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

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