tools/perf/util/arm-spe.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Arm Statistical Profiling Extensions (SPE) support
   4  * Copyright (c) 2017-2018, Arm Ltd.
   5  */
   6
   7 #include <byteswap.h>
   8 #include <endian.h>
   9 #include <errno.h>
  10 #include <inttypes.h>
  11 #include <linux/bitops.h>
  12 #include <linux/kernel.h>
  13 #include <linux/log2.h>
  14 #include <linux/types.h>
  15 #include <linux/zalloc.h>
  16 #include <stdlib.h>
  17 #include <unistd.h>
  18
  19 #include "auxtrace.h"
  20 #include "color.h"
  21 #include "debug.h"
  22 #include "evlist.h"
  23 #include "evsel.h"
  24 #include "machine.h"
  25 #include "session.h"
  26 #include "symbol.h"
  27 #include "thread.h"
  28 #include "thread-stack.h"
  29 #include "tool.h"
  30 #include "util/synthetic-events.h"
  31
  32 #include "arm-spe.h"
  33 #include "arm-spe-decoder/arm-spe-decoder.h"
  34 #include "arm-spe-decoder/arm-spe-pkt-decoder.h"
  35
  36 #define MAX_TIMESTAMP (~0ULL)
  37
  38 struct arm_spe {
  39         struct auxtrace                 auxtrace;
  40         struct auxtrace_queues          queues;
  41         struct auxtrace_heap            heap;
  42         struct itrace_synth_opts        synth_opts;
  43         u32                             auxtrace_type;
  44         struct perf_session             *session;
  45         struct machine                  *machine;
  46         u32                             pmu_type;
  47
  48         u8                              timeless_decoding;
  49         u8                              data_queued;
  50
  51         u8                              sample_flc;
  52         u8                              sample_llc;
  53         u8                              sample_tlb;
  54         u8                              sample_branch;
  55         u8                              sample_remote_access;
  56
  57         u64                             l1d_miss_id;
  58         u64                             l1d_access_id;
  59         u64                             llc_miss_id;
  60         u64                             llc_access_id;
  61         u64                             tlb_miss_id;
  62         u64                             tlb_access_id;
  63         u64                             branch_miss_id;
  64         u64                             remote_access_id;
  65
  66         u64                             kernel_start;
  67
  68         unsigned long                   num_events;
  69 };
  70
  71 struct arm_spe_queue {
  72         struct arm_spe                  *spe;
  73         unsigned int                    queue_nr;
  74         struct auxtrace_buffer          *buffer;
  75         struct auxtrace_buffer          *old_buffer;
  76         union perf_event                *event_buf;
  77         bool                            on_heap;
  78         bool                            done;
  79         pid_t                           pid;
  80         pid_t                           tid;
  81         int                             cpu;
  82         struct arm_spe_decoder          *decoder;
  83         u64                             time;
  84         u64                             timestamp;
  85         struct thread                   *thread;
  86 };
  87
  88 static void arm_spe_dump(struct arm_spe *spe __maybe_unused,
  89                          unsigned char *buf, size_t len)
  90 {
  91         struct arm_spe_pkt packet;
  92         size_t pos = 0;
  93         int ret, pkt_len, i;
  94         char desc[ARM_SPE_PKT_DESC_MAX];
  95         const char *color = PERF_COLOR_BLUE;
  96
  97         color_fprintf(stdout, color,
  98                       ". ... ARM SPE data: size %zu bytes\n",
  99                       len);
 100
 101         while (len) {
 102                 ret = arm_spe_get_packet(buf, len, &packet);
 103                 if (ret > 0)
 104                         pkt_len = ret;
 105                 else
 106                         pkt_len = 1;
