arch/x86/events/intel/ds.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/bitops.h>
   3 #include <linux/types.h>
   4 #include <linux/slab.h>
   5
   6 #include <asm/cpu_entry_area.h>
   7 #include <asm/perf_event.h>
   8 #include <asm/tlbflush.h>
   9 #include <asm/insn.h>
  10
  11 #include "../perf_event.h"
  12
  13 /* Waste a full page so it can be mapped into the cpu_entry_area */
  14 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
  15
  16 /* The size of a BTS record in bytes: */
  17 #define BTS_RECORD_SIZE         24
  18
  19 #define PEBS_FIXUP_SIZE         PAGE_SIZE
  20
  21 /*
  22  * pebs_record_32 for p4 and core not supported
  23
  24 struct pebs_record_32 {
  25         u32 flags, ip;
  26         u32 ax, bc, cx, dx;
  27         u32 si, di, bp, sp;
  28 };
  29
  30  */
  31
  32 union intel_x86_pebs_dse {
  33         u64 val;
  34         struct {
  35                 unsigned int ld_dse:4;
  36                 unsigned int ld_stlb_miss:1;
  37                 unsigned int ld_locked:1;
  38                 unsigned int ld_reserved:26;
  39         };
  40         struct {
  41                 unsigned int st_l1d_hit:1;
  42                 unsigned int st_reserved1:3;
  43                 unsigned int st_stlb_miss:1;
  44                 unsigned int st_locked:1;
  45                 unsigned int st_reserved2:26;
  46         };
  47 };
  48
  49
  50 /*
  51  * Map PEBS Load Latency Data Source encodings to generic
  52  * memory data source information
  53  */
  54 #define P(a, b) PERF_MEM_S(a, b)
  55 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
  56 #define LEVEL(x) P(LVLNUM, x)
  57 #define REM P(REMOTE, REMOTE)
  58 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
  59
  60 /* Version for Sandy Bridge and later */
  61 static u64 pebs_data_source[] = {
  62         P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
  63         OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
  64         OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
  65         OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
  66         OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
  67         OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
  68         OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
  69         OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
  70         OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
  71         OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
  72         OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
  73         OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
  74         OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
  75         OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
  76         OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
  77         OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
  78 };
  79
  80 /* Patch up minor differences in the bits */
  81 void __init intel_pmu_pebs_data_source_nhm(void)
  82 {
  83         pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
  84         pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
  85         pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
  86 }
  87
  88 void __init intel_pmu_pebs_data_source_skl(bool pmem)
  89 {
  90         u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
  91
  92         pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
  93         pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
  94         pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
  95         pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
  96         pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
  97 }
  98
  99 static u64 precise_store_data(u64 status)
 100 {
 101         union intel_x86_pebs_dse dse;
 102         u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
 103
 104         dse.val = status;
 105
 106         /*
 107          * bit 4: TLB access
 108          * 1 = stored missed 2nd level TLB
 109          *
 110          * so it either hit the walker or the OS
 111          * otherwise hit 2nd level TLB
 112          */
 113         if (dse.st_stlb_miss)
 114                 val |= P(TLB, MISS);
 115         else
 116                 val |= P(TLB, HIT);
 117
 118         /*
 119          * bit 0: hit L1 data cache
 120          * if not set, then all we know is that
 121          * it missed L1D
 122          */
 123         if (dse.st_l1d_hit)
 124                 val |= P(LVL, HIT);
 125         else
 126                 val |= P(LVL, MISS);
 127
 128         /*
 129          * bit 5: Locked prefix
 130          */
 131         if (dse.st_locked)
 132                 val |= P(LOCK, LOCKED);
 133
 134         return val;
 135 }
 136
 137 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
 138 {
 139         union perf_mem_data_src dse;
 140
 141         dse.val = PERF_MEM_NA;
 142
 143         if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
 144                 dse.mem_op = PERF_MEM_OP_STORE;
 145         else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
 146                 dse.mem_op = PERF_MEM_OP_LOAD;
 147
 148         /*
 149          * L1 info only valid for following events:
 150          *
 151          * MEM_UOPS_RETIRED.STLB_MISS_STORES
 152          * MEM_UOPS_RETIRED.LOCK_STORES
 153          * MEM_UOPS_RETIRED.SPLIT_STORES
 154          * MEM_UOPS_RETIRED.ALL_STORES
 155          */
 156         if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
 157                 if (status & 1)
 158                         dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
 159                 else
 160                         dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
 161         }
 162         return dse.val;
 163 }
 164
 165 static u64 load_latency_data(u64 status)
 166 {
 167         union intel_x86_pebs_dse dse;
 168         u64 val;
 169
 170         dse.val = status;
 171
 172         /*
 173          * use the mapping table for bit 0-3
 174          */
 175         val = pebs_data_source[dse.ld_dse];
 176
 177         /*
 178          * Nehalem models do not support TLB, Lock infos
 179          */
 180         if (x86_pmu.pebs_no_tlb) {
 181                 val |= P(TLB, NA) | P(LOCK, NA);
 182                 return val;
 183         }
 184         /*
 185          * bit 4: TLB access
 186          * 0 = did not miss 2nd level TLB
 187          * 1 = missed 2nd level TLB
 188          */
 189         if (dse.ld_stlb_miss)
 190                 val |= P(TLB, MISS) | P(TLB, L2);
 191         else
 192                 val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
 193
 194         /*
 195          * bit 5: locked prefix
 196          */
 197         if (dse.ld_locked)
 198                 val |= P(LOCK, LOCKED);
 199
 200         return val;
 201 }
 202
 203 struct pebs_record_core {
 204         u64 flags, ip;
 205         u64 ax, bx, cx, dx;
 206         u64 si, di, bp, sp;
 207         u64 r8,  r9,  r10, r11;
 208         u64 r12, r13, r14, r15;
 209 };
 210
 211 struct pebs_record_nhm {
 212         u64 flags, ip;
 213         u64 ax, bx, cx, dx;
 214         u64 si, di, bp, sp;
 215         u64 r8,  r9,  r10, r11;
 216         u64 r12, r13, r14, r15;
 217         u64 status, dla, dse, lat;
 218 };
 219
 220 /*
 221  * Same as pebs_record_nhm, with two additional fields.
