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

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * intel_pt_decoder.c: Intel Processor Trace support
   4  * Copyright (c) 2013-2014, Intel Corporation.
   5  */
   6
   7 #ifndef _GNU_SOURCE
   8 #define _GNU_SOURCE
   9 #endif
  10 #include <stdlib.h>
  11 #include <stdbool.h>
  12 #include <string.h>
  13 #include <errno.h>
  14 #include <stdint.h>
  15 #include <inttypes.h>
  16 #include <linux/compiler.h>
  17 #include <linux/zalloc.h>
  18
  19 #include "../cache.h"
  20 #include "../auxtrace.h"
  21
  22 #include "intel-pt-insn-decoder.h"
  23 #include "intel-pt-pkt-decoder.h"
  24 #include "intel-pt-decoder.h"
  25 #include "intel-pt-log.h"
  26
  27 #define INTEL_PT_BLK_SIZE 1024
  28
  29 #define BIT63 (((uint64_t)1 << 63))
  30
  31 #define INTEL_PT_RETURN 1
  32
  33 /* Maximum number of loops with no packets consumed i.e. stuck in a loop */
  34 #define INTEL_PT_MAX_LOOPS 10000
  35
  36 struct intel_pt_blk {
  37         struct intel_pt_blk *prev;
  38         uint64_t ip[INTEL_PT_BLK_SIZE];
  39 };
  40
  41 struct intel_pt_stack {
  42         struct intel_pt_blk *blk;
  43         struct intel_pt_blk *spare;
  44         int pos;
  45 };
  46
  47 enum intel_pt_pkt_state {
  48         INTEL_PT_STATE_NO_PSB,
  49         INTEL_PT_STATE_NO_IP,
  50         INTEL_PT_STATE_ERR_RESYNC,
  51         INTEL_PT_STATE_IN_SYNC,
  52         INTEL_PT_STATE_TNT_CONT,
  53         INTEL_PT_STATE_TNT,
  54         INTEL_PT_STATE_TIP,
  55         INTEL_PT_STATE_TIP_PGD,
  56         INTEL_PT_STATE_FUP,
  57         INTEL_PT_STATE_FUP_NO_TIP,
  58 };
  59
  60 static inline bool intel_pt_sample_time(enum intel_pt_pkt_state pkt_state)
  61 {
  62         switch (pkt_state) {
  63         case INTEL_PT_STATE_NO_PSB:
  64         case INTEL_PT_STATE_NO_IP:
  65         case INTEL_PT_STATE_ERR_RESYNC:
  66         case INTEL_PT_STATE_IN_SYNC:
  67         case INTEL_PT_STATE_TNT_CONT:
  68                 return true;
  69         case INTEL_PT_STATE_TNT:
  70         case INTEL_PT_STATE_TIP:
  71         case INTEL_PT_STATE_TIP_PGD:
  72         case INTEL_PT_STATE_FUP:
  73         case INTEL_PT_STATE_FUP_NO_TIP:
  74                 return false;
  75         default:
  76                 return true;
  77         };
  78 }
  79
  80 #ifdef INTEL_PT_STRICT
  81 #define INTEL_PT_STATE_ERR1     INTEL_PT_STATE_NO_PSB
  82 #define INTEL_PT_STATE_ERR2     INTEL_PT_STATE_NO_PSB
  83 #define INTEL_PT_STATE_ERR3     INTEL_PT_STATE_NO_PSB
  84 #define INTEL_PT_STATE_ERR4     INTEL_PT_STATE_NO_PSB
  85 #else
  86 #define INTEL_PT_STATE_ERR1     (decoder->pkt_state)
  87 #define INTEL_PT_STATE_ERR2     INTEL_PT_STATE_NO_IP
  88 #define INTEL_PT_STATE_ERR3     INTEL_PT_STATE_ERR_RESYNC
  89 #define INTEL_PT_STATE_ERR4     INTEL_PT_STATE_IN_SYNC
  90 #endif
  91
  92 struct intel_pt_decoder {
  93         int (*get_trace)(struct intel_pt_buffer *buffer, void *data);
  94         int (*walk_insn)(struct intel_pt_insn *intel_pt_insn,
  95                          uint64_t *insn_cnt_ptr, uint64_t *ip, uint64_t to_ip,
  96                          uint64_t max_insn_cnt, void *data);
  97         bool (*pgd_ip)(uint64_t ip, void *data);
  98         int (*lookahead)(void *data, intel_pt_lookahead_cb_t cb, void *cb_data);
  99         void *data;
 100         struct intel_pt_state state;
 101         const unsigned char *buf;
 102         size_t len;
 103         bool return_compression;
 104         bool branch_enable;
 105         bool mtc_insn;
 106         bool pge;
 107         bool have_tma;
 108         bool have_cyc;
 109         bool fixup_last_mtc;
 110         bool have_last_ip;
 111         bool in_psb;
 112         enum intel_pt_param_flags flags;
 113         uint64_t pos;
 114         uint64_t last_ip;
 115         uint64_t ip;
 116         uint64_t cr3;
 117         uint64_t timestamp;
 118         uint64_t tsc_timestamp;
 119         uint64_t ref_timestamp;
 120         uint64_t buf_timestamp;
 121         uint64_t sample_timestamp;
 122         uint64_t ret_addr;
 123         uint64_t ctc_timestamp;
 124         uint64_t ctc_delta;
 125         uint64_t cycle_cnt;
 126         uint64_t cyc_ref_timestamp;
 127         uint32_t last_mtc;
 128         uint32_t tsc_ctc_ratio_n;
 129         uint32_t tsc_ctc_ratio_d;
 130         uint32_t tsc_ctc_mult;
 131         uint32_t tsc_slip;
 132         uint32_t ctc_rem_mask;
 133         int mtc_shift;
 134         struct intel_pt_stack stack;
 135         enum intel_pt_pkt_state pkt_state;
 136         enum intel_pt_pkt_ctx pkt_ctx;
 137         enum intel_pt_pkt_ctx prev_pkt_ctx;
 138         enum intel_pt_blk_type blk_type;
 139         int blk_type_pos;
 140         struct intel_pt_pkt packet;
 141         struct intel_pt_pkt tnt;
 142         int pkt_step;
 143         int pkt_len;
 144         int last_packet_type;
 145         unsigned int cbr;
 146         unsigned int cbr_seen;
 147         unsigned int max_non_turbo_ratio;
 148         double max_non_turbo_ratio_fp;
 149         double cbr_cyc_to_tsc;
 150         double calc_cyc_to_tsc;
 151         bool have_calc_cyc_to_tsc;
 152         int exec_mode;
 153         unsigned int insn_bytes;
 154         uint64_t period;
 155         enum intel_pt_period_type period_type;
 156         uint64_t tot_insn_cnt;
 157         uint64_t period_insn_cnt;
 158         uint64_t period_mask;
 159         uint64_t period_ticks;
 160         uint64_t last_masked_timestamp;
 161         uint64_t tot_cyc_cnt;
 162         uint64_t sample_tot_cyc_cnt;
 163         uint64_t base_cyc_cnt;
 164         uint64_t cyc_cnt_timestamp;
 165         double tsc_to_cyc;
 166         bool continuous_period;
 167         bool overflow;
 168         bool set_fup_tx_flags;
 169         bool set_fup_ptw;
 170         bool set_fup_mwait;
 171         bool set_fup_pwre;
 172         bool set_fup_exstop;
 173         bool set_fup_bep;
 174         bool sample_cyc;
 175         unsigned int fup_tx_flags;
 176         unsigned int tx_flags;
 177         uint64_t fup_ptw_payload;
 178         uint64_t fup_mwait_payload;
 179         uint64_t fup_pwre_payload;
 180         uint64_t cbr_payload;
 181         uint64_t timestamp_insn_cnt;
 182         uint64_t sample_insn_cnt;
 183         uint64_t stuck_ip;
 184         int no_progress;
 185         int stuck_ip_prd;
 186         int stuck_ip_cnt;
 187         const unsigned char *next_buf;
 188         size_t next_len;
 189         unsigned char temp_buf[INTEL_PT_PKT_MAX_SZ];
 190 };
 191
 192 static uint64_t intel_pt_lower_power_of_2(uint64_t x)
 193 {
 194         int i;
 195
 196         for (i = 0; x != 1; i++)
 197                 x >>= 1;
 198
 199         return x << i;
 200 }
 201
 202 static void intel_pt_setup_period(struct intel_pt_decoder *decoder)
 203 {
 204         if (decoder->period_type == INTEL_PT_PERIOD_TICKS) {
 205                 uint64_t period;
 206
 207                 period = intel_pt_lower_power_of_2(decoder->period);
 208                 decoder->period_mask  = ~(period - 1);
 209                 decoder->period_ticks = period;
 210         }
 211 }
 212
 213 static uint64_t multdiv(uint64_t t, uint32_t n, uint32_t d)
 214 {
 215         if (!d)
 216                 return 0;
 217         return (t / d) * n + ((t % d) * n) / d;
 218 }
 219
 220 struct intel_pt_decoder *intel_pt_decoder_new(struct intel_pt_params *params)
 221 {
 222         struct intel_pt_decoder *decoder;
 223
 224         if (!params->get_trace || !params->walk_insn)
 225                 return NULL;
 226
 227         decoder = zalloc(sizeof(struct intel_pt_decoder));
 228         if (!decoder)
 229                 return NULL;
 230
 231         decoder->get_trace          = params->get_trace;
 232         decoder->walk_insn          = params->walk_insn;
 233         decoder->pgd_ip             = params->pgd_ip;
 234         decoder->lookahead          = params->lookahead;
 235         decoder->data               = params->data;
 236         decoder->return_compression = params->return_compression;
 237         decoder->branch_enable      = params->branch_enable;
 238
 239         decoder->flags              = params->flags;
 240
 241         decoder->period             = params->period;
 242         decoder->period_type        = params->period_type;
 243
 244         decoder->max_non_turbo_ratio    = params->max_non_turbo_ratio;
 245         decoder->max_non_turbo_ratio_fp = params->max_non_turbo_ratio;
 246
 247         intel_pt_setup_period(decoder);
 248
 249         decoder->mtc_shift = params->mtc_period;
 250         decoder->ctc_rem_mask = (1 << decoder->mtc_shift) - 1;
 251
 252         decoder->tsc_ctc_ratio_n = params->tsc_ctc_ratio_n;
 253         decoder->tsc_ctc_ratio_d = params->tsc_ctc_ratio_d;
 254
 255         if (!decoder->tsc_ctc_ratio_n)
 256                 decoder->tsc_ctc_ratio_d = 0;
 257
 258         if (decoder->tsc_ctc_ratio_d) {
 259                 if (!(decoder->tsc_ctc_ratio_n % decoder->tsc_ctc_ratio_d))
 260                         decoder->tsc_ctc_mult = decoder->tsc_ctc_ratio_n /
 261                                                 decoder->tsc_ctc_ratio_d;
 262         }
 263
 264         /*
 265          * A TSC packet can slip past MTC packets so that the timestamp appears
 266          * to go backwards. One estimate is that can be up to about 40 CPU
 267          * cycles, which is certainly less than 0x1000 TSC ticks, but accept
 268          * slippage an order of magnitude more to be on the safe side.
