root/arch/x86/events/intel/lbr.c

/* [<][>][^][v][top][bottom][index][help] */

DEFINITIONS

This source file includes following definitions.
  1. __intel_pmu_lbr_enable
  2. __intel_pmu_lbr_disable
  3. intel_pmu_lbr_reset_32
  4. intel_pmu_lbr_reset_64
  5. intel_pmu_lbr_reset
  6. intel_pmu_lbr_tos
  7. lbr_from_signext_quirk_needed
  8. lbr_from_signext_quirk_wr
  9. lbr_from_signext_quirk_rd
  10. wrlbr_from
  11. wrlbr_to
  12. rdlbr_from
  13. rdlbr_to
  14. __intel_pmu_lbr_restore
  15. __intel_pmu_lbr_save
  16. intel_pmu_lbr_sched_task
  17. branch_user_callstack
  18. intel_pmu_lbr_add
  19. intel_pmu_lbr_del
  20. intel_pmu_lbr_enable_all
  21. intel_pmu_lbr_disable_all
  22. intel_pmu_lbr_read_32
  23. intel_pmu_lbr_read_64
  24. intel_pmu_lbr_read
  25. intel_pmu_setup_sw_lbr_filter
  26. intel_pmu_setup_hw_lbr_filter
  27. intel_pmu_setup_lbr_filter
  28. branch_type
  29. common_branch_type
  30. intel_pmu_lbr_filter
  31. intel_pmu_store_pebs_lbrs
  32. intel_pmu_lbr_init_core
  33. intel_pmu_lbr_init_nhm
  34. intel_pmu_lbr_init_snb
  35. intel_pmu_lbr_init_hsw
  36. intel_pmu_lbr_init_skl
  37. intel_pmu_lbr_init_atom
  38. intel_pmu_lbr_init_slm
  39. intel_pmu_lbr_init_knl

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/perf_event.h>
   3 #include <linux/types.h>
   4 
   5 #include <asm/perf_event.h>
   6 #include <asm/msr.h>
   7 #include <asm/insn.h>
   8 
   9 #include "../perf_event.h"
  10 
  11 enum {
  12         LBR_FORMAT_32           = 0x00,
  13         LBR_FORMAT_LIP          = 0x01,
  14         LBR_FORMAT_EIP          = 0x02,
  15         LBR_FORMAT_EIP_FLAGS    = 0x03,
  16         LBR_FORMAT_EIP_FLAGS2   = 0x04,
  17         LBR_FORMAT_INFO         = 0x05,
  18         LBR_FORMAT_TIME         = 0x06,
  19         LBR_FORMAT_MAX_KNOWN    = LBR_FORMAT_TIME,
  20 };
  21 
  22 static const enum {
  23         LBR_EIP_FLAGS           = 1,
  24         LBR_TSX                 = 2,
  25 } lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
  26         [LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
  27         [LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
  28 };
  29 
  30 /*
  31  * Intel LBR_SELECT bits
  32  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
  33  *
  34  * Hardware branch filter (not available on all CPUs)
  35  */
  36 #define LBR_KERNEL_BIT          0 /* do not capture at ring0 */
  37 #define LBR_USER_BIT            1 /* do not capture at ring > 0 */
  38 #define LBR_JCC_BIT             2 /* do not capture conditional branches */
  39 #define LBR_REL_CALL_BIT        3 /* do not capture relative calls */
  40 #define LBR_IND_CALL_BIT        4 /* do not capture indirect calls */
  41 #define LBR_RETURN_BIT          5 /* do not capture near returns */
  42 #define LBR_IND_JMP_BIT         6 /* do not capture indirect jumps */
  43 #define LBR_REL_JMP_BIT         7 /* do not capture relative jumps */
  44 #define LBR_FAR_BIT             8 /* do not capture far branches */
  45 #define LBR_CALL_STACK_BIT      9 /* enable call stack */
  46 
  47 /*
  48  * Following bit only exists in Linux; we mask it out before writing it to
  49  * the actual MSR. But it helps the constraint perf code to understand
  50  * that this is a separate configuration.
  51  */
  52 #define LBR_NO_INFO_BIT        63 /* don't read LBR_INFO. */
  53 
  54 #define LBR_KERNEL      (1 << LBR_KERNEL_BIT)
  55 #define LBR_USER        (1 << LBR_USER_BIT)
  56 #define LBR_JCC         (1 << LBR_JCC_BIT)
  57 #define LBR_REL_CALL    (1 << LBR_REL_CALL_BIT)
  58 #define LBR_IND_CALL    (1 << LBR_IND_CALL_BIT)
  59 #define LBR_RETURN      (1 << LBR_RETURN_BIT)
  60 #define LBR_REL_JMP     (1 << LBR_REL_JMP_BIT)
  61 #define LBR_IND_JMP     (1 << LBR_IND_JMP_BIT)
  62 #define LBR_FAR         (1 << LBR_FAR_BIT)
  63 #define LBR_CALL_STACK  (1 << LBR_CALL_STACK_BIT)
  64 #define LBR_NO_INFO     (1ULL << LBR_NO_INFO_BIT)
  65 
  66 #define LBR_PLM (LBR_KERNEL | LBR_USER)
  67 
  68 #define LBR_SEL_MASK    0x3ff   /* valid bits in LBR_SELECT */
  69 #define LBR_NOT_SUPP    -1      /* LBR filter not supported */
  70 #define LBR_IGN         0       /* ignored */
  71 
  72 #define LBR_ANY          \
  73         (LBR_JCC        |\
  74          LBR_REL_CALL   |\
  75          LBR_IND_CALL   |\
  76          LBR_RETURN     |\
  77          LBR_REL_JMP    |\
  78          LBR_IND_JMP    |\
  79          LBR_FAR)
  80 
  81 #define LBR_FROM_FLAG_MISPRED   BIT_ULL(63)
  82 #define LBR_FROM_FLAG_IN_TX     BIT_ULL(62)
  83 #define LBR_FROM_FLAG_ABORT     BIT_ULL(61)
  84 
  85 #define LBR_FROM_SIGNEXT_2MSB   (BIT_ULL(60) | BIT_ULL(59))
  86 
  87 /*
  88  * x86control flow change classification
  89  * x86control flow changes include branches, interrupts, traps, faults
  90  */
  91 enum {
  92         X86_BR_NONE             = 0,      /* unknown */
  93 
  94         X86_BR_USER             = 1 << 0, /* branch target is user */
  95         X86_BR_KERNEL           = 1 << 1, /* branch target is kernel */
  96 
  97         X86_BR_CALL             = 1 << 2, /* call */
  98         X86_BR_RET              = 1 << 3, /* return */
  99         X86_BR_SYSCALL          = 1 << 4, /* syscall */
 100         X86_BR_SYSRET           = 1 << 5, /* syscall return */
 101         X86_BR_INT              = 1 << 6, /* sw interrupt */
 102         X86_BR_IRET             = 1 << 7, /* return from interrupt */
