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

   1 /*
   2  * intel_pt_decoder.c: Intel Processor Trace support
   3  * Copyright (c) 2013-2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  */
  15
  16 #ifndef _GNU_SOURCE
  17 #define _GNU_SOURCE
  18 #endif
  19 #include <stdlib.h>
  20 #include <stdbool.h>
  21 #include <string.h>
  22 #include <errno.h>
  23 #include <stdint.h>
  24 #include <inttypes.h>
  25 #include <linux/compiler.h>
  26
  27 #include "../cache.h"
  28 #include "../util.h"
  29 #include "../auxtrace.h"
  30
  31 #include "intel-pt-insn-decoder.h"
  32 #include "intel-pt-pkt-decoder.h"
  33 #include "intel-pt-decoder.h"
  34 #include "intel-pt-log.h"
  35
  36 #define INTEL_PT_BLK_SIZE 1024
  37
  38 #define BIT63 (((uint64_t)1 << 63))
  39
  40 #define INTEL_PT_RETURN 1
  41
  42 /* Maximum number of loops with no packets consumed i.e. stuck in a loop */
  43 #define INTEL_PT_MAX_LOOPS 10000
  44
  45 struct intel_pt_blk {
  46         struct intel_pt_blk *prev;
  47         uint64_t ip[INTEL_PT_BLK_SIZE];
  48 };
  49
  50 struct intel_pt_stack {
  51         struct intel_pt_blk *blk;
  52         struct intel_pt_blk *spare;
  53         int pos;
  54 };
  55
  56 enum intel_pt_pkt_state {
  57         INTEL_PT_STATE_NO_PSB,
  58         INTEL_PT_STATE_NO_IP,
  59         INTEL_PT_STATE_ERR_RESYNC,
  60         INTEL_PT_STATE_IN_SYNC,
  61         INTEL_PT_STATE_TNT,
  62         INTEL_PT_STATE_TIP,
  63         INTEL_PT_STATE_TIP_PGD,
  64         INTEL_PT_STATE_FUP,
  65         INTEL_PT_STATE_FUP_NO_TIP,
  66 };
  67
  68 static inline bool intel_pt_sample_time(enum intel_pt_pkt_state pkt_state)
  69 {
  70         switch (pkt_state) {
  71         case INTEL_PT_STATE_NO_PSB:
  72         case INTEL_PT_STATE_NO_IP:
  73         case INTEL_PT_STATE_ERR_RESYNC:
  74         case INTEL_PT_STATE_IN_SYNC:
  75         case INTEL_PT_STATE_TNT:
  76                 return true;
  77         case INTEL_PT_STATE_TIP:
  78         case INTEL_PT_STATE_TIP_PGD:
  79         case INTEL_PT_STATE_FUP:
  80         case INTEL_PT_STATE_FUP_NO_TIP:
  81                 return false;
  82         default:
  83                 return true;
  84         };
  85 }
  86
  87 #ifdef INTEL_PT_STRICT
  88 #define INTEL_PT_STATE_ERR1     INTEL_PT_STATE_NO_PSB
  89 #define INTEL_PT_STATE_ERR2     INTEL_PT_STATE_NO_PSB
  90 #define INTEL_PT_STATE_ERR3     INTEL_PT_STATE_NO_PSB
  91 #define INTEL_PT_STATE_ERR4     INTEL_PT_STATE_NO_PSB
  92 #else
  93 #define INTEL_PT_STATE_ERR1     (decoder->pkt_state)
  94 #define INTEL_PT_STATE_ERR2     INTEL_PT_STATE_NO_IP
  95 #define INTEL_PT_STATE_ERR3     INTEL_PT_STATE_ERR_RESYNC
  96 #define INTEL_PT_STATE_ERR4     INTEL_PT_STATE_IN_SYNC
  97 #endif
  98
  99 struct intel_pt_decoder {
 100         int (*get_trace)(struct intel_pt_buffer *buffer, void *data);
 101         int (*walk_insn)(struct intel_pt_insn *intel_pt_insn,
 102                          uint64_t *insn_cnt_ptr, uint64_t *ip, uint64_t to_ip,
 103                          uint64_t max_insn_cnt, void *data);
 104         bool (*pgd_ip)(uint64_t ip, void *data);
 105         void *data;
 106         struct intel_pt_state state;
 107         const unsigned char *buf;
 108         size_t len;
 109         bool return_compression;
 110         bool branch_enable;
 111         bool mtc_insn;
 112         bool pge;
 113         bool have_tma;
 114         bool have_cyc;
 115         bool fixup_last_mtc;
 116         bool have_last_ip;
 117         enum intel_pt_param_flags flags;
 118         uint64_t pos;
 119         uint64_t last_ip;
 120         uint64_t ip;
 121         uint64_t cr3;
 122         uint64_t timestamp;
 123         uint64_t tsc_timestamp;
 124         uint64_t ref_timestamp;
 125         uint64_t sample_timestamp;
 126         uint64_t ret_addr;
 127         uint64_t ctc_timestamp;
 128         uint64_t ctc_delta;
 129         uint64_t cycle_cnt;
 130         uint64_t cyc_ref_timestamp;
 131         uint32_t last_mtc;
 132         uint32_t tsc_ctc_ratio_n;
 133         uint32_t tsc_ctc_ratio_d;
 134         uint32_t tsc_ctc_mult;
 135         uint32_t tsc_slip;
 136         uint32_t ctc_rem_mask;
 137         int mtc_shift;
 138         struct intel_pt_stack stack;
 139         enum intel_pt_pkt_state pkt_state;
 140         struct intel_pt_pkt packet;
 141         struct intel_pt_pkt tnt;
 142         int pkt_step;
 143         int pkt_len;
 144         int last_packet_type;
 145         unsigned int cbr;
 146         unsigned int cbr_seen;
 147         unsigned int max_non_turbo_ratio;
 148         double max_non_turbo_ratio_fp;
 149         double cbr_cyc_to_tsc;
 150         double calc_cyc_to_tsc;
 151         bool have_calc_cyc_to_tsc;
 152         int exec_mode;
 153         unsigned int insn_bytes;
 154         uint64_t period;
 155         enum intel_pt_period_type period_type;
 156         uint64_t tot_insn_cnt;
 157         uint64_t period_insn_cnt;
 158         uint64_t period_mask;
 159         uint64_t period_ticks;
 160         uint64_t last_masked_timestamp;
 161         bool continuous_period;
 162         bool overflow;
 163         bool set_fup_tx_flags;
 164         bool set_fup_ptw;
 165         bool set_fup_mwait;
 166         bool set_fup_pwre;
 167         bool set_fup_exstop;
 168         unsigned int fup_tx_flags;
 169         unsigned int tx_flags;
 170         uint64_t fup_ptw_payload;
 171         uint64_t fup_mwait_payload;
 172         uint64_t fup_pwre_payload;
 173         uint64_t cbr_payload;
 174         uint64_t timestamp_insn_cnt;
 175         uint64_t sample_insn_cnt;
 176         uint64_t stuck_ip;
 177         int no_progress;
 178         int stuck_ip_prd;
 179         int stuck_ip_cnt;
 180         const unsigned char *next_buf;
 181         size_t next_len;
 182         unsigned char temp_buf[INTEL_PT_PKT_MAX_SZ];
 183 };
 184
 185 static uint64_t intel_pt_lower_power_of_2(uint64_t x)
 186 {
 187         int i;
 188
 189         for (i = 0; x != 1; i++)
 190                 x >>= 1;
 191
 192         return x << i;
 193 }
 194
 195 static void intel_pt_setup_period(struct intel_pt_decoder *decoder)
 196 {
 197         if (decoder->period_type == INTEL_PT_PERIOD_TICKS) {
 198                 uint64_t period;
 199
 200                 period = intel_pt_lower_power_of_2(decoder->period);
 201                 decoder->period_mask  = ~(period - 1);
 202                 decoder->period_ticks = period;
 203         }
 204 }
 205
 206 static uint64_t multdiv(uint64_t t, uint32_t n, uint32_t d)
 207 {
 208         if (!d)
 209                 return 0;
 210         return (t / d) * n + ((t % d) * n) / d;
 211 }
 212
 213 struct intel_pt_decoder *intel_pt_decoder_new(struct intel_pt_params *params)
 214 {
 215         struct intel_pt_decoder *decoder;
 216
 217         if (!params->get_trace || !params->walk_insn)
 218                 return NULL;
 219
 220         decoder = zalloc(sizeof(struct intel_pt_decoder));
 221         if (!decoder)
 222                 return NULL;
 223
 224         decoder->get_trace          = params->get_trace;
 225         decoder->walk_insn          = params->walk_insn;
 226         decoder->pgd_ip             = params->pgd_ip;
 227         decoder->data               = params->data;
 228         decoder->return_compression = params->return_compression;
 229         decoder->branch_enable      = params->branch_enable;
 230
 231         decoder->flags              = params->flags;
 232
 233         decoder->period             = params->period;
 234         decoder->period_type        = params->period_type;
 235
 236         decoder->max_non_turbo_ratio    = params->max_non_turbo_ratio;
 237         decoder->max_non_turbo_ratio_fp = params->max_non_turbo_ratio;
 238
 239         intel_pt_setup_period(decoder);
 240
 241         decoder->mtc_shift = params->mtc_period;
 242         decoder->ctc_rem_mask = (1 << decoder->mtc_shift) - 1;
 243
 244         decoder->tsc_ctc_ratio_n = params->tsc_ctc_ratio_n;
 245         decoder->tsc_ctc_ratio_d = params->tsc_ctc_ratio_d;
 246
 247         if (!decoder->tsc_ctc_ratio_n)
 248                 decoder->tsc_ctc_ratio_d = 0;
 249
 250         if (decoder->tsc_ctc_ratio_d) {
 251                 if (!(decoder->tsc_ctc_ratio_n % decoder->tsc_ctc_ratio_d))
 252                         decoder->tsc_ctc_mult = decoder->tsc_ctc_ratio_n /
 253                                                 decoder->tsc_ctc_ratio_d;
 254
 255                 /*
 256                  * Allow for timestamps appearing to backwards because a TSC
 257                  * packet has slipped past a MTC packet, so allow 2 MTC ticks
 258                  * or ...
