cpu >= 0) && is_attr_for_this_pmu(kobj, attr) && cpu_has(&cpu_data(cpu), X86_FEATURE_RTM) ? attr->mode : 0;
5320 }
5321
5322 static umode_t hybrid_format_is_visible(struct kobject *kobj,
5323                                         struct attribute *attr, int i)
5324 {
5325         struct device *dev = kobj_to_dev(kobj);
5326         struct x86_hybrid_pmu *pmu =
5327                 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
5328         struct perf_pmu_format_hybrid_attr *pmu_attr =
5329                 container_of(attr, struct perf_pmu_format_hybrid_attr, attr.attr);
5330         int cpu = hybrid_find_supported_cpu(pmu);
5331
5332         return (cpu >= 0) && (pmu->cpu_type & pmu_attr->pmu_type) ? attr->mode : 0;
5333 }
5334
5335 static struct attribute_group hybrid_group_events_td  = {
5336         .name           = "events",
5337         .is_visible     = hybrid_events_is_visible,
5338 };
5339
5340 static struct attribute_group hybrid_group_events_mem = {
5341         .name           = "events",
5342         .is_visible     = hybrid_events_is_visible,
5343 };
5344
5345 static struct attribute_group hybrid_group_events_tsx = {
5346         .name           = "events",
5347         .is_visible     = hybrid_tsx_is_visible,
5348 };
5349
5350 static struct attribute_group hybrid_group_format_extra = {
5351         .name           = "format",
5352         .is_visible     = hybrid_format_is_visible,
5353 };
5354
5355 static ssize_t intel_hybrid_get_attr_cpus(struct device *dev,
5356                                           struct device_attribute *attr,
5357                                           char *buf)
5358 {
5359         struct x86_hybrid_pmu *pmu =
5360                 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
5361
5362         return cpumap_print_to_pagebuf(true, buf, &pmu->supported_cpus);
5363 }
5364
5365 static DEVICE_ATTR(cpus, S_IRUGO, intel_hybrid_get_attr_cpus, NULL);
5366 static struct attribute *intel_hybrid_cpus_attrs[] = {
5367         &dev_attr_cpus.attr,
5368         NULL,
5369 };
5370
5371 static struct attribute_group hybrid_group_cpus = {
5372         .attrs          = intel_hybrid_cpus_attrs,
5373 };
5374
5375 static const struct attribute_group *hybrid_attr_update[] = {
5376         &hybrid_group_events_td,
5377         &hybrid_group_events_mem,
5378         &hybrid_group_events_tsx,
5379         &group_caps_gen,
5380         &group_caps_lbr,
5381         &hybrid_group_format_extra,
5382         &group_default,
5383         &hybrid_group_cpus,
5384         NULL,
5385 };
5386
5387 static struct attribute *empty_attrs;
5388
5389 static void intel_pmu_check_num_counters(int *num_counters,
5390                                          int *num_counters_fixed,
5391                                          u64 *intel_ctrl, u64 fixed_mask)
5392 {
5393         if (*num_counters > INTEL_PMC_MAX_GENERIC) {
5394                 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
5395                      *num_counters, INTEL_PMC_MAX_GENERIC);
5396                 *num_counters = INTEL_PMC_MAX_GENERIC;
5397         }
5398         *intel_ctrl = (1ULL << *num_counters) - 1;
5399
5400         if (*num_counters_fixed > INTEL_PMC_MAX_FIXED) {
5401                 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
5402                      *num_counters_fixed, INTEL_PMC_MAX_FIXED);
5403                 *num_counters_fixed = INTEL_PMC_MAX_FIXED;
5404         }
5405
5406         *intel_ctrl |= fixed_mask << INTEL_PMC_IDX_FIXED;
5407 }
5408
5409 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints,
5410                                               int num_counters,
5411                                               int num_counters_fixed,
5412                                               u64 intel_ctrl)
5413 {
5414         struct event_constraint *c;
5415
5416         if (!event_constraints)
5417                 return;
5418
5419         /*
5420          * event on fixed counter2 (REF_CYCLES) only works on this
5421          * counter, so do not extend mask to generic counters
5422          */
5423         for_each_event_constraint(c, event_constraints) {
5424                 /*
5425                  * Don't extend the topdown slots and metrics
5426                  * events to the generic counters.
5427                  */
5428                 if (c->idxmsk64 & INTEL_PMC_MSK_TOPDOWN) {
5429                         /*
5430                          * Disable topdown slots and metrics events,
5431                          * if slots event is not in CPUID.
5432                          */
5433                         if (!(INTEL_PMC_MSK_FIXED_SLOTS & intel_ctrl))
5434                                 c->idxmsk64 = 0;
5435                         c->weight = hweight64(c->idxmsk64);
5436                         continue;
5437                 }
5438
5439                 if (c->cmask == FIXED_EVENT_FLAGS) {
5440                         /* Disabled fixed counters which are not in CPUID */
5441                         c->idxmsk64 &= intel_ctrl;
5442
5443                         if (c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES)
5444                                 c->idxmsk64 |= (1ULL << num_counters) - 1;
5445                 }
5446                 c->idxmsk64 &=
5447                         ~(~0ULL << (INTEL_PMC_IDX_FIXED + num_counters_fixed));
5448                 c->weight = hweight64(c->idxmsk64);
5449         }
5450 }
5451
5452 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs)
5453 {
5454         struct extra_reg *er;
5455
5456         /*
5457          * Access extra MSR may cause #GP under certain circumstances.