 107                 printf(".");
 108                 color_fprintf(stdout, color, "  %08x: ", pos);
 109                 for (i = 0; i < pkt_len; i++)
 110                         color_fprintf(stdout, color, " %02x", buf[i]);
 111                 for (; i < 16; i++)
 112                         color_fprintf(stdout, color, "   ");
 113                 if (ret > 0) {
 114                         ret = arm_spe_pkt_desc(&packet, desc,
 115                                                ARM_SPE_PKT_DESC_MAX);
 116                         if (ret > 0)
 117                                 color_fprintf(stdout, color, " %s\n", desc);
 118                 } else {
 119                         color_fprintf(stdout, color, " Bad packet!\n");
 120                 }
 121                 pos += pkt_len;
 122                 buf += pkt_len;
 123                 len -= pkt_len;
 124         }
 125 }
 126
 127 static void arm_spe_dump_event(struct arm_spe *spe, unsigned char *buf,
 128                                size_t len)
 129 {
 130         printf(".\n");
 131         arm_spe_dump(spe, buf, len);
 132 }
 133
 134 static int arm_spe_get_trace(struct arm_spe_buffer *b, void *data)
 135 {
 136         struct arm_spe_queue *speq = data;
 137         struct auxtrace_buffer *buffer = speq->buffer;
 138         struct auxtrace_buffer *old_buffer = speq->old_buffer;
 139         struct auxtrace_queue *queue;
 140
 141         queue = &speq->spe->queues.queue_array[speq->queue_nr];
 142
 143         buffer = auxtrace_buffer__next(queue, buffer);
 144         /* If no more data, drop the previous auxtrace_buffer and return */
 145         if (!buffer) {
 146                 if (old_buffer)
 147                         auxtrace_buffer__drop_data(old_buffer);
 148                 b->len = 0;
 149                 return 0;
 150         }
 151
 152         speq->buffer = buffer;
 153
 154         /* If the aux_buffer doesn't have data associated, try to load it */
 155         if (!buffer->data) {
 156                 /* get the file desc associated with the perf data file */
 157                 int fd = perf_data__fd(speq->spe->session->data);
 158
 159                 buffer->data = auxtrace_buffer__get_data(buffer, fd);
 160                 if (!buffer->data)
 161                         return -ENOMEM;
 162         }
 163
 164         b->len = buffer->size;
 165         b->buf = buffer->data;
 166
 167         if (b->len) {
 168                 if (old_buffer)
 169                         auxtrace_buffer__drop_data(old_buffer);
 170                 speq->old_buffer = buffer;
 171         } else {
 172                 auxtrace_buffer__drop_data(buffer);
 173                 return arm_spe_get_trace(b, data);
 174         }
 175
 176         return 0;
 177 }
 178
 179 static struct arm_spe_queue *arm_spe__alloc_queue(struct arm_spe *spe,
 180                 unsigned int queue_nr)
 181 {
 182         struct arm_spe_params params = { .get_trace = 0, };
 183         struct arm_spe_queue *speq;
 184
 185         speq = zalloc(sizeof(*speq));
 186         if (!speq)
 187                 return NULL;
 188
 189         speq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
 190         if (!speq->event_buf)
 191                 goto out_free;
 192
 193         speq->spe = spe;
 194         speq->queue_nr = queue_nr;
 195         speq->pid = -1;
 196         speq->tid = -1;
 197         speq->cpu = -1;
 198
 199         /* params set */
 200         params.get_trace = arm_spe_get_trace;
 201         params.data = speq;
 202
 203         /* create new decoder */
 204         speq->decoder = arm_spe_decoder_new(&params);
 205         if (!speq->decoder)
 206                 goto out_free;
 207
 208         return speq;
 209
 210 out_free:
 211         zfree(&speq->event_buf);
 212         free(speq);
 213
 214         return NULL;
 215 }
 216
 217 static inline u8 arm_spe_cpumode(struct arm_spe *spe, u64 ip)
 218 {
 219         return ip >= spe->kernel_start ?
