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

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