tools/perf/util/intel-pt-decoder/intel-pt-decoder.c

   1 /*
   2  * intel_pt_decoder.c: Intel Processor Trace support
   3  * Copyright (c) 2013-2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  */
  15
  16 #ifndef _GNU_SOURCE
  17 #define _GNU_SOURCE
  18 #endif
  19 #include <stdlib.h>
  20 #include <stdbool.h>
  21 #include <string.h>
  22 #include <errno.h>
  23 #include <stdint.h>
  24 #include <inttypes.h>
  25 #include <linux/compiler.h>
  26
  27 #include "../cache.h"
  28 #include "../util.h"
  29
  30 #include "intel-pt-insn-decoder.h"
  31 #include "intel-pt-pkt-decoder.h"
  32 #include "intel-pt-decoder.h"
  33 #include "intel-pt-log.h"
  34
  35 #define INTEL_PT_BLK_SIZE 1024
  36
  37 #define BIT63 (((uint64_t)1 << 63))
  38
  39 #define INTEL_PT_RETURN 1
  40
  41 /* Maximum number of loops with no packets consumed i.e. stuck in a loop */
  42 #define INTEL_PT_MAX_LOOPS 10000
  43
  44 struct intel_pt_blk {
  45         struct intel_pt_blk *prev;
  46         uint64_t ip[INTEL_PT_BLK_SIZE];
  47 };
  48
  49 struct intel_pt_stack {
  50         struct intel_pt_blk *blk;
  51         struct intel_pt_blk *spare;
  52         int pos;
  53 };
  54
  55 enum intel_pt_pkt_state {
  56         INTEL_PT_STATE_NO_PSB,
  57         INTEL_PT_STATE_NO_IP,
  58         INTEL_PT_STATE_ERR_RESYNC,
  59         INTEL_PT_STATE_IN_SYNC,
  60         INTEL_PT_STATE_TNT,
  61         INTEL_PT_STATE_TIP,
  62         INTEL_PT_STATE_TIP_PGD,
  63         INTEL_PT_STATE_FUP,
  64         INTEL_PT_STATE_FUP_NO_TIP,
  65 };
  66
  67 static inline bool intel_pt_sample_time(enum intel_pt_pkt_state pkt_state)
  68 {
  69         switch (pkt_state) {
  70         case INTEL_PT_STATE_NO_PSB:
  71         case INTEL_PT_STATE_NO_IP:
  72         case INTEL_PT_STATE_ERR_RESYNC:
  73         case INTEL_PT_STATE_IN_SYNC:
  74         case INTEL_PT_STATE_TNT:
  75                 return true;
  76         case INTEL_PT_STATE_TIP:
  77         case INTEL_PT_STATE_TIP_PGD:
  78         case INTEL_PT_STATE_FUP:
  79         case INTEL_PT_STATE_FUP_NO_TIP:
  80                 return false;
  81         default:
  82                 return true;
  83         };
  84 }
  85
  86 #ifdef INTEL_PT_STRICT
  87 #define INTEL_PT_STATE_ERR1     INTEL_PT_STATE_NO_PSB
  88 #define INTEL_PT_STATE_ERR2     INTEL_PT_STATE_NO_PSB
  89 #define INTEL_PT_STATE_ERR3     INTEL_PT_STATE_NO_PSB
  90 #define INTEL_PT_STATE_ERR4     INTEL_PT_STATE_NO_PSB
  91 #else
  92 #define INTEL_PT_STATE_ERR1     (decoder->pkt_state)
  93 #define INTEL_PT_STATE_ERR2     INTEL_PT_STATE_NO_IP
  94 #define INTEL_PT_STATE_ERR3     INTEL_PT_STATE_ERR_RESYNC
  95 #define INTEL_PT_STATE_ERR4     INTEL_PT_STATE_IN_SYNC
  96 #endif
  97
  98 struct intel_pt_decoder {
  99         int (*get_trace)(struct intel_pt_buffer *buffer, void *data);
 100         int (*walk_insn)(struct intel_pt_insn *intel_pt_insn,
 101                          uint64_t *insn_cnt_ptr, uint64_t *ip, uint64_t to_ip,
 102                          uint64_t max_insn_cnt, void *data);
 103         bool (*pgd_ip)(uint64_t ip, void *data);
 104         void *data;
 105         struct intel_pt_state state;
 106         const unsigned char *buf;
 107         size_t len;
 108         bool return_compression;
 109         bool branch_enable;
 110         bool mtc_insn;
 111         bool pge;
 112         bool have_tma;
 113         bool have_cyc;
 114         bool fixup_last_mtc;
 115         bool have_last_ip;
 116         enum intel_pt_param_flags flags;
 117         uint64_t pos;
 118         uint64_t last_ip;
 119         uint64_t ip;
 120         uint64_t cr3;
 121         uint64_t timestamp;
 122         uint64_t tsc_timestamp;
 123         uint64_t ref_timestamp;
 124         uint64_t sample_timestamp;
 125         uint64_t ret_addr;
 126         uint64_t ctc_timestamp;
 127         uint64_t ctc_delta;
 128         uint64_t cycle_cnt;
 129         uint64_t cyc_ref_timestamp;
 130         uint32_t last_mtc;
 131         uint32_t tsc_ctc_ratio_n;
 132         uint32_t tsc_ctc_ratio_d;
 133         uint32_t tsc_ctc_mult;
 134         uint32_t tsc_slip;
 135         uint32_t ctc_rem_mask;
 136         int mtc_shift;
 137         struct intel_pt_stack stack;
 138         enum intel_pt_pkt_state pkt_state;
 139         struct intel_pt_pkt packet;
 140         struct intel_pt_pkt tnt;
 141         int pkt_step;
 142         int pkt_len;
 143         int last_packet_type;
 144         unsigned int cbr;
 145         unsigned int cbr_seen;
 146         unsigned int max_non_turbo_ratio;
 147         double max_non_turbo_ratio_fp;
 148         double cbr_cyc_to_tsc;
 149         double calc_cyc_to_tsc;
 150         bool have_calc_cyc_to_tsc;
 151         int exec_mode;
 152         unsigned int insn_bytes;
 153         uint64_t period;
 154         enum intel_pt_period_type period_type;
 155         uint64_t tot_insn_cnt;
 156         uint64_t period_insn_cnt;
 157         uint64_t period_mask;
 158         uint64_t period_ticks;
 159         uint64_t last_masked_timestamp;
 160         bool continuous_period;
 161         bool overflow;
 162         bool set_fup_tx_flags;
 163         bool set_fup_ptw;
 164         bool set_fup_mwait;
 165         bool set_fup_pwre;
 166         bool set_fup_exstop;
 167         unsigned int fup_tx_flags;
 168         unsigned int tx_flags;
 169         uint64_t fup_ptw_payload;
 170         uint64_t fup_mwait_payload;
 171         uint64_t fup_pwre_payload;
 172         uint64_t cbr_payload;
 173         uint64_t timestamp_insn_cnt;
 174         uint64_t sample_insn_cnt;
 175         uint64_t stuck_ip;
 176         int no_progress;
 177         int stuck_ip_prd;
 178         int stuck_ip_cnt;
 179         const unsigned char *next_buf;
 180         size_t next_len;
 181         unsigned char temp_buf[INTEL_PT_PKT_MAX_SZ];
 182 };
 183
 184 static uint64_t intel_pt_lower_power_of_2(uint64_t x)
 185 {
 186         int i;
 187
 188         for (i = 0; x != 1; i++)
 189                 x >>= 1;
 190
 191         return x << i;
 192 }
 193
 194 static void intel_pt_setup_period(struct intel_pt_decoder *decoder)
 195 {
 196         if (decoder->period_type == INTEL_PT_PERIOD_TICKS) {
 197                 uint64_t period;
 198
 199                 period = intel_pt_lower_power_of_2(decoder->period);
 200                 decoder->period_mask  = ~(period - 1);
 201                 decoder->period_ticks = period;
 202         }
 203 }
 204
 205 static uint64_t multdiv(uint64_t t, uint32_t n, uint32_t d)
 206 {
 207         if (!d)
 208                 return 0;
 209         return (t / d) * n + ((t % d) * n) / d;
 210 }
 211
 212 struct intel_pt_decoder *intel_pt_decoder_new(struct intel_pt_params *params)
 213 {
 214         struct intel_pt_decoder *decoder;
 215
 216         if (!params->get_trace || !params->walk_insn)
 217                 return NULL;
 218
 219         decoder = zalloc(sizeof(struct intel_pt_decoder));
 220         if (!decoder)
 221                 return NULL;
 222
 223         decoder->get_trace          = params->get_trace;
 224         decoder->walk_insn          = params->walk_insn;
 225         decoder->pgd_ip             = params->pgd_ip;
 226         decoder->data               = params->data;
 227         decoder->return_compression = params->return_compression;
 228         decoder->branch_enable      = params->branch_enable;
 229
 230         decoder->flags              = params->flags;
 231
 232         decoder->period             = params->period;
 233         decoder->period_type        = params->period_type;
 234
 235         decoder->max_non_turbo_ratio    = params->max_non_turbo_ratio;
 236         decoder->max_non_turbo_ratio_fp = params->max_non_turbo_ratio;
 237
 238         intel_pt_setup_period(decoder);
 239
 240         decoder->mtc_shift = params->mtc_period;
 241         decoder->ctc_rem_mask = (1 << decoder->mtc_shift) - 1;
 242
 243         decoder->tsc_ctc_ratio_n = params->tsc_ctc_ratio_n;
 244         decoder->tsc_ctc_ratio_d = params->tsc_ctc_ratio_d;
 245
 246         if (!decoder->tsc_ctc_ratio_n)
 247                 decoder->tsc_ctc_ratio_d = 0;
 248
 249         if (decoder->tsc_ctc_ratio_d) {
 250                 if (!(decoder->tsc_ctc_ratio_n % decoder->tsc_ctc_ratio_d))
 251                         decoder->tsc_ctc_mult = decoder->tsc_ctc_ratio_n /
 252                                                 decoder->tsc_ctc_ratio_d;
 253
 254                 /*
 255                  * Allow for timestamps appearing to backwards because a TSC
 256                  * packet has slipped past a MTC packet, so allow 2 MTC ticks
 257                  * or ...
