1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright(C) 2015-2018 Linaro Limited. 4 * 5 * Author: Tor Jeremiassen <tor@ti.com> 6 * Author: Mathieu Poirier <mathieu.poirier@linaro.org> 7 */ 8 9 #include <linux/bitops.h> 10 #include <linux/err.h> 11 #include <linux/kernel.h> 12 #include <linux/log2.h> 13 #include <linux/types.h> 14 #include <linux/zalloc.h> 15 16 #include <opencsd/ocsd_if_types.h> 17 #include <stdlib.h> 18 19 #include "auxtrace.h" 20 #include "color.h" 21 #include "cs-etm.h" 22 #include "cs-etm-decoder/cs-etm-decoder.h" 23 #include "debug.h" 24 #include "dso.h" 25 #include "evlist.h" 26 #include "intlist.h" 27 #include "machine.h" 28 #include "map.h" 29 #include "perf.h" 30 #include "session.h" 31 #include "map_symbol.h" 32 #include "branch.h" 33 #include "symbol.h" 34 #include "tool.h" 35 #include "thread.h" 36 #include "thread-stack.h" 37 #include <tools/libc_compat.h> 38 #include "util/synthetic-events.h" 39 40 #define MAX_TIMESTAMP (~0ULL) 41 42 struct cs_etm_auxtrace { 43 struct auxtrace auxtrace; 44 struct auxtrace_queues queues; 45 struct auxtrace_heap heap; 46 struct itrace_synth_opts synth_opts; 47 struct perf_session *session; 48 struct machine *machine; 49 struct thread *unknown_thread; 50 51 u8 timeless_decoding; 52 u8 snapshot_mode; 53 u8 data_queued; 54 u8 sample_branches; 55 u8 sample_instructions; 56 57 int num_cpu; 58 u32 auxtrace_type; 59 u64 branches_sample_type; 60 u64 branches_id; 61 u64 instructions_sample_type; 62 u64 instructions_sample_period; 63 u64 instructions_id; 64 u64 **metadata; 65 u64 kernel_start; 66 unsigned int pmu_type; 67 }; 68 69 struct cs_etm_traceid_queue { 70 u8 trace_chan_id; 71 pid_t pid, tid; 72 u64 period_instructions; 73 size_t last_branch_pos; 74 union perf_event *event_buf; 75 struct thread *thread; 76 struct branch_stack *last_branch; 77 struct branch_stack *last_branch_rb; 78 struct cs_etm_packet *prev_packet; 79 struct cs_etm_packet *packet; 80 struct cs_etm_packet_queue packet_queue; 81 }; 82 83 struct cs_etm_queue { 84 struct cs_etm_auxtrace *etm; 85 struct cs_etm_decoder *decoder; 86 struct auxtrace_buffer *buffer; 87 unsigned int queue_nr; 88 u8 pending_timestamp; 89 u64 offset; 90 const unsigned char *buf; 91 size_t buf_len, buf_used; 92 /* Conversion between traceID and index in traceid_queues array */ 93 struct intlist *traceid_queues_list; 94 struct cs_etm_traceid_queue **traceid_queues; 95 }; 96 97 static int cs_etm__update_queues(struct cs_etm_auxtrace *etm); 98 static int cs_etm__process_queues(struct cs_etm_auxtrace *etm); 99 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm, 100 pid_t tid); 101 static int cs_etm__get_data_block(struct cs_etm_queue *etmq); 102 static int cs_etm__decode_data_block(struct cs_etm_queue *etmq); 103 104 /* PTMs ETMIDR [11:8] set to b0011 */ 105 #define ETMIDR_PTM_VERSION 0x00000300 106 107 /* 108 * A struct auxtrace_heap_item only has a queue_nr and a timestamp to 109 * work with. One option is to modify to auxtrace_heap_XYZ() API or simply 110 * encode the etm queue number as the upper 16 bit and the channel as 111 * the lower 16 bit. 112 */ 113 #define TO_CS_QUEUE_NR(queue_nr, trace_chan_id) \ 114 (queue_nr << 16 | trace_chan_id) 115 #define TO_QUEUE_NR(cs_queue_nr) (cs_queue_nr >> 16) 116 #define TO_TRACE_CHAN_ID(cs_queue_nr) (cs_queue_nr & 0x0000ffff) 117 118 static u32 cs_etm__get_v7_protocol_version(u32 etmidr) 119 { 120 etmidr &= ETMIDR_PTM_VERSION; 121 122 if (etmidr == ETMIDR_PTM_VERSION) 123 return CS_ETM_PROTO_PTM; 124 125 return CS_ETM_PROTO_ETMV3; 126 } 127 128 static int cs_etm__get_magic(u8 trace_chan_id, u64 *magic) 129 { 130 struct int_node *inode; 131 u64 *metadata; 132 133 inode = intlist__find(traceid_list, trace_chan_id); 134 if (!inode) 135 return -EINVAL; 136 137 metadata = inode->priv; 138 *magic = metadata[CS_ETM_MAGIC]; 139 return 0; 140 } 141 142 int cs_etm__get_cpu(u8 trace_chan_id, int *cpu) 143 { 144 struct int_node *inode; 145 u64 *metadata; 146 147 inode = intlist__find(traceid_list, trace_chan_id); 148 if (!inode) 149 return -EINVAL; 150 151 metadata = inode->priv; 152 *cpu = (int)metadata[CS_ETM_CPU]; 153 return 0; 154 } 155 156 void cs_etm__etmq_set_traceid_queue_timestamp(struct cs_etm_queue *etmq, 157 u8 trace_chan_id) 158 { 159 /* 160 * Wnen a timestamp packet is encountered the backend code 161 * is stopped so that the front end has time to process packets 162 * that were accumulated in the traceID queue. Since there can 163 * be more than one channel per cs_etm_queue, we need to specify 164 * what traceID queue needs servicing. 165 */ 166 etmq->pending_timestamp = trace_chan_id; 167 } 168 169 static u64 cs_etm__etmq_get_timestamp(struct cs_etm_queue *etmq, 170 u8 *trace_chan_id) 171 { 172 struct cs_etm_packet_queue *packet_queue; 173 174 if (!etmq->pending_timestamp) 175 return 0; 176 177 if (trace_chan_id) 178 *trace_chan_id = etmq->pending_timestamp; 179 180 packet_queue = cs_etm__etmq_get_packet_queue(etmq, 181 etmq->pending_timestamp); 182 if (!packet_queue) 183 return 0; 184 185 /* Acknowledge pending status */ 186 etmq->pending_timestamp = 0; 187 188 /* See function cs_etm_decoder__do_{hard|soft}_timestamp() */ 189 return packet_queue->timestamp; 190 } 191 192 static void cs_etm__clear_packet_queue(struct cs_etm_packet_queue *queue) 193 { 194 int i; 195 196 queue->head = 0; 197 queue->tail = 0; 198 queue->packet_count = 0; 199 for (i = 0; i < CS_ETM_PACKET_MAX_BUFFER; i++) { 200 queue->packet_buffer[i].isa = CS_ETM_ISA_UNKNOWN; 201 queue->packet_buffer[i].start_addr = CS_ETM_INVAL_ADDR; 202 queue->packet_buffer[i].end_addr = CS_ETM_INVAL_ADDR; 203 queue->packet_buffer[i].instr_count = 0; 204 queue->packet_buffer[i].last_instr_taken_branch = false; 205 queue->packet_buffer[i].last_instr_size = 0; 206 queue->packet_buffer[i].last_instr_type = 0; 207 queue->packet_buffer[i].last_instr_subtype = 0; 208 queue->packet_buffer[i].last_instr_cond = 0; 209 queue->packet_buffer[i].flags = 0; 210 queue->packet_buffer[i].exception_number = UINT32_MAX; 211 queue->packet_buffer[i].trace_chan_id = UINT8_MAX; 212 queue->packet_buffer[i].cpu = INT_MIN; 213 } 214 } 215 216 static void cs_etm__clear_all_packet_queues(struct cs_etm_queue *etmq) 217 { 218 int idx; 219 struct int_node *inode; 220 struct cs_etm_traceid_queue *tidq; 221 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 222 223 intlist__for_each_entry(inode, traceid_queues_list) { 224 idx = (int)(intptr_t)inode->priv; 225 tidq = etmq->traceid_queues[idx]; 226 cs_etm__clear_packet_queue(&tidq->packet_queue); 227 } 228 } 229 230 static int cs_etm__init_traceid_queue(struct cs_etm_queue *etmq, 231 struct cs_etm_traceid_queue *tidq, 232 u8 trace_chan_id) 233 { 234 int rc = -ENOMEM; 235 struct auxtrace_queue *queue; 236 struct cs_etm_auxtrace *etm = etmq->etm; 237 238 cs_etm__clear_packet_queue(&tidq->packet_queue); 239 240 queue = &etmq->etm->queues.queue_array[etmq->queue_nr]; 241 tidq->tid = queue->tid; 242 tidq->pid = -1; 243 tidq->trace_chan_id = trace_chan_id; 244 245 tidq->packet = zalloc(sizeof(struct cs_etm_packet)); 246 if (!tidq->packet) 247 goto out; 248 249 tidq->prev_packet = zalloc(sizeof(struct cs_etm_packet)); 250 if (!tidq->prev_packet) 251 goto out_free; 252 253 if (etm->synth_opts.last_branch) { 254 size_t sz = sizeof(struct branch_stack); 255 256 sz += etm->synth_opts.last_branch_sz * 257 sizeof(struct branch_entry); 258 tidq->last_branch = zalloc(sz); 259 if (!tidq->last_branch) 260 goto out_free; 261 tidq->last_branch_rb = zalloc(sz); 262 if (!tidq->last_branch_rb) 263 goto out_free; 264 } 265 266 tidq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE); 267 if (!tidq->event_buf) 268 goto out_free; 269 270 return 0; 271 272 out_free: 273 zfree(&tidq->last_branch_rb); 274 zfree(&tidq->last_branch); 275 zfree(&tidq->prev_packet); 276 zfree(&tidq->packet); 277 out: 278 return rc; 279 } 280 281 static struct cs_etm_traceid_queue 282 *cs_etm__etmq_get_traceid_queue(struct cs_etm_queue *etmq, u8 trace_chan_id) 283 { 284 int idx; 285 struct int_node *inode; 286 struct intlist *traceid_queues_list; 287 struct cs_etm_traceid_queue *tidq, **traceid_queues; 288 struct cs_etm_auxtrace *etm = etmq->etm; 289 290 if (etm->timeless_decoding) 291 trace_chan_id = CS_ETM_PER_THREAD_TRACEID; 292 293 traceid_queues_list = etmq->traceid_queues_list; 294 295 /* 296 * Check if the traceid_queue exist for this traceID by looking 297 * in the queue list. 298 */ 299 inode = intlist__find(traceid_queues_list, trace_chan_id); 300 if (inode) { 301 idx = (int)(intptr_t)inode->priv; 302 return etmq->traceid_queues[idx]; 303 } 304 305 /* We couldn't find a traceid_queue for this traceID, allocate one */ 306 tidq = malloc(sizeof(*tidq)); 307 if (!tidq) 308 return NULL; 309 310 memset(tidq, 0, sizeof(*tidq)); 311 312 /* Get a valid index for the new traceid_queue */ 313 idx = intlist__nr_entries(traceid_queues_list); 314 /* Memory for the inode is free'ed in cs_etm_free_traceid_queues () */ 315 inode = intlist__findnew(traceid_queues_list, trace_chan_id); 316 if (!inode) 317 goto out_free; 318 319 /* Associate this traceID with this index */ 320 inode->priv = (void *)(intptr_t)idx; 321 322 if (cs_etm__init_traceid_queue(etmq, tidq, trace_chan_id)) 323 goto out_free; 324 325 /* Grow the traceid_queues array by one unit */ 326 traceid_queues = etmq->traceid_queues; 327 traceid_queues = reallocarray(traceid_queues, 328 idx + 1, 329 sizeof(*traceid_queues)); 330 331 /* 332 * On failure reallocarray() returns NULL and the original block of 333 * memory is left untouched. 