 259                  */
 260                 decoder->tsc_slip = multdiv(2 << decoder->mtc_shift,
 261                                         decoder->tsc_ctc_ratio_n,
 262                                         decoder->tsc_ctc_ratio_d);
 263         }
 264         /* ... or 0x100 paranoia */
 265         if (decoder->tsc_slip < 0x100)
 266                 decoder->tsc_slip = 0x100;
 267
 268         intel_pt_log("timestamp: mtc_shift %u\n", decoder->mtc_shift);
 269         intel_pt_log("timestamp: tsc_ctc_ratio_n %u\n", decoder->tsc_ctc_ratio_n);
 270         intel_pt_log("timestamp: tsc_ctc_ratio_d %u\n", decoder->tsc_ctc_ratio_d);
 271         intel_pt_log("timestamp: tsc_ctc_mult %u\n", decoder->tsc_ctc_mult);
 272         intel_pt_log("timestamp: tsc_slip %#x\n", decoder->tsc_slip);
 273
 274         return decoder;
 275 }
 276
 277 static void intel_pt_pop_blk(struct intel_pt_stack *stack)
 278 {
 279         struct intel_pt_blk *blk = stack->blk;
 280
 281         stack->blk = blk->prev;
 282         if (!stack->spare)
 283                 stack->spare = blk;
 284         else
 285                 free(blk);
 286 }
 287
 288 static uint64_t intel_pt_pop(struct intel_pt_stack *stack)
 289 {
 290         if (!stack->pos) {
 291                 if (!stack->blk)
 292                         return 0;
 293                 intel_pt_pop_blk(stack);
 294                 if (!stack->blk)
 295                         return 0;
 296                 stack->pos = INTEL_PT_BLK_SIZE;
 297         }
 298         return stack->blk->ip[--stack->pos];
 299 }
 300
 301 static int intel_pt_alloc_blk(struct intel_pt_stack *stack)
 302 {
 303         struct intel_pt_blk *blk;
 304
 305         if (stack->spare) {
 306                 blk = stack->spare;
 307                 stack->spare = NULL;
 308         } else {
 309                 blk = malloc(sizeof(struct intel_pt_blk));
 310                 if (!blk)
 311                         return -ENOMEM;
 312         }
 313
 314         blk->prev = stack->blk;
 315         stack->blk = blk;
 316         stack->pos = 0;
 317         return 0;
 318 }
 319
 320 static int intel_pt_push(struct intel_pt_stack *stack, uint64_t ip)
 321 {
 322         int err;
 323
 324         if (!stack->blk || stack->pos == INTEL_PT_BLK_SIZE) {
 325                 err = intel_pt_alloc_blk(stack);
 326                 if (err)
 327                         return err;
 328         }
 329
 330         stack->blk->ip[stack->pos++] = ip;
 331         return 0;
 332 }
 333
 334 static void intel_pt_clear_stack(struct intel_pt_stack *stack)
 335 {
 336         while (stack->blk)
 337                 intel_pt_pop_blk(stack);
 338         stack->pos = 0;
 339 }
 340
 341 static void intel_pt_free_stack(struct intel_pt_stack *stack)
 342 {
 343         intel_pt_clear_stack(stack);
 344         zfree(&stack->blk);
 345         zfree(&stack->spare);
 346 }
 347
 348 void intel_pt_decoder_free(struct intel_pt_decoder *decoder)
 349 {
 350         intel_pt_free_stack(&decoder->stack);
 351         free(decoder);
 352 }
 353
 354 static int intel_pt_ext_err(int code)
 355 {
 356         switch (code) {
 357         case -ENOMEM:
 358                 return INTEL_PT_ERR_NOMEM;
 359         case -ENOSYS:
 360                 return INTEL_PT_ERR_INTERN;
 361         case -EBADMSG:
 362                 return INTEL_PT_ERR_BADPKT;
 363         case -ENODATA:
 364                 return INTEL_PT_ERR_NODATA;
 365         case -EILSEQ:
 366                 return INTEL_PT_ERR_NOINSN;
 367         case -ENOENT:
 368                 return INTEL_PT_ERR_MISMAT;
 369         case -EOVERFLOW:
 370                 return INTEL_PT_ERR_OVR;
 371         case -ENOSPC:
 372                 return INTEL_PT_ERR_LOST;
 373         case -ELOOP:
 374                 return INTEL_PT_ERR_NELOOP;
 375         default:
 376                 return INTEL_PT_ERR_UNK;
 377         }
 378 }
 379
 380 static const char *intel_pt_err_msgs[] = {
 381         [INTEL_PT_ERR_NOMEM]  = "Memory allocation failed",
 382         [INTEL_PT_ERR_INTERN] = "Internal error",
 383         [INTEL_PT_ERR_BADPKT] = "Bad packet",
 384         [INTEL_PT_ERR_NODATA] = "No more data",
 385         [INTEL_PT_ERR_NOINSN] = "Failed to get instruction",
 386         [INTEL_PT_ERR_MISMAT] = "Trace doesn't match instruction",
 387         [INTEL_PT_ERR_OVR]    = "Overflow packet",
 388         [INTEL_PT_ERR_LOST]   = "Lost trace data",
 389         [INTEL_PT_ERR_UNK]    = "Unknown error!",
 390         [INTEL_PT_ERR_NELOOP] = "Never-ending loop",
 391 };
 392
 393 int intel_pt__strerror(int code, char *buf, size_t buflen)
 394 {
 395         if (code < 1 || code >= INTEL_PT_ERR_MAX)
 396                 code = INTEL_PT_ERR_UNK;
 397         strlcpy(buf, intel_pt_err_msgs[code], buflen);
 398         return 0;
 399 }
 400
 401 static uint64_t intel_pt_calc_ip(const struct intel_pt_pkt *packet,
 402                                  uint64_t last_ip)
 403 {
 404         uint64_t ip;
 405
 406         switch (packet->count) {
 407         case 1:
 408                 ip = (last_ip & (uint64_t)0xffffffffffff0000ULL) |
 409                      packet->payload;
 410                 break;
 411         case 2:
 412                 ip = (last_ip & (uint64_t)0xffffffff00000000ULL) |
 413                      packet->payload;
 414                 break;
 415         case 3:
 416                 ip = packet->payload;
 417                 /* Sign-extend 6-byte ip */
 418                 if (ip & (uint64_t)0x800000000000ULL)
 419                         ip |= (uint64_t)0xffff000000000000ULL;
 420                 break;
 421         case 4:
 422                 ip = (last_ip & (uint64_t)0xffff000000000000ULL) |
 423                      packet->payload;
 424                 break;
 425         case 6:
 426                 ip = packet->payload;
 427                 break;
 428         default:
 429                 return 0;
 430         }
 431
 432         return ip;
 433 }
 434
 435 static inline void intel_pt_set_last_ip(struct intel_pt_decoder *decoder)
 436 {
 437         decoder->last_ip = intel_pt_calc_ip(&decoder->packet, decoder->last_ip);
 438         decoder->have_last_ip = true;
 439 }
 440
 441 static inline void intel_pt_set_ip(struct intel_pt_decoder *decoder)
 442 {
 443         intel_pt_set_last_ip(decoder);
 444         decoder->ip = decoder->last_ip;
 445 }
 446
 447 static void intel_pt_decoder_log_packet(struct intel_pt_decoder *decoder)
 448 {
 449         intel_pt_log_packet(&decoder->packet, decoder->pkt_len, decoder->pos,
 450                             decoder->buf);
 451 }
 452
 453 static int intel_pt_bug(struct intel_pt_decoder *decoder)
 454 {
 455         intel_pt_log("ERROR: Internal error\n");
 456         decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
 457         return -ENOSYS;
 458 }
 459
 460 static inline void intel_pt_clear_tx_flags(struct intel_pt_decoder *decoder)
 461 {
 462         decoder->tx_flags = 0;
 463 }
 464
 465 static inline void intel_pt_update_in_tx(struct intel_pt_decoder *decoder)
 466 {
 467         decoder->tx_flags = decoder->packet.payload & INTEL_PT_IN_TX;
 468 }
 469
 470 static int intel_pt_bad_packet(struct intel_pt_decoder *decoder)
 471 {
 472         intel_pt_clear_tx_flags(decoder);
 473         decoder->have_tma = false;
 474         decoder->pkt_len = 1;
 475         decoder->pkt_step = 1;
 476         intel_pt_decoder_log_packet(decoder);
 477         if (decoder->pkt_state != INTEL_PT_STATE_NO_PSB) {
 478                 intel_pt_log("ERROR: Bad packet\n");
 479                 decoder->pkt_state = INTEL_PT_STATE_ERR1;
 480         }
 481         return -EBADMSG;
 482 }
 483
 484 static int intel_pt_get_data(struct intel_pt_decoder *decoder)
 485 {
 486         struct intel_pt_buffer buffer = { .buf = 0, };
 487         int ret;
 488
 489         decoder->pkt_step = 0;
 490
 491         intel_pt_log("Getting more data\n");
 492         ret = decoder->get_trace(&buffer, decoder->data);
 493         if (ret)
 494                 return ret;
 495         decoder->buf = buffer.buf;
 496         decoder->len = buffer.len;
 497         if (!decoder->len) {
 498                 intel_pt_log("No more data\n");
 499                 return -ENODATA;
 500         }
 501         if (!buffer.consecutive) {
 502                 decoder->ip = 0;
 503                 decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
 504                 decoder->ref_timestamp = buffer.ref_timestamp;
 505                 decoder->timestamp = 0;
 506                 decoder->have_tma = false;
 507                 decoder->state.trace_nr = buffer.trace_nr;
 508                 intel_pt_log("Reference timestamp 0x%" PRIx64 "\n",
 509                              decoder->ref_timestamp);
 510                 return -ENOLINK;
 511         }
 512
 513         return 0;
 514 }
 515
 516 static int intel_pt_get_next_data(struct intel_pt_decoder *decoder)
 517 {
 518         if (!decoder->next_buf)
 519                 return intel_pt_get_data(decoder);
 520
 521         decoder->buf = decoder->next_buf;
 522         decoder->len = decoder->next_len;
 523         decoder->next_buf = 0;
 524         decoder->next_len = 0;
 525         return 0;
 526 }
 527
 528 static int intel_pt_get_split_packet(struct intel_pt_decoder *decoder)
 529 {
 530         unsigned char *buf = decoder->temp_buf;
 531         size_t old_len, len, n;
 532         int ret;
 533
 534         old_len = decoder->len;
 535         len = decoder->len;
 536         memcpy(buf, decoder->buf, len);
 537
 538         ret = intel_pt_get_data(decoder);
 539         if (ret) {
 540                 decoder->pos += old_len;
 541                 return ret < 0 ? ret : -EINVAL;
 542         }
 543
 544         n = INTEL_PT_PKT_MAX_SZ - len;
 545         if (n > decoder->len)
 546                 n = decoder->len;
 547         memcpy(buf + len, decoder->buf, n);
 548         len += n;
 549
 550         ret = intel_pt_get_packet(buf, len, &decoder->packet);
 551         if (ret < (int)old_len) {
 552                 decoder->next_buf = decoder->buf;
 553                 decoder->next_len = decoder->len;
 554                 decoder->buf = buf;
 555                 decoder->len = old_len;
 556                 return intel_pt_bad_packet(decoder);
 557         }
 558
 559         decoder->next_buf = decoder->buf + (ret - old_len);
 560         decoder->next_len = decoder->len - (ret - old_len);
 561
 562         decoder->buf = buf;
 563         decoder->len = ret;
 564
 565         return ret;
 566 }
 567
 568 struct intel_pt_pkt_info {
 569         struct intel_pt_decoder   *decoder;
 570         struct intel_pt_pkt       packet;
 571         uint64_t                  pos;
 572         int                       pkt_len;
 573         int                       last_packet_type;
 574         void                      *data;
 575 };
 576
 577 typedef int (*intel_pt_pkt_cb_t)(struct intel_pt_pkt_info *pkt_info);
 578
 579 /* Lookahead packets in current buffer */
 580 static int intel_pt_pkt_lookahead(struct intel_pt_decoder *decoder,
 581                                   intel_pt_pkt_cb_t cb, void *data)
 582 {
 583         struct intel_pt_pkt_info pkt_info;
 584         const unsigned char *buf = decoder->buf;
 585         size_t len = decoder->len;
 586         int ret;
 587
 588         pkt_info.decoder          = decoder;
 589         pkt_info.pos              = decoder->pos;
 590         pkt_info.pkt_len          = decoder->pkt_step;
 591         pkt_info.last_packet_type = decoder->last_packet_type;
 592         pkt_info.data             = data;
 593
 594         while (1) {
 595                 do {
 596                         pkt_info.pos += pkt_info.pkt_len;
 597                         buf          += pkt_info.pkt_len;
 598                         len          -= pkt_info.pkt_len;
 599
 600                         if (!len)
 601                                 return INTEL_PT_NEED_MORE_BYTES;
 602
 603                         ret = intel_pt_get_packet(buf, len, &pkt_info.packet);
 604                         if (!ret)
 605                                 return INTEL_PT_NEED_MORE_BYTES;
 606                         if (ret < 0)
 607                                 return ret;
 608
 609                         pkt_info.pkt_len = ret;
 610                 } while (pkt_info.packet.type == INTEL_PT_PAD);
 611
 612                 ret = cb(&pkt_info);
 613                 if (ret)
 614                         return 0;
 615
 616                 pkt_info.last_packet_type = pkt_info.packet.type;
 617         }
 618 }
 619
 620 struct intel_pt_calc_cyc_to_tsc_info {
 621         uint64_t        cycle_cnt;
 622         unsigned int    cbr;
 623         uint32_t        last_mtc;
 624         uint64_t        ctc_timestamp;
 625         uint64_t        ctc_delta;
 626         uint64_t        tsc_timestamp;
 627         uint64_t        timestamp;
 628         bool            have_tma;
 629         bool            fixup_last_mtc;
 630         bool            from_mtc;
 631         double          cbr_cyc_to_tsc;
 632 };
 633
 634 /*
 635  * MTC provides a 8-bit slice of CTC but the TMA packet only provides the lower
 636  * 16 bits of CTC. If mtc_shift > 8 then some of the MTC bits are not in the CTC
 637  * provided by the TMA packet. Fix-up the last_mtc calculated from the TMA
 638  * packet by copying the missing bits from the current MTC assuming the least
 639  * difference between the two, and that the current MTC comes after last_mtc.