 269          */
 270         decoder->tsc_slip = 0x10000;
 271
 272         intel_pt_log("timestamp: mtc_shift %u\n", decoder->mtc_shift);
 273         intel_pt_log("timestamp: tsc_ctc_ratio_n %u\n", decoder->tsc_ctc_ratio_n);
 274         intel_pt_log("timestamp: tsc_ctc_ratio_d %u\n", decoder->tsc_ctc_ratio_d);
 275         intel_pt_log("timestamp: tsc_ctc_mult %u\n", decoder->tsc_ctc_mult);
 276         intel_pt_log("timestamp: tsc_slip %#x\n", decoder->tsc_slip);
 277
 278         return decoder;
 279 }
 280
 281 static void intel_pt_pop_blk(struct intel_pt_stack *stack)
 282 {
 283         struct intel_pt_blk *blk = stack->blk;
 284
 285         stack->blk = blk->prev;
 286         if (!stack->spare)
 287                 stack->spare = blk;
 288         else
 289                 free(blk);
 290 }
 291
 292 static uint64_t intel_pt_pop(struct intel_pt_stack *stack)
 293 {
 294         if (!stack->pos) {
 295                 if (!stack->blk)
 296                         return 0;
 297                 intel_pt_pop_blk(stack);
 298                 if (!stack->blk)
 299                         return 0;
 300                 stack->pos = INTEL_PT_BLK_SIZE;
 301         }
 302         return stack->blk->ip[--stack->pos];
 303 }
 304
 305 static int intel_pt_alloc_blk(struct intel_pt_stack *stack)
 306 {
 307         struct intel_pt_blk *blk;
 308
 309         if (stack->spare) {
 310                 blk = stack->spare;
 311                 stack->spare = NULL;
 312         } else {
 313                 blk = malloc(sizeof(struct intel_pt_blk));
 314                 if (!blk)
 315                         return -ENOMEM;
 316         }
 317
 318         blk->prev = stack->blk;
 319         stack->blk = blk;
 320         stack->pos = 0;
 321         return 0;
 322 }
 323
 324 static int intel_pt_push(struct intel_pt_stack *stack, uint64_t ip)
 325 {
 326         int err;
 327
 328         if (!stack->blk || stack->pos == INTEL_PT_BLK_SIZE) {
 329                 err = intel_pt_alloc_blk(stack);
 330                 if (err)
 331                         return err;
 332         }
 333
 334         stack->blk->ip[stack->pos++] = ip;
 335         return 0;
 336 }
 337
 338 static void intel_pt_clear_stack(struct intel_pt_stack *stack)
 339 {
 340         while (stack->blk)
 341                 intel_pt_pop_blk(stack);
 342         stack->pos = 0;
 343 }
 344
 345 static void intel_pt_free_stack(struct intel_pt_stack *stack)
 346 {
 347         intel_pt_clear_stack(stack);
 348         zfree(&stack->blk);
 349         zfree(&stack->spare);
 350 }
 351
 352 void intel_pt_decoder_free(struct intel_pt_decoder *decoder)
 353 {
 354         intel_pt_free_stack(&decoder->stack);
 355         free(decoder);
 356 }
 357
 358 static int intel_pt_ext_err(int code)
 359 {
 360         switch (code) {
 361         case -ENOMEM:
 362                 return INTEL_PT_ERR_NOMEM;
 363         case -ENOSYS:
 364                 return INTEL_PT_ERR_INTERN;
 365         case -EBADMSG:
 366                 return INTEL_PT_ERR_BADPKT;
 367         case -ENODATA:
 368                 return INTEL_PT_ERR_NODATA;
 369         case -EILSEQ:
 370                 return INTEL_PT_ERR_NOINSN;
 371         case -ENOENT:
 372                 return INTEL_PT_ERR_MISMAT;
 373         case -EOVERFLOW:
 374                 return INTEL_PT_ERR_OVR;
 375         case -ENOSPC:
 376                 return INTEL_PT_ERR_LOST;
 377         case -ELOOP:
 378                 return INTEL_PT_ERR_NELOOP;
 379         default:
 380                 return INTEL_PT_ERR_UNK;
 381         }
 382 }
 383
 384 static const char *intel_pt_err_msgs[] = {
 385         [INTEL_PT_ERR_NOMEM]  = "Memory allocation failed",
 386         [INTEL_PT_ERR_INTERN] = "Internal error",
 387         [INTEL_PT_ERR_BADPKT] = "Bad packet",
 388         [INTEL_PT_ERR_NODATA] = "No more data",
 389         [INTEL_PT_ERR_NOINSN] = "Failed to get instruction",
 390         [INTEL_PT_ERR_MISMAT] = "Trace doesn't match instruction",
 391         [INTEL_PT_ERR_OVR]    = "Overflow packet",
 392         [INTEL_PT_ERR_LOST]   = "Lost trace data",
 393         [INTEL_PT_ERR_UNK]    = "Unknown error!",
 394         [INTEL_PT_ERR_NELOOP] = "Never-ending loop",
 395 };
 396
 397 int intel_pt__strerror(int code, char *buf, size_t buflen)
 398 {
 399         if (code < 1 || code >= INTEL_PT_ERR_MAX)
 400                 code = INTEL_PT_ERR_UNK;
 401         strlcpy(buf, intel_pt_err_msgs[code], buflen);
 402         return 0;
 403 }
 404
 405 static uint64_t intel_pt_calc_ip(const struct intel_pt_pkt *packet,
 406                                  uint64_t last_ip)
 407 {
 408         uint64_t ip;
 409
 410         switch (packet->count) {
 411         case 1:
 412                 ip = (last_ip & (uint64_t)0xffffffffffff0000ULL) |
 413                      packet->payload;
 414                 break;
 415         case 2:
 416                 ip = (last_ip & (uint64_t)0xffffffff00000000ULL) |
 417                      packet->payload;
 418                 break;
 419         case 3:
 420                 ip = packet->payload;
 421                 /* Sign-extend 6-byte ip */
 422                 if (ip & (uint64_t)0x800000000000ULL)
 423                         ip |= (uint64_t)0xffff000000000000ULL;
 424                 break;
 425         case 4:
 426                 ip = (last_ip & (uint64_t)0xffff000000000000ULL) |
 427                      packet->payload;
 428                 break;
 429         case 6:
 430                 ip = packet->payload;
 431                 break;
 432         default:
 433                 return 0;
 434         }
 435
 436         return ip;
 437 }
 438
 439 static inline void intel_pt_set_last_ip(struct intel_pt_decoder *decoder)
 440 {
 441         decoder->last_ip = intel_pt_calc_ip(&decoder->packet, decoder->last_ip);
 442         decoder->have_last_ip = true;
 443 }
 444
 445 static inline void intel_pt_set_ip(struct intel_pt_decoder *decoder)
 446 {
 447         intel_pt_set_last_ip(decoder);
 448         decoder->ip = decoder->last_ip;
 449 }
 450
 451 static void intel_pt_decoder_log_packet(struct intel_pt_decoder *decoder)
 452 {
 453         intel_pt_log_packet(&decoder->packet, decoder->pkt_len, decoder->pos,
 454                             decoder->buf);
 455 }
 456
 457 static int intel_pt_bug(struct intel_pt_decoder *decoder)
 458 {
 459         intel_pt_log("ERROR: Internal error\n");
 460         decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
 461         return -ENOSYS;
 462 }
 463
 464 static inline void intel_pt_clear_tx_flags(struct intel_pt_decoder *decoder)
 465 {
 466         decoder->tx_flags = 0;
 467 }
 468
 469 static inline void intel_pt_update_in_tx(struct intel_pt_decoder *decoder)
 470 {
 471         decoder->tx_flags = decoder->packet.payload & INTEL_PT_IN_TX;
 472 }
 473
 474 static int intel_pt_bad_packet(struct intel_pt_decoder *decoder)
 475 {
 476         intel_pt_clear_tx_flags(decoder);
 477         decoder->have_tma = false;
 478         decoder->pkt_len = 1;
 479         decoder->pkt_step = 1;
 480         intel_pt_decoder_log_packet(decoder);
 481         if (decoder->pkt_state != INTEL_PT_STATE_NO_PSB) {
 482                 intel_pt_log("ERROR: Bad packet\n");
 483                 decoder->pkt_state = INTEL_PT_STATE_ERR1;
 484         }
 485         return -EBADMSG;
 486 }
 487
 488 static inline void intel_pt_update_sample_time(struct intel_pt_decoder *decoder)
 489 {
 490         decoder->sample_timestamp = decoder->timestamp;
 491         decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
 492 }
 493
 494 static void intel_pt_reposition(struct intel_pt_decoder *decoder)
 495 {
 496         decoder->ip = 0;
 497         decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
 498         decoder->timestamp = 0;
 499         decoder->have_tma = false;
 500 }
 501
 502 static int intel_pt_get_data(struct intel_pt_decoder *decoder, bool reposition)
 503 {
 504         struct intel_pt_buffer buffer = { .buf = 0, };
 505         int ret;
 506
 507         decoder->pkt_step = 0;
 508
 509         intel_pt_log("Getting more data\n");
 510         ret = decoder->get_trace(&buffer, decoder->data);
 511         if (ret)
 512                 return ret;
 513         decoder->buf = buffer.buf;
 514         decoder->len = buffer.len;
 515         if (!decoder->len) {
 516                 intel_pt_log("No more data\n");
 517                 return -ENODATA;
 518         }
 519         decoder->buf_timestamp = buffer.ref_timestamp;
 520         if (!buffer.consecutive || reposition) {
 521                 intel_pt_reposition(decoder);
 522                 decoder->ref_timestamp = buffer.ref_timestamp;
 523                 decoder->state.trace_nr = buffer.trace_nr;
 524                 intel_pt_log("Reference timestamp 0x%" PRIx64 "\n",
 525                              decoder->ref_timestamp);
 526                 return -ENOLINK;
 527         }
 528
 529         return 0;
 530 }
 531
 532 static int intel_pt_get_next_data(struct intel_pt_decoder *decoder,
 533                                   bool reposition)
 534 {
 535         if (!decoder->next_buf)
 536                 return intel_pt_get_data(decoder, reposition);
 537
 538         decoder->buf = decoder->next_buf;
 539         decoder->len = decoder->next_len;
 540         decoder->next_buf = 0;
 541         decoder->next_len = 0;
 542         return 0;
 543 }
 544
 545 static int intel_pt_get_split_packet(struct intel_pt_decoder *decoder)
 546 {
 547         unsigned char *buf = decoder->temp_buf;
 548         size_t old_len, len, n;
 549         int ret;
 550
 551         old_len = decoder->len;
 552         len = decoder->len;
 553         memcpy(buf, decoder->buf, len);
 554
 555         ret = intel_pt_get_data(decoder, false);
 556         if (ret) {
 557                 decoder->pos += old_len;
 558                 return ret < 0 ? ret : -EINVAL;
 559         }
 560
 561         n = INTEL_PT_PKT_MAX_SZ - len;
 562         if (n > decoder->len)
 563                 n = decoder->len;
 564         memcpy(buf + len, decoder->buf, n);
 565         len += n;
 566
 567         decoder->prev_pkt_ctx = decoder->pkt_ctx;
 568         ret = intel_pt_get_packet(buf, len, &decoder->packet, &decoder->pkt_ctx);
 569         if (ret < (int)old_len) {
 570                 decoder->next_buf = decoder->buf;
 571                 decoder->next_len = decoder->len;
 572                 decoder->buf = buf;
 573                 decoder->len = old_len;
 574                 return intel_pt_bad_packet(decoder);
 575         }
 576
 577         decoder->next_buf = decoder->buf + (ret - old_len);
 578         decoder->next_len = decoder->len - (ret - old_len);
 579
 580         decoder->buf = buf;
 581         decoder->len = ret;
 582
 583         return ret;
 584 }
 585
 586 struct intel_pt_pkt_info {
 587         struct intel_pt_decoder   *decoder;
 588         struct intel_pt_pkt       packet;
 589         uint64_t                  pos;
 590         int                       pkt_len;
 591         int                       last_packet_type;
 592         void                      *data;
 593 };
 594
 595 typedef int (*intel_pt_pkt_cb_t)(struct intel_pt_pkt_info *pkt_info);
 596
 597 /* Lookahead packets in current buffer */
 598 static int intel_pt_pkt_lookahead(struct intel_pt_decoder *decoder,
 599                                   intel_pt_pkt_cb_t cb, void *data)
 600 {
 601         struct intel_pt_pkt_info pkt_info;
 602         const unsigned char *buf = decoder->buf;
 603         enum intel_pt_pkt_ctx pkt_ctx = decoder->pkt_ctx;
 604         size_t len = decoder->len;
 605         int ret;
 606
 607         pkt_info.decoder          = decoder;
 608         pkt_info.pos              = decoder->pos;
 609         pkt_info.pkt_len          = decoder->pkt_step;
 610         pkt_info.last_packet_type = decoder->last_packet_type;
 611         pkt_info.data             = data;
 612
 613         while (1) {
 614                 do {
 615                         pkt_info.pos += pkt_info.pkt_len;
 616                         buf          += pkt_info.pkt_len;
 617                         len          -= pkt_info.pkt_len;
 618
 619                         if (!len)
 620                                 return INTEL_PT_NEED_MORE_BYTES;
 621
 622                         ret = intel_pt_get_packet(buf, len, &pkt_info.packet,
 623                                                   &pkt_ctx);
 624                         if (!ret)
 625                                 return INTEL_PT_NEED_MORE_BYTES;
 626                         if (ret < 0)
 627                                 return ret;
 628
 629                         pkt_info.pkt_len = ret;
 630                 } while (pkt_info.packet.type == INTEL_PT_PAD);
 631
 632                 ret = cb(&pkt_info);
 633                 if (ret)
 634                         return 0;
 635
 636                 pkt_info.last_packet_type = pkt_info.packet.type;
 637         }
 638 }
 639
 640 struct intel_pt_calc_cyc_to_tsc_info {
 641         uint64_t        cycle_cnt;
 642         unsigned int    cbr;
 643         uint32_t        last_mtc;
 644         uint64_t        ctc_timestamp;
 645         uint64_t        ctc_delta;
 646         uint64_t        tsc_timestamp;
 647         uint64_t        timestamp;
 648         bool            have_tma;
 649         bool            fixup_last_mtc;
 650         bool            from_mtc;
 651         double          cbr_cyc_to_tsc;
 652 };
 653
 654 /*
 655  * MTC provides a 8-bit slice of CTC but the TMA packet only provides the lower
 656  * 16 bits of CTC. If mtc_shift > 8 then some of the MTC bits are not in the CTC
 657  * provided by the TMA packet. Fix-up the last_mtc calculated from the TMA
 658  * packet by copying the missing bits from the current MTC assuming the least
 659  * difference between the two, and that the current MTC comes after last_mtc.
 660  */
 661 static void intel_pt_fixup_last_mtc(uint32_t mtc, int mtc_shift,
 662                                     uint32_t *last_mtc)
 663 {
 664         uint32_t first_missing_bit = 1U << (16 - mtc_shift);
 665         uint32_t mask = ~(first_missing_bit - 1);
 666
 667         *last_mtc |= mtc & mask;
 668         if (*last_mtc >= mtc) {
 669                 *last_mtc -= first_missing_bit;
 670                 *last_mtc &= 0xff;
 671         }
 672 }
 673
 674 static int intel_pt_calc_cyc_cb(struct intel_pt_pkt_info *pkt_info)
 675 {
 676         struct intel_pt_decoder *decoder = pkt_info->decoder;
 677         struct intel_pt_calc_cyc_to_tsc_info *data = pkt_info->data;
 678         uint64_t timestamp;
 679         double cyc_to_tsc;
 680         unsigned int cbr;
 681         uint32_t mtc, mtc_delta, ctc, fc, ctc_rem;
 682
 683         switch (pkt_info->packet.type) {
 684         case INTEL_PT_TNT:
 685         case INTEL_PT_TIP_PGE:
 686         case INTEL_PT_TIP:
 687         case INTEL_PT_FUP:
 688         case INTEL_PT_PSB:
 689         case INTEL_PT_PIP:
 690         case INTEL_PT_MODE_EXEC:
 691         case INTEL_PT_MODE_TSX:
 692         case INTEL_PT_PSBEND:
 693         case INTEL_PT_PAD:
 694         case INTEL_PT_VMCS:
 695         case INTEL_PT_MNT:
 696         case INTEL_PT_PTWRITE:
 697         case INTEL_PT_PTWRITE_IP:
 698         case INTEL_PT_BBP:
 699         case INTEL_PT_BIP:
 700         case INTEL_PT_BEP:
 701         case INTEL_PT_BEP_IP:
 702                 return 0;
 703
 704         case INTEL_PT_MTC:
 705                 if (!data->have_tma)
 706                         return 0;
 707
 708                 mtc = pkt_info->packet.payload;
 709                 if (decoder->mtc_shift > 8 && data->fixup_last_mtc) {
 710                         data->fixup_last_mtc = false;
 711                         intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
 712                                                 &data->last_mtc);
 713                 }
 714                 if (mtc > data->last_mtc)
 715                         mtc_delta = mtc - data->last_mtc;
 716                 else
 717                         mtc_delta = mtc + 256 - data->last_mtc;
 718                 data->ctc_delta += mtc_delta << decoder->mtc_shift;
 719                 data->last_mtc = mtc;
 720
 721                 if (decoder->tsc_ctc_mult) {
 722                         timestamp = data->ctc_timestamp +
 723                                 data->ctc_delta * decoder->tsc_ctc_mult;
 724                 } else {
 725                         timestamp = data->ctc_timestamp +
 726                                 multdiv(data->ctc_delta,
 727                                         decoder->tsc_ctc_ratio_n,
 728                                         decoder->tsc_ctc_ratio_d);
 729                 }
 730
 731                 if (timestamp < data->timestamp)
 732                         return 1;
 733
 734                 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
 735                         data->timestamp = timestamp;
 736                         return 0;
 737                 }
 738
 739                 break;
 740
 741         case INTEL_PT_TSC:
 742                 /*
 743                  * For now, do not support using TSC packets - refer
 744                  * intel_pt_calc_cyc_to_tsc().