 222  */
 223 struct pebs_record_hsw {
 224         u64 flags, ip;
 225         u64 ax, bx, cx, dx;
 226         u64 si, di, bp, sp;
 227         u64 r8,  r9,  r10, r11;
 228         u64 r12, r13, r14, r15;
 229         u64 status, dla, dse, lat;
 230         u64 real_ip, tsx_tuning;
 231 };
 232
 233 union hsw_tsx_tuning {
 234         struct {
 235                 u32 cycles_last_block     : 32,
 236                     hle_abort             : 1,
 237                     rtm_abort             : 1,
 238                     instruction_abort     : 1,
 239                     non_instruction_abort : 1,
 240                     retry                 : 1,
 241                     data_conflict         : 1,
 242                     capacity_writes       : 1,
 243                     capacity_reads        : 1;
 244         };
 245         u64         value;
 246 };
 247
 248 #define PEBS_HSW_TSX_FLAGS      0xff00000000ULL
 249
 250 /* Same as HSW, plus TSC */
 251
 252 struct pebs_record_skl {
 253         u64 flags, ip;
 254         u64 ax, bx, cx, dx;
 255         u64 si, di, bp, sp;
 256         u64 r8,  r9,  r10, r11;
 257         u64 r12, r13, r14, r15;
 258         u64 status, dla, dse, lat;
 259         u64 real_ip, tsx_tuning;
 260         u64 tsc;
 261 };
 262
 263 void init_debug_store_on_cpu(int cpu)
 264 {
 265         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
 266
 267         if (!ds)
 268                 return;
 269
 270         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
 271                      (u32)((u64)(unsigned long)ds),
 272                      (u32)((u64)(unsigned long)ds >> 32));
 273 }
 274
 275 void fini_debug_store_on_cpu(int cpu)
 276 {
 277         if (!per_cpu(cpu_hw_events, cpu).ds)
 278                 return;
 279
 280         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
 281 }
 282
 283 static DEFINE_PER_CPU(void *, insn_buffer);
 284
 285 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
 286 {
 287         unsigned long start = (unsigned long)cea;
 288         phys_addr_t pa;
 289         size_t msz = 0;
 290
 291         pa = virt_to_phys(addr);
 292
 293         preempt_disable();
 294         for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
 295                 cea_set_pte(cea, pa, prot);
 296
 297         /*
 298          * This is a cross-CPU update of the cpu_entry_area, we must shoot down
 299          * all TLB entries for it.
 300          */
 301         flush_tlb_kernel_range(start, start + size);
 302         preempt_enable();
 303 }
 304
 305 static void ds_clear_cea(void *cea, size_t size)
 306 {
 307         unsigned long start = (unsigned long)cea;
 308         size_t msz = 0;
 309
 310         preempt_disable();
 311         for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
 312                 cea_set_pte(cea, 0, PAGE_NONE);
 313
 314         flush_tlb_kernel_range(start, start + size);
 315         preempt_enable();
 316 }
 317
 318 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
 319 {
 320         unsigned int order = get_order(size);
 321         int node = cpu_to_node(cpu);
 322         struct page *page;
 323
 324         page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
 325         return page ? page_address(page) : NULL;
 326 }
 327
 328 static void dsfree_pages(const void *buffer, size_t size)
 329 {
 330         if (buffer)
 331                 free_pages((unsigned long)buffer, get_order(size));
 332 }
 333
 334 static int alloc_pebs_buffer(int cpu)
 335 {
 336         struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
 337         struct debug_store *ds = hwev->ds;
 338         size_t bsiz = x86_pmu.pebs_buffer_size;
 339         int max, node = cpu_to_node(cpu);
 340         void *buffer, *ibuffer, *cea;
 341
 342         if (!x86_pmu.pebs)
 343                 return 0;
 344
 345         buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
 346         if (unlikely(!buffer))
 347                 return -ENOMEM;
 348
 349         /*
 350          * HSW+ already provides us the eventing ip; no need to allocate this
 351          * buffer then.
 352          */
 353         if (x86_pmu.intel_cap.pebs_format < 2) {
 354                 ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
 355                 if (!ibuffer) {
 356                         dsfree_pages(buffer, bsiz);
 357                         return -ENOMEM;
 358                 }
 359                 per_cpu(insn_buffer, cpu) = ibuffer;
 360         }
 361         hwev->ds_pebs_vaddr = buffer;
 362         /* Update the cpu entry area mapping */
 363         cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
 364         ds->pebs_buffer_base = (unsigned long) cea;
 365         ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
 366         ds->pebs_index = ds->pebs_buffer_base;
 367         max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
 368         ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
 369         return 0;
 370 }
 371
 372 static void release_pebs_buffer(int cpu)
 373 {
 374         struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
 375         void *cea;
 376
 377         if (!x86_pmu.pebs)
 378                 return;
 379
 380         kfree(per_cpu(insn_buffer, cpu));
 381         per_cpu(insn_buffer, cpu) = NULL;
 382
 383         /* Clear the fixmap */
 384         cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
 385         ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
 386         dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
 387         hwev->ds_pebs_vaddr = NULL;
 388 }
 389
 390 static int alloc_bts_buffer(int cpu)
 391 {
 392         struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
 393         struct debug_store *ds = hwev->ds;
 394         void *buffer, *cea;
 395         int max;
 396
 397         if (!x86_pmu.bts)
 398                 return 0;
 399
 400         buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
 401         if (unlikely(!buffer)) {
 402                 WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
 403                 return -ENOMEM;
 404         }
 405         hwev->ds_bts_vaddr = buffer;
 406         /* Update the fixmap */
 407         cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
 408         ds->bts_buffer_base = (unsigned long) cea;
 409         ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
 410         ds->bts_index = ds->bts_buffer_base;
 411         max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
 412         ds->bts_absolute_maximum = ds->bts_buffer_base +
 413                                         max * BTS_RECORD_SIZE;
 414         ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
 415                                         (max / 16) * BTS_RECORD_SIZE;
 416         return 0;
 417 }
 418
 419 static void release_bts_buffer(int cpu)
 420 {
 421         struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
 422         void *cea;
 423
 424         if (!x86_pmu.bts)
 425                 return;
 426
 427         /* Clear the fixmap */
 428         cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
 429         ds_clear_cea(cea, BTS_BUFFER_SIZE);
 430         dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
 431         hwev->ds_bts_vaddr = NULL;
 432 }
 433
 434 static int alloc_ds_buffer(int cpu)
 435 {
 436         struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
 437
 438         memset(ds, 0, sizeof(*ds));
 439         per_cpu(cpu_hw_events, cpu).ds = ds;
 440         return 0;
 441 }
 442
 443 static void release_ds_buffer(int cpu)
 444 {
 445         per_cpu(cpu_hw_events, cpu).ds = NULL;
 446 }
 447
 448 void release_ds_buffers(void)
 449 {
 450         int cpu;
 451
 452         if (!x86_pmu.bts && !x86_pmu.pebs)
 453                 return;
 454
 455         for_each_possible_cpu(cpu)
 456                 release_ds_buffer(cpu);
 457
 458         for_each_possible_cpu(cpu) {
 459                 /*
 460                  * Again, ignore errors from offline CPUs, they will no longer
 461                  * observe cpu_hw_events.ds and not program the DS_AREA when
 462                  * they come up.