 640  */
 641 static void intel_pt_fixup_last_mtc(uint32_t mtc, int mtc_shift,
 642                                     uint32_t *last_mtc)
 643 {
 644         uint32_t first_missing_bit = 1U << (16 - mtc_shift);
 645         uint32_t mask = ~(first_missing_bit - 1);
 646
 647         *last_mtc |= mtc & mask;
 648         if (*last_mtc >= mtc) {
 649                 *last_mtc -= first_missing_bit;
 650                 *last_mtc &= 0xff;
 651         }
 652 }
 653
 654 static int intel_pt_calc_cyc_cb(struct intel_pt_pkt_info *pkt_info)
 655 {
 656         struct intel_pt_decoder *decoder = pkt_info->decoder;
 657         struct intel_pt_calc_cyc_to_tsc_info *data = pkt_info->data;
 658         uint64_t timestamp;
 659         double cyc_to_tsc;
 660         unsigned int cbr;
 661         uint32_t mtc, mtc_delta, ctc, fc, ctc_rem;
 662
 663         switch (pkt_info->packet.type) {
 664         case INTEL_PT_TNT:
 665         case INTEL_PT_TIP_PGE:
 666         case INTEL_PT_TIP:
 667         case INTEL_PT_FUP:
 668         case INTEL_PT_PSB:
 669         case INTEL_PT_PIP:
 670         case INTEL_PT_MODE_EXEC:
 671         case INTEL_PT_MODE_TSX:
 672         case INTEL_PT_PSBEND:
 673         case INTEL_PT_PAD:
 674         case INTEL_PT_VMCS:
 675         case INTEL_PT_MNT:
 676         case INTEL_PT_PTWRITE:
 677         case INTEL_PT_PTWRITE_IP:
 678                 return 0;
 679
 680         case INTEL_PT_MTC:
 681                 if (!data->have_tma)
 682                         return 0;
 683
 684                 mtc = pkt_info->packet.payload;
 685                 if (decoder->mtc_shift > 8 && data->fixup_last_mtc) {
 686                         data->fixup_last_mtc = false;
 687                         intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
 688                                                 &data->last_mtc);
 689                 }
 690                 if (mtc > data->last_mtc)
 691                         mtc_delta = mtc - data->last_mtc;
 692                 else
 693                         mtc_delta = mtc + 256 - data->last_mtc;
 694                 data->ctc_delta += mtc_delta << decoder->mtc_shift;
 695                 data->last_mtc = mtc;
 696
 697                 if (decoder->tsc_ctc_mult) {
 698                         timestamp = data->ctc_timestamp +
 699                                 data->ctc_delta * decoder->tsc_ctc_mult;
 700                 } else {
 701                         timestamp = data->ctc_timestamp +
 702                                 multdiv(data->ctc_delta,
 703                                         decoder->tsc_ctc_ratio_n,
 704                                         decoder->tsc_ctc_ratio_d);
 705                 }
 706
 707                 if (timestamp < data->timestamp)
 708                         return 1;
 709
 710                 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
 711                         data->timestamp = timestamp;
 712                         return 0;
 713                 }
 714
 715                 break;
 716
 717         case INTEL_PT_TSC:
 718                 /*
 719                  * For now, do not support using TSC packets - refer
 720                  * intel_pt_calc_cyc_to_tsc().
 721                  */
 722                 if (data->from_mtc)
 723                         return 1;
 724                 timestamp = pkt_info->packet.payload |
 725                             (data->timestamp & (0xffULL << 56));
 726                 if (data->from_mtc && timestamp < data->timestamp &&
 727                     data->timestamp - timestamp < decoder->tsc_slip)
 728                         return 1;
 729                 if (timestamp < data->timestamp)
 730                         timestamp += (1ULL << 56);
 731                 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
 732                         if (data->from_mtc)
 733                                 return 1;
 734                         data->tsc_timestamp = timestamp;
 735                         data->timestamp = timestamp;
 736                         return 0;
 737                 }
 738                 break;
 739
 740         case INTEL_PT_TMA:
 741                 if (data->from_mtc)
 742                         return 1;
 743
 744                 if (!decoder->tsc_ctc_ratio_d)
 745                         return 0;
 746
 747                 ctc = pkt_info->packet.payload;
 748                 fc = pkt_info->packet.count;
 749                 ctc_rem = ctc & decoder->ctc_rem_mask;
 750
 751                 data->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
 752
 753                 data->ctc_timestamp = data->tsc_timestamp - fc;
 754                 if (decoder->tsc_ctc_mult) {
 755                         data->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
 756                 } else {
 757                         data->ctc_timestamp -=
 758                                 multdiv(ctc_rem, decoder->tsc_ctc_ratio_n,
 759                                         decoder->tsc_ctc_ratio_d);
 760                 }
 761
 762                 data->ctc_delta = 0;
 763                 data->have_tma = true;
 764                 data->fixup_last_mtc = true;
 765
 766                 return 0;
 767
 768         case INTEL_PT_CYC:
 769                 data->cycle_cnt += pkt_info->packet.payload;
 770                 return 0;
 771
 772         case INTEL_PT_CBR:
 773                 cbr = pkt_info->packet.payload;
 774                 if (data->cbr && data->cbr != cbr)
 775                         return 1;
 776                 data->cbr = cbr;
 777                 data->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
 778                 return 0;
 779
 780         case INTEL_PT_TIP_PGD:
 781         case INTEL_PT_TRACESTOP:
 782         case INTEL_PT_EXSTOP:
 783         case INTEL_PT_EXSTOP_IP:
 784         case INTEL_PT_MWAIT:
 785         case INTEL_PT_PWRE:
 786         case INTEL_PT_PWRX:
 787         case INTEL_PT_OVF:
 788         case INTEL_PT_BAD: /* Does not happen */
 789         default:
 790                 return 1;
 791         }
 792
 793         if (!data->cbr && decoder->cbr) {
 794                 data->cbr = decoder->cbr;
 795                 data->cbr_cyc_to_tsc = decoder->cbr_cyc_to_tsc;
 796         }
 797
 798         if (!data->cycle_cnt)
 799                 return 1;
 800
 801         cyc_to_tsc = (double)(timestamp - decoder->timestamp) / data->cycle_cnt;
 802
 803         if (data->cbr && cyc_to_tsc > data->cbr_cyc_to_tsc &&
 804             cyc_to_tsc / data->cbr_cyc_to_tsc > 1.25) {
 805                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle too big (c.f. CBR-based value %g), pos " x64_fmt "\n",
 806                              cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
 807                 return 1;
 808         }
 809
 810         decoder->calc_cyc_to_tsc = cyc_to_tsc;
 811         decoder->have_calc_cyc_to_tsc = true;
 812
 813         if (data->cbr) {
 814                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. CBR-based value %g, pos " x64_fmt "\n",
 815                              cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
 816         } else {
 817                 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. unknown CBR-based value, pos " x64_fmt "\n",
 818                              cyc_to_tsc, pkt_info->pos);
 819         }
 820
 821         return 1;
 822 }
 823
 824 static void intel_pt_calc_cyc_to_tsc(struct intel_pt_decoder *decoder,
 825                                      bool from_mtc)
 826 {
 827         struct intel_pt_calc_cyc_to_tsc_info data = {
 828                 .cycle_cnt      = 0,
 829                 .cbr            = 0,
 830                 .last_mtc       = decoder->last_mtc,
 831                 .ctc_timestamp  = decoder->ctc_timestamp,
 832                 .ctc_delta      = decoder->ctc_delta,
 833                 .tsc_timestamp  = decoder->tsc_timestamp,
 834                 .timestamp      = decoder->timestamp,
 835                 .have_tma       = decoder->have_tma,
 836                 .fixup_last_mtc = decoder->fixup_last_mtc,
 837                 .from_mtc       = from_mtc,
 838                 .cbr_cyc_to_tsc = 0,
 839         };
 840
 841         /*
 842          * For now, do not support using TSC packets for at least the reasons:
 843          * 1) timing might have stopped
 844          * 2) TSC packets within PSB+ can slip against CYC packets
 845          */
 846         if (!from_mtc)
 847                 return;
 848
 849         intel_pt_pkt_lookahead(decoder, intel_pt_calc_cyc_cb, &data);
 850 }
 851
 852 static int intel_pt_get_next_packet(struct intel_pt_decoder *decoder)
 853 {
 854         int ret;
 855
 856         decoder->last_packet_type = decoder->packet.type;
 857
 858         do {
 859                 decoder->pos += decoder->pkt_step;
 860                 decoder->buf += decoder->pkt_step;
 861                 decoder->len -= decoder->pkt_step;
 862
 863                 if (!decoder->len) {
 864                         ret = intel_pt_get_next_data(decoder);
 865                         if (ret)
 866                                 return ret;
 867                 }
 868
 869                 ret = intel_pt_get_packet(decoder->buf, decoder->len,
 870                                           &decoder->packet);
 871                 if (ret == INTEL_PT_NEED_MORE_BYTES && BITS_PER_LONG == 32 &&
 872                     decoder->len < INTEL_PT_PKT_MAX_SZ && !decoder->next_buf) {
 873                         ret = intel_pt_get_split_packet(decoder);
 874                         if (ret < 0)
 875                                 return ret;
 876                 }
 877                 if (ret <= 0)
 878                         return intel_pt_bad_packet(decoder);
 879
 880                 decoder->pkt_len = ret;
 881                 decoder->pkt_step = ret;
 882                 intel_pt_decoder_log_packet(decoder);
 883         } while (decoder->packet.type == INTEL_PT_PAD);
 884
 885         return 0;
 886 }
 887
 888 static uint64_t intel_pt_next_period(struct intel_pt_decoder *decoder)
 889 {
 890         uint64_t timestamp, masked_timestamp;
 891
 892         timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
 893         masked_timestamp = timestamp & decoder->period_mask;
 894         if (decoder->continuous_period) {
 895                 if (masked_timestamp != decoder->last_masked_timestamp)
 896                         return 1;
 897         } else {
 898                 timestamp += 1;
 899                 masked_timestamp = timestamp & decoder->period_mask;
 900                 if (masked_timestamp != decoder->last_masked_timestamp) {
 901                         decoder->last_masked_timestamp = masked_timestamp;
 902                         decoder->continuous_period = true;
 903                 }
 904         }
 905         return decoder->period_ticks - (timestamp - masked_timestamp);
 906 }
 907
 908 static uint64_t intel_pt_next_sample(struct intel_pt_decoder *decoder)
 909 {
 910         switch (decoder->period_type) {
 911         case INTEL_PT_PERIOD_INSTRUCTIONS:
 912                 return decoder->period - decoder->period_insn_cnt;
 913         case INTEL_PT_PERIOD_TICKS:
 914                 return intel_pt_next_period(decoder);
 915         case INTEL_PT_PERIOD_NONE:
 916         case INTEL_PT_PERIOD_MTC:
 917         default:
 918                 return 0;
 919         }
 920 }
 921
 922 static void intel_pt_sample_insn(struct intel_pt_decoder *decoder)
 923 {
 924         uint64_t timestamp, masked_timestamp;
 925
 926         switch (decoder->period_type) {
 927         case INTEL_PT_PERIOD_INSTRUCTIONS:
 928                 decoder->period_insn_cnt = 0;
 929                 break;
 930         case INTEL_PT_PERIOD_TICKS:
 931                 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
 932                 masked_timestamp = timestamp & decoder->period_mask;
 933                 decoder->last_masked_timestamp = masked_timestamp;
 934                 break;
 935         case INTEL_PT_PERIOD_NONE:
 936         case INTEL_PT_PERIOD_MTC:
 937         default:
 938                 break;
 939         }
 940
 941         decoder->state.type |= INTEL_PT_INSTRUCTION;
 942 }
 943
 944 static int intel_pt_walk_insn(struct intel_pt_decoder *decoder,
 945                               struct intel_pt_insn *intel_pt_insn, uint64_t ip)
 946 {
 947         uint64_t max_insn_cnt, insn_cnt = 0;
 948         int err;
 949
 950         if (!decoder->mtc_insn)
 951                 decoder->mtc_insn = true;
 952
 953         max_insn_cnt = intel_pt_next_sample(decoder);
 954
 955         err = decoder->walk_insn(intel_pt_insn, &insn_cnt, &decoder->ip, ip,
 956                                  max_insn_cnt, decoder->data);
 957
 958         decoder->tot_insn_cnt += insn_cnt;
 959         decoder->timestamp_insn_cnt += insn_cnt;
 960         decoder->sample_insn_cnt += insn_cnt;
 961         decoder->period_insn_cnt += insn_cnt;
 962
 963         if (err) {
 964                 decoder->no_progress = 0;
 965                 decoder->pkt_state = INTEL_PT_STATE_ERR2;
 966                 intel_pt_log_at("ERROR: Failed to get instruction",
 967                                 decoder->ip);
 968                 if (err == -ENOENT)
 969                         return -ENOLINK;
 970                 return -EILSEQ;
 971         }
 972
 973         if (ip && decoder->ip == ip) {
 974                 err = -EAGAIN;
 975                 goto out;
 976         }
 977
 978         if (max_insn_cnt && insn_cnt >= max_insn_cnt)
 979                 intel_pt_sample_insn(decoder);
 980
 981         if (intel_pt_insn->branch == INTEL_PT_BR_NO_BRANCH) {
 982                 decoder->state.type = INTEL_PT_INSTRUCTION;
 983                 decoder->state.from_ip = decoder->ip;
 984                 decoder->state.to_ip = 0;
 985                 decoder->ip += intel_pt_insn->length;
 986                 err = INTEL_PT_RETURN;
 987                 goto out;
 988         }
 989
 990         if (intel_pt_insn->op == INTEL_PT_OP_CALL) {
 991                 /* Zero-length calls are excluded */
 992                 if (intel_pt_insn->branch != INTEL_PT_BR_UNCONDITIONAL ||
 993                     intel_pt_insn->rel) {
 994                         err = intel_pt_push(&decoder->stack, decoder->ip +
 995                                             intel_pt_insn->length);
 996                         if (err)
 997                                 goto out;
 998                 }
 999         } else if (intel_pt_insn->op == INTEL_PT_OP_RET) {
1000                 decoder->ret_addr = intel_pt_pop(&decoder->stack);
1001         }
1002
1003         if (intel_pt_insn->branch == INTEL_PT_BR_UNCONDITIONAL) {
1004                 int cnt = decoder->no_progress++;
1005
1006                 decoder->state.from_ip = decoder->ip;
1007                 decoder->ip += intel_pt_insn->length +
1008                                 intel_pt_insn->rel;
1009                 decoder->state.to_ip = decoder->ip;
1010                 err = INTEL_PT_RETURN;
1011
1012                 /*
1013                  * Check for being stuck in a loop.  This can happen if a
1014                  * decoder error results in the decoder erroneously setting the
1015                  * ip to an address that is itself in an infinite loop that
1016                  * consumes no packets.  When that happens, there must be an
1017                  * unconditional branch.