 745                  */
 746                 if (data->from_mtc)
 747                         return 1;
 748                 timestamp = pkt_info->packet.payload |
 749                             (data->timestamp & (0xffULL << 56));
 750                 if (data->from_mtc && timestamp < data->timestamp &&
 751                     data->timestamp - timestamp < decoder->tsc_slip)
 752                         return 1;
 753                 if (timestamp < data->timestamp)
 754                         timestamp += (1ULL << 56);
 755                 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
 756                         if (data->from_mtc)
 757                                 return 1;
 758                         data->tsc_timestamp = timestamp;
 759                         data->timestamp = timestamp;
 760                         return 0;
 761                 }
 762                 break;
 763
 764         case INTEL_PT_TMA:
 765                 if (data->from_mtc)
 766                         return 1;
 767
 768                 if (!decoder->tsc_ctc_ratio_d)
 769                         return 0;
 770
 771                 ctc = pkt_info->packet.payload;
 772                 fc = pkt_info->packet.count;
 773                 ctc_rem = ctc & decoder->ctc_rem_mask;
 774
 775                 data->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
 776
 777                 data->ctc_timestamp = data->tsc_timestamp - fc;
 778                 if (decoder->tsc_ctc_mult) {
 779                         data->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
 780                 } else {
 781                         data->ctc_timestamp -=
 782                                 multdiv(ctc_rem, decoder->tsc_ctc_ratio_n,
 783                                         decoder->tsc_ctc_ratio_d);
 784                 }
 785
 786                 data->ctc_delta = 0;
 787                 data->have_tma = true;
 788                 data->fixup_last_mtc = true;
 789
 790                 return 0;
 791
 792         case INTEL_PT_CYC:
 793                 data->cycle_cnt += pkt_info->packet.payload;
 794                 return 0;
 795
 796         case INTEL_PT_CBR:
 797                 cbr = pkt_info->packet.payload;
 798                 if (data->cbr && data->cbr != cbr)
 799                         return 1;
 800                 data->cbr = cbr;
 801                 data->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
 802                 return 0;
 803
 804         case INTEL_PT_TIP_PGD:
 805         case INTEL_PT_TRACESTOP:
 806         case INTEL_PT_EXSTOP:
 807         case INTEL_PT_EXSTOP_IP:
 808         case INTEL_PT_MWAIT:
 809         case INTEL_PT_PWRE:
 810         case INTEL_PT_PWRX:
 811         case INTEL_PT_OVF:
 812         case INTEL_PT_BAD: /* Does not happen */
 813         default:
 814                 return 1;
 815         }
 816
 817         if (!data->cbr && decoder->cbr) {
 818                 data->cbr = decoder->cbr;
 819                 data->cbr_cyc_to_tsc = decoder->cbr_cyc_to_tsc;
 820         }
 821
 822         if (!data->cycle_cnt)
 823                 return 1;
 824
 825         cyc_to_tsc = (double)(timestamp - decoder->timestamp) / data->cycle_cnt;
 826
 827         if (data->cbr && cyc_to_tsc > data->cbr_cyc_to_tsc &&
 828             cyc_to_tsc / data->cbr_cyc_to_tsc > 1.25) {
 829                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle too big (c.f. CBR-based value %g), pos " x64_fmt "\n",
 830                              cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
 831                 return 1;
 832         }
 833
 834         decoder->calc_cyc_to_tsc = cyc_to_tsc;
 835         decoder->have_calc_cyc_to_tsc = true;
 836
 837         if (data->cbr) {
 838                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. CBR-based value %g, pos " x64_fmt "\n",
 839                              cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
 840         } else {
 841                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. unknown CBR-based value, pos " x64_fmt "\n",
 842                              cyc_to_tsc, pkt_info->pos);
 843         }
 844
 845         return 1;
 846 }
 847
 848 static void intel_pt_calc_cyc_to_tsc(struct intel_pt_decoder *decoder,
 849                                      bool from_mtc)
 850 {
 851         struct intel_pt_calc_cyc_to_tsc_info data = {
 852                 .cycle_cnt      = 0,
 853                 .cbr            = 0,
 854                 .last_mtc       = decoder->last_mtc,
 855                 .ctc_timestamp  = decoder->ctc_timestamp,
 856                 .ctc_delta      = decoder->ctc_delta,
 857                 .tsc_timestamp  = decoder->tsc_timestamp,
 858                 .timestamp      = decoder->timestamp,
 859                 .have_tma       = decoder->have_tma,
 860                 .fixup_last_mtc = decoder->fixup_last_mtc,
 861                 .from_mtc       = from_mtc,
 862                 .cbr_cyc_to_tsc = 0,
 863         };
 864
 865         /*
 866          * For now, do not support using TSC packets for at least the reasons:
 867          * 1) timing might have stopped
 868          * 2) TSC packets within PSB+ can slip against CYC packets
 869          */
 870         if (!from_mtc)
 871                 return;
 872
 873         intel_pt_pkt_lookahead(decoder, intel_pt_calc_cyc_cb, &data);
 874 }
 875
 876 static int intel_pt_get_next_packet(struct intel_pt_decoder *decoder)
 877 {
 878         int ret;
 879
 880         decoder->last_packet_type = decoder->packet.type;
 881
 882         do {
 883                 decoder->pos += decoder->pkt_step;
 884                 decoder->buf += decoder->pkt_step;
 885                 decoder->len -= decoder->pkt_step;
 886
 887                 if (!decoder->len) {
 888                         ret = intel_pt_get_next_data(decoder, false);
 889                         if (ret)
 890                                 return ret;
 891                 }
 892
 893                 decoder->prev_pkt_ctx = decoder->pkt_ctx;
 894                 ret = intel_pt_get_packet(decoder->buf, decoder->len,
 895                                           &decoder->packet, &decoder->pkt_ctx);
 896                 if (ret == INTEL_PT_NEED_MORE_BYTES && BITS_PER_LONG == 32 &&
 897                     decoder->len < INTEL_PT_PKT_MAX_SZ && !decoder->next_buf) {
 898                         ret = intel_pt_get_split_packet(decoder);
 899                         if (ret < 0)
 900                                 return ret;
 901                 }
 902                 if (ret <= 0)
 903                         return intel_pt_bad_packet(decoder);
 904
 905                 decoder->pkt_len = ret;
 906                 decoder->pkt_step = ret;
 907                 intel_pt_decoder_log_packet(decoder);
 908         } while (decoder->packet.type == INTEL_PT_PAD);
 909
 910         return 0;
 911 }
 912
 913 static uint64_t intel_pt_next_period(struct intel_pt_decoder *decoder)
 914 {
 915         uint64_t timestamp, masked_timestamp;
 916
 917         timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
 918         masked_timestamp = timestamp & decoder->period_mask;
 919         if (decoder->continuous_period) {
 920                 if (masked_timestamp > decoder->last_masked_timestamp)
 921                         return 1;
 922         } else {
 923                 timestamp += 1;
 924                 masked_timestamp = timestamp & decoder->period_mask;
 925                 if (masked_timestamp > decoder->last_masked_timestamp) {
 926                         decoder->last_masked_timestamp = masked_timestamp;
 927                         decoder->continuous_period = true;
 928                 }
 929         }
 930
 931         if (masked_timestamp < decoder->last_masked_timestamp)
 932                 return decoder->period_ticks;
 933
 934         return decoder->period_ticks - (timestamp - masked_timestamp);
 935 }
 936
 937 static uint64_t intel_pt_next_sample(struct intel_pt_decoder *decoder)
 938 {
 939         switch (decoder->period_type) {
 940         case INTEL_PT_PERIOD_INSTRUCTIONS:
 941                 return decoder->period - decoder->period_insn_cnt;
 942         case INTEL_PT_PERIOD_TICKS:
 943                 return intel_pt_next_period(decoder);
 944         case INTEL_PT_PERIOD_NONE:
 945         case INTEL_PT_PERIOD_MTC:
 946         default:
 947                 return 0;
 948         }
 949 }
 950
 951 static void intel_pt_sample_insn(struct intel_pt_decoder *decoder)
 952 {
 953         uint64_t timestamp, masked_timestamp;
 954
 955         switch (decoder->period_type) {
 956         case INTEL_PT_PERIOD_INSTRUCTIONS:
 957                 decoder->period_insn_cnt = 0;
 958                 break;
 959         case INTEL_PT_PERIOD_TICKS:
 960                 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
 961                 masked_timestamp = timestamp & decoder->period_mask;
 962                 if (masked_timestamp > decoder->last_masked_timestamp)
 963                         decoder->last_masked_timestamp = masked_timestamp;
 964                 else
 965                         decoder->last_masked_timestamp += decoder->period_ticks;
 966                 break;
 967         case INTEL_PT_PERIOD_NONE:
 968         case INTEL_PT_PERIOD_MTC:
 969         default:
 970                 break;
 971         }
 972
 973         decoder->state.type |= INTEL_PT_INSTRUCTION;
 974 }
 975
 976 static int intel_pt_walk_insn(struct intel_pt_decoder *decoder,
 977                               struct intel_pt_insn *intel_pt_insn, uint64_t ip)
 978 {
 979         uint64_t max_insn_cnt, insn_cnt = 0;
 980         int err;
 981
 982         if (!decoder->mtc_insn)
 983                 decoder->mtc_insn = true;
 984
 985         max_insn_cnt = intel_pt_next_sample(decoder);
 986
 987         err = decoder->walk_insn(intel_pt_insn, &insn_cnt, &decoder->ip, ip,
 988                                  max_insn_cnt, decoder->data);
 989
 990         decoder->tot_insn_cnt += insn_cnt;
 991         decoder->timestamp_insn_cnt += insn_cnt;
 992         decoder->sample_insn_cnt += insn_cnt;
 993         decoder->period_insn_cnt += insn_cnt;
 994
 995         if (err) {
 996                 decoder->no_progress = 0;
 997                 decoder->pkt_state = INTEL_PT_STATE_ERR2;
 998                 intel_pt_log_at("ERROR: Failed to get instruction",
 999                                 decoder->ip);
1000                 if (err == -ENOENT)
1001                         return -ENOLINK;
1002                 return -EILSEQ;
1003         }
1004
1005         if (ip && decoder->ip == ip) {
1006                 err = -EAGAIN;
1007                 goto out;
1008         }
1009
1010         if (max_insn_cnt && insn_cnt >= max_insn_cnt)
1011                 intel_pt_sample_insn(decoder);
1012
1013         if (intel_pt_insn->branch == INTEL_PT_BR_NO_BRANCH) {
1014                 decoder->state.type = INTEL_PT_INSTRUCTION;
1015                 decoder->state.from_ip = decoder->ip;
1016                 decoder->state.to_ip = 0;
1017                 decoder->ip += intel_pt_insn->length;
1018                 err = INTEL_PT_RETURN;
1019                 goto out;
1020         }
1021
1022         if (intel_pt_insn->op == INTEL_PT_OP_CALL) {
1023                 /* Zero-length calls are excluded */
1024                 if (intel_pt_insn->branch != INTEL_PT_BR_UNCONDITIONAL ||
1025                     intel_pt_insn->rel) {
1026                         err = intel_pt_push(&decoder->stack, decoder->ip +
1027                                             intel_pt_insn->length);
1028                         if (err)
1029                                 goto out;
1030                 }
1031         } else if (intel_pt_insn->op == INTEL_PT_OP_RET) {
1032                 decoder->ret_addr = intel_pt_pop(&decoder->stack);
1033         }
1034
1035         if (intel_pt_insn->branch == INTEL_PT_BR_UNCONDITIONAL) {
1036                 int cnt = decoder->no_progress++;
1037
1038                 decoder->state.from_ip = decoder->ip;
1039                 decoder->ip += intel_pt_insn->length +
1040                                 intel_pt_insn->rel;
1041                 decoder->state.to_ip = decoder->ip;
1042                 err = INTEL_PT_RETURN;
1043
1044                 /*
1045                  * Check for being stuck in a loop.  This can happen if a
1046                  * decoder error results in the decoder erroneously setting the
1047                  * ip to an address that is itself in an infinite loop that
1048                  * consumes no packets.  When that happens, there must be an
1049                  * unconditional branch.