 220                 PERF_RECORD_MISC_KERNEL :
 221                 PERF_RECORD_MISC_USER;
 222 }
 223
 224 static void arm_spe_prep_sample(struct arm_spe *spe,
 225                                 struct arm_spe_queue *speq,
 226                                 union perf_event *event,
 227                                 struct perf_sample *sample)
 228 {
 229         struct arm_spe_record *record = &speq->decoder->record;
 230
 231         if (!spe->timeless_decoding)
 232                 sample->time = speq->timestamp;
 233
 234         sample->ip = record->from_ip;
 235         sample->cpumode = arm_spe_cpumode(spe, sample->ip);
 236         sample->pid = speq->pid;
 237         sample->tid = speq->tid;
 238         sample->addr = record->to_ip;
 239         sample->period = 1;
 240         sample->cpu = speq->cpu;
 241
 242         event->sample.header.type = PERF_RECORD_SAMPLE;
 243         event->sample.header.misc = sample->cpumode;
 244         event->sample.header.size = sizeof(struct perf_event_header);
 245 }
 246
 247 static inline int
 248 arm_spe_deliver_synth_event(struct arm_spe *spe,
 249                             struct arm_spe_queue *speq __maybe_unused,
 250                             union perf_event *event,
 251                             struct perf_sample *sample)
 252 {
 253         int ret;
 254
 255         ret = perf_session__deliver_synth_event(spe->session, event, sample);
 256         if (ret)
 257                 pr_err("ARM SPE: failed to deliver event, error %d\n", ret);
 258
 259         return ret;
 260 }
 261
 262 static int
 263 arm_spe_synth_spe_events_sample(struct arm_spe_queue *speq,
 264                                 u64 spe_events_id)
 265 {
 266         struct arm_spe *spe = speq->spe;
 267         union perf_event *event = speq->event_buf;
 268         struct perf_sample sample = { .ip = 0, };
 269
 270         arm_spe_prep_sample(spe, speq, event, &sample);
 271
 272         sample.id = spe_events_id;
 273         sample.stream_id = spe_events_id;
 274
 275         return arm_spe_deliver_synth_event(spe, speq, event, &sample);
 276 }
 277
 278 static int arm_spe_sample(struct arm_spe_queue *speq)
 279 {
 280         const struct arm_spe_record *record = &speq->decoder->record;
 281         struct arm_spe *spe = speq->spe;
 282         int err;
 283
 284         if (spe->sample_flc) {
 285                 if (record->type & ARM_SPE_L1D_MISS) {
 286                         err = arm_spe_synth_spe_events_sample(
 287                                         speq, spe->l1d_miss_id);
 288                         if (err)
 289                                 return err;
 290                 }
 291
 292                 if (record->type & ARM_SPE_L1D_ACCESS) {
 293                         err = arm_spe_synth_spe_events_sample(
 294                                         speq, spe->l1d_access_id);
 295                         if (err)
 296                                 return err;
 297                 }
 298         }
 299
 300         if (spe->sample_llc) {
 301                 if (record->type & ARM_SPE_LLC_MISS) {
 302                         err = arm_spe_synth_spe_events_sample(
 303                                         speq, spe->llc_miss_id);
 304                         if (err)
 305                                 return err;
 306                 }
 307
 308                 if (record->type & ARM_SPE_LLC_ACCESS) {
 309                         err = arm_spe_synth_spe_events_sample(
 310                                         speq, spe->llc_access_id);
 311                         if (err)
 312                                 return err;
 313                 }
 314         }
 315
 316         if (spe->sample_tlb) {
 317                 if (record->type & ARM_SPE_TLB_MISS) {
 318                         err = arm_spe_synth_spe_events_sample(
 319                                         speq, spe->tlb_miss_id);
 320                         if (err)
 321                                 return err;
 322                 }
 323
 324                 if (record->type & ARM_SPE_TLB_ACCESS) {
 325                         err = arm_spe_synth_spe_events_sample(
 326                                         speq, spe->tlb_access_id);
 327                         if (err)
 328                                 return err;
 329                 }
 330         }
 331
 332         if (spe->sample_branch && (record->type & ARM_SPE_BRANCH_MISS)) {
 333                 err = arm_spe_synth_spe_events_sample(speq,
 334                                                       spe->branch_miss_id);
 335                 if (err)
 336                         return err;
 337         }
 338
 339         if (spe->sample_remote_access &&
 340             (record->type & ARM_SPE_REMOTE_ACCESS)) {
 341                 err = arm_spe_synth_spe_events_sample(speq,
 342                                                       spe->remote_access_id);
 343                 if (err)
 344                         return err;
 345         }
 346
 347         return 0;
 348 }
 349
 350 static int arm_spe_run_decoder(struct arm_spe_queue *speq, u64 *timestamp)
 351 {
 352         struct arm_spe *spe = speq->spe;
 353         int ret;
 354
 355         if (!spe->kernel_start)
 356                 spe->kernel_start = machine__kernel_start(spe->machine);
 357
 358         while (1) {
 359                 ret = arm_spe_decode(speq->decoder);
 360                 if (!ret) {
 361                         pr_debug("No data or all data has been processed.\n");
 362                         return 1;
 363                 }
 364
 365                 /*
 366                  * Error is detected when decode SPE trace data, continue to
 367                  * the next trace data and find out more records.