 258                  */
 259                 decoder->tsc_slip = multdiv(2 << decoder->mtc_shift,
 260                                         decoder->tsc_ctc_ratio_n,
 261                                         decoder->tsc_ctc_ratio_d);
 262         }
 263         /* ... or 0x100 paranoia */
 264         if (decoder->tsc_slip < 0x100)
 265                 decoder->tsc_slip = 0x100;
 266
 267         intel_pt_log("timestamp: mtc_shift %u\n", decoder->mtc_shift);
 268         intel_pt_log("timestamp: tsc_ctc_ratio_n %u\n", decoder->tsc_ctc_ratio_n);
 269         intel_pt_log("timestamp: tsc_ctc_ratio_d %u\n", decoder->tsc_ctc_ratio_d);
 270         intel_pt_log("timestamp: tsc_ctc_mult %u\n", decoder->tsc_ctc_mult);
 271         intel_pt_log("timestamp: tsc_slip %#x\n", decoder->tsc_slip);
 272
 273         return decoder;
 274 }
 275
 276 static void intel_pt_pop_blk(struct intel_pt_stack *stack)
 277 {
 278         struct intel_pt_blk *blk = stack->blk;
 279
 280         stack->blk = blk->prev;
 281         if (!stack->spare)
 282                 stack->spare = blk;
 283         else
 284                 free(blk);
 285 }
 286
 287 static uint64_t intel_pt_pop(struct intel_pt_stack *stack)
 288 {
 289         if (!stack->pos) {
 290                 if (!stack->blk)
 291                         return 0;
 292                 intel_pt_pop_blk(stack);
 293                 if (!stack->blk)
 294                         return 0;
 295                 stack->pos = INTEL_PT_BLK_SIZE;
 296         }
 297         return stack->blk->ip[--stack->pos];
 298 }
 299
 300 static int intel_pt_alloc_blk(struct intel_pt_stack *stack)
 301 {
 302         struct intel_pt_blk *blk;
 303
 304         if (stack->spare) {
 305                 blk = stack->spare;
 306                 stack->spare = NULL;
 307         } else {
 308                 blk = malloc(sizeof(struct intel_pt_blk));
 309                 if (!blk)
 310                         return -ENOMEM;
 311         }
 312
 313         blk->prev = stack->blk;
 314         stack->blk = blk;
 315         stack->pos = 0;
 316         return 0;
 317 }
 318
 319 static int intel_pt_push(struct intel_pt_stack *stack, uint64_t ip)
 320 {
 321         int err;
 322
 323         if (!stack->blk || stack->pos == INTEL_PT_BLK_SIZE) {
 324                 err = intel_pt_alloc_blk(stack);
 325                 if (err)
 326                         return err;
 327         }
 328
 329         stack->blk->ip[stack->pos++] = ip;
 330         return 0;
 331 }
 332
 333 static void intel_pt_clear_stack(struct intel_pt_stack *stack)
 334 {
 335         while (stack->blk)
 336                 intel_pt_pop_blk(stack);
 337         stack->pos = 0;
 338 }
 339
 340 static void intel_pt_free_stack(struct intel_pt_stack *stack)
 341 {
 342         intel_pt_clear_stack(stack);
 343         zfree(&stack->blk);
 344         zfree(&stack->spare);
 345 }
 346
 347 void intel_pt_decoder_free(struct intel_pt_decoder *decoder)
 348 {
 349         intel_pt_free_stack(&decoder->stack);
 350         free(decoder);
 351 }
 352
 353 static int intel_pt_ext_err(int code)
 354 {
 355         switch (code) {
 356         case -ENOMEM:
 357                 return INTEL_PT_ERR_NOMEM;
 358         case -ENOSYS:
 359                 return INTEL_PT_ERR_INTERN;
 360         case -EBADMSG:
 361                 return INTEL_PT_ERR_BADPKT;
 362         case -ENODATA:
 363                 return INTEL_PT_ERR_NODATA;
 364         case -EILSEQ:
 365                 return INTEL_PT_ERR_NOINSN;
 366         case -ENOENT:
 367                 return INTEL_PT_ERR_MISMAT;
 368         case -EOVERFLOW:
 369                 return INTEL_PT_ERR_OVR;
 370         case -ENOSPC:
 371                 return INTEL_PT_ERR_LOST;
 372         case -ELOOP:
 373                 return INTEL_PT_ERR_NELOOP;
 374         default:
 375                 return INTEL_PT_ERR_UNK;
 376         }
 377 }
 378
 379 static const char *intel_pt_err_msgs[] = {
 380         [INTEL_PT_ERR_NOMEM]  = "Memory allocation failed",
 381         [INTEL_PT_ERR_INTERN] = "Internal error",
 382         [INTEL_PT_ERR_BADPKT] = "Bad packet",
 383         [INTEL_PT_ERR_NODATA] = "No more data",
 384         [INTEL_PT_ERR_NOINSN] = "Failed to get instruction",
 385         [INTEL_PT_ERR_MISMAT] = "Trace doesn't match instruction",
 386         [INTEL_PT_ERR_OVR]    = "Overflow packet",
 387         [INTEL_PT_ERR_LOST]   = "Lost trace data",
 388         [INTEL_PT_ERR_UNK]    = "Unknown error!",
 389         [INTEL_PT_ERR_NELOOP] = "Never-ending loop",
 390 };
 391
 392 int intel_pt__strerror(int code, char *buf, size_t buflen)
 393 {
 394         if (code < 1 || code >= INTEL_PT_ERR_MAX)
 395                 code = INTEL_PT_ERR_UNK;
 396         strlcpy(buf, intel_pt_err_msgs[code], buflen);
 397         return 0;
 398 }
 399
 400 static uint64_t intel_pt_calc_ip(const struct intel_pt_pkt *packet,
 401                                  uint64_t last_ip)
 402 {
 403         uint64_t ip;
 404
 405         switch (packet->count) {
 406         case 1:
 407                 ip = (last_ip & (uint64_t)0xffffffffffff0000ULL) |
 408                      packet->payload;
 409                 break;
 410         case 2:
 411                 ip = (last_ip & (uint64_t)0xffffffff00000000ULL) |
 412                      packet->payload;
 413                 break;
 414         case 3:
 415                 ip = packet->payload;
 416                 /* Sign-extend 6-byte ip */
 417                 if (ip & (uint64_t)0x800000000000ULL)
 418                         ip |= (uint64_t)0xffff000000000000ULL;
 419                 break;
 420         case 4:
 421                 ip = (last_ip & (uint64_t)0xffff000000000000ULL) |
 422                      packet->payload;
 423                 break;
 424         case 6:
 425                 ip = packet->payload;
 426                 break;
 427         default:
 428                 return 0;
 429         }
 430
 431         return ip;
 432 }
 433
 434 static inline void intel_pt_set_last_ip(struct intel_pt_decoder *decoder)
 435 {
 436         decoder->last_ip = intel_pt_calc_ip(&decoder->packet, decoder->last_ip);
 437         decoder->have_last_ip = true;
 438 }
 439
 440 static inline void intel_pt_set_ip(struct intel_pt_decoder *decoder)
 441 {
 442         intel_pt_set_last_ip(decoder);
 443         decoder->ip = decoder->last_ip;
 444 }
 445
 446 static void intel_pt_decoder_log_packet(struct intel_pt_decoder *decoder)
 447 {
 448         intel_pt_log_packet(&decoder->packet, decoder->pkt_len, decoder->pos,
 449                             decoder->buf);
 450 }
 451
 452 static int intel_pt_bug(struct intel_pt_decoder *decoder)
 453 {
 454         intel_pt_log("ERROR: Internal error\n");
 455         decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
 456         return -ENOSYS;
 457 }
 458
 459 static inline void intel_pt_clear_tx_flags(struct intel_pt_decoder *decoder)
 460 {
 461         decoder->tx_flags = 0;
 462 }
 463
 464 static inline void intel_pt_update_in_tx(struct intel_pt_decoder *decoder)
 465 {
 466         decoder->tx_flags = decoder->packet.payload & INTEL_PT_IN_TX;
 467 }
 468
 469 static int intel_pt_bad_packet(struct intel_pt_decoder *decoder)
 470 {
 471         intel_pt_clear_tx_flags(decoder);
 472         decoder->have_tma = false;
 473         decoder->pkt_len = 1;
 474         decoder->pkt_step = 1;
 475         intel_pt_decoder_log_packet(decoder);
 476         if (decoder->pkt_state != INTEL_PT_STATE_NO_PSB) {
 477                 intel_pt_log("ERROR: Bad packet\n");
 478                 decoder->pkt_state = INTEL_PT_STATE_ERR1;
 479         }
 480         return -EBADMSG;
 481 }
 482
 483 static int intel_pt_get_data(struct intel_pt_decoder *decoder)
 484 {
 485         struct intel_pt_buffer buffer = { .buf = 0, };
 486         int ret;
 487
 488         decoder->pkt_step = 0;
 489
 490         intel_pt_log("Getting more data\n");
 491         ret = decoder->get_trace(&buffer, decoder->data);
 492         if (ret)
 493                 return ret;
 494         decoder->buf = buffer.buf;
 495         decoder->len = buffer.len;
 496         if (!decoder->len) {
 497                 intel_pt_log("No more data\n");
 498                 return -ENODATA;
 499         }
 500         if (!buffer.consecutive) {
 501                 decoder->ip = 0;
 502                 decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
 503                 decoder->ref_timestamp = buffer.ref_timestamp;
 504                 decoder->timestamp = 0;
 505                 decoder->have_tma = false;
 506                 decoder->state.trace_nr = buffer.trace_nr;
 507                 intel_pt_log("Reference timestamp 0x%" PRIx64 "\n",
 508                              decoder->ref_timestamp);
 509                 return -ENOLINK;
 510         }
 511
 512         return 0;
 513 }
 514
 515 static int intel_pt_get_next_data(struct intel_pt_decoder *decoder)
 516 {
 517         if (!decoder->next_buf)
 518                 return intel_pt_get_data(decoder);
 519
 520         decoder->buf = decoder->next_buf;
 521         decoder->len = decoder->next_len;
 522         decoder->next_buf = 0;
 523         decoder->next_len = 0;
 524         return 0;
 525 }
 526
 527 static int intel_pt_get_split_packet(struct intel_pt_decoder *decoder)
 528 {
 529         unsigned char *buf = decoder->temp_buf;
 530         size_t old_len, len, n;
 531         int ret;
 532
 533         old_len = decoder->len;
 534         len = decoder->len;
 535         memcpy(buf, decoder->buf, len);
 536
 537         ret = intel_pt_get_data(decoder);
 538         if (ret) {
 539                 decoder->pos += old_len;
 540                 return ret < 0 ? ret : -EINVAL;
 541         }
 542
 543         n = INTEL_PT_PKT_MAX_SZ - len;
 544         if (n > decoder->len)
 545                 n = decoder->len;
 546         memcpy(buf + len, decoder->buf, n);
 547         len += n;
 548
 549         ret = intel_pt_get_packet(buf, len, &decoder->packet);
 550         if (ret < (int)old_len) {
 551                 decoder->next_buf = decoder->buf;
 552                 decoder->next_len = decoder->len;
 553                 decoder->buf = buf;
 554                 decoder->len = old_len;
 555                 return intel_pt_bad_packet(decoder);
 556         }
 557
 558         decoder->next_buf = decoder->buf + (ret - old_len);
 559         decoder->next_len = decoder->len - (ret - old_len);
 560
 561         decoder->buf = buf;
 562         decoder->len = ret;
 563
 564         return ret;
 565 }
 566
 567 struct intel_pt_pkt_info {
 568         struct intel_pt_decoder   *decoder;
 569         struct intel_pt_pkt       packet;
 570         uint64_t                  pos;
 571         int                       pkt_len;
 572         int                       last_packet_type;
 573         void                      *data;
 574 };
 575
 576 typedef int (*intel_pt_pkt_cb_t)(struct intel_pt_pkt_info *pkt_info);
 577
 578 /* Lookahead packets in current buffer */
 579 static int intel_pt_pkt_lookahead(struct intel_pt_decoder *decoder,
 580                                   intel_pt_pkt_cb_t cb, void *data)
 581 {
 582         struct intel_pt_pkt_info pkt_info;
 583         const unsigned char *buf = decoder->buf;
 584         size_t len = decoder->len;
 585         int ret;
 586
 587         pkt_info.decoder          = decoder;
 588         pkt_info.pos              = decoder->pos;
 589         pkt_info.pkt_len          = decoder->pkt_step;
 590         pkt_info.last_packet_type = decoder->last_packet_type;
 591         pkt_info.data             = data;
 592
 593         while (1) {
 594                 do {
 595                         pkt_info.pos += pkt_info.pkt_len;
 596                         buf          += pkt_info.pkt_len;
 597                         len          -= pkt_info.pkt_len;
 598
 599                         if (!len)
 600                                 return INTEL_PT_NEED_MORE_BYTES;
 601
 602                         ret = intel_pt_get_packet(buf, len, &pkt_info.packet);
 603                         if (!ret)
 604                                 return INTEL_PT_NEED_MORE_BYTES;
 605                         if (ret < 0)
 606                                 return ret;
 607
 608                         pkt_info.pkt_len = ret;
 609                 } while (pkt_info.packet.type == INTEL_PT_PAD);
 610
 611                 ret = cb(&pkt_info);
 612                 if (ret)
 613                         return 0;
 614
 615                 pkt_info.last_packet_type = pkt_info.packet.type;
 616         }
 617 }
 618
 619 struct intel_pt_calc_cyc_to_tsc_info {
 620         uint64_t        cycle_cnt;
 621         unsigned int    cbr;
 622         uint32_t        last_mtc;
 623         uint64_t        ctc_timestamp;
 624         uint64_t        ctc_delta;
 625         uint64_t        tsc_timestamp;
 626         uint64_t        timestamp;
 627         bool            have_tma;
 628         bool            fixup_last_mtc;
 629         bool            from_mtc;
 630         double          cbr_cyc_to_tsc;
 631 };
 632
 633 /*
 634  * MTC provides a 8-bit slice of CTC but the TMA packet only provides the lower
 635  * 16 bits of CTC. If mtc_shift > 8 then some of the MTC bits are not in the CTC
 636  * provided by the TMA packet. Fix-up the last_mtc calculated from the TMA
 637  * packet by copying the missing bits from the current MTC assuming the least
 638  * difference between the two, and that the current MTC comes after last_mtc.