5458          * E.g. KVM doesn't support offcore event
5459          * Check all extra_regs here.
5460          */
5461         if (!extra_regs)
5462                 return;
5463
5464         for (er = extra_regs; er->msr; er++) {
5465                 er->extra_msr_access = check_msr(er->msr, 0x11UL);
5466                 /* Disable LBR select mapping */
5467                 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
5468                         x86_pmu.lbr_sel_map = NULL;
5469         }
5470 }
5471
5472 static void intel_pmu_check_hybrid_pmus(u64 fixed_mask)
5473 {
5474         struct x86_hybrid_pmu *pmu;
5475         int i;
5476
5477         for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) {
5478                 pmu = &x86_pmu.hybrid_pmu[i];
5479
5480                 intel_pmu_check_num_counters(&pmu->num_counters,
5481                                              &pmu->num_counters_fixed,
5482                                              &pmu->intel_ctrl,
5483                                              fixed_mask);
5484
5485                 if (pmu->intel_cap.perf_metrics) {
5486                         pmu->intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS;
5487                         pmu->intel_ctrl |= INTEL_PMC_MSK_FIXED_SLOTS;
5488                 }
5489
5490                 if (pmu->intel_cap.pebs_output_pt_available)
5491                         pmu->pmu.capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
5492
5493                 intel_pmu_check_event_constraints(pmu->event_constraints,
5494                                                   pmu->num_counters,
5495                                                   pmu->num_counters_fixed,
5496                                                   pmu->intel_ctrl);
5497
5498                 intel_pmu_check_extra_regs(pmu->extra_regs);
5499         }
5500 }
5501
5502 __init int intel_pmu_init(void)
5503 {
5504         struct attribute **extra_skl_attr = &empty_attrs;
5505         struct attribute **extra_attr = &empty_attrs;
5506         struct attribute **td_attr    = &empty_attrs;
5507         struct attribute **mem_attr   = &empty_attrs;
5508         struct attribute **tsx_attr   = &empty_attrs;
5509         union cpuid10_edx edx;
5510         union cpuid10_eax eax;
5511         union cpuid10_ebx ebx;
5512         unsigned int fixed_mask;
5513         bool pmem = false;
5514         int version, i;
5515         char *name;
5516         struct x86_hybrid_pmu *pmu;
5517
5518         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
5519                 switch (boot_cpu_data.x86) {
5520                 case 0x6:
5521                         return p6_pmu_init();
5522                 case 0xb:
5523                         return knc_pmu_init();
5524                 case 0xf:
5525                         return p4_pmu_init();
5526                 }
5527                 return -ENODEV;
5528         }
5529
5530         /*
5531          * Check whether the Architectural PerfMon supports
5532          * Branch Misses Retired hw_event or not.
5533          */
5534         cpuid(10, &eax.full, &ebx.full, &fixed_mask, &edx.full);
5535         if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
5536                 return -ENODEV;
5537
5538         version = eax.split.version_id;
5539         if (version < 2)
5540                 x86_pmu = core_pmu;
5541         else
5542                 x86_pmu = intel_pmu;
5543
5544         x86_pmu.version                 = version;
5545         x86_pmu.num_counters            = eax.split.num_counters;
5546         x86_pmu.cntval_bits             = eax.split.bit_width;
5547         x86_pmu.cntval_mask             = (1ULL << eax.split.bit_width) - 1;
5548
5549         x86_pmu.events_maskl            = ebx.full;
5550         x86_pmu.events_mask_len         = eax.split.mask_length;
5551
5552         x86_pmu.max_pebs_events         = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
5553
5554         /*
5555          * Quirk: v2 perfmon does not report fixed-purpose events, so
5556          * assume at least 3 events, when not running in a hypervisor:
5557          */
5558         if (version > 1 && version < 5) {
5559                 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR);
5560
5561                 x86_pmu.num_counters_fixed =
5562                         max((int)edx.split.num_counters_fixed, assume);
5563
5564                 fixed_mask = (1L << x86_pmu.num_counters_fixed) - 1;
5565         } else if (version >= 5)
5566                 x86_pmu.num_counters_fixed = fls(fixed_mask);
5567
5568         if (boot_cpu_has(X86_FEATURE_PDCM)) {
5569                 u64 capabilities;
5570
5571                 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
5572                 x86_pmu.intel_cap.capabilities = capabilities;
5573         }
5574
5575         if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32) {
5576                 x86_pmu.lbr_reset = intel_pmu_lbr_reset_32;
5577                 x86_pmu.lbr_read = intel_pmu_lbr_read_32;
5578         }
5579
5580         if (boot_cpu_has(X86_FEATURE_ARCH_LBR))
5581                 intel_pmu_arch_lbr_init();
5582
5583         intel_ds_init();
5584
5585         x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
5586
5587         if (version >= 5) {
5588                 x86_pmu.intel_cap.anythread_deprecated = edx.split.anythread_deprecated;
5589                 if (x86_pmu.intel_cap.anythread_deprecated)
5590                         pr_cont(" AnyThread deprecated, ");
5591         }
5592
5593         /*
5594          * Install the hw-cache-events table:
5595          */
5596         switch (boot_cpu_data.x86_model) {
5597         case INTEL_FAM6_CORE_YONAH:
5598                 pr_cont("Core events, ");
5599                 name = "core";
5600                 break;
5601
5602         case INTEL_FAM6_CORE2_MEROM:
5603                 x86_add_quirk(intel_clovertown_quirk);
5604                 fallthrough;
5605
5606         case INTEL_FAM6_CORE2_MEROM_L:
5607         case INTEL_FAM6_CORE2_PENRYN:
5608         case INTEL_FAM6_CORE2_DUNNINGTON:
5609                 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
5610                        sizeof(hw_cache_event_ids));
5611
5612                 intel_pmu_lbr_init_core();
5613
5614                 x86_pmu.event_constraints = intel_core2_event_constraints;
5615                 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
5616                 pr_cont("Core2 events, ");
5617                 name = "core2";
5618                 break;
5619
5620         case INTEL_FAM6_NEHALEM:
5621         case INTEL_FAM6_NEHALEM_EP:
5622         case INTEL_FAM6_NEHALEM_EX:
5623                 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
5624                        sizeof(hw_cache_event_ids));
5625                 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
5626                        sizeof(hw_cache_extra_regs));
5627
5628                 intel_pmu_lbr_init_nhm();
5629
5630                 x86_pmu.event_constraints = intel_nehalem_event_constraints;
5631                 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
5632                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
5633                 x86_pmu.extra_regs = intel_nehalem_extra_regs;
5634                 x86_pmu.limit_period = nhm_limit_period;
5635
5636                 mem_attr = nhm_mem_events_attrs;
5637
5638                 /* UOPS_ISSUED.STALLED_CYCLES */
5639                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5640                         X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5641                 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
5642                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
5643                         X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
5644
5645                 intel_pmu_pebs_data_source_nhm();
5646                 x86_add_quirk(intel_nehalem_quirk);
5647                 x86_pmu.pebs_no_tlb = 1;
5648                 extra_attr = nhm_format_attr;
5649
5650                 pr_cont("Nehalem events, ");
5651                 name = "nehalem";
5652                 break;
5653
5654         case INTEL_FAM6_ATOM_BONNELL:
5655         case INTEL_FAM6_ATOM_BONNELL_MID:
5656         case INTEL_FAM6_ATOM_SALTWELL:
5657         case INTEL_FAM6_ATOM_SALTWELL_MID:
5658         case INTEL_FAM6_ATOM_SALTWELL_TABLET:
5659                 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
5660                        sizeof(hw_cache_event_ids));
5661
5662                 intel_pmu_lbr_init_atom();
5663
5664                 x86_pmu.event_constraints = intel_gen_event_constraints;
5665                 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
5666                 x86_pmu.pebs_aliases = intel_pebs_aliases_core2;
5667                 pr_cont("Atom events, ");
5668                 name = "bonnell";
5669                 break;
5670
5671         case INTEL_FAM6_ATOM_SILVERMONT:
5672         case INTEL_FAM6_ATOM_SILVERMONT_D:
5673         case INTEL_FAM6_ATOM_SILVERMONT_MID:
5674         case INTEL_FAM6_ATOM_AIRMONT:
5675         case INTEL_FAM6_ATOM_AIRMONT_MID:
5676                 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids,
5677                         sizeof(hw_cache_event_ids));
5678                 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs,
5679                        sizeof(hw_cache_extra_regs));
5680
5681                 intel_pmu_lbr_init_slm();
5682
5683                 x86_pmu.event_constraints = intel_slm_event_constraints;
5684                 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
5685                 x86_pmu.extra_regs = intel_slm_extra_regs;
5686                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5687                 td_attr = slm_events_attrs;
5688                 extra_attr = slm_format_attr;
5689                 pr_cont("Silvermont events, ");
5690                 name = "silvermont";
5691                 break;
5692
5693         case INTEL_FAM6_ATOM_GOLDMONT:
5694         case INTEL_FAM6_ATOM_GOLDMONT_D:
5695                 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids,
5696                        sizeof(hw_cache_event_ids));
5697                 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs,
5698                        sizeof(hw_cache_extra_regs));
5699
5700                 intel_pmu_lbr_init_skl();
5701
5702                 x86_pmu.event_constraints = intel_slm_event_constraints;
5703                 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints;
5704                 x86_pmu.extra_regs = intel_glm_extra_regs;
5705                 /*
5706                  * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
5707                  * for precise cycles.
5708                  * :pp is identical to :ppp
5709                  */
5710                 x86_pmu.pebs_aliases = NULL;
5711                 x86_pmu.pebs_prec_dist = true;
5712                 x86_pmu.lbr_pt_coexist = true;
5713                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5714                 td_attr = glm_events_attrs;
5715                 extra_attr = slm_format_attr;
5716                 pr_cont("Goldmont events, ");
5717                 name = "goldmont";
5718                 break;
5719
5720         case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
5721                 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
5722                        sizeof(hw_cache_event_ids));
5723                 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs,
5724                        sizeof(hw_cache_extra_regs));
5725
5726                 intel_pmu_lbr_init_skl();
5727
5728                 x86_pmu.event_constraints = intel_slm_event_constraints;
5729                 x86_pmu.extra_regs = intel_glm_extra_regs;
5730                 /*
5731                  * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
5732                  * for precise cycles.