 368                  */
 369                 if (ret < 0)
 370                         continue;
 371
 372                 ret = arm_spe_sample(speq);
 373                 if (ret)
 374                         return ret;
 375
 376                 if (!spe->timeless_decoding && speq->timestamp >= *timestamp) {
 377                         *timestamp = speq->timestamp;
 378                         return 0;
 379                 }
 380         }
 381
 382         return 0;
 383 }
 384
 385 static int arm_spe__setup_queue(struct arm_spe *spe,
 386                                struct auxtrace_queue *queue,
 387                                unsigned int queue_nr)
 388 {
 389         struct arm_spe_queue *speq = queue->priv;
 390         struct arm_spe_record *record;
 391
 392         if (list_empty(&queue->head) || speq)
 393                 return 0;
 394
 395         speq = arm_spe__alloc_queue(spe, queue_nr);
 396
 397         if (!speq)
 398                 return -ENOMEM;
 399
 400         queue->priv = speq;
 401
 402         if (queue->cpu != -1)
 403                 speq->cpu = queue->cpu;
 404
 405         if (!speq->on_heap) {
 406                 int ret;
 407
 408                 if (spe->timeless_decoding)
 409                         return 0;
 410
 411 retry:
 412                 ret = arm_spe_decode(speq->decoder);
 413
 414                 if (!ret)
 415                         return 0;
 416
 417                 if (ret < 0)
 418                         goto retry;
 419
 420                 record = &speq->decoder->record;
 421
 422                 speq->timestamp = record->timestamp;
 423                 ret = auxtrace_heap__add(&spe->heap, queue_nr, speq->timestamp);
 424                 if (ret)
 425                         return ret;
 426                 speq->on_heap = true;
 427         }
 428
 429         return 0;
 430 }
 431
 432 static int arm_spe__setup_queues(struct arm_spe *spe)
 433 {
 434         unsigned int i;
 435         int ret;
 436
 437         for (i = 0; i < spe->queues.nr_queues; i++) {
 438                 ret = arm_spe__setup_queue(spe, &spe->queues.queue_array[i], i);
 439                 if (ret)
 440                         return ret;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static int arm_spe__update_queues(struct arm_spe *spe)
 447 {
 448         if (spe->queues.new_data) {
 449                 spe->queues.new_data = false;
 450                 return arm_spe__setup_queues(spe);
 451         }
 452
 453         return 0;
 454 }
 455
 456 static bool arm_spe__is_timeless_decoding(struct arm_spe *spe)
 457 {
 458         struct evsel *evsel;
 459         struct evlist *evlist = spe->session->evlist;
 460         bool timeless_decoding = true;
 461
 462         /*
 463          * Circle through the list of event and complain if we find one
 464          * with the time bit set.
