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/coresight-pmu.h> 11 #include <linux/err.h> 12 #include <linux/kernel.h> 13 #include <linux/log2.h> 14 #include <linux/types.h> 15 #include <linux/zalloc.h> 16 17 #include <opencsd/ocsd_if_types.h> 18 #include <stdlib.h> 19 20 #include "auxtrace.h" 21 #include "color.h" 22 #include "cs-etm.h" 23 #include "cs-etm-decoder/cs-etm-decoder.h" 24 #include "debug.h" 25 #include "dso.h" 26 #include "evlist.h" 27 #include "intlist.h" 28 #include "machine.h" 29 #include "map.h" 30 #include "perf.h" 31 #include "session.h" 32 #include "map_symbol.h" 33 #include "branch.h" 34 #include "symbol.h" 35 #include "tool.h" 36 #include "thread.h" 37 #include "thread-stack.h" 38 #include <tools/libc_compat.h> 39 #include "util/synthetic-events.h" 40 41 struct cs_etm_auxtrace { 42 struct auxtrace auxtrace; 43 struct auxtrace_queues queues; 44 struct auxtrace_heap heap; 45 struct itrace_synth_opts synth_opts; 46 struct perf_session *session; 47 struct machine *machine; 48 struct thread *unknown_thread; 49 50 u8 timeless_decoding; 51 u8 snapshot_mode; 52 u8 data_queued; 53 u8 sample_branches; 54 u8 sample_instructions; 55 56 int num_cpu; 57 u64 latest_kernel_timestamp; 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 unsigned int pmu_type; 66 }; 67 68 struct cs_etm_traceid_queue { 69 u8 trace_chan_id; 70 pid_t pid, tid; 71 u64 period_instructions; 72 size_t last_branch_pos; 73 union perf_event *event_buf; 74 struct thread *thread; 75 struct branch_stack *last_branch; 76 struct branch_stack *last_branch_rb; 77 struct cs_etm_packet *prev_packet; 78 struct cs_etm_packet *packet; 79 struct cs_etm_packet_queue packet_queue; 80 }; 81 82 struct cs_etm_queue { 83 struct cs_etm_auxtrace *etm; 84 struct cs_etm_decoder *decoder; 85 struct auxtrace_buffer *buffer; 86 unsigned int queue_nr; 87 u8 pending_timestamp_chan_id; 88 u64 offset; 89 const unsigned char *buf; 90 size_t buf_len, buf_used; 91 /* Conversion between traceID and index in traceid_queues array */ 92 struct intlist *traceid_queues_list; 93 struct cs_etm_traceid_queue **traceid_queues; 94 }; 95 96 /* RB tree for quick conversion between traceID and metadata pointers */ 97 static struct intlist *traceid_list; 98 99 static int cs_etm__process_queues(struct cs_etm_auxtrace *etm); 100 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm, 101 pid_t tid); 102 static int cs_etm__get_data_block(struct cs_etm_queue *etmq); 103 static int cs_etm__decode_data_block(struct cs_etm_queue *etmq); 104 105 /* PTMs ETMIDR [11:8] set to b0011 */ 106 #define ETMIDR_PTM_VERSION 0x00000300 107 108 /* 109 * A struct auxtrace_heap_item only has a queue_nr and a timestamp to 110 * work with. One option is to modify to auxtrace_heap_XYZ() API or simply 111 * encode the etm queue number as the upper 16 bit and the channel as 112 * the lower 16 bit. 113 */ 114 #define TO_CS_QUEUE_NR(queue_nr, trace_chan_id) \ 115 (queue_nr << 16 | trace_chan_id) 116 #define TO_QUEUE_NR(cs_queue_nr) (cs_queue_nr >> 16) 117 #define TO_TRACE_CHAN_ID(cs_queue_nr) (cs_queue_nr & 0x0000ffff) 118 119 static u32 cs_etm__get_v7_protocol_version(u32 etmidr) 120 { 121 etmidr &= ETMIDR_PTM_VERSION; 122 123 if (etmidr == ETMIDR_PTM_VERSION) 124 return CS_ETM_PROTO_PTM; 125 126 return CS_ETM_PROTO_ETMV3; 127 } 128 129 static int cs_etm__get_magic(u8 trace_chan_id, u64 *magic) 130 { 131 struct int_node *inode; 132 u64 *metadata; 133 134 inode = intlist__find(traceid_list, trace_chan_id); 135 if (!inode) 136 return -EINVAL; 137 138 metadata = inode->priv; 139 *magic = metadata[CS_ETM_MAGIC]; 140 return 0; 141 } 142 143 int cs_etm__get_cpu(u8 trace_chan_id, int *cpu) 144 { 145 struct int_node *inode; 146 u64 *metadata; 147 148 inode = intlist__find(traceid_list, trace_chan_id); 149 if (!inode) 150 return -EINVAL; 151 152 metadata = inode->priv; 153 *cpu = (int)metadata[CS_ETM_CPU]; 154 return 0; 155 } 156 157 /* 158 * The returned PID format is presented by two bits: 159 * 160 * Bit ETM_OPT_CTXTID: CONTEXTIDR or CONTEXTIDR_EL1 is traced; 161 * Bit ETM_OPT_CTXTID2: CONTEXTIDR_EL2 is traced. 162 * 163 * It's possible that the two bits ETM_OPT_CTXTID and ETM_OPT_CTXTID2 164 * are enabled at the same time when the session runs on an EL2 kernel. 165 * This means the CONTEXTIDR_EL1 and CONTEXTIDR_EL2 both will be 166 * recorded in the trace data, the tool will selectively use 167 * CONTEXTIDR_EL2 as PID. 168 */ 169 int cs_etm__get_pid_fmt(u8 trace_chan_id, u64 *pid_fmt) 170 { 171 struct int_node *inode; 172 u64 *metadata, val; 173 174 inode = intlist__find(traceid_list, trace_chan_id); 175 if (!inode) 176 return -EINVAL; 177 178 metadata = inode->priv; 179 180 if (metadata[CS_ETM_MAGIC] == __perf_cs_etmv3_magic) { 181 val = metadata[CS_ETM_ETMCR]; 182 /* CONTEXTIDR is traced */ 183 if (val & BIT(ETM_OPT_CTXTID)) 184 *pid_fmt = BIT(ETM_OPT_CTXTID); 185 } else { 186 val = metadata[CS_ETMV4_TRCCONFIGR]; 187 /* CONTEXTIDR_EL2 is traced */ 188 if (val & (BIT(ETM4_CFG_BIT_VMID) | BIT(ETM4_CFG_BIT_VMID_OPT))) 189 *pid_fmt = BIT(ETM_OPT_CTXTID2); 190 /* CONTEXTIDR_EL1 is traced */ 191 else if (val & BIT(ETM4_CFG_BIT_CTXTID)) 192 *pid_fmt = BIT(ETM_OPT_CTXTID); 193 } 194 195 return 0; 196 } 197 198 void cs_etm__etmq_set_traceid_queue_timestamp(struct cs_etm_queue *etmq, 199 u8 trace_chan_id) 200 { 201 /* 202 * When a timestamp packet is encountered the backend code 203 * is stopped so that the front end has time to process packets 204 * that were accumulated in the traceID queue. Since there can 205 * be more than one channel per cs_etm_queue, we need to specify 206 * what traceID queue needs servicing. 