 463                  */
 464                 fini_debug_store_on_cpu(cpu);
 465         }
 466
 467         for_each_possible_cpu(cpu) {
 468                 release_pebs_buffer(cpu);
 469                 release_bts_buffer(cpu);
 470         }
 471 }
 472
 473 void reserve_ds_buffers(void)
 474 {
 475         int bts_err = 0, pebs_err = 0;
 476         int cpu;
 477
 478         x86_pmu.bts_active = 0;
 479         x86_pmu.pebs_active = 0;
 480
 481         if (!x86_pmu.bts && !x86_pmu.pebs)
 482                 return;
 483
 484         if (!x86_pmu.bts)
 485                 bts_err = 1;
 486
 487         if (!x86_pmu.pebs)
 488                 pebs_err = 1;
 489
 490         for_each_possible_cpu(cpu) {
 491                 if (alloc_ds_buffer(cpu)) {
 492                         bts_err = 1;
 493                         pebs_err = 1;
 494                 }
 495
 496                 if (!bts_err && alloc_bts_buffer(cpu))
 497                         bts_err = 1;
 498
 499                 if (!pebs_err && alloc_pebs_buffer(cpu))
 500                         pebs_err = 1;
 501
 502                 if (bts_err && pebs_err)
 503                         break;
 504         }
 505
 506         if (bts_err) {
 507                 for_each_possible_cpu(cpu)
 508                         release_bts_buffer(cpu);
 509         }
 510
 511         if (pebs_err) {
 512                 for_each_possible_cpu(cpu)
 513                         release_pebs_buffer(cpu);
 514         }
 515
 516         if (bts_err && pebs_err) {
 517                 for_each_possible_cpu(cpu)
 518                         release_ds_buffer(cpu);
 519         } else {
 520                 if (x86_pmu.bts && !bts_err)
 521                         x86_pmu.bts_active = 1;
 522
 523                 if (x86_pmu.pebs && !pebs_err)
 524                         x86_pmu.pebs_active = 1;
 525
 526                 for_each_possible_cpu(cpu) {
 527                         /*
 528                          * Ignores wrmsr_on_cpu() errors for offline CPUs they
 529                          * will get this call through intel_pmu_cpu_starting().
 530                          */
 531                         init_debug_store_on_cpu(cpu);
 532                 }
 533         }
 534 }
 535
 536 /*
 537  * BTS
 538  */
 539
 540 struct event_constraint bts_constraint =
 541         EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
 542
 543 void intel_pmu_enable_bts(u64 config)
 544 {
 545         unsigned long debugctlmsr;
 546
 547         debugctlmsr = get_debugctlmsr();
 548
 549         debugctlmsr |= DEBUGCTLMSR_TR;
 550         debugctlmsr |= DEBUGCTLMSR_BTS;
 551         if (config & ARCH_PERFMON_EVENTSEL_INT)
 552                 debugctlmsr |= DEBUGCTLMSR_BTINT;
 553
 554         if (!(config & ARCH_PERFMON_EVENTSEL_OS))
 555                 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
 556
 557         if (!(config & ARCH_PERFMON_EVENTSEL_USR))
 558                 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
 559
 560         update_debugctlmsr(debugctlmsr);
 561 }
 562
 563 void intel_pmu_disable_bts(void)
 564 {
 565         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 566         unsigned long debugctlmsr;
 567
 568         if (!cpuc->ds)
 569                 return;
 570
 571         debugctlmsr = get_debugctlmsr();
 572
 573         debugctlmsr &=
 574                 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
 575                   DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
 576
 577         update_debugctlmsr(debugctlmsr);
 578 }
 579
 580 int intel_pmu_drain_bts_buffer(void)
 581 {
 582         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 583         struct debug_store *ds = cpuc->ds;
 584         struct bts_record {
 585                 u64     from;
 586                 u64     to;
 587                 u64     flags;
 588         };
 589         struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
 590         struct bts_record *at, *base, *top;
 591         struct perf_output_handle handle;
 592         struct perf_event_header header;
 593         struct perf_sample_data data;
 594         unsigned long skip = 0;
 595         struct pt_regs regs;
 596
 597         if (!event)
 598                 return 0;
 599
 600         if (!x86_pmu.bts_active)
 601                 return 0;
 602
 603         base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
 604         top  = (struct bts_record *)(unsigned long)ds->bts_index;
 605
 606         if (top <= base)
 607                 return 0;
 608
 609         memset(&regs, 0, sizeof(regs));
 610
 611         ds->bts_index = ds->bts_buffer_base;
 612
 613         perf_sample_data_init(&data, 0, event->hw.last_period);
 614
 615         /*
 616          * BTS leaks kernel addresses in branches across the cpl boundary,
 617          * such as traps or system calls, so unless the user is asking for
 618          * kernel tracing (and right now it's not possible), we'd need to
 619          * filter them out. But first we need to count how many of those we
 620          * have in the current batch. This is an extra O(n) pass, however,
 621          * it's much faster than the other one especially considering that
 622          * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
 623          * alloc_bts_buffer()).