 639  */
 640 static void intel_pt_fixup_last_mtc(uint32_t mtc, int mtc_shift,
 641                                     uint32_t *last_mtc)
 642 {
 643         uint32_t first_missing_bit = 1U << (16 - mtc_shift);
 644         uint32_t mask = ~(first_missing_bit - 1);
 645
 646         *last_mtc |= mtc & mask;
 647         if (*last_mtc >= mtc) {
 648                 *last_mtc -= first_missing_bit;
 649                 *last_mtc &= 0xff;
 650         }
 651 }
 652
 653 static int intel_pt_calc_cyc_cb(struct intel_pt_pkt_info *pkt_info)
 654 {
 655         struct intel_pt_decoder *decoder = pkt_info->decoder;
 656         struct intel_pt_calc_cyc_to_tsc_info *data = pkt_info->data;
 657         uint64_t timestamp;
 658         double cyc_to_tsc;
 659         unsigned int cbr;
 660         uint32_t mtc, mtc_delta, ctc, fc, ctc_rem;
 661
 662         switch (pkt_info->packet.type) {
 663         case INTEL_PT_TNT:
 664         case INTEL_PT_TIP_PGE:
 665         case INTEL_PT_TIP:
 666         case INTEL_PT_FUP:
 667         case INTEL_PT_PSB:
 668         case INTEL_PT_PIP:
 669         case INTEL_PT_MODE_EXEC:
 670         case INTEL_PT_MODE_TSX:
 671         case INTEL_PT_PSBEND:
 672         case INTEL_PT_PAD:
 673         case INTEL_PT_VMCS:
 674         case INTEL_PT_MNT:
 675         case INTEL_PT_PTWRITE:
 676         case INTEL_PT_PTWRITE_IP:
 677                 return 0;
 678
 679         case INTEL_PT_MTC:
 680                 if (!data->have_tma)
 681                         return 0;
 682
 683                 mtc = pkt_info->packet.payload;
 684                 if (decoder->mtc_shift > 8 && data->fixup_last_mtc) {
 685                         data->fixup_last_mtc = false;
 686                         intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
 687                                                 &data->last_mtc);
 688                 }
 689                 if (mtc > data->last_mtc)
 690                         mtc_delta = mtc - data->last_mtc;
 691                 else
 692                         mtc_delta = mtc + 256 - data->last_mtc;
 693                 data->ctc_delta += mtc_delta << decoder->mtc_shift;
 694                 data->last_mtc = mtc;
 695
 696                 if (decoder->tsc_ctc_mult) {
 697                         timestamp = data->ctc_timestamp +
 698                                 data->ctc_delta * decoder->tsc_ctc_mult;
 699                 } else {
 700                         timestamp = data->ctc_timestamp +
 701                                 multdiv(data->ctc_delta,
 702                                         decoder->tsc_ctc_ratio_n,
 703                                         decoder->tsc_ctc_ratio_d);
 704                 }
 705
 706                 if (timestamp < data->timestamp)
 707                         return 1;
 708
 709                 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
 710                         data->timestamp = timestamp;
 711                         return 0;
 712                 }
 713
 714                 break;
 715
 716         case INTEL_PT_TSC:
 717                 /*
 718                  * For now, do not support using TSC packets - refer
 719                  * intel_pt_calc_cyc_to_tsc().
 720                  */
 721                 if (data->from_mtc)
 722                         return 1;
 723                 timestamp = pkt_info->packet.payload |
 724                             (data->timestamp & (0xffULL << 56));
 725                 if (data->from_mtc && timestamp < data->timestamp &&
 726                     data->timestamp - timestamp < decoder->tsc_slip)
 727                         return 1;
 728                 if (timestamp < data->timestamp)
 729                         timestamp += (1ULL << 56);
 730                 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
 731                         if (data->from_mtc)
 732                                 return 1;
 733                         data->tsc_timestamp = timestamp;
 734                         data->timestamp = timestamp;
 735                         return 0;
 736                 }
 737                 break;
 738
 739         case INTEL_PT_TMA:
 740                 if (data->from_mtc)
 741                         return 1;
 742
 743                 if (!decoder->tsc_ctc_ratio_d)
 744                         return 0;
 745
 746                 ctc = pkt_info->packet.payload;
 747                 fc = pkt_info->packet.count;
 748                 ctc_rem = ctc & decoder->ctc_rem_mask;
 749
 750                 data->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
 751
 752                 data->ctc_timestamp = data->tsc_timestamp - fc;
 753                 if (decoder->tsc_ctc_mult) {
 754                         data->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
 755                 } else {
 756                         data->ctc_timestamp -=
 757                                 multdiv(ctc_rem, decoder->tsc_ctc_ratio_n,
 758                                         decoder->tsc_ctc_ratio_d);
 759                 }
 760
 761                 data->ctc_delta = 0;
 762                 data->have_tma = true;
 763                 data->fixup_last_mtc = true;
 764
 765                 return 0;
 766
 767         case INTEL_PT_CYC:
 768                 data->cycle_cnt += pkt_info->packet.payload;
 769                 return 0;
 770
 771         case INTEL_PT_CBR:
 772                 cbr = pkt_info->packet.payload;
 773                 if (data->cbr && data->cbr != cbr)
 774                         return 1;
 775                 data->cbr = cbr;
 776                 data->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
 777                 return 0;
 778
 779         case INTEL_PT_TIP_PGD:
 780         case INTEL_PT_TRACESTOP:
 781         case INTEL_PT_EXSTOP:
 782         case INTEL_PT_EXSTOP_IP:
 783         case INTEL_PT_MWAIT:
 784         case INTEL_PT_PWRE:
 785         case INTEL_PT_PWRX:
 786         case INTEL_PT_OVF:
 787         case INTEL_PT_BAD: /* Does not happen */
 788         default:
 789                 return 1;
 790         }
 791
 792         if (!data->cbr && decoder->cbr) {
 793                 data->cbr = decoder->cbr;
 794                 data->cbr_cyc_to_tsc = decoder->cbr_cyc_to_tsc;
 795         }
 796
 797         if (!data->cycle_cnt)
 798                 return 1;
 799
 800         cyc_to_tsc = (double)(timestamp - decoder->timestamp) / data->cycle_cnt;
 801
 802         if (data->cbr && cyc_to_tsc > data->cbr_cyc_to_tsc &&
 803             cyc_to_tsc / data->cbr_cyc_to_tsc > 1.25) {
 804                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle too big (c.f. CBR-based value %g), pos " x64_fmt "\n",
 805                              cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
 806                 return 1;
 807         }
 808
 809         decoder->calc_cyc_to_tsc = cyc_to_tsc;
 810         decoder->have_calc_cyc_to_tsc = true;
 811
 812         if (data->cbr) {
 813                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. CBR-based value %g, pos " x64_fmt "\n",
 814                              cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
 815         } else {
 816                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. unknown CBR-based value, pos " x64_fmt "\n",
 817                              cyc_to_tsc, pkt_info->pos);
 818         }
 819
 820         return 1;
 821 }
 822
 823 static void intel_pt_calc_cyc_to_tsc(struct intel_pt_decoder *decoder,
 824                                      bool from_mtc)
 825 {
 826         struct intel_pt_calc_cyc_to_tsc_info data = {
 827                 .cycle_cnt      = 0,
 828                 .cbr            = 0,
 829                 .last_mtc       = decoder->last_mtc,
 830                 .ctc_timestamp  = decoder->ctc_timestamp,
 831                 .ctc_delta      = decoder->ctc_delta,
 832                 .tsc_timestamp  = decoder->tsc_timestamp,
 833                 .timestamp      = decoder->timestamp,
 834                 .have_tma       = decoder->have_tma,
 835                 .fixup_last_mtc = decoder->fixup_last_mtc,
 836                 .from_mtc       = from_mtc,
 837                 .cbr_cyc_to_tsc = 0,
 838         };
 839
 840         /*
 841          * For now, do not support using TSC packets for at least the reasons:
 842          * 1) timing might have stopped
 843          * 2) TSC packets within PSB+ can slip against CYC packets
 844          */
 845         if (!from_mtc)
 846                 return;
 847
 848         intel_pt_pkt_lookahead(decoder, intel_pt_calc_cyc_cb, &data);
 849 }
 850
 851 static int intel_pt_get_next_packet(struct intel_pt_decoder *decoder)
 852 {
 853         int ret;
 854
 855         decoder->last_packet_type = decoder->packet.type;
 856
 857         do {
 858                 decoder->pos += decoder->pkt_step;
 859                 decoder->buf += decoder->pkt_step;
 860                 decoder->len -= decoder->pkt_step;
 861
 862                 if (!decoder->len) {
 863                         ret = intel_pt_get_next_data(decoder);
 864                         if (ret)
 865                                 return ret;
 866                 }
 867
 868                 ret = intel_pt_get_packet(decoder->buf, decoder->len,
 869                                           &decoder->packet);
 870                 if (ret == INTEL_PT_NEED_MORE_BYTES &&
 871                     decoder->len < INTEL_PT_PKT_MAX_SZ && !decoder->next_buf) {
 872                         ret = intel_pt_get_split_packet(decoder);
 873                         if (ret < 0)
 874                                 return ret;
 875                 }
 876                 if (ret <= 0)
 877                         return intel_pt_bad_packet(decoder);
 878
 879                 decoder->pkt_len = ret;
 880                 decoder->pkt_step = ret;
 881                 intel_pt_decoder_log_packet(decoder);
 882         } while (decoder->packet.type == INTEL_PT_PAD);
 883
 884         return 0;
 885 }
 886
 887 static uint64_t intel_pt_next_period(struct intel_pt_decoder *decoder)
 888 {
 889         uint64_t timestamp, masked_timestamp;
 890
 891         timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
 892         masked_timestamp = timestamp & decoder->period_mask;
 893         if (decoder->continuous_period) {
 894                 if (masked_timestamp != decoder->last_masked_timestamp)
 895                         return 1;
 896         } else {
 897                 timestamp += 1;
 898                 masked_timestamp = timestamp & decoder->period_mask;
 899                 if (masked_timestamp != decoder->last_masked_timestamp) {
 900                         decoder->last_masked_timestamp = masked_timestamp;
 901                         decoder->continuous_period = true;
 902                 }
 903         }
 904         return decoder->period_ticks - (timestamp - masked_timestamp);
 905 }
 906
 907 static uint64_t intel_pt_next_sample(struct intel_pt_decoder *decoder)
 908 {
 909         switch (decoder->period_type) {
 910         case INTEL_PT_PERIOD_INSTRUCTIONS:
 911                 return decoder->period - decoder->period_insn_cnt;
 912         case INTEL_PT_PERIOD_TICKS:
 913                 return intel_pt_next_period(decoder);
 914         case INTEL_PT_PERIOD_NONE:
 915         case INTEL_PT_PERIOD_MTC:
 916         default:
 917                 return 0;
 918         }
 919 }
 920
 921 static void intel_pt_sample_insn(struct intel_pt_decoder *decoder)
 922 {
 923         uint64_t timestamp, masked_timestamp;
 924
 925         switch (decoder->period_type) {
 926         case INTEL_PT_PERIOD_INSTRUCTIONS:
 927                 decoder->period_insn_cnt = 0;
 928                 break;
 929         case INTEL_PT_PERIOD_TICKS:
 930                 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
 931                 masked_timestamp = timestamp & decoder->period_mask;
 932                 decoder->last_masked_timestamp = masked_timestamp;
 933                 break;
 934         case INTEL_PT_PERIOD_NONE:
 935         case INTEL_PT_PERIOD_MTC:
 936         default:
 937                 break;
 938         }
 939
 940         decoder->state.type |= INTEL_PT_INSTRUCTION;
 941 }
 942
 943 static int intel_pt_walk_insn(struct intel_pt_decoder *decoder,
 944                               struct intel_pt_insn *intel_pt_insn, uint64_t ip)
 945 {
 946         uint64_t max_insn_cnt, insn_cnt = 0;
 947         int err;
 948
 949         if (!decoder->mtc_insn)
 950                 decoder->mtc_insn = true;
 951
 952         max_insn_cnt = intel_pt_next_sample(decoder);
 953
 954         err = decoder->walk_insn(intel_pt_insn, &insn_cnt, &decoder->ip, ip,
 955                                  max_insn_cnt, decoder->data);
 956
 957         decoder->tot_insn_cnt += insn_cnt;
 958         decoder->timestamp_insn_cnt += insn_cnt;
 959         decoder->sample_insn_cnt += insn_cnt;
 960         decoder->period_insn_cnt += insn_cnt;
 961
 962         if (err) {
 963                 decoder->no_progress = 0;
 964                 decoder->pkt_state = INTEL_PT_STATE_ERR2;
 965                 intel_pt_log_at("ERROR: Failed to get instruction",
 966                                 decoder->ip);
 967                 if (err == -ENOENT)
 968                         return -ENOLINK;
 969                 return -EILSEQ;
 970         }
 971
 972         if (ip && decoder->ip == ip) {
 973                 err = -EAGAIN;
 974                 goto out;
 975         }
 976
 977         if (max_insn_cnt && insn_cnt >= max_insn_cnt)
 978                 intel_pt_sample_insn(decoder);
 979
 980         if (intel_pt_insn->branch == INTEL_PT_BR_NO_BRANCH) {
 981                 decoder->state.type = INTEL_PT_INSTRUCTION;
 982                 decoder->state.from_ip = decoder->ip;
 983                 decoder->state.to_ip = 0;
 984                 decoder->ip += intel_pt_insn->length;
 985                 err = INTEL_PT_RETURN;
 986                 goto out;
 987         }
 988
 989         if (intel_pt_insn->op == INTEL_PT_OP_CALL) {
 990                 /* Zero-length calls are excluded */
 991                 if (intel_pt_insn->branch != INTEL_PT_BR_UNCONDITIONAL ||
 992                     intel_pt_insn->rel) {
 993                         err = intel_pt_push(&decoder->stack, decoder->ip +
 994                                             intel_pt_insn->length);
 995                         if (err)
 996                                 goto out;
 997                 }
 998         } else if (intel_pt_insn->op == INTEL_PT_OP_RET) {
 999                 decoder->ret_addr = intel_pt_pop(&decoder->stack);
1000         }
1001
1002         if (intel_pt_insn->branch == INTEL_PT_BR_UNCONDITIONAL) {
1003                 int cnt = decoder->no_progress++;
1004
1005                 decoder->state.from_ip = decoder->ip;
1006                 decoder->ip += intel_pt_insn->length +
1007                                 intel_pt_insn->rel;
1008                 decoder->state.to_ip = decoder->ip;
1009                 err = INTEL_PT_RETURN;
1010
1011                 /*
1012                  * Check for being stuck in a loop.  This can happen if a
1013                  * decoder error results in the decoder erroneously setting the
1014                  * ip to an address that is itself in an infinite loop that
1015                  * consumes no packets.  When that happens, there must be an
1016                  * unconditional branch.