207 */ 208 etmq->pending_timestamp_chan_id = trace_chan_id; 209 } 210 211 static u64 cs_etm__etmq_get_timestamp(struct cs_etm_queue *etmq, 212 u8 *trace_chan_id) 213 { 214 struct cs_etm_packet_queue *packet_queue; 215 216 if (!etmq->pending_timestamp_chan_id) 217 return 0; 218 219 if (trace_chan_id) 220 *trace_chan_id = etmq->pending_timestamp_chan_id; 221 222 packet_queue = cs_etm__etmq_get_packet_queue(etmq, 223 etmq->pending_timestamp_chan_id); 224 if (!packet_queue) 225 return 0; 226 227 /* Acknowledge pending status */ 228 etmq->pending_timestamp_chan_id = 0; 229 230 /* See function cs_etm_decoder__do_{hard|soft}_timestamp() */ 231 return packet_queue->cs_timestamp; 232 } 233 234 static void cs_etm__clear_packet_queue(struct cs_etm_packet_queue *queue) 235 { 236 int i; 237 238 queue->head = 0; 239 queue->tail = 0; 240 queue->packet_count = 0; 241 for (i = 0; i < CS_ETM_PACKET_MAX_BUFFER; i++) { 242 queue->packet_buffer[i].isa = CS_ETM_ISA_UNKNOWN; 243 queue->packet_buffer[i].start_addr = CS_ETM_INVAL_ADDR; 244 queue->packet_buffer[i].end_addr = CS_ETM_INVAL_ADDR; 245 queue->packet_buffer[i].instr_count = 0; 246 queue->packet_buffer[i].last_instr_taken_branch = false; 247 queue->packet_buffer[i].last_instr_size = 0; 248 queue->packet_buffer[i].last_instr_type = 0; 249 queue->packet_buffer[i].last_instr_subtype = 0; 250 queue->packet_buffer[i].last_instr_cond = 0; 251 queue->packet_buffer[i].flags = 0; 252 queue->packet_buffer[i].exception_number = UINT32_MAX; 253 queue->packet_buffer[i].trace_chan_id = UINT8_MAX; 254 queue->packet_buffer[i].cpu = INT_MIN; 255 } 256 } 257 258 static void cs_etm__clear_all_packet_queues(struct cs_etm_queue *etmq) 259 { 260 int idx; 261 struct int_node *inode; 262 struct cs_etm_traceid_queue *tidq; 263 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 264 265 intlist__for_each_entry(inode, traceid_queues_list) { 266 idx = (int)(intptr_t)inode->priv; 267 tidq = etmq->traceid_queues[idx]; 268 cs_etm__clear_packet_queue(&tidq->packet_queue); 269 } 270 } 271 272 static int cs_etm__init_traceid_queue(struct cs_etm_queue *etmq, 273 struct cs_etm_traceid_queue *tidq, 274 u8 trace_chan_id) 275 { 276 int rc = -ENOMEM; 277 struct auxtrace_queue *queue; 278 struct cs_etm_auxtrace *etm = etmq->etm; 279 280 cs_etm__clear_packet_queue(&tidq->packet_queue); 281 282 queue = &etmq->etm->queues.queue_array[etmq->queue_nr]; 283 tidq->tid = queue->tid; 284 tidq->pid = -1; 285 tidq->trace_chan_id = trace_chan_id; 286 287 tidq->packet = zalloc(sizeof(struct cs_etm_packet)); 288 if (!tidq->packet) 289 goto out; 290 291 tidq->prev_packet = zalloc(sizeof(struct cs_etm_packet)); 292 if (!tidq->prev_packet) 293 goto out_free; 294 295 if (etm->synth_opts.last_branch) { 296 size_t sz = sizeof(struct branch_stack); 297 298 sz += etm->synth_opts.last_branch_sz * 299 sizeof(struct branch_entry); 300 tidq->last_branch = zalloc(sz); 301 if (!tidq->last_branch) 302 goto out_free; 303 tidq->last_branch_rb = zalloc(sz); 304 if (!tidq->last_branch_rb) 305 goto out_free; 306 } 307 308 tidq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE); 309 if (!tidq->event_buf) 310 goto out_free; 311 312 return 0; 313 314 out_free: 315 zfree(&tidq->last_branch_rb); 316 zfree(&tidq->last_branch); 317 zfree(&tidq->prev_packet); 318 zfree(&tidq->packet); 319 out: 320 return rc; 321 } 322 323 static struct cs_etm_traceid_queue 324 *cs_etm__etmq_get_traceid_queue(struct cs_etm_queue *etmq, u8 trace_chan_id) 325 { 326 int idx; 327 struct int_node *inode; 328 struct intlist *traceid_queues_list; 329 struct cs_etm_traceid_queue *tidq, **traceid_queues; 330 struct cs_etm_auxtrace *etm = etmq->etm; 331 332 if (etm->timeless_decoding) 333 trace_chan_id = CS_ETM_PER_THREAD_TRACEID; 334 335 traceid_queues_list = etmq->traceid_queues_list; 336 337 /* 338 * Check if the traceid_queue exist for this traceID by looking 339 * in the queue list. 340 */ 341 inode = intlist__find(traceid_queues_list, trace_chan_id); 342 if (inode) { 343 idx = (int)(intptr_t)inode->priv; 344 return etmq->traceid_queues[idx]; 345 } 346 347 /* We couldn't find a traceid_queue for this traceID, allocate one */ 348 tidq = malloc(sizeof(*tidq)); 349 if (!tidq) 350 return NULL; 351 352 memset(tidq, 0, sizeof(*tidq)); 353 354 /* Get a valid index for the new traceid_queue */ 355 idx = intlist__nr_entries(traceid_queues_list); 356 /* Memory for the inode is free'ed in cs_etm_free_traceid_queues () */ 357 inode = intlist__findnew(traceid_queues_list, trace_chan_id); 358 if (!inode) 359 goto out_free; 360 361 /* Associate this traceID with this index */ 362 inode->priv = (void *)(intptr_t)idx; 363 364 if (cs_etm__init_traceid_queue(etmq, tidq, trace_chan_id)) 365 goto out_free; 366 367 /* Grow the traceid_queues array by one unit */ 368 traceid_queues = etmq->traceid_queues; 369 traceid_queues = reallocarray(traceid_queues, 370 idx + 1, 371 sizeof(*traceid_queues)); 372 373 /* 374 * On failure reallocarray() returns NULL and the original block of 375 * memory is left untouched. 