 103         X86_BR_JCC              = 1 << 8, /* conditional */
 104         X86_BR_JMP              = 1 << 9, /* jump */
 105         X86_BR_IRQ              = 1 << 10,/* hw interrupt or trap or fault */
 106         X86_BR_IND_CALL         = 1 << 11,/* indirect calls */
 107         X86_BR_ABORT            = 1 << 12,/* transaction abort */
 108         X86_BR_IN_TX            = 1 << 13,/* in transaction */
 109         X86_BR_NO_TX            = 1 << 14,/* not in transaction */
 110         X86_BR_ZERO_CALL        = 1 << 15,/* zero length call */
 111         X86_BR_CALL_STACK       = 1 << 16,/* call stack */
 112         X86_BR_IND_JMP          = 1 << 17,/* indirect jump */
 113 
 114         X86_BR_TYPE_SAVE        = 1 << 18,/* indicate to save branch type */
 115 
 116 };
 117 
 118 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
 119 #define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
 120 
 121 #define X86_BR_ANY       \
 122         (X86_BR_CALL    |\
 123          X86_BR_RET     |\
 124          X86_BR_SYSCALL |\
 125          X86_BR_SYSRET  |\
 126          X86_BR_INT     |\
 127          X86_BR_IRET    |\
 128          X86_BR_JCC     |\
 129          X86_BR_JMP      |\
 130          X86_BR_IRQ      |\
 131          X86_BR_ABORT    |\
 132          X86_BR_IND_CALL |\
 133          X86_BR_IND_JMP  |\
 134          X86_BR_ZERO_CALL)
 135 
 136 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
 137 
 138 #define X86_BR_ANY_CALL          \
 139         (X86_BR_CALL            |\
 140          X86_BR_IND_CALL        |\
 141          X86_BR_ZERO_CALL       |\
 142          X86_BR_SYSCALL         |\
 143          X86_BR_IRQ             |\
 144          X86_BR_INT)
 145 
 146 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
 147 
 148 /*
 149  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
 150  * otherwise it becomes near impossible to get a reliable stack.
 151  */
 152 
 153 static void __intel_pmu_lbr_enable(bool pmi)
 154 {
 155         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 156         u64 debugctl, lbr_select = 0, orig_debugctl;
 157 
 158         /*
 159          * No need to unfreeze manually, as v4 can do that as part
 160          * of the GLOBAL_STATUS ack.
 161          */
 162         if (pmi && x86_pmu.version >= 4)
 163                 return;
 164 
 165         /*
 166          * No need to reprogram LBR_SELECT in a PMI, as it
 167          * did not change.
 168          */
 169         if (cpuc->lbr_sel)
 170                 lbr_select = cpuc->lbr_sel->config & x86_pmu.lbr_sel_mask;
 171         if (!pmi && cpuc->lbr_sel)
 172                 wrmsrl(MSR_LBR_SELECT, lbr_select);
 173 
 174         rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
 175         orig_debugctl = debugctl;
 176         debugctl |= DEBUGCTLMSR_LBR;
 177         /*
 178          * LBR callstack does not work well with FREEZE_LBRS_ON_PMI.
 179          * If FREEZE_LBRS_ON_PMI is set, PMI near call/return instructions
 180          * may cause superfluous increase/decrease of LBR_TOS.
 181          */
 182         if (!(lbr_select & LBR_CALL_STACK))
 183                 debugctl |= DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
 184         if (orig_debugctl != debugctl)
 185                 wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
 186 }
 187 
 188 static void __intel_pmu_lbr_disable(void)
 189 {
 190         u64 debugctl;
 191 
 192         rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
 193         debugctl &= ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
 194         wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
 195 }
 196 
 197 static void intel_pmu_lbr_reset_32(void)
 198 {
 199         int i;
 200 
 201         for (i = 0; i < x86_pmu.lbr_nr; i++)
 202                 wrmsrl(x86_pmu.lbr_from + i, 0);
 203 }
 204 
 205 static void intel_pmu_lbr_reset_64(void)
 206 {
 207         int i;
 208 
 209         for (i = 0; i < x86_pmu.lbr_nr; i++) {
 210                 wrmsrl(x86_pmu.lbr_from + i, 0);
 211                 wrmsrl(x86_pmu.lbr_to   + i, 0);
 212                 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
 213                         wrmsrl(MSR_LBR_INFO_0 + i, 0);
 214         }
 215 }
 216 
 217 void intel_pmu_lbr_reset(void)
 218 {
 219         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 220 
 221         if (!x86_pmu.lbr_nr)
 222                 return;
 223 
 224         if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
 225                 intel_pmu_lbr_reset_32();
 226         else
 227                 intel_pmu_lbr_reset_64();
 228 
 229         cpuc->last_task_ctx = NULL;
 230         cpuc->last_log_id = 0;
 231 }
 232 
 233 /*
 234  * TOS = most recently recorded branch
 235  */
 236 static inline u64 intel_pmu_lbr_tos(void)
 237 {
 238         u64 tos;
 239 
 240         rdmsrl(x86_pmu.lbr_tos, tos);
 241         return tos;
 242 }
 243 
 244 enum {
 245         LBR_NONE,
 246         LBR_VALID,
 247 };
 248 
 249 /*
 250  * For formats with LBR_TSX flags (e.g. LBR_FORMAT_EIP_FLAGS2), bits 61:62 in
 251  * MSR_LAST_BRANCH_FROM_x are the TSX flags when TSX is supported, but when
 252  * TSX is not supported they have no consistent behavior:
 253  *
 254  *   - For wrmsr(), bits 61:62 are considered part of the sign extension.
 255  *   - For HW updates (branch captures) bits 61:62 are always OFF and are not
 256  *     part of the sign extension.