5733                  */
5734                 x86_pmu.pebs_aliases = NULL;
5735                 x86_pmu.pebs_prec_dist = true;
5736                 x86_pmu.lbr_pt_coexist = true;
5737                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5738                 x86_pmu.flags |= PMU_FL_PEBS_ALL;
5739                 x86_pmu.get_event_constraints = glp_get_event_constraints;
5740                 td_attr = glm_events_attrs;
5741                 /* Goldmont Plus has 4-wide pipeline */
5742                 event_attr_td_total_slots_scale_glm.event_str = "4";
5743                 extra_attr = slm_format_attr;
5744                 pr_cont("Goldmont plus events, ");
5745                 name = "goldmont_plus";
5746                 break;
5747
5748         case INTEL_FAM6_ATOM_TREMONT_D:
5749         case INTEL_FAM6_ATOM_TREMONT:
5750         case INTEL_FAM6_ATOM_TREMONT_L:
5751                 x86_pmu.late_ack = true;
5752                 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
5753                        sizeof(hw_cache_event_ids));
5754                 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs,
5755                        sizeof(hw_cache_extra_regs));
5756                 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
5757
5758                 intel_pmu_lbr_init_skl();
5759
5760                 x86_pmu.event_constraints = intel_slm_event_constraints;
5761                 x86_pmu.extra_regs = intel_tnt_extra_regs;
5762                 /*
5763                  * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
5764                  * for precise cycles.
5765                  */
5766                 x86_pmu.pebs_aliases = NULL;
5767                 x86_pmu.pebs_prec_dist = true;
5768                 x86_pmu.lbr_pt_coexist = true;
5769                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5770                 x86_pmu.get_event_constraints = tnt_get_event_constraints;
5771                 td_attr = tnt_events_attrs;
5772                 extra_attr = slm_format_attr;
5773                 pr_cont("Tremont events, ");
5774                 name = "Tremont";
5775                 break;
5776
5777         case INTEL_FAM6_WESTMERE:
5778         case INTEL_FAM6_WESTMERE_EP:
5779         case INTEL_FAM6_WESTMERE_EX:
5780                 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
5781                        sizeof(hw_cache_event_ids));
5782                 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
5783                        sizeof(hw_cache_extra_regs));
5784
5785                 intel_pmu_lbr_init_nhm();
5786
5787                 x86_pmu.event_constraints = intel_westmere_event_constraints;
5788                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
5789                 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
5790                 x86_pmu.extra_regs = intel_westmere_extra_regs;
5791                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5792
5793                 mem_attr = nhm_mem_events_attrs;
5794
5795                 /* UOPS_ISSUED.STALLED_CYCLES */
5796                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5797                         X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5798                 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
5799                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
5800                         X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
5801
5802                 intel_pmu_pebs_data_source_nhm();
5803                 extra_attr = nhm_format_attr;
5804                 pr_cont("Westmere events, ");
5805                 name = "westmere";
5806                 break;
5807
5808         case INTEL_FAM6_SANDYBRIDGE:
5809         case INTEL_FAM6_SANDYBRIDGE_X:
5810                 x86_add_quirk(intel_sandybridge_quirk);
5811                 x86_add_quirk(intel_ht_bug);
5812                 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
5813                        sizeof(hw_cache_event_ids));
5814                 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
5815                        sizeof(hw_cache_extra_regs));
5816
5817                 intel_pmu_lbr_init_snb();
5818
5819                 x86_pmu.event_constraints = intel_snb_event_constraints;
5820                 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
5821                 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
5822                 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X)
5823                         x86_pmu.extra_regs = intel_snbep_extra_regs;
5824                 else
5825                         x86_pmu.extra_regs = intel_snb_extra_regs;
5826
5827
5828                 /* all extra regs are per-cpu when HT is on */
5829                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5830                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5831
5832                 td_attr  = snb_events_attrs;
5833                 mem_attr = snb_mem_events_attrs;
5834
5835                 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
5836                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5837                         X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5838                 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
5839                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
5840                         X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
5841
5842                 extra_attr = nhm_format_attr;
5843
5844                 pr_cont("SandyBridge events, ");
5845                 name = "sandybridge";
5846                 break;
5847
5848         case INTEL_FAM6_IVYBRIDGE:
5849         case INTEL_FAM6_IVYBRIDGE_X:
5850                 x86_add_quirk(intel_ht_bug);
5851                 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
5852                        sizeof(hw_cache_event_ids));
5853                 /* dTLB-load-misses on IVB is different than SNB */
5854                 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
5855
5856                 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
5857                        sizeof(hw_cache_extra_regs));
5858
5859                 intel_pmu_lbr_init_snb();
5860
5861                 x86_pmu.event_constraints = intel_ivb_event_constraints;
5862                 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
5863                 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
5864                 x86_pmu.pebs_prec_dist = true;
5865                 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X)
5866                         x86_pmu.extra_regs = intel_snbep_extra_regs;
5867                 else
5868                         x86_pmu.extra_regs = intel_snb_extra_regs;
5869                 /* all extra regs are per-cpu when HT is on */
5870                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5871                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5872
5873                 td_attr  = snb_events_attrs;
5874                 mem_attr = snb_mem_events_attrs;
5875
5876                 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
5877                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5878                         X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5879
5880                 extra_attr = nhm_format_attr;
5881
5882                 pr_cont("IvyBridge events, ");
5883                 name = "ivybridge";