 624          */
 625         for (at = base; at < top; at++) {
 626                 /*
 627                  * Note that right now *this* BTS code only works if
 628                  * attr::exclude_kernel is set, but let's keep this extra
 629                  * check here in case that changes.
 630                  */
 631                 if (event->attr.exclude_kernel &&
 632                     (kernel_ip(at->from) || kernel_ip(at->to)))
 633                         skip++;
 634         }
 635
 636         /*
 637          * Prepare a generic sample, i.e. fill in the invariant fields.
 638          * We will overwrite the from and to address before we output
 639          * the sample.
 640          */
 641         rcu_read_lock();
 642         perf_prepare_sample(&header, &data, event, &regs);
 643
 644         if (perf_output_begin(&handle, event, header.size *
 645                               (top - base - skip)))
 646                 goto unlock;
 647
 648         for (at = base; at < top; at++) {
 649                 /* Filter out any records that contain kernel addresses. */
 650                 if (event->attr.exclude_kernel &&
 651                     (kernel_ip(at->from) || kernel_ip(at->to)))
 652                         continue;
 653
 654                 data.ip         = at->from;
 655                 data.addr       = at->to;
 656
 657                 perf_output_sample(&handle, &header, &data, event);
 658         }
 659
 660         perf_output_end(&handle);
 661
 662         /* There's new data available. */
 663         event->hw.interrupts++;
 664         event->pending_kill = POLL_IN;
 665 unlock:
 666         rcu_read_unlock();
 667         return 1;
 668 }
 669
 670 static inline void intel_pmu_drain_pebs_buffer(void)
 671 {
 672         struct pt_regs regs;
 673
 674         x86_pmu.drain_pebs(&regs);
 675 }
 676
 677 /*
 678  * PEBS
 679  */
 680 struct event_constraint intel_core2_pebs_event_constraints[] = {
 681         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
 682         INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
 683         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
 684         INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
 685         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
 686         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
 687         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
 688         EVENT_CONSTRAINT_END
 689 };
 690
 691 struct event_constraint intel_atom_pebs_event_constraints[] = {
 692         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
 693         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
 694         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
 695         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
 696         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
 697         /* Allow all events as PEBS with no flags */
 698         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
 699         EVENT_CONSTRAINT_END
 700 };
 701
 702 struct event_constraint intel_slm_pebs_event_constraints[] = {
 703         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
 704         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
 705         /* Allow all events as PEBS with no flags */
 706         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
 707         EVENT_CONSTRAINT_END
 708 };
 709
 710 struct event_constraint intel_glm_pebs_event_constraints[] = {
 711         /* Allow all events as PEBS with no flags */
 712         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
 713         EVENT_CONSTRAINT_END
 714 };
 715
 716 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
 717         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
 718         INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
 719         INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
 720         INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
 721         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
 722         INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
 723         INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
 724         INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
 725         INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
 726         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
 727         INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
 728         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
 729         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
 730         EVENT_CONSTRAINT_END
 731 };
 732
 733 struct event_constraint intel_westmere_pebs_event_constraints[] = {
 734         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
 735         INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
 736         INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
 737         INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
 738         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
 739         INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
 740         INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
 741         INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
 742         INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
 743         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
 744         INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
 745         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
 746         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
 747         EVENT_CONSTRAINT_END
 748 };
 749
 750 struct event_constraint intel_snb_pebs_event_constraints[] = {
 751         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
 752         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
 753         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
 754         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
 755         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
 756         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
 757         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
 758         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
 759         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
 760         /* Allow all events as PEBS with no flags */
 761         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
 762         EVENT_CONSTRAINT_END
 763 };
 764
 765 struct event_constraint intel_ivb_pebs_event_constraints[] = {
 766         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
 767         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
 768         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
 769         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
 770         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
 771         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
 772         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
 773         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
 774         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
 775         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
 776         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
 777         /* Allow all events as PEBS with no flags */
 778         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
 779         EVENT_CONSTRAINT_END
 780 };
 781
 782 struct event_constraint intel_hsw_pebs_event_constraints[] = {
 783         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
 784         INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
 785         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
 786         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
 787         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
 788         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
 789         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
 790         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
 791         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
 792         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
 793         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
 794         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
 795         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
 796         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
 797         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
 798         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
 799         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
 800         /* Allow all events as PEBS with no flags */
 801         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
 802         EVENT_CONSTRAINT_END
 803 };
 804
 805 struct event_constraint intel_bdw_pebs_event_constraints[] = {
 806         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
 807         INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
 808         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
 809         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
 810         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
 811         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
 812         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
 813         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
 814         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
 815         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
 816         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
 817         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
 818         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
 819         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
 820         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
 821         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
 822         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
 823         /* Allow all events as PEBS with no flags */
 824         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
 825         EVENT_CONSTRAINT_END
 826 };
 827
 828
 829 struct event_constraint intel_skl_pebs_event_constraints[] = {
 830         INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),      /* INST_RETIRED.PREC_DIST */
 831         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
 832         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
 833         /* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
 834         INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
 835         INTEL_PLD_CONSTRAINT(0x1cd, 0xf),                     /* MEM_TRANS_RETIRED.* */
 836         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
 837         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
 838         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
 839         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
 840         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
 841         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
 842         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
 843         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
 844         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
 845         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
 846         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
 847         /* Allow all events as PEBS with no flags */
 848         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
 849         EVENT_CONSTRAINT_END
 850 };
 851
 852 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
 853 {
 854         struct event_constraint *c;
 855
 856         if (!event->attr.precise_ip)
 857                 return NULL;
 858
 859         if (x86_pmu.pebs_constraints) {
 860                 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
 861                         if ((event->hw.config & c->cmask) == c->code) {
 862                                 event->hw.flags |= c->flags;
 863                                 return c;
 864                         }
 865                 }
 866         }
 867
 868         /*
 869          * Extended PEBS support
 870          * Makes the PEBS code search the normal constraints.
 871          */
 872         if (x86_pmu.flags & PMU_FL_PEBS_ALL)
 873                 return NULL;
 874
 875         return &emptyconstraint;
 876 }
 877
 878 /*
 879  * We need the sched_task callback even for per-cpu events when we use
 880  * the large interrupt threshold, such that we can provide PID and TID
 881  * to PEBS samples.