 257  *
 258  * Therefore, if:
 259  *
 260  *   1) LBR has TSX format
 261  *   2) CPU has no TSX support enabled
 262  *
 263  * ... then any value passed to wrmsr() must be sign extended to 63 bits and any
 264  * value from rdmsr() must be converted to have a 61 bits sign extension,
 265  * ignoring the TSX flags.
 266  */
 267 static inline bool lbr_from_signext_quirk_needed(void)
 268 {
 269         int lbr_format = x86_pmu.intel_cap.lbr_format;
 270         bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
 271                            boot_cpu_has(X86_FEATURE_RTM);
 272 
 273         return !tsx_support && (lbr_desc[lbr_format] & LBR_TSX);
 274 }
 275 
 276 static DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
 277 
 278 /* If quirk is enabled, ensure sign extension is 63 bits: */
 279 inline u64 lbr_from_signext_quirk_wr(u64 val)
 280 {
 281         if (static_branch_unlikely(&lbr_from_quirk_key)) {
 282                 /*
 283                  * Sign extend into bits 61:62 while preserving bit 63.
 284                  *
 285                  * Quirk is enabled when TSX is disabled. Therefore TSX bits
 286                  * in val are always OFF and must be changed to be sign
 287                  * extension bits. Since bits 59:60 are guaranteed to be
 288                  * part of the sign extension bits, we can just copy them
 289                  * to 61:62.
 290                  */
 291                 val |= (LBR_FROM_SIGNEXT_2MSB & val) << 2;
 292         }
 293         return val;
 294 }
 295 
 296 /*
 297  * If quirk is needed, ensure sign extension is 61 bits:
 298  */
 299 static u64 lbr_from_signext_quirk_rd(u64 val)
 300 {
 301         if (static_branch_unlikely(&lbr_from_quirk_key)) {
 302                 /*
 303                  * Quirk is on when TSX is not enabled. Therefore TSX
 304                  * flags must be read as OFF.
 305                  */
 306                 val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);
 307         }
 308         return val;
 309 }
 310 
 311 static inline void wrlbr_from(unsigned int idx, u64 val)
 312 {
 313         val = lbr_from_signext_quirk_wr(val);
 314         wrmsrl(x86_pmu.lbr_from + idx, val);
 315 }
 316 
 317 static inline void wrlbr_to(unsigned int idx, u64 val)
 318 {
 319         wrmsrl(x86_pmu.lbr_to + idx, val);
 320 }
 321 
 322 static inline u64 rdlbr_from(unsigned int idx)
 323 {
 324         u64 val;
 325 
 326         rdmsrl(x86_pmu.lbr_from + idx, val);
 327 
 328         return lbr_from_signext_quirk_rd(val);
 329 }
 330 
 331 static inline u64 rdlbr_to(unsigned int idx)
 332 {
 333         u64 val;
 334 
 335         rdmsrl(x86_pmu.lbr_to + idx, val);
 336 
 337         return val;
 338 }
 339 
 340 static void __intel_pmu_lbr_restore(struct x86_perf_task_context *task_ctx)
 341 {
 342         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 343         int i;
 344         unsigned lbr_idx, mask;
 345         u64 tos;
 346 
 347         if (task_ctx->lbr_callstack_users == 0 ||
 348             task_ctx->lbr_stack_state == LBR_NONE) {
 349                 intel_pmu_lbr_reset();
 350                 return;
 351         }
 352 
 353         tos = task_ctx->tos;
 354         /*
 355          * Does not restore the LBR registers, if
 356          * - No one else touched them, and
 357          * - Did not enter C6
 358          */
 359         if ((task_ctx == cpuc->last_task_ctx) &&
 360             (task_ctx->log_id == cpuc->last_log_id) &&
 361             rdlbr_from(tos)) {
 362                 task_ctx->lbr_stack_state = LBR_NONE;
 363                 return;
 364         }
 365 
 366         mask = x86_pmu.lbr_nr - 1;
 367         for (i = 0; i < task_ctx->valid_lbrs; i++) {
 368                 lbr_idx = (tos - i) & mask;
 369                 wrlbr_from(lbr_idx, task_ctx->lbr_from[i]);
 370                 wrlbr_to  (lbr_idx, task_ctx->lbr_to[i]);
 371 
 372                 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
 373                         wrmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
 374         }
 375 
 376         for (; i < x86_pmu.lbr_nr; i++) {
 377                 lbr_idx = (tos - i) & mask;
 378                 wrlbr_from(lbr_idx, 0);
 379                 wrlbr_to(lbr_idx, 0);
 380                 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
 381                         wrmsrl(MSR_LBR_INFO_0 + lbr_idx, 0);
 382         }
 383 
 384         wrmsrl(x86_pmu.lbr_tos, tos);
 385         task_ctx->lbr_stack_state = LBR_NONE;
 386 }
 387 
 388 static void __intel_pmu_lbr_save(struct x86_perf_task_context *task_ctx)
 389 {
 390         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 391         unsigned lbr_idx, mask;
 392         u64 tos, from;
 393         int i;
 394 
 395         if (task_ctx->lbr_callstack_users == 0) {
 396                 task_ctx->lbr_stack_state = LBR_NONE;
 397                 return;
 398         }
 399 
 400         mask = x86_pmu.lbr_nr - 1;
 401         tos = intel_pmu_lbr_tos();
 402         for (i = 0; i < x86_pmu.lbr_nr; i++) {
 403                 lbr_idx = (tos - i) & mask;
 404                 from = rdlbr_from(lbr_idx);
 405                 if (!from)
 406                         break;
 407                 task_ctx->lbr_from[i] = from;
 408                 task_ctx->lbr_to[i]   = rdlbr_to(lbr_idx);
 409                 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
 410                         rdmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
 411         }
 412         task_ctx->valid_lbrs = i;
 413         task_ctx->tos = tos;
 414         task_ctx->lbr_stack_state = LBR_VALID;
 415 
 416         cpuc->last_task_ctx = task_ctx;
 417         cpuc->last_log_id = ++task_ctx->log_id;
 418 }
 419 
 420 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
 421 {
 422         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 423         struct x86_perf_task_context *task_ctx;
 424 
 425         if (!cpuc->lbr_users)
 426                 return;
 427 
 428         /*
 429          * If LBR callstack feature is enabled and the stack was saved when
 430          * the task was scheduled out, restore the stack. Otherwise flush
 431          * the LBR stack.