334 */ 335 if (!traceid_queues) 336 goto out_free; 337 338 traceid_queues[idx] = tidq; 339 etmq->traceid_queues = traceid_queues; 340 341 return etmq->traceid_queues[idx]; 342 343 out_free: 344 /* 345 * Function intlist__remove() removes the inode from the list 346 * and delete the memory associated to it. 347 */ 348 intlist__remove(traceid_queues_list, inode); 349 free(tidq); 350 351 return NULL; 352 } 353 354 struct cs_etm_packet_queue 355 *cs_etm__etmq_get_packet_queue(struct cs_etm_queue *etmq, u8 trace_chan_id) 356 { 357 struct cs_etm_traceid_queue *tidq; 358 359 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 360 if (tidq) 361 return &tidq->packet_queue; 362 363 return NULL; 364 } 365 366 static void cs_etm__packet_swap(struct cs_etm_auxtrace *etm, 367 struct cs_etm_traceid_queue *tidq) 368 { 369 struct cs_etm_packet *tmp; 370 371 if (etm->sample_branches || etm->synth_opts.last_branch || 372 etm->sample_instructions) { 373 /* 374 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for 375 * the next incoming packet. 376 */ 377 tmp = tidq->packet; 378 tidq->packet = tidq->prev_packet; 379 tidq->prev_packet = tmp; 380 } 381 } 382 383 static void cs_etm__packet_dump(const char *pkt_string) 384 { 385 const char *color = PERF_COLOR_BLUE; 386 int len = strlen(pkt_string); 387 388 if (len && (pkt_string[len-1] == '\n')) 389 color_fprintf(stdout, color, " %s", pkt_string); 390 else 391 color_fprintf(stdout, color, " %s\n", pkt_string); 392 393 fflush(stdout); 394 } 395 396 static void cs_etm__set_trace_param_etmv3(struct cs_etm_trace_params *t_params, 397 struct cs_etm_auxtrace *etm, int idx, 398 u32 etmidr) 399 { 400 u64 **metadata = etm->metadata; 401 402 t_params[idx].protocol = cs_etm__get_v7_protocol_version(etmidr); 403 t_params[idx].etmv3.reg_ctrl = metadata[idx][CS_ETM_ETMCR]; 404 t_params[idx].etmv3.reg_trc_id = metadata[idx][CS_ETM_ETMTRACEIDR]; 405 } 406 407 static void cs_etm__set_trace_param_etmv4(struct cs_etm_trace_params *t_params, 408 struct cs_etm_auxtrace *etm, int idx) 409 { 410 u64 **metadata = etm->metadata; 411 412 t_params[idx].protocol = CS_ETM_PROTO_ETMV4i; 413 t_params[idx].etmv4.reg_idr0 = metadata[idx][CS_ETMV4_TRCIDR0]; 414 t_params[idx].etmv4.reg_idr1 = metadata[idx][CS_ETMV4_TRCIDR1]; 415 t_params[idx].etmv4.reg_idr2 = metadata[idx][CS_ETMV4_TRCIDR2]; 416 t_params[idx].etmv4.reg_idr8 = metadata[idx][CS_ETMV4_TRCIDR8]; 417 t_params[idx].etmv4.reg_configr = metadata[idx][CS_ETMV4_TRCCONFIGR]; 418 t_params[idx].etmv4.reg_traceidr = metadata[idx][CS_ETMV4_TRCTRACEIDR]; 419 } 420 421 static int cs_etm__init_trace_params(struct cs_etm_trace_params *t_params, 422 struct cs_etm_auxtrace *etm) 423 { 424 int i; 425 u32 etmidr; 426 u64 architecture; 427 428 for (i = 0; i < etm->num_cpu; i++) { 429 architecture = etm->metadata[i][CS_ETM_MAGIC]; 430 431 switch (architecture) { 432 case __perf_cs_etmv3_magic: 433 etmidr = etm->metadata[i][CS_ETM_ETMIDR]; 434 cs_etm__set_trace_param_etmv3(t_params, etm, i, etmidr); 435 break; 436 case __perf_cs_etmv4_magic: 437 cs_etm__set_trace_param_etmv4(t_params, etm, i); 438 break; 439 default: 440 return -EINVAL; 441 } 442 } 443 444 return 0; 445 } 446 447 static int cs_etm__init_decoder_params(struct cs_etm_decoder_params *d_params, 448 struct cs_etm_queue *etmq, 449 enum cs_etm_decoder_operation mode) 450 { 451 int ret = -EINVAL; 452 453 if (!(mode < CS_ETM_OPERATION_MAX)) 454 goto out; 455 456 d_params->packet_printer = cs_etm__packet_dump; 457 d_params->operation = mode; 458 d_params->data = etmq; 459 d_params->formatted = true; 460 d_params->fsyncs = false; 461 d_params->hsyncs = false; 462 d_params->frame_aligned = true; 463 464 ret = 0; 465 out: 466 return ret; 467 } 468 469 static void cs_etm__dump_event(struct cs_etm_auxtrace *etm, 470 struct auxtrace_buffer *buffer) 471 { 472 int ret; 473 const char *color = PERF_COLOR_BLUE; 474 struct cs_etm_decoder_params d_params; 475 struct cs_etm_trace_params *t_params; 476 struct cs_etm_decoder *decoder; 477 size_t buffer_used = 0; 478 479 fprintf(stdout, "\n"); 480 color_fprintf(stdout, color, 481 ". ... CoreSight ETM Trace data: size %zu bytes\n", 482 buffer->size); 483 484 /* Use metadata to fill in trace parameters for trace decoder */ 485 t_params = zalloc(sizeof(*t_params) * etm->num_cpu); 486 487 if (!t_params) 488 return; 489 490 if (cs_etm__init_trace_params(t_params, etm)) 491 goto out_free; 492 493 /* Set decoder parameters to simply print the trace packets */ 494 if (cs_etm__init_decoder_params(&d_params, NULL, 495 CS_ETM_OPERATION_PRINT)) 496 goto out_free; 497 498 decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params); 499 500 if (!decoder) 501 goto out_free; 502 do { 503 size_t consumed; 504 505 ret = cs_etm_decoder__process_data_block( 506 decoder, buffer->offset, 507 &((u8 *)buffer->data)[buffer_used], 508 buffer->size - buffer_used, &consumed); 509 if (ret) 510 break; 511 512 buffer_used += consumed; 513 } while (buffer_used < buffer->size); 514 515 cs_etm_decoder__free(decoder); 516 517 out_free: 518 zfree(&t_params); 519 } 520 521 static int cs_etm__flush_events(struct perf_session *session, 522 struct perf_tool *tool) 523 { 524 int ret; 525 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 526 struct cs_etm_auxtrace, 527 auxtrace); 528 if (dump_trace) 529 return 0; 530 531 if (!tool->ordered_events) 532 return -EINVAL; 533 534 ret = cs_etm__update_queues(etm); 535 536 if (ret < 0) 537 return ret; 538 539 if (etm->timeless_decoding) 540 return cs_etm__process_timeless_queues(etm, -1); 541 542 return cs_etm__process_queues(etm); 543 } 544 545 static void cs_etm__free_traceid_queues(struct cs_etm_queue *etmq) 546 { 547 int idx; 548 uintptr_t priv; 549 struct int_node *inode, *tmp; 550 struct cs_etm_traceid_queue *tidq; 551 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 552 553 intlist__for_each_entry_safe(inode, tmp, traceid_queues_list) { 554 priv = (uintptr_t)inode->priv; 555 idx = priv; 556 557 /* Free this traceid_queue from the array */ 558 tidq = etmq->traceid_queues[idx]; 559 thread__zput(tidq->thread); 560 zfree(&tidq->event_buf); 561 zfree(&tidq->last_branch); 562 zfree(&tidq->last_branch_rb); 563 zfree(&tidq->prev_packet); 564 zfree(&tidq->packet); 565 zfree(&tidq); 566 567 /* 568 * Function intlist__remove() removes the inode from the list 569 * and delete the memory associated to it. 570 */ 571 intlist__remove(traceid_queues_list, inode); 572 } 573 574 /* Then the RB tree itself */ 575 intlist__delete(traceid_queues_list); 576 etmq->traceid_queues_list = NULL; 577 578 /* finally free the traceid_queues array */ 579 zfree(&etmq->traceid_queues); 580 } 581 582 static void cs_etm__free_queue(void *priv) 583 { 584 struct cs_etm_queue *etmq = priv; 585 586 if (!etmq) 587 return; 588 589 cs_etm_decoder__free(etmq->decoder); 590 cs_etm__free_traceid_queues(etmq); 591 free(etmq); 592 } 593 594 static void cs_etm__free_events(struct perf_session *session) 595 { 596 unsigned int i; 597 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 598 struct cs_etm_auxtrace, 599 auxtrace); 600 struct auxtrace_queues *queues = &aux->queues; 601 602 for (i = 0; i < queues->nr_queues; i++) { 603 cs_etm__free_queue(queues->queue_array[i].priv); 604 queues->queue_array[i].priv = NULL; 605 } 606 607 auxtrace_queues__free(queues); 608 } 609 610 static void cs_etm__free(struct perf_session *session) 611 { 612 int i; 613 struct int_node *inode, *tmp; 614 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 615 struct cs_etm_auxtrace, 616 auxtrace); 617 cs_etm__free_events(session); 618 session->auxtrace = NULL; 619 620 /* First remove all traceID/metadata nodes for the RB tree */ 621 intlist__for_each_entry_safe(inode, tmp, traceid_list) 622 intlist__remove(traceid_list, inode); 623 /* Then the RB tree itself */ 624 intlist__delete(traceid_list); 625 626 for (i = 0; i < aux->num_cpu; i++) 627 zfree(&aux->metadata[i]); 628 629 thread__zput(aux->unknown_thread); 630 zfree(&aux->metadata); 631 zfree(&aux); 632 } 633 634 static u8 cs_etm__cpu_mode(struct cs_etm_queue *etmq, u64 address) 635 { 636 struct machine *machine; 637 638 machine = etmq->etm->machine; 639 640 if (address >= etmq->etm->kernel_start) { 641 if (machine__is_host(machine)) 642 return PERF_RECORD_MISC_KERNEL; 643 else 644 return PERF_RECORD_MISC_GUEST_KERNEL; 645 } else { 646 if (machine__is_host(machine)) 647 return PERF_RECORD_MISC_USER; 648 else if (perf_guest) 649 return PERF_RECORD_MISC_GUEST_USER; 650 else 651 return PERF_RECORD_MISC_HYPERVISOR; 652 } 653 } 654 655 static u32 cs_etm__mem_access(struct cs_etm_queue *etmq, u8 trace_chan_id, 656 u64 address, size_t size, u8 *buffer) 657 { 658 u8 cpumode; 659 u64 offset; 660 int len; 661 struct thread *thread; 662 struct machine *machine; 663 struct addr_location al; 664 struct cs_etm_traceid_queue *tidq; 665 666 if (!etmq) 667 return 0; 668 669 machine = etmq->etm->machine; 670 cpumode = cs_etm__cpu_mode(etmq, address); 671 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 672 if (!tidq) 673 return 0; 674 675 thread = tidq->thread; 676 if (!thread) { 677 if (cpumode != PERF_RECORD_MISC_KERNEL) 678 return 0; 679 thread = etmq->etm->unknown_thread; 680 } 681 682 if (!thread__find_map(thread, cpumode, address, &al) || !al.map->dso) 683 return 0; 684 685 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR && 686 dso__data_status_seen(al.map->dso, DSO_DATA_STATUS_SEEN_ITRACE)) 687 return 0; 688 689 offset = al.map->map_ip(al.map, address); 690 691 map__load(al.map); 692 693 len = dso__data_read_offset(al.map->dso, machine, offset, buffer, size); 694 695 if (len <= 0) 696 return 0; 697 698 return len; 699 } 700 701 static struct cs_etm_queue *cs_etm__alloc_queue(struct cs_etm_auxtrace *etm) 702 { 703 struct cs_etm_decoder_params d_params; 704 struct cs_etm_trace_params *t_params = NULL; 705 struct cs_etm_queue *etmq; 706 707 etmq = zalloc(sizeof(*etmq)); 708 if (!etmq) 709 return NULL; 710 711 etmq->traceid_queues_list = intlist__new(NULL); 712 if (!etmq->traceid_queues_list) 713 goto out_free; 714 715 /* Use metadata to fill in trace parameters for trace decoder */ 716 t_params = zalloc(sizeof(*t_params) * etm->num_cpu); 717 718 if (!t_params) 719 goto out_free; 720 721 if (cs_etm__init_trace_params(t_params, etm)) 722 goto out_free; 723 724 /* Set decoder parameters to decode trace packets */ 725 if (cs_etm__init_decoder_params(&d_params, etmq, 726 CS_ETM_OPERATION_DECODE)) 727 goto out_free; 728 729 etmq->decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params); 730 731 if (!etmq->decoder) 732 goto out_free; 733 734 /* 735 * Register a function to handle all memory accesses required by 736 * the trace decoder library. 