1017                  */
1018                 if (cnt) {
1019                         if (cnt == 1) {
1020                                 decoder->stuck_ip = decoder->state.to_ip;
1021                                 decoder->stuck_ip_prd = 1;
1022                                 decoder->stuck_ip_cnt = 1;
1023                         } else if (cnt > INTEL_PT_MAX_LOOPS ||
1024                                    decoder->state.to_ip == decoder->stuck_ip) {
1025                                 intel_pt_log_at("ERROR: Never-ending loop",
1026                                                 decoder->state.to_ip);
1027                                 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1028                                 err = -ELOOP;
1029                                 goto out;
1030                         } else if (!--decoder->stuck_ip_cnt) {
1031                                 decoder->stuck_ip_prd += 1;
1032                                 decoder->stuck_ip_cnt = decoder->stuck_ip_prd;
1033                                 decoder->stuck_ip = decoder->state.to_ip;
1034                         }
1035                 }
1036                 goto out_no_progress;
1037         }
1038 out:
1039         decoder->no_progress = 0;
1040 out_no_progress:
1041         decoder->state.insn_op = intel_pt_insn->op;
1042         decoder->state.insn_len = intel_pt_insn->length;
1043         memcpy(decoder->state.insn, intel_pt_insn->buf,
1044                INTEL_PT_INSN_BUF_SZ);
1045
1046         if (decoder->tx_flags & INTEL_PT_IN_TX)
1047                 decoder->state.flags |= INTEL_PT_IN_TX;
1048
1049         return err;
1050 }
1051
1052 static bool intel_pt_fup_event(struct intel_pt_decoder *decoder)
1053 {
1054         bool ret = false;
1055
1056         if (decoder->set_fup_tx_flags) {
1057                 decoder->set_fup_tx_flags = false;
1058                 decoder->tx_flags = decoder->fup_tx_flags;
1059                 decoder->state.type = INTEL_PT_TRANSACTION;
1060                 decoder->state.from_ip = decoder->ip;
1061                 decoder->state.to_ip = 0;
1062                 decoder->state.flags = decoder->fup_tx_flags;
1063                 return true;
1064         }
1065         if (decoder->set_fup_ptw) {
1066                 decoder->set_fup_ptw = false;
1067                 decoder->state.type = INTEL_PT_PTW;
1068                 decoder->state.flags |= INTEL_PT_FUP_IP;
1069                 decoder->state.from_ip = decoder->ip;
1070                 decoder->state.to_ip = 0;
1071                 decoder->state.ptw_payload = decoder->fup_ptw_payload;
1072                 return true;
1073         }
1074         if (decoder->set_fup_mwait) {
1075                 decoder->set_fup_mwait = false;
1076                 decoder->state.type = INTEL_PT_MWAIT_OP;
1077                 decoder->state.from_ip = decoder->ip;
1078                 decoder->state.to_ip = 0;
1079                 decoder->state.mwait_payload = decoder->fup_mwait_payload;
1080                 ret = true;
1081         }
1082         if (decoder->set_fup_pwre) {
1083                 decoder->set_fup_pwre = false;
1084                 decoder->state.type |= INTEL_PT_PWR_ENTRY;
1085                 decoder->state.type &= ~INTEL_PT_BRANCH;
1086                 decoder->state.from_ip = decoder->ip;
1087                 decoder->state.to_ip = 0;
1088                 decoder->state.pwre_payload = decoder->fup_pwre_payload;
1089                 ret = true;
1090         }
1091         if (decoder->set_fup_exstop) {
1092                 decoder->set_fup_exstop = false;
1093                 decoder->state.type |= INTEL_PT_EX_STOP;
1094                 decoder->state.type &= ~INTEL_PT_BRANCH;
1095                 decoder->state.flags |= INTEL_PT_FUP_IP;
1096                 decoder->state.from_ip = decoder->ip;
1097                 decoder->state.to_ip = 0;
1098                 ret = true;
1099         }
1100         return ret;
1101 }
1102
1103 static inline bool intel_pt_fup_with_nlip(struct intel_pt_decoder *decoder,
1104                                           struct intel_pt_insn *intel_pt_insn,
1105                                           uint64_t ip, int err)
1106 {
1107         return decoder->flags & INTEL_PT_FUP_WITH_NLIP && !err &&
1108                intel_pt_insn->branch == INTEL_PT_BR_INDIRECT &&
1109                ip == decoder->ip + intel_pt_insn->length;
1110 }
1111
1112 static int intel_pt_walk_fup(struct intel_pt_decoder *decoder)
1113 {
1114         struct intel_pt_insn intel_pt_insn;
1115         uint64_t ip;
1116         int err;
1117
1118         ip = decoder->last_ip;
1119
1120         while (1) {
1121                 err = intel_pt_walk_insn(decoder, &intel_pt_insn, ip);
1122                 if (err == INTEL_PT_RETURN)
1123                         return 0;
1124                 if (err == -EAGAIN ||
1125                     intel_pt_fup_with_nlip(decoder, &intel_pt_insn, ip, err)) {
1126                         if (intel_pt_fup_event(decoder))
1127                                 return 0;
1128                         return -EAGAIN;
1129                 }
1130                 decoder->set_fup_tx_flags = false;
1131                 if (err)
1132                         return err;
1133
1134                 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1135                         intel_pt_log_at("ERROR: Unexpected indirect branch",
1136                                         decoder->ip);
1137                         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1138                         return -ENOENT;
1139                 }
1140
1141                 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1142                         intel_pt_log_at("ERROR: Unexpected conditional branch",
1143                                         decoder->ip);
1144                         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1145                         return -ENOENT;
1146                 }
1147
1148                 intel_pt_bug(decoder);
1149         }
1150 }
1151
1152 static int intel_pt_walk_tip(struct intel_pt_decoder *decoder)
1153 {
1154         struct intel_pt_insn intel_pt_insn;
1155         int err;
1156
1157         err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1158         if (err == INTEL_PT_RETURN &&
1159             decoder->pgd_ip &&
1160             decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1161             (decoder->state.type & INTEL_PT_BRANCH) &&
1162             decoder->pgd_ip(decoder->state.to_ip, decoder->data)) {
1163                 /* Unconditional branch leaving filter region */
1164                 decoder->no_progress = 0;
1165                 decoder->pge = false;
1166                 decoder->continuous_period = false;
1167                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1168                 decoder->state.to_ip = 0;
1169                 return 0;
1170         }
1171         if (err == INTEL_PT_RETURN)
1172                 return 0;
1173         if (err)
1174                 return err;
1175
1176         if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1177                 if (decoder->pkt_state == INTEL_PT_STATE_TIP_PGD) {
1178                         decoder->pge = false;
1179                         decoder->continuous_period = false;
1180                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1181                         decoder->state.from_ip = decoder->ip;
1182                         decoder->state.to_ip = 0;
1183                         if (decoder->packet.count != 0)
1184                                 decoder->ip = decoder->last_ip;
1185                 } else {
1186                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1187                         decoder->state.from_ip = decoder->ip;
1188                         if (decoder->packet.count == 0) {
1189                                 decoder->state.to_ip = 0;
1190                         } else {
1191                                 decoder->state.to_ip = decoder->last_ip;
1192                                 decoder->ip = decoder->last_ip;
1193                         }
1194                 }
1195                 return 0;
1196         }
1197
1198         if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1199                 uint64_t to_ip = decoder->ip + intel_pt_insn.length +
1200                                  intel_pt_insn.rel;
1201
1202                 if (decoder->pgd_ip &&
1203                     decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1204                     decoder->pgd_ip(to_ip, decoder->data)) {
1205                         /* Conditional branch leaving filter region */
1206                         decoder->pge = false;
1207                         decoder->continuous_period = false;
1208                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1209                         decoder->ip = to_ip;
1210                         decoder->state.from_ip = decoder->ip;
1211                         decoder->state.to_ip = 0;
1212                         return 0;
1213                 }
1214                 intel_pt_log_at("ERROR: Conditional branch when expecting indirect branch",
1215                                 decoder->ip);
1216                 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1217                 return -ENOENT;
1218         }
1219
1220         return intel_pt_bug(decoder);
1221 }
1222
1223 static int intel_pt_walk_tnt(struct intel_pt_decoder *decoder)
1224 {
1225         struct intel_pt_insn intel_pt_insn;
1226         int err;
1227
1228         while (1) {
1229                 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1230                 if (err == INTEL_PT_RETURN)
1231                         return 0;
1232                 if (err)
1233                         return err;
1234
1235                 if (intel_pt_insn.op == INTEL_PT_OP_RET) {
1236                         if (!decoder->return_compression) {
1237                                 intel_pt_log_at("ERROR: RET when expecting conditional branch",
1238                                                 decoder->ip);
1239                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1240                                 return -ENOENT;
1241                         }
1242                         if (!decoder->ret_addr) {
1243                                 intel_pt_log_at("ERROR: Bad RET compression (stack empty)",
1244                                                 decoder->ip);
1245                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1246                                 return -ENOENT;
1247                         }
1248                         if (!(decoder->tnt.payload & BIT63)) {
1249                                 intel_pt_log_at("ERROR: Bad RET compression (TNT=N)",
1250                                                 decoder->ip);
1251                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1252                                 return -ENOENT;
1253                         }
1254                         decoder->tnt.count -= 1;
1255                         if (!decoder->tnt.count)
1256                                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1257                         decoder->tnt.payload <<= 1;
1258                         decoder->state.from_ip = decoder->ip;
1259                         decoder->ip = decoder->ret_addr;
1260                         decoder->state.to_ip = decoder->ip;
1261                         return 0;
1262                 }
1263
1264                 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1265                         /* Handle deferred TIPs */
1266                         err = intel_pt_get_next_packet(decoder);
1267                         if (err)
1268                                 return err;
1269                         if (decoder->packet.type != INTEL_PT_TIP ||
1270                             decoder->packet.count == 0) {
1271                                 intel_pt_log_at("ERROR: Missing deferred TIP for indirect branch",
1272                                                 decoder->ip);
1273                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1274                                 decoder->pkt_step = 0;
1275                                 return -ENOENT;
1276                         }
1277                         intel_pt_set_last_ip(decoder);
1278                         decoder->state.from_ip = decoder->ip;
1279                         decoder->state.to_ip = decoder->last_ip;
1280                         decoder->ip = decoder->last_ip;
1281                         return 0;
1282                 }
1283
1284                 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1285                         decoder->tnt.count -= 1;
1286                         if (!decoder->tnt.count)
1287                                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1288                         if (decoder->tnt.payload & BIT63) {
1289                                 decoder->tnt.payload <<= 1;
1290                                 decoder->state.from_ip = decoder->ip;
1291                                 decoder->ip += intel_pt_insn.length +
1292                                                intel_pt_insn.rel;
1293                                 decoder->state.to_ip = decoder->ip;
1294                                 return 0;
1295                         }
1296                         /* Instruction sample for a non-taken branch */
1297                         if (decoder->state.type & INTEL_PT_INSTRUCTION) {
1298                                 decoder->tnt.payload <<= 1;
1299                                 decoder->state.type = INTEL_PT_INSTRUCTION;
1300                                 decoder->state.from_ip = decoder->ip;
1301                                 decoder->state.to_ip = 0;
1302                                 decoder->ip += intel_pt_insn.length;
1303                                 return 0;
1304                         }
1305                         decoder->ip += intel_pt_insn.length;
1306                         if (!decoder->tnt.count)
1307                                 return -EAGAIN;
1308                         decoder->tnt.payload <<= 1;
1309                         continue;
1310                 }
1311
1312                 return intel_pt_bug(decoder);
1313         }
1314 }
1315
1316 static int intel_pt_mode_tsx(struct intel_pt_decoder *decoder, bool *no_tip)
1317 {
1318         unsigned int fup_tx_flags;
1319         int err;
1320
1321         fup_tx_flags = decoder->packet.payload &
1322                        (INTEL_PT_IN_TX | INTEL_PT_ABORT_TX);
1323         err = intel_pt_get_next_packet(decoder);
1324         if (err)
1325                 return err;
1326         if (decoder->packet.type == INTEL_PT_FUP) {
1327                 decoder->fup_tx_flags = fup_tx_flags;