1018                  */
1019                 if (cnt) {
1020                         if (cnt == 1) {
1021                                 decoder->stuck_ip = decoder->state.to_ip;
1022                                 decoder->stuck_ip_prd = 1;
1023                                 decoder->stuck_ip_cnt = 1;
1024                         } else if (cnt > INTEL_PT_MAX_LOOPS ||
1025                                    decoder->state.to_ip == decoder->stuck_ip) {
1026                                 intel_pt_log_at("ERROR: Never-ending loop",
1027                                                 decoder->state.to_ip);
1028                                 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1029                                 err = -ELOOP;
1030                                 goto out;
1031                         } else if (!--decoder->stuck_ip_cnt) {
1032                                 decoder->stuck_ip_prd += 1;
1033                                 decoder->stuck_ip_cnt = decoder->stuck_ip_prd;
1034                                 decoder->stuck_ip = decoder->state.to_ip;
1035                         }
1036                 }
1037                 goto out_no_progress;
1038         }
1039 out:
1040         decoder->no_progress = 0;
1041 out_no_progress:
1042         decoder->state.insn_op = intel_pt_insn->op;
1043         decoder->state.insn_len = intel_pt_insn->length;
1044         memcpy(decoder->state.insn, intel_pt_insn->buf,
1045                INTEL_PT_INSN_BUF_SZ);
1046
1047         if (decoder->tx_flags & INTEL_PT_IN_TX)
1048                 decoder->state.flags |= INTEL_PT_IN_TX;
1049
1050         return err;
1051 }
1052
1053 static bool intel_pt_fup_event(struct intel_pt_decoder *decoder)
1054 {
1055         bool ret = false;
1056
1057         if (decoder->set_fup_tx_flags) {
1058                 decoder->set_fup_tx_flags = false;
1059                 decoder->tx_flags = decoder->fup_tx_flags;
1060                 decoder->state.type = INTEL_PT_TRANSACTION;
1061                 decoder->state.from_ip = decoder->ip;
1062                 decoder->state.to_ip = 0;
1063                 decoder->state.flags = decoder->fup_tx_flags;
1064                 return true;
1065         }
1066         if (decoder->set_fup_ptw) {
1067                 decoder->set_fup_ptw = false;
1068                 decoder->state.type = INTEL_PT_PTW;
1069                 decoder->state.flags |= INTEL_PT_FUP_IP;
1070                 decoder->state.from_ip = decoder->ip;
1071                 decoder->state.to_ip = 0;
1072                 decoder->state.ptw_payload = decoder->fup_ptw_payload;
1073                 return true;
1074         }
1075         if (decoder->set_fup_mwait) {
1076                 decoder->set_fup_mwait = false;
1077                 decoder->state.type = INTEL_PT_MWAIT_OP;
1078                 decoder->state.from_ip = decoder->ip;
1079                 decoder->state.to_ip = 0;
1080                 decoder->state.mwait_payload = decoder->fup_mwait_payload;
1081                 ret = true;
1082         }
1083         if (decoder->set_fup_pwre) {
1084                 decoder->set_fup_pwre = false;
1085                 decoder->state.type |= INTEL_PT_PWR_ENTRY;
1086                 decoder->state.type &= ~INTEL_PT_BRANCH;
1087                 decoder->state.from_ip = decoder->ip;
1088                 decoder->state.to_ip = 0;
1089                 decoder->state.pwre_payload = decoder->fup_pwre_payload;
1090                 ret = true;
1091         }
1092         if (decoder->set_fup_exstop) {
1093                 decoder->set_fup_exstop = false;
1094                 decoder->state.type |= INTEL_PT_EX_STOP;
1095                 decoder->state.type &= ~INTEL_PT_BRANCH;
1096                 decoder->state.flags |= INTEL_PT_FUP_IP;
1097                 decoder->state.from_ip = decoder->ip;
1098                 decoder->state.to_ip = 0;
1099                 ret = true;
1100         }
1101         return ret;
1102 }
1103
1104 static inline bool intel_pt_fup_with_nlip(struct intel_pt_decoder *decoder,
1105                                           struct intel_pt_insn *intel_pt_insn,
1106                                           uint64_t ip, int err)
1107 {
1108         return decoder->flags & INTEL_PT_FUP_WITH_NLIP && !err &&
1109                intel_pt_insn->branch == INTEL_PT_BR_INDIRECT &&
1110                ip == decoder->ip + intel_pt_insn->length;
1111 }
1112
1113 static int intel_pt_walk_fup(struct intel_pt_decoder *decoder)
1114 {
1115         struct intel_pt_insn intel_pt_insn;
1116         uint64_t ip;
1117         int err;
1118
1119         ip = decoder->last_ip;
1120
1121         while (1) {
1122                 err = intel_pt_walk_insn(decoder, &intel_pt_insn, ip);
1123                 if (err == INTEL_PT_RETURN)
1124                         return 0;
1125                 if (err == -EAGAIN ||
1126                     intel_pt_fup_with_nlip(decoder, &intel_pt_insn, ip, err)) {
1127                         if (intel_pt_fup_event(decoder))
1128                                 return 0;
1129                         return -EAGAIN;
1130                 }
1131                 decoder->set_fup_tx_flags = false;
1132                 if (err)
1133                         return err;
1134
1135                 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1136                         intel_pt_log_at("ERROR: Unexpected indirect branch",
1137                                         decoder->ip);
1138                         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1139                         return -ENOENT;
1140                 }
1141
1142                 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1143                         intel_pt_log_at("ERROR: Unexpected conditional branch",
1144                                         decoder->ip);
1145                         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1146                         return -ENOENT;
1147                 }
1148
1149                 intel_pt_bug(decoder);
1150         }
1151 }
1152
1153 static int intel_pt_walk_tip(struct intel_pt_decoder *decoder)
1154 {
1155         struct intel_pt_insn intel_pt_insn;
1156         int err;
1157
1158         err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1159         if (err == INTEL_PT_RETURN &&
1160             decoder->pgd_ip &&
1161             decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1162             (decoder->state.type & INTEL_PT_BRANCH) &&
1163             decoder->pgd_ip(decoder->state.to_ip, decoder->data)) {
1164                 /* Unconditional branch leaving filter region */
1165                 decoder->no_progress = 0;
1166                 decoder->pge = false;
1167                 decoder->continuous_period = false;
1168                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1169                 decoder->state.type |= INTEL_PT_TRACE_END;
1170                 return 0;
1171         }
1172         if (err == INTEL_PT_RETURN)
1173                 return 0;
1174         if (err)
1175                 return err;
1176
1177         if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1178                 if (decoder->pkt_state == INTEL_PT_STATE_TIP_PGD) {
1179                         decoder->pge = false;
1180                         decoder->continuous_period = false;
1181                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1182                         decoder->state.from_ip = decoder->ip;
1183                         if (decoder->packet.count == 0) {
1184                                 decoder->state.to_ip = 0;
1185                         } else {
1186                                 decoder->state.to_ip = decoder->last_ip;
1187                                 decoder->ip = decoder->last_ip;
1188                         }
1189                         decoder->state.type |= INTEL_PT_TRACE_END;
1190                 } else {
1191                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1192                         decoder->state.from_ip = decoder->ip;
1193                         if (decoder->packet.count == 0) {
1194                                 decoder->state.to_ip = 0;
1195                         } else {
1196                                 decoder->state.to_ip = decoder->last_ip;
1197                                 decoder->ip = decoder->last_ip;
1198                         }
1199                 }
1200                 return 0;
1201         }
1202
1203         if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1204                 uint64_t to_ip = decoder->ip + intel_pt_insn.length +
1205                                  intel_pt_insn.rel;
1206
1207                 if (decoder->pgd_ip &&
1208                     decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1209                     decoder->pgd_ip(to_ip, decoder->data)) {
1210                         /* Conditional branch leaving filter region */
1211                         decoder->pge = false;
1212                         decoder->continuous_period = false;
1213                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1214                         decoder->ip = to_ip;
1215                         decoder->state.from_ip = decoder->ip;
1216                         decoder->state.to_ip = to_ip;
1217                         decoder->state.type |= INTEL_PT_TRACE_END;
1218                         return 0;
1219                 }
1220                 intel_pt_log_at("ERROR: Conditional branch when expecting indirect branch",
1221                                 decoder->ip);
1222                 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1223                 return -ENOENT;
1224         }
1225
1226         return intel_pt_bug(decoder);
1227 }
1228
1229 static int intel_pt_walk_tnt(struct intel_pt_decoder *decoder)
1230 {
1231         struct intel_pt_insn intel_pt_insn;
1232         int err;
1233
1234         while (1) {
1235                 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1236                 if (err == INTEL_PT_RETURN)
1237                         return 0;
1238                 if (err)
1239                         return err;
1240
1241                 if (intel_pt_insn.op == INTEL_PT_OP_RET) {
1242                         if (!decoder->return_compression) {
1243                                 intel_pt_log_at("ERROR: RET when expecting conditional branch",
1244                                                 decoder->ip);
1245                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1246                                 return -ENOENT;
1247                         }
1248                         if (!decoder->ret_addr) {
1249                                 intel_pt_log_at("ERROR: Bad RET compression (stack empty)",
1250                                                 decoder->ip);
1251                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1252                                 return -ENOENT;
1253                         }
1254                         if (!(decoder->tnt.payload & BIT63)) {
1255                                 intel_pt_log_at("ERROR: Bad RET compression (TNT=N)",
1256                                                 decoder->ip);
1257                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1258                                 return -ENOENT;
1259                         }
1260                         decoder->tnt.count -= 1;
1261                         if (!decoder->tnt.count)
1262                                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1263                         decoder->tnt.payload <<= 1;
1264                         decoder->state.from_ip = decoder->ip;
1265                         decoder->ip = decoder->ret_addr;
1266                         decoder->state.to_ip = decoder->ip;
1267                         return 0;
1268                 }
1269
1270                 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1271                         /* Handle deferred TIPs */
1272                         err = intel_pt_get_next_packet(decoder);
1273                         if (err)
1274                                 return err;
1275                         if (decoder->packet.type != INTEL_PT_TIP ||
1276                             decoder->packet.count == 0) {
1277                                 intel_pt_log_at("ERROR: Missing deferred TIP for indirect branch",
1278                                                 decoder->ip);
1279                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1280                                 decoder->pkt_step = 0;
1281                                 return -ENOENT;
1282                         }
1283                         intel_pt_set_last_ip(decoder);
1284                         decoder->state.from_ip = decoder->ip;
1285                         decoder->state.to_ip = decoder->last_ip;
1286                         decoder->ip = decoder->last_ip;
1287                         return 0;
1288                 }
1289
1290                 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1291                         decoder->tnt.count -= 1;
1292                         if (!decoder->tnt.count)
1293                                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1294                         if (decoder->tnt.payload & BIT63) {
1295                                 decoder->tnt.payload <<= 1;
1296                                 decoder->state.from_ip = decoder->ip;
1297                                 decoder->ip += intel_pt_insn.length +
1298                                                intel_pt_insn.rel;
1299                                 decoder->state.to_ip = decoder->ip;
1300                                 return 0;
1301                         }
1302                         /* Instruction sample for a non-taken branch */
1303                         if (decoder->state.type & INTEL_PT_INSTRUCTION) {
1304                                 decoder->tnt.payload <<= 1;
1305                                 decoder->state.type = INTEL_PT_INSTRUCTION;
1306                                 decoder->state.from_ip = decoder->ip;
1307                                 decoder->state.to_ip = 0;
1308                                 decoder->ip += intel_pt_insn.length;
1309                                 return 0;
1310                         }
1311                         decoder->ip += intel_pt_insn.length;
1312                         if (!decoder->tnt.count)
1313                                 return -EAGAIN;
1314                         decoder->tnt.payload <<= 1;
1315                         continue;
1316                 }
1317
1318                 return intel_pt_bug(decoder);
1319         }
1320 }
1321
1322 static int intel_pt_mode_tsx(struct intel_pt_decoder *decoder, bool *no_tip)
1323 {
1324         unsigned int fup_tx_flags;
1325         int err;
1326
1327         fup_tx_flags = decoder->packet.payload &
1328                        (INTEL_PT_IN_TX | INTEL_PT_ABORT_TX);
1329         err = intel_pt_get_next_packet(decoder);
1330         if (err)
1331                 return err;
1332         if (decoder->packet.type == INTEL_PT_FUP) {