737 */ 738 if (cs_etm_decoder__add_mem_access_cb(etmq->decoder, 739 0x0L, ((u64) -1L), 740 cs_etm__mem_access)) 741 goto out_free_decoder; 742 743 zfree(&t_params); 744 return etmq; 745 746 out_free_decoder: 747 cs_etm_decoder__free(etmq->decoder); 748 out_free: 749 intlist__delete(etmq->traceid_queues_list); 750 free(etmq); 751 752 return NULL; 753 } 754 755 static int cs_etm__setup_queue(struct cs_etm_auxtrace *etm, 756 struct auxtrace_queue *queue, 757 unsigned int queue_nr) 758 { 759 int ret = 0; 760 unsigned int cs_queue_nr; 761 u8 trace_chan_id; 762 u64 timestamp; 763 struct cs_etm_queue *etmq = queue->priv; 764 765 if (list_empty(&queue->head) || etmq) 766 goto out; 767 768 etmq = cs_etm__alloc_queue(etm); 769 770 if (!etmq) { 771 ret = -ENOMEM; 772 goto out; 773 } 774 775 queue->priv = etmq; 776 etmq->etm = etm; 777 etmq->queue_nr = queue_nr; 778 etmq->offset = 0; 779 780 if (etm->timeless_decoding) 781 goto out; 782 783 /* 784 * We are under a CPU-wide trace scenario. As such we need to know 785 * when the code that generated the traces started to execute so that 786 * it can be correlated with execution on other CPUs. So we get a 787 * handle on the beginning of traces and decode until we find a 788 * timestamp. The timestamp is then added to the auxtrace min heap 789 * in order to know what nibble (of all the etmqs) to decode first. 790 */ 791 while (1) { 792 /* 793 * Fetch an aux_buffer from this etmq. Bail if no more 794 * blocks or an error has been encountered. 795 */ 796 ret = cs_etm__get_data_block(etmq); 797 if (ret <= 0) 798 goto out; 799 800 /* 801 * Run decoder on the trace block. The decoder will stop when 802 * encountering a timestamp, a full packet queue or the end of 803 * trace for that block. 804 */ 805 ret = cs_etm__decode_data_block(etmq); 806 if (ret) 807 goto out; 808 809 /* 810 * Function cs_etm_decoder__do_{hard|soft}_timestamp() does all 811 * the timestamp calculation for us. 812 */ 813 timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id); 814 815 /* We found a timestamp, no need to continue. */ 816 if (timestamp) 817 break; 818 819 /* 820 * We didn't find a timestamp so empty all the traceid packet 821 * queues before looking for another timestamp packet, either 822 * in the current data block or a new one. Packets that were 823 * just decoded are useless since no timestamp has been 824 * associated with them. As such simply discard them. 825 */ 826 cs_etm__clear_all_packet_queues(etmq); 827 } 828 829 /* 830 * We have a timestamp. Add it to the min heap to reflect when 831 * instructions conveyed by the range packets of this traceID queue 832 * started to execute. Once the same has been done for all the traceID 833 * queues of each etmq, redenring and decoding can start in 834 * chronological order. 835 * 836 * Note that packets decoded above are still in the traceID's packet 837 * queue and will be processed in cs_etm__process_queues(). 838 */ 839 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id); 840 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, timestamp); 841 out: 842 return ret; 843 } 844 845 static int cs_etm__setup_queues(struct cs_etm_auxtrace *etm) 846 { 847 unsigned int i; 848 int ret; 849 850 if (!etm->kernel_start) 851 etm->kernel_start = machine__kernel_start(etm->machine); 852 853 for (i = 0; i < etm->queues.nr_queues; i++) { 854 ret = cs_etm__setup_queue(etm, &etm->queues.queue_array[i], i); 855 if (ret) 856 return ret; 857 } 858 859 return 0; 860 } 861 862 static int cs_etm__update_queues(struct cs_etm_auxtrace *etm) 863 { 864 if (etm->queues.new_data) { 865 etm->queues.new_data = false; 866 return cs_etm__setup_queues(etm); 867 } 868 869 return 0; 870 } 871 872 static inline 873 void cs_etm__copy_last_branch_rb(struct cs_etm_queue *etmq, 874 struct cs_etm_traceid_queue *tidq) 875 { 876 struct branch_stack *bs_src = tidq->last_branch_rb; 877 struct branch_stack *bs_dst = tidq->last_branch; 878 size_t nr = 0; 879 880 /* 881 * Set the number of records before early exit: ->nr is used to 882 * determine how many branches to copy from ->entries. 883 */ 884 bs_dst->nr = bs_src->nr; 885 886 /* 887 * Early exit when there is nothing to copy. 888 */ 889 if (!bs_src->nr) 890 return; 891 892 /* 893 * As bs_src->entries is a circular buffer, we need to copy from it in 894 * two steps. First, copy the branches from the most recently inserted 895 * branch ->last_branch_pos until the end of bs_src->entries buffer. 896 */ 897 nr = etmq->etm->synth_opts.last_branch_sz - tidq->last_branch_pos; 898 memcpy(&bs_dst->entries[0], 899 &bs_src->entries[tidq->last_branch_pos], 900 sizeof(struct branch_entry) * nr); 901 902 /* 903 * If we wrapped around at least once, the branches from the beginning 904 * of the bs_src->entries buffer and until the ->last_branch_pos element 905 * are older valid branches: copy them over. The total number of 906 * branches copied over will be equal to the number of branches asked by 907 * the user in last_branch_sz. 908 */ 909 if (bs_src->nr >= etmq->etm->synth_opts.last_branch_sz) { 910 memcpy(&bs_dst->entries[nr], 911 &bs_src->entries[0], 912 sizeof(struct branch_entry) * tidq->last_branch_pos); 913 } 914 } 915 916 static inline 917 void cs_etm__reset_last_branch_rb(struct cs_etm_traceid_queue *tidq) 918 { 919 tidq->last_branch_pos = 0; 920 tidq->last_branch_rb->nr = 0; 921 } 922 923 static inline int cs_etm__t32_instr_size(struct cs_etm_queue *etmq, 924 u8 trace_chan_id, u64 addr) 925 { 926 u8 instrBytes[2]; 927 928 cs_etm__mem_access(etmq, trace_chan_id, addr, 929 ARRAY_SIZE(instrBytes), instrBytes); 930 /* 931 * T32 instruction size is indicated by bits[15:11] of the first 932 * 16-bit word of the instruction: 0b11101, 0b11110 and 0b11111 933 * denote a 32-bit instruction. 934 */ 935 return ((instrBytes[1] & 0xF8) >= 0xE8) ? 4 : 2; 936 } 937 938 static inline u64 cs_etm__first_executed_instr(struct cs_etm_packet *packet) 939 { 940 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */ 941 if (packet->sample_type == CS_ETM_DISCONTINUITY) 942 return 0; 943 944 return packet->start_addr; 945 } 946 947 static inline 948 u64 cs_etm__last_executed_instr(const struct cs_etm_packet *packet) 949 { 950 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */ 951 if (packet->sample_type == CS_ETM_DISCONTINUITY) 952 return 0; 953 954 return packet->end_addr - packet->last_instr_size; 955 } 956 957 static inline u64 cs_etm__instr_addr(struct cs_etm_queue *etmq, 958 u64 trace_chan_id, 959 const struct cs_etm_packet *packet, 960 u64 offset) 961 { 962 if (packet->isa == CS_ETM_ISA_T32) { 963 u64 addr = packet->start_addr; 964 965 while (offset) { 966 addr += cs_etm__t32_instr_size(etmq, 967 trace_chan_id, addr); 968 offset--; 969 } 970 return addr; 971 } 972 973 /* Assume a 4 byte instruction size (A32/A64) */ 974 return packet->start_addr + offset * 4; 975 } 976 977 static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq, 978 struct cs_etm_traceid_queue *tidq) 979 { 980 struct branch_stack *bs = tidq->last_branch_rb; 981 struct branch_entry *be; 982 983 /* 984 * The branches are recorded in a circular buffer in reverse 985 * chronological order: we start recording from the last element of the 986 * buffer down. After writing the first element of the stack, move the 987 * insert position back to the end of the buffer. 988 */ 989 if (!tidq->last_branch_pos) 990 tidq->last_branch_pos = etmq->etm->synth_opts.last_branch_sz; 991 992 tidq->last_branch_pos -= 1; 993 994 be = &bs->entries[tidq->last_branch_pos]; 995 be->from = cs_etm__last_executed_instr(tidq->prev_packet); 996 be->to = cs_etm__first_executed_instr(tidq->packet); 997 /* No support for mispredict */ 998 be->flags.mispred = 0; 999 be->flags.predicted = 1; 1000 1001 /* 1002 * Increment bs->nr until reaching the number of last branches asked by 1003 * the user on the command line. 1004 */ 1005 if (bs->nr < etmq->etm->synth_opts.last_branch_sz) 1006 bs->nr += 1; 1007 } 1008 1009 static int cs_etm__inject_event(union perf_event *event, 1010 struct perf_sample *sample, u64 type) 1011 { 1012 event->header.size = perf_event__sample_event_size(sample, type, 0); 1013 return perf_event__synthesize_sample(event, type, 0, sample); 1014 } 1015 1016 1017 static int 1018 cs_etm__get_trace(struct cs_etm_queue *etmq) 1019 { 1020 struct auxtrace_buffer *aux_buffer = etmq->buffer; 1021 struct auxtrace_buffer *old_buffer = aux_buffer; 1022 struct auxtrace_queue *queue; 1023 1024 queue = &etmq->etm->queues.queue_array[etmq->queue_nr]; 1025 1026 aux_buffer = auxtrace_buffer__next(queue, aux_buffer); 1027 1028 /* If no more data, drop the previous auxtrace_buffer and return */ 1029 if (!aux_buffer) { 1030 if (old_buffer) 1031 auxtrace_buffer__drop_data(old_buffer); 1032 etmq->buf_len = 0; 1033 return 0; 1034 } 1035 1036 etmq->buffer = aux_buffer; 1037 1038 /* If the aux_buffer doesn't have data associated, try to load it */ 1039 if (!aux_buffer->data) { 1040 /* get the file desc associated with the perf data file */ 1041 int fd = perf_data__fd(etmq->etm->session->data); 1042 1043 aux_buffer->data = auxtrace_buffer__get_data(aux_buffer, fd); 1044 if (!aux_buffer->data) 1045 return -ENOMEM; 1046 } 1047 1048 /* If valid, drop the previous buffer */ 1049 if (old_buffer) 1050 auxtrace_buffer__drop_data(old_buffer); 1051 1052 etmq->buf_used = 0; 1053 etmq->buf_len = aux_buffer->size; 1054 etmq->buf = aux_buffer->data; 1055 1056 return etmq->buf_len; 1057 } 1058 1059 static void cs_etm__set_pid_tid_cpu(struct cs_etm_auxtrace *etm, 1060 struct cs_etm_traceid_queue *tidq) 1061 { 1062 if ((!tidq->thread) && (tidq->tid != -1)) 1063 tidq->thread = machine__find_thread(etm->machine, -1, 1064 tidq->tid); 1065 1066 if (tidq->thread) 1067 tidq->pid = tidq->thread->pid_; 1068 } 1069 1070 int cs_etm__etmq_set_tid(struct cs_etm_queue *etmq, 1071 pid_t tid, u8 trace_chan_id) 1072 { 1073 int cpu, err = -EINVAL; 1074 struct cs_etm_auxtrace *etm = etmq->etm; 1075 struct cs_etm_traceid_queue *tidq; 1076 1077 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 1078 if (!tidq) 1079 return err; 1080 1081 if (cs_etm__get_cpu(trace_chan_id, &cpu) < 0) 1082 return err; 1083 1084 err = machine__set_current_tid(etm->machine, cpu, tid, tid); 1085 if (err) 1086 return err; 1087 1088 tidq->tid = tid; 1089 thread__zput(tidq->thread); 1090 1091 cs_etm__set_pid_tid_cpu(etm, tidq); 1092 return 0; 1093 } 1094 1095 bool cs_etm__etmq_is_timeless(struct cs_etm_queue *etmq) 1096 { 1097 return !!etmq->etm->timeless_decoding; 1098 } 1099 1100 static void cs_etm__copy_insn(struct cs_etm_queue *etmq, 1101 u64 trace_chan_id, 1102 const struct cs_etm_packet *packet, 1103 struct perf_sample *sample) 1104 { 1105 /* 1106 * It's pointless to read instructions for the CS_ETM_DISCONTINUITY 1107 * packet, so directly bail out with 'insn_len' = 0. 