 432          */
 433         task_ctx = ctx ? ctx->task_ctx_data : NULL;
 434         if (task_ctx) {
 435                 if (sched_in)
 436                         __intel_pmu_lbr_restore(task_ctx);
 437                 else
 438                         __intel_pmu_lbr_save(task_ctx);
 439                 return;
 440         }
 441 
 442         /*
 443          * Since a context switch can flip the address space and LBR entries
 444          * are not tagged with an identifier, we need to wipe the LBR, even for
 445          * per-cpu events. You simply cannot resolve the branches from the old
 446          * address space.
 447          */
 448         if (sched_in)
 449                 intel_pmu_lbr_reset();
 450 }
 451 
 452 static inline bool branch_user_callstack(unsigned br_sel)
 453 {
 454         return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
 455 }
 456 
 457 void intel_pmu_lbr_add(struct perf_event *event)
 458 {
 459         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 460         struct x86_perf_task_context *task_ctx;
 461 
 462         if (!x86_pmu.lbr_nr)
 463                 return;
 464 
 465         cpuc->br_sel = event->hw.branch_reg.reg;
 466 
 467         if (branch_user_callstack(cpuc->br_sel) && event->ctx->task_ctx_data) {
 468                 task_ctx = event->ctx->task_ctx_data;
 469                 task_ctx->lbr_callstack_users++;
 470         }
 471 
 472         /*
 473          * Request pmu::sched_task() callback, which will fire inside the
 474          * regular perf event scheduling, so that call will:
 475          *
 476          *  - restore or wipe; when LBR-callstack,
 477          *  - wipe; otherwise,
 478          *
 479          * when this is from __perf_event_task_sched_in().
 480          *
 481          * However, if this is from perf_install_in_context(), no such callback
 482          * will follow and we'll need to reset the LBR here if this is the
 483          * first LBR event.
 484          *
 485          * The problem is, we cannot tell these cases apart... but we can
 486          * exclude the biggest chunk of cases by looking at
 487          * event->total_time_running. An event that has accrued runtime cannot
 488          * be 'new'. Conversely, a new event can get installed through the
 489          * context switch path for the first time.
 490          */
 491         if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
 492                 cpuc->lbr_pebs_users++;
 493         perf_sched_cb_inc(event->ctx->pmu);
 494         if (!cpuc->lbr_users++ && !event->total_time_running)
 495                 intel_pmu_lbr_reset();
 496 }
 497 
 498 void intel_pmu_lbr_del(struct perf_event *event)
 499 {
 500         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 501         struct x86_perf_task_context *task_ctx;
 502 
 503         if (!x86_pmu.lbr_nr)
 504                 return;
 505 
 506         if (branch_user_callstack(cpuc->br_sel) &&
 507             event->ctx->task_ctx_data) {
 508                 task_ctx = event->ctx->task_ctx_data;
 509                 task_ctx->lbr_callstack_users--;
 510         }
 511 
 512         if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
 513                 cpuc->lbr_pebs_users--;
 514         cpuc->lbr_users--;
 515         WARN_ON_ONCE(cpuc->lbr_users < 0);
 516         WARN_ON_ONCE(cpuc->lbr_pebs_users < 0);
 517         perf_sched_cb_dec(event->ctx->pmu);
 518 }
 519 
 520 void intel_pmu_lbr_enable_all(bool pmi)
 521 {
 522         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 523 
 524         if (cpuc->lbr_users)
 525                 __intel_pmu_lbr_enable(pmi);
 526 }
 527 
 528 void intel_pmu_lbr_disable_all(void)
 529 {
 530         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 531 
 532         if (cpuc->lbr_users)
 533                 __intel_pmu_lbr_disable();
 534 }
 535 
 536 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
 537 {
 538         unsigned long mask = x86_pmu.lbr_nr - 1;
 539         u64 tos = intel_pmu_lbr_tos();
 540         int i;
 541 
 542         for (i = 0; i < x86_pmu.lbr_nr; i++) {
 543                 unsigned long lbr_idx = (tos - i) & mask;
 544                 union {
 545                         struct {
 546                                 u32 from;
 547                                 u32 to;
 548                         };
 549                         u64     lbr;
 550                 } msr_lastbranch;
 551 
 552                 rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
 553 
 554                 cpuc->lbr_entries[i].from       = msr_lastbranch.from;
 555                 cpuc->lbr_entries[i].to         = msr_lastbranch.to;
 556                 cpuc->lbr_entries[i].mispred    = 0;
 557                 cpuc->lbr_entries[i].predicted  = 0;
 558                 cpuc->lbr_entries[i].in_tx      = 0;
 559                 cpuc->lbr_entries[i].abort      = 0;
 560                 cpuc->lbr_entries[i].cycles     = 0;
 561                 cpuc->lbr_entries[i].type       = 0;
 562                 cpuc->lbr_entries[i].reserved   = 0;
 563         }
 564         cpuc->lbr_stack.nr = i;
 565 }
 566 
 567 /*
 568  * Due to lack of segmentation in Linux the effective address (offset)
 569  * is the same as the linear address, allowing us to merge the LIP and EIP
 570  * LBR formats.
 571  */
 572 static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
 573 {
 574         bool need_info = false, call_stack = false;
 575         unsigned long mask = x86_pmu.lbr_nr - 1;
 576         int lbr_format = x86_pmu.intel_cap.lbr_format;
 577         u64 tos = intel_pmu_lbr_tos();
 578         int i;
 579         int out = 0;
 580         int num = x86_pmu.lbr_nr;
 581 
 582         if (cpuc->lbr_sel) {
 583                 need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
 584                 if (cpuc->lbr_sel->config & LBR_CALL_STACK)
 585                         call_stack = true;
 586         }
 587 
 588         for (i = 0; i < num; i++) {
 589                 unsigned long lbr_idx = (tos - i) & mask;
 590                 u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
 591                 int skip = 0;
 592                 u16 cycles = 0;
 593                 int lbr_flags = lbr_desc[lbr_format];
 594 
 595                 from = rdlbr_from(lbr_idx);
 596                 to   = rdlbr_to(lbr_idx);
 597 
 598                 /*
 599                  * Read LBR call stack entries
 600                  * until invalid entry (0s) is detected.