1333                 decoder->fup_tx_flags = fup_tx_flags;
1334                 decoder->set_fup_tx_flags = true;
1335                 if (!(decoder->fup_tx_flags & INTEL_PT_ABORT_TX))
1336                         *no_tip = true;
1337         } else {
1338                 intel_pt_log_at("ERROR: Missing FUP after MODE.TSX",
1339                                 decoder->pos);
1340                 intel_pt_update_in_tx(decoder);
1341         }
1342         return 0;
1343 }
1344
1345 static void intel_pt_calc_tsc_timestamp(struct intel_pt_decoder *decoder)
1346 {
1347         uint64_t timestamp;
1348
1349         decoder->have_tma = false;
1350
1351         if (decoder->ref_timestamp) {
1352                 timestamp = decoder->packet.payload |
1353                             (decoder->ref_timestamp & (0xffULL << 56));
1354                 if (timestamp < decoder->ref_timestamp) {
1355                         if (decoder->ref_timestamp - timestamp > (1ULL << 55))
1356                                 timestamp += (1ULL << 56);
1357                 } else {
1358                         if (timestamp - decoder->ref_timestamp > (1ULL << 55))
1359                                 timestamp -= (1ULL << 56);
1360                 }
1361                 decoder->tsc_timestamp = timestamp;
1362                 decoder->timestamp = timestamp;
1363                 decoder->ref_timestamp = 0;
1364                 decoder->timestamp_insn_cnt = 0;
1365         } else if (decoder->timestamp) {
1366                 timestamp = decoder->packet.payload |
1367                             (decoder->timestamp & (0xffULL << 56));
1368                 decoder->tsc_timestamp = timestamp;
1369                 if (timestamp < decoder->timestamp &&
1370                     decoder->timestamp - timestamp < decoder->tsc_slip) {
1371                         intel_pt_log_to("Suppressing backwards timestamp",
1372                                         timestamp);
1373                         timestamp = decoder->timestamp;
1374                 }
1375                 if (timestamp < decoder->timestamp) {
1376                         intel_pt_log_to("Wraparound timestamp", timestamp);
1377                         timestamp += (1ULL << 56);
1378                         decoder->tsc_timestamp = timestamp;
1379                 }
1380                 decoder->timestamp = timestamp;
1381                 decoder->timestamp_insn_cnt = 0;
1382         }
1383
1384         if (decoder->last_packet_type == INTEL_PT_CYC) {
1385                 decoder->cyc_ref_timestamp = decoder->timestamp;
1386                 decoder->cycle_cnt = 0;
1387                 decoder->have_calc_cyc_to_tsc = false;
1388                 intel_pt_calc_cyc_to_tsc(decoder, false);
1389         }
1390
1391         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1392 }
1393
1394 static int intel_pt_overflow(struct intel_pt_decoder *decoder)
1395 {
1396         intel_pt_log("ERROR: Buffer overflow\n");
1397         intel_pt_clear_tx_flags(decoder);
1398         decoder->timestamp_insn_cnt = 0;
1399         decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1400         decoder->overflow = true;
1401         return -EOVERFLOW;
1402 }
1403
1404 static void intel_pt_calc_tma(struct intel_pt_decoder *decoder)
1405 {
1406         uint32_t ctc = decoder->packet.payload;
1407         uint32_t fc = decoder->packet.count;
1408         uint32_t ctc_rem = ctc & decoder->ctc_rem_mask;
1409
1410         if (!decoder->tsc_ctc_ratio_d)
1411                 return;
1412
1413         decoder->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
1414         decoder->ctc_timestamp = decoder->tsc_timestamp - fc;
1415         if (decoder->tsc_ctc_mult) {
1416                 decoder->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
1417         } else {
1418                 decoder->ctc_timestamp -= multdiv(ctc_rem,
1419                                                   decoder->tsc_ctc_ratio_n,
1420                                                   decoder->tsc_ctc_ratio_d);
1421         }
1422         decoder->ctc_delta = 0;
1423         decoder->have_tma = true;
1424         decoder->fixup_last_mtc = true;
1425         intel_pt_log("CTC timestamp " x64_fmt " last MTC %#x  CTC rem %#x\n",
1426                      decoder->ctc_timestamp, decoder->last_mtc, ctc_rem);
1427 }
1428
1429 static void intel_pt_calc_mtc_timestamp(struct intel_pt_decoder *decoder)
1430 {
1431         uint64_t timestamp;
1432         uint32_t mtc, mtc_delta;
1433
1434         if (!decoder->have_tma)
1435                 return;
1436
1437         mtc = decoder->packet.payload;
1438
1439         if (decoder->mtc_shift > 8 && decoder->fixup_last_mtc) {
1440                 decoder->fixup_last_mtc = false;
1441                 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
1442                                         &decoder->last_mtc);
1443         }
1444
1445         if (mtc > decoder->last_mtc)
1446                 mtc_delta = mtc - decoder->last_mtc;
1447         else
1448                 mtc_delta = mtc + 256 - decoder->last_mtc;
1449
1450         decoder->ctc_delta += mtc_delta << decoder->mtc_shift;
1451
1452         if (decoder->tsc_ctc_mult) {
1453                 timestamp = decoder->ctc_timestamp +
1454                             decoder->ctc_delta * decoder->tsc_ctc_mult;
1455         } else {
1456                 timestamp = decoder->ctc_timestamp +
1457                             multdiv(decoder->ctc_delta,
1458                                     decoder->tsc_ctc_ratio_n,
1459                                     decoder->tsc_ctc_ratio_d);
1460         }
1461
1462         if (timestamp < decoder->timestamp)
1463                 intel_pt_log("Suppressing MTC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1464                              timestamp, decoder->timestamp);
1465         else
1466                 decoder->timestamp = timestamp;
1467
1468         decoder->timestamp_insn_cnt = 0;
1469         decoder->last_mtc = mtc;
1470
1471         if (decoder->last_packet_type == INTEL_PT_CYC) {
1472                 decoder->cyc_ref_timestamp = decoder->timestamp;
1473                 decoder->cycle_cnt = 0;
1474                 decoder->have_calc_cyc_to_tsc = false;
1475                 intel_pt_calc_cyc_to_tsc(decoder, true);
1476         }
1477
1478         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1479 }
1480
1481 static void intel_pt_calc_cbr(struct intel_pt_decoder *decoder)
1482 {
1483         unsigned int cbr = decoder->packet.payload & 0xff;
1484
1485         decoder->cbr_payload = decoder->packet.payload;
1486
1487         if (decoder->cbr == cbr)
1488                 return;
1489
1490         decoder->cbr = cbr;
1491         decoder->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
1492 }
1493
1494 static void intel_pt_calc_cyc_timestamp(struct intel_pt_decoder *decoder)
1495 {
1496         uint64_t timestamp = decoder->cyc_ref_timestamp;
1497
1498         decoder->have_cyc = true;
1499
1500         decoder->cycle_cnt += decoder->packet.payload;
1501
1502         if (!decoder->cyc_ref_timestamp)
1503                 return;
1504
1505         if (decoder->have_calc_cyc_to_tsc)
1506                 timestamp += decoder->cycle_cnt * decoder->calc_cyc_to_tsc;
1507         else if (decoder->cbr)
1508                 timestamp += decoder->cycle_cnt * decoder->cbr_cyc_to_tsc;
1509         else
1510                 return;
1511
1512         if (timestamp < decoder->timestamp)
1513                 intel_pt_log("Suppressing CYC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1514                              timestamp, decoder->timestamp);
1515         else
1516                 decoder->timestamp = timestamp;
1517
1518         decoder->timestamp_insn_cnt = 0;
1519
1520         intel_pt_log_to("Setting timestamp", decoder->timestamp);
1521 }
1522
1523 /* Walk PSB+ packets when already in sync. */
1524 static int intel_pt_walk_psbend(struct intel_pt_decoder *decoder)
1525 {
1526         int err;
1527
1528         while (1) {
1529                 err = intel_pt_get_next_packet(decoder);
1530                 if (err)
1531                         return err;
1532
1533                 switch (decoder->packet.type) {
1534                 case INTEL_PT_PSBEND:
1535                         return 0;
1536
1537                 case INTEL_PT_TIP_PGD:
1538                 case INTEL_PT_TIP_PGE:
1539                 case INTEL_PT_TIP:
1540                 case INTEL_PT_TNT:
1541                 case INTEL_PT_TRACESTOP:
1542                 case INTEL_PT_BAD:
1543                 case INTEL_PT_PSB:
1544                 case INTEL_PT_PTWRITE:
1545                 case INTEL_PT_PTWRITE_IP:
1546                 case INTEL_PT_EXSTOP:
1547                 case INTEL_PT_EXSTOP_IP:
1548                 case INTEL_PT_MWAIT:
1549                 case INTEL_PT_PWRE:
1550                 case INTEL_PT_PWRX:
1551                         decoder->have_tma = false;
1552                         intel_pt_log("ERROR: Unexpected packet\n");
1553                         return -EAGAIN;
1554
1555                 case INTEL_PT_OVF:
1556                         return intel_pt_overflow(decoder);
1557
1558                 case INTEL_PT_TSC:
1559                         intel_pt_calc_tsc_timestamp(decoder);
1560                         break;
1561
1562                 case INTEL_PT_TMA:
1563                         intel_pt_calc_tma(decoder);
1564                         break;
1565
1566                 case INTEL_PT_CBR:
1567                         intel_pt_calc_cbr(decoder);
1568                         break;
1569
1570                 case INTEL_PT_MODE_EXEC:
1571                         decoder->exec_mode = decoder->packet.payload;
1572                         break;
1573
1574                 case INTEL_PT_PIP:
1575                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1576                         break;
1577
1578                 case INTEL_PT_FUP:
1579                         decoder->pge = true;
1580                         if (decoder->packet.count)
1581                                 intel_pt_set_last_ip(decoder);
1582                         break;
1583
1584                 case INTEL_PT_MODE_TSX:
1585                         intel_pt_update_in_tx(decoder);
1586                         break;
1587
1588                 case INTEL_PT_MTC:
1589                         intel_pt_calc_mtc_timestamp(decoder);
1590                         if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1591                                 decoder->state.type |= INTEL_PT_INSTRUCTION;
1592                         break;
1593
1594                 case INTEL_PT_CYC:
1595                 case INTEL_PT_VMCS:
1596                 case INTEL_PT_MNT:
1597                 case INTEL_PT_PAD:
1598                 default:
1599                         break;
1600                 }
1601         }
1602 }
1603
1604 static int intel_pt_walk_fup_tip(struct intel_pt_decoder *decoder)
1605 {
1606         int err;
1607
1608         if (decoder->tx_flags & INTEL_PT_ABORT_TX) {
1609                 decoder->tx_flags = 0;
1610                 decoder->state.flags &= ~INTEL_PT_IN_TX;
1611                 decoder->state.flags |= INTEL_PT_ABORT_TX;
1612         } else {
1613                 decoder->state.flags |= INTEL_PT_ASYNC;
1614         }
1615
1616         while (1) {
1617                 err = intel_pt_get_next_packet(decoder);
1618                 if (err)
1619                         return err;
1620
1621                 switch (decoder->packet.type) {
1622                 case INTEL_PT_TNT:
1623                 case INTEL_PT_FUP:
1624                 case INTEL_PT_TRACESTOP:
1625                 case INTEL_PT_PSB:
1626                 case INTEL_PT_TSC:
1627                 case INTEL_PT_TMA:
1628                 case INTEL_PT_MODE_TSX:
1629                 case INTEL_PT_BAD:
1630                 case INTEL_PT_PSBEND:
1631                 case INTEL_PT_PTWRITE:
1632                 case INTEL_PT_PTWRITE_IP:
1633                 case INTEL_PT_EXSTOP:
1634                 case INTEL_PT_EXSTOP_IP:
1635                 case INTEL_PT_MWAIT:
1636                 case INTEL_PT_PWRE:
1637                 case INTEL_PT_PWRX:
1638                         intel_pt_log("ERROR: Missing TIP after FUP\n");
1639                         decoder->pkt_state = INTEL_PT_STATE_ERR3;
1640                         decoder->pkt_step = 0;
1641                         return -ENOENT;
1642
1643                 case INTEL_PT_CBR:
1644                         intel_pt_calc_cbr(decoder);
1645                         break;
1646
1647                 case INTEL_PT_OVF:
1648                         return intel_pt_overflow(decoder);
1649
1650                 case INTEL_PT_TIP_PGD:
1651                         decoder->state.from_ip = decoder->ip;
1652                         if (decoder->packet.count == 0) {
1653                                 decoder->state.to_ip = 0;
1654                         } else {
1655                                 intel_pt_set_ip(decoder);
1656                                 decoder->state.to_ip = decoder->ip;
1657                         }
1658                         decoder->pge = false;
1659                         decoder->continuous_period = false;
1660                         decoder->state.type |= INTEL_PT_TRACE_END;
1661                         return 0;
1662
1663                 case INTEL_PT_TIP_PGE:
1664                         decoder->pge = true;
1665                         intel_pt_log("Omitting PGE ip " x64_fmt "\n",
1666                                      decoder->ip);
1667                         decoder->state.from_ip = 0;
1668                         if (decoder->packet.count == 0) {
1669                                 decoder->state.to_ip = 0;
1670                         } else {
1671                                 intel_pt_set_ip(decoder);
1672                                 decoder->state.to_ip = decoder->ip;
1673                         }
1674                         decoder->state.type |= INTEL_PT_TRACE_BEGIN;
1675                         return 0;
1676
1677                 case INTEL_PT_TIP:
1678                         decoder->state.from_ip = decoder->ip;
1679                         if (decoder->packet.count == 0) {
1680                                 decoder->state.to_ip = 0;
1681                         } else {
1682                                 intel_pt_set_ip(decoder);
1683                                 decoder->state.to_ip = decoder->ip;
1684                         }
1685                         return 0;
1686
1687                 case INTEL_PT_PIP:
1688                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1689                         break;
1690
1691                 case INTEL_PT_MTC:
1692                         intel_pt_calc_mtc_timestamp(decoder);
1693                         if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1694                                 decoder->state.type |= INTEL_PT_INSTRUCTION;
1695                         break;
1696
1697                 case INTEL_PT_CYC:
1698                         intel_pt_calc_cyc_timestamp(decoder);