 465          */
 466         evlist__for_each_entry(evlist, evsel) {
 467                 if ((evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
 468                         timeless_decoding = false;
 469         }
 470
 471         return timeless_decoding;
 472 }
 473
 474 static void arm_spe_set_pid_tid_cpu(struct arm_spe *spe,
 475                                     struct auxtrace_queue *queue)
 476 {
 477         struct arm_spe_queue *speq = queue->priv;
 478         pid_t tid;
 479
 480         tid = machine__get_current_tid(spe->machine, speq->cpu);
 481         if (tid != -1) {
 482                 speq->tid = tid;
 483                 thread__zput(speq->thread);
 484         } else
 485                 speq->tid = queue->tid;
 486
 487         if ((!speq->thread) && (speq->tid != -1)) {
 488                 speq->thread = machine__find_thread(spe->machine, -1,
 489                                                     speq->tid);
 490         }
 491
 492         if (speq->thread) {
 493                 speq->pid = speq->thread->pid_;
 494                 if (queue->cpu == -1)
 495                         speq->cpu = speq->thread->cpu;
 496         }
 497 }
 498
 499 static int arm_spe_process_queues(struct arm_spe *spe, u64 timestamp)
 500 {
 501         unsigned int queue_nr;
 502         u64 ts;
 503         int ret;
 504
 505         while (1) {
 506                 struct auxtrace_queue *queue;
 507                 struct arm_spe_queue *speq;
 508
 509                 if (!spe->heap.heap_cnt)
 510                         return 0;
 511
 512                 if (spe->heap.heap_array[0].ordinal >= timestamp)
 513                         return 0;
 514
 515                 queue_nr = spe->heap.heap_array[0].queue_nr;
 516                 queue = &spe->queues.queue_array[queue_nr];
 517                 speq = queue->priv;
 518
 519                 auxtrace_heap__pop(&spe->heap);
 520
 521                 if (spe->heap.heap_cnt) {
 522                         ts = spe->heap.heap_array[0].ordinal + 1;
 523                         if (ts > timestamp)
 524                                 ts = timestamp;
 525                 } else {
 526                         ts = timestamp;
 527                 }
 528
 529                 arm_spe_set_pid_tid_cpu(spe, queue);
 530
 531                 ret = arm_spe_run_decoder(speq, &ts);
 532                 if (ret < 0) {
 533                         auxtrace_heap__add(&spe->heap, queue_nr, ts);
 534                         return ret;
 535                 }
 536
 537                 if (!ret) {
 538                         ret = auxtrace_heap__add(&spe->heap, queue_nr, ts);
 539                         if (ret < 0)
 540                                 return ret;
 541                 } else {
 542                         speq->on_heap = false;
 543                 }
 544         }
 545
 546         return 0;
 547 }
 548
 549 static int arm_spe_process_timeless_queues(struct arm_spe *spe, pid_t tid,
 550                                             u64 time_)
 551 {
 552         struct auxtrace_queues *queues = &spe->queues;
 553         unsigned int i;
 554         u64 ts = 0;
 555
 556         for (i = 0; i < queues->nr_queues; i++) {
 557                 struct auxtrace_queue *queue = &spe->queues.queue_array[i];
 558                 struct arm_spe_queue *speq = queue->priv;
 559
 560                 if (speq && (tid == -1 || speq->tid == tid)) {
 561                         speq->time = time_;
 562                         arm_spe_set_pid_tid_cpu(spe, queue);
 563                         arm_spe_run_decoder(speq, &ts);
 564                 }
 565         }
 566         return 0;
 567 }
 568
 569 static int arm_spe_process_event(struct perf_session *session,
 570                                  union perf_event *event,
 571                                  struct perf_sample *sample,
 572                                  struct perf_tool *tool)
 573 {
 574         int err = 0;
 575         u64 timestamp;
 576         struct arm_spe *spe = container_of(session->auxtrace,
 577                         struct arm_spe, auxtrace);
 578
 579         if (dump_trace)
 580                 return 0;
 581
 582         if (!tool->ordered_events) {
 583                 pr_err("SPE trace requires ordered events\n");
 584                 return -EINVAL;
 585         }
 586
 587         if (sample->time && (sample->time != (u64) -1))
 588                 timestamp = sample->time;
 589         else
 590                 timestamp = 0;
 591
 592         if (timestamp || spe->timeless_decoding) {
 593                 err = arm_spe__update_queues(spe);
 594                 if (err)
 595                         return err;
 596         }
 597
 598         if (spe->timeless_decoding) {