5884                 break;
5885
5886
5887         case INTEL_FAM6_HASWELL:
5888         case INTEL_FAM6_HASWELL_X:
5889         case INTEL_FAM6_HASWELL_L:
5890         case INTEL_FAM6_HASWELL_G:
5891                 x86_add_quirk(intel_ht_bug);
5892                 x86_add_quirk(intel_pebs_isolation_quirk);
5893                 x86_pmu.late_ack = true;
5894                 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5895                 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5896
5897                 intel_pmu_lbr_init_hsw();
5898
5899                 x86_pmu.event_constraints = intel_hsw_event_constraints;
5900                 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
5901                 x86_pmu.extra_regs = intel_snbep_extra_regs;
5902                 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
5903                 x86_pmu.pebs_prec_dist = true;
5904                 /* all extra regs are per-cpu when HT is on */
5905                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5906                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5907
5908                 x86_pmu.hw_config = hsw_hw_config;
5909                 x86_pmu.get_event_constraints = hsw_get_event_constraints;
5910                 x86_pmu.lbr_double_abort = true;
5911                 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5912                         hsw_format_attr : nhm_format_attr;
5913                 td_attr  = hsw_events_attrs;
5914                 mem_attr = hsw_mem_events_attrs;
5915                 tsx_attr = hsw_tsx_events_attrs;
5916                 pr_cont("Haswell events, ");
5917                 name = "haswell";
5918                 break;
5919
5920         case INTEL_FAM6_BROADWELL:
5921         case INTEL_FAM6_BROADWELL_D:
5922         case INTEL_FAM6_BROADWELL_G:
5923         case INTEL_FAM6_BROADWELL_X:
5924                 x86_add_quirk(intel_pebs_isolation_quirk);
5925                 x86_pmu.late_ack = true;
5926                 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5927                 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5928
5929                 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
5930                 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ |
5931                                                                          BDW_L3_MISS|HSW_SNOOP_DRAM;
5932                 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS|
5933                                                                           HSW_SNOOP_DRAM;
5934                 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ|
5935                                                                              BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
5936                 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE|
5937                                                                               BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
5938
5939                 intel_pmu_lbr_init_hsw();
5940
5941                 x86_pmu.event_constraints = intel_bdw_event_constraints;
5942                 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints;
5943                 x86_pmu.extra_regs = intel_snbep_extra_regs;
5944                 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
5945                 x86_pmu.pebs_prec_dist = true;
5946                 /* all extra regs are per-cpu when HT is on */
5947                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5948                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5949
5950                 x86_pmu.hw_config = hsw_hw_config;
5951                 x86_pmu.get_event_constraints = hsw_get_event_constraints;
5952                 x86_pmu.limit_period = bdw_limit_period;
5953                 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5954                         hsw_format_attr : nhm_format_attr;
5955                 td_attr  = hsw_events_attrs;
5956                 mem_attr = hsw_mem_events_attrs;
5957                 tsx_attr = hsw_tsx_events_attrs;
5958                 pr_cont("Broadwell events, ");
5959                 name = "broadwell";
5960                 break;
5961
5962         case INTEL_FAM6_XEON_PHI_KNL:
5963         case INTEL_FAM6_XEON_PHI_KNM:
5964                 memcpy(hw_cache_event_ids,
5965                        slm_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5966                 memcpy(hw_cache_extra_regs,
5967                        knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5968                 intel_pmu_lbr_init_knl();
5969
5970                 x86_pmu.event_constraints = intel_slm_event_constraints;
5971                 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
5972                 x86_pmu.extra_regs = intel_knl_extra_regs;
5973
5974                 /* all extra regs are per-cpu when HT is on */
5975                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5976                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5977                 extra_attr = slm_format_attr;
5978                 pr_cont("Knights Landing/Mill events, ");
5979                 name = "knights-landing";
5980                 break;
5981
5982         case INTEL_FAM6_SKYLAKE_X:
5983                 pmem = true;
5984                 fallthrough;
5985         case INTEL_FAM6_SKYLAKE_L:
5986         case INTEL_FAM6_SKYLAKE:
5987         case INTEL_FAM6_KABYLAKE_L:
5988         case INTEL_FAM6_KABYLAKE:
5989         case INTEL_FAM6_COMETLAKE_L:
5990         case INTEL_FAM6_COMETLAKE:
5991                 x86_add_quirk(intel_pebs_isolation_quirk);
5992                 x86_pmu.late_ack = true;
5993                 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5994                 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5995                 intel_pmu_lbr_init_skl();
5996
5997                 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */
5998                 event_attr_td_recovery_bubbles.event_str_noht =
5999                         "event=0xd,umask=0x1,cmask=1";
6000                 event_attr_td_recovery_bubbles.event_str_ht =
6001                         "event=0xd,umask=0x1,cmask=1,any=1";
6002
6003                 x86_pmu.event_constraints = intel_skl_event_constraints;
6004                 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints;
6005                 x86_pmu.extra_regs = intel_skl_extra_regs;
6006                 x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
6007                 x86_pmu.pebs_prec_dist = true;
6008                 /* all extra regs are per-cpu when HT is on */
6009                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6010                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6011
6012                 x86_pmu.hw_config = hsw_hw_config;
6013                 x86_pmu.get_event_constraints = hsw_get_event_constraints;
6014                 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
6015                         hsw_format_attr : nhm_format_attr;
6016                 extra_skl_attr = skl_format_attr;
6017                 td_attr  = hsw_events_attrs;
6018                 mem_attr = hsw_mem_events_attrs;
6019                 tsx_attr = hsw_tsx_events_attrs;
6020                 intel_pmu_pebs_data_source_skl(pmem);
6021
6022                 /*
6023                  * Processors with CPUID.RTM_ALWAYS_ABORT have TSX deprecated by default.