 601                  */
 602                 if (call_stack && !from)
 603                         break;
 604 
 605                 if (lbr_format == LBR_FORMAT_INFO && need_info) {
 606                         u64 info;
 607 
 608                         rdmsrl(MSR_LBR_INFO_0 + lbr_idx, info);
 609                         mis = !!(info & LBR_INFO_MISPRED);
 610                         pred = !mis;
 611                         in_tx = !!(info & LBR_INFO_IN_TX);
 612                         abort = !!(info & LBR_INFO_ABORT);
 613                         cycles = (info & LBR_INFO_CYCLES);
 614                 }
 615 
 616                 if (lbr_format == LBR_FORMAT_TIME) {
 617                         mis = !!(from & LBR_FROM_FLAG_MISPRED);
 618                         pred = !mis;
 619                         skip = 1;
 620                         cycles = ((to >> 48) & LBR_INFO_CYCLES);
 621 
 622                         to = (u64)((((s64)to) << 16) >> 16);
 623                 }
 624 
 625                 if (lbr_flags & LBR_EIP_FLAGS) {
 626                         mis = !!(from & LBR_FROM_FLAG_MISPRED);
 627                         pred = !mis;
 628                         skip = 1;
 629                 }
 630                 if (lbr_flags & LBR_TSX) {
 631                         in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
 632                         abort = !!(from & LBR_FROM_FLAG_ABORT);
 633                         skip = 3;
 634                 }
 635                 from = (u64)((((s64)from) << skip) >> skip);
 636 
 637                 /*
 638                  * Some CPUs report duplicated abort records,
 639                  * with the second entry not having an abort bit set.
 640                  * Skip them here. This loop runs backwards,
 641                  * so we need to undo the previous record.
 642                  * If the abort just happened outside the window
 643                  * the extra entry cannot be removed.
 644                  */
 645                 if (abort && x86_pmu.lbr_double_abort && out > 0)
 646                         out--;
 647 
 648                 cpuc->lbr_entries[out].from      = from;
 649                 cpuc->lbr_entries[out].to        = to;
 650                 cpuc->lbr_entries[out].mispred   = mis;
 651                 cpuc->lbr_entries[out].predicted = pred;
 652                 cpuc->lbr_entries[out].in_tx     = in_tx;
 653                 cpuc->lbr_entries[out].abort     = abort;
 654                 cpuc->lbr_entries[out].cycles    = cycles;
 655                 cpuc->lbr_entries[out].type      = 0;
 656                 cpuc->lbr_entries[out].reserved  = 0;
 657                 out++;
 658         }
 659         cpuc->lbr_stack.nr = out;
 660 }
 661 
 662 void intel_pmu_lbr_read(void)
 663 {
 664         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 665 
 666         /*
 667          * Don't read when all LBRs users are using adaptive PEBS.
 668          *
 669          * This could be smarter and actually check the event,
 670          * but this simple approach seems to work for now.
 671          */
 672         if (!cpuc->lbr_users || cpuc->lbr_users == cpuc->lbr_pebs_users)
 673                 return;
 674 
 675         if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
 676                 intel_pmu_lbr_read_32(cpuc);
 677         else
 678                 intel_pmu_lbr_read_64(cpuc);
 679 
 680         intel_pmu_lbr_filter(cpuc);
 681 }
 682 
 683 /*
 684  * SW filter is used:
 685  * - in case there is no HW filter
 686  * - in case the HW filter has errata or limitations
 687  */
 688 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
 689 {
 690         u64 br_type = event->attr.branch_sample_type;
 691         int mask = 0;
 692 
 693         if (br_type & PERF_SAMPLE_BRANCH_USER)
 694                 mask |= X86_BR_USER;
 695 
 696         if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
 697                 mask |= X86_BR_KERNEL;
 698 
 699         /* we ignore BRANCH_HV here */
 700 
 701         if (br_type & PERF_SAMPLE_BRANCH_ANY)
 702                 mask |= X86_BR_ANY;
 703 
 704         if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
 705                 mask |= X86_BR_ANY_CALL;
 706 
 707         if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
 708                 mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
 709 
 710         if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
 711                 mask |= X86_BR_IND_CALL;
 712 
 713         if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
 714                 mask |= X86_BR_ABORT;
 715 
 716         if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
 717                 mask |= X86_BR_IN_TX;
 718 
 719         if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
 720                 mask |= X86_BR_NO_TX;
 721 
 722         if (br_type & PERF_SAMPLE_BRANCH_COND)
 723                 mask |= X86_BR_JCC;
 724 
 725         if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
 726                 if (!x86_pmu_has_lbr_callstack())
 727                         return -EOPNOTSUPP;
 728                 if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
 729                         return -EINVAL;
 730                 mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
 731                         X86_BR_CALL_STACK;
 732         }
 733 
 734         if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
 735                 mask |= X86_BR_IND_JMP;
 736 
 737         if (br_type & PERF_SAMPLE_BRANCH_CALL)
 738                 mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
 739 
 740         if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
 741                 mask |= X86_BR_TYPE_SAVE;
 742 
 743         /*
 744          * stash actual user request into reg, it may
 745          * be used by fixup code for some CPU
 746          */
 747         event->hw.branch_reg.reg = mask;
 748         return 0;
 749 }
 750 
 751 /*
 752  * setup the HW LBR filter
 753  * Used only when available, may not be enough to disambiguate
 754  * all branches, may need the help of the SW filter
 755  */
 756 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
 757 {
 758         struct hw_perf_event_extra *reg;
 759         u64 br_type = event->attr.branch_sample_type;
 760         u64 mask = 0, v;
 761         int i;
 762 
 763         for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
 764                 if (!(br_type & (1ULL << i)))
 765                         continue;
 766 
 767                 v = x86_pmu.lbr_sel_map[i];
 768                 if (v == LBR_NOT_SUPP)
 769                         return -EOPNOTSUPP;
 770 
 771                 if (v != LBR_IGN)
 772                         mask |= v;
 773         }
 774 
 775         reg = &event->hw.branch_reg;
 776         reg->idx = EXTRA_REG_LBR;
 777 
 778         /*
 779          * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
 780          * in suppress mode. So LBR_SELECT should be set to
 781          * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
 782          * But the 10th bit LBR_CALL_STACK does not operate
 783          * in suppress mode.