 599                 if (event->header.type == PERF_RECORD_EXIT) {
 600                         err = arm_spe_process_timeless_queues(spe,
 601                                         event->fork.tid,
 602                                         sample->time);
 603                 }
 604         } else if (timestamp) {
 605                 if (event->header.type == PERF_RECORD_EXIT) {
 606                         err = arm_spe_process_queues(spe, timestamp);
 607                         if (err)
 608                                 return err;
 609                 }
 610         }
 611
 612         return err;
 613 }
 614
 615 static int arm_spe_process_auxtrace_event(struct perf_session *session,
 616                                           union perf_event *event,
 617                                           struct perf_tool *tool __maybe_unused)
 618 {
 619         struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
 620                                              auxtrace);
 621
 622         if (!spe->data_queued) {
 623                 struct auxtrace_buffer *buffer;
 624                 off_t data_offset;
 625                 int fd = perf_data__fd(session->data);
 626                 int err;
 627
 628                 if (perf_data__is_pipe(session->data)) {
 629                         data_offset = 0;
 630                 } else {
 631                         data_offset = lseek(fd, 0, SEEK_CUR);
 632                         if (data_offset == -1)
 633                                 return -errno;
 634                 }
 635
 636                 err = auxtrace_queues__add_event(&spe->queues, session, event,
 637                                 data_offset, &buffer);
 638                 if (err)
 639                         return err;
 640
 641                 /* Dump here now we have copied a piped trace out of the pipe */
 642                 if (dump_trace) {
 643                         if (auxtrace_buffer__get_data(buffer, fd)) {
 644                                 arm_spe_dump_event(spe, buffer->data,
 645                                                 buffer->size);
 646                                 auxtrace_buffer__put_data(buffer);
 647                         }
 648                 }
 649         }
 650
 651         return 0;
 652 }
 653
 654 static int arm_spe_flush(struct perf_session *session __maybe_unused,
 655                          struct perf_tool *tool __maybe_unused)
 656 {
 657         struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
 658                         auxtrace);
 659         int ret;
 660
 661         if (dump_trace)
 662                 return 0;
 663
 664         if (!tool->ordered_events)
 665                 return -EINVAL;
 666
 667         ret = arm_spe__update_queues(spe);
 668         if (ret < 0)
 669                 return ret;
 670
 671         if (spe->timeless_decoding)
 672                 return arm_spe_process_timeless_queues(spe, -1,
 673                                 MAX_TIMESTAMP - 1);
 674
 675         return arm_spe_process_queues(spe, MAX_TIMESTAMP);
 676 }
 677
 678 static void arm_spe_free_queue(void *priv)
 679 {
 680         struct arm_spe_queue *speq = priv;
 681
 682         if (!speq)
 683                 return;
 684         thread__zput(speq->thread);
 685         arm_spe_decoder_free(speq->decoder);
 686         zfree(&speq->event_buf);
 687         free(speq);
 688 }
 689
 690 static void arm_spe_free_events(struct perf_session *session)
 691 {
 692         struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
 693                                              auxtrace);
 694         struct auxtrace_queues *queues = &spe->queues;
 695         unsigned int i;
 696
 697         for (i = 0; i < queues->nr_queues; i++) {
 698                 arm_spe_free_queue(queues->queue_array[i].priv);
 699                 queues->queue_array[i].priv = NULL;
 700         }
 701         auxtrace_queues__free(queues);
 702 }
 703
 704 static void arm_spe_free(struct perf_session *session)
 705 {
 706         struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
 707                                              auxtrace);
 708
 709         auxtrace_heap__free(&spe->heap);
 710         arm_spe_free_events(session);
 711         session->auxtrace = NULL;
 712         free(spe);
 713 }
 714
 715 static bool arm_spe_evsel_is_auxtrace(struct perf_session *session,
 716                                       struct evsel *evsel)
 717 {
 718         struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe, auxtrace);
 719
 720         return evsel->core.attr.type == spe->pmu_type;
 721 }
 722
 723 static const char * const arm_spe_info_fmts[] = {
 724         [ARM_SPE_PMU_TYPE]              = "  PMU Type           %"PRId64"\n",