1108 */ 1109 if (packet->sample_type == CS_ETM_DISCONTINUITY) { 1110 sample->insn_len = 0; 1111 return; 1112 } 1113 1114 /* 1115 * T32 instruction size might be 32-bit or 16-bit, decide by calling 1116 * cs_etm__t32_instr_size(). 1117 */ 1118 if (packet->isa == CS_ETM_ISA_T32) 1119 sample->insn_len = cs_etm__t32_instr_size(etmq, trace_chan_id, 1120 sample->ip); 1121 /* Otherwise, A64 and A32 instruction size are always 32-bit. */ 1122 else 1123 sample->insn_len = 4; 1124 1125 cs_etm__mem_access(etmq, trace_chan_id, sample->ip, 1126 sample->insn_len, (void *)sample->insn); 1127 } 1128 1129 static int cs_etm__synth_instruction_sample(struct cs_etm_queue *etmq, 1130 struct cs_etm_traceid_queue *tidq, 1131 u64 addr, u64 period) 1132 { 1133 int ret = 0; 1134 struct cs_etm_auxtrace *etm = etmq->etm; 1135 union perf_event *event = tidq->event_buf; 1136 struct perf_sample sample = {.ip = 0,}; 1137 1138 event->sample.header.type = PERF_RECORD_SAMPLE; 1139 event->sample.header.misc = cs_etm__cpu_mode(etmq, addr); 1140 event->sample.header.size = sizeof(struct perf_event_header); 1141 1142 sample.ip = addr; 1143 sample.pid = tidq->pid; 1144 sample.tid = tidq->tid; 1145 sample.id = etmq->etm->instructions_id; 1146 sample.stream_id = etmq->etm->instructions_id; 1147 sample.period = period; 1148 sample.cpu = tidq->packet->cpu; 1149 sample.flags = tidq->prev_packet->flags; 1150 sample.cpumode = event->sample.header.misc; 1151 1152 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->packet, &sample); 1153 1154 if (etm->synth_opts.last_branch) 1155 sample.branch_stack = tidq->last_branch; 1156 1157 if (etm->synth_opts.inject) { 1158 ret = cs_etm__inject_event(event, &sample, 1159 etm->instructions_sample_type); 1160 if (ret) 1161 return ret; 1162 } 1163 1164 ret = perf_session__deliver_synth_event(etm->session, event, &sample); 1165 1166 if (ret) 1167 pr_err( 1168 "CS ETM Trace: failed to deliver instruction event, error %d\n", 1169 ret); 1170 1171 return ret; 1172 } 1173 1174 /* 1175 * The cs etm packet encodes an instruction range between a branch target 1176 * and the next taken branch. Generate sample accordingly. 1177 */ 1178 static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq, 1179 struct cs_etm_traceid_queue *tidq) 1180 { 1181 int ret = 0; 1182 struct cs_etm_auxtrace *etm = etmq->etm; 1183 struct perf_sample sample = {.ip = 0,}; 1184 union perf_event *event = tidq->event_buf; 1185 struct dummy_branch_stack { 1186 u64 nr; 1187 u64 hw_idx; 1188 struct branch_entry entries; 1189 } dummy_bs; 1190 u64 ip; 1191 1192 ip = cs_etm__last_executed_instr(tidq->prev_packet); 1193 1194 event->sample.header.type = PERF_RECORD_SAMPLE; 1195 event->sample.header.misc = cs_etm__cpu_mode(etmq, ip); 1196 event->sample.header.size = sizeof(struct perf_event_header); 1197 1198 sample.ip = ip; 1199 sample.pid = tidq->pid; 1200 sample.tid = tidq->tid; 1201 sample.addr = cs_etm__first_executed_instr(tidq->packet); 1202 sample.id = etmq->etm->branches_id; 1203 sample.stream_id = etmq->etm->branches_id; 1204 sample.period = 1; 1205 sample.cpu = tidq->packet->cpu; 1206 sample.flags = tidq->prev_packet->flags; 1207 sample.cpumode = event->sample.header.misc; 1208 1209 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->prev_packet, 1210 &sample); 1211 1212 /* 1213 * perf report cannot handle events without a branch stack 1214 */ 1215 if (etm->synth_opts.last_branch) { 1216 dummy_bs = (struct dummy_branch_stack){ 1217 .nr = 1, 1218 .hw_idx = -1ULL, 1219 .entries = { 1220 .from = sample.ip, 1221 .to = sample.addr, 1222 }, 1223 }; 1224 sample.branch_stack = (struct branch_stack *)&dummy_bs; 1225 } 1226 1227 if (etm->synth_opts.inject) { 1228 ret = cs_etm__inject_event(event, &sample, 1229 etm->branches_sample_type); 1230 if (ret) 1231 return ret; 1232 } 1233 1234 ret = perf_session__deliver_synth_event(etm->session, event, &sample); 1235 1236 if (ret) 1237 pr_err( 1238 "CS ETM Trace: failed to deliver instruction event, error %d\n", 1239 ret); 1240 1241 return ret; 1242 } 1243 1244 struct cs_etm_synth { 1245 struct perf_tool dummy_tool; 1246 struct perf_session *session; 1247 }; 1248 1249 static int cs_etm__event_synth(struct perf_tool *tool, 1250 union perf_event *event, 1251 struct perf_sample *sample __maybe_unused, 1252 struct machine *machine __maybe_unused) 1253 { 1254 struct cs_etm_synth *cs_etm_synth = 1255 container_of(tool, struct cs_etm_synth, dummy_tool); 1256 1257 return perf_session__deliver_synth_event(cs_etm_synth->session, 1258 event, NULL); 1259 } 1260 1261 static int cs_etm__synth_event(struct perf_session *session, 1262 struct perf_event_attr *attr, u64 id) 1263 { 1264 struct cs_etm_synth cs_etm_synth; 1265 1266 memset(&cs_etm_synth, 0, sizeof(struct cs_etm_synth)); 1267 cs_etm_synth.session = session; 1268 1269 return perf_event__synthesize_attr(&cs_etm_synth.dummy_tool, attr, 1, 1270 &id, cs_etm__event_synth); 1271 } 1272 1273 static int cs_etm__synth_events(struct cs_etm_auxtrace *etm, 1274 struct perf_session *session) 1275 { 1276 struct evlist *evlist = session->evlist; 1277 struct evsel *evsel; 1278 struct perf_event_attr attr; 1279 bool found = false; 1280 u64 id; 1281 int err; 1282 1283 evlist__for_each_entry(evlist, evsel) { 1284 if (evsel->core.attr.type == etm->pmu_type) { 1285 found = true; 1286 break; 1287 } 1288 } 1289 1290 if (!found) { 1291 pr_debug("No selected events with CoreSight Trace data\n"); 1292 return 0; 1293 } 1294 1295 memset(&attr, 0, sizeof(struct perf_event_attr)); 1296 attr.size = sizeof(struct perf_event_attr); 1297 attr.type = PERF_TYPE_HARDWARE; 1298 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK; 1299 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID | 1300 PERF_SAMPLE_PERIOD; 1301 if (etm->timeless_decoding) 1302 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME; 1303 else 1304 attr.sample_type |= PERF_SAMPLE_TIME; 1305 1306 attr.exclude_user = evsel->core.attr.exclude_user; 1307 attr.exclude_kernel = evsel->core.attr.exclude_kernel; 1308 attr.exclude_hv = evsel->core.attr.exclude_hv; 1309 attr.exclude_host = evsel->core.attr.exclude_host; 1310 attr.exclude_guest = evsel->core.attr.exclude_guest; 1311 attr.sample_id_all = evsel->core.attr.sample_id_all; 1312 attr.read_format = evsel->core.attr.read_format; 1313 1314 /* create new id val to be a fixed offset from evsel id */ 1315 id = evsel->core.id[0] + 1000000000; 1316 1317 if (!id) 1318 id = 1; 1319 1320 if (etm->synth_opts.branches) { 1321 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; 1322 attr.sample_period = 1; 1323 attr.sample_type |= PERF_SAMPLE_ADDR; 1324 err = cs_etm__synth_event(session, &attr, id); 1325 if (err) 1326 return err; 1327 etm->sample_branches = true; 1328 etm->branches_sample_type = attr.sample_type; 1329 etm->branches_id = id; 1330 id += 1; 1331 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR; 1332 } 1333 1334 if (etm->synth_opts.last_branch) 1335 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK; 1336 1337 if (etm->synth_opts.instructions) { 1338 attr.config = PERF_COUNT_HW_INSTRUCTIONS; 1339 attr.sample_period = etm->synth_opts.period; 1340 etm->instructions_sample_period = attr.sample_period; 1341 err = cs_etm__synth_event(session, &attr, id); 1342 if (err) 1343 return err; 1344 etm->sample_instructions = true; 1345 etm->instructions_sample_type = attr.sample_type; 1346 etm->instructions_id = id; 1347 id += 1; 1348 } 1349 1350 return 0; 1351 } 1352 1353 static int cs_etm__sample(struct cs_etm_queue *etmq, 1354 struct cs_etm_traceid_queue *tidq) 1355 { 1356 struct cs_etm_auxtrace *etm = etmq->etm; 1357 int ret; 1358 u8 trace_chan_id = tidq->trace_chan_id; 1359 u64 instrs_prev; 1360 1361 /* Get instructions remainder from previous packet */ 1362 instrs_prev = tidq->period_instructions; 1363 1364 tidq->period_instructions += tidq->packet->instr_count; 1365 1366 /* 1367 * Record a branch when the last instruction in 1368 * PREV_PACKET is a branch. 