 784          */
 785         reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
 786 
 787         if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
 788             (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
 789             (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
 790                 reg->config |= LBR_NO_INFO;
 791 
 792         return 0;
 793 }
 794 
 795 int intel_pmu_setup_lbr_filter(struct perf_event *event)
 796 {
 797         int ret = 0;
 798 
 799         /*
 800          * no LBR on this PMU
 801          */
 802         if (!x86_pmu.lbr_nr)
 803                 return -EOPNOTSUPP;
 804 
 805         /*
 806          * setup SW LBR filter
 807          */
 808         ret = intel_pmu_setup_sw_lbr_filter(event);
 809         if (ret)
 810                 return ret;
 811 
 812         /*
 813          * setup HW LBR filter, if any
 814          */
 815         if (x86_pmu.lbr_sel_map)
 816                 ret = intel_pmu_setup_hw_lbr_filter(event);
 817 
 818         return ret;
 819 }
 820 
 821 /*
 822  * return the type of control flow change at address "from"
 823  * instruction is not necessarily a branch (in case of interrupt).
 824  *
 825  * The branch type returned also includes the priv level of the
 826  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
 827  *
 828  * If a branch type is unknown OR the instruction cannot be
 829  * decoded (e.g., text page not present), then X86_BR_NONE is
 830  * returned.
 831  */
 832 static int branch_type(unsigned long from, unsigned long to, int abort)
 833 {
 834         struct insn insn;
 835         void *addr;
 836         int bytes_read, bytes_left;
 837         int ret = X86_BR_NONE;
 838         int ext, to_plm, from_plm;
 839         u8 buf[MAX_INSN_SIZE];
 840         int is64 = 0;
 841 
 842         to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
 843         from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
 844 
 845         /*
 846          * maybe zero if lbr did not fill up after a reset by the time
 847          * we get a PMU interrupt
 848          */
 849         if (from == 0 || to == 0)
 850                 return X86_BR_NONE;
 851 
 852         if (abort)
 853                 return X86_BR_ABORT | to_plm;
 854 
 855         if (from_plm == X86_BR_USER) {
 856                 /*
 857                  * can happen if measuring at the user level only
 858                  * and we interrupt in a kernel thread, e.g., idle.
 859                  */
 860                 if (!current->mm)
 861                         return X86_BR_NONE;
 862 
 863                 /* may fail if text not present */
 864                 bytes_left = copy_from_user_nmi(buf, (void __user *)from,
 865                                                 MAX_INSN_SIZE);
 866                 bytes_read = MAX_INSN_SIZE - bytes_left;
 867                 if (!bytes_read)
 868                         return X86_BR_NONE;
 869 
 870                 addr = buf;
 871         } else {
 872                 /*
 873                  * The LBR logs any address in the IP, even if the IP just
 874                  * faulted. This means userspace can control the from address.
 875                  * Ensure we don't blindy read any address by validating it is
 876                  * a known text address.
 877                  */
 878                 if (kernel_text_address(from)) {
 879                         addr = (void *)from;
 880                         /*
 881                          * Assume we can get the maximum possible size
 882                          * when grabbing kernel data.  This is not
 883                          * _strictly_ true since we could possibly be
 884                          * executing up next to a memory hole, but
 885                          * it is very unlikely to be a problem.
 886                          */
 887                         bytes_read = MAX_INSN_SIZE;
 888                 } else {
 889                         return X86_BR_NONE;
 890                 }
 891         }
 892 
 893         /*
 894          * decoder needs to know the ABI especially
 895          * on 64-bit systems running 32-bit apps
 896          */
 897 #ifdef CONFIG_X86_64
 898         is64 = kernel_ip((unsigned long)addr) || !test_thread_flag(TIF_IA32);
 899 #endif
 900         insn_init(&insn, addr, bytes_read, is64);
 901         insn_get_opcode(&insn);
 902         if (!insn.opcode.got)
 903                 return X86_BR_ABORT;
 904 
 905         switch (insn.opcode.bytes[0]) {
 906         case 0xf:
 907                 switch (insn.opcode.bytes[1]) {
 908                 case 0x05: /* syscall */
 909                 case 0x34: /* sysenter */
 910                         ret = X86_BR_SYSCALL;
 911                         break;
 912                 case 0x07: /* sysret */
 913                 case 0x35: /* sysexit */
 914                         ret = X86_BR_SYSRET;
 915                         break;
 916                 case 0x80 ... 0x8f: /* conditional */
 917                         ret = X86_BR_JCC;
 918                         break;
 919                 default:
 920                         ret = X86_BR_NONE;
 921                 }
 922                 break;
 923         case 0x70 ... 0x7f: /* conditional */
 924                 ret = X86_BR_JCC;
 925                 break;
 926         case 0xc2: /* near ret */
 927         case 0xc3: /* near ret */
 928         case 0xca: /* far ret */
 929         case 0xcb: /* far ret */
 930                 ret = X86_BR_RET;
 931                 break;
 932         case 0xcf: /* iret */
 933                 ret = X86_BR_IRET;
 934                 break;
 935         case 0xcc ... 0xce: /* int */
 936                 ret = X86_BR_INT;
 937                 break;
 938         case 0xe8: /* call near rel */
 939                 insn_get_immediate(&insn);
 940                 if (insn.immediate1.value == 0) {
 941                         /* zero length call */
 942                         ret = X86_BR_ZERO_CALL;
 943                         break;
 944                 }
 945                 /* fall through */
 946         case 0x9a: /* call far absolute */
 947                 ret = X86_BR_CALL;
 948                 break;
 949         case 0xe0 ... 0xe3: /* loop jmp */
 950                 ret = X86_BR_JCC;
 951                 break;
 952         case 0xe9 ... 0xeb: /* jmp */
 953                 ret = X86_BR_JMP;
 954                 break;
 955         case 0xff: /* call near absolute, call far absolute ind */
 956                 insn_get_modrm(&insn);
 957                 ext = (insn.modrm.bytes[0] >> 3) & 0x7;
 958                 switch (ext) {
 959                 case 2: /* near ind call */
 960                 case 3: /* far ind call */
 961                         ret = X86_BR_IND_CALL;
 962                         break;
 963                 case 4:
 964                 case 5:
 965                         ret = X86_BR_IND_JMP;
 966                         break;
 967                 }
 968                 break;
 969         default:
 970                 ret = X86_BR_NONE;
 971         }
 972         /*
 973          * interrupts, traps, faults (and thus ring transition) may
 974          * occur on any instructions. Thus, to classify them correctly,
 975          * we need to first look at the from and to priv levels. If they
 976          * are different and to is in the kernel, then it indicates
 977          * a ring transition. If the from instruction is not a ring
 978          * transition instr (syscall, systenter, int), then it means
 979          * it was a irq, trap or fault.