 882  */
 883 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
 884 {
 885         return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
 886 }
 887
 888 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
 889 {
 890         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 891
 892         if (!sched_in && pebs_needs_sched_cb(cpuc))
 893                 intel_pmu_drain_pebs_buffer();
 894 }
 895
 896 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
 897 {
 898         struct debug_store *ds = cpuc->ds;
 899         u64 threshold;
 900         int reserved;
 901
 902         if (x86_pmu.flags & PMU_FL_PEBS_ALL)
 903                 reserved = x86_pmu.max_pebs_events + x86_pmu.num_counters_fixed;
 904         else
 905                 reserved = x86_pmu.max_pebs_events;
 906
 907         if (cpuc->n_pebs == cpuc->n_large_pebs) {
 908                 threshold = ds->pebs_absolute_maximum -
 909                         reserved * x86_pmu.pebs_record_size;
 910         } else {
 911                 threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
 912         }
 913
 914         ds->pebs_interrupt_threshold = threshold;
 915 }
 916
 917 static void
 918 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
 919 {
 920         /*
 921          * Make sure we get updated with the first PEBS
 922          * event. It will trigger also during removal, but
 923          * that does not hurt:
 924          */
 925         bool update = cpuc->n_pebs == 1;
 926
 927         if (needed_cb != pebs_needs_sched_cb(cpuc)) {
 928                 if (!needed_cb)
 929                         perf_sched_cb_inc(pmu);
 930                 else
 931                         perf_sched_cb_dec(pmu);
 932
 933                 update = true;
 934         }
 935
 936         if (update)
 937                 pebs_update_threshold(cpuc);
 938 }
 939
 940 void intel_pmu_pebs_add(struct perf_event *event)
 941 {
 942         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 943         struct hw_perf_event *hwc = &event->hw;
 944         bool needed_cb = pebs_needs_sched_cb(cpuc);
 945
 946         cpuc->n_pebs++;
 947         if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
 948                 cpuc->n_large_pebs++;
 949
 950         pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
 951 }
 952
 953 void intel_pmu_pebs_enable(struct perf_event *event)
 954 {
 955         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 956         struct hw_perf_event *hwc = &event->hw;
 957         struct debug_store *ds = cpuc->ds;
 958
 959         hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
 960
 961         cpuc->pebs_enabled |= 1ULL << hwc->idx;
 962
 963         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
 964                 cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
 965         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
 966                 cpuc->pebs_enabled |= 1ULL << 63;
 967
 968         /*
 969          * Use auto-reload if possible to save a MSR write in the PMI.
 970          * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
 971          */
 972         if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
 973                 unsigned int idx = hwc->idx;
 974
 975                 if (idx >= INTEL_PMC_IDX_FIXED)
 976                         idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
 977                 ds->pebs_event_reset[idx] =
 978                         (u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
 979         } else {
 980                 ds->pebs_event_reset[hwc->idx] = 0;
 981         }
 982 }
 983
 984 void intel_pmu_pebs_del(struct perf_event *event)
 985 {
 986         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 987         struct hw_perf_event *hwc = &event->hw;
 988         bool needed_cb = pebs_needs_sched_cb(cpuc);
 989
 990         cpuc->n_pebs--;
 991         if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
 992                 cpuc->n_large_pebs--;
 993
 994         pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
 995 }
 996
 997 void intel_pmu_pebs_disable(struct perf_event *event)
 998 {
 999         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1000         struct hw_perf_event *hwc = &event->hw;
1001
1002         if (cpuc->n_pebs == cpuc->n_large_pebs)
1003                 intel_pmu_drain_pebs_buffer();
1004
1005         cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1006
1007         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
1008                 cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1009         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1010                 cpuc->pebs_enabled &= ~(1ULL << 63);
1011
1012         if (cpuc->enabled)
1013                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1014
1015         hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1016 }
1017
1018 void intel_pmu_pebs_enable_all(void)
1019 {
1020         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1021
1022         if (cpuc->pebs_enabled)
1023                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1024 }
1025
1026 void intel_pmu_pebs_disable_all(void)
1027 {
1028         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1029
1030         if (cpuc->pebs_enabled)
1031                 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1032 }
1033
1034 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1035 {
1036         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1037         unsigned long from = cpuc->lbr_entries[0].from;
1038         unsigned long old_to, to = cpuc->lbr_entries[0].to;
1039         unsigned long ip = regs->ip;
1040         int is_64bit = 0;
1041         void *kaddr;
1042         int size;
1043
1044         /*
1045          * We don't need to fixup if the PEBS assist is fault like
1046          */
1047         if (!x86_pmu.intel_cap.pebs_trap)
1048                 return 1;
1049
1050         /*
1051          * No LBR entry, no basic block, no rewinding
1052          */
1053         if (!cpuc->lbr_stack.nr || !from || !to)
1054                 return 0;
1055
1056         /*
1057          * Basic blocks should never cross user/kernel boundaries
1058          */
1059         if (kernel_ip(ip) != kernel_ip(to))
1060                 return 0;
1061
1062         /*
1063          * unsigned math, either ip is before the start (impossible) or
1064          * the basic block is larger than 1 page (sanity)
1065          */
1066         if ((ip - to) > PEBS_FIXUP_SIZE)
1067                 return 0;
1068
1069         /*
1070          * We sampled a branch insn, rewind using the LBR stack
1071          */
1072         if (ip == to) {
1073                 set_linear_ip(regs, from);
1074                 return 1;
1075         }
1076
1077         size = ip - to;
1078         if (!kernel_ip(ip)) {
1079                 int bytes;
1080                 u8 *buf = this_cpu_read(insn_buffer);
1081
1082                 /* 'size' must fit our buffer, see above */
1083                 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1084                 if (bytes != 0)
1085                         return 0;
1086
1087                 kaddr = buf;
1088         } else {
1089                 kaddr = (void *)to;
1090         }
1091
1092         do {
1093                 struct insn insn;
1094
1095                 old_to = to;
1096
1097 #ifdef CONFIG_X86_64
1098                 is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1099 #endif
1100                 insn_init(&insn, kaddr, size, is_64bit);
1101                 insn_get_length(&insn);
1102                 /*
1103                  * Make sure there was not a problem decoding the
1104                  * instruction and getting the length.  This is
1105                  * doubly important because we have an infinite
1106                  * loop if insn.length=0.