1050                  */
1051                 if (cnt) {
1052                         if (cnt == 1) {
1053                                 decoder->stuck_ip = decoder->state.to_ip;
1054                                 decoder->stuck_ip_prd = 1;
1055                                 decoder->stuck_ip_cnt = 1;
1056                         } else if (cnt > INTEL_PT_MAX_LOOPS ||
1057                                    decoder->state.to_ip == decoder->stuck_ip) {
1058                                 intel_pt_log_at("ERROR: Never-ending loop",
1059                                                 decoder->state.to_ip);
1060                                 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1061                                 err = -ELOOP;
1062                                 goto out;
1063                         } else if (!--decoder->stuck_ip_cnt) {
1064                                 decoder->stuck_ip_prd += 1;
1065                                 decoder->stuck_ip_cnt = decoder->stuck_ip_prd;
1066                                 decoder->stuck_ip = decoder->state.to_ip;
1067                         }
1068                 }
1069                 goto out_no_progress;
1070         }
1071 out:
1072         decoder->no_progress = 0;
1073 out_no_progress:
1074         decoder->state.insn_op = intel_pt_insn->op;
1075         decoder->state.insn_len = intel_pt_insn->length;
1076         memcpy(decoder->state.insn, intel_pt_insn->buf,
1077                INTEL_PT_INSN_BUF_SZ);
1078
1079         if (decoder->tx_flags & INTEL_PT_IN_TX)
1080                 decoder->state.flags |= INTEL_PT_IN_TX;
1081
1082         return err;
1083 }
1084
1085 static bool intel_pt_fup_event(struct intel_pt_decoder *decoder)
1086 {
1087         bool ret = false;
1088
1089         if (decoder->set_fup_tx_flags) {
1090                 decoder->set_fup_tx_flags = false;
1091                 decoder->tx_flags = decoder->fup_tx_flags;
1092                 decoder->state.type = INTEL_PT_TRANSACTION;
1093                 decoder->state.from_ip = decoder->ip;
1094                 decoder->state.to_ip = 0;
1095                 decoder->state.flags = decoder->fup_tx_flags;
1096                 return true;
1097         }
1098         if (decoder->set_fup_ptw) {
1099                 decoder->set_fup_ptw = false;
1100                 decoder->state.type = INTEL_PT_PTW;
1101                 decoder->state.flags |= INTEL_PT_FUP_IP;
1102                 decoder->state.from_ip = decoder->ip;
1103                 decoder->state.to_ip = 0;
1104                 decoder->state.ptw_payload = decoder->fup_ptw_payload;
1105                 return true;
1106         }
1107         if (decoder->set_fup_mwait) {
1108                 decoder->set_fup_mwait = false;
1109                 decoder->state.type = INTEL_PT_MWAIT_OP;
1110                 decoder->state.from_ip = decoder->ip;
1111                 decoder->state.to_ip = 0;
1112                 decoder->state.mwait_payload = decoder->fup_mwait_payload;
1113                 ret = true;
1114         }
1115         if (decoder->set_fup_pwre) {
1116                 decoder->set_fup_pwre = false;
1117                 decoder->state.type |= INTEL_PT_PWR_ENTRY;
1118                 decoder->state.type &= ~INTEL_PT_BRANCH;
1119                 decoder->state.from_ip = decoder->ip;
1120                 decoder->state.to_ip = 0;
1121                 decoder->state.pwre_payload = decoder->fup_pwre_payload;
1122                 ret = true;
1123         }
1124         if (decoder->set_fup_exstop) {
1125                 decoder->set_fup_exstop = false;
1126                 decoder->state.type |= INTEL_PT_EX_STOP;
1127                 decoder->state.type &= ~INTEL_PT_BRANCH;
1128                 decoder->state.flags |= INTEL_PT_FUP_IP;
1129                 decoder->state.from_ip = decoder->ip;
1130                 decoder->state.to_ip = 0;
1131                 ret = true;
1132         }
1133         if (decoder->set_fup_bep) {
1134                 decoder->set_fup_bep = false;
1135                 decoder->state.type |= INTEL_PT_BLK_ITEMS;
1136                 decoder->state.type &= ~INTEL_PT_BRANCH;
1137                 decoder->state.from_ip = decoder->ip;
1138                 decoder->state.to_ip = 0;
1139                 ret = true;
1140         }
1141         return ret;
1142 }
1143
1144 static inline bool intel_pt_fup_with_nlip(struct intel_pt_decoder *decoder,
1145                                           struct intel_pt_insn *intel_pt_insn,
1146                                           uint64_t ip, int err)
1147 {
1148         return decoder->flags & INTEL_PT_FUP_WITH_NLIP && !err &&
1149                intel_pt_insn->branch == INTEL_PT_BR_INDIRECT &&
1150                ip == decoder->ip + intel_pt_insn->length;
1151 }
1152
1153 static int intel_pt_walk_fup(struct intel_pt_decoder *decoder)
1154 {
1155         struct intel_pt_insn intel_pt_insn;
1156         uint64_t ip;
1157         int err;
1158
1159         ip = decoder->last_ip;
1160
1161         while (1) {
1162                 err = intel_pt_walk_insn(decoder, &intel_pt_insn, ip);
1163                 if (err == INTEL_PT_RETURN)
1164                         return 0;
1165                 if (err == -EAGAIN ||
1166                     intel_pt_fup_with_nlip(decoder, &intel_pt_insn, ip, err)) {
1167                         if (intel_pt_fup_event(decoder))
1168                                 return 0;
1169                         return -EAGAIN;
1170                 }
1171                 decoder->set_fup_tx_flags = false;
1172                 if (err)
1173                         return err;
1174
1175                 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1176                         intel_pt_log_at("ERROR: Unexpected indirect branch",
1177                                         decoder->ip);
1178                         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1179                         return -ENOENT;
1180                 }
1181
1182                 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1183                         intel_pt_log_at("ERROR: Unexpected conditional branch",
1184                                         decoder->ip);
1185                         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1186                         return -ENOENT;
1187                 }
1188
1189                 intel_pt_bug(decoder);
1190         }
1191 }
1192
1193 static int intel_pt_walk_tip(struct intel_pt_decoder *decoder)
1194 {
1195         struct intel_pt_insn intel_pt_insn;
1196         int err;
1197
1198         err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1199         if (err == INTEL_PT_RETURN &&
1200             decoder->pgd_ip &&
1201             decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1202             (decoder->state.type & INTEL_PT_BRANCH) &&
1203             decoder->pgd_ip(decoder->state.to_ip, decoder->data)) {
1204                 /* Unconditional branch leaving filter region */
1205                 decoder->no_progress = 0;
1206                 decoder->pge = false;
1207                 decoder->continuous_period = false;
1208                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1209                 decoder->state.type |= INTEL_PT_TRACE_END;
1210                 return 0;
1211         }
1212         if (err == INTEL_PT_RETURN)
1213                 return 0;
1214         if (err)
1215                 return err;
1216
1217         if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1218                 if (decoder->pkt_state == INTEL_PT_STATE_TIP_PGD) {
1219                         decoder->pge = false;
1220                         decoder->continuous_period = false;
1221                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1222                         decoder->state.from_ip = decoder->ip;
1223                         if (decoder->packet.count == 0) {
1224                                 decoder->state.to_ip = 0;
1225                         } else {
1226                                 decoder->state.to_ip = decoder->last_ip;
1227                                 decoder->ip = decoder->last_ip;
1228                         }
1229                         decoder->state.type |= INTEL_PT_TRACE_END;
1230                 } else {
1231                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1232                         decoder->state.from_ip = decoder->ip;
1233                         if (decoder->packet.count == 0) {
1234                                 decoder->state.to_ip = 0;
1235                         } else {
1236                                 decoder->state.to_ip = decoder->last_ip;
1237                                 decoder->ip = decoder->last_ip;
1238                         }
1239                 }
1240                 return 0;
1241         }
1242
1243         if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1244                 uint64_t to_ip = decoder->ip + intel_pt_insn.length +
1245                                  intel_pt_insn.rel;
1246
1247                 if (decoder->pgd_ip &&
1248                     decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1249                     decoder->pgd_ip(to_ip, decoder->data)) {
1250                         /* Conditional branch leaving filter region */
1251                         decoder->pge = false;
1252                         decoder->continuous_period = false;
1253                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1254                         decoder->ip = to_ip;
1255                         decoder->state.from_ip = decoder->ip;
1256                         decoder->state.to_ip = to_ip;
1257                         decoder->state.type |= INTEL_PT_TRACE_END;
1258                         return 0;
1259                 }
1260                 intel_pt_log_at("ERROR: Conditional branch when expecting indirect branch",
1261                                 decoder->ip);
1262                 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1263                 return -ENOENT;
1264         }
1265
1266         return intel_pt_bug(decoder);
1267 }
1268
1269 static int intel_pt_walk_tnt(struct intel_pt_decoder *decoder)
1270 {
1271         struct intel_pt_insn intel_pt_insn;
1272         int err;
1273
1274         while (1) {
1275                 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1276                 if (err == INTEL_PT_RETURN)
1277                         return 0;
1278                 if (err)
1279                         return err;
1280
1281                 if (intel_pt_insn.op == INTEL_PT_OP_RET) {
1282                         if (!decoder->return_compression) {
1283                                 intel_pt_log_at("ERROR: RET when expecting conditional branch",
1284                                                 decoder->ip);
1285                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1286                                 return -ENOENT;
1287                         }
1288                         if (!decoder->ret_addr) {
1289                                 intel_pt_log_at("ERROR: Bad RET compression (stack empty)",
1290                                                 decoder->ip);
1291                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1292                                 return -ENOENT;
1293                         }
1294                         if (!(decoder->tnt.payload & BIT63)) {
1295                                 intel_pt_log_at("ERROR: Bad RET compression (TNT=N)",
1296                                                 decoder->ip);
1297                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1298                                 return -ENOENT;
1299                         }
1300                         decoder->tnt.count -= 1;
1301                         if (decoder->tnt.count)
1302                                 decoder->pkt_state = INTEL_PT_STATE_TNT_CONT;
1303                         else
1304                                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1305                         decoder->tnt.payload <<= 1;
1306                         decoder->state.from_ip = decoder->ip;
1307                         decoder->ip = decoder->ret_addr;
1308                         decoder->state.to_ip = decoder->ip;
1309                         return 0;
1310                 }
1311
1312                 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1313                         /* Handle deferred TIPs */
1314                         err = intel_pt_get_next_packet(decoder);
1315                         if (err)
1316                                 return err;
1317                         if (decoder->packet.type != INTEL_PT_TIP ||
1318                             decoder->packet.count == 0) {
1319                                 intel_pt_log_at("ERROR: Missing deferred TIP for indirect branch",
1320                                                 decoder->ip);
1321                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1322                                 decoder->pkt_step = 0;
1323                                 return -ENOENT;
1324                         }
1325                         intel_pt_set_last_ip(decoder);
1326                         decoder->state.from_ip = decoder->ip;
1327                         decoder->state.to_ip = decoder->last_ip;
1328                         decoder->ip = decoder->last_ip;
1329                         return 0;
1330                 }
1331
1332                 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1333                         decoder->tnt.count -= 1;
1334                         if (decoder->tnt.count)
1335                                 decoder->pkt_state = INTEL_PT_STATE_TNT_CONT;
1336                         else
1337                                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1338                         if (decoder->tnt.payload & BIT63) {
1339                                 decoder->tnt.payload <<= 1;
1340                                 decoder->state.from_ip = decoder->ip;
1341                                 decoder->ip += intel_pt_insn.length +
1342                                                intel_pt_insn.rel;
1343                                 decoder->state.to_ip = decoder->ip;
1344                                 return 0;
1345                         }
1346                         /* Instruction sample for a non-taken branch */
1347                         if (decoder->state.type & INTEL_PT_INSTRUCTION) {
1348                                 decoder->tnt.payload <<= 1;
1349                                 decoder->state.type = INTEL_PT_INSTRUCTION;
1350                                 decoder->state.from_ip = decoder->ip;
1351                                 decoder->state.to_ip = 0;
1352                                 decoder->ip += intel_pt_insn.length;
1353                                 return 0;
1354                         }
1355                         decoder->sample_cyc = false;
1356                         decoder->ip += intel_pt_insn.length;
1357                         if (!decoder->tnt.count) {
1358                                 intel_pt_update_sample_time(decoder);
1359                                 return -EAGAIN;
1360                         }
1361                         decoder->tnt.payload <<= 1;
1362                         continue;
1363                 }
1364
1365                 return intel_pt_bug(decoder);
1366         }
1367 }
1368
1369 static int intel_pt_mode_tsx(struct intel_pt_decoder *decoder, bool *no_tip)
1370 {
1371         unsigned int fup_tx_flags;
1372         int err;
1373
1374         fup_tx_flags = decoder->packet.payload &
1375                        (INTEL_PT_IN_TX | INTEL_PT_ABORT_TX);
1376         err = intel_pt_get_next_packet(decoder);
1377         if (err)
1378                 return err;
1379         if (decoder->packet.type == INTEL_PT_FUP) {
1380                 decoder->fup_tx_flags = fup_tx_flags;
1381                 decoder->set_fup_tx_flags = true;
1382                 if (!(decoder->fup_tx_flags & INTEL_PT_ABORT_TX))
1383                         *no_tip = true;
1384         } else {
1385                 intel_pt_log_at("ERROR: Missing FUP after MODE.TSX",
1386                                 decoder->pos);
1387                 intel_pt_update_in_tx(decoder);
1388         }
1389         return 0;
1390 }
1391
1392 static uint64_t intel_pt_8b_tsc(uint64_t timestamp, uint64_t ref_timestamp)
1393 {
1394         timestamp |= (ref_timestamp & (0xffULL << 56));
1395
1396         if (timestamp < ref_timestamp) {
1397                 if (ref_timestamp - timestamp > (1ULL << 55))
1398                         timestamp += (1ULL << 56);
1399         } else {
1400                 if (timestamp - ref_timestamp > (1ULL << 55))
1401                         timestamp -= (1ULL << 56);
1402         }
1403
1404         return timestamp;
1405 }
1406
1407 static void intel_pt_calc_tsc_timestamp(struct intel_pt_decoder *decoder)
1408 {
1409         uint64_t timestamp;
1410
1411         decoder->have_tma = false;
1412
1413         if (decoder->ref_timestamp) {
1414                 timestamp = intel_pt_8b_tsc(decoder->packet.payload,
1415                                             decoder->ref_timestamp);
1416                 decoder->tsc_timestamp = timestamp;
1417                 decoder->timestamp = timestamp;
1418                 decoder->ref_timestamp = 0;
1419                 decoder->timestamp_insn_cnt = 0;
1420         } else if (decoder->timestamp) {
1421                 timestamp = decoder->packet.payload |
1422                             (decoder->timestamp & (0xffULL << 56));
1423                 decoder->tsc_timestamp = timestamp;
1424                 if (timestamp < decoder->timestamp &&
1425                     decoder->timestamp - timestamp < decoder->tsc_slip) {
1426                         intel_pt_log_to("Suppressing backwards timestamp",
1427                                         timestamp);
1428                         timestamp = decoder->timestamp;
1429                 }
1430                 if (timestamp < decoder->timestamp) {
1431                         intel_pt_log_to("Wraparound timestamp", timestamp);
1432                         timestamp += (1ULL << 56);
1433                         decoder->tsc_timestamp = timestamp;
1434                 }
1435                 decoder->timestamp = timestamp;
1436                 decoder->timestamp_insn_cnt = 0;
1437         }
1438
1439         if (decoder->last_packet_type == INTEL_PT_CYC) {
1440                 decoder->cyc_ref_timestamp = decoder->timestamp;
1441                 decoder->cycle_cnt = 0;
1442                 decoder->have_calc_cyc_to_tsc = false;
1443                 intel_pt_calc_cyc_to_tsc(decoder, false);
1444         }
1445
1446         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1447 }
1448
1449 static int intel_pt_overflow(struct intel_pt_decoder *decoder)
1450 {
1451         intel_pt_log("ERROR: Buffer overflow\n");
1452         intel_pt_clear_tx_flags(decoder);
1453         decoder->timestamp_insn_cnt = 0;
1454         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1455         decoder->overflow = true;
1456         return -EOVERFLOW;
1457 }
1458
1459 static inline void intel_pt_mtc_cyc_cnt_pge(struct intel_pt_decoder *decoder)
1460 {
1461         if (decoder->have_cyc)
1462                 return;
1463
1464         decoder->cyc_cnt_timestamp = decoder->timestamp;
1465         decoder->base_cyc_cnt = decoder->tot_cyc_cnt;
1466 }
1467
1468 static inline void intel_pt_mtc_cyc_cnt_cbr(struct intel_pt_decoder *decoder)
1469 {
1470         decoder->tsc_to_cyc = decoder->cbr / decoder->max_non_turbo_ratio_fp;
1471
1472         if (decoder->pge)
1473                 intel_pt_mtc_cyc_cnt_pge(decoder);
1474 }
1475
1476 static inline void intel_pt_mtc_cyc_cnt_upd(struct intel_pt_decoder *decoder)
1477 {
1478         uint64_t tot_cyc_cnt, tsc_delta;
1479
1480         if (decoder->have_cyc)
1481                 return;
1482
1483         decoder->sample_cyc = true;
1484
1485         if (!decoder->pge || decoder->timestamp <= decoder->cyc_cnt_timestamp)
1486                 return;
1487
1488         tsc_delta = decoder->timestamp - decoder->cyc_cnt_timestamp;
1489         tot_cyc_cnt = tsc_delta * decoder->tsc_to_cyc + decoder->base_cyc_cnt;
1490
1491         if (tot_cyc_cnt > decoder->tot_cyc_cnt)