6024                  * TSX force abort hooks are not required on these systems. Only deploy
6025                  * workaround when microcode has not enabled X86_FEATURE_RTM_ALWAYS_ABORT.
6026                  */
6027                 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT) &&
6028                    !boot_cpu_has(X86_FEATURE_RTM_ALWAYS_ABORT)) {
6029                         x86_pmu.flags |= PMU_FL_TFA;
6030                         x86_pmu.get_event_constraints = tfa_get_event_constraints;
6031                         x86_pmu.enable_all = intel_tfa_pmu_enable_all;
6032                         x86_pmu.commit_scheduling = intel_tfa_commit_scheduling;
6033                 }
6034
6035                 pr_cont("Skylake events, ");
6036                 name = "skylake";
6037                 break;
6038
6039         case INTEL_FAM6_ICELAKE_X:
6040         case INTEL_FAM6_ICELAKE_D:
6041                 pmem = true;
6042                 fallthrough;
6043         case INTEL_FAM6_ICELAKE_L:
6044         case INTEL_FAM6_ICELAKE:
6045         case INTEL_FAM6_TIGERLAKE_L:
6046         case INTEL_FAM6_TIGERLAKE:
6047         case INTEL_FAM6_ROCKETLAKE:
6048                 x86_pmu.late_ack = true;
6049                 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
6050                 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
6051                 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
6052                 intel_pmu_lbr_init_skl();
6053
6054                 x86_pmu.event_constraints = intel_icl_event_constraints;
6055                 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints;
6056                 x86_pmu.extra_regs = intel_icl_extra_regs;
6057                 x86_pmu.pebs_aliases = NULL;
6058                 x86_pmu.pebs_prec_dist = true;
6059                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6060                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6061
6062                 x86_pmu.hw_config = hsw_hw_config;
6063                 x86_pmu.get_event_constraints = icl_get_event_constraints;
6064                 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
6065                         hsw_format_attr : nhm_format_attr;
6066                 extra_skl_attr = skl_format_attr;
6067                 mem_attr = icl_events_attrs;
6068                 td_attr = icl_td_events_attrs;
6069                 tsx_attr = icl_tsx_events_attrs;
6070                 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
6071                 x86_pmu.lbr_pt_coexist = true;
6072                 intel_pmu_pebs_data_source_skl(pmem);
6073                 x86_pmu.num_topdown_events = 4;
6074                 x86_pmu.update_topdown_event = icl_update_topdown_event;
6075                 x86_pmu.set_topdown_event_period = icl_set_topdown_event_period;
6076                 pr_cont("Icelake events, ");
6077                 name = "icelake";
6078                 break;
6079
6080         case INTEL_FAM6_SAPPHIRERAPIDS_X:
6081                 pmem = true;
6082                 x86_pmu.late_ack = true;
6083                 memcpy(hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(hw_cache_event_ids));
6084                 memcpy(hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
6085
6086                 x86_pmu.event_constraints = intel_spr_event_constraints;
6087                 x86_pmu.pebs_constraints = intel_spr_pebs_event_constraints;
6088                 x86_pmu.extra_regs = intel_spr_extra_regs;
6089                 x86_pmu.limit_period = spr_limit_period;
6090                 x86_pmu.pebs_aliases = NULL;
6091                 x86_pmu.pebs_prec_dist = true;
6092                 x86_pmu.pebs_block = true;
6093                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6094                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6095                 x86_pmu.flags |= PMU_FL_PEBS_ALL;
6096                 x86_pmu.flags |= PMU_FL_INSTR_LATENCY;
6097                 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX;
6098
6099                 x86_pmu.hw_config = hsw_hw_config;
6100                 x86_pmu.get_event_constraints = spr_get_event_constraints;
6101                 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
6102                         hsw_format_attr : nhm_format_attr;
6103                 extra_skl_attr = skl_format_attr;
6104                 mem_attr = spr_events_attrs;
6105                 td_attr = spr_td_events_attrs;
6106                 tsx_attr = spr_tsx_events_attrs;
6107                 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
6108                 x86_pmu.lbr_pt_coexist = true;
6109                 intel_pmu_pebs_data_source_skl(pmem);
6110                 x86_pmu.num_topdown_events = 8;
6111                 x86_pmu.update_topdown_event = icl_update_topdown_event;
6112                 x86_pmu.set_topdown_event_period = icl_set_topdown_event_period;
6113                 pr_cont("Sapphire Rapids events, ");
6114                 name = "sapphire_rapids";
6115                 break;
6116
6117         case INTEL_FAM6_ALDERLAKE:
6118         case INTEL_FAM6_ALDERLAKE_L:
6119                 /*
6120                  * Alder Lake has 2 types of CPU, core and atom.