1328                 decoder->set_fup_tx_flags = true;
1329                 if (!(decoder->fup_tx_flags & INTEL_PT_ABORT_TX))
1330                         *no_tip = true;
1331         } else {
1332                 intel_pt_log_at("ERROR: Missing FUP after MODE.TSX",
1333                                 decoder->pos);
1334                 intel_pt_update_in_tx(decoder);
1335         }
1336         return 0;
1337 }
1338
1339 static void intel_pt_calc_tsc_timestamp(struct intel_pt_decoder *decoder)
1340 {
1341         uint64_t timestamp;
1342
1343         decoder->have_tma = false;
1344
1345         if (decoder->ref_timestamp) {
1346                 timestamp = decoder->packet.payload |
1347                             (decoder->ref_timestamp & (0xffULL << 56));
1348                 if (timestamp < decoder->ref_timestamp) {
1349                         if (decoder->ref_timestamp - timestamp > (1ULL << 55))
1350                                 timestamp += (1ULL << 56);
1351                 } else {
1352                         if (timestamp - decoder->ref_timestamp > (1ULL << 55))
1353                                 timestamp -= (1ULL << 56);
1354                 }
1355                 decoder->tsc_timestamp = timestamp;
1356                 decoder->timestamp = timestamp;
1357                 decoder->ref_timestamp = 0;
1358                 decoder->timestamp_insn_cnt = 0;
1359         } else if (decoder->timestamp) {
1360                 timestamp = decoder->packet.payload |
1361                             (decoder->timestamp & (0xffULL << 56));
1362                 decoder->tsc_timestamp = timestamp;
1363                 if (timestamp < decoder->timestamp &&
1364                     decoder->timestamp - timestamp < decoder->tsc_slip) {
1365                         intel_pt_log_to("Suppressing backwards timestamp",
1366                                         timestamp);
1367                         timestamp = decoder->timestamp;
1368                 }
1369                 if (timestamp < decoder->timestamp) {
1370                         intel_pt_log_to("Wraparound timestamp", timestamp);
1371                         timestamp += (1ULL << 56);
1372                         decoder->tsc_timestamp = timestamp;
1373                 }
1374                 decoder->timestamp = timestamp;
1375                 decoder->timestamp_insn_cnt = 0;
1376         }
1377
1378         if (decoder->last_packet_type == INTEL_PT_CYC) {
1379                 decoder->cyc_ref_timestamp = decoder->timestamp;
1380                 decoder->cycle_cnt = 0;
1381                 decoder->have_calc_cyc_to_tsc = false;
1382                 intel_pt_calc_cyc_to_tsc(decoder, false);
1383         }
1384
1385         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1386 }
1387
1388 static int intel_pt_overflow(struct intel_pt_decoder *decoder)
1389 {
1390         intel_pt_log("ERROR: Buffer overflow\n");
1391         intel_pt_clear_tx_flags(decoder);
1392         decoder->cbr = 0;
1393         decoder->timestamp_insn_cnt = 0;
1394         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1395         decoder->overflow = true;
1396         return -EOVERFLOW;
1397 }
1398
1399 static void intel_pt_calc_tma(struct intel_pt_decoder *decoder)
1400 {
1401         uint32_t ctc = decoder->packet.payload;
1402         uint32_t fc = decoder->packet.count;
1403         uint32_t ctc_rem = ctc & decoder->ctc_rem_mask;
1404
1405         if (!decoder->tsc_ctc_ratio_d)
1406                 return;
1407
1408         decoder->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
1409         decoder->ctc_timestamp = decoder->tsc_timestamp - fc;
1410         if (decoder->tsc_ctc_mult) {
1411                 decoder->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
1412         } else {
1413                 decoder->ctc_timestamp -= multdiv(ctc_rem,
1414                                                   decoder->tsc_ctc_ratio_n,
1415                                                   decoder->tsc_ctc_ratio_d);
1416         }
1417         decoder->ctc_delta = 0;
1418         decoder->have_tma = true;
1419         decoder->fixup_last_mtc = true;
1420         intel_pt_log("CTC timestamp " x64_fmt " last MTC %#x  CTC rem %#x\n",
1421                      decoder->ctc_timestamp, decoder->last_mtc, ctc_rem);
1422 }
1423
1424 static void intel_pt_calc_mtc_timestamp(struct intel_pt_decoder *decoder)
1425 {
1426         uint64_t timestamp;
1427         uint32_t mtc, mtc_delta;
1428
1429         if (!decoder->have_tma)
1430                 return;
1431
1432         mtc = decoder->packet.payload;
1433
1434         if (decoder->mtc_shift > 8 && decoder->fixup_last_mtc) {
1435                 decoder->fixup_last_mtc = false;
1436                 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
1437                                         &decoder->last_mtc);
1438         }
1439
1440         if (mtc > decoder->last_mtc)
1441                 mtc_delta = mtc - decoder->last_mtc;
1442         else
1443                 mtc_delta = mtc + 256 - decoder->last_mtc;
1444
1445         decoder->ctc_delta += mtc_delta << decoder->mtc_shift;
1446
1447         if (decoder->tsc_ctc_mult) {
1448                 timestamp = decoder->ctc_timestamp +
1449                             decoder->ctc_delta * decoder->tsc_ctc_mult;
1450         } else {
1451                 timestamp = decoder->ctc_timestamp +
1452                             multdiv(decoder->ctc_delta,
1453                                     decoder->tsc_ctc_ratio_n,
1454                                     decoder->tsc_ctc_ratio_d);
1455         }
1456
1457         if (timestamp < decoder->timestamp)
1458                 intel_pt_log("Suppressing MTC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1459                              timestamp, decoder->timestamp);
1460         else
1461                 decoder->timestamp = timestamp;
1462
1463         decoder->timestamp_insn_cnt = 0;
1464         decoder->last_mtc = mtc;
1465
1466         if (decoder->last_packet_type == INTEL_PT_CYC) {
1467                 decoder->cyc_ref_timestamp = decoder->timestamp;
1468                 decoder->cycle_cnt = 0;
1469                 decoder->have_calc_cyc_to_tsc = false;
1470                 intel_pt_calc_cyc_to_tsc(decoder, true);
1471         }
1472 }
1473
1474 static void intel_pt_calc_cbr(struct intel_pt_decoder *decoder)
1475 {
1476         unsigned int cbr = decoder->packet.payload & 0xff;
1477
1478         decoder->cbr_payload = decoder->packet.payload;
1479
1480         if (decoder->cbr == cbr)
1481                 return;
1482
1483         decoder->cbr = cbr;
1484         decoder->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
1485 }
1486
1487 static void intel_pt_calc_cyc_timestamp(struct intel_pt_decoder *decoder)
1488 {
1489         uint64_t timestamp = decoder->cyc_ref_timestamp;
1490
1491         decoder->have_cyc = true;
1492
1493         decoder->cycle_cnt += decoder->packet.payload;
1494
1495         if (!decoder->cyc_ref_timestamp)
1496                 return;
1497
1498         if (decoder->have_calc_cyc_to_tsc)
1499                 timestamp += decoder->cycle_cnt * decoder->calc_cyc_to_tsc;
1500         else if (decoder->cbr)
1501                 timestamp += decoder->cycle_cnt * decoder->cbr_cyc_to_tsc;
1502         else
1503                 return;
1504
1505         if (timestamp < decoder->timestamp)
1506                 intel_pt_log("Suppressing CYC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1507                              timestamp, decoder->timestamp);
1508         else
1509                 decoder->timestamp = timestamp;
1510
1511         decoder->timestamp_insn_cnt = 0;
1512 }
1513
1514 /* Walk PSB+ packets when already in sync. */
1515 static int intel_pt_walk_psbend(struct intel_pt_decoder *decoder)
1516 {
1517         int err;
1518
1519         while (1) {
1520                 err = intel_pt_get_next_packet(decoder);
1521                 if (err)
1522                         return err;
1523
1524                 switch (decoder->packet.type) {
1525                 case INTEL_PT_PSBEND:
1526                         return 0;
1527
1528                 case INTEL_PT_TIP_PGD:
1529                 case INTEL_PT_TIP_PGE:
1530                 case INTEL_PT_TIP:
1531                 case INTEL_PT_TNT:
1532                 case INTEL_PT_TRACESTOP:
1533                 case INTEL_PT_BAD:
1534                 case INTEL_PT_PSB:
1535                 case INTEL_PT_PTWRITE:
1536                 case INTEL_PT_PTWRITE_IP:
1537                 case INTEL_PT_EXSTOP:
1538                 case INTEL_PT_EXSTOP_IP:
1539                 case INTEL_PT_MWAIT:
1540                 case INTEL_PT_PWRE:
1541                 case INTEL_PT_PWRX:
1542                         decoder->have_tma = false;
1543                         intel_pt_log("ERROR: Unexpected packet\n");
1544                         return -EAGAIN;
1545
1546                 case INTEL_PT_OVF:
1547                         return intel_pt_overflow(decoder);
1548
1549                 case INTEL_PT_TSC:
1550                         intel_pt_calc_tsc_timestamp(decoder);
1551                         break;
1552
1553                 case INTEL_PT_TMA:
1554                         intel_pt_calc_tma(decoder);
1555                         break;
1556
1557                 case INTEL_PT_CBR:
1558                         intel_pt_calc_cbr(decoder);
1559                         break;
1560
1561                 case INTEL_PT_MODE_EXEC:
1562                         decoder->exec_mode = decoder->packet.payload;
1563                         break;
1564
1565                 case INTEL_PT_PIP:
1566                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1567                         break;
1568
1569                 case INTEL_PT_FUP:
1570                         decoder->pge = true;
1571                         if (decoder->packet.count)
1572                                 intel_pt_set_last_ip(decoder);
1573                         break;
1574
1575                 case INTEL_PT_MODE_TSX:
1576                         intel_pt_update_in_tx(decoder);
1577                         break;
1578
1579                 case INTEL_PT_MTC:
1580                         intel_pt_calc_mtc_timestamp(decoder);
1581                         if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1582                                 decoder->state.type |= INTEL_PT_INSTRUCTION;
1583                         break;
1584
1585                 case INTEL_PT_CYC:
1586                 case INTEL_PT_VMCS:
1587                 case INTEL_PT_MNT:
1588                 case INTEL_PT_PAD:
1589                 default:
1590                         break;
1591                 }
1592         }
1593 }
1594
1595 static int intel_pt_walk_fup_tip(struct intel_pt_decoder *decoder)
1596 {
1597         int err;
1598
1599         if (decoder->tx_flags & INTEL_PT_ABORT_TX) {
1600                 decoder->tx_flags = 0;
1601                 decoder->state.flags &= ~INTEL_PT_IN_TX;
1602                 decoder->state.flags |= INTEL_PT_ABORT_TX;
1603         } else {
1604                 decoder->state.flags |= INTEL_PT_ASYNC;
1605         }
1606
1607         while (1) {
1608                 err = intel_pt_get_next_packet(decoder);
1609                 if (err)
1610                         return err;
1611
1612                 switch (decoder->packet.type) {
1613                 case INTEL_PT_TNT:
1614                 case INTEL_PT_FUP:
1615                 case INTEL_PT_TRACESTOP:
1616                 case INTEL_PT_PSB:
1617                 case INTEL_PT_TSC:
1618                 case INTEL_PT_TMA:
1619                 case INTEL_PT_MODE_TSX:
1620                 case INTEL_PT_BAD:
1621                 case INTEL_PT_PSBEND:
1622                 case INTEL_PT_PTWRITE:
1623                 case INTEL_PT_PTWRITE_IP:
1624                 case INTEL_PT_EXSTOP:
1625                 case INTEL_PT_EXSTOP_IP:
1626                 case INTEL_PT_MWAIT:
1627                 case INTEL_PT_PWRE:
1628                 case INTEL_PT_PWRX:
1629                         intel_pt_log("ERROR: Missing TIP after FUP\n");
1630                         decoder->pkt_state = INTEL_PT_STATE_ERR3;
1631                         decoder->pkt_step = 0;
1632                         return -ENOENT;
1633
1634                 case INTEL_PT_CBR:
1635                         intel_pt_calc_cbr(decoder);
1636                         break;
1637
1638                 case INTEL_PT_OVF:
1639                         return intel_pt_overflow(decoder);
1640
1641                 case INTEL_PT_TIP_PGD:
1642                         decoder->state.from_ip = decoder->ip;
1643                         decoder->state.to_ip = 0;
1644                         if (decoder->packet.count != 0) {
1645                                 intel_pt_set_ip(decoder);
1646                                 intel_pt_log("Omitting PGD ip " x64_fmt "\n",
1647                                              decoder->ip);
1648                         }
1649                         decoder->pge = false;
1650                         decoder->continuous_period = false;
1651                         return 0;
1652
1653                 case INTEL_PT_TIP_PGE:
1654                         decoder->pge = true;
1655                         intel_pt_log("Omitting PGE ip " x64_fmt "\n",
1656                                      decoder->ip);
1657                         decoder->state.from_ip = 0;
1658                         if (decoder->packet.count == 0) {
1659                                 decoder->state.to_ip = 0;
1660                         } else {
1661                                 intel_pt_set_ip(decoder);
1662                                 decoder->state.to_ip = decoder->ip;
1663                         }
1664                         return 0;
1665
1666                 case INTEL_PT_TIP:
1667                         decoder->state.from_ip = decoder->ip;
1668                         if (decoder->packet.count == 0) {
1669                                 decoder->state.to_ip = 0;
1670                         } else {
1671                                 intel_pt_set_ip(decoder);
1672                                 decoder->state.to_ip = decoder->ip;
1673                         }
1674                         return 0;
1675
1676                 case INTEL_PT_PIP:
1677                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1678                         break;
1679
1680                 case INTEL_PT_MTC:
1681                         intel_pt_calc_mtc_timestamp(decoder);
1682                         if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1683                                 decoder->state.type |= INTEL_PT_INSTRUCTION;
1684                         break;
1685
1686                 case INTEL_PT_CYC:
1687                         intel_pt_calc_cyc_timestamp(decoder);
1688                         break;
1689