1492                 decoder->tot_cyc_cnt = tot_cyc_cnt;
1493 }
1494
1495 static void intel_pt_calc_tma(struct intel_pt_decoder *decoder)
1496 {
1497         uint32_t ctc = decoder->packet.payload;
1498         uint32_t fc = decoder->packet.count;
1499         uint32_t ctc_rem = ctc & decoder->ctc_rem_mask;
1500
1501         if (!decoder->tsc_ctc_ratio_d)
1502                 return;
1503
1504         if (decoder->pge && !decoder->in_psb)
1505                 intel_pt_mtc_cyc_cnt_pge(decoder);
1506         else
1507                 intel_pt_mtc_cyc_cnt_upd(decoder);
1508
1509         decoder->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
1510         decoder->ctc_timestamp = decoder->tsc_timestamp - fc;
1511         if (decoder->tsc_ctc_mult) {
1512                 decoder->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
1513         } else {
1514                 decoder->ctc_timestamp -= multdiv(ctc_rem,
1515                                                   decoder->tsc_ctc_ratio_n,
1516                                                   decoder->tsc_ctc_ratio_d);
1517         }
1518         decoder->ctc_delta = 0;
1519         decoder->have_tma = true;
1520         decoder->fixup_last_mtc = true;
1521         intel_pt_log("CTC timestamp " x64_fmt " last MTC %#x  CTC rem %#x\n",
1522                      decoder->ctc_timestamp, decoder->last_mtc, ctc_rem);
1523 }
1524
1525 static void intel_pt_calc_mtc_timestamp(struct intel_pt_decoder *decoder)
1526 {
1527         uint64_t timestamp;
1528         uint32_t mtc, mtc_delta;
1529
1530         if (!decoder->have_tma)
1531                 return;
1532
1533         mtc = decoder->packet.payload;
1534
1535         if (decoder->mtc_shift > 8 && decoder->fixup_last_mtc) {
1536                 decoder->fixup_last_mtc = false;
1537                 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
1538                                         &decoder->last_mtc);
1539         }
1540
1541         if (mtc > decoder->last_mtc)
1542                 mtc_delta = mtc - decoder->last_mtc;
1543         else
1544                 mtc_delta = mtc + 256 - decoder->last_mtc;
1545
1546         decoder->ctc_delta += mtc_delta << decoder->mtc_shift;
1547
1548         if (decoder->tsc_ctc_mult) {
1549                 timestamp = decoder->ctc_timestamp +
1550                             decoder->ctc_delta * decoder->tsc_ctc_mult;
1551         } else {
1552                 timestamp = decoder->ctc_timestamp +
1553                             multdiv(decoder->ctc_delta,
1554                                     decoder->tsc_ctc_ratio_n,
1555                                     decoder->tsc_ctc_ratio_d);
1556         }
1557
1558         if (timestamp < decoder->timestamp)
1559                 intel_pt_log("Suppressing MTC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1560                              timestamp, decoder->timestamp);
1561         else
1562                 decoder->timestamp = timestamp;
1563
1564         intel_pt_mtc_cyc_cnt_upd(decoder);
1565
1566         decoder->timestamp_insn_cnt = 0;
1567         decoder->last_mtc = mtc;
1568
1569         if (decoder->last_packet_type == INTEL_PT_CYC) {
1570                 decoder->cyc_ref_timestamp = decoder->timestamp;
1571                 decoder->cycle_cnt = 0;
1572                 decoder->have_calc_cyc_to_tsc = false;
1573                 intel_pt_calc_cyc_to_tsc(decoder, true);
1574         }
1575
1576         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1577 }
1578
1579 static void intel_pt_calc_cbr(struct intel_pt_decoder *decoder)
1580 {
1581         unsigned int cbr = decoder->packet.payload & 0xff;
1582
1583         decoder->cbr_payload = decoder->packet.payload;
1584
1585         if (decoder->cbr == cbr)
1586                 return;
1587
1588         decoder->cbr = cbr;
1589         decoder->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
1590
1591         intel_pt_mtc_cyc_cnt_cbr(decoder);
1592 }
1593
1594 static void intel_pt_calc_cyc_timestamp(struct intel_pt_decoder *decoder)
1595 {
1596         uint64_t timestamp = decoder->cyc_ref_timestamp;
1597
1598         decoder->have_cyc = true;
1599
1600         decoder->cycle_cnt += decoder->packet.payload;
1601         if (decoder->pge)
1602                 decoder->tot_cyc_cnt += decoder->packet.payload;
1603         decoder->sample_cyc = true;
1604
1605         if (!decoder->cyc_ref_timestamp)
1606                 return;
1607
1608         if (decoder->have_calc_cyc_to_tsc)
1609                 timestamp += decoder->cycle_cnt * decoder->calc_cyc_to_tsc;
1610         else if (decoder->cbr)
1611                 timestamp += decoder->cycle_cnt * decoder->cbr_cyc_to_tsc;
1612         else
1613                 return;
1614
1615         if (timestamp < decoder->timestamp)
1616                 intel_pt_log("Suppressing CYC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1617                              timestamp, decoder->timestamp);
1618         else
1619                 decoder->timestamp = timestamp;
1620
1621         decoder->timestamp_insn_cnt = 0;
1622
1623         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1624 }
1625
1626 static void intel_pt_bbp(struct intel_pt_decoder *decoder)
1627 {
1628         if (decoder->prev_pkt_ctx == INTEL_PT_NO_CTX) {
1629                 memset(decoder->state.items.mask, 0, sizeof(decoder->state.items.mask));
1630                 decoder->state.items.is_32_bit = false;
1631         }
1632         decoder->blk_type = decoder->packet.payload;
1633         decoder->blk_type_pos = intel_pt_blk_type_pos(decoder->blk_type);
1634         if (decoder->blk_type == INTEL_PT_GP_REGS)
1635                 decoder->state.items.is_32_bit = decoder->packet.count;
1636         if (decoder->blk_type_pos < 0) {
1637                 intel_pt_log("WARNING: Unknown block type %u\n",
1638                              decoder->blk_type);
1639         } else if (decoder->state.items.mask[decoder->blk_type_pos]) {
1640                 intel_pt_log("WARNING: Duplicate block type %u\n",
1641                              decoder->blk_type);
1642         }
1643 }
1644
1645 static void intel_pt_bip(struct intel_pt_decoder *decoder)
1646 {
1647         uint32_t id = decoder->packet.count;
1648         uint32_t bit = 1 << id;
1649         int pos = decoder->blk_type_pos;
1650
1651         if (pos < 0 || id >= INTEL_PT_BLK_ITEM_ID_CNT) {
1652                 intel_pt_log("WARNING: Unknown block item %u type %d\n",
1653                              id, decoder->blk_type);
1654                 return;
1655         }
1656
1657         if (decoder->state.items.mask[pos] & bit) {
1658                 intel_pt_log("WARNING: Duplicate block item %u type %d\n",
1659                              id, decoder->blk_type);
1660         }
1661
1662         decoder->state.items.mask[pos] |= bit;
1663         decoder->state.items.val[pos][id] = decoder->packet.payload;
1664 }
1665
1666 /* Walk PSB+ packets when already in sync. */
1667 static int intel_pt_walk_psbend(struct intel_pt_decoder *decoder)
1668 {
1669         int err;
1670
1671         decoder->in_psb = true;
1672
1673         while (1) {
1674                 err = intel_pt_get_next_packet(decoder);
1675                 if (err)
1676                         goto out;
1677
1678                 switch (decoder->packet.type) {
1679                 case INTEL_PT_PSBEND:
1680                         err = 0;
1681                         goto out;
1682
1683                 case INTEL_PT_TIP_PGD:
1684                 case INTEL_PT_TIP_PGE:
1685                 case INTEL_PT_TIP:
1686                 case INTEL_PT_TNT:
1687                 case INTEL_PT_TRACESTOP:
1688                 case INTEL_PT_BAD:
1689                 case INTEL_PT_PSB:
1690                 case INTEL_PT_PTWRITE:
1691                 case INTEL_PT_PTWRITE_IP:
1692                 case INTEL_PT_EXSTOP:
1693                 case INTEL_PT_EXSTOP_IP:
1694                 case INTEL_PT_MWAIT:
1695                 case INTEL_PT_PWRE:
1696                 case INTEL_PT_PWRX:
1697                 case INTEL_PT_BBP:
1698                 case INTEL_PT_BIP:
1699                 case INTEL_PT_BEP:
1700                 case INTEL_PT_BEP_IP:
1701                         decoder->have_tma = false;
1702                         intel_pt_log("ERROR: Unexpected packet\n");
1703                         err = -EAGAIN;
1704                         goto out;
1705
1706                 case INTEL_PT_OVF:
1707                         err = intel_pt_overflow(decoder);
1708                         goto out;
1709
1710                 case INTEL_PT_TSC:
1711                         intel_pt_calc_tsc_timestamp(decoder);
1712                         break;
1713
1714                 case INTEL_PT_TMA:
1715                         intel_pt_calc_tma(decoder);
1716                         break;
1717
1718                 case INTEL_PT_CBR:
1719                         intel_pt_calc_cbr(decoder);
1720                         break;
1721
1722                 case INTEL_PT_MODE_EXEC:
1723                         decoder->exec_mode = decoder->packet.payload;
1724                         break;
1725
1726                 case INTEL_PT_PIP:
1727                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1728                         break;
1729
1730                 case INTEL_PT_FUP:
1731                         decoder->pge = true;
1732                         if (decoder->packet.count)
1733                                 intel_pt_set_last_ip(decoder);
1734                         break;
1735
1736                 case INTEL_PT_MODE_TSX:
1737                         intel_pt_update_in_tx(decoder);
1738                         break;
1739
1740                 case INTEL_PT_MTC:
1741                         intel_pt_calc_mtc_timestamp(decoder);
1742                         if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1743                                 decoder->state.type |= INTEL_PT_INSTRUCTION;
1744                         break;
1745
1746                 case INTEL_PT_CYC:
1747                 case INTEL_PT_VMCS:
1748                 case INTEL_PT_MNT:
1749                 case INTEL_PT_PAD:
1750                 default:
1751                         break;
1752                 }
1753         }
1754 out:
1755         decoder->in_psb = false;
1756
1757         return err;
1758 }
1759
1760 static int intel_pt_walk_fup_tip(struct intel_pt_decoder *decoder)
1761 {
1762         int err;
1763
1764         if (decoder->tx_flags & INTEL_PT_ABORT_TX) {
1765                 decoder->tx_flags = 0;
1766                 decoder->state.flags &= ~INTEL_PT_IN_TX;
1767                 decoder->state.flags |= INTEL_PT_ABORT_TX;
1768         } else {
1769                 decoder->state.flags |= INTEL_PT_ASYNC;
1770         }
1771
1772         while (1) {
1773                 err = intel_pt_get_next_packet(decoder);
1774                 if (err)
1775                         return err;
1776
1777                 switch (decoder->packet.type) {
1778                 case INTEL_PT_TNT:
1779                 case INTEL_PT_FUP:
1780                 case INTEL_PT_TRACESTOP:
1781                 case INTEL_PT_PSB:
1782                 case INTEL_PT_TSC:
1783                 case INTEL_PT_TMA:
1784                 case INTEL_PT_MODE_TSX:
1785                 case INTEL_PT_BAD:
1786                 case INTEL_PT_PSBEND:
1787                 case INTEL_PT_PTWRITE:
1788                 case INTEL_PT_PTWRITE_IP:
1789                 case INTEL_PT_EXSTOP:
1790                 case INTEL_PT_EXSTOP_IP:
1791                 case INTEL_PT_MWAIT:
1792                 case INTEL_PT_PWRE:
1793                 case INTEL_PT_PWRX:
1794                 case INTEL_PT_BBP:
1795                 case INTEL_PT_BIP:
1796                 case INTEL_PT_BEP:
1797                 case INTEL_PT_BEP_IP:
1798                         intel_pt_log("ERROR: Missing TIP after FUP\n");
1799                         decoder->pkt_state = INTEL_PT_STATE_ERR3;
1800                         decoder->pkt_step = 0;
1801                         return -ENOENT;
1802
1803                 case INTEL_PT_CBR:
1804                         intel_pt_calc_cbr(decoder);
1805                         break;
1806
1807                 case INTEL_PT_OVF:
1808                         return intel_pt_overflow(decoder);
1809
1810                 case INTEL_PT_TIP_PGD:
1811                         decoder->state.from_ip = decoder->ip;
1812                         if (decoder->packet.count == 0) {
1813                                 decoder->state.to_ip = 0;
1814                         } else {
1815                                 intel_pt_set_ip(decoder);
1816                                 decoder->state.to_ip = decoder->ip;
1817                         }
1818                         decoder->pge = false;
1819                         decoder->continuous_period = false;
1820                         decoder->state.type |= INTEL_PT_TRACE_END;
1821                         return 0;
1822
1823                 case INTEL_PT_TIP_PGE:
1824                         decoder->pge = true;
1825                         intel_pt_log("Omitting PGE ip " x64_fmt "\n",
1826                                      decoder->ip);
1827                         decoder->state.from_ip = 0;
1828                         if (decoder->packet.count == 0) {
1829                                 decoder->state.to_ip = 0;
1830                         } else {
1831                                 intel_pt_set_ip(decoder);
1832                                 decoder->state.to_ip = decoder->ip;
1833                         }
1834                         decoder->state.type |= INTEL_PT_TRACE_BEGIN;
1835                         intel_pt_mtc_cyc_cnt_pge(decoder);
1836                         return 0;
1837
1838                 case INTEL_PT_TIP:
1839                         decoder->state.from_ip = decoder->ip;
1840                         if (decoder->packet.count == 0) {
1841                                 decoder->state.to_ip = 0;
1842                         } else {
1843                                 intel_pt_set_ip(decoder);
1844                                 decoder->state.to_ip = decoder->ip;
1845                         }
1846                         return 0;
1847
1848                 case INTEL_PT_PIP:
1849                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1850                         break;
1851
1852                 case INTEL_PT_MTC:
1853                         intel_pt_calc_mtc_timestamp(decoder);
1854                         if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1855                                 decoder->state.type |= INTEL_PT_INSTRUCTION;
1856                         break;
1857
1858                 case INTEL_PT_CYC:
1859                         intel_pt_calc_cyc_timestamp(decoder);
1860                         break;
1861
1862                 case INTEL_PT_MODE_EXEC:
1863                         decoder->exec_mode = decoder->packet.payload;
1864                         break;
1865
1866                 case INTEL_PT_VMCS:
1867                 case INTEL_PT_MNT:
1868                 case INTEL_PT_PAD:
1869                         break;
1870
1871                 default:
1872                         return intel_pt_bug(decoder);
1873                 }
1874         }
1875 }
1876
1877 static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)
1878 {
1879         bool no_tip = false;
1880         int err;
1881
1882         while (1) {
1883                 err = intel_pt_get_next_packet(decoder);
1884                 if (err)
1885                         return err;
1886 next:
1887                 switch (decoder->packet.type) {
1888                 case INTEL_PT_TNT:
1889                         if (!decoder->packet.count)
1890                                 break;
1891                         decoder->tnt = decoder->packet;
1892                         decoder->pkt_state = INTEL_PT_STATE_TNT;
1893                         err = intel_pt_walk_tnt(decoder);
1894                         if (err == -EAGAIN)
1895                                 break;
1896                         return err;
1897
1898                 case INTEL_PT_TIP_PGD:
1899                         if (decoder->packet.count != 0)
1900                                 intel_pt_set_last_ip(decoder);
1901                         decoder->pkt_state = INTEL_PT_STATE_TIP_PGD;
1902                         return intel_pt_walk_tip(decoder);
1903
1904                 case INTEL_PT_TIP_PGE: {
1905                         decoder->pge = true;
1906                         intel_pt_mtc_cyc_cnt_pge(decoder);
1907                         if (decoder->packet.count == 0) {
1908                                 intel_pt_log_at("Skipping zero TIP.PGE",
1909                                                 decoder->pos);
1910                                 break;
1911                         }
1912                         intel_pt_set_ip(decoder);
1913                         decoder->state.from_ip = 0;
1914                         decoder->state.to_ip = decoder->ip;
1915                         decoder->state.type |= INTEL_PT_TRACE_BEGIN;
1916                         return 0;
1917                 }
1918
1919                 case INTEL_PT_OVF:
1920                         return intel_pt_overflow(decoder);
1921
1922                 case INTEL_PT_TIP:
1923                         if (decoder->packet.count != 0)
1924                                 intel_pt_set_last_ip(decoder);
1925                         decoder->pkt_state = INTEL_PT_STATE_TIP;
1926                         return intel_pt_walk_tip(decoder);
1927
1928                 case INTEL_PT_FUP:
1929                         if (decoder->packet.count == 0) {
1930                                 intel_pt_log_at("Skipping zero FUP",
1931                                                 decoder->pos);
1932                                 no_tip = false;
1933                                 break;
1934                         }
1935                         intel_pt_set_last_ip(decoder);
1936                         if (!decoder->branch_enable) {
1937                                 decoder->ip = decoder->last_ip;
1938                                 if (intel_pt_fup_event(decoder))
1939                                         return 0;
1940                                 no_tip = false;
1941                                 break;
1942                         }
1943                         if (decoder->set_fup_mwait)
1944                                 no_tip = true;
1945                         err = intel_pt_walk_fup(decoder);
1946                         if (err != -EAGAIN) {
1947                                 if (err)
1948                                         return err;
1949                                 if (no_tip)
1950                                         decoder->pkt_state =
1951                                                 INTEL_PT_STATE_FUP_NO_TIP;
1952                                 else
1953                                         decoder->pkt_state = INTEL_PT_STATE_FUP;
1954                                 return 0;
1955                         }
1956                         if (no_tip) {
1957                                 no_tip = false;
1958                                 break;
1959                         }
1960                         return intel_pt_walk_fup_tip(decoder);
1961
1962                 case INTEL_PT_TRACESTOP:
1963                         decoder->pge = false;
1964                         decoder->continuous_period = false;
1965                         intel_pt_clear_tx_flags(decoder);
1966                         decoder->have_tma = false;
1967                         break;
1968
1969                 case INTEL_PT_PSB:
1970                         decoder->last_ip = 0;
1971                         decoder->have_last_ip = true;
1972                         intel_pt_clear_stack(&decoder->stack);
1973                         err = intel_pt_walk_psbend(decoder);
1974                         if (err == -EAGAIN)
1975                                 goto next;
1976                         if (err)
1977                                 return err;
1978                         /*
1979                          * PSB+ CBR will not have changed but cater for the
1980                          * possibility of another CBR change that gets caught up
1981                          * in the PSB+.
