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

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