 725 };
 726
 727 static void arm_spe_print_info(__u64 *arr)
 728 {
 729         if (!dump_trace)
 730                 return;
 731
 732         fprintf(stdout, arm_spe_info_fmts[ARM_SPE_PMU_TYPE], arr[ARM_SPE_PMU_TYPE]);
 733 }
 734
 735 struct arm_spe_synth {
 736         struct perf_tool dummy_tool;
 737         struct perf_session *session;
 738 };
 739
 740 static int arm_spe_event_synth(struct perf_tool *tool,
 741                                union perf_event *event,
 742                                struct perf_sample *sample __maybe_unused,
 743                                struct machine *machine __maybe_unused)
 744 {
 745         struct arm_spe_synth *arm_spe_synth =
 746                       container_of(tool, struct arm_spe_synth, dummy_tool);
 747
 748         return perf_session__deliver_synth_event(arm_spe_synth->session,
 749                                                  event, NULL);
 750 }
 751
 752 static int arm_spe_synth_event(struct perf_session *session,
 753                                struct perf_event_attr *attr, u64 id)
 754 {
 755         struct arm_spe_synth arm_spe_synth;
 756
 757         memset(&arm_spe_synth, 0, sizeof(struct arm_spe_synth));
 758         arm_spe_synth.session = session;
 759
 760         return perf_event__synthesize_attr(&arm_spe_synth.dummy_tool, attr, 1,
 761                                            &id, arm_spe_event_synth);
 762 }
 763
 764 static void arm_spe_set_event_name(struct evlist *evlist, u64 id,
 765                                     const char *name)
 766 {
 767         struct evsel *evsel;
 768
 769         evlist__for_each_entry(evlist, evsel) {
 770                 if (evsel->core.id && evsel->core.id[0] == id) {
 771                         if (evsel->name)
 772                                 zfree(&evsel->name);
 773                         evsel->name = strdup(name);
 774                         break;
 775                 }
 776         }
 777 }
 778
 779 static int
 780 arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 781 {
 782         struct evlist *evlist = session->evlist;
 783         struct evsel *evsel;
 784         struct perf_event_attr attr;
 785         bool found = false;
 786         u64 id;
 787         int err;
 788
 789         evlist__for_each_entry(evlist, evsel) {
 790                 if (evsel->core.attr.type == spe->pmu_type) {
 791                         found = true;
 792                         break;
 793                 }
 794         }
 795
 796         if (!found) {
 797                 pr_debug("No selected events with SPE trace data\n");
 798                 return 0;
 799         }
 800
 801         memset(&attr, 0, sizeof(struct perf_event_attr));
 802         attr.size = sizeof(struct perf_event_attr);
 803         attr.type = PERF_TYPE_HARDWARE;
 804         attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
 805         attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
 806                 PERF_SAMPLE_PERIOD;
 807         if (spe->timeless_decoding)
 808                 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
 809         else
 810                 attr.sample_type |= PERF_SAMPLE_TIME;
 811
 812         attr.exclude_user = evsel->core.attr.exclude_user;
 813         attr.exclude_kernel = evsel->core.attr.exclude_kernel;
 814         attr.exclude_hv = evsel->core.attr.exclude_hv;
 815         attr.exclude_host = evsel->core.attr.exclude_host;
 816         attr.exclude_guest = evsel->core.attr.exclude_guest;
 817         attr.sample_id_all = evsel->core.attr.sample_id_all;
 818         attr.read_format = evsel->core.attr.read_format;
 819
 820         /* create new id val to be a fixed offset from evsel id */
 821         id = evsel->core.id[0] + 1000000000;
 822
 823         if (!id)
 824                 id = 1;
 825
 826         if (spe->synth_opts.flc) {
 827                 spe->sample_flc = true;
 828
 829                 /* Level 1 data cache miss */
 830                 err = arm_spe_synth_event(session, &attr, id);
 831                 if (err)
 832                         return err;
 833                 spe->l1d_miss_id = id;
 834                 arm_spe_set_event_name(evlist, id, "l1d-miss");
 835                 id += 1;
 836
 837                 /* Level 1 data cache access */
 838                 err = arm_spe_synth_event(session, &attr, id);
 839                 if (err)
 840                         return err;
 841                 spe->l1d_access_id = id;
 842                 arm_spe_set_event_name(evlist, id, "l1d-access");
 843                 id += 1;
 844         }
 845
 846         if (spe->synth_opts.llc) {
 847                 spe->sample_llc = true;
 848
 849                 /* Last level cache miss */