 980          *
 981          * we have no way of detecting kernel to kernel faults.
 982          */
 983         if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
 984             && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
 985                 ret = X86_BR_IRQ;
 986 
 987         /*
 988          * branch priv level determined by target as
 989          * is done by HW when LBR_SELECT is implemented
 990          */
 991         if (ret != X86_BR_NONE)
 992                 ret |= to_plm;
 993 
 994         return ret;
 995 }
 996 
 997 #define X86_BR_TYPE_MAP_MAX     16
 998 
 999 static int branch_map[X86_BR_TYPE_MAP_MAX] = {
1000         PERF_BR_CALL,           /* X86_BR_CALL */
1001         PERF_BR_RET,            /* X86_BR_RET */
1002         PERF_BR_SYSCALL,        /* X86_BR_SYSCALL */
1003         PERF_BR_SYSRET,         /* X86_BR_SYSRET */
1004         PERF_BR_UNKNOWN,        /* X86_BR_INT */
1005         PERF_BR_UNKNOWN,        /* X86_BR_IRET */
1006         PERF_BR_COND,           /* X86_BR_JCC */
1007         PERF_BR_UNCOND,         /* X86_BR_JMP */
1008         PERF_BR_UNKNOWN,        /* X86_BR_IRQ */
1009         PERF_BR_IND_CALL,       /* X86_BR_IND_CALL */
1010         PERF_BR_UNKNOWN,        /* X86_BR_ABORT */
1011         PERF_BR_UNKNOWN,        /* X86_BR_IN_TX */
1012         PERF_BR_UNKNOWN,        /* X86_BR_NO_TX */
1013         PERF_BR_CALL,           /* X86_BR_ZERO_CALL */
1014         PERF_BR_UNKNOWN,        /* X86_BR_CALL_STACK */
1015         PERF_BR_IND,            /* X86_BR_IND_JMP */
1016 };
1017 
1018 static int
1019 common_branch_type(int type)
1020 {
1021         int i;
1022 
1023         type >>= 2; /* skip X86_BR_USER and X86_BR_KERNEL */
1024 
1025         if (type) {
1026                 i = __ffs(type);
1027                 if (i < X86_BR_TYPE_MAP_MAX)
1028                         return branch_map[i];
1029         }
1030 
1031         return PERF_BR_UNKNOWN;
1032 }
1033 
1034 /*
1035  * implement actual branch filter based on user demand.
1036  * Hardware may not exactly satisfy that request, thus
1037  * we need to inspect opcodes. Mismatched branches are
1038  * discarded. Therefore, the number of branches returned
1039  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1040  */
1041 static void
1042 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1043 {
1044         u64 from, to;
1045         int br_sel = cpuc->br_sel;
1046         int i, j, type;
1047         bool compress = false;
1048 
1049         /* if sampling all branches, then nothing to filter */
1050         if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1051             ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1052                 return;
1053 
1054         for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1055 
1056                 from = cpuc->lbr_entries[i].from;
1057                 to = cpuc->lbr_entries[i].to;
1058 
1059                 type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1060                 if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1061                         if (cpuc->lbr_entries[i].in_tx)
1062                                 type |= X86_BR_IN_TX;
1063                         else
1064                                 type |= X86_BR_NO_TX;
1065                 }
1066 
1067                 /* if type does not correspond, then discard */
1068                 if (type == X86_BR_NONE || (br_sel & type) != type) {
1069                         cpuc->lbr_entries[i].from = 0;
1070                         compress = true;
1071                 }
1072 
1073                 if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1074                         cpuc->lbr_entries[i].type = common_branch_type(type);
1075         }
1076 
1077         if (!compress)
1078                 return;
1079 
1080         /* remove all entries with from=0 */
1081         for (i = 0; i < cpuc->lbr_stack.nr; ) {
1082                 if (!cpuc->lbr_entries[i].from) {
1083                         j = i;
1084                         while (++j < cpuc->lbr_stack.nr)
1085                                 cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1086                         cpuc->lbr_stack.nr--;
1087                         if (!cpuc->lbr_entries[i].from)
1088                                 continue;
1089                 }
1090                 i++;
1091         }
1092 }
1093 
1094 void intel_pmu_store_pebs_lbrs(struct pebs_lbr *lbr)
1095 {
1096         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1097         int i;
1098 
1099         cpuc->lbr_stack.nr = x86_pmu.lbr_nr;
1100         for (i = 0; i < x86_pmu.lbr_nr; i++) {
1101                 u64 info = lbr->lbr[i].info;
1102                 struct perf_branch_entry *e = &cpuc->lbr_entries[i];
1103 
1104                 e->from         = lbr->lbr[i].from;
1105                 e->to           = lbr->lbr[i].to;
1106                 e->mispred      = !!(info & LBR_INFO_MISPRED);
1107                 e->predicted    = !(info & LBR_INFO_MISPRED);
1108                 e->in_tx        = !!(info & LBR_INFO_IN_TX);
1109                 e->abort        = !!(info & LBR_INFO_ABORT);
1110                 e->cycles       = info & LBR_INFO_CYCLES;
1111                 e->reserved     = 0;
1112         }
1113         intel_pmu_lbr_filter(cpuc);
1114 }
1115 
1116 /*
1117  * Map interface branch filters onto LBR filters
1118  */
1119 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1120         [PERF_SAMPLE_BRANCH_ANY_SHIFT]          = LBR_ANY,
1121         [PERF_SAMPLE_BRANCH_USER_SHIFT]         = LBR_USER,
1122         [PERF_SAMPLE_BRANCH_KERNEL_SHIFT]       = LBR_KERNEL,
1123         [PERF_SAMPLE_BRANCH_HV_SHIFT]           = LBR_IGN,
1124         [PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]   = LBR_RETURN | LBR_REL_JMP
1125                                                 | LBR_IND_JMP | LBR_FAR,
1126         /*
1127          * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1128          */
1129         [PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1130          LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1131         /*
1132          * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1133          */
1134         [PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1135         [PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1136         [PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1137 };
1138 
1139 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1140         [PERF_SAMPLE_BRANCH_ANY_SHIFT]          = LBR_ANY,
1141         [PERF_SAMPLE_BRANCH_USER_SHIFT]         = LBR_USER,
1142         [PERF_SAMPLE_BRANCH_KERNEL_SHIFT]       = LBR_KERNEL,
1143         [PERF_SAMPLE_BRANCH_HV_SHIFT]           = LBR_IGN,