6121                  *
6122                  * Initialize the common PerfMon capabilities here.
6123                  */
6124                 x86_pmu.hybrid_pmu = kcalloc(X86_HYBRID_NUM_PMUS,
6125                                              sizeof(struct x86_hybrid_pmu),
6126                                              GFP_KERNEL);
6127                 if (!x86_pmu.hybrid_pmu)
6128                         return -ENOMEM;
6129                 static_branch_enable(&perf_is_hybrid);
6130                 x86_pmu.num_hybrid_pmus = X86_HYBRID_NUM_PMUS;
6131
6132                 x86_pmu.late_ack = true;
6133                 x86_pmu.pebs_aliases = NULL;
6134                 x86_pmu.pebs_prec_dist = true;
6135                 x86_pmu.pebs_block = true;
6136                 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6137                 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6138                 x86_pmu.flags |= PMU_FL_PEBS_ALL;
6139                 x86_pmu.flags |= PMU_FL_INSTR_LATENCY;
6140                 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX;
6141                 x86_pmu.lbr_pt_coexist = true;
6142                 intel_pmu_pebs_data_source_skl(false);
6143                 x86_pmu.num_topdown_events = 8;
6144                 x86_pmu.update_topdown_event = adl_update_topdown_event;
6145                 x86_pmu.set_topdown_event_period = adl_set_topdown_event_period;
6146
6147                 x86_pmu.filter_match = intel_pmu_filter_match;
6148                 x86_pmu.get_event_constraints = adl_get_event_constraints;
6149                 x86_pmu.hw_config = adl_hw_config;
6150                 x86_pmu.limit_period = spr_limit_period;
6151                 x86_pmu.get_hybrid_cpu_type = adl_get_hybrid_cpu_type;
6152                 /*
6153                  * The rtm_abort_event is used to check whether to enable GPRs
6154                  * for the RTM abort event. Atom doesn't have the RTM abort
6155                  * event. There is no harmful to set it in the common
6156                  * x86_pmu.rtm_abort_event.
6157                  */
6158                 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
6159
6160                 td_attr = adl_hybrid_events_attrs;
6161                 mem_attr = adl_hybrid_mem_attrs;
6162                 tsx_attr = adl_hybrid_tsx_attrs;
6163                 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
6164                         adl_hybrid_extra_attr_rtm : adl_hybrid_extra_attr;
6165
6166                 /* Initialize big core specific PerfMon capabilities.*/
6167                 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX];
6168                 pmu->name = "cpu_core";
6169                 pmu->cpu_type = hybrid_big;
6170                 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) {
6171                         pmu->num_counters = x86_pmu.num_counters + 2;
6172                         pmu->num_counters_fixed = x86_pmu.num_counters_fixed + 1;
6173                 } else {
6174                         pmu->num_counters = x86_pmu.num_counters;
6175                         pmu->num_counters_fixed = x86_pmu.num_counters_fixed;
6176                 }
6177                 pmu->max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, pmu->num_counters);
6178                 pmu->unconstrained = (struct event_constraint)
6179                                         __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1,
6180                                                            0, pmu->num_counters, 0, 0);
6181                 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities;
6182                 pmu->intel_cap.perf_metrics = 1;
6183                 pmu->intel_cap.pebs_output_pt_available = 0;
6184
6185                 memcpy(pmu->hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(pmu->hw_cache_event_ids));
6186                 memcpy(pmu->hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(pmu->hw_cache_extra_regs));
6187                 pmu->event_constraints = intel_spr_event_constraints;
6188                 pmu->pebs_constraints = intel_spr_pebs_event_constraints;
6189                 pmu->extra_regs = intel_spr_extra_regs;
6190
6191                 /* Initialize Atom core specific PerfMon capabilities.*/
6192                 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX];
6193                 pmu->name = "cpu_atom";
6194                 pmu->cpu_type = hybrid_small;
6195                 pmu->num_counters = x86_pmu.num_counters;
6196                 pmu->num_counters_fixed = x86_pmu.num_counters_fixed;
6197                 pmu->max_pebs_events = x86_pmu.max_pebs_events;
6198                 pmu->unconstrained = (struct event_constraint)
6199                                         __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1,
6200                                                            0, pmu->num_counters, 0, 0);
6201                 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities;
6202                 pmu->intel_cap.perf_metrics = 0;
6203                 pmu->intel_cap.pebs_output_pt_available = 1;
6204
6205                 memcpy(pmu->hw_cache_event_ids, glp_hw_cache_event_ids, sizeof(pmu->hw_cache_event_ids));
6206                 memcpy(pmu->hw_cache_extra_regs, tnt_hw_cache_extra_regs, sizeof(pmu->hw_cache_extra_regs));
6207                 pmu->hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