1107                  */
1108                 if (!insn.length)
1109                         break;
1110
1111                 to += insn.length;
1112                 kaddr += insn.length;
1113                 size -= insn.length;
1114         } while (to < ip);
1115
1116         if (to == ip) {
1117                 set_linear_ip(regs, old_to);
1118                 return 1;
1119         }
1120
1121         /*
1122          * Even though we decoded the basic block, the instruction stream
1123          * never matched the given IP, either the TO or the IP got corrupted.
1124          */
1125         return 0;
1126 }
1127
1128 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1129 {
1130         if (pebs->tsx_tuning) {
1131                 union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1132                 return tsx.cycles_last_block;
1133         }
1134         return 0;
1135 }
1136
1137 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1138 {
1139         u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1140
1141         /* For RTM XABORTs also log the abort code from AX */
1142         if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1143                 txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1144         return txn;
1145 }
1146
1147 static void setup_pebs_sample_data(struct perf_event *event,
1148                                    struct pt_regs *iregs, void *__pebs,
1149                                    struct perf_sample_data *data,
1150                                    struct pt_regs *regs)
1151 {
1152 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1153                 (PERF_X86_EVENT_PEBS_ST_HSW | \
1154                  PERF_X86_EVENT_PEBS_LD_HSW | \
1155                  PERF_X86_EVENT_PEBS_NA_HSW)
1156         /*
1157          * We cast to the biggest pebs_record but are careful not to
1158          * unconditionally access the 'extra' entries.
1159          */
1160         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1161         struct pebs_record_skl *pebs = __pebs;
1162         u64 sample_type;
1163         int fll, fst, dsrc;
1164         int fl = event->hw.flags;
1165
1166         if (pebs == NULL)
1167                 return;
1168
1169         sample_type = event->attr.sample_type;
1170         dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1171
1172         fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1173         fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1174
1175         perf_sample_data_init(data, 0, event->hw.last_period);
1176
1177         data->period = event->hw.last_period;
1178
1179         /*
1180          * Use latency for weight (only avail with PEBS-LL)
1181          */
1182         if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1183                 data->weight = pebs->lat;
1184
1185         /*
1186          * data.data_src encodes the data source
1187          */
1188         if (dsrc) {
1189                 u64 val = PERF_MEM_NA;
1190                 if (fll)
1191                         val = load_latency_data(pebs->dse);
1192                 else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1193                         val = precise_datala_hsw(event, pebs->dse);
1194                 else if (fst)
1195                         val = precise_store_data(pebs->dse);
1196                 data->data_src.val = val;
1197         }
1198
1199         /*
1200          * We must however always use iregs for the unwinder to stay sane; the
1201          * record BP,SP,IP can point into thin air when the record is from a
1202          * previous PMI context or an (I)RET happend between the record and
1203          * PMI.
1204          */
1205         if (sample_type & PERF_SAMPLE_CALLCHAIN)
1206                 data->callchain = perf_callchain(event, iregs);
1207
1208         /*
1209          * We use the interrupt regs as a base because the PEBS record does not
1210          * contain a full regs set, specifically it seems to lack segment
1211          * descriptors, which get used by things like user_mode().
1212          *
1213          * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1214          */
1215         *regs = *iregs;
1216
1217         /*
1218          * Initialize regs_>flags from PEBS,
1219          * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1220          * i.e., do not rely on it being zero:
1221          */
1222         regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1223
1224         if (sample_type & PERF_SAMPLE_REGS_INTR) {
1225                 regs->ax = pebs->ax;
1226                 regs->bx = pebs->bx;
1227                 regs->cx = pebs->cx;
1228                 regs->dx = pebs->dx;
1229                 regs->si = pebs->si;
1230                 regs->di = pebs->di;
1231
1232                 regs->bp = pebs->bp;
1233                 regs->sp = pebs->sp;
1234
1235 #ifndef CONFIG_X86_32
1236                 regs->r8 = pebs->r8;
1237                 regs->r9 = pebs->r9;
1238                 regs->r10 = pebs->r10;
1239                 regs->r11 = pebs->r11;
1240                 regs->r12 = pebs->r12;
1241                 regs->r13 = pebs->r13;
1242                 regs->r14 = pebs->r14;
1243                 regs->r15 = pebs->r15;
1244 #endif
1245         }
1246
1247         if (event->attr.precise_ip > 1) {
1248                 /*
1249                  * Haswell and later processors have an 'eventing IP'
1250                  * (real IP) which fixes the off-by-1 skid in hardware.
1251                  * Use it when precise_ip >= 2 :
1252                  */
1253                 if (x86_pmu.intel_cap.pebs_format >= 2) {
1254                         set_linear_ip(regs, pebs->real_ip);
1255                         regs->flags |= PERF_EFLAGS_EXACT;
1256                 } else {
1257                         /* Otherwise, use PEBS off-by-1 IP: */
1258                         set_linear_ip(regs, pebs->ip);
1259
1260                         /*
1261                          * With precise_ip >= 2, try to fix up the off-by-1 IP
1262                          * using the LBR. If successful, the fixup function
1263                          * corrects regs->ip and calls set_linear_ip() on regs:
1264                          */
1265                         if (intel_pmu_pebs_fixup_ip(regs))
1266                                 regs->flags |= PERF_EFLAGS_EXACT;
1267                 }
1268         } else {
1269                 /*
1270                  * When precise_ip == 1, return the PEBS off-by-1 IP,
1271                  * no fixup attempted:
1272                  */
1273                 set_linear_ip(regs, pebs->ip);
1274         }
1275
1276
1277         if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1278             x86_pmu.intel_cap.pebs_format >= 1)
1279                 data->addr = pebs->dla;
1280
1281         if (x86_pmu.intel_cap.pebs_format >= 2) {
1282                 /* Only set the TSX weight when no memory weight. */
1283                 if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1284                         data->weight = intel_hsw_weight(pebs);
1285
1286                 if (sample_type & PERF_SAMPLE_TRANSACTION)
1287                         data->txn = intel_hsw_transaction(pebs);
1288         }
1289
1290         /*
1291          * v3 supplies an accurate time stamp, so we use that
1292          * for the time stamp.