1144         [PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]   = LBR_RETURN | LBR_FAR,
1145         [PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]     = LBR_REL_CALL | LBR_IND_CALL
1146                                                 | LBR_FAR,
1147         [PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]     = LBR_IND_CALL,
1148         [PERF_SAMPLE_BRANCH_COND_SHIFT]         = LBR_JCC,
1149         [PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]     = LBR_IND_JMP,
1150         [PERF_SAMPLE_BRANCH_CALL_SHIFT]         = LBR_REL_CALL,
1151 };
1152 
1153 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1154         [PERF_SAMPLE_BRANCH_ANY_SHIFT]          = LBR_ANY,
1155         [PERF_SAMPLE_BRANCH_USER_SHIFT]         = LBR_USER,
1156         [PERF_SAMPLE_BRANCH_KERNEL_SHIFT]       = LBR_KERNEL,
1157         [PERF_SAMPLE_BRANCH_HV_SHIFT]           = LBR_IGN,
1158         [PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]   = LBR_RETURN | LBR_FAR,
1159         [PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]     = LBR_REL_CALL | LBR_IND_CALL
1160                                                 | LBR_FAR,
1161         [PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]     = LBR_IND_CALL,
1162         [PERF_SAMPLE_BRANCH_COND_SHIFT]         = LBR_JCC,
1163         [PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]   = LBR_REL_CALL | LBR_IND_CALL
1164                                                 | LBR_RETURN | LBR_CALL_STACK,
1165         [PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]     = LBR_IND_JMP,
1166         [PERF_SAMPLE_BRANCH_CALL_SHIFT]         = LBR_REL_CALL,
1167 };
1168 
1169 /* core */
1170 void __init intel_pmu_lbr_init_core(void)
1171 {
1172         x86_pmu.lbr_nr     = 4;
1173         x86_pmu.lbr_tos    = MSR_LBR_TOS;
1174         x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1175         x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1176 
1177         /*
1178          * SW branch filter usage:
1179          * - compensate for lack of HW filter
1180          */
1181 }
1182 
1183 /* nehalem/westmere */
1184 void __init intel_pmu_lbr_init_nhm(void)
1185 {
1186         x86_pmu.lbr_nr     = 16;
1187         x86_pmu.lbr_tos    = MSR_LBR_TOS;
1188         x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1189         x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1190 
1191         x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1192         x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1193 
1194         /*
1195          * SW branch filter usage:
1196          * - workaround LBR_SEL errata (see above)
1197          * - support syscall, sysret capture.
1198          *   That requires LBR_FAR but that means far
1199          *   jmp need to be filtered out
1200          */
1201 }
1202 
1203 /* sandy bridge */
1204 void __init intel_pmu_lbr_init_snb(void)
1205 {
1206         x86_pmu.lbr_nr   = 16;
1207         x86_pmu.lbr_tos  = MSR_LBR_TOS;
1208         x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1209         x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1210 
1211         x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1212         x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1213 
1214         /*
1215          * SW branch filter usage:
1216          * - support syscall, sysret capture.
1217          *   That requires LBR_FAR but that means far
1218          *   jmp need to be filtered out
1219          */
1220 }
1221 
1222 /* haswell */
1223 void intel_pmu_lbr_init_hsw(void)
1224 {
1225         x86_pmu.lbr_nr   = 16;
1226         x86_pmu.lbr_tos  = MSR_LBR_TOS;
1227         x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1228         x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1229 
1230         x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1231         x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1232 
1233         if (lbr_from_signext_quirk_needed())
1234                 static_branch_enable(&lbr_from_quirk_key);
1235 }
1236 
1237 /* skylake */
1238 __init void intel_pmu_lbr_init_skl(void)
1239 {
1240         x86_pmu.lbr_nr   = 32;
1241         x86_pmu.lbr_tos  = MSR_LBR_TOS;
1242         x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1243         x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1244 
1245         x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1246         x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1247 
1248         /*
1249          * SW branch filter usage:
1250          * - support syscall, sysret capture.
1251          *   That requires LBR_FAR but that means far
1252          *   jmp need to be filtered out
1253          */
1254 }
1255 
1256 /* atom */
1257 void __init intel_pmu_lbr_init_atom(void)
1258 {
1259         /*
1260          * only models starting at stepping 10 seems
1261          * to have an operational LBR which can freeze
1262          * on PMU interrupt
1263          */
1264         if (boot_cpu_data.x86_model == 28
1265             && boot_cpu_data.x86_stepping < 10) {
1266                 pr_cont("LBR disabled due to erratum");
1267                 return;
1268         }
1269 
1270         x86_pmu.lbr_nr     = 8;
1271         x86_pmu.lbr_tos    = MSR_LBR_TOS;
1272         x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1273         x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1274 
1275         /*
1276          * SW branch filter usage:
1277          * - compensate for lack of HW filter
1278          */
1279 }
1280 
1281 /* slm */
1282 void __init intel_pmu_lbr_init_slm(void)
1283 {
1284         x86_pmu.lbr_nr     = 8;
1285         x86_pmu.lbr_tos    = MSR_LBR_TOS;
1286         x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1287         x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1288 
1289         x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1290         x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1291 
1292         /*
1293          * SW branch filter usage:
1294          * - compensate for lack of HW filter
1295          */
1296         pr_cont("8-deep LBR, ");
1297 }
1298 
1299 /* Knights Landing */
1300 void intel_pmu_lbr_init_knl(void)
1301 {
1302         x86_pmu.lbr_nr     = 8;
1303         x86_pmu.lbr_tos    = MSR_LBR_TOS;
1304         x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1305         x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1306 
1307         x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1308         x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1309 
1310         /* Knights Landing does have MISPREDICT bit */
1311         if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1312                 x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1313 }

/* [<][>][^][v][top][bottom][index][help] */