376 */ 377 if (!traceid_queues) 378 goto out_free; 379 380 traceid_queues[idx] = tidq; 381 etmq->traceid_queues = traceid_queues; 382 383 return etmq->traceid_queues[idx]; 384 385 out_free: 386 /* 387 * Function intlist__remove() removes the inode from the list 388 * and delete the memory associated to it. 389 */ 390 intlist__remove(traceid_queues_list, inode); 391 free(tidq); 392 393 return NULL; 394 } 395 396 struct cs_etm_packet_queue 397 *cs_etm__etmq_get_packet_queue(struct cs_etm_queue *etmq, u8 trace_chan_id) 398 { 399 struct cs_etm_traceid_queue *tidq; 400 401 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 402 if (tidq) 403 return &tidq->packet_queue; 404 405 return NULL; 406 } 407 408 static void cs_etm__packet_swap(struct cs_etm_auxtrace *etm, 409 struct cs_etm_traceid_queue *tidq) 410 { 411 struct cs_etm_packet *tmp; 412 413 if (etm->sample_branches || etm->synth_opts.last_branch || 414 etm->sample_instructions) { 415 /* 416 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for 417 * the next incoming packet. 418 */ 419 tmp = tidq->packet; 420 tidq->packet = tidq->prev_packet; 421 tidq->prev_packet = tmp; 422 } 423 } 424 425 static void cs_etm__packet_dump(const char *pkt_string) 426 { 427 const char *color = PERF_COLOR_BLUE; 428 int len = strlen(pkt_string); 429 430 if (len && (pkt_string[len-1] == '\n')) 431 color_fprintf(stdout, color, " %s", pkt_string); 432 else 433 color_fprintf(stdout, color, " %s\n", pkt_string); 434 435 fflush(stdout); 436 } 437 438 static void cs_etm__set_trace_param_etmv3(struct cs_etm_trace_params *t_params, 439 struct cs_etm_auxtrace *etm, int idx, 440 u32 etmidr) 441 { 442 u64 **metadata = etm->metadata; 443 444 t_params[idx].protocol = cs_etm__get_v7_protocol_version(etmidr); 445 t_params[idx].etmv3.reg_ctrl = metadata[idx][CS_ETM_ETMCR]; 446 t_params[idx].etmv3.reg_trc_id = metadata[idx][CS_ETM_ETMTRACEIDR]; 447 } 448 449 static void cs_etm__set_trace_param_etmv4(struct cs_etm_trace_params *t_params, 450 struct cs_etm_auxtrace *etm, int idx) 451 { 452 u64 **metadata = etm->metadata; 453 454 t_params[idx].protocol = CS_ETM_PROTO_ETMV4i; 455 t_params[idx].etmv4.reg_idr0 = metadata[idx][CS_ETMV4_TRCIDR0]; 456 t_params[idx].etmv4.reg_idr1 = metadata[idx][CS_ETMV4_TRCIDR1]; 457 t_params[idx].etmv4.reg_idr2 = metadata[idx][CS_ETMV4_TRCIDR2]; 458 t_params[idx].etmv4.reg_idr8 = metadata[idx][CS_ETMV4_TRCIDR8]; 459 t_params[idx].etmv4.reg_configr = metadata[idx][CS_ETMV4_TRCCONFIGR]; 460 t_params[idx].etmv4.reg_traceidr = metadata[idx][CS_ETMV4_TRCTRACEIDR]; 461 } 462 463 static int cs_etm__init_trace_params(struct cs_etm_trace_params *t_params, 464 struct cs_etm_auxtrace *etm, 465 int decoders) 466 { 467 int i; 468 u32 etmidr; 469 u64 architecture; 470 471 for (i = 0; i < decoders; i++) { 472 architecture = etm->metadata[i][CS_ETM_MAGIC]; 473 474 switch (architecture) { 475 case __perf_cs_etmv3_magic: 476 etmidr = etm->metadata[i][CS_ETM_ETMIDR]; 477 cs_etm__set_trace_param_etmv3(t_params, etm, i, etmidr); 478 break; 479 case __perf_cs_etmv4_magic: 480 cs_etm__set_trace_param_etmv4(t_params, etm, i); 481 break; 482 default: 483 return -EINVAL; 484 } 485 } 486 487 return 0; 488 } 489 490 static int cs_etm__init_decoder_params(struct cs_etm_decoder_params *d_params, 491 struct cs_etm_queue *etmq, 492 enum cs_etm_decoder_operation mode, 493 bool formatted) 494 { 495 int ret = -EINVAL; 496 497 if (!(mode < CS_ETM_OPERATION_MAX)) 498 goto out; 499 500 d_params->packet_printer = cs_etm__packet_dump; 501 d_params->operation = mode; 502 d_params->data = etmq; 503 d_params->formatted = formatted; 504 d_params->fsyncs = false; 505 d_params->hsyncs = false; 506 d_params->frame_aligned = true; 507 508 ret = 0; 509 out: 510 return ret; 511 } 512 513 static void cs_etm__dump_event(struct cs_etm_queue *etmq, 514 struct auxtrace_buffer *buffer) 515 { 516 int ret; 517 const char *color = PERF_COLOR_BLUE; 518 size_t buffer_used = 0; 519 520 fprintf(stdout, "\n"); 521 color_fprintf(stdout, color, 522 ". ... CoreSight ETM Trace data: size %zu bytes\n", 523 buffer->size); 524 525 do { 526 size_t consumed; 527 528 ret = cs_etm_decoder__process_data_block( 529 etmq->decoder, buffer->offset, 530 &((u8 *)buffer->data)[buffer_used], 531 buffer->size - buffer_used, &consumed); 532 if (ret) 533 break; 534 535 buffer_used += consumed; 536 } while (buffer_used < buffer->size); 537 538 cs_etm_decoder__reset(etmq->decoder); 539 } 540 541 static int cs_etm__flush_events(struct perf_session *session, 542 struct perf_tool *tool) 543 { 544 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 545 struct cs_etm_auxtrace, 546 auxtrace); 547 if (dump_trace) 548 return 0; 549 550 if (!tool->ordered_events) 551 return -EINVAL; 552 553 if (etm->timeless_decoding) 554 return cs_etm__process_timeless_queues(etm, -1); 555 556 return cs_etm__process_queues(etm); 557 } 558 559 static void cs_etm__free_traceid_queues(struct cs_etm_queue *etmq) 560 { 561 int idx; 562 uintptr_t priv; 563 struct int_node *inode, *tmp; 564 struct cs_etm_traceid_queue *tidq; 565 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 566 567 intlist__for_each_entry_safe(inode, tmp, traceid_queues_list) { 568 priv = (uintptr_t)inode->priv; 569 idx = priv; 570 571 /* Free this traceid_queue from the array */ 572 tidq = etmq->traceid_queues[idx]; 573 thread__zput(tidq->thread); 574 zfree(&tidq->event_buf); 575 zfree(&tidq->last_branch); 576 