1982                          */
1983                         if (decoder->cbr != decoder->cbr_seen)
1984                                 return 0;
1985                         break;
1986
1987                 case INTEL_PT_PIP:
1988                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1989                         break;
1990
1991                 case INTEL_PT_MTC:
1992                         intel_pt_calc_mtc_timestamp(decoder);
1993                         if (decoder->period_type != INTEL_PT_PERIOD_MTC)
1994                                 break;
1995                         /*
1996                          * Ensure that there has been an instruction since the
1997                          * last MTC.
1998                          */
1999                         if (!decoder->mtc_insn)
2000                                 break;
2001                         decoder->mtc_insn = false;
2002                         /* Ensure that there is a timestamp */
2003                         if (!decoder->timestamp)
2004                                 break;
2005                         decoder->state.type = INTEL_PT_INSTRUCTION;
2006                         decoder->state.from_ip = decoder->ip;
2007                         decoder->state.to_ip = 0;
2008                         decoder->mtc_insn = false;
2009                         return 0;
2010
2011                 case INTEL_PT_TSC:
2012                         intel_pt_calc_tsc_timestamp(decoder);
2013                         break;
2014
2015                 case INTEL_PT_TMA:
2016                         intel_pt_calc_tma(decoder);
2017                         break;
2018
2019                 case INTEL_PT_CYC:
2020                         intel_pt_calc_cyc_timestamp(decoder);
2021                         break;
2022
2023                 case INTEL_PT_CBR:
2024                         intel_pt_calc_cbr(decoder);
2025                         if (decoder->cbr != decoder->cbr_seen)
2026                                 return 0;
2027                         break;
2028
2029                 case INTEL_PT_MODE_EXEC:
2030                         decoder->exec_mode = decoder->packet.payload;
2031                         break;
2032
2033                 case INTEL_PT_MODE_TSX:
2034                         /* MODE_TSX need not be followed by FUP */
2035                         if (!decoder->pge) {
2036                                 intel_pt_update_in_tx(decoder);
2037                                 break;
2038                         }
2039                         err = intel_pt_mode_tsx(decoder, &no_tip);
2040                         if (err)
2041                                 return err;
2042                         goto next;
2043
2044                 case INTEL_PT_BAD: /* Does not happen */
2045                         return intel_pt_bug(decoder);
2046
2047                 case INTEL_PT_PSBEND:
2048                 case INTEL_PT_VMCS:
2049                 case INTEL_PT_MNT:
2050                 case INTEL_PT_PAD:
2051                         break;
2052
2053                 case INTEL_PT_PTWRITE_IP:
2054                         decoder->fup_ptw_payload = decoder->packet.payload;
2055                         err = intel_pt_get_next_packet(decoder);
2056                         if (err)
2057                                 return err;
2058                         if (decoder->packet.type == INTEL_PT_FUP) {
2059                                 decoder->set_fup_ptw = true;
2060                                 no_tip = true;
2061                         } else {
2062                                 intel_pt_log_at("ERROR: Missing FUP after PTWRITE",
2063                                                 decoder->pos);
2064                         }
2065                         goto next;
2066
2067                 case INTEL_PT_PTWRITE:
2068                         decoder->state.type = INTEL_PT_PTW;
2069                         decoder->state.from_ip = decoder->ip;
2070                         decoder->state.to_ip = 0;
2071                         decoder->state.ptw_payload = decoder->packet.payload;
2072                         return 0;
2073
2074                 case INTEL_PT_MWAIT:
2075                         decoder->fup_mwait_payload = decoder->packet.payload;
2076                         decoder->set_fup_mwait = true;
2077                         break;
2078
2079                 case INTEL_PT_PWRE:
2080                         if (decoder->set_fup_mwait) {
2081                                 decoder->fup_pwre_payload =
2082                                                         decoder->packet.payload;
2083                                 decoder->set_fup_pwre = true;
2084                                 break;
2085                         }
2086                         decoder->state.type = INTEL_PT_PWR_ENTRY;
2087                         decoder->state.from_ip = decoder->ip;
2088                         decoder->state.to_ip = 0;
2089                         decoder->state.pwrx_payload = decoder->packet.payload;
2090                         return 0;
2091
2092                 case INTEL_PT_EXSTOP_IP:
2093                         err = intel_pt_get_next_packet(decoder);
2094                         if (err)
2095                                 return err;
2096                         if (decoder->packet.type == INTEL_PT_FUP) {
2097                                 decoder->set_fup_exstop = true;
2098                                 no_tip = true;
2099                         } else {
2100                                 intel_pt_log_at("ERROR: Missing FUP after EXSTOP",
2101                                                 decoder->pos);
2102                         }
2103                         goto next;
2104
2105                 case INTEL_PT_EXSTOP:
2106                         decoder->state.type = INTEL_PT_EX_STOP;
2107                         decoder->state.from_ip = decoder->ip;
2108                         decoder->state.to_ip = 0;
2109                         return 0;
2110
2111                 case INTEL_PT_PWRX:
2112                         decoder->state.type = INTEL_PT_PWR_EXIT;
2113                         decoder->state.from_ip = decoder->ip;
2114                         decoder->state.to_ip = 0;
2115                         decoder->state.pwrx_payload = decoder->packet.payload;
2116                         return 0;
2117
2118                 case INTEL_PT_BBP:
2119                         intel_pt_bbp(decoder);
2120                         break;
2121
2122                 case INTEL_PT_BIP:
2123                         intel_pt_bip(decoder);
2124                         break;
2125
2126                 case INTEL_PT_BEP:
2127                         decoder->state.type = INTEL_PT_BLK_ITEMS;
2128                         decoder->state.from_ip = decoder->ip;
2129                         decoder->state.to_ip = 0;
2130                         return 0;
2131
2132                 case INTEL_PT_BEP_IP:
2133                         err = intel_pt_get_next_packet(decoder);
2134                         if (err)
2135                                 return err;
2136                         if (decoder->packet.type == INTEL_PT_FUP) {
2137                                 decoder->set_fup_bep = true;
2138                                 no_tip = true;
2139                         } else {
2140                                 intel_pt_log_at("ERROR: Missing FUP after BEP",
2141                                                 decoder->pos);
2142                         }
2143                         goto next;
2144
2145                 default:
2146                         return intel_pt_bug(decoder);
2147                 }
2148         }
2149 }
2150
2151 static inline bool intel_pt_have_ip(struct intel_pt_decoder *decoder)
2152 {
2153         return decoder->packet.count &&
2154                (decoder->have_last_ip || decoder->packet.count == 3 ||
2155                 decoder->packet.count == 6);
2156 }
2157
2158 /* Walk PSB+ packets to get in sync. */
2159 static int intel_pt_walk_psb(struct intel_pt_decoder *decoder)
2160 {
2161         int err;
2162
2163         decoder->in_psb = true;
2164
2165         while (1) {
2166                 err = intel_pt_get_next_packet(decoder);
2167                 if (err)
2168                         goto out;
2169
2170                 switch (decoder->packet.type) {
2171                 case INTEL_PT_TIP_PGD:
2172                         decoder->continuous_period = false;
2173                         __fallthrough;
2174                 case INTEL_PT_TIP_PGE:
2175                 case INTEL_PT_TIP:
2176                 case INTEL_PT_PTWRITE:
2177                 case INTEL_PT_PTWRITE_IP:
2178                 case INTEL_PT_EXSTOP:
2179                 case INTEL_PT_EXSTOP_IP:
2180                 case INTEL_PT_MWAIT:
2181                 case INTEL_PT_PWRE:
2182                 case INTEL_PT_PWRX:
2183                 case INTEL_PT_BBP:
2184                 case INTEL_PT_BIP:
2185                 case INTEL_PT_BEP:
2186                 case INTEL_PT_BEP_IP:
2187                         intel_pt_log("ERROR: Unexpected packet\n");
2188                         err = -ENOENT;
2189                         goto out;
2190
2191                 case INTEL_PT_FUP:
2192                         decoder->pge = true;
2193                         if (intel_pt_have_ip(decoder)) {
2194                                 uint64_t current_ip = decoder->ip;
2195
2196                                 intel_pt_set_ip(decoder);
2197                                 if (current_ip)
2198                                         intel_pt_log_to("Setting IP",
2199                                                         decoder->ip);
2200                         }
2201                         break;
2202
2203                 case INTEL_PT_MTC:
2204                         intel_pt_calc_mtc_timestamp(decoder);
2205                         break;
2206
2207                 case INTEL_PT_TSC:
2208                         intel_pt_calc_tsc_timestamp(decoder);
2209                         break;
2210
2211                 case INTEL_PT_TMA:
2212                         intel_pt_calc_tma(decoder);
2213                         break;
2214
2215                 case INTEL_PT_CYC:
2216                         intel_pt_calc_cyc_timestamp(decoder);
2217                         break;
2218
2219                 case INTEL_PT_CBR:
2220                         intel_pt_calc_cbr(decoder);
2221                         break;
2222
2223                 case INTEL_PT_PIP:
2224                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2225                         break;
2226
2227                 case INTEL_PT_MODE_EXEC:
2228                         decoder->exec_mode = decoder->packet.payload;
2229                         break;
2230
2231                 case INTEL_PT_MODE_TSX:
2232                         intel_pt_update_in_tx(decoder);
2233                         break;
2234
2235                 case INTEL_PT_TRACESTOP:
2236                         decoder->pge = false;
2237                         decoder->continuous_period = false;
2238                         intel_pt_clear_tx_flags(decoder);
2239                         __fallthrough;
2240
2241                 case INTEL_PT_TNT:
2242                         decoder->have_tma = false;
2243                         intel_pt_log("ERROR: Unexpected packet\n");
2244                         if (decoder->ip)
2245                                 decoder->pkt_state = INTEL_PT_STATE_ERR4;
2246                         else
2247                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
2248                         err = -ENOENT;
2249                         goto out;
2250
2251                 case INTEL_PT_BAD: /* Does not happen */
2252                         err = intel_pt_bug(decoder);
2253                         goto out;
2254
2255                 case INTEL_PT_OVF:
2256                         err = intel_pt_overflow(decoder);
2257                         goto out;
2258
2259                 case INTEL_PT_PSBEND:
2260                         err = 0;
2261                         goto out;
2262
2263                 case INTEL_PT_PSB:
2264                 case INTEL_PT_VMCS:
2265                 case INTEL_PT_MNT:
2266                 case INTEL_PT_PAD:
2267                 default:
2268                         break;
2269                 }
2270         }
2271 out:
2272         decoder->in_psb = false;
2273
2274         return err;
2275 }
2276
2277 static int intel_pt_walk_to_ip(struct intel_pt_decoder *decoder)
2278 {
2279         int err;
2280
2281         while (1) {
2282                 err = intel_pt_get_next_packet(decoder);
2283                 if (err)
2284                         return err;
2285
2286                 switch (decoder->packet.type) {
2287                 case INTEL_PT_TIP_PGD:
2288                         decoder->continuous_period = false;
2289                         decoder->pge = false;
2290                         if (intel_pt_have_ip(decoder))
2291                                 intel_pt_set_ip(decoder);
2292                         if (!decoder->ip)
2293                                 break;
2294                         decoder->state.type |= INTEL_PT_TRACE_END;
2295                         return 0;
2296
2297                 case INTEL_PT_TIP_PGE:
2298                         decoder->pge = true;
2299                         intel_pt_mtc_cyc_cnt_pge(decoder);
2300                         if (intel_pt_have_ip(decoder))
2301                                 intel_pt_set_ip(decoder);
2302                         if (!decoder->ip)
2303                                 break;
2304                         decoder->state.type |= INTEL_PT_TRACE_BEGIN;
2305                         return 0;
2306
2307                 case INTEL_PT_TIP:
2308                         decoder->pge = true;
2309                         if (intel_pt_have_ip(decoder))
2310                                 intel_pt_set_ip(decoder);
2311                         if (!decoder->ip)
2312                                 break;
2313                         return 0;
2314
2315                 case INTEL_PT_FUP:
2316                         if (intel_pt_have_ip(decoder))
2317                                 intel_pt_set_ip(decoder);
2318                         if (decoder->ip)
