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

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