1699                         break;
1700
1701                 case INTEL_PT_MODE_EXEC:
1702                         decoder->exec_mode = decoder->packet.payload;
1703                         break;
1704
1705                 case INTEL_PT_VMCS:
1706                 case INTEL_PT_MNT:
1707                 case INTEL_PT_PAD:
1708                         break;
1709
1710                 default:
1711                         return intel_pt_bug(decoder);
1712                 }
1713         }
1714 }
1715
1716 static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)
1717 {
1718         bool no_tip = false;
1719         int err;
1720
1721         while (1) {
1722                 err = intel_pt_get_next_packet(decoder);
1723                 if (err)
1724                         return err;
1725 next:
1726                 switch (decoder->packet.type) {
1727                 case INTEL_PT_TNT:
1728                         if (!decoder->packet.count)
1729                                 break;
1730                         decoder->tnt = decoder->packet;
1731                         decoder->pkt_state = INTEL_PT_STATE_TNT;
1732                         err = intel_pt_walk_tnt(decoder);
1733                         if (err == -EAGAIN)
1734                                 break;
1735                         return err;
1736
1737                 case INTEL_PT_TIP_PGD:
1738                         if (decoder->packet.count != 0)
1739                                 intel_pt_set_last_ip(decoder);
1740                         decoder->pkt_state = INTEL_PT_STATE_TIP_PGD;
1741                         return intel_pt_walk_tip(decoder);
1742
1743                 case INTEL_PT_TIP_PGE: {
1744                         decoder->pge = true;
1745                         if (decoder->packet.count == 0) {
1746                                 intel_pt_log_at("Skipping zero TIP.PGE",
1747                                                 decoder->pos);
1748                                 break;
1749                         }
1750                         intel_pt_set_ip(decoder);
1751                         decoder->state.from_ip = 0;
1752                         decoder->state.to_ip = decoder->ip;
1753                         decoder->state.type |= INTEL_PT_TRACE_BEGIN;
1754                         return 0;
1755                 }
1756
1757                 case INTEL_PT_OVF:
1758                         return intel_pt_overflow(decoder);
1759
1760                 case INTEL_PT_TIP:
1761                         if (decoder->packet.count != 0)
1762                                 intel_pt_set_last_ip(decoder);
1763                         decoder->pkt_state = INTEL_PT_STATE_TIP;
1764                         return intel_pt_walk_tip(decoder);
1765
1766                 case INTEL_PT_FUP:
1767                         if (decoder->packet.count == 0) {
1768                                 intel_pt_log_at("Skipping zero FUP",
1769                                                 decoder->pos);
1770                                 no_tip = false;
1771                                 break;
1772                         }
1773                         intel_pt_set_last_ip(decoder);
1774                         if (!decoder->branch_enable) {
1775                                 decoder->ip = decoder->last_ip;
1776                                 if (intel_pt_fup_event(decoder))
1777                                         return 0;
1778                                 no_tip = false;
1779                                 break;
1780                         }
1781                         if (decoder->set_fup_mwait)
1782                                 no_tip = true;
1783                         err = intel_pt_walk_fup(decoder);
1784                         if (err != -EAGAIN) {
1785                                 if (err)
1786                                         return err;
1787                                 if (no_tip)
1788                                         decoder->pkt_state =
1789                                                 INTEL_PT_STATE_FUP_NO_TIP;
1790                                 else
1791                                         decoder->pkt_state = INTEL_PT_STATE_FUP;
1792                                 return 0;
1793                         }
1794                         if (no_tip) {
1795                                 no_tip = false;
1796                                 break;
1797                         }
1798                         return intel_pt_walk_fup_tip(decoder);
1799
1800                 case INTEL_PT_TRACESTOP:
1801                         decoder->pge = false;
1802                         decoder->continuous_period = false;
1803                         intel_pt_clear_tx_flags(decoder);
1804                         decoder->have_tma = false;
1805                         break;
1806
1807                 case INTEL_PT_PSB:
1808                         decoder->last_ip = 0;
1809                         decoder->have_last_ip = true;
1810                         intel_pt_clear_stack(&decoder->stack);
1811                         err = intel_pt_walk_psbend(decoder);
1812                         if (err == -EAGAIN)
1813                                 goto next;
1814                         if (err)
1815                                 return err;
1816                         break;
1817
1818                 case INTEL_PT_PIP:
1819                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1820                         break;
1821
1822                 case INTEL_PT_MTC:
1823                         intel_pt_calc_mtc_timestamp(decoder);
1824                         if (decoder->period_type != INTEL_PT_PERIOD_MTC)
1825                                 break;
1826                         /*
1827                          * Ensure that there has been an instruction since the
1828                          * last MTC.
1829                          */
1830                         if (!decoder->mtc_insn)
1831                                 break;
1832                         decoder->mtc_insn = false;
1833                         /* Ensure that there is a timestamp */
1834                         if (!decoder->timestamp)
1835                                 break;
1836                         decoder->state.type = INTEL_PT_INSTRUCTION;
1837                         decoder->state.from_ip = decoder->ip;
1838                         decoder->state.to_ip = 0;
1839                         decoder->mtc_insn = false;
1840                         return 0;
1841
1842                 case INTEL_PT_TSC:
1843                         intel_pt_calc_tsc_timestamp(decoder);
1844                         break;
1845
1846                 case INTEL_PT_TMA:
1847                         intel_pt_calc_tma(decoder);
1848                         break;
1849
1850                 case INTEL_PT_CYC:
1851                         intel_pt_calc_cyc_timestamp(decoder);
1852                         break;
1853
1854                 case INTEL_PT_CBR:
1855                         intel_pt_calc_cbr(decoder);
1856                         if (!decoder->branch_enable &&
1857                             decoder->cbr != decoder->cbr_seen) {
1858                                 decoder->cbr_seen = decoder->cbr;
1859                                 decoder->state.type = INTEL_PT_CBR_CHG;
1860                                 decoder->state.from_ip = decoder->ip;
1861                                 decoder->state.to_ip = 0;
1862                                 decoder->state.cbr_payload =
1863                                                         decoder->packet.payload;
1864                                 return 0;
1865                         }
1866                         break;
1867
1868                 case INTEL_PT_MODE_EXEC:
1869                         decoder->exec_mode = decoder->packet.payload;
1870                         break;
1871
1872                 case INTEL_PT_MODE_TSX:
1873                         /* MODE_TSX need not be followed by FUP */
1874                         if (!decoder->pge) {
1875                                 intel_pt_update_in_tx(decoder);
1876                                 break;
1877                         }
1878                         err = intel_pt_mode_tsx(decoder, &no_tip);
1879                         if (err)
1880                                 return err;
1881                         goto next;
1882
1883                 case INTEL_PT_BAD: /* Does not happen */
1884                         return intel_pt_bug(decoder);
1885
1886                 case INTEL_PT_PSBEND:
1887                 case INTEL_PT_VMCS:
1888                 case INTEL_PT_MNT:
1889                 case INTEL_PT_PAD:
1890                         break;
1891
1892                 case INTEL_PT_PTWRITE_IP:
1893                         decoder->fup_ptw_payload = decoder->packet.payload;
1894                         err = intel_pt_get_next_packet(decoder);
1895                         if (err)
1896                                 return err;
1897                         if (decoder->packet.type == INTEL_PT_FUP) {
1898                                 decoder->set_fup_ptw = true;
1899                                 no_tip = true;
1900                         } else {
1901                                 intel_pt_log_at("ERROR: Missing FUP after PTWRITE",
1902                                                 decoder->pos);
1903                         }
1904                         goto next;
1905
1906                 case INTEL_PT_PTWRITE:
1907                         decoder->state.type = INTEL_PT_PTW;
1908                         decoder->state.from_ip = decoder->ip;
1909                         decoder->state.to_ip = 0;
1910                         decoder->state.ptw_payload = decoder->packet.payload;
1911                         return 0;
1912
1913                 case INTEL_PT_MWAIT:
1914                         decoder->fup_mwait_payload = decoder->packet.payload;
1915                         decoder->set_fup_mwait = true;
1916                         break;
1917
1918                 case INTEL_PT_PWRE:
1919                         if (decoder->set_fup_mwait) {
1920                                 decoder->fup_pwre_payload =
1921                                                         decoder->packet.payload;
1922                                 decoder->set_fup_pwre = true;
1923                                 break;
1924                         }
1925                         decoder->state.type = INTEL_PT_PWR_ENTRY;
1926                         decoder->state.from_ip = decoder->ip;
1927                         decoder->state.to_ip = 0;
1928                         decoder->state.pwrx_payload = decoder->packet.payload;
1929                         return 0;
1930
1931                 case INTEL_PT_EXSTOP_IP:
1932                         err = intel_pt_get_next_packet(decoder);
1933                         if (err)
1934                                 return err;
1935                         if (decoder->packet.type == INTEL_PT_FUP) {
1936                                 decoder->set_fup_exstop = true;
1937                                 no_tip = true;
1938                         } else {
1939                                 intel_pt_log_at("ERROR: Missing FUP after EXSTOP",
1940                                                 decoder->pos);
1941                         }
1942                         goto next;
1943
1944                 case INTEL_PT_EXSTOP:
1945                         decoder->state.type = INTEL_PT_EX_STOP;
1946                         decoder->state.from_ip = decoder->ip;
1947                         decoder->state.to_ip = 0;
1948                         return 0;
1949
1950                 case INTEL_PT_PWRX:
1951                         decoder->state.type = INTEL_PT_PWR_EXIT;
1952                         decoder->state.from_ip = decoder->ip;
1953                         decoder->state.to_ip = 0;
1954                         decoder->state.pwrx_payload = decoder->packet.payload;
1955                         return 0;
1956
1957                 default:
1958                         return intel_pt_bug(decoder);
1959                 }
1960         }
1961 }
1962
1963 static inline bool intel_pt_have_ip(struct intel_pt_decoder *decoder)
1964 {
1965         return decoder->packet.count &&
1966                (decoder->have_last_ip || decoder->packet.count == 3 ||
1967                 decoder->packet.count == 6);
1968 }
1969
1970 /* Walk PSB+ packets to get in sync. */
1971 static int intel_pt_walk_psb(struct intel_pt_decoder *decoder)
1972 {
1973         int err;
1974
1975         while (1) {
1976                 err = intel_pt_get_next_packet(decoder);
1977                 if (err)
1978                         return err;
1979
1980                 switch (decoder->packet.type) {
1981                 case INTEL_PT_TIP_PGD:
1982                         decoder->continuous_period = false;
1983                         __fallthrough;
1984                 case INTEL_PT_TIP_PGE:
1985                 case INTEL_PT_TIP:
1986                 case INTEL_PT_PTWRITE:
1987                 case INTEL_PT_PTWRITE_IP:
1988                 case INTEL_PT_EXSTOP:
1989                 case INTEL_PT_EXSTOP_IP:
1990                 case INTEL_PT_MWAIT:
1991                 case INTEL_PT_PWRE:
1992                 case INTEL_PT_PWRX:
1993                         intel_pt_log("ERROR: Unexpected packet\n");
1994                         return -ENOENT;
1995
1996                 case INTEL_PT_FUP:
1997                         decoder->pge = true;
1998                         if (intel_pt_have_ip(decoder)) {
1999                                 uint64_t current_ip = decoder->ip;
2000
2001                                 intel_pt_set_ip(decoder);
2002                                 if (current_ip)
2003                                         intel_pt_log_to("Setting IP",
2004                                                         decoder->ip);
2005                         }
2006                         break;
2007
2008                 case INTEL_PT_MTC:
2009                         intel_pt_calc_mtc_timestamp(decoder);
2010                         break;
2011
2012                 case INTEL_PT_TSC:
2013                         intel_pt_calc_tsc_timestamp(decoder);
2014                         break;
2015
2016                 case INTEL_PT_TMA:
2017                         intel_pt_calc_tma(decoder);
2018                         break;
2019
2020                 case INTEL_PT_CYC:
2021                         intel_pt_calc_cyc_timestamp(decoder);
2022                         break;
2023
2024                 case INTEL_PT_CBR:
2025                         intel_pt_calc_cbr(decoder);
2026                         break;
2027
2028                 case INTEL_PT_PIP:
2029                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2030                         break;
2031
2032                 case INTEL_PT_MODE_EXEC:
2033                         decoder->exec_mode = decoder->packet.payload;
2034                         break;
2035
2036                 case INTEL_PT_MODE_TSX:
2037                         intel_pt_update_in_tx(decoder);