1369 */ 1370 if (etm->synth_opts.last_branch && 1371 tidq->prev_packet->sample_type == CS_ETM_RANGE && 1372 tidq->prev_packet->last_instr_taken_branch) 1373 cs_etm__update_last_branch_rb(etmq, tidq); 1374 1375 if (etm->sample_instructions && 1376 tidq->period_instructions >= etm->instructions_sample_period) { 1377 /* 1378 * Emit instruction sample periodically 1379 * TODO: allow period to be defined in cycles and clock time 1380 */ 1381 1382 /* 1383 * Below diagram demonstrates the instruction samples 1384 * generation flows: 1385 * 1386 * Instrs Instrs Instrs Instrs 1387 * Sample(n) Sample(n+1) Sample(n+2) Sample(n+3) 1388 * | | | | 1389 * V V V V 1390 * -------------------------------------------------- 1391 * ^ ^ 1392 * | | 1393 * Period Period 1394 * instructions(Pi) instructions(Pi') 1395 * 1396 * | | 1397 * \---------------- -----------------/ 1398 * V 1399 * tidq->packet->instr_count 1400 * 1401 * Instrs Sample(n...) are the synthesised samples occurring 1402 * every etm->instructions_sample_period instructions - as 1403 * defined on the perf command line. Sample(n) is being the 1404 * last sample before the current etm packet, n+1 to n+3 1405 * samples are generated from the current etm packet. 1406 * 1407 * tidq->packet->instr_count represents the number of 1408 * instructions in the current etm packet. 1409 * 1410 * Period instructions (Pi) contains the the number of 1411 * instructions executed after the sample point(n) from the 1412 * previous etm packet. This will always be less than 1413 * etm->instructions_sample_period. 1414 * 1415 * When generate new samples, it combines with two parts 1416 * instructions, one is the tail of the old packet and another 1417 * is the head of the new coming packet, to generate 1418 * sample(n+1); sample(n+2) and sample(n+3) consume the 1419 * instructions with sample period. After sample(n+3), the rest 1420 * instructions will be used by later packet and it is assigned 1421 * to tidq->period_instructions for next round calculation. 1422 */ 1423 1424 /* 1425 * Get the initial offset into the current packet instructions; 1426 * entry conditions ensure that instrs_prev is less than 1427 * etm->instructions_sample_period. 1428 */ 1429 u64 offset = etm->instructions_sample_period - instrs_prev; 1430 u64 addr; 1431 1432 /* Prepare last branches for instruction sample */ 1433 if (etm->synth_opts.last_branch) 1434 cs_etm__copy_last_branch_rb(etmq, tidq); 1435 1436 while (tidq->period_instructions >= 1437 etm->instructions_sample_period) { 1438 /* 1439 * Calculate the address of the sampled instruction (-1 1440 * as sample is reported as though instruction has just 1441 * been executed, but PC has not advanced to next 1442 * instruction) 1443 */ 1444 addr = cs_etm__instr_addr(etmq, trace_chan_id, 1445 tidq->packet, offset - 1); 1446 ret = cs_etm__synth_instruction_sample( 1447 etmq, tidq, addr, 1448 etm->instructions_sample_period); 1449 if (ret) 1450 return ret; 1451 1452 offset += etm->instructions_sample_period; 1453 tidq->period_instructions -= 1454 etm->instructions_sample_period; 1455 } 1456 } 1457 1458 if (etm->sample_branches) { 1459 bool generate_sample = false; 1460 1461 /* Generate sample for tracing on packet */ 1462 if (tidq->prev_packet->sample_type == CS_ETM_DISCONTINUITY) 1463 generate_sample = true; 1464 1465 /* Generate sample for branch taken packet */ 1466 if (tidq->prev_packet->sample_type == CS_ETM_RANGE && 1467 tidq->prev_packet->last_instr_taken_branch) 1468 generate_sample = true; 1469 1470 if (generate_sample) { 1471 ret = cs_etm__synth_branch_sample(etmq, tidq); 1472 if (ret) 1473 return ret; 1474 } 1475 } 1476 1477 cs_etm__packet_swap(etm, tidq); 1478 1479 return 0; 1480 } 1481 1482 static int cs_etm__exception(struct cs_etm_traceid_queue *tidq) 1483 { 1484 /* 1485 * When the exception packet is inserted, whether the last instruction 1486 * in previous range packet is taken branch or not, we need to force 1487 * to set 'prev_packet->last_instr_taken_branch' to true. This ensures 1488 * to generate branch sample for the instruction range before the 1489 * exception is trapped to kernel or before the exception returning. 1490 * 1491 * The exception packet includes the dummy address values, so don't 1492 * swap PACKET with PREV_PACKET. This keeps PREV_PACKET to be useful 1493 * for generating instruction and branch samples. 1494 */ 1495 if (tidq->prev_packet->sample_type == CS_ETM_RANGE) 1496 tidq->prev_packet->last_instr_taken_branch = true; 1497 1498 return 0; 1499 } 1500 1501 static int cs_etm__flush(struct cs_etm_queue *etmq, 1502 struct cs_etm_traceid_queue *tidq) 1503 { 1504 int err = 0; 1505 struct cs_etm_auxtrace *etm = etmq->etm; 1506 1507 /* Handle start tracing packet */ 1508 if (tidq->prev_packet->sample_type == CS_ETM_EMPTY) 1509 goto swap_packet; 1510 1511 if (etmq->etm->synth_opts.last_branch && 1512 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1513 u64 addr; 1514 1515 /* Prepare last branches for instruction sample */ 1516 cs_etm__copy_last_branch_rb(etmq, tidq); 1517 1518 /* 1519 * Generate a last branch event for the branches left in the 1520 * circular buffer at the end of the trace. 1521 * 1522 * Use the address of the end of the last reported execution 1523 * range 1524 */ 1525 addr = cs_etm__last_executed_instr(tidq->prev_packet); 1526 1527 err = cs_etm__synth_instruction_sample( 1528 etmq, tidq, addr, 1529 tidq->period_instructions); 1530 if (err) 1531 return err; 1532 1533 tidq->period_instructions = 0; 1534 1535 } 1536 1537 if (etm->sample_branches && 1538 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1539 err = cs_etm__synth_branch_sample(etmq, tidq); 1540 if (err) 1541 return err; 1542 } 1543 1544 swap_packet: 1545 cs_etm__packet_swap(etm, tidq); 1546 1547 /* Reset last branches after flush the trace */ 1548 if (etm->synth_opts.last_branch) 1549 cs_etm__reset_last_branch_rb(tidq); 1550 1551 return err; 1552 } 1553 1554 static int cs_etm__end_block(struct cs_etm_queue *etmq, 1555 struct cs_etm_traceid_queue *tidq) 1556 { 1557 int err; 1558 1559 /* 1560 * It has no new packet coming and 'etmq->packet' contains the stale 1561 * packet which was set at the previous time with packets swapping; 1562 * so skip to generate branch sample to avoid stale packet. 1563 * 1564 * For this case only flush branch stack and generate a last branch 1565 * event for the branches left in the circular buffer at the end of 1566 * the trace. 1567 */ 1568 if (etmq->etm->synth_opts.last_branch && 1569 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1570 u64 addr; 1571 1572 /* Prepare last branches for instruction sample */ 1573 cs_etm__copy_last_branch_rb(etmq, tidq); 1574 1575 /* 1576 * Use the address of the end of the last reported execution 1577 * range. 1578 */ 1579 addr = cs_etm__last_executed_instr(tidq->prev_packet); 1580 1581 err = cs_etm__synth_instruction_sample( 1582 etmq, tidq, addr, 1583 tidq->period_instructions); 1584 if (err) 1585 return err; 1586 1587 tidq->period_instructions = 0; 1588 } 1589 1590 return 0; 1591 } 1592 /* 1593 * cs_etm__get_data_block: Fetch a block from the auxtrace_buffer queue 1594 * if need be. 1595 * Returns: < 0 if error 1596 * = 0 if no more auxtrace_buffer to read 1597 * > 0 if the current buffer isn't empty yet 1598 */ 1599 static int cs_etm__get_data_block(struct cs_etm_queue *etmq) 1600 { 1601 int ret; 1602 1603 if (!etmq->buf_len) { 1604 ret = cs_etm__get_trace(etmq); 1605 if (ret <= 0) 1606 return ret; 1607 /* 1608 * We cannot assume consecutive blocks in the data file 1609 * are contiguous, reset the decoder to force re-sync. 1610 */ 1611 ret = cs_etm_decoder__reset(etmq->decoder); 1612 if (ret) 1613 return ret; 1614 } 1615 1616 return etmq->buf_len; 1617 } 1618 1619 static bool cs_etm__is_svc_instr(struct cs_etm_queue *etmq, u8 trace_chan_id, 1620 struct cs_etm_packet *packet, 1621 u64 end_addr) 1622 { 1623 /* Initialise to keep compiler happy */ 1624 u16 instr16 = 0; 1625 u32 instr32 = 0; 1626 u64 addr; 1627 1628 switch (packet->isa) { 1629 case CS_ETM_ISA_T32: 1630 /* 1631 * The SVC of T32 is defined in ARM DDI 0487D.a, F5.1.247: 1632 * 1633 * b'15 b'8 1634 * +-----------------+--------+ 1635 * | 1 1 0 1 1 1 1 1 | imm8 | 1636 * +-----------------+--------+ 1637 * 1638 * According to the specifiction, it only defines SVC for T32 1639 * with 16 bits instruction and has no definition for 32bits; 1640 * so below only read 2 bytes as instruction size for T32. 1641 */ 1642 addr = end_addr - 2; 1643 cs_etm__mem_access(etmq, trace_chan_id, addr, 1644 sizeof(instr16), (u8 *)&instr16); 1645 if ((instr16 & 0xFF00) == 0xDF00) 1646 return true; 1647 1648 break; 1649 case CS_ETM_ISA_A32: 1650 /* 1651 * The SVC of A32 is defined in ARM DDI 0487D.a, F5.1.247: 1652 * 1653 * b'31 b'28 b'27 b'24 1654 * +---------+---------+-------------------------+ 1655 * | !1111 | 1 1 1 1 | imm24 | 1656 * +---------+---------+-------------------------+ 1657 */ 1658 addr = end_addr - 4; 1659 cs_etm__mem_access(etmq, trace_chan_id, addr, 1660 sizeof(instr32), (u8 *)&instr32); 1661 if ((instr32 & 0x0F000000) == 0x0F000000 && 1662 (instr32 & 0xF0000000) != 0xF0000000) 1663 return true; 1664 1665 break; 1666 case CS_ETM_ISA_A64: 1667 /* 1668 * The SVC of A64 is defined in ARM DDI 0487D.a, C6.2.294: 1669 * 1670 * b'31 b'21 b'4 b'0 1671 * +-----------------------+---------+-----------+ 1672 * | 1 1 0 1 0 1 0 0 0 0 0 | imm16 | 0 0 0 0 1 | 1673 * +-----------------------+---------+-----------+ 1674 */ 1675 addr = end_addr - 4; 1676 cs_etm__mem_access(etmq, trace_chan_id, addr, 1677 sizeof(instr32), (u8 *)&instr32); 1678 if ((instr32 & 0xFFE0001F) == 0xd4000001) 1679 return true; 1680 1681 break; 1682 case CS_ETM_ISA_UNKNOWN: 1683 default: 1684 break; 1685 } 1686 1687 return false; 1688 } 1689 1690 static bool cs_etm__is_syscall(struct cs_etm_queue *etmq, 1691 struct cs_etm_traceid_queue *tidq, u64 magic) 1692 { 1693 u8 trace_chan_id = tidq->trace_chan_id; 1694 struct cs_etm_packet *packet = tidq->packet; 1695 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1696 1697 if (magic == __perf_cs_etmv3_magic) 1698 if (packet->exception_number == CS_ETMV3_EXC_SVC) 1699 return true; 1700 1701 /* 1702 * ETMv4 exception type CS_ETMV4_EXC_CALL covers SVC, SMC and 1703 * HVC cases; need to check if it's SVC instruction based on 1704 * packet address. 