zfree(&tidq->last_branch_rb); 577 zfree(&tidq->prev_packet); 578 zfree(&tidq->packet); 579 zfree(&tidq); 580 581 /* 582 * Function intlist__remove() removes the inode from the list 583 * and delete the memory associated to it. 584 */ 585 intlist__remove(traceid_queues_list, inode); 586 } 587 588 /* Then the RB tree itself */ 589 intlist__delete(traceid_queues_list); 590 etmq->traceid_queues_list = NULL; 591 592 /* finally free the traceid_queues array */ 593 zfree(&etmq->traceid_queues); 594 } 595 596 static void cs_etm__free_queue(void *priv) 597 { 598 struct cs_etm_queue *etmq = priv; 599 600 if (!etmq) 601 return; 602 603 cs_etm_decoder__free(etmq->decoder); 604 cs_etm__free_traceid_queues(etmq); 605 free(etmq); 606 } 607 608 static void cs_etm__free_events(struct perf_session *session) 609 { 610 unsigned int i; 611 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 612 struct cs_etm_auxtrace, 613 auxtrace); 614 struct auxtrace_queues *queues = &aux->queues; 615 616 for (i = 0; i < queues->nr_queues; i++) { 617 cs_etm__free_queue(queues->queue_array[i].priv); 618 queues->queue_array[i].priv = NULL; 619 } 620 621 auxtrace_queues__free(queues); 622 } 623 624 static void cs_etm__free(struct perf_session *session) 625 { 626 int i; 627 struct int_node *inode, *tmp; 628 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 629 struct cs_etm_auxtrace, 630 auxtrace); 631 cs_etm__free_events(session); 632 session->auxtrace = NULL; 633 634 /* First remove all traceID/metadata nodes for the RB tree */ 635 intlist__for_each_entry_safe(inode, tmp, traceid_list) 636 intlist__remove(traceid_list, inode); 637 /* Then the RB tree itself */ 638 intlist__delete(traceid_list); 639 640 for (i = 0; i < aux->num_cpu; i++) 641 zfree(&aux->metadata[i]); 642 643 thread__zput(aux->unknown_thread); 644 zfree(&aux->metadata); 645 zfree(&aux); 646 } 647 648 static bool cs_etm__evsel_is_auxtrace(struct perf_session *session, 649 struct evsel *evsel) 650 { 651 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 652 struct cs_etm_auxtrace, 653 auxtrace); 654 655 return evsel->core.attr.type == aux->pmu_type; 656 } 657 658 static u8 cs_etm__cpu_mode(struct cs_etm_queue *etmq, u64 address) 659 { 660 struct machine *machine; 661 662 machine = etmq->etm->machine; 663 664 if (address >= machine__kernel_start(machine)) { 665 if (machine__is_host(machine)) 666 return PERF_RECORD_MISC_KERNEL; 667 else 668 return PERF_RECORD_MISC_GUEST_KERNEL; 669 } else { 670 if (machine__is_host(machine)) 671 return PERF_RECORD_MISC_USER; 672 else if (perf_guest) 673 return PERF_RECORD_MISC_GUEST_USER; 674 else 675 return PERF_RECORD_MISC_HYPERVISOR; 676 } 677 } 678 679 static u32 cs_etm__mem_access(struct cs_etm_queue *etmq, u8 trace_chan_id, 680 u64 address, size_t size, u8 *buffer) 681 { 682 u8 cpumode; 683 u64 offset; 684 int len; 685 struct thread *thread; 686 struct machine *machine; 687 struct addr_location al; 688 struct cs_etm_traceid_queue *tidq; 689 690 if (!etmq) 691 return 0; 692 693 machine = etmq->etm->machine; 694 cpumode = cs_etm__cpu_mode(etmq, address); 695 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 696 if (!tidq) 697 return 0; 698 699 thread = tidq->thread; 700 if (!thread) { 701 if (cpumode != PERF_RECORD_MISC_KERNEL) 702 return 0; 703 thread = etmq->etm->unknown_thread; 704 } 705 706 if (!thread__find_map(thread, cpumode, address, &al) || !al.map->dso) 707 return 0; 708 709 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR && 710 dso__data_status_seen(al.map->dso, DSO_DATA_STATUS_SEEN_ITRACE)) 711 return 0; 712 713 offset = al.map->map_ip(al.map, address); 714 715 map__load(al.map); 716 717 len = dso__data_read_offset(al.map->dso, machine, offset, buffer, size); 718 719 if (len <= 0) 720 return 0; 721 722 return len; 723 } 724 725 static struct cs_etm_queue *cs_etm__alloc_queue(struct cs_etm_auxtrace *etm, 726 bool formatted) 727 { 728 struct cs_etm_decoder_params d_params; 729 struct cs_etm_trace_params *t_params = NULL; 730 struct cs_etm_queue *etmq; 731 /* 732 * Each queue can only contain data from one CPU when unformatted, so only one decoder is 733 * needed. 734 */ 735 int decoders = formatted ? etm->num_cpu : 1; 736 737 etmq = zalloc(sizeof(*etmq)); 738 if (!etmq) 739 return NULL; 740 741 etmq->traceid_queues_list = intlist__new(NULL); 742 if (!etmq->traceid_queues_list) 743 goto out_free; 744 745 /* Use metadata to fill in trace parameters for trace decoder */ 746 t_params = zalloc(sizeof(*t_params) * decoders); 747 748 if (!t_params) 749 goto out_free; 750 751 if (cs_etm__init_trace_params(t_params, etm, decoders)) 752 goto out_free; 753 754 /* Set decoder parameters to decode trace packets */ 755 if (cs_etm__init_decoder_params(&d_params, etmq, 756 dump_trace ? CS_ETM_OPERATION_PRINT : 757 CS_ETM_OPERATION_DECODE, 758 formatted)) 759 goto out_free; 760 761 etmq->decoder = cs_etm_decoder__new(decoders, &d_params, 762 t_params); 763 764 if (!etmq->decoder) 765 goto out_free; 766 767 /* 768 * Register a function to handle all memory accesses required by 769 * the trace decoder library. 