2038                         break;
2039
2040                 case INTEL_PT_TRACESTOP:
2041                         decoder->pge = false;
2042                         decoder->continuous_period = false;
2043                         intel_pt_clear_tx_flags(decoder);
2044                         __fallthrough;
2045
2046                 case INTEL_PT_TNT:
2047                         decoder->have_tma = false;
2048                         intel_pt_log("ERROR: Unexpected packet\n");
2049                         if (decoder->ip)
2050                                 decoder->pkt_state = INTEL_PT_STATE_ERR4;
2051                         else
2052                                 decoder->pkt_state = INTEL_PT_STATE_ERR3;
2053                         return -ENOENT;
2054
2055                 case INTEL_PT_BAD: /* Does not happen */
2056                         return intel_pt_bug(decoder);
2057
2058                 case INTEL_PT_OVF:
2059                         return intel_pt_overflow(decoder);
2060
2061                 case INTEL_PT_PSBEND:
2062                         return 0;
2063
2064                 case INTEL_PT_PSB:
2065                 case INTEL_PT_VMCS:
2066                 case INTEL_PT_MNT:
2067                 case INTEL_PT_PAD:
2068                 default:
2069                         break;
2070                 }
2071         }
2072 }
2073
2074 static int intel_pt_walk_to_ip(struct intel_pt_decoder *decoder)
2075 {
2076         int err;
2077
2078         while (1) {
2079                 err = intel_pt_get_next_packet(decoder);
2080                 if (err)
2081                         return err;
2082
2083                 switch (decoder->packet.type) {
2084                 case INTEL_PT_TIP_PGD:
2085                         decoder->continuous_period = false;
2086                         __fallthrough;
2087                 case INTEL_PT_TIP_PGE:
2088                 case INTEL_PT_TIP:
2089                         decoder->pge = decoder->packet.type != INTEL_PT_TIP_PGD;
2090                         if (intel_pt_have_ip(decoder))
2091                                 intel_pt_set_ip(decoder);
2092                         if (!decoder->ip)
2093                                 break;
2094                         if (decoder->packet.type == INTEL_PT_TIP_PGE)
2095                                 decoder->state.type |= INTEL_PT_TRACE_BEGIN;
2096                         if (decoder->packet.type == INTEL_PT_TIP_PGD)
2097                                 decoder->state.type |= INTEL_PT_TRACE_END;
2098                         return 0;
2099
2100                 case INTEL_PT_FUP:
2101                         if (intel_pt_have_ip(decoder))
2102                                 intel_pt_set_ip(decoder);
2103                         if (decoder->ip)
2104                                 return 0;
2105                         break;
2106
2107                 case INTEL_PT_MTC:
2108                         intel_pt_calc_mtc_timestamp(decoder);
2109                         break;
2110
2111                 case INTEL_PT_TSC:
2112                         intel_pt_calc_tsc_timestamp(decoder);
2113                         break;
2114
2115                 case INTEL_PT_TMA:
2116                         intel_pt_calc_tma(decoder);
2117                         break;
2118
2119                 case INTEL_PT_CYC:
2120                         intel_pt_calc_cyc_timestamp(decoder);
2121                         break;
2122
2123                 case INTEL_PT_CBR:
2124                         intel_pt_calc_cbr(decoder);
2125                         break;
2126
2127                 case INTEL_PT_PIP:
2128                         decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2129                         break;
2130
2131                 case INTEL_PT_MODE_EXEC:
2132                         decoder->exec_mode = decoder->packet.payload;
2133                         break;
2134
2135                 case INTEL_PT_MODE_TSX:
2136                         intel_pt_update_in_tx(decoder);
2137                         break;
2138
2139                 case INTEL_PT_OVF:
2140                         return intel_pt_overflow(decoder);
2141
2142                 case INTEL_PT_BAD: /* Does not happen */
2143                         return intel_pt_bug(decoder);
2144
2145                 case INTEL_PT_TRACESTOP:
2146                         decoder->pge = false;
2147                         decoder->continuous_period = false;
2148                         intel_pt_clear_tx_flags(decoder);
2149                         decoder->have_tma = false;
2150                         break;
2151
2152                 case INTEL_PT_PSB:
2153                         decoder->last_ip = 0;
2154                         decoder->have_last_ip = true;
2155                         intel_pt_clear_stack(&decoder->stack);
2156                         err = intel_pt_walk_psb(decoder);
2157                         if (err)
2158                                 return err;
2159                         if (decoder->ip) {
2160                                 /* Do not have a sample */
2161                                 decoder->state.type = 0;
2162                                 return 0;
2163                         }
2164                         break;
2165
2166                 case INTEL_PT_TNT:
2167                 case INTEL_PT_PSBEND:
2168                 case INTEL_PT_VMCS:
2169                 case INTEL_PT_MNT:
2170                 case INTEL_PT_PAD:
2171                 case INTEL_PT_PTWRITE:
2172                 case INTEL_PT_PTWRITE_IP:
2173                 case INTEL_PT_EXSTOP:
2174                 case INTEL_PT_EXSTOP_IP:
2175                 case INTEL_PT_MWAIT:
2176                 case INTEL_PT_PWRE:
2177                 case INTEL_PT_PWRX:
2178                 default:
2179                         break;
2180                 }
2181         }
2182 }
2183
2184 static int intel_pt_sync_ip(struct intel_pt_decoder *decoder)
2185 {
2186         int err;
2187
2188         decoder->set_fup_tx_flags = false;
2189         decoder->set_fup_ptw = false;
2190         decoder->set_fup_mwait = false;
2191         decoder->set_fup_pwre = false;
2192         decoder->set_fup_exstop = false;
2193
2194         if (!decoder->branch_enable) {
2195                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2196                 decoder->overflow = false;
2197                 decoder->state.type = 0; /* Do not have a sample */
2198                 return 0;
2199         }
2200
2201         intel_pt_log("Scanning for full IP\n");
2202         err = intel_pt_walk_to_ip(decoder);
2203         if (err)
2204                 return err;
2205
2206         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2207         decoder->overflow = false;
2208
2209         decoder->state.from_ip = 0;
2210         decoder->state.to_ip = decoder->ip;
2211         intel_pt_log_to("Setting IP", decoder->ip);
2212
2213         return 0;
2214 }
2215
2216 static int intel_pt_part_psb(struct intel_pt_decoder *decoder)
2217 {
2218         const unsigned char *end = decoder->buf + decoder->len;
2219         size_t i;
2220
2221         for (i = INTEL_PT_PSB_LEN - 1; i; i--) {
2222                 if (i > decoder->len)
2223                         continue;
2224                 if (!memcmp(end - i, INTEL_PT_PSB_STR, i))
2225                         return i;
2226         }
2227         return 0;
2228 }
2229
2230 static int intel_pt_rest_psb(struct intel_pt_decoder *decoder, int part_psb)
2231 {
2232         size_t rest_psb = INTEL_PT_PSB_LEN - part_psb;
2233         const char *psb = INTEL_PT_PSB_STR;
2234
2235         if (rest_psb > decoder->len ||
2236             memcmp(decoder->buf, psb + part_psb, rest_psb))
2237                 return 0;
2238
2239         return rest_psb;
2240 }
2241
2242 static int intel_pt_get_split_psb(struct intel_pt_decoder *decoder,
2243                                   int part_psb)
2244 {
2245         int rest_psb, ret;
2246
2247         decoder->pos += decoder->len;
2248         decoder->len = 0;
2249
2250         ret = intel_pt_get_next_data(decoder);
2251         if (ret)
2252                 return ret;
2253
2254         rest_psb = intel_pt_rest_psb(decoder, part_psb);
2255         if (!rest_psb)
2256                 return 0;
2257
2258         decoder->pos -= part_psb;
2259         decoder->next_buf = decoder->buf + rest_psb;
2260         decoder->next_len = decoder->len - rest_psb;
2261         memcpy(decoder->temp_buf, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2262         decoder->buf = decoder->temp_buf;
2263         decoder->len = INTEL_PT_PSB_LEN;
2264
2265         return 0;
2266 }
2267
2268 static int intel_pt_scan_for_psb(struct intel_pt_decoder *decoder)
2269 {
2270         unsigned char *next;
2271         int ret;
2272
2273         intel_pt_log("Scanning for PSB\n");
2274         while (1) {
2275                 if (!decoder->len) {
2276                         ret = intel_pt_get_next_data(decoder);
2277                         if (ret)
2278                                 return ret;
2279                 }
2280
2281                 next = memmem(decoder->buf, decoder->len, INTEL_PT_PSB_STR,
2282                               INTEL_PT_PSB_LEN);
2283                 if (!next) {
2284                         int part_psb;
2285
2286                         part_psb = intel_pt_part_psb(decoder);
2287                         if (part_psb) {
2288                                 ret = intel_pt_get_split_psb(decoder, part_psb);
2289                                 if (ret)
2290                                         return ret;
2291                         } else {
2292                                 decoder->pos += decoder->len;
2293                                 decoder->len = 0;
2294                         }
2295                         continue;
2296                 }
2297
2298                 decoder->pkt_step = next - decoder->buf;
2299                 return intel_pt_get_next_packet(decoder);
2300         }
2301 }
2302
2303 static int intel_pt_sync(struct intel_pt_decoder *decoder)
2304 {
2305         int err;
2306
2307         decoder->pge = false;
2308         decoder->continuous_period = false;
2309         decoder->have_last_ip = false;
2310         decoder->last_ip = 0;
2311         decoder->ip = 0;
2312         intel_pt_clear_stack(&decoder->stack);
2313
2314         err = intel_pt_scan_for_psb(decoder);
2315         if (err)
2316                 return err;
2317
2318         decoder->have_last_ip = true;
2319         decoder->pkt_state = INTEL_PT_STATE_NO_IP;
2320
2321         err = intel_pt_walk_psb(decoder);
2322         if (err)
2323                 return err;
2324
2325         if (decoder->ip) {
2326                 decoder->state.type = 0; /* Do not have a sample */
2327                 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2328         } else {
2329                 return intel_pt_sync_ip(decoder);
2330         }
2331
2332         return 0;
2333 }
2334
2335 static uint64_t intel_pt_est_timestamp(struct intel_pt_decoder *decoder)
2336 {
2337         uint64_t est = decoder->sample_insn_cnt << 1;
2338
2339         if (!decoder->cbr || !decoder->max_non_turbo_ratio)
2340                 goto out;
2341
2342         est *= decoder->max_non_turbo_ratio;
2343         est /= decoder->cbr;
2344 out:
2345         return decoder->sample_timestamp + est;
2346 }
2347
2348 const struct intel_pt_state *intel_pt_decode(struct intel_pt_decoder *decoder)
2349 {
2350         int err;
2351
2352         do {
2353                 decoder->state.type = INTEL_PT_BRANCH;
2354                 decoder->state.flags = 0;
2355
2356                 switch (decoder->pkt_state) {
2357                 case INTEL_PT_STATE_NO_PSB:
2358                         err = intel_pt_sync(decoder);
2359                         break;
2360                 case INTEL_PT_STATE_NO_IP:
2361                         decoder->have_last_ip = false;
2362                         decoder->last_ip = 0;
2363                         decoder->ip = 0;
2364                         __fallthrough;
2365                 case INTEL_PT_STATE_ERR_RESYNC:
2366                         err = intel_pt_sync_ip(decoder);
2367                         break;
2368                 case INTEL_PT_STATE_IN_SYNC:
2369                         err = intel_pt_walk_trace(decoder);
2370                         break;
2371                 case INTEL_PT_STATE_TNT:
2372                         err = intel_pt_walk_tnt(decoder);
2373                         if (err == -EAGAIN)
2374                                 err = intel_pt_walk_trace(decoder);
2375                         break;
2376                 case INTEL_PT_STATE_TIP:
2377                 case INTEL_PT_STATE_TIP_PGD:
2378                         err = intel_pt_walk_tip(decoder);
2379                         break;
2380                 case INTEL_PT_STATE_FUP:
2381                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2382                         err = intel_pt_walk_fup(decoder);
2383                         if (err == -EAGAIN)
2384                                 err = intel_pt_walk_fup_tip(decoder);
2385                         else if (!err)
2386                                 decoder->pkt_state = INTEL_PT_STATE_FUP;
2387                         break;
2388                 case INTEL_PT_STATE_FUP_NO_TIP:
2389                         decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2390                         err = intel_pt_walk_fup(decoder);
2391                         if (err == -EAGAIN)
2392                                 err = intel_pt_walk_trace(decoder);
2393                         break;
2394                 default:
2395                         err = intel_pt_bug(decoder);
2396                         break;
2397                 }
2398         } while (err == -ENOLINK);
2399
2400         if (err) {
2401                 decoder->state.err = intel_pt_ext_err(err);
2402                 decoder->state.from_ip = decoder->ip;
2403                 decoder->sample_timestamp = decoder->timestamp;
2404                 decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
2405         } else {
2406                 decoder->state.err = 0;
2407                 if (decoder->cbr != decoder->cbr_seen && decoder->state.type) {
2408                         decoder->cbr_seen = decoder->cbr;
2409                         decoder->state.type |= INTEL_PT_CBR_CHG;
2410                         decoder->state.cbr_payload = decoder->cbr_payload;
2411                 }
2412                 if (intel_pt_sample_time(decoder->pkt_state)) {
2413                         decoder->sample_timestamp = decoder->timestamp;
2414                         decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
2415                 }
2416         }
2417
2418         decoder->state.timestamp = decoder->sample_timestamp;
2419         decoder->state.est_timestamp = intel_pt_est_timestamp(decoder);
2420         decoder->state.cr3 = decoder->cr3;
2421         decoder->state.tot_insn_cnt = decoder->tot_insn_cnt;
2422
2423         return &decoder->state;
2424 }
2425
2426 /**
2427  * intel_pt_next_psb - move buffer pointer to the start of the next PSB packet.