1293          *
1294          * We can only do this for the default trace clock.
1295          */
1296         if (x86_pmu.intel_cap.pebs_format >= 3 &&
1297                 event->attr.use_clockid == 0)
1298                 data->time = native_sched_clock_from_tsc(pebs->tsc);
1299
1300         if (has_branch_stack(event))
1301                 data->br_stack = &cpuc->lbr_stack;
1302 }
1303
1304 static inline void *
1305 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1306 {
1307         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1308         void *at;
1309         u64 pebs_status;
1310
1311         /*
1312          * fmt0 does not have a status bitfield (does not use
1313          * perf_record_nhm format)
1314          */
1315         if (x86_pmu.intel_cap.pebs_format < 1)
1316                 return base;
1317
1318         if (base == NULL)
1319                 return NULL;
1320
1321         for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1322                 struct pebs_record_nhm *p = at;
1323
1324                 if (test_bit(bit, (unsigned long *)&p->status)) {
1325                         /* PEBS v3 has accurate status bits */
1326                         if (x86_pmu.intel_cap.pebs_format >= 3)
1327                                 return at;
1328
1329                         if (p->status == (1 << bit))
1330                                 return at;
1331
1332                         /* clear non-PEBS bit and re-check */
1333                         pebs_status = p->status & cpuc->pebs_enabled;
1334                         pebs_status &= PEBS_COUNTER_MASK;
1335                         if (pebs_status == (1 << bit))
1336                                 return at;
1337                 }
1338         }
1339         return NULL;
1340 }
1341
1342 void intel_pmu_auto_reload_read(struct perf_event *event)
1343 {
1344         WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1345
1346         perf_pmu_disable(event->pmu);
1347         intel_pmu_drain_pebs_buffer();
1348         perf_pmu_enable(event->pmu);
1349 }
1350
1351 /*
1352  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1353  */
1354 static int
1355 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1356 {
1357         struct hw_perf_event *hwc = &event->hw;
1358         int shift = 64 - x86_pmu.cntval_bits;
1359         u64 period = hwc->sample_period;
1360         u64 prev_raw_count, new_raw_count;
1361         s64 new, old;
1362
1363         WARN_ON(!period);
1364
1365         /*
1366          * drain_pebs() only happens when the PMU is disabled.
1367          */
1368         WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1369
1370         prev_raw_count = local64_read(&hwc->prev_count);
1371         rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1372         local64_set(&hwc->prev_count, new_raw_count);
1373
1374         /*
1375          * Since the counter increments a negative counter value and
1376          * overflows on the sign switch, giving the interval:
1377          *
1378          *   [-period, 0]
1379          *
1380          * the difference between two consequtive reads is:
1381          *
1382          *   A) value2 - value1;
1383          *      when no overflows have happened in between,
1384          *
1385          *   B) (0 - value1) + (value2 - (-period));
1386          *      when one overflow happened in between,
1387          *
1388          *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1389          *      when @n overflows happened in between.
1390          *
1391          * Here A) is the obvious difference, B) is the extension to the
1392          * discrete interval, where the first term is to the top of the
1393          * interval and the second term is from the bottom of the next
1394          * interval and C) the extension to multiple intervals, where the
1395          * middle term is the whole intervals covered.
1396          *
1397          * An equivalent of C, by reduction, is:
1398          *
1399          *   value2 - value1 + n * period
1400          */
1401         new = ((s64)(new_raw_count << shift) >> shift);
1402         old = ((s64)(prev_raw_count << shift) >> shift);
1403         local64_add(new - old + count * period, &event->count);
1404
1405         local64_set(&hwc->period_left, -new);
1406
1407         perf_event_update_userpage(event);
1408
1409         return 0;
1410 }
1411
1412 static void __intel_pmu_pebs_event(struct perf_event *event,
1413                                    struct pt_regs *iregs,
1414                                    void *base, void *top,
1415                                    int bit, int count)
1416 {
1417         struct hw_perf_event *hwc = &event->hw;
1418         struct perf_sample_data data;
1419         struct pt_regs regs;
1420         void *at = get_next_pebs_record_by_bit(base, top, bit);
1421
1422         if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1423                 /*
1424                  * Now, auto-reload is only enabled in fixed period mode.
1425                  * The reload value is always hwc->sample_period.
1426                  * May need to change it, if auto-reload is enabled in
1427                  * freq mode later.
1428                  */
1429                 intel_pmu_save_and_restart_reload(event, count);
1430         } else if (!intel_pmu_save_and_restart(event))
1431                 return;
1432
1433         while (count > 1) {
1434                 setup_pebs_sample_data(event, iregs, at, &data, &regs);
1435                 perf_event_output(event, &data, &regs);
1436                 at += x86_pmu.pebs_record_size;
1437                 at = get_next_pebs_record_by_bit(at, top, bit);
1438                 count--;
1439         }
1440
1441         setup_pebs_sample_data(event, iregs, at, &data, &regs);
1442
1443         /*
1444          * All but the last records are processed.
1445          * The last one is left to be able to call the overflow handler.