1690                 case INTEL_PT_MODE_EXEC:
1691                         decoder->exec_mode = decoder->packet.payload;
1692                         break;
1693
1694                 case INTEL_PT_VMCS:
1695                 case INTEL_PT_MNT:
1696                 case INTEL_PT_PAD:
1697                         break;
1698
1699                 default:
1700                         return intel_pt_bug(decoder);
1701                 }
1702         }
1703 }
1704
1705 static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)
1706 {
1707         bool no_tip = false;
1708         int err;
1709
1710         while (1) {
1711                 err = intel_pt_get_next_packet(decoder);
1712                 if (err)
1713                         return err;
1714 next:
1715                 switch (decoder->packet.type) {
1716                 case INTEL_PT_TNT:
1717                         if (!decoder->packet.count)
1718                                 break;
1719                         decoder->tnt = decoder->packet;
1720                         decoder->pkt_state = INTEL_PT_STATE_TNT;
1721                         err = intel_pt_walk_tnt(decoder);
1722                         if (err == -EAGAIN)
1723                                 break;
1724                         return err;
1725
1726                 case INTEL_PT_TIP_PGD:
1727                         if (decoder->packet.count != 0)
1728                                 intel_pt_set_last_ip(decoder);
1729                         decoder->pkt_state = INTEL_PT_STATE_TIP_PGD;
1730                         return intel_pt_walk_tip(decoder);
1731
1732                 case INTEL_PT_TIP_PGE: {
1733                         decoder->pge = true;
1734                         if (decoder->packet.count == 0) {
1735                                 intel_pt_log_at("Skipping zero TIP.PGE",
1736                                                 decoder->pos);
1737                                 break;
1738                         }
1739                         intel_pt_set_ip(decoder);
1740                         decoder->state.from_ip = 0;
1741                         decoder->state.to_ip = decoder->ip;
1742                         return 0;
1743                 }
1744
1745                 case INTEL_PT_OVF:
1746                         return intel_pt_overflow(decoder);
1747
1748                 case INTEL_PT_TIP:
1749                         if (decoder->packet.count != 0)
1750                                 intel_pt_set_last_ip(decoder);
1751                         decoder->pkt_state = INTEL_PT_STATE_TIP;
1752                         return intel_pt_walk_tip(decoder);
1753
1754                 case INTEL_PT_FUP:
1755                         if (decoder->packet.count == 0) {
1756                                 intel_pt_log_at("Skipping zero FUP",
1757                                                 decoder->pos);
1758                                 no_tip = false;
1759                                 break;
1760                         }
1761                         intel_pt_set_last_ip(decoder);
1762                         if (!decoder->branch_enable) {
1763                                 decoder->ip = decoder->last_ip;
1764                                 if (intel_pt_fup_event(decoder))
1765                                         return 0;
1766                                 no_tip = false;
1767                                 break;
1768                         }
1769                         if (decoder->set_fup_mwait)
1770                                 no_tip = true;
1771                         err = intel_pt_walk_fup(decoder);
1772                         if (err != -EAGAIN) {
1773                                 if (err)
1774                                         return err;
1775                                 if (no_tip)
1776                                         decoder->pkt_state =
1777                                                 INTEL_PT_STATE_FUP_NO_TIP;
1778                                 else
1779                                         decoder->pkt_state = INTEL_PT_STATE_FUP;
1780                                 return 0;
1781                         }
1782                         if (no_tip) {
1783                                 no_tip = false;
1784                                 break;
1785                         }
1786                         return intel_pt_walk_fup_tip(decoder);
1787
1788                 case INTEL_PT_TRACESTOP:
1789                         decoder->pge = false;
1790                         decoder->continuous_period = false;
1791                         intel_pt_clear_tx_flags(decoder);
1792                         decoder->have_tma = false;
1793                         break;
1794
1795                 case INTEL_PT_PSB:
1796                         decoder->last_ip = 0;
1797                         decoder->have_last_ip = true;
1798                         intel_pt_clear_stack(&decoder->stack);
1799                         err = intel_pt_walk_psbend(decoder);
1800                         if (err == -EAGAIN)
1801                                 goto next;
1802                         if (err)
1803                                 return err;
1804                         break;
1805
1806                 case INTEL_PT_PIP:
1807                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1808                         break;
1809
1810                 case INTEL_PT_MTC:
1811                         intel_pt_calc_mtc_timestamp(decoder);
1812                         if (decoder->period_type != INTEL_PT_PERIOD_MTC)
1813                                 break;
1814                         /*
1815                          * Ensure that there has been an instruction since the
1816                          * last MTC.
1817                          */
1818                         if (!decoder->mtc_insn)
1819                                 break;
1820                         decoder->mtc_insn = false;
1821                         /* Ensure that there is a timestamp */
1822                         if (!decoder->timestamp)
1823                                 break;
1824                         decoder->state.type = INTEL_PT_INSTRUCTION;
1825                         decoder->state.from_ip = decoder->ip;
1826                         decoder->state.to_ip = 0;
1827                         decoder->mtc_insn = false;
1828                         return 0;
1829
1830                 case INTEL_PT_TSC:
1831                         intel_pt_calc_tsc_timestamp(decoder);
1832                         break;
1833
1834                 case INTEL_PT_TMA:
1835                         intel_pt_calc_tma(decoder);
1836                         break;
1837
1838                 case INTEL_PT_CYC:
1839                         intel_pt_calc_cyc_timestamp(decoder);
1840                         break;
1841
1842                 case INTEL_PT_CBR:
1843                         intel_pt_calc_cbr(decoder);
1844                         if (!decoder->branch_enable &&
1845                             decoder->cbr != decoder->cbr_seen) {
1846                                 decoder->cbr_seen = decoder->cbr;
1847                                 decoder->state.type = INTEL_PT_CBR_CHG;
1848                                 decoder->state.from_ip = decoder->ip;
1849                                 decoder->state.to_ip = 0;
1850                                 decoder->state.cbr_payload =
1851                                                         decoder->packet.payload;
1852                                 return 0;
1853                         }
1854                         break;
1855
1856                 case INTEL_PT_MODE_EXEC:
1857                         decoder->exec_mode = decoder->packet.payload;
1858                         break;
1859
1860                 case INTEL_PT_MODE_TSX:
1861                         /* MODE_TSX need not be followed by FUP */
1862                         if (!decoder->pge) {
1863                                 intel_pt_update_in_tx(decoder);
1864                                 break;
1865                         }
1866                         err = intel_pt_mode_tsx(decoder, &no_tip);
1867                         if (err)
1868                                 return err;
1869                         goto next;
1870
1871                 case INTEL_PT_BAD: /* Does not happen */
1872                         return intel_pt_bug(decoder);
1873
1874                 case INTEL_PT_PSBEND:
1875                 case INTEL_PT_VMCS:
1876                 case INTEL_PT_MNT:
1877                 case INTEL_PT_PAD:
1878                         break;
1879
1880                 case INTEL_PT_PTWRITE_IP:
1881                         decoder->fup_ptw_payload = decoder->packet.payload;
1882                         err = intel_pt_get_next_packet(decoder);
1883                         if (err)
1884                                 return err;
1885                         if (decoder->packet.type == INTEL_PT_FUP) {
1886                                 decoder->set_fup_ptw = true;
1887                                 no_tip = true;
1888                         } else {
1889                                 intel_pt_log_at("ERROR: Missing FUP after PTWRITE",
1890                                                 decoder->pos);
1891                         }
1892                         goto next;
1893
1894                 case INTEL_PT_PTWRITE:
1895                         decoder->state.type = INTEL_PT_PTW;
1896                         decoder->state.from_ip = decoder->ip;
1897                         decoder->state.to_ip = 0;
1898                         decoder->state.ptw_payload = decoder->packet.payload;
1899                         return 0;
1900
1901                 case INTEL_PT_MWAIT:
1902                         decoder->fup_mwait_payload = decoder->packet.payload;
1903                         decoder->set_fup_mwait = true;
1904                         break;
1905
1906                 case INTEL_PT_PWRE:
1907                         if (decoder->set_fup_mwait) {
1908                                 decoder->fup_pwre_payload =
1909                                                         decoder->packet.payload;
1910                                 decoder->set_fup_pwre = true;
1911                                 break;
1912                         }
1913                         decoder->state.type = INTEL_PT_PWR_ENTRY;
1914                         decoder->state.from_ip = decoder->ip;
1915                         decoder->state.to_ip = 0;
1916                         decoder->state.pwrx_payload = decoder->packet.payload;
1917                         return 0;
1918
1919                 case INTEL_PT_EXSTOP_IP:
1920                         err = intel_pt_get_next_packet(decoder);
1921                         if (err)
1922                                 return err;
1923                         if (decoder->packet.type == INTEL_PT_FUP) {
1924                                 decoder->set_fup_exstop = true;
1925                                 no_tip = true;
1926                         } else {
1927                                 intel_pt_log_at("ERROR: Missing FUP after EXSTOP",
1928                                                 decoder->pos);
1929                         }
1930                         goto next;
1931
1932                 case INTEL_PT_EXSTOP:
1933                         decoder->state.type = INTEL_PT_EX_STOP;
1934                         decoder->state.from_ip = decoder->ip;
1935                         decoder->state.to_ip = 0;
1936                         return 0;
1937
1938                 case INTEL_PT_PWRX:
1939                         decoder->state.type = INTEL_PT_PWR_EXIT;
1940                         decoder->state.from_ip = decoder->ip;
1941                         decoder->state.to_ip = 0;
1942                         decoder->state.pwrx_payload = decoder->packet.payload;
1943                         return 0;
1944
1945                 default:
1946                         return intel_pt_bug(decoder);
1947                 }
1948         }
1949 }
1950
1951 static inline bool intel_pt_have_ip(struct intel_pt_decoder *decoder)
1952 {
1953         return decoder->packet.count &&
1954                (decoder->have_last_ip || decoder->packet.count == 3 ||
1955                 decoder->packet.count == 6);
1956 }
1957
1958 /* Walk PSB+ packets to get in sync. */
1959 static int intel_pt_walk_psb(struct intel_pt_decoder *decoder)
1960 {
1961         int err;
1962
1963         while (1) {
1964                 err = intel_pt_get_next_packet(decoder);
1965                 if (err)
1966                         return err;
1967
1968                 switch (decoder->packet.type) {
1969                 case INTEL_PT_TIP_PGD:
1970                         decoder->continuous_period = false;
1971                         __fallthrough;
1972                 case INTEL_PT_TIP_PGE:
1973                 case INTEL_PT_TIP:
1974                 case INTEL_PT_PTWRITE:
1975                 case INTEL_PT_PTWRITE_IP:
1976                 case INTEL_PT_EXSTOP:
1977                 case INTEL_PT_EXSTOP_IP:
1978                 case INTEL_PT_MWAIT:
1979                 case INTEL_PT_PWRE:
1980                 case INTEL_PT_PWRX:
1981                         intel_pt_log("ERROR: Unexpected packet\n");
1982                         return -ENOENT;
1983
1984                 case INTEL_PT_FUP:
1985                         decoder->pge = true;
1986                         if (intel_pt_have_ip(decoder)) {
1987                                 uint64_t current_ip = decoder->ip;
1988
1989                                 intel_pt_set_ip(decoder);
1990                                 if (current_ip)
1991                                         intel_pt_log_to("Setting IP",
1992                                                         decoder->ip);
1993                         }
1994                         break;
1995
1996                 case INTEL_PT_MTC:
1997                         intel_pt_calc_mtc_timestamp(decoder);
1998                         break;
1999
2000                 case INTEL_PT_TSC:
2001                         intel_pt_calc_tsc_timestamp(decoder);
2002                         break;
2003
2004                 case INTEL_PT_TMA:
2005                         intel_pt_calc_tma(decoder);
2006                         break;
2007
2008                 case INTEL_PT_CYC:
2009                         intel_pt_calc_cyc_timestamp(decoder);
2010                         break;
2011
2012                 case INTEL_PT_CBR:
2013                         intel_pt_calc_cbr(decoder);
2014                         break;
2015
2016                 case INTEL_PT_PIP:
2017                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2018                         break;
2019
2020                 case INTEL_PT_MODE_EXEC:
2021                         decoder->exec_mode = decoder->packet.payload;
2022                         break;
2023