2428  * @buf: pointer to buffer pointer
2429  * @len: size of buffer
2430  *
2431  * Updates the buffer pointer to point to the start of the next PSB packet if
2432  * there is one, otherwise the buffer pointer is unchanged.  If @buf is updated,
2433  * @len is adjusted accordingly.
2434  *
2435  * Return: %true if a PSB packet is found, %false otherwise.
2436  */
2437 static bool intel_pt_next_psb(unsigned char **buf, size_t *len)
2438 {
2439         unsigned char *next;
2440
2441         next = memmem(*buf, *len, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2442         if (next) {
2443                 *len -= next - *buf;
2444                 *buf = next;
2445                 return true;
2446         }
2447         return false;
2448 }
2449
2450 /**
2451  * intel_pt_step_psb - move buffer pointer to the start of the following PSB
2452  *                     packet.
2453  * @buf: pointer to buffer pointer
2454  * @len: size of buffer
2455  *
2456  * Updates the buffer pointer to point to the start of the following PSB packet
2457  * (skipping the PSB at @buf itself) if there is one, otherwise the buffer
2458  * pointer is unchanged.  If @buf is updated, @len is adjusted accordingly.
2459  *
2460  * Return: %true if a PSB packet is found, %false otherwise.
2461  */
2462 static bool intel_pt_step_psb(unsigned char **buf, size_t *len)
2463 {
2464         unsigned char *next;
2465
2466         if (!*len)
2467                 return false;
2468
2469         next = memmem(*buf + 1, *len - 1, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2470         if (next) {
2471                 *len -= next - *buf;
2472                 *buf = next;
2473                 return true;
2474         }
2475         return false;
2476 }
2477
2478 /**
2479  * intel_pt_last_psb - find the last PSB packet in a buffer.
2480  * @buf: buffer
2481  * @len: size of buffer
2482  *
2483  * This function finds the last PSB in a buffer.
2484  *
2485  * Return: A pointer to the last PSB in @buf if found, %NULL otherwise.
2486  */
2487 static unsigned char *intel_pt_last_psb(unsigned char *buf, size_t len)
2488 {
2489         const char *n = INTEL_PT_PSB_STR;
2490         unsigned char *p;
2491         size_t k;
2492
2493         if (len < INTEL_PT_PSB_LEN)
2494                 return NULL;
2495
2496         k = len - INTEL_PT_PSB_LEN + 1;
2497         while (1) {
2498                 p = memrchr(buf, n[0], k);
2499                 if (!p)
2500                         return NULL;
2501                 if (!memcmp(p + 1, n + 1, INTEL_PT_PSB_LEN - 1))
2502                         return p;
2503                 k = p - buf;
2504                 if (!k)
2505                         return NULL;
2506         }
2507 }
2508
2509 /**
2510  * intel_pt_next_tsc - find and return next TSC.
2511  * @buf: buffer
2512  * @len: size of buffer
2513  * @tsc: TSC value returned
2514  * @rem: returns remaining size when TSC is found
2515  *
2516  * Find a TSC packet in @buf and return the TSC value.  This function assumes
2517  * that @buf starts at a PSB and that PSB+ will contain TSC and so stops if a
2518  * PSBEND packet is found.
2519  *
2520  * Return: %true if TSC is found, false otherwise.
2521  */
2522 static bool intel_pt_next_tsc(unsigned char *buf, size_t len, uint64_t *tsc,
2523                               size_t *rem)
2524 {
2525         struct intel_pt_pkt packet;
2526         int ret;
2527
2528         while (len) {
2529                 ret = intel_pt_get_packet(buf, len, &packet);
2530                 if (ret <= 0)
2531                         return false;
2532                 if (packet.type == INTEL_PT_TSC) {
2533                         *tsc = packet.payload;
2534                         *rem = len;
2535                         return true;
2536                 }
2537                 if (packet.type == INTEL_PT_PSBEND)
2538                         return false;
2539                 buf += ret;
2540                 len -= ret;
2541         }
2542         return false;
2543 }
2544
2545 /**
2546  * intel_pt_tsc_cmp - compare 7-byte TSCs.
2547  * @tsc1: first TSC to compare
2548  * @tsc2: second TSC to compare
2549  *
2550  * This function compares 7-byte TSC values allowing for the possibility that
2551  * TSC wrapped around.  Generally it is not possible to know if TSC has wrapped
2552  * around so for that purpose this function assumes the absolute difference is
2553  * less than half the maximum difference.
2554  *
2555  * Return: %-1 if @tsc1 is before @tsc2, %0 if @tsc1 == @tsc2, %1 if @tsc1 is
2556  * after @tsc2.
2557  */
2558 static int intel_pt_tsc_cmp(uint64_t tsc1, uint64_t tsc2)
2559 {
2560         const uint64_t halfway = (1ULL << 55);
2561
2562         if (tsc1 == tsc2)
2563                 return 0;
2564
2565         if (tsc1 < tsc2) {
2566                 if (tsc2 - tsc1 < halfway)
2567                         return -1;
2568                 else
2569                         return 1;
2570         } else {
2571                 if (tsc1 - tsc2 < halfway)
2572                         return 1;
2573                 else
2574                         return -1;
2575         }
2576 }
2577
2578 #define MAX_PADDING (PERF_AUXTRACE_RECORD_ALIGNMENT - 1)
2579
2580 /**
2581  * adj_for_padding - adjust overlap to account for padding.
2582  * @buf_b: second buffer
2583  * @buf_a: first buffer
2584  * @len_a: size of first buffer
2585  *
2586  * @buf_a might have up to 7 bytes of padding appended. Adjust the overlap
2587  * accordingly.
2588  *
2589  * Return: A pointer into @buf_b from where non-overlapped data starts
2590  */
2591 static unsigned char *adj_for_padding(unsigned char *buf_b,
2592                                       unsigned char *buf_a, size_t len_a)
2593 {
2594         unsigned char *p = buf_b - MAX_PADDING;
2595         unsigned char *q = buf_a + len_a - MAX_PADDING;
2596         int i;
2597
2598         for (i = MAX_PADDING; i; i--, p++, q++) {
2599                 if (*p != *q)
2600                         break;
2601         }
2602
2603         return p;
2604 }
2605
2606 /**
2607  * intel_pt_find_overlap_tsc - determine start of non-overlapped trace data
2608  *                             using TSC.
2609  * @buf_a: first buffer
2610  * @len_a: size of first buffer
2611  * @buf_b: second buffer
2612  * @len_b: size of second buffer
2613  * @consecutive: returns true if there is data in buf_b that is consecutive
2614  *               to buf_a
2615  *
2616  * If the trace contains TSC we can look at the last TSC of @buf_a and the
2617  * first TSC of @buf_b in order to determine if the buffers overlap, and then
2618  * walk forward in @buf_b until a later TSC is found.  A precondition is that
2619  * @buf_a and @buf_b are positioned at a PSB.
2620  *
2621  * Return: A pointer into @buf_b from where non-overlapped data starts, or
2622  * @buf_b + @len_b if there is no non-overlapped data.
2623  */
2624 static unsigned char *intel_pt_find_overlap_tsc(unsigned char *buf_a,
2625                                                 size_t len_a,
2626                                                 unsigned char *buf_b,
2627                                                 size_t len_b, bool *consecutive)
2628 {
2629         uint64_t tsc_a, tsc_b;
2630         unsigned char *p;
2631         size_t len, rem_a, rem_b;
2632
2633         p = intel_pt_last_psb(buf_a, len_a);
2634         if (!p)
2635                 return buf_b; /* No PSB in buf_a => no overlap */
2636
2637         len = len_a - (p - buf_a);
2638         if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a)) {
2639                 /* The last PSB+ in buf_a is incomplete, so go back one more */
2640                 len_a -= len;
2641                 p = intel_pt_last_psb(buf_a, len_a);
2642                 if (!p)
2643                         return buf_b; /* No full PSB+ => assume no overlap */
2644                 len = len_a - (p - buf_a);
2645                 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a))
2646                         return buf_b; /* No TSC in buf_a => assume no overlap */
2647         }
2648
2649         while (1) {
2650                 /* Ignore PSB+ with no TSC */
2651                 if (intel_pt_next_tsc(buf_b, len_b, &tsc_b, &rem_b)) {
2652                         int cmp = intel_pt_tsc_cmp(tsc_a, tsc_b);
2653
2654                         /* Same TSC, so buffers are consecutive */
2655                         if (!cmp && rem_b >= rem_a) {
2656                                 unsigned char *start;
2657
2658                                 *consecutive = true;
2659                                 start = buf_b + len_b - (rem_b - rem_a);
2660                                 return adj_for_padding(start, buf_a, len_a);
2661                         }
2662                         if (cmp < 0)
2663                                 return buf_b; /* tsc_a < tsc_b => no overlap */
2664                 }
2665
2666                 if (!intel_pt_step_psb(&buf_b, &len_b))
2667                         return buf_b + len_b; /* No PSB in buf_b => no data */
2668         }
2669 }
2670
2671 /**
2672  * intel_pt_find_overlap - determine start of non-overlapped trace data.
2673  * @buf_a: first buffer
2674  * @len_a: size of first buffer
2675  * @buf_b: second buffer
2676  * @len_b: size of second buffer
2677  * @have_tsc: can use TSC packets to detect overlap
2678  * @consecutive: returns true if there is data in buf_b that is consecutive
2679  *               to buf_a
2680  *
2681  * When trace samples or snapshots are recorded there is the possibility that
2682  * the data overlaps.  Note that, for the purposes of decoding, data is only
2683  * useful if it begins with a PSB packet.
2684  *
2685  * Return: A pointer into @buf_b from where non-overlapped data starts, or
2686  * @buf_b + @len_b if there is no non-overlapped data.
2687  */
2688 unsigned char *intel_pt_find_overlap(unsigned char *buf_a, size_t len_a,
2689                                      unsigned char *buf_b, size_t len_b,
2690                                      bool have_tsc, bool *consecutive)
2691 {
2692         unsigned char *found;
2693
2694         /* Buffer 'b' must start at PSB so throw away everything before that */
2695         if (!intel_pt_next_psb(&buf_b, &len_b))
2696                 return buf_b + len_b; /* No PSB */
2697
2698         if (!intel_pt_next_psb(&buf_a, &len_a))
2699                 return buf_b; /* No overlap */
2700
2701         if (have_tsc) {
2702                 found = intel_pt_find_overlap_tsc(buf_a, len_a, buf_b, len_b,
2703                                                   consecutive);
2704                 if (found)
2705                         return found;
2706         }
2707
2708         /*
2709          * Buffer 'b' cannot end within buffer 'a' so, for comparison purposes,
2710          * we can ignore the first part of buffer 'a'.
2711          */
2712         while (len_b < len_a) {
2713                 if (!intel_pt_step_psb(&buf_a, &len_a))
2714                         return buf_b; /* No overlap */
2715         }
2716
2717         /* Now len_b >= len_a */
2718         while (1) {
2719                 /* Potential overlap so check the bytes */
2720                 found = memmem(buf_a, len_a, buf_b, len_a);
2721                 if (found) {
2722                         *consecutive = true;
2723                         return adj_for_padding(buf_b + len_a, buf_a, len_a);
2724                 }
2725
2726                 /* Try again at next PSB in buffer 'a' */
2727                 if (!intel_pt_step_psb(&buf_a, &len_a))
2728                         return buf_b; /* No overlap */
2729         }
2730 }