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