770 */ 771 if (cs_etm_decoder__add_mem_access_cb(etmq->decoder, 772 0x0L, ((u64) -1L), 773 cs_etm__mem_access)) 774 goto out_free_decoder; 775 776 zfree(&t_params); 777 return etmq; 778 779 out_free_decoder: 780 cs_etm_decoder__free(etmq->decoder); 781 out_free: 782 intlist__delete(etmq->traceid_queues_list); 783 free(etmq); 784 785 return NULL; 786 } 787 788 static int cs_etm__setup_queue(struct cs_etm_auxtrace *etm, 789 struct auxtrace_queue *queue, 790 unsigned int queue_nr, 791 bool formatted) 792 { 793 struct cs_etm_queue *etmq = queue->priv; 794 795 if (list_empty(&queue->head) || etmq) 796 return 0; 797 798 etmq = cs_etm__alloc_queue(etm, formatted); 799 800 if (!etmq) 801 return -ENOMEM; 802 803 queue->priv = etmq; 804 etmq->etm = etm; 805 etmq->queue_nr = queue_nr; 806 etmq->offset = 0; 807 808 return 0; 809 } 810 811 static int cs_etm__queue_first_cs_timestamp(struct cs_etm_auxtrace *etm, 812 struct cs_etm_queue *etmq, 813 unsigned int queue_nr) 814 { 815 int ret = 0; 816 unsigned int cs_queue_nr; 817 u8 trace_chan_id; 818 u64 cs_timestamp; 819 820 /* 821 * We are under a CPU-wide trace scenario. As such we need to know 822 * when the code that generated the traces started to execute so that 823 * it can be correlated with execution on other CPUs. So we get a 824 * handle on the beginning of traces and decode until we find a 825 * timestamp. The timestamp is then added to the auxtrace min heap 826 * in order to know what nibble (of all the etmqs) to decode first. 827 */ 828 while (1) { 829 /* 830 * Fetch an aux_buffer from this etmq. Bail if no more 831 * blocks or an error has been encountered. 832 */ 833 ret = cs_etm__get_data_block(etmq); 834 if (ret <= 0) 835 goto out; 836 837 /* 838 * Run decoder on the trace block. The decoder will stop when 839 * encountering a CS timestamp, a full packet queue or the end of 840 * trace for that block. 841 */ 842 ret = cs_etm__decode_data_block(etmq); 843 if (ret) 844 goto out; 845 846 /* 847 * Function cs_etm_decoder__do_{hard|soft}_timestamp() does all 848 * the timestamp calculation for us. 849 */ 850 cs_timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id); 851 852 /* We found a timestamp, no need to continue. */ 853 if (cs_timestamp) 854 break; 855 856 /* 857 * We didn't find a timestamp so empty all the traceid packet 858 * queues before looking for another timestamp packet, either 859 * in the current data block or a new one. Packets that were 860 * just decoded are useless since no timestamp has been 861 * associated with them. As such simply discard them. 862 */ 863 cs_etm__clear_all_packet_queues(etmq); 864 } 865 866 /* 867 * We have a timestamp. Add it to the min heap to reflect when 868 * instructions conveyed by the range packets of this traceID queue 869 * started to execute. Once the same has been done for all the traceID 870 * queues of each etmq, redenring and decoding can start in 871 * chronological order. 872 * 873 * Note that packets decoded above are still in the traceID's packet 874 * queue and will be processed in cs_etm__process_queues(). 875 */ 876 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id); 877 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, cs_timestamp); 878 out: 879 return ret; 880 } 881 882 static inline 883 void cs_etm__copy_last_branch_rb(struct cs_etm_queue *etmq, 884 struct cs_etm_traceid_queue *tidq) 885 { 886 struct branch_stack *bs_src = tidq->last_branch_rb; 887 struct branch_stack *bs_dst = tidq->last_branch; 888 size_t nr = 0; 889 890 /* 891 * Set the number of records before early exit: ->nr is used to 892 * determine how many branches to copy from ->entries. 893 */ 894 bs_dst->nr = bs_src->nr; 895 896 /* 897 * Early exit when there is nothing to copy. 898 */ 899 if (!bs_src->nr) 900 return; 901 902 /* 903 * As bs_src->entries is a circular buffer, we need to copy from it in 904 * two steps. First, copy the branches from the most recently inserted 905 * branch ->last_branch_pos until the end of bs_src->entries buffer. 906 */ 907 nr = etmq->etm->synth_opts.last_branch_sz - tidq->last_branch_pos; 908 memcpy(&bs_dst->entries[0], 909 &bs_src->entries[tidq->last_branch_pos], 910 sizeof(struct branch_entry) * nr); 911 912 /* 913 * If we wrapped around at least once, the branches from the beginning 914 * of the bs_src->entries buffer and until the ->last_branch_pos element 915 * are older valid branches: copy them over. The total number of 916 * branches copied over will be equal to the number of branches asked by 917 * the user in last_branch_sz. 918 */ 919 if (bs_src->nr >= etmq->etm->synth_opts.last_branch_sz) { 920 memcpy(&bs_dst->entries[nr], 921 &bs_src->entries[0], 922 sizeof(struct branch_entry) * tidq->last_branch_pos); 923 } 924 } 925 926 static inline 927 void cs_etm__reset_last_branch_rb(struct cs_etm_traceid_queue *tidq) 928 { 929 tidq->last_branch_pos = 0; 930 tidq->last_branch_rb->nr = 0; 931 } 932 933 static inline int cs_etm__t32_instr_size(struct cs_etm_queue *etmq, 934 u8 trace_chan_id, u64 addr) 935 { 936 u8 instrBytes[2]; 937 938 cs_etm__mem_access(etmq, trace_chan_id, addr, 939 ARRAY_SIZE(instrBytes), instrBytes); 940 /* 941 * T32 instruction size is indicated by bits[15:11] of the first 942 * 16-bit word of the instruction: 0b11101, 0b11110 and 0b11111 943 * denote a 32-bit instruction. 944 */ 945 return ((instrBytes[1] & 0xF8) >= 0xE8) ? 