2319                                 return 0;
2320                         break;
2321
2322                 case INTEL_PT_MTC:
2323                         intel_pt_calc_mtc_timestamp(decoder);
2324                         break;
2325
2326                 case INTEL_PT_TSC:
2327                         intel_pt_calc_tsc_timestamp(decoder);
2328                         break;
2329
2330                 case INTEL_PT_TMA:
2331                         intel_pt_calc_tma(decoder);
2332                         break;
2333
2334                 case INTEL_PT_CYC:
2335                         intel_pt_calc_cyc_timestamp(decoder);
2336                         break;
2337
2338                 case INTEL_PT_CBR:
2339                         intel_pt_calc_cbr(decoder);
2340                         break;
2341
2342                 case INTEL_PT_PIP:
2343                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2344                         break;
2345
2346                 case INTEL_PT_MODE_EXEC:
2347                         decoder->exec_mode = decoder->packet.payload;
2348                         break;
2349
2350                 case INTEL_PT_MODE_TSX:
2351                         intel_pt_update_in_tx(decoder);
2352                         break;
2353
2354                 case INTEL_PT_OVF:
2355                         return intel_pt_overflow(decoder);
2356
2357                 case INTEL_PT_BAD: /* Does not happen */
2358                         return intel_pt_bug(decoder);
2359
2360                 case INTEL_PT_TRACESTOP:
2361                         decoder->pge = false;
2362                         decoder->continuous_period = false;
2363                         intel_pt_clear_tx_flags(decoder);
2364                         decoder->have_tma = false;
2365                         break;
2366
2367                 case INTEL_PT_PSB:
2368                         decoder->last_ip = 0;
2369                         decoder->have_last_ip = true;
2370                         intel_pt_clear_stack(&decoder->stack);
2371                         err = intel_pt_walk_psb(decoder);
2372                         if (err)
2373                                 return err;
2374                         if (decoder->ip) {
2375                                 /* Do not have a sample */
2376                                 decoder->state.type = 0;
2377                                 return 0;
2378                         }
2379                         break;
2380
2381                 case INTEL_PT_TNT:
2382                 case INTEL_PT_PSBEND:
2383                 case INTEL_PT_VMCS:
2384                 case INTEL_PT_MNT:
2385                 case INTEL_PT_PAD:
2386                 case INTEL_PT_PTWRITE:
2387                 case INTEL_PT_PTWRITE_IP:
2388                 case INTEL_PT_EXSTOP:
2389                 case INTEL_PT_EXSTOP_IP:
2390                 case INTEL_PT_MWAIT:
2391                 case INTEL_PT_PWRE:
2392                 case INTEL_PT_PWRX:
2393                 case INTEL_PT_BBP:
2394                 case INTEL_PT_BIP:
2395                 case INTEL_PT_BEP:
2396                 case INTEL_PT_BEP_IP:
2397                 default:
2398                         break;
2399                 }
2400         }
2401 }
2402
2403 static int intel_pt_sync_ip(struct intel_pt_decoder *decoder)
2404 {
2405         int err;
2406
2407         decoder->set_fup_tx_flags = false;
2408         decoder->set_fup_ptw = false;
2409         decoder->set_fup_mwait = false;
2410         decoder->set_fup_pwre = false;
2411         decoder->set_fup_exstop = false;
2412         decoder->set_fup_bep = false;
2413
2414         if (!decoder->branch_enable) {
2415                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2416                 decoder->overflow = false;
2417                 decoder->state.type = 0; /* Do not have a sample */
2418                 return 0;
2419         }
2420
2421         intel_pt_log("Scanning for full IP\n");
2422         err = intel_pt_walk_to_ip(decoder);
2423         if (err)
2424                 return err;
2425
2426         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2427         decoder->overflow = false;
2428
2429         decoder->state.from_ip = 0;
2430         decoder->state.to_ip = decoder->ip;
2431         intel_pt_log_to("Setting IP", decoder->ip);
2432
2433         return 0;
2434 }
2435
2436 static int intel_pt_part_psb(struct intel_pt_decoder *decoder)
2437 {
2438         const unsigned char *end = decoder->buf + decoder->len;
2439         size_t i;
2440
2441         for (i = INTEL_PT_PSB_LEN - 1; i; i--) {
2442                 if (i > decoder->len)
2443                         continue;
2444                 if (!memcmp(end - i, INTEL_PT_PSB_STR, i))
2445                         return i;
2446         }
2447         return 0;
2448 }
2449
2450 static int intel_pt_rest_psb(struct intel_pt_decoder *decoder, int part_psb)
2451 {
2452         size_t rest_psb = INTEL_PT_PSB_LEN - part_psb;
2453         const char *psb = INTEL_PT_PSB_STR;
2454
2455         if (rest_psb > decoder->len ||
2456             memcmp(decoder->buf, psb + part_psb, rest_psb))
2457                 return 0;
2458
2459         return rest_psb;
2460 }
2461
2462 static int intel_pt_get_split_psb(struct intel_pt_decoder *decoder,
2463                                   int part_psb)
2464 {
2465         int rest_psb, ret;
2466
2467         decoder->pos += decoder->len;
2468         decoder->len = 0;
2469
2470         ret = intel_pt_get_next_data(decoder, false);
2471         if (ret)
2472                 return ret;
2473
2474         rest_psb = intel_pt_rest_psb(decoder, part_psb);
2475         if (!rest_psb)
2476                 return 0;
2477
2478         decoder->pos -= part_psb;
2479         decoder->next_buf = decoder->buf + rest_psb;
2480         decoder->next_len = decoder->len - rest_psb;
2481         memcpy(decoder->temp_buf, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2482         decoder->buf = decoder->temp_buf;
2483         decoder->len = INTEL_PT_PSB_LEN;
2484
2485         return 0;
2486 }
2487
2488 static int intel_pt_scan_for_psb(struct intel_pt_decoder *decoder)
2489 {
2490         unsigned char *next;
2491         int ret;
2492
2493         intel_pt_log("Scanning for PSB\n");
2494         while (1) {
2495                 if (!decoder->len) {
2496                         ret = intel_pt_get_next_data(decoder, false);
2497                         if (ret)
2498                                 return ret;
2499                 }
2500
2501                 next = memmem(decoder->buf, decoder->len, INTEL_PT_PSB_STR,
2502                               INTEL_PT_PSB_LEN);
2503                 if (!next) {
2504                         int part_psb;
2505
2506                         part_psb = intel_pt_part_psb(decoder);
2507                         if (part_psb) {
2508                                 ret = intel_pt_get_split_psb(decoder, part_psb);
2509                                 if (ret)
2510                                         return ret;
2511                         } else {
2512                                 decoder->pos += decoder->len;
2513                                 decoder->len = 0;
2514                         }
2515                         continue;
2516                 }
2517
2518                 decoder->pkt_step = next - decoder->buf;
2519                 return intel_pt_get_next_packet(decoder);
2520         }
2521 }
2522
2523 static int intel_pt_sync(struct intel_pt_decoder *decoder)
2524 {
2525         int err;
2526
2527         decoder->pge = false;
2528         decoder->continuous_period = false;
2529         decoder->have_last_ip = false;
2530         decoder->last_ip = 0;
2531         decoder->ip = 0;
2532         intel_pt_clear_stack(&decoder->stack);
2533
2534         err = intel_pt_scan_for_psb(decoder);
2535         if (err)
2536                 return err;
2537
2538         decoder->have_last_ip = true;
2539         decoder->pkt_state = INTEL_PT_STATE_NO_IP;
2540
2541         err = intel_pt_walk_psb(decoder);
2542         if (err)
2543                 return err;
2544
2545         if (decoder->ip) {
2546                 decoder->state.type = 0; /* Do not have a sample */
2547                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2548         } else {
2549                 return intel_pt_sync_ip(decoder);
2550         }
2551
2552         return 0;
2553 }
2554
2555 static uint64_t intel_pt_est_timestamp(struct intel_pt_decoder *decoder)
2556 {
2557         uint64_t est = decoder->sample_insn_cnt << 1;
2558
2559         if (!decoder->cbr || !decoder->max_non_turbo_ratio)
2560                 goto out;
2561
2562         est *= decoder->max_non_turbo_ratio;
2563         est /= decoder->cbr;
2564 out:
2565         return decoder->sample_timestamp + est;
2566 }
2567
2568 const struct intel_pt_state *intel_pt_decode(struct intel_pt_decoder *decoder)
2569 {
2570         int err;
2571
2572         do {
2573                 decoder->state.type = INTEL_PT_BRANCH;
2574                 decoder->state.flags = 0;
2575
2576                 switch (decoder->pkt_state) {
2577                 case INTEL_PT_STATE_NO_PSB:
2578                         err = intel_pt_sync(decoder);
2579                         break;
2580                 case INTEL_PT_STATE_NO_IP:
2581                         decoder->have_last_ip = false;
2582                         decoder->last_ip = 0;
2583                         decoder->ip = 0;
2584                         __fallthrough;
2585                 case INTEL_PT_STATE_ERR_RESYNC:
2586                         err = intel_pt_sync_ip(decoder);
2587                         break;
2588                 case INTEL_PT_STATE_IN_SYNC:
2589                         err = intel_pt_walk_trace(decoder);
2590                         break;
2591                 case INTEL_PT_STATE_TNT:
2592                 case INTEL_PT_STATE_TNT_CONT:
2593                         err = intel_pt_walk_tnt(decoder);
2594                         if (err == -EAGAIN)
2595                                 err = intel_pt_walk_trace(decoder);
2596                         break;
2597                 case INTEL_PT_STATE_TIP:
2598                 case INTEL_PT_STATE_TIP_PGD:
2599                         err = intel_pt_walk_tip(decoder);
2600                         break;
2601                 case INTEL_PT_STATE_FUP:
2602                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2603                         err = intel_pt_walk_fup(decoder);
2604                         if (err == -EAGAIN)
2605                                 err = intel_pt_walk_fup_tip(decoder);
2606                         else if (!err)
2607                                 decoder->pkt_state = INTEL_PT_STATE_FUP;
2608                         break;
2609                 case INTEL_PT_STATE_FUP_NO_TIP:
2610                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2611                         err = intel_pt_walk_fup(decoder);
2612                         if (err == -EAGAIN)
2613                                 err = intel_pt_walk_trace(decoder);
2614                         break;
2615                 default:
2616                         err = intel_pt_bug(decoder);
2617                         break;
2618                 }
2619         } while (err == -ENOLINK);
2620
2621         if (err) {
2622                 decoder->state.err = intel_pt_ext_err(err);
2623                 decoder->state.from_ip = decoder->ip;
2624                 intel_pt_update_sample_time(decoder);
2625                 decoder->sample_tot_cyc_cnt = decoder->tot_cyc_cnt;
2626         } else {
2627                 decoder->state.err = 0;
2628                 if (decoder->cbr != decoder->cbr_seen) {
2629                         decoder->cbr_seen = decoder->cbr;
2630                         if (!decoder->state.type) {
2631                                 decoder->state.from_ip = decoder->ip;
2632                                 decoder->state.to_ip = 0;
2633                         }
2634                         decoder->state.type |= INTEL_PT_CBR_CHG;
2635                         decoder->state.cbr_payload = decoder->cbr_payload;
2636                         decoder->state.cbr = decoder->cbr;
2637                 }
2638                 if (intel_pt_sample_time(decoder->pkt_state)) {
2639                         intel_pt_update_sample_time(decoder);
2640                         if (decoder->sample_cyc)
2641                                 decoder->sample_tot_cyc_cnt = decoder->tot_cyc_cnt;
2642                 }
2643         }
2644
2645         decoder->state.timestamp = decoder->sample_timestamp;
2646         decoder->state.est_timestamp = intel_pt_est_timestamp(decoder);
2647         decoder->state.cr3 = decoder->cr3;
2648         decoder->state.tot_insn_cnt = decoder->tot_insn_cnt;
2649         decoder->state.tot_cyc_cnt = decoder->sample_tot_cyc_cnt;
2650
2651         return &decoder->state;
2652 }
2653
2654 /**
2655  * intel_pt_next_psb - move buffer pointer to the start of the next PSB packet.
2656  * @buf: pointer to buffer pointer
2657  * @len: size of buffer
2658  *
2659  * Updates the buffer pointer to point to the start of the next PSB packet if
2660  * there is one, otherwise the buffer pointer is unchanged.  If @buf is updated,
2661  * @len is adjusted accordingly.
2662  *
2663  * Return: %true if a PSB packet is found, %false otherwise.
2664  */
2665 static bool intel_pt_next_psb(unsigned char **buf, size_t *len)
2666 {
2667         unsigned char *next;
2668
2669         next = memmem(*buf, *len, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2670         if (next) {
2671                 *len -= next - *buf;
2672                 *buf = next;
2673                 return true;
2674         }
2675         return false;
2676 }
2677
2678 /**
2679  * intel_pt_step_psb - move buffer pointer to the start of the following PSB
2680  *                     packet.
2681  * @buf: pointer to buffer pointer
2682  * @len: size of buffer
2683  *
2684  * Updates the buffer pointer to point to the start of the following PSB packet
2685  * (skipping the PSB at @buf itself) if there is one, otherwise the buffer
2686  * pointer is unchanged.  If @buf is updated, @len is adjusted accordingly.
2687  *
2688  * Return: %true if a PSB packet is found, %false otherwise.