1705 */ 1706 if (magic == __perf_cs_etmv4_magic) { 1707 if (packet->exception_number == CS_ETMV4_EXC_CALL && 1708 cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet, 1709 prev_packet->end_addr)) 1710 return true; 1711 } 1712 1713 return false; 1714 } 1715 1716 static bool cs_etm__is_async_exception(struct cs_etm_traceid_queue *tidq, 1717 u64 magic) 1718 { 1719 struct cs_etm_packet *packet = tidq->packet; 1720 1721 if (magic == __perf_cs_etmv3_magic) 1722 if (packet->exception_number == CS_ETMV3_EXC_DEBUG_HALT || 1723 packet->exception_number == CS_ETMV3_EXC_ASYNC_DATA_ABORT || 1724 packet->exception_number == CS_ETMV3_EXC_PE_RESET || 1725 packet->exception_number == CS_ETMV3_EXC_IRQ || 1726 packet->exception_number == CS_ETMV3_EXC_FIQ) 1727 return true; 1728 1729 if (magic == __perf_cs_etmv4_magic) 1730 if (packet->exception_number == CS_ETMV4_EXC_RESET || 1731 packet->exception_number == CS_ETMV4_EXC_DEBUG_HALT || 1732 packet->exception_number == CS_ETMV4_EXC_SYSTEM_ERROR || 1733 packet->exception_number == CS_ETMV4_EXC_INST_DEBUG || 1734 packet->exception_number == CS_ETMV4_EXC_DATA_DEBUG || 1735 packet->exception_number == CS_ETMV4_EXC_IRQ || 1736 packet->exception_number == CS_ETMV4_EXC_FIQ) 1737 return true; 1738 1739 return false; 1740 } 1741 1742 static bool cs_etm__is_sync_exception(struct cs_etm_queue *etmq, 1743 struct cs_etm_traceid_queue *tidq, 1744 u64 magic) 1745 { 1746 u8 trace_chan_id = tidq->trace_chan_id; 1747 struct cs_etm_packet *packet = tidq->packet; 1748 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1749 1750 if (magic == __perf_cs_etmv3_magic) 1751 if (packet->exception_number == CS_ETMV3_EXC_SMC || 1752 packet->exception_number == CS_ETMV3_EXC_HYP || 1753 packet->exception_number == CS_ETMV3_EXC_JAZELLE_THUMBEE || 1754 packet->exception_number == CS_ETMV3_EXC_UNDEFINED_INSTR || 1755 packet->exception_number == CS_ETMV3_EXC_PREFETCH_ABORT || 1756 packet->exception_number == CS_ETMV3_EXC_DATA_FAULT || 1757 packet->exception_number == CS_ETMV3_EXC_GENERIC) 1758 return true; 1759 1760 if (magic == __perf_cs_etmv4_magic) { 1761 if (packet->exception_number == CS_ETMV4_EXC_TRAP || 1762 packet->exception_number == CS_ETMV4_EXC_ALIGNMENT || 1763 packet->exception_number == CS_ETMV4_EXC_INST_FAULT || 1764 packet->exception_number == CS_ETMV4_EXC_DATA_FAULT) 1765 return true; 1766 1767 /* 1768 * For CS_ETMV4_EXC_CALL, except SVC other instructions 1769 * (SMC, HVC) are taken as sync exceptions. 1770 */ 1771 if (packet->exception_number == CS_ETMV4_EXC_CALL && 1772 !cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet, 1773 prev_packet->end_addr)) 1774 return true; 1775 1776 /* 1777 * ETMv4 has 5 bits for exception number; if the numbers 1778 * are in the range ( CS_ETMV4_EXC_FIQ, CS_ETMV4_EXC_END ] 1779 * they are implementation defined exceptions. 1780 * 1781 * For this case, simply take it as sync exception. 1782 */ 1783 if (packet->exception_number > CS_ETMV4_EXC_FIQ && 1784 packet->exception_number <= CS_ETMV4_EXC_END) 1785 return true; 1786 } 1787 1788 return false; 1789 } 1790 1791 static int cs_etm__set_sample_flags(struct cs_etm_queue *etmq, 1792 struct cs_etm_traceid_queue *tidq) 1793 { 1794 struct cs_etm_packet *packet = tidq->packet; 1795 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1796 u8 trace_chan_id = tidq->trace_chan_id; 1797 u64 magic; 1798 int ret; 1799 1800 switch (packet->sample_type) { 1801 case CS_ETM_RANGE: 1802 /* 1803 * Immediate branch instruction without neither link nor 1804 * return flag, it's normal branch instruction within 1805 * the function. 1806 */ 1807 if (packet->last_instr_type == OCSD_INSTR_BR && 1808 packet->last_instr_subtype == OCSD_S_INSTR_NONE) { 1809 packet->flags = PERF_IP_FLAG_BRANCH; 1810 1811 if (packet->last_instr_cond) 1812 packet->flags |= PERF_IP_FLAG_CONDITIONAL; 1813 } 1814 1815 /* 1816 * Immediate branch instruction with link (e.g. BL), this is 1817 * branch instruction for function call. 1818 */ 1819 if (packet->last_instr_type == OCSD_INSTR_BR && 1820 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK) 1821 packet->flags = PERF_IP_FLAG_BRANCH | 1822 PERF_IP_FLAG_CALL; 1823 1824 /* 1825 * Indirect branch instruction with link (e.g. BLR), this is 1826 * branch instruction for function call. 1827 */ 1828 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1829 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK) 1830 packet->flags = PERF_IP_FLAG_BRANCH | 1831 PERF_IP_FLAG_CALL; 1832 1833 /* 1834 * Indirect branch instruction with subtype of 1835 * OCSD_S_INSTR_V7_IMPLIED_RET, this is explicit hint for 1836 * function return for A32/T32. 1837 */ 1838 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1839 packet->last_instr_subtype == OCSD_S_INSTR_V7_IMPLIED_RET) 1840 packet->flags = PERF_IP_FLAG_BRANCH | 1841 PERF_IP_FLAG_RETURN; 1842 1843 /* 1844 * Indirect branch instruction without link (e.g. BR), usually 1845 * this is used for function return, especially for functions 1846 * within dynamic link lib. 1847 */ 1848 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1849 packet->last_instr_subtype == OCSD_S_INSTR_NONE) 1850 packet->flags = PERF_IP_FLAG_BRANCH | 1851 PERF_IP_FLAG_RETURN; 1852 1853 /* Return instruction for function return. */ 1854 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1855 packet->last_instr_subtype == OCSD_S_INSTR_V8_RET) 1856 packet->flags = PERF_IP_FLAG_BRANCH | 1857 PERF_IP_FLAG_RETURN; 1858 1859 /* 1860 * Decoder might insert a discontinuity in the middle of 1861 * instruction packets, fixup prev_packet with flag 1862 * PERF_IP_FLAG_TRACE_BEGIN to indicate restarting trace. 1863 */ 1864 if (prev_packet->sample_type == CS_ETM_DISCONTINUITY) 1865 prev_packet->flags |= PERF_IP_FLAG_BRANCH | 1866 PERF_IP_FLAG_TRACE_BEGIN; 1867 1868 /* 1869 * If the previous packet is an exception return packet 1870 * and the return address just follows SVC instuction, 1871 * it needs to calibrate the previous packet sample flags 1872 * as PERF_IP_FLAG_SYSCALLRET. 1873 */ 1874 if (prev_packet->flags == (PERF_IP_FLAG_BRANCH | 1875 PERF_IP_FLAG_RETURN | 1876 PERF_IP_FLAG_INTERRUPT) && 1877 cs_etm__is_svc_instr(etmq, trace_chan_id, 1878 packet, packet->start_addr)) 1879 prev_packet->flags = PERF_IP_FLAG_BRANCH | 1880 PERF_IP_FLAG_RETURN | 1881 PERF_IP_FLAG_SYSCALLRET; 1882 break; 1883 case CS_ETM_DISCONTINUITY: 1884 /* 1885 * The trace is discontinuous, if the previous packet is 1886 * instruction packet, set flag PERF_IP_FLAG_TRACE_END 1887 * for previous packet. 1888 */ 1889 if (prev_packet->sample_type == CS_ETM_RANGE) 1890 prev_packet->flags |= PERF_IP_FLAG_BRANCH | 1891 PERF_IP_FLAG_TRACE_END; 1892 break; 1893 case CS_ETM_EXCEPTION: 1894 ret = cs_etm__get_magic(packet->trace_chan_id, &magic); 1895 if (ret) 1896 return ret; 1897 1898 /* The exception is for system call. */ 1899 if (cs_etm__is_syscall(etmq, tidq, magic)) 1900 packet->flags = PERF_IP_FLAG_BRANCH | 1901 PERF_IP_FLAG_CALL | 1902 PERF_IP_FLAG_SYSCALLRET; 1903 /* 1904 * The exceptions are triggered by external signals from bus, 1905 * interrupt controller, debug module, PE reset or halt. 1906 */ 1907 else if (cs_etm__is_async_exception(tidq, magic)) 1908 packet->flags = PERF_IP_FLAG_BRANCH | 1909 PERF_IP_FLAG_CALL | 1910 PERF_IP_FLAG_ASYNC | 1911 PERF_IP_FLAG_INTERRUPT; 1912 /* 1913 * Otherwise, exception is caused by trap, instruction & 1914 * data fault, or alignment errors. 1915 */ 1916 else if (cs_etm__is_sync_exception(etmq, tidq, magic)) 1917 packet->flags = PERF_IP_FLAG_BRANCH | 1918 PERF_IP_FLAG_CALL | 1919 PERF_IP_FLAG_INTERRUPT; 1920 1921 /* 1922 * When the exception packet is inserted, since exception 1923 * packet is not used standalone for generating samples 1924 * and it's affiliation to the previous instruction range 1925 * packet; so set previous range packet flags to tell perf 1926 * it is an exception taken branch. 1927 */ 1928 if (prev_packet->sample_type == CS_ETM_RANGE) 1929 prev_packet->flags = packet->flags; 1930 break; 1931 case CS_ETM_EXCEPTION_RET: 1932 /* 1933 * When the exception return packet is inserted, since 1934 * exception return packet is not used standalone for 1935 * generating samples and it's affiliation to the previous 1936 * instruction range packet; so set previous range packet 1937 * flags to tell perf it is an exception return branch. 1938 * 1939 * The exception return can be for either system call or 1940 * other exception types; unfortunately the packet doesn't 1941 * contain exception type related info so we cannot decide 1942 * the exception type purely based on exception return packet. 1943 * If we record the exception number from exception packet and 1944 * reuse it for excpetion return packet, this is not reliable 1945 * due the trace can be discontinuity or the interrupt can 1946 * be nested, thus the recorded exception number cannot be 1947 * used for exception return packet for these two cases. 1948 * 1949 * For exception return packet, we only need to distinguish the 1950 * packet is for system call or for other types. Thus the 1951 * decision can be deferred when receive the next packet which 1952 * contains the return address, based on the return address we 1953 * can read out the previous instruction and check if it's a 1954 * system call instruction and then calibrate the sample flag 1955 * as needed. 1956 */ 1957 if (prev_packet->sample_type == CS_ETM_RANGE) 1958 prev_packet->flags = PERF_IP_FLAG_BRANCH | 1959 PERF_IP_FLAG_RETURN | 1960 PERF_IP_FLAG_INTERRUPT; 1961 break; 1962 case CS_ETM_EMPTY: 1963 default: 1964 break; 1965 } 1966 1967 return 0; 1968 } 1969 1970 static int cs_etm__decode_data_block(struct cs_etm_queue *etmq) 1971 { 1972 int ret = 0; 1973 size_t processed = 0; 1974 1975 /* 1976 * Packets are decoded and added to the decoder's packet queue 1977 * until the decoder packet processing callback has requested that 1978 * processing stops or there is nothing left in the buffer. Normal 1979 * operations that stop processing are a timestamp packet or a full 1980 * decoder buffer queue. 