4 : 2; 946 } 947 948 static inline u64 cs_etm__first_executed_instr(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->start_addr; 955 } 956 957 static inline 958 u64 cs_etm__last_executed_instr(const struct cs_etm_packet *packet) 959 { 960 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */ 961 if (packet->sample_type == CS_ETM_DISCONTINUITY) 962 return 0; 963 964 return packet->end_addr - packet->last_instr_size; 965 } 966 967 static inline u64 cs_etm__instr_addr(struct cs_etm_queue *etmq, 968 u64 trace_chan_id, 969 const struct cs_etm_packet *packet, 970 u64 offset) 971 { 972 if (packet->isa == CS_ETM_ISA_T32) { 973 u64 addr = packet->start_addr; 974 975 while (offset) { 976 addr += cs_etm__t32_instr_size(etmq, 977 trace_chan_id, addr); 978 offset--; 979 } 980 return addr; 981 } 982 983 /* Assume a 4 byte instruction size (A32/A64) */ 984 return packet->start_addr + offset * 4; 985 } 986 987 static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq, 988 struct cs_etm_traceid_queue *tidq) 989 { 990 struct branch_stack *bs = tidq->last_branch_rb; 991 struct branch_entry *be; 992 993 /* 994 * The branches are recorded in a circular buffer in reverse 995 * chronological order: we start recording from the last element of the 996 * buffer down. After writing the first element of the stack, move the 997 * insert position back to the end of the buffer. 998 */ 999 if (!tidq->last_branch_pos) 1000 tidq->last_branch_pos = etmq->etm->synth_opts.last_branch_sz; 1001 1002 tidq->last_branch_pos -= 1; 1003 1004 be = &bs->entries[tidq->last_branch_pos]; 1005 be->from = cs_etm__last_executed_instr(tidq->prev_packet); 1006 be->to = cs_etm__first_executed_instr(tidq->packet); 1007 /* No support for mispredict */ 1008 be->flags.mispred = 0; 1009 be->flags.predicted = 1; 1010 1011 /* 1012 * Increment bs->nr until reaching the number of last branches asked by 1013 * the user on the command line. 1014 */ 1015 if (bs->nr < etmq->etm->synth_opts.last_branch_sz) 1016 bs->nr += 1; 1017 } 1018 1019 static int cs_etm__inject_event(union perf_event *event, 1020 struct perf_sample *sample, u64 type) 1021 { 1022 event->header.size = perf_event__sample_event_size(sample, type, 0); 1023 return perf_event__synthesize_sample(event, type, 0, sample); 1024 } 1025 1026 1027 static int 1028 cs_etm__get_trace(struct cs_etm_queue *etmq) 1029 { 1030 struct auxtrace_buffer *aux_buffer = etmq->buffer; 1031 struct auxtrace_buffer *old_buffer = aux_buffer; 1032 struct auxtrace_queue *queue; 1033 1034 queue = &etmq->etm->queues.queue_array[etmq->queue_nr]; 1035 1036 aux_buffer = auxtrace_buffer__next(queue, aux_buffer); 1037 1038 /* If no more data, drop the previous auxtrace_buffer and return */ 1039 if (!aux_buffer) { 1040 if (old_buffer) 1041 auxtrace_buffer__drop_data(old_buffer); 1042 etmq->buf_len = 0; 1043 return 0; 1044 } 1045 1046 etmq->buffer = aux_buffer; 1047 1048 /* If the aux_buffer doesn't have data associated, try to load it */ 1049 if (!aux_buffer->data) { 1050 /* get the file desc associated with the perf data file */ 1051 int fd = perf_data__fd(etmq->etm->session->data); 1052 1053 aux_buffer->data = auxtrace_buffer__get_data(aux_buffer, fd); 1054 if (!aux_buffer->data) 1055 return -ENOMEM; 1056 } 1057 1058 /* If valid, drop the previous buffer */ 1059 if (old_buffer) 1060 auxtrace_buffer__drop_data(old_buffer); 1061 1062 etmq->buf_used = 0; 1063 etmq->buf_len = aux_buffer->size; 1064 etmq->buf = aux_buffer->data; 1065 1066 return etmq->buf_len; 1067 } 1068 1069 static void cs_etm__set_pid_tid_cpu(struct cs_etm_auxtrace *etm, 1070 struct cs_etm_traceid_queue *tidq) 1071 { 1072 if ((!tidq->thread) && (tidq->tid != -1)) 1073 tidq->thread = machine__find_thread(etm->machine, -1, 1074 tidq->tid); 1075 1076 if (tidq->thread) 1077 tidq->pid = tidq->thread->pid_; 1078 } 1079 1080 int cs_etm__etmq_set_tid(struct cs_etm_queue *etmq, 1081 pid_t tid, u8 trace_chan_id) 1082 { 1083 int cpu, err = -EINVAL; 1084 struct cs_etm_auxtrace *etm = etmq->etm; 1085 struct cs_etm_traceid_queue *tidq; 1086 1087 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 1088 if (!tidq) 1089 return err; 1090 1091 if (cs_etm__get_cpu(trace_chan_id, &cpu) < 0) 1092 return err; 1093 1094 err = machine__set_current_tid(etm->machine, cpu, tid, tid); 1095 if (err) 1096 return err; 1097 1098 tidq->tid = tid; 1099 thread__zput(tidq->thread); 1100 1101 cs_etm__set_pid_tid_cpu(etm, tidq); 1102 return 0; 1103 } 1104 1105 bool cs_etm__etmq_is_timeless(struct cs_etm_queue *etmq) 1106 { 1107 return !!etmq->etm->timeless_decoding; 1108 } 1109 1110 static void cs_etm__copy_insn(struct cs_etm_queue *etmq, 1111 u64 trace_chan_id, 1112 const struct cs_etm_packet *packet, 1113 struct perf_sample *sample) 1114 { 1115 /* 1116 * It's pointless to read instructions for the CS_ETM_DISCONTINUITY 1117 * packet, so directly bail out with 'insn_len' = 0. 1118 */ 1119 if (packet->sample_type == CS_ETM_DISCONTINUITY) { 1120 sample->insn_len = 0; 1121 return; 1122 } 1123 1124 /* 1125 * T32 instruction size might be 32-bit or 16-bit, decide by calling 1126 * cs_etm__t32_instr_size(). 