6208                 pmu->event_constraints = intel_slm_event_constraints;
6209                 pmu->pebs_constraints = intel_grt_pebs_event_constraints;
6210                 pmu->extra_regs = intel_grt_extra_regs;
6211                 pr_cont("Alderlake Hybrid events, ");
6212                 name = "alderlake_hybrid";
6213                 break;
6214
6215         default:
6216                 switch (x86_pmu.version) {
6217                 case 1:
6218                         x86_pmu.event_constraints = intel_v1_event_constraints;
6219                         pr_cont("generic architected perfmon v1, ");
6220                         name = "generic_arch_v1";
6221                         break;
6222                 default:
6223                         /*
6224                          * default constraints for v2 and up
6225                          */
6226                         x86_pmu.event_constraints = intel_gen_event_constraints;
6227                         pr_cont("generic architected perfmon, ");
6228                         name = "generic_arch_v2+";
6229                         break;
6230                 }
6231         }
6232
6233         snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name);
6234
6235         if (!is_hybrid()) {
6236                 group_events_td.attrs  = td_attr;
6237                 group_events_mem.attrs = mem_attr;
6238                 group_events_tsx.attrs = tsx_attr;
6239                 group_format_extra.attrs = extra_attr;
6240                 group_format_extra_skl.attrs = extra_skl_attr;
6241
6242                 x86_pmu.attr_update = attr_update;
6243         } else {
6244                 hybrid_group_events_td.attrs  = td_attr;
6245                 hybrid_group_events_mem.attrs = mem_attr;
6246                 hybrid_group_events_tsx.attrs = tsx_attr;
6247                 hybrid_group_format_extra.attrs = extra_attr;
6248
6249                 x86_pmu.attr_update = hybrid_attr_update;
6250         }
6251
6252         intel_pmu_check_num_counters(&x86_pmu.num_counters,
6253                                      &x86_pmu.num_counters_fixed,
6254                                      &x86_pmu.intel_ctrl,
6255                                      (u64)fixed_mask);
6256
6257         /* AnyThread may be deprecated on arch perfmon v5 or later */
6258         if (x86_pmu.intel_cap.anythread_deprecated)
6259                 x86_pmu.format_attrs = intel_arch_formats_attr;
6260
6261         intel_pmu_check_event_constraints(x86_pmu.event_constraints,
6262                                           x86_pmu.num_counters,
6263                                           x86_pmu.num_counters_fixed,
6264                                           x86_pmu.intel_ctrl);
6265         /*
6266          * Access LBR MSR may cause #GP under certain circumstances.
6267          * E.g. KVM doesn't support LBR MSR
6268          * Check all LBT MSR here.
6269          * Disable LBR access if any LBR MSRs can not be accessed.
6270          */
6271         if (x86_pmu.lbr_tos && !check_msr(x86_pmu.lbr_tos, 0x3UL))
6272                 x86_pmu.lbr_nr = 0;
6273         for (i = 0; i < x86_pmu.lbr_nr; i++) {
6274                 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) &&
6275                       check_msr(x86_pmu.lbr_to + i, 0xffffUL)))
6276                         x86_pmu.lbr_nr = 0;
6277         }
6278
6279         if (x86_pmu.lbr_nr)
6280                 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr);
6281
6282         intel_pmu_check_extra_regs(x86_pmu.extra_regs);
6283
6284         /* Support full width counters using alternative MSR range */
6285         if (x86_pmu.intel_cap.full_width_write) {
6286                 x86_pmu.max_period = x86_pmu.cntval_mask >> 1;
6287                 x86_pmu.perfctr = MSR_IA32_PMC0;
6288                 pr_cont("full-width counters, ");
6289         }
6290
6291         if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics)
6292                 x86_pmu.intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS;
6293
6294         if (is_hybrid())
6295                 intel_pmu_check_hybrid_pmus((u64)fixed_mask);
6296
6297         return 0;
6298 }
6299
6300 /*
6301  * HT bug: phase 2 init
6302  * Called once we have valid topology information to check
6303  * whether or not HT is enabled
6304  * If HT is off, then we disable the workaround
6305  */
6306 static __init int fixup_ht_bug(void)
6307 {
6308         int c;
6309         /*
6310          * problem not present on this CPU model, nothing to do
6311          */
6312         if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED))
6313                 return 0;
6314
6315         if (topology_max_smt_threads() > 1) {
6316                 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
6317                 return 0;
6318         }
6319
6320         cpus_read_lock();
6321
6322         hardlockup_detector_perf_stop();
6323
6324         x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED);
6325
6326         x86_pmu.start_scheduling = NULL;
6327         x86_pmu.commit_scheduling = NULL;
6328         x86_pmu.stop_scheduling = NULL;
6329
6330         hardlockup_detector_perf_restart();
6331
6332         for_each_online_cpu(c)
6333                 free_excl_cntrs(&per_cpu(cpu_hw_events, c));
6334
6335         cpus_read_unlock();
6336         pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
6337         return 0;
6338 }
6339 subsys_initcall(fixup_ht_bug)