1981 */ 1982 ret = cs_etm_decoder__process_data_block(etmq->decoder, 1983 etmq->offset, 1984 &etmq->buf[etmq->buf_used], 1985 etmq->buf_len, 1986 &processed); 1987 if (ret) 1988 goto out; 1989 1990 etmq->offset += processed; 1991 etmq->buf_used += processed; 1992 etmq->buf_len -= processed; 1993 1994 out: 1995 return ret; 1996 } 1997 1998 static int cs_etm__process_traceid_queue(struct cs_etm_queue *etmq, 1999 struct cs_etm_traceid_queue *tidq) 2000 { 2001 int ret; 2002 struct cs_etm_packet_queue *packet_queue; 2003 2004 packet_queue = &tidq->packet_queue; 2005 2006 /* Process each packet in this chunk */ 2007 while (1) { 2008 ret = cs_etm_decoder__get_packet(packet_queue, 2009 tidq->packet); 2010 if (ret <= 0) 2011 /* 2012 * Stop processing this chunk on 2013 * end of data or error 2014 */ 2015 break; 2016 2017 /* 2018 * Since packet addresses are swapped in packet 2019 * handling within below switch() statements, 2020 * thus setting sample flags must be called 2021 * prior to switch() statement to use address 2022 * information before packets swapping. 2023 */ 2024 ret = cs_etm__set_sample_flags(etmq, tidq); 2025 if (ret < 0) 2026 break; 2027 2028 switch (tidq->packet->sample_type) { 2029 case CS_ETM_RANGE: 2030 /* 2031 * If the packet contains an instruction 2032 * range, generate instruction sequence 2033 * events. 2034 */ 2035 cs_etm__sample(etmq, tidq); 2036 break; 2037 case CS_ETM_EXCEPTION: 2038 case CS_ETM_EXCEPTION_RET: 2039 /* 2040 * If the exception packet is coming, 2041 * make sure the previous instruction 2042 * range packet to be handled properly. 2043 */ 2044 cs_etm__exception(tidq); 2045 break; 2046 case CS_ETM_DISCONTINUITY: 2047 /* 2048 * Discontinuity in trace, flush 2049 * previous branch stack 2050 */ 2051 cs_etm__flush(etmq, tidq); 2052 break; 2053 case CS_ETM_EMPTY: 2054 /* 2055 * Should not receive empty packet, 2056 * report error. 2057 */ 2058 pr_err("CS ETM Trace: empty packet\n"); 2059 return -EINVAL; 2060 default: 2061 break; 2062 } 2063 } 2064 2065 return ret; 2066 } 2067 2068 static void cs_etm__clear_all_traceid_queues(struct cs_etm_queue *etmq) 2069 { 2070 int idx; 2071 struct int_node *inode; 2072 struct cs_etm_traceid_queue *tidq; 2073 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 2074 2075 intlist__for_each_entry(inode, traceid_queues_list) { 2076 idx = (int)(intptr_t)inode->priv; 2077 tidq = etmq->traceid_queues[idx]; 2078 2079 /* Ignore return value */ 2080 cs_etm__process_traceid_queue(etmq, tidq); 2081 2082 /* 2083 * Generate an instruction sample with the remaining 2084 * branchstack entries. 2085 */ 2086 cs_etm__flush(etmq, tidq); 2087 } 2088 } 2089 2090 static int cs_etm__run_decoder(struct cs_etm_queue *etmq) 2091 { 2092 int err = 0; 2093 struct cs_etm_traceid_queue *tidq; 2094 2095 tidq = cs_etm__etmq_get_traceid_queue(etmq, CS_ETM_PER_THREAD_TRACEID); 2096 if (!tidq) 2097 return -EINVAL; 2098 2099 /* Go through each buffer in the queue and decode them one by one */ 2100 while (1) { 2101 err = cs_etm__get_data_block(etmq); 2102 if (err <= 0) 2103 return err; 2104 2105 /* Run trace decoder until buffer consumed or end of trace */ 2106 do { 2107 err = cs_etm__decode_data_block(etmq); 2108 if (err) 2109 return err; 2110 2111 /* 2112 * Process each packet in this chunk, nothing to do if 2113 * an error occurs other than hoping the next one will 2114 * be better. 2115 */ 2116 err = cs_etm__process_traceid_queue(etmq, tidq); 2117 2118 } while (etmq->buf_len); 2119 2120 if (err == 0) 2121 /* Flush any remaining branch stack entries */ 2122 err = cs_etm__end_block(etmq, tidq); 2123 } 2124 2125 return err; 2126 } 2127 2128 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm, 2129 pid_t tid) 2130 { 2131 unsigned int i; 2132 struct auxtrace_queues *queues = &etm->queues; 2133 2134 for (i = 0; i < queues->nr_queues; i++) { 2135 struct auxtrace_queue *queue = &etm->queues.queue_array[i]; 2136 struct cs_etm_queue *etmq = queue->priv; 2137 struct cs_etm_traceid_queue *tidq; 2138 2139 if (!etmq) 2140 continue; 2141 2142 tidq = cs_etm__etmq_get_traceid_queue(etmq, 2143 CS_ETM_PER_THREAD_TRACEID); 2144 2145 if (!tidq) 2146 continue; 2147 2148 if ((tid == -1) || (tidq->tid == tid)) { 2149 cs_etm__set_pid_tid_cpu(etm, tidq); 2150 cs_etm__run_decoder(etmq); 2151 } 2152 } 2153 2154 return 0; 2155 } 2156 2157 static int cs_etm__process_queues(struct cs_etm_auxtrace *etm) 2158 { 2159 int ret = 0; 2160 unsigned int cs_queue_nr, queue_nr; 2161 u8 trace_chan_id; 2162 u64 timestamp; 2163 struct auxtrace_queue *queue; 2164 struct cs_etm_queue *etmq; 2165 struct cs_etm_traceid_queue *tidq; 2166 2167 while (1) { 2168 if (!etm->heap.heap_cnt) 2169 goto out; 2170 2171 /* Take the entry at the top of the min heap */ 2172 cs_queue_nr = etm->heap.heap_array[0].queue_nr; 2173 queue_nr = TO_QUEUE_NR(cs_queue_nr); 2174 trace_chan_id = TO_TRACE_CHAN_ID(cs_queue_nr); 2175 queue = &etm->queues.queue_array[queue_nr]; 2176 etmq = queue->priv; 2177 2178 /* 2179 * Remove the top entry from the heap since we are about 2180 * to process it. 2181 */ 2182 auxtrace_heap__pop(&etm->heap); 2183 2184 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 2185 if (!tidq) { 2186 /* 2187 * No traceID queue has been allocated for this traceID, 2188 * which means something somewhere went very wrong. No 2189 * other choice than simply exit. 2190 */ 2191 ret = -EINVAL; 2192 goto out; 2193 } 2194 2195 /* 2196 * Packets associated with this timestamp are already in 2197 * the etmq's traceID queue, so process them. 2198 */ 2199 ret = cs_etm__process_traceid_queue(etmq, tidq); 2200 if (ret < 0) 2201 goto out; 2202 2203 /* 2204 * Packets for this timestamp have been processed, time to 2205 * move on to the next timestamp, fetching a new auxtrace_buffer 2206 * if need be. 2207 */ 2208 refetch: 2209 ret = cs_etm__get_data_block(etmq); 2210 if (ret < 0) 2211 goto out; 2212 2213 /* 2214 * No more auxtrace_buffers to process in this etmq, simply 2215 * move on to another entry in the auxtrace_heap. 2216 */ 2217 if (!ret) 2218 continue; 2219 2220 ret = cs_etm__decode_data_block(etmq); 2221 if (ret) 2222 goto out; 2223 2224 timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id); 2225 2226 if (!timestamp) { 2227 /* 2228 * Function cs_etm__decode_data_block() returns when 2229 * there is no more traces to decode in the current 2230 * auxtrace_buffer OR when a timestamp has been 2231 * encountered on any of the traceID queues. Since we 2232 * did not get a timestamp, there is no more traces to 2233 * process in this auxtrace_buffer. As such empty and 2234 * flush all traceID queues. 2235 */ 2236 cs_etm__clear_all_traceid_queues(etmq); 2237 2238 /* Fetch another auxtrace_buffer for this etmq */ 2239 goto refetch; 2240 } 2241 2242 /* 2243 * Add to the min heap the timestamp for packets that have 2244 * just been decoded. They will be processed and synthesized 2245 * during the next call to cs_etm__process_traceid_queue() for 2246 * this queue/traceID. 2247 */ 2248 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id); 2249 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, timestamp); 2250 } 2251 2252 out: 2253 return ret; 2254 } 2255 2256 static int cs_etm__process_itrace_start(struct cs_etm_auxtrace *etm, 2257 union perf_event *event) 2258 { 2259 struct thread *th; 2260 2261 if (etm->timeless_decoding) 2262 return 0; 2263 2264 /* 2265 * Add the tid/pid to the log so that we can get a match when 2266 * we get a contextID from the decoder. 2267 */ 2268 th = machine__findnew_thread(etm->machine, 2269 event->itrace_start.pid, 2270 event->itrace_start.tid); 2271 if (!th) 2272 return -ENOMEM; 2273 2274 thread__put(th); 2275 2276 return 0; 2277 } 2278 2279 static int cs_etm__process_switch_cpu_wide(struct cs_etm_auxtrace *etm, 2280 union perf_event *event) 2281 { 2282 struct thread *th; 2283 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT; 2284 2285 /* 2286 * Context switch in per-thread mode are irrelevant since perf 2287 * will start/stop tracing as the process is scheduled. 2288 */ 2289 if (etm->timeless_decoding) 2290 return 0; 2291 2292 /* 2293 * SWITCH_IN events carry the next process to be switched out while 2294 * SWITCH_OUT events carry the process to be switched in. As such 2295 * we don't care about IN events. 2296 */ 2297 if (!out) 2298 return 0; 2299 2300 /* 2301 * Add the tid/pid to the log so that we can get a match when 2302 * we get a contextID from the decoder. 