 850                 err = arm_spe_synth_event(session, &attr, id);
 851                 if (err)
 852                         return err;
 853                 spe->llc_miss_id = id;
 854                 arm_spe_set_event_name(evlist, id, "llc-miss");
 855                 id += 1;
 856
 857                 /* Last level cache access */
 858                 err = arm_spe_synth_event(session, &attr, id);
 859                 if (err)
 860                         return err;
 861                 spe->llc_access_id = id;
 862                 arm_spe_set_event_name(evlist, id, "llc-access");
 863                 id += 1;
 864         }
 865
 866         if (spe->synth_opts.tlb) {
 867                 spe->sample_tlb = true;
 868
 869                 /* TLB miss */
 870                 err = arm_spe_synth_event(session, &attr, id);
 871                 if (err)
 872                         return err;
 873                 spe->tlb_miss_id = id;
 874                 arm_spe_set_event_name(evlist, id, "tlb-miss");
 875                 id += 1;
 876
 877                 /* TLB access */
 878                 err = arm_spe_synth_event(session, &attr, id);
 879                 if (err)
 880                         return err;
 881                 spe->tlb_access_id = id;
 882                 arm_spe_set_event_name(evlist, id, "tlb-access");
 883                 id += 1;
 884         }
 885
 886         if (spe->synth_opts.branches) {
 887                 spe->sample_branch = true;
 888
 889                 /* Branch miss */
 890                 err = arm_spe_synth_event(session, &attr, id);
 891                 if (err)
 892                         return err;
 893                 spe->branch_miss_id = id;
 894                 arm_spe_set_event_name(evlist, id, "branch-miss");
 895                 id += 1;
 896         }
 897
 898         if (spe->synth_opts.remote_access) {
 899                 spe->sample_remote_access = true;
 900
 901                 /* Remote access */
 902                 err = arm_spe_synth_event(session, &attr, id);
 903                 if (err)
 904                         return err;
 905                 spe->remote_access_id = id;
 906                 arm_spe_set_event_name(evlist, id, "remote-access");
 907                 id += 1;
 908         }
 909
 910         return 0;
 911 }
 912
 913 int arm_spe_process_auxtrace_info(union perf_event *event,
 914                                   struct perf_session *session)
 915 {
 916         struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
 917         size_t min_sz = sizeof(u64) * ARM_SPE_AUXTRACE_PRIV_MAX;
 918         struct arm_spe *spe;
 919         int err;
 920
 921         if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
 922                                         min_sz)
 923                 return -EINVAL;
 924
 925         spe = zalloc(sizeof(struct arm_spe));
 926         if (!spe)
 927                 return -ENOMEM;
 928
 929         err = auxtrace_queues__init(&spe->queues);
 930         if (err)
 931                 goto err_free;
 932
 933         spe->session = session;
 934         spe->machine = &session->machines.host; /* No kvm support */
 935         spe->auxtrace_type = auxtrace_info->type;
 936         spe->pmu_type = auxtrace_info->priv[ARM_SPE_PMU_TYPE];
 937
 938         spe->timeless_decoding = arm_spe__is_timeless_decoding(spe);
 939         spe->auxtrace.process_event = arm_spe_process_event;
 940         spe->auxtrace.process_auxtrace_event = arm_spe_process_auxtrace_event;
 941         spe->auxtrace.flush_events = arm_spe_flush;
 942         spe->auxtrace.free_events = arm_spe_free_events;
 943         spe->auxtrace.free = arm_spe_free;
 944         spe->auxtrace.evsel_is_auxtrace = arm_spe_evsel_is_auxtrace;
 945         session->auxtrace = &spe->auxtrace;
 946
 947         arm_spe_print_info(&auxtrace_info->priv[0]);
 948
 949         if (dump_trace)
 950                 return 0;
 951
 952         if (session->itrace_synth_opts && session->itrace_synth_opts->set)
 953                 spe->synth_opts = *session->itrace_synth_opts;
 954         else
 955                 itrace_synth_opts__set_default(&spe->synth_opts, false);
 956
 957         err = arm_spe_synth_events(spe, session);
 958         if (err)
 959                 goto err_free_queues;
 960
 961         err = auxtrace_queues__process_index(&spe->queues, session);
 962         if (err)
 963                 goto err_free_queues;
 964
 965         if (spe->queues.populated)
 966                 spe->data_queued = true;
 967
 968         return 0;
 969
 970 err_free_queues:
 971         auxtrace_queues__free(&spe->queues);
 972         session->auxtrace = NULL;
 973 err_free:
 974         free(spe);
 975         return err;
 976 }