2024                 case INTEL_PT_MODE_TSX:
2025                         intel_pt_update_in_tx(decoder);
2026                         break;
2027
2028                 case INTEL_PT_TRACESTOP:
2029                         decoder->pge = false;
2030                         decoder->continuous_period = false;
2031                         intel_pt_clear_tx_flags(decoder);
2032                         __fallthrough;
2033
2034                 case INTEL_PT_TNT:
2035                         decoder->have_tma = false;
2036                         intel_pt_log("ERROR: Unexpected packet\n");
2037                         if (decoder->ip)
2038                                 decoder->pkt_state = INTEL_PT_STATE_ERR4;
2039                         else
2040                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
2041                         return -ENOENT;
2042
2043                 case INTEL_PT_BAD: /* Does not happen */
2044                         return intel_pt_bug(decoder);
2045
2046                 case INTEL_PT_OVF:
2047                         return intel_pt_overflow(decoder);
2048
2049                 case INTEL_PT_PSBEND:
2050                         return 0;
2051
2052                 case INTEL_PT_PSB:
2053                 case INTEL_PT_VMCS:
2054                 case INTEL_PT_MNT:
2055                 case INTEL_PT_PAD:
2056                 default:
2057                         break;
2058                 }
2059         }
2060 }
2061
2062 static int intel_pt_walk_to_ip(struct intel_pt_decoder *decoder)
2063 {
2064         int err;
2065
2066         while (1) {
2067                 err = intel_pt_get_next_packet(decoder);
2068                 if (err)
2069                         return err;
2070
2071                 switch (decoder->packet.type) {
2072                 case INTEL_PT_TIP_PGD:
2073                         decoder->continuous_period = false;
2074                         __fallthrough;
2075                 case INTEL_PT_TIP_PGE:
2076                 case INTEL_PT_TIP:
2077                         decoder->pge = decoder->packet.type != INTEL_PT_TIP_PGD;
2078                         if (intel_pt_have_ip(decoder))
2079                                 intel_pt_set_ip(decoder);
2080                         if (decoder->ip)
2081                                 return 0;
2082                         break;
2083
2084                 case INTEL_PT_FUP:
2085                         if (intel_pt_have_ip(decoder))
2086                                 intel_pt_set_ip(decoder);
2087                         if (decoder->ip)
2088                                 return 0;
2089                         break;
2090
2091                 case INTEL_PT_MTC:
2092                         intel_pt_calc_mtc_timestamp(decoder);
2093                         break;
2094
2095                 case INTEL_PT_TSC:
2096                         intel_pt_calc_tsc_timestamp(decoder);
2097                         break;
2098
2099                 case INTEL_PT_TMA:
2100                         intel_pt_calc_tma(decoder);
2101                         break;
2102
2103                 case INTEL_PT_CYC:
2104                         intel_pt_calc_cyc_timestamp(decoder);
2105                         break;
2106
2107                 case INTEL_PT_CBR:
2108                         intel_pt_calc_cbr(decoder);
2109                         break;
2110
2111                 case INTEL_PT_PIP:
2112                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2113                         break;
2114
2115                 case INTEL_PT_MODE_EXEC:
2116                         decoder->exec_mode = decoder->packet.payload;
2117                         break;
2118
2119                 case INTEL_PT_MODE_TSX:
2120                         intel_pt_update_in_tx(decoder);
2121                         break;
2122
2123                 case INTEL_PT_OVF:
2124                         return intel_pt_overflow(decoder);
2125
2126                 case INTEL_PT_BAD: /* Does not happen */
2127                         return intel_pt_bug(decoder);
2128
2129                 case INTEL_PT_TRACESTOP:
2130                         decoder->pge = false;
2131                         decoder->continuous_period = false;
2132                         intel_pt_clear_tx_flags(decoder);
2133                         decoder->have_tma = false;
2134                         break;
2135
2136                 case INTEL_PT_PSB:
2137                         decoder->last_ip = 0;
2138                         decoder->have_last_ip = true;
2139                         intel_pt_clear_stack(&decoder->stack);
2140                         err = intel_pt_walk_psb(decoder);
2141                         if (err)
2142                                 return err;
2143                         if (decoder->ip) {
2144                                 /* Do not have a sample */
2145                                 decoder->state.type = 0;
2146                                 return 0;
2147                         }
2148                         break;
2149
2150                 case INTEL_PT_TNT:
2151                 case INTEL_PT_PSBEND:
2152                 case INTEL_PT_VMCS:
2153                 case INTEL_PT_MNT:
2154                 case INTEL_PT_PAD:
2155                 case INTEL_PT_PTWRITE:
2156                 case INTEL_PT_PTWRITE_IP:
2157                 case INTEL_PT_EXSTOP:
2158                 case INTEL_PT_EXSTOP_IP:
2159                 case INTEL_PT_MWAIT:
2160                 case INTEL_PT_PWRE:
2161                 case INTEL_PT_PWRX:
2162                 default:
2163                         break;
2164                 }
2165         }
2166 }
2167
2168 static int intel_pt_sync_ip(struct intel_pt_decoder *decoder)
2169 {
2170         int err;
2171
2172         decoder->set_fup_tx_flags = false;
2173         decoder->set_fup_ptw = false;
2174         decoder->set_fup_mwait = false;
2175         decoder->set_fup_pwre = false;
2176         decoder->set_fup_exstop = false;
2177
2178         if (!decoder->branch_enable) {
2179                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2180                 decoder->overflow = false;
2181                 decoder->state.type = 0; /* Do not have a sample */
2182                 return 0;
2183         }
2184
2185         intel_pt_log("Scanning for full IP\n");
2186         err = intel_pt_walk_to_ip(decoder);
2187         if (err)
2188                 return err;
2189
2190         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2191         decoder->overflow = false;
2192
2193         decoder->state.from_ip = 0;
2194         decoder->state.to_ip = decoder->ip;
2195         intel_pt_log_to("Setting IP", decoder->ip);
2196
2197         return 0;
2198 }
2199
2200 static int intel_pt_part_psb(struct intel_pt_decoder *decoder)
2201 {
2202         const unsigned char *end = decoder->buf + decoder->len;
2203         size_t i;
2204
2205         for (i = INTEL_PT_PSB_LEN - 1; i; i--) {
2206                 if (i > decoder->len)
2207                         continue;
2208                 if (!memcmp(end - i, INTEL_PT_PSB_STR, i))
2209                         return i;
2210         }
2211         return 0;
2212 }
2213
2214 static int intel_pt_rest_psb(struct intel_pt_decoder *decoder, int part_psb)
2215 {
2216         size_t rest_psb = INTEL_PT_PSB_LEN - part_psb;
2217         const char *psb = INTEL_PT_PSB_STR;
2218
2219         if (rest_psb > decoder->len ||
2220             memcmp(decoder->buf, psb + part_psb, rest_psb))
2221                 return 0;
2222
2223         return rest_psb;
2224 }
2225
2226 static int intel_pt_get_split_psb(struct intel_pt_decoder *decoder,
2227                                   int part_psb)
2228 {
2229         int rest_psb, ret;
2230
2231         decoder->pos += decoder->len;
2232         decoder->len = 0;
2233
2234         ret = intel_pt_get_next_data(decoder);
2235         if (ret)
2236                 return ret;
2237
2238         rest_psb = intel_pt_rest_psb(decoder, part_psb);
2239         if (!rest_psb)
2240                 return 0;
2241
2242         decoder->pos -= part_psb;
2243         decoder->next_buf = decoder->buf + rest_psb;
2244         decoder->next_len = decoder->len - rest_psb;
2245         memcpy(decoder->temp_buf, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2246         decoder->buf = decoder->temp_buf;
2247         decoder->len = INTEL_PT_PSB_LEN;
2248
2249         return 0;
2250 }
2251
2252 static int intel_pt_scan_for_psb(struct intel_pt_decoder *decoder)
2253 {
2254         unsigned char *next;
2255         int ret;
2256
2257         intel_pt_log("Scanning for PSB\n");
2258         while (1) {
2259                 if (!decoder->len) {
2260                         ret = intel_pt_get_next_data(decoder);
2261                         if (ret)
2262                                 return ret;
2263                 }
2264
2265                 next = memmem(decoder->buf, decoder->len, INTEL_PT_PSB_STR,
2266                               INTEL_PT_PSB_LEN);
2267                 if (!next) {
2268                         int part_psb;
2269
2270                         part_psb = intel_pt_part_psb(decoder);
2271                         if (part_psb) {
2272                                 ret = intel_pt_get_split_psb(decoder, part_psb);
2273                                 if (ret)
2274                                         return ret;
2275                         } else {
2276                                 decoder->pos += decoder->len;
2277                                 decoder->len = 0;
2278                         }
2279                         continue;
2280                 }
2281
2282                 decoder->pkt_step = next - decoder->buf;
2283                 return intel_pt_get_next_packet(decoder);
2284         }
2285 }
2286
2287 static int intel_pt_sync(struct intel_pt_decoder *decoder)
2288 {
2289         int err;
2290
2291         decoder->pge = false;
2292         decoder->continuous_period = false;
2293         decoder->have_last_ip = false;
2294         decoder->last_ip = 0;
2295         decoder->ip = 0;
2296         intel_pt_clear_stack(&decoder->stack);
2297
2298         err = intel_pt_scan_for_psb(decoder);
2299         if (err)
2300                 return err;
2301
2302         decoder->have_last_ip = true;
2303         decoder->pkt_state = INTEL_PT_STATE_NO_IP;
2304
2305         err = intel_pt_walk_psb(decoder);
2306         if (err)
2307                 return err;
2308
2309         if (decoder->ip) {
2310                 decoder->state.type = 0; /* Do not have a sample */
2311                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2312         } else {
2313                 return intel_pt_sync_ip(decoder);
2314         }
2315
2316         return 0;
2317 }
2318
2319 static uint64_t intel_pt_est_timestamp(struct intel_pt_decoder *decoder)
2320 {
2321         uint64_t est = decoder->sample_insn_cnt << 1;
2322
2323         if (!decoder->cbr || !decoder->max_non_turbo_ratio)
2324                 goto out;
2325
2326         est *= decoder->max_non_turbo_ratio;
2327         est /= decoder->cbr;
2328 out:
2329         return decoder->sample_timestamp + est;
2330 }
2331
2332 const struct intel_pt_state *intel_pt_decode(struct intel_pt_decoder *decoder)
2333 {
2334         int err;
2335
2336         do {
2337                 decoder->state.type = INTEL_PT_BRANCH;
2338                 decoder->state.flags = 0;
2339
2340                 switch (decoder->pkt_state) {
2341                 case INTEL_PT_STATE_NO_PSB:
2342                         err = intel_pt_sync(decoder);
2343                         break;
2344                 case INTEL_PT_STATE_NO_IP:
2345                         decoder->have_last_ip = false;
2346                         decoder->last_ip = 0;
2347                         decoder->ip = 0;
2348                         __fallthrough;
2349                 case INTEL_PT_STATE_ERR_RESYNC:
2350                         err = intel_pt_sync_ip(decoder);
2351                         break;
2352                 case INTEL_PT_STATE_IN_SYNC:
2353                         err = intel_pt_walk_trace(decoder);
2354                         break;
2355                 case INTEL_PT_STATE_TNT:
2356                         err = intel_pt_walk_tnt(decoder);
2357                         if (err == -EAGAIN)
2358                                 err = intel_pt_walk_trace(decoder);
2359                         break;
2360                 case INTEL_PT_STATE_TIP:
2361                 case INTEL_PT_STATE_TIP_PGD:
2362                         err = intel_pt_walk_tip(decoder);
2363                         break;
2364                 case INTEL_PT_STATE_FUP:
2365                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2366                         err = intel_pt_walk_fup(decoder);
2367                         if (err == -EAGAIN)
2368                                 err = intel_pt_walk_fup_tip(decoder);
2369                         else if (!err)
2370                                 decoder->pkt_state = INTEL_PT_STATE_FUP;
2371                         break;
2372                 case INTEL_PT_STATE_FUP_NO_TIP:
2373                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2374                         err = intel_pt_walk_fup(decoder);
2375                         if (err == -EAGAIN)
2376                                 err = intel_pt_walk_trace(decoder);
2377                         break;
2378                 default:
2379                         err = intel_pt_bug(decoder);
2380                         break;
2381                 }
2382         } while (err == -ENOLINK);
2383
2384         if (err) {
2385                 decoder->state.err = intel_pt_ext_err(err);
2386                 decoder->state.from_ip = decoder->ip;
2387                 decoder->sample_timestamp = decoder->timestamp;
2388                 decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
2389         } else {
2390                 decoder->state.err = 0;
2391                 if (decoder->cbr != decoder->cbr_seen && decoder->state.type) {
2392                         decoder->cbr_seen = decoder->cbr;
2393                         decoder->state.type |= INTEL_PT_CBR_CHG;
2394                         decoder->state.cbr_payload = decoder->cbr_payload;
2395                 }
2396                 if (intel_pt_sample_time(decoder->pkt_state)) {
2397                         decoder->sample_timestamp = decoder->timestamp;
2398                         decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
2399                 }
2400         }
2401
2402         decoder->state.timestamp = decoder->sample_timestamp;
2403         decoder->state.est_timestamp = intel_pt_est_timestamp(decoder);
2404         decoder->state.cr3 = decoder->cr3;
2405         decoder->state.tot_insn_cnt = decoder->tot_insn_cnt;
2406
2407         return &decoder->state;
2408 }
2409
2410 /**
2411  * intel_pt_next_psb - move buffer pointer to the start of the next PSB packet.