1127 */ 1128 if (packet->isa == CS_ETM_ISA_T32) 1129 sample->insn_len = cs_etm__t32_instr_size(etmq, trace_chan_id, 1130 sample->ip); 1131 /* Otherwise, A64 and A32 instruction size are always 32-bit. */ 1132 else 1133 sample->insn_len = 4; 1134 1135 cs_etm__mem_access(etmq, trace_chan_id, sample->ip, 1136 sample->insn_len, (void *)sample->insn); 1137 } 1138 1139 static int cs_etm__synth_instruction_sample(struct cs_etm_queue *etmq, 1140 struct cs_etm_traceid_queue *tidq, 1141 u64 addr, u64 period) 1142 { 1143 int ret = 0; 1144 struct cs_etm_auxtrace *etm = etmq->etm; 1145 union perf_event *event = tidq->event_buf; 1146 struct perf_sample sample = {.ip = 0,}; 1147 1148 event->sample.header.type = PERF_RECORD_SAMPLE; 1149 event->sample.header.misc = cs_etm__cpu_mode(etmq, addr); 1150 event->sample.header.size = sizeof(struct perf_event_header); 1151 1152 if (!etm->timeless_decoding) 1153 sample.time = etm->latest_kernel_timestamp; 1154 sample.ip = addr; 1155 sample.pid = tidq->pid; 1156 sample.tid = tidq->tid; 1157 sample.id = etmq->etm->instructions_id; 1158 sample.stream_id = etmq->etm->instructions_id; 1159 sample.period = period; 1160 sample.cpu = tidq->packet->cpu; 1161 sample.flags = tidq->prev_packet->flags; 1162 sample.cpumode = event->sample.header.misc; 1163 1164 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->packet, &sample); 1165 1166 if (etm->synth_opts.last_branch) 1167 sample.branch_stack = tidq->last_branch; 1168 1169 if (etm->synth_opts.inject) { 1170 ret = cs_etm__inject_event(event, &sample, 1171 etm->instructions_sample_type); 1172 if (ret) 1173 return ret; 1174 } 1175 1176 ret = perf_session__deliver_synth_event(etm->session, event, &sample); 1177 1178 if (ret) 1179 pr_err( 1180 "CS ETM Trace: failed to deliver instruction event, error %d\n", 1181 ret); 1182 1183 return ret; 1184 } 1185 1186 /* 1187 * The cs etm packet encodes an instruction range between a branch target 1188 * and the next taken branch. Generate sample accordingly. 1189 */ 1190 static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq, 1191 struct cs_etm_traceid_queue *tidq) 1192 { 1193 int ret = 0; 1194 struct cs_etm_auxtrace *etm = etmq->etm; 1195 struct perf_sample sample = {.ip = 0,}; 1196 union perf_event *event = tidq->event_buf; 1197 struct dummy_branch_stack { 1198 u64 nr; 1199 u64 hw_idx; 1200 struct branch_entry entries; 1201 } dummy_bs; 1202 u64 ip; 1203 1204 ip = cs_etm__last_executed_instr(tidq->prev_packet); 1205 1206 event->sample.header.type = PERF_RECORD_SAMPLE; 1207 event->sample.header.misc = cs_etm__cpu_mode(etmq, ip); 1208 event->sample.header.size = sizeof(struct perf_event_header); 1209 1210 if (!etm->timeless_decoding) 1211 sample.time = etm->latest_kernel_timestamp; 1212 sample.ip = ip; 1213 sample.pid = tidq->pid; 1214 sample.tid = tidq->tid; 1215 sample.addr = cs_etm__first_executed_instr(tidq->packet); 1216 sample.id = etmq->etm->branches_id; 1217 sample.stream_id = etmq->etm->branches_id; 1218 sample.period = 1; 1219 sample.cpu = tidq->packet->cpu; 1220 sample.flags = tidq->prev_packet->flags; 1221 sample.cpumode = event->sample.header.misc; 1222 1223 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->prev_packet, 1224 &sample); 1225 1226 /* 1227 * perf report cannot handle events without a branch stack 1228 */ 1229 if (etm->synth_opts.last_branch) { 1230 dummy_bs = (struct dummy_branch_stack){ 1231 .nr = 1, 1232 .hw_idx = -1ULL, 1233 .entries = { 1234 .from = sample.ip, 1235 .to = sample.addr, 1236 }, 1237 }; 1238 sample.branch_stack = (struct branch_stack *)&dummy_bs; 1239 } 1240 1241 if (etm->synth_opts.inject) { 1242 ret = cs_etm__inject_event(event, &sample, 1243 etm->branches_sample_type); 1244 if (ret) 1245 return ret; 1246 } 1247 1248 ret = perf_session__deliver_synth_event(etm->session, event, &sample); 1249 1250 if (ret) 1251 pr_err( 1252 "CS ETM Trace: failed to deliver instruction event, error %d\n", 1253 ret); 1254 1255 return ret; 1256 } 1257 1258 struct cs_etm_synth { 1259 struct perf_tool dummy_tool; 1260 struct perf_session *session; 1261 }; 1262 1263 static int cs_etm__event_synth(struct perf_tool *tool, 1264 union perf_event *event, 1265 struct perf_sample *sample __maybe_unused, 1266 struct machine *machine __maybe_unused) 1267 { 1268 struct cs_etm_synth *cs_etm_synth = 1269 container_of(tool, struct cs_etm_synth, dummy_tool); 1270 1271 return perf_session__deliver_synth_event(cs_etm_synth->session, 1272 event, NULL); 1273 } 1274 1275 static int cs_etm__synth_event(struct perf_session *session, 1276 struct perf_event_attr *attr, u64 id) 1277 { 1278 struct cs_etm_synth cs_etm_synth; 1279 1280 memset(&cs_etm_synth, 0, sizeof(struct cs_etm_synth)); 1281 cs_etm_synth.session = session; 1282 1283 return perf_event__synthesize_attr(&cs_etm_synth.dummy_tool, attr, 1, 1284 &id, cs_etm__event_synth); 1285 } 1286 1287 static int cs_etm__synth_events(struct cs_etm_auxtrace *etm, 1288 struct perf_session *session) 1289 { 1290 struct evlist *evlist = session->evlist; 1291 struct evsel *evsel; 1292 struct perf_event_attr attr; 1293 bool found = false; 1294 u64 id; 1295 int err; 1296 1297 evlist__for_each_entry(evlist, evsel) { 1298 if (evsel->core.attr.type == etm->pmu_type) { 1299 found = true; 1300 break; 1301 } 1302 } 1303 1304 if (!found) { 1305 pr_debug("No selected events with CoreSight Trace data\n"); 1306 return 0; 1307 } 1308 1309 memset(&attr, 0, sizeof(struct perf_event_attr)); 1310 attr.size = sizeof(struct perf_event_attr); 1311 attr.type = PERF_TYPE_HARDWARE; 1312 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK; 1313 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID | 1314 PERF_SAMPLE_PERIOD; 1315 if (etm->timeless_decoding) 1316 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME; 1317 else 1318 attr.sample_type |= PERF_SAMPLE_TIME; 1319 1320 attr.exclude_user = evsel->core.attr.exclude_user; 