2689  */
2690 static bool intel_pt_step_psb(unsigned char **buf, size_t *len)
2691 {
2692         unsigned char *next;
2693
2694         if (!*len)
2695                 return false;
2696
2697         next = memmem(*buf + 1, *len - 1, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2698         if (next) {
2699                 *len -= next - *buf;
2700                 *buf = next;
2701                 return true;
2702         }
2703         return false;
2704 }
2705
2706 /**
2707  * intel_pt_last_psb - find the last PSB packet in a buffer.
2708  * @buf: buffer
2709  * @len: size of buffer
2710  *
2711  * This function finds the last PSB in a buffer.
2712  *
2713  * Return: A pointer to the last PSB in @buf if found, %NULL otherwise.
2714  */
2715 static unsigned char *intel_pt_last_psb(unsigned char *buf, size_t len)
2716 {
2717         const char *n = INTEL_PT_PSB_STR;
2718         unsigned char *p;
2719         size_t k;
2720
2721         if (len < INTEL_PT_PSB_LEN)
2722                 return NULL;
2723
2724         k = len - INTEL_PT_PSB_LEN + 1;
2725         while (1) {
2726                 p = memrchr(buf, n[0], k);
2727                 if (!p)
2728                         return NULL;
2729                 if (!memcmp(p + 1, n + 1, INTEL_PT_PSB_LEN - 1))
2730                         return p;
2731                 k = p - buf;
2732                 if (!k)
2733                         return NULL;
2734         }
2735 }
2736
2737 /**
2738  * intel_pt_next_tsc - find and return next TSC.
2739  * @buf: buffer
2740  * @len: size of buffer
2741  * @tsc: TSC value returned
2742  * @rem: returns remaining size when TSC is found
2743  *
2744  * Find a TSC packet in @buf and return the TSC value.  This function assumes
2745  * that @buf starts at a PSB and that PSB+ will contain TSC and so stops if a
2746  * PSBEND packet is found.
2747  *
2748  * Return: %true if TSC is found, false otherwise.
2749  */
2750 static bool intel_pt_next_tsc(unsigned char *buf, size_t len, uint64_t *tsc,
2751                               size_t *rem)
2752 {
2753         enum intel_pt_pkt_ctx ctx = INTEL_PT_NO_CTX;
2754         struct intel_pt_pkt packet;
2755         int ret;
2756
2757         while (len) {
2758                 ret = intel_pt_get_packet(buf, len, &packet, &ctx);
2759                 if (ret <= 0)
2760                         return false;
2761                 if (packet.type == INTEL_PT_TSC) {
2762                         *tsc = packet.payload;
2763                         *rem = len;
2764                         return true;
2765                 }
2766                 if (packet.type == INTEL_PT_PSBEND)
2767                         return false;
2768                 buf += ret;
2769                 len -= ret;
2770         }
2771         return false;
2772 }
2773
2774 /**
2775  * intel_pt_tsc_cmp - compare 7-byte TSCs.
2776  * @tsc1: first TSC to compare
2777  * @tsc2: second TSC to compare
2778  *
2779  * This function compares 7-byte TSC values allowing for the possibility that
2780  * TSC wrapped around.  Generally it is not possible to know if TSC has wrapped
2781  * around so for that purpose this function assumes the absolute difference is
2782  * less than half the maximum difference.
2783  *
2784  * Return: %-1 if @tsc1 is before @tsc2, %0 if @tsc1 == @tsc2, %1 if @tsc1 is
2785  * after @tsc2.
2786  */
2787 static int intel_pt_tsc_cmp(uint64_t tsc1, uint64_t tsc2)
2788 {
2789         const uint64_t halfway = (1ULL << 55);
2790
2791         if (tsc1 == tsc2)
2792                 return 0;
2793
2794         if (tsc1 < tsc2) {
2795                 if (tsc2 - tsc1 < halfway)
2796                         return -1;
2797                 else
2798                         return 1;
2799         } else {
2800                 if (tsc1 - tsc2 < halfway)
2801                         return 1;
2802                 else
2803                         return -1;
2804         }
2805 }
2806
2807 #define MAX_PADDING (PERF_AUXTRACE_RECORD_ALIGNMENT - 1)
2808
2809 /**
2810  * adj_for_padding - adjust overlap to account for padding.
2811  * @buf_b: second buffer
2812  * @buf_a: first buffer
2813  * @len_a: size of first buffer
2814  *
2815  * @buf_a might have up to 7 bytes of padding appended. Adjust the overlap
2816  * accordingly.
2817  *
2818  * Return: A pointer into @buf_b from where non-overlapped data starts
2819  */
2820 static unsigned char *adj_for_padding(unsigned char *buf_b,
2821                                       unsigned char *buf_a, size_t len_a)
2822 {
2823         unsigned char *p = buf_b - MAX_PADDING;
2824         unsigned char *q = buf_a + len_a - MAX_PADDING;
2825         int i;
2826
2827         for (i = MAX_PADDING; i; i--, p++, q++) {
2828                 if (*p != *q)
2829                         break;
2830         }
2831
2832         return p;
2833 }
2834
2835 /**
2836  * intel_pt_find_overlap_tsc - determine start of non-overlapped trace data
2837  *                             using TSC.
2838  * @buf_a: first buffer
2839  * @len_a: size of first buffer
2840  * @buf_b: second buffer
2841  * @len_b: size of second buffer
2842  * @consecutive: returns true if there is data in buf_b that is consecutive
2843  *               to buf_a
2844  *
2845  * If the trace contains TSC we can look at the last TSC of @buf_a and the
2846  * first TSC of @buf_b in order to determine if the buffers overlap, and then
2847  * walk forward in @buf_b until a later TSC is found.  A precondition is that
2848  * @buf_a and @buf_b are positioned at a PSB.
2849  *
2850  * Return: A pointer into @buf_b from where non-overlapped data starts, or
2851  * @buf_b + @len_b if there is no non-overlapped data.
2852  */
2853 static unsigned char *intel_pt_find_overlap_tsc(unsigned char *buf_a,
2854                                                 size_t len_a,
2855                                                 unsigned char *buf_b,
2856                                                 size_t len_b, bool *consecutive)
2857 {
2858         uint64_t tsc_a, tsc_b;
2859         unsigned char *p;
2860         size_t len, rem_a, rem_b;
2861
2862         p = intel_pt_last_psb(buf_a, len_a);
2863         if (!p)
2864                 return buf_b; /* No PSB in buf_a => no overlap */
2865
2866         len = len_a - (p - buf_a);
2867         if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a)) {
2868                 /* The last PSB+ in buf_a is incomplete, so go back one more */
2869                 len_a -= len;
2870                 p = intel_pt_last_psb(buf_a, len_a);
2871                 if (!p)
2872                         return buf_b; /* No full PSB+ => assume no overlap */
2873                 len = len_a - (p - buf_a);
2874                 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a))
2875                         return buf_b; /* No TSC in buf_a => assume no overlap */
2876         }
2877
2878         while (1) {
2879                 /* Ignore PSB+ with no TSC */
2880                 if (intel_pt_next_tsc(buf_b, len_b, &tsc_b, &rem_b)) {
2881                         int cmp = intel_pt_tsc_cmp(tsc_a, tsc_b);
2882
2883                         /* Same TSC, so buffers are consecutive */
2884                         if (!cmp && rem_b >= rem_a) {
2885                                 unsigned char *start;
2886
2887                                 *consecutive = true;
2888                                 start = buf_b + len_b - (rem_b - rem_a);
2889                                 return adj_for_padding(start, buf_a, len_a);
2890                         }
2891                         if (cmp < 0)
2892                                 return buf_b; /* tsc_a < tsc_b => no overlap */
2893                 }
2894
2895                 if (!intel_pt_step_psb(&buf_b, &len_b))
2896                         return buf_b + len_b; /* No PSB in buf_b => no data */
2897         }
2898 }
2899
2900 /**
2901  * intel_pt_find_overlap - determine start of non-overlapped trace data.
2902  * @buf_a: first buffer
2903  * @len_a: size of first buffer
2904  * @buf_b: second buffer
2905  * @len_b: size of second buffer
2906  * @have_tsc: can use TSC packets to detect overlap
2907  * @consecutive: returns true if there is data in buf_b that is consecutive
2908  *               to buf_a
2909  *
2910  * When trace samples or snapshots are recorded there is the possibility that
2911  * the data overlaps.  Note that, for the purposes of decoding, data is only
2912  * useful if it begins with a PSB packet.
2913  *
2914  * Return: A pointer into @buf_b from where non-overlapped data starts, or
2915  * @buf_b + @len_b if there is no non-overlapped data.
2916  */
2917 unsigned char *intel_pt_find_overlap(unsigned char *buf_a, size_t len_a,
2918                                      unsigned char *buf_b, size_t len_b,
2919                                      bool have_tsc, bool *consecutive)
2920 {
2921         unsigned char *found;
2922
2923         /* Buffer 'b' must start at PSB so throw away everything before that */
2924         if (!intel_pt_next_psb(&buf_b, &len_b))
2925                 return buf_b + len_b; /* No PSB */
2926
2927         if (!intel_pt_next_psb(&buf_a, &len_a))
2928                 return buf_b; /* No overlap */
2929
2930         if (have_tsc) {
2931                 found = intel_pt_find_overlap_tsc(buf_a, len_a, buf_b, len_b,
2932                                                   consecutive);
2933                 if (found)
2934                         return found;
2935         }
2936
2937         /*
2938          * Buffer 'b' cannot end within buffer 'a' so, for comparison purposes,
2939          * we can ignore the first part of buffer 'a'.
2940          */
2941         while (len_b < len_a) {
2942                 if (!intel_pt_step_psb(&buf_a, &len_a))
2943                         return buf_b; /* No overlap */
2944         }
2945
2946         /* Now len_b >= len_a */
2947         while (1) {
2948                 /* Potential overlap so check the bytes */
2949                 found = memmem(buf_a, len_a, buf_b, len_a);
2950                 if (found) {
2951                         *consecutive = true;
2952                         return adj_for_padding(buf_b + len_a, buf_a, len_a);
2953                 }
2954
2955                 /* Try again at next PSB in buffer 'a' */
2956                 if (!intel_pt_step_psb(&buf_a, &len_a))
2957                         return buf_b; /* No overlap */
2958         }
2959 }
2960
2961 /**
2962  * struct fast_forward_data - data used by intel_pt_ff_cb().
2963  * @timestamp: timestamp to fast forward towards
2964  * @buf_timestamp: buffer timestamp of last buffer with trace data earlier than
2965  *                 the fast forward timestamp.
2966  */
2967 struct fast_forward_data {
2968         uint64_t timestamp;
2969         uint64_t buf_timestamp;
2970 };
2971
2972 /**
2973  * intel_pt_ff_cb - fast forward lookahead callback.
2974  * @buffer: Intel PT trace buffer
2975  * @data: opaque pointer to fast forward data (struct fast_forward_data)
2976  *
2977  * Determine if @buffer trace is past the fast forward timestamp.
2978  *
2979  * Return: 1 (stop lookahead) if @buffer trace is past the fast forward
2980  *         timestamp, and 0 otherwise.
2981  */
2982 static int intel_pt_ff_cb(struct intel_pt_buffer *buffer, void *data)
2983 {
2984         struct fast_forward_data *d = data;
2985         unsigned char *buf;
2986         uint64_t tsc;
2987         size_t rem;
2988         size_t len;
2989
2990         buf = (unsigned char *)buffer->buf;
2991         len = buffer->len;
2992
2993         if (!intel_pt_next_psb(&buf, &len) ||
2994             !intel_pt_next_tsc(buf, len, &tsc, &rem))
2995                 return 0;
2996
2997         tsc = intel_pt_8b_tsc(tsc, buffer->ref_timestamp);
2998
2999         intel_pt_log("Buffer 1st timestamp " x64_fmt " ref timestamp " x64_fmt "\n",
3000                      tsc, buffer->ref_timestamp);
3001
3002         /*
3003          * If the buffer contains a timestamp earlier that the fast forward
3004          * timestamp, then record it, else stop.
3005          */
3006         if (tsc < d->timestamp)
3007                 d->buf_timestamp = buffer->ref_timestamp;
3008         else
3009                 return 1;
3010
3011         return 0;
3012 }
3013
3014 /**
3015  * intel_pt_fast_forward - reposition decoder forwards.
3016  * @decoder: Intel PT decoder
3017  * @timestamp: timestamp to fast forward towards
3018  *
3019  * Reposition decoder at the last PSB with a timestamp earlier than @timestamp.
3020  *
3021  * Return: 0 on success or negative error code on failure.
3022  */
3023 int intel_pt_fast_forward(struct intel_pt_decoder *decoder, uint64_t timestamp)
3024 {
3025         struct fast_forward_data d = { .timestamp = timestamp };
3026         unsigned char *buf;
3027         size_t len;
3028         int err;
3029
3030         intel_pt_log("Fast forward towards timestamp " x64_fmt "\n", timestamp);
3031
3032         /* Find buffer timestamp of buffer to fast forward to */
3033         err = decoder->lookahead(decoder->data, intel_pt_ff_cb, &d);
3034         if (err < 0)
3035                 return err;
3036
3037         /* Walk to buffer with same buffer timestamp */
3038         if (d.buf_timestamp) {
3039                 do {
3040                         decoder->pos += decoder->len;
3041                         decoder->len = 0;
3042                         err = intel_pt_get_next_data(decoder, true);
3043                         /* -ENOLINK means non-consecutive trace */
3044                         if (err && err != -ENOLINK)
3045                                 return err;
3046                 } while (decoder->buf_timestamp != d.buf_timestamp);
3047         }
3048
3049         if (!decoder->buf)
3050                 return 0;
3051
3052         buf = (unsigned char *)decoder->buf;
3053         len = decoder->len;
3054
3055         if (!intel_pt_next_psb(&buf, &len))
3056                 return 0;
3057
3058         /*
3059          * Walk PSBs while the PSB timestamp is less than the fast forward
3060          * timestamp.
3061          */
3062         do {
3063                 uint64_t tsc;
3064                 size_t rem;
3065
3066                 if (!intel_pt_next_tsc(buf, len, &tsc, &rem))
3067                         break;
3068                 tsc = intel_pt_8b_tsc(tsc, decoder->buf_timestamp);
3069                 /*
3070                  * A TSC packet can slip past MTC packets but, after fast
3071                  * forward, decoding starts at the TSC timestamp. That means
3072                  * the timestamps may not be exactly the same as the timestamps
3073                  * that would have been decoded without fast forward.
3074                  */
3075                 if (tsc < timestamp) {
3076                         intel_pt_log("Fast forward to next PSB timestamp " x64_fmt "\n", tsc);
3077                         decoder->pos += decoder->len - len;
3078                         decoder->buf = buf;
3079                         decoder->len = len;
3080                         intel_pt_reposition(decoder);
3081                 } else {
3082                         break;
3083                 }
3084         } while (intel_pt_step_psb(&buf, &len));
3085
3086         return 0;
3087 }