2412  * @buf: pointer to buffer pointer
2413  * @len: size of buffer
2414  *
2415  * Updates the buffer pointer to point to the start of the next PSB packet if
2416  * there is one, otherwise the buffer pointer is unchanged.  If @buf is updated,
2417  * @len is adjusted accordingly.
2418  *
2419  * Return: %true if a PSB packet is found, %false otherwise.
2420  */
2421 static bool intel_pt_next_psb(unsigned char **buf, size_t *len)
2422 {
2423         unsigned char *next;
2424
2425         next = memmem(*buf, *len, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2426         if (next) {
2427                 *len -= next - *buf;
2428                 *buf = next;
2429                 return true;
2430         }
2431         return false;
2432 }
2433
2434 /**
2435  * intel_pt_step_psb - move buffer pointer to the start of the following PSB
2436  *                     packet.
2437  * @buf: pointer to buffer pointer
2438  * @len: size of buffer
2439  *
2440  * Updates the buffer pointer to point to the start of the following PSB packet
2441  * (skipping the PSB at @buf itself) if there is one, otherwise the buffer
2442  * pointer is unchanged.  If @buf is updated, @len is adjusted accordingly.
2443  *
2444  * Return: %true if a PSB packet is found, %false otherwise.
2445  */
2446 static bool intel_pt_step_psb(unsigned char **buf, size_t *len)
2447 {
2448         unsigned char *next;
2449
2450         if (!*len)
2451                 return false;
2452
2453         next = memmem(*buf + 1, *len - 1, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2454         if (next) {
2455                 *len -= next - *buf;
2456                 *buf = next;
2457                 return true;
2458         }
2459         return false;
2460 }
2461
2462 /**
2463  * intel_pt_last_psb - find the last PSB packet in a buffer.
2464  * @buf: buffer
2465  * @len: size of buffer
2466  *
2467  * This function finds the last PSB in a buffer.
2468  *
2469  * Return: A pointer to the last PSB in @buf if found, %NULL otherwise.
2470  */
2471 static unsigned char *intel_pt_last_psb(unsigned char *buf, size_t len)
2472 {
2473         const char *n = INTEL_PT_PSB_STR;
2474         unsigned char *p;
2475         size_t k;
2476
2477         if (len < INTEL_PT_PSB_LEN)
2478                 return NULL;
2479
2480         k = len - INTEL_PT_PSB_LEN + 1;
2481         while (1) {
2482                 p = memrchr(buf, n[0], k);
2483                 if (!p)
2484                         return NULL;
2485                 if (!memcmp(p + 1, n + 1, INTEL_PT_PSB_LEN - 1))
2486                         return p;
2487                 k = p - buf;
2488                 if (!k)
2489                         return NULL;
2490         }
2491 }
2492
2493 /**
2494  * intel_pt_next_tsc - find and return next TSC.
2495  * @buf: buffer
2496  * @len: size of buffer
2497  * @tsc: TSC value returned
2498  * @rem: returns remaining size when TSC is found
2499  *
2500  * Find a TSC packet in @buf and return the TSC value.  This function assumes
2501  * that @buf starts at a PSB and that PSB+ will contain TSC and so stops if a
2502  * PSBEND packet is found.
2503  *
2504  * Return: %true if TSC is found, false otherwise.
2505  */
2506 static bool intel_pt_next_tsc(unsigned char *buf, size_t len, uint64_t *tsc,
2507                               size_t *rem)
2508 {
2509         struct intel_pt_pkt packet;
2510         int ret;
2511
2512         while (len) {
2513                 ret = intel_pt_get_packet(buf, len, &packet);
2514                 if (ret <= 0)
2515                         return false;
2516                 if (packet.type == INTEL_PT_TSC) {
2517                         *tsc = packet.payload;
2518                         *rem = len;
2519                         return true;
2520                 }
2521                 if (packet.type == INTEL_PT_PSBEND)
2522                         return false;
2523                 buf += ret;
2524                 len -= ret;
2525         }
2526         return false;
2527 }
2528
2529 /**
2530  * intel_pt_tsc_cmp - compare 7-byte TSCs.
2531  * @tsc1: first TSC to compare
2532  * @tsc2: second TSC to compare
2533  *
2534  * This function compares 7-byte TSC values allowing for the possibility that
2535  * TSC wrapped around.  Generally it is not possible to know if TSC has wrapped
2536  * around so for that purpose this function assumes the absolute difference is
2537  * less than half the maximum difference.
2538  *
2539  * Return: %-1 if @tsc1 is before @tsc2, %0 if @tsc1 == @tsc2, %1 if @tsc1 is
2540  * after @tsc2.
2541  */
2542 static int intel_pt_tsc_cmp(uint64_t tsc1, uint64_t tsc2)
2543 {
2544         const uint64_t halfway = (1ULL << 55);
2545
2546         if (tsc1 == tsc2)
2547                 return 0;
2548
2549         if (tsc1 < tsc2) {
2550                 if (tsc2 - tsc1 < halfway)
2551                         return -1;
2552                 else
2553                         return 1;
2554         } else {
2555                 if (tsc1 - tsc2 < halfway)
2556                         return 1;
2557                 else
2558                         return -1;
2559         }
2560 }
2561
2562 /**
2563  * intel_pt_find_overlap_tsc - determine start of non-overlapped trace data
2564  *                             using TSC.
2565  * @buf_a: first buffer
2566  * @len_a: size of first buffer
2567  * @buf_b: second buffer
2568  * @len_b: size of second buffer
2569  * @consecutive: returns true if there is data in buf_b that is consecutive
2570  *               to buf_a
2571  *
2572  * If the trace contains TSC we can look at the last TSC of @buf_a and the
2573  * first TSC of @buf_b in order to determine if the buffers overlap, and then
2574  * walk forward in @buf_b until a later TSC is found.  A precondition is that
2575  * @buf_a and @buf_b are positioned at a PSB.
2576  *
2577  * Return: A pointer into @buf_b from where non-overlapped data starts, or
2578  * @buf_b + @len_b if there is no non-overlapped data.
2579  */
2580 static unsigned char *intel_pt_find_overlap_tsc(unsigned char *buf_a,
2581                                                 size_t len_a,
2582                                                 unsigned char *buf_b,
2583                                                 size_t len_b, bool *consecutive)
2584 {
2585         uint64_t tsc_a, tsc_b;
2586         unsigned char *p;
2587         size_t len, rem_a, rem_b;
2588
2589         p = intel_pt_last_psb(buf_a, len_a);
2590         if (!p)
2591                 return buf_b; /* No PSB in buf_a => no overlap */
2592
2593         len = len_a - (p - buf_a);
2594         if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a)) {
2595                 /* The last PSB+ in buf_a is incomplete, so go back one more */
2596                 len_a -= len;
2597                 p = intel_pt_last_psb(buf_a, len_a);
2598                 if (!p)
2599                         return buf_b; /* No full PSB+ => assume no overlap */
2600                 len = len_a - (p - buf_a);
2601                 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a))
2602                         return buf_b; /* No TSC in buf_a => assume no overlap */
2603         }
2604
2605         while (1) {
2606                 /* Ignore PSB+ with no TSC */
2607                 if (intel_pt_next_tsc(buf_b, len_b, &tsc_b, &rem_b)) {
2608                         int cmp = intel_pt_tsc_cmp(tsc_a, tsc_b);
2609
2610                         /* Same TSC, so buffers are consecutive */
2611                         if (!cmp && rem_b >= rem_a) {
2612                                 *consecutive = true;
2613                                 return buf_b + len_b - (rem_b - rem_a);
2614                         }
2615                         if (cmp < 0)
2616                                 return buf_b; /* tsc_a < tsc_b => no overlap */
2617                 }
2618
2619                 if (!intel_pt_step_psb(&buf_b, &len_b))
2620                         return buf_b + len_b; /* No PSB in buf_b => no data */
2621         }
2622 }
2623
2624 /**
2625  * intel_pt_find_overlap - determine start of non-overlapped trace data.
2626  * @buf_a: first buffer
2627  * @len_a: size of first buffer
2628  * @buf_b: second buffer
2629  * @len_b: size of second buffer
2630  * @have_tsc: can use TSC packets to detect overlap
2631  * @consecutive: returns true if there is data in buf_b that is consecutive
2632  *               to buf_a
2633  *
2634  * When trace samples or snapshots are recorded there is the possibility that
2635  * the data overlaps.  Note that, for the purposes of decoding, data is only
2636  * useful if it begins with a PSB packet.
2637  *
2638  * Return: A pointer into @buf_b from where non-overlapped data starts, or
2639  * @buf_b + @len_b if there is no non-overlapped data.
2640  */
2641 unsigned char *intel_pt_find_overlap(unsigned char *buf_a, size_t len_a,
2642                                      unsigned char *buf_b, size_t len_b,
2643                                      bool have_tsc, bool *consecutive)
2644 {
2645         unsigned char *found;
2646
2647         /* Buffer 'b' must start at PSB so throw away everything before that */
2648         if (!intel_pt_next_psb(&buf_b, &len_b))
2649                 return buf_b + len_b; /* No PSB */
2650
2651         if (!intel_pt_next_psb(&buf_a, &len_a))
2652                 return buf_b; /* No overlap */
2653
2654         if (have_tsc) {
2655                 found = intel_pt_find_overlap_tsc(buf_a, len_a, buf_b, len_b,
2656                                                   consecutive);
2657                 if (found)
2658                         return found;
2659         }
2660
2661         /*
2662          * Buffer 'b' cannot end within buffer 'a' so, for comparison purposes,
2663          * we can ignore the first part of buffer 'a'.
2664          */
2665         while (len_b < len_a) {
2666                 if (!intel_pt_step_psb(&buf_a, &len_a))
2667                         return buf_b; /* No overlap */
2668         }
2669
2670         /* Now len_b >= len_a */
2671         while (1) {
2672                 /* Potential overlap so check the bytes */
2673                 found = memmem(buf_a, len_a, buf_b, len_a);
2674                 if (found) {
2675                         *consecutive = true;
2676                         return buf_b + len_a;
2677                 }
2678
2679                 /* Try again at next PSB in buffer 'a' */
2680                 if (!intel_pt_step_psb(&buf_a, &len_a))
2681                         return buf_b; /* No overlap */
2682         }
2683 }