2303 */ 2304 th = machine__findnew_thread(etm->machine, 2305 event->context_switch.next_prev_pid, 2306 event->context_switch.next_prev_tid); 2307 if (!th) 2308 return -ENOMEM; 2309 2310 thread__put(th); 2311 2312 return 0; 2313 } 2314 2315 static int cs_etm__process_event(struct perf_session *session, 2316 union perf_event *event, 2317 struct perf_sample *sample, 2318 struct perf_tool *tool) 2319 { 2320 int err = 0; 2321 u64 timestamp; 2322 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2323 struct cs_etm_auxtrace, 2324 auxtrace); 2325 2326 if (dump_trace) 2327 return 0; 2328 2329 if (!tool->ordered_events) { 2330 pr_err("CoreSight ETM Trace requires ordered events\n"); 2331 return -EINVAL; 2332 } 2333 2334 if (sample->time && (sample->time != (u64) -1)) 2335 timestamp = sample->time; 2336 else 2337 timestamp = 0; 2338 2339 if (timestamp || etm->timeless_decoding) { 2340 err = cs_etm__update_queues(etm); 2341 if (err) 2342 return err; 2343 } 2344 2345 if (etm->timeless_decoding && 2346 event->header.type == PERF_RECORD_EXIT) 2347 return cs_etm__process_timeless_queues(etm, 2348 event->fork.tid); 2349 2350 if (event->header.type == PERF_RECORD_ITRACE_START) 2351 return cs_etm__process_itrace_start(etm, event); 2352 else if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) 2353 return cs_etm__process_switch_cpu_wide(etm, event); 2354 2355 if (!etm->timeless_decoding && 2356 event->header.type == PERF_RECORD_AUX) 2357 return cs_etm__process_queues(etm); 2358 2359 return 0; 2360 } 2361 2362 static int cs_etm__process_auxtrace_event(struct perf_session *session, 2363 union perf_event *event, 2364 struct perf_tool *tool __maybe_unused) 2365 { 2366 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2367 struct cs_etm_auxtrace, 2368 auxtrace); 2369 if (!etm->data_queued) { 2370 struct auxtrace_buffer *buffer; 2371 off_t data_offset; 2372 int fd = perf_data__fd(session->data); 2373 bool is_pipe = perf_data__is_pipe(session->data); 2374 int err; 2375 2376 if (is_pipe) 2377 data_offset = 0; 2378 else { 2379 data_offset = lseek(fd, 0, SEEK_CUR); 2380 if (data_offset == -1) 2381 return -errno; 2382 } 2383 2384 err = auxtrace_queues__add_event(&etm->queues, session, 2385 event, data_offset, &buffer); 2386 if (err) 2387 return err; 2388 2389 if (dump_trace) 2390 if (auxtrace_buffer__get_data(buffer, fd)) { 2391 cs_etm__dump_event(etm, buffer); 2392 auxtrace_buffer__put_data(buffer); 2393 } 2394 } 2395 2396 return 0; 2397 } 2398 2399 static bool cs_etm__is_timeless_decoding(struct cs_etm_auxtrace *etm) 2400 { 2401 struct evsel *evsel; 2402 struct evlist *evlist = etm->session->evlist; 2403 bool timeless_decoding = true; 2404 2405 /* 2406 * Circle through the list of event and complain if we find one 2407 * with the time bit set. 2408 */ 2409 evlist__for_each_entry(evlist, evsel) { 2410 if ((evsel->core.attr.sample_type & PERF_SAMPLE_TIME)) 2411 timeless_decoding = false; 2412 } 2413 2414 return timeless_decoding; 2415 } 2416 2417 static const char * const cs_etm_global_header_fmts[] = { 2418 [CS_HEADER_VERSION_0] = " Header version %llx\n", 2419 [CS_PMU_TYPE_CPUS] = " PMU type/num cpus %llx\n", 2420 [CS_ETM_SNAPSHOT] = " Snapshot %llx\n", 2421 }; 2422 2423 static const char * const cs_etm_priv_fmts[] = { 2424 [CS_ETM_MAGIC] = " Magic number %llx\n", 2425 [CS_ETM_CPU] = " CPU %lld\n", 2426 [CS_ETM_ETMCR] = " ETMCR %llx\n", 2427 [CS_ETM_ETMTRACEIDR] = " ETMTRACEIDR %llx\n", 2428 [CS_ETM_ETMCCER] = " ETMCCER %llx\n", 2429 [CS_ETM_ETMIDR] = " ETMIDR %llx\n", 2430 }; 2431 2432 static const char * const cs_etmv4_priv_fmts[] = { 2433 [CS_ETM_MAGIC] = " Magic number %llx\n", 2434 [CS_ETM_CPU] = " CPU %lld\n", 2435 [CS_ETMV4_TRCCONFIGR] = " TRCCONFIGR %llx\n", 2436 [CS_ETMV4_TRCTRACEIDR] = " TRCTRACEIDR %llx\n", 2437 [CS_ETMV4_TRCIDR0] = " TRCIDR0 %llx\n", 2438 [CS_ETMV4_TRCIDR1] = " TRCIDR1 %llx\n", 2439 [CS_ETMV4_TRCIDR2] = " TRCIDR2 %llx\n", 2440 [CS_ETMV4_TRCIDR8] = " TRCIDR8 %llx\n", 2441 [CS_ETMV4_TRCAUTHSTATUS] = " TRCAUTHSTATUS %llx\n", 2442 }; 2443 2444 static void cs_etm__print_auxtrace_info(__u64 *val, int num) 2445 { 2446 int i, j, cpu = 0; 2447 2448 for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++) 2449 fprintf(stdout, cs_etm_global_header_fmts[i], val[i]); 2450 2451 for (i = CS_HEADER_VERSION_0_MAX; cpu < num; cpu++) { 2452 if (val[i] == __perf_cs_etmv3_magic) 2453 for (j = 0; j < CS_ETM_PRIV_MAX; j++, i++) 2454 fprintf(stdout, cs_etm_priv_fmts[j], val[i]); 2455 else if (val[i] == __perf_cs_etmv4_magic) 2456 for (j = 0; j < CS_ETMV4_PRIV_MAX; j++, i++) 2457 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]); 2458 else 2459 /* failure.. return */ 2460 return; 2461 } 2462 } 2463 2464 int cs_etm__process_auxtrace_info(union perf_event *event, 2465 struct perf_session *session) 2466 { 2467 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info; 2468 struct cs_etm_auxtrace *etm = NULL; 2469 struct int_node *inode; 2470 unsigned int pmu_type; 2471 int event_header_size = sizeof(struct perf_event_header); 2472 int info_header_size; 2473 int total_size = auxtrace_info->header.size; 2474 int priv_size = 0; 2475 int num_cpu; 2476 int err = 0, idx = -1; 2477 int i, j, k; 2478 u64 *ptr, *hdr = NULL; 2479 u64 **metadata = NULL; 2480 2481 /* 2482 * sizeof(auxtrace_info_event::type) + 2483 * sizeof(auxtrace_info_event::reserved) == 8 2484 */ 2485 info_header_size = 8; 2486 2487 if (total_size < (event_header_size + info_header_size)) 2488 return -EINVAL; 2489 2490 priv_size = total_size - event_header_size - info_header_size; 2491 2492 /* First the global part */ 2493 ptr = (u64 *) auxtrace_info->priv; 2494 2495 /* Look for version '0' of the header */ 2496 if (ptr[0] != 0) 2497 return -EINVAL; 2498 2499 hdr = zalloc(sizeof(*hdr) * CS_HEADER_VERSION_0_MAX); 2500 if (!hdr) 2501 return -ENOMEM; 2502 2503 /* Extract header information - see cs-etm.h for format */ 2504 for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++) 2505 hdr[i] = ptr[i]; 2506 num_cpu = hdr[CS_PMU_TYPE_CPUS] & 0xffffffff; 2507 pmu_type = (unsigned int) ((hdr[CS_PMU_TYPE_CPUS] >> 32) & 2508 0xffffffff); 2509 2510 /* 2511 * Create an RB tree for traceID-metadata tuple. Since the conversion 2512 * has to be made for each packet that gets decoded, optimizing access 2513 * in anything other than a sequential array is worth doing. 2514 */ 2515 traceid_list = intlist__new(NULL); 2516 if (!traceid_list) { 2517 err = -ENOMEM; 2518 goto err_free_hdr; 2519 } 2520 2521 metadata = zalloc(sizeof(*metadata) * num_cpu); 2522 if (!metadata) { 2523 err = -ENOMEM; 2524 goto err_free_traceid_list; 2525 } 2526 2527 /* 2528 * The metadata is stored in the auxtrace_info section and encodes 2529 * the configuration of the ARM embedded trace macrocell which is 2530 * required by the trace decoder to properly decode the trace due 2531 * to its highly compressed nature. 2532 */ 2533 for (j = 0; j < num_cpu; j++) { 2534 if (ptr[i] == __perf_cs_etmv3_magic) { 2535 metadata[j] = zalloc(sizeof(*metadata[j]) * 2536 CS_ETM_PRIV_MAX); 2537 if (!metadata[j]) { 2538 err = -ENOMEM; 2539 goto err_free_metadata; 2540 } 2541 for (k = 0; k < CS_ETM_PRIV_MAX; k++) 2542 metadata[j][k] = ptr[i + k]; 2543 2544 /* The traceID is our handle */ 2545 idx = metadata[j][CS_ETM_ETMTRACEIDR]; 2546 i += CS_ETM_PRIV_MAX; 2547 } else if (ptr[i] == __perf_cs_etmv4_magic) { 2548 metadata[j] = zalloc(sizeof(*metadata[j]) * 2549 CS_ETMV4_PRIV_MAX); 2550 if (!metadata[j]) { 2551 err = -ENOMEM; 2552 goto err_free_metadata; 2553 } 2554 for (k = 0; k < CS_ETMV4_PRIV_MAX; k++) 2555 metadata[j][k] = ptr[i + k]; 2556 2557 /* The traceID is our handle */ 2558 idx = metadata[j][CS_ETMV4_TRCTRACEIDR]; 2559 i += CS_ETMV4_PRIV_MAX; 2560 } 2561 2562 /* Get an RB node for this CPU */ 2563 inode = intlist__findnew(traceid_list, idx); 2564 2565 /* Something went wrong, no need to continue */ 2566 if (!inode) { 2567 err = -ENOMEM; 2568 goto err_free_metadata; 2569 } 2570 2571 /* 2572 * The node for that CPU should not be taken. 2573 * Back out if that's the case. 2574 */ 2575 if (inode->priv) { 2576 err = -EINVAL; 2577 goto err_free_metadata; 2578 } 2579 /* All good, associate the traceID with the metadata pointer */ 2580 inode->priv = metadata[j]; 2581 } 2582 2583 /* 2584 * Each of CS_HEADER_VERSION_0_MAX, CS_ETM_PRIV_MAX and 2585 * CS_ETMV4_PRIV_MAX mark how many double words are in the 2586 * global metadata, and each cpu's metadata respectively. 2587 * The following tests if the correct number of double words was 2588 * present in the auxtrace info section. 2589 */ 2590 if (i * 8 != priv_size) { 2591 err = -EINVAL; 2592 goto err_free_metadata; 2593 } 2594 2595 etm = zalloc(sizeof(*etm)); 2596 2597 if (!etm) { 2598 err = -ENOMEM; 2599 goto err_free_metadata; 2600 } 2601 2602 err = auxtrace_queues__init(&etm->queues); 2603 if (err) 2604 goto err_free_etm; 2605 2606 etm->session = session; 2607 etm->machine = &session->machines.host; 2608 2609 etm->num_cpu = num_cpu; 2610 etm->pmu_type = pmu_type; 2611 etm->snapshot_mode = (hdr[CS_ETM_SNAPSHOT] != 0); 2612 etm->metadata = metadata; 2613 etm->auxtrace_type = auxtrace_info->type; 2614 etm->timeless_decoding = cs_etm__is_timeless_decoding(etm); 2615 2616 etm->auxtrace.process_event = cs_etm__process_event; 2617 etm->auxtrace.process_auxtrace_event = cs_etm__process_auxtrace_event; 2618 etm->auxtrace.flush_events = cs_etm__flush_events; 2619 etm->auxtrace.free_events = cs_etm__free_events; 2620 etm->auxtrace.free = cs_etm__free; 2621 session->auxtrace = &etm->auxtrace; 2622 2623 etm->unknown_thread = thread__new(999999999, 999999999); 2624 if (!etm->unknown_thread) { 2625 err = -ENOMEM; 2626 goto err_free_queues; 2627 } 2628 2629 /* 2630 * Initialize list node so that at thread__zput() we can avoid 2631 * segmentation fault at list_del_init(). 2632 */ 2633 INIT_LIST_HEAD(&etm->unknown_thread->node); 2634 2635 err = thread__set_comm(etm->unknown_thread, "unknown", 0); 2636 if (err) 2637 goto err_delete_thread; 2638 2639 if (thread__init_maps(etm->unknown_thread, etm->machine)) { 2640 err = -ENOMEM; 2641 goto err_delete_thread; 2642 } 2643 2644 if (dump_trace) { 2645 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu); 2646 return 0; 2647 } 2648 2649 if (session->itrace_synth_opts->set) { 2650 etm->synth_opts = *session->itrace_synth_opts; 2651 } else { 2652 itrace_synth_opts__set_default(&etm->synth_opts, 2653 session->itrace_synth_opts->default_no_sample); 2654 etm->synth_opts.callchain = false; 2655 } 2656 2657 err = cs_etm__synth_events(etm, session); 2658 if (err) 2659 goto err_delete_thread; 2660 2661 err = auxtrace_queues__process_index(&etm->queues, session); 2662 if (err) 2663 goto err_delete_thread; 2664 2665 etm->data_queued = etm->queues.populated; 2666 2667 return 0; 2668 2669 err_delete_thread: 2670 thread__zput(etm->unknown_thread); 2671 err_free_queues: 2672 auxtrace_queues__free(&etm->queues); 2673 session->auxtrace = NULL; 2674 err_free_etm: 2675 zfree(&etm); 2676 err_free_metadata: 2677 /* No need to check @metadata[j], free(NULL) is supported */ 2678 for (j = 0; j < num_cpu; j++) 2679 zfree(&metadata[j]); 2680 zfree(&metadata); 2681 err_free_traceid_list: 2682 intlist__delete(traceid_list); 2683 err_free_hdr: 2684 zfree(&hdr); 2685 2686 return err; 2687 } 2688