1446          */
1447         if (perf_event_overflow(event, &data, &regs)) {
1448                 x86_pmu_stop(event, 0);
1449                 return;
1450         }
1451
1452 }
1453
1454 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1455 {
1456         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1457         struct debug_store *ds = cpuc->ds;
1458         struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1459         struct pebs_record_core *at, *top;
1460         int n;
1461
1462         if (!x86_pmu.pebs_active)
1463                 return;
1464
1465         at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1466         top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1467
1468         /*
1469          * Whatever else happens, drain the thing
1470          */
1471         ds->pebs_index = ds->pebs_buffer_base;
1472
1473         if (!test_bit(0, cpuc->active_mask))
1474                 return;
1475
1476         WARN_ON_ONCE(!event);
1477
1478         if (!event->attr.precise_ip)
1479                 return;
1480
1481         n = top - at;
1482         if (n <= 0) {
1483                 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1484                         intel_pmu_save_and_restart_reload(event, 0);
1485                 return;
1486         }
1487
1488         __intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1489 }
1490
1491 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1492 {
1493         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1494         struct debug_store *ds = cpuc->ds;
1495         struct perf_event *event;
1496         void *base, *at, *top;
1497         short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1498         short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1499         int bit, i, size;
1500         u64 mask;
1501
1502         if (!x86_pmu.pebs_active)
1503                 return;
1504
1505         base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1506         top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1507
1508         ds->pebs_index = ds->pebs_buffer_base;
1509
1510         mask = (1ULL << x86_pmu.max_pebs_events) - 1;
1511         size = x86_pmu.max_pebs_events;
1512         if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
1513                 mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
1514                 size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
1515         }
1516
1517         if (unlikely(base >= top)) {
1518                 /*
1519                  * The drain_pebs() could be called twice in a short period
1520                  * for auto-reload event in pmu::read(). There are no
1521                  * overflows have happened in between.
1522                  * It needs to call intel_pmu_save_and_restart_reload() to
1523                  * update the event->count for this case.
1524                  */
1525                 for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled,
1526                                  size) {
1527                         event = cpuc->events[bit];
1528                         if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1529                                 intel_pmu_save_and_restart_reload(event, 0);
1530                 }
1531                 return;
1532         }
1533
1534         for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1535                 struct pebs_record_nhm *p = at;
1536                 u64 pebs_status;
1537
1538                 pebs_status = p->status & cpuc->pebs_enabled;
1539                 pebs_status &= mask;
1540
1541                 /* PEBS v3 has more accurate status bits */
1542                 if (x86_pmu.intel_cap.pebs_format >= 3) {
1543                         for_each_set_bit(bit, (unsigned long *)&pebs_status,
1544                                          size)
1545                                 counts[bit]++;
1546
1547                         continue;
1548                 }
1549
1550                 /*
1551                  * On some CPUs the PEBS status can be zero when PEBS is
1552                  * racing with clearing of GLOBAL_STATUS.
1553                  *
1554                  * Normally we would drop that record, but in the
1555                  * case when there is only a single active PEBS event
1556                  * we can assume it's for that event.
1557                  */
1558                 if (!pebs_status && cpuc->pebs_enabled &&
1559                         !(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1560                         pebs_status = p->status = cpuc->pebs_enabled;
1561
1562                 bit = find_first_bit((unsigned long *)&pebs_status,
1563                                         x86_pmu.max_pebs_events);
1564                 if (bit >= x86_pmu.max_pebs_events)
1565                         continue;
1566
1567                 /*
1568                  * The PEBS hardware does not deal well with the situation
1569                  * when events happen near to each other and multiple bits
1570                  * are set. But it should happen rarely.
1571                  *
1572                  * If these events include one PEBS and multiple non-PEBS
1573                  * events, it doesn't impact PEBS record. The record will
1574                  * be handled normally. (slow path)
1575                  *
1576                  * If these events include two or more PEBS events, the
1577                  * records for the events can be collapsed into a single
1578                  * one, and it's not possible to reconstruct all events
1579                  * that caused the PEBS record. It's called collision.
1580                  * If collision happened, the record will be dropped.
1581                  */
1582                 if (p->status != (1ULL << bit)) {
1583                         for_each_set_bit(i, (unsigned long *)&pebs_status,
1584                                          x86_pmu.max_pebs_events)
1585                                 error[i]++;
1586                         continue;
1587                 }
1588
1589                 counts[bit]++;
1590         }
1591
1592         for (bit = 0; bit < size; bit++) {
1593                 if ((counts[bit] == 0) && (error[bit] == 0))
1594                         continue;
1595
1596                 event = cpuc->events[bit];
1597                 if (WARN_ON_ONCE(!event))
1598                         continue;
1599
1600                 if (WARN_ON_ONCE(!event->attr.precise_ip))
1601                         continue;
1602
1603                 /* log dropped samples number */
1604                 if (error[bit]) {
1605                         perf_log_lost_samples(event, error[bit]);
1606
1607                         if (perf_event_account_interrupt(event))
1608                                 x86_pmu_stop(event, 0);
1609                 }
1610
1611                 if (counts[bit]) {
1612                         __intel_pmu_pebs_event(event, iregs, base,
1613                                                top, bit, counts[bit]);
1614                 }
1615         }
1616 }
1617
1618 /*
1619  * BTS, PEBS probe and setup
1620  */
1621
1622 void __init intel_ds_init(void)
1623 {
1624         /*
1625          * No support for 32bit formats
1626          */
1627         if (!boot_cpu_has(X86_FEATURE_DTES64))
1628                 return;
1629
1630         x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1631         x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1632         x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1633         if (x86_pmu.pebs) {
1634                 char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1635                 int format = x86_pmu.intel_cap.pebs_format;
1636
1637                 switch (format) {
1638                 case 0:
1639                         pr_cont("PEBS fmt0%c, ", pebs_type);
1640                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1641                         /*
1642                          * Using >PAGE_SIZE buffers makes the WRMSR to
1643                          * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1644                          * mysteriously hang on Core2.
1645                          *
1646                          * As a workaround, we don't do this.
1647                          */
1648                         x86_pmu.pebs_buffer_size = PAGE_SIZE;
1649                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1650                         break;
1651
1652                 case 1:
1653                         pr_cont("PEBS fmt1%c, ", pebs_type);
1654                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1655                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1656                         break;
1657
1658                 case 2:
1659                         pr_cont("PEBS fmt2%c, ", pebs_type);
1660                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1661                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1662                         break;
1663
1664                 case 3:
1665                         pr_cont("PEBS fmt3%c, ", pebs_type);
1666                         x86_pmu.pebs_record_size =
1667                                                 sizeof(struct pebs_record_skl);
1668                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1669                         x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
1670                         break;
1671
1672                 default:
1673                         pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1674                         x86_pmu.pebs = 0;
1675                 }
1676         }
1677 }
1678
1679 void perf_restore_debug_store(void)
1680 {
1681         struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1682
1683         if (!x86_pmu.bts && !x86_pmu.pebs)
1684                 return;
1685
1686         wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1687 }