1321 attr.exclude_kernel = evsel->core.attr.exclude_kernel; 1322 attr.exclude_hv = evsel->core.attr.exclude_hv; 1323 attr.exclude_host = evsel->core.attr.exclude_host; 1324 attr.exclude_guest = evsel->core.attr.exclude_guest; 1325 attr.sample_id_all = evsel->core.attr.sample_id_all; 1326 attr.read_format = evsel->core.attr.read_format; 1327 1328 /* create new id val to be a fixed offset from evsel id */ 1329 id = evsel->core.id[0] + 1000000000; 1330 1331 if (!id) 1332 id = 1; 1333 1334 if (etm->synth_opts.branches) { 1335 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; 1336 attr.sample_period = 1; 1337 attr.sample_type |= PERF_SAMPLE_ADDR; 1338 err = cs_etm__synth_event(session, &attr, id); 1339 if (err) 1340 return err; 1341 etm->sample_branches = true; 1342 etm->branches_sample_type = attr.sample_type; 1343 etm->branches_id = id; 1344 id += 1; 1345 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR; 1346 } 1347 1348 if (etm->synth_opts.last_branch) { 1349 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK; 1350 /* 1351 * We don't use the hardware index, but the sample generation 1352 * code uses the new format branch_stack with this field, 1353 * so the event attributes must indicate that it's present. 1354 */ 1355 attr.branch_sample_type |= PERF_SAMPLE_BRANCH_HW_INDEX; 1356 } 1357 1358 if (etm->synth_opts.instructions) { 1359 attr.config = PERF_COUNT_HW_INSTRUCTIONS; 1360 attr.sample_period = etm->synth_opts.period; 1361 etm->instructions_sample_period = attr.sample_period; 1362 err = cs_etm__synth_event(session, &attr, id); 1363 if (err) 1364 return err; 1365 etm->sample_instructions = true; 1366 etm->instructions_sample_type = attr.sample_type; 1367 etm->instructions_id = id; 1368 id += 1; 1369 } 1370 1371 return 0; 1372 } 1373 1374 static int cs_etm__sample(struct cs_etm_queue *etmq, 1375 struct cs_etm_traceid_queue *tidq) 1376 { 1377 struct cs_etm_auxtrace *etm = etmq->etm; 1378 int ret; 1379 u8 trace_chan_id = tidq->trace_chan_id; 1380 u64 instrs_prev; 1381 1382 /* Get instructions remainder from previous packet */ 1383 instrs_prev = tidq->period_instructions; 1384 1385 tidq->period_instructions += tidq->packet->instr_count; 1386 1387 /* 1388 * Record a branch when the last instruction in 1389 * PREV_PACKET is a branch. 1390 */ 1391 if (etm->synth_opts.last_branch && 1392 tidq->prev_packet->sample_type == CS_ETM_RANGE && 1393 tidq->prev_packet->last_instr_taken_branch) 1394 cs_etm__update_last_branch_rb(etmq, tidq); 1395 1396 if (etm->sample_instructions && 1397 tidq->period_instructions >= etm->instructions_sample_period) { 1398 /* 1399 * Emit instruction sample periodically 1400 * TODO: allow period to be defined in cycles and clock time 1401 */ 1402 1403 /* 1404 * Below diagram demonstrates the instruction samples 1405 * generation flows: 1406 * 1407 * Instrs Instrs Instrs Instrs 1408 * Sample(n) Sample(n+1) Sample(n+2) Sample(n+3) 1409 * | | | | 1410 * V V V V 1411 * -------------------------------------------------- 1412 * ^ ^ 1413 * | | 1414 * Period Period 1415 * instructions(Pi) instructions(Pi') 1416 * 1417 * | | 1418 * \---------------- -----------------/ 1419 * V 1420 * tidq->packet->instr_count 1421 * 1422 * Instrs Sample(n...) are the synthesised samples occurring 1423 * every etm->instructions_sample_period instructions - as 1424 * defined on the perf command line. Sample(n) is being the 1425 * last sample before the current etm packet, n+1 to n+3 1426 * samples are generated from the current etm packet. 1427 * 1428 * tidq->packet->instr_count represents the number of 1429 * instructions in the current etm packet. 1430 * 1431 * Period instructions (Pi) contains the the number of 1432 * instructions executed after the sample point(n) from the 1433 * previous etm packet. This will always be less than 1434 * etm->instructions_sample_period. 1435 * 1436 * When generate new samples, it combines with two parts 1437 * instructions, one is the tail of the old packet and another 1438 * is the head of the new coming packet, to generate 1439 * sample(n+1); sample(n+2) and sample(n+3) consume the 1440 * instructions with sample period. After sample(n+3), the rest 1441 * instructions will be used by later packet and it is assigned 1442 * to tidq->period_instructions for next round calculation. 1443 */ 1444 1445 /* 1446 * Get the initial offset into the current packet instructions; 1447 * entry conditions ensure that instrs_prev is less than 1448 * etm->instructions_sample_period. 1449 */ 1450 u64 offset = etm->instructions_sample_period - instrs_prev; 1451 u64 addr; 1452 1453 /* Prepare last branches for instruction sample */ 1454 if (etm->synth_opts.last_branch) 1455 cs_etm__copy_last_branch_rb(etmq, tidq); 1456 1457 while (tidq->period_instructions >= 1458 etm->instructions_sample_period) { 1459 /* 1460 * Calculate the address of the sampled instruction (-1 1461 * as sample is reported as though instruction has just 1462 * been executed, but PC has not advanced to next 1463 * instruction) 1464 */ 1465 addr = cs_etm__instr_addr(etmq, trace_chan_id, 1466 tidq->packet, offset - 1); 1467 ret = cs_etm__synth_instruction_sample( 1468 etmq, tidq, addr, 1469 etm->instructions_sample_period); 1470 if (ret) 1471 return ret; 1472 1473 offset += etm->instructions_sample_period; 1474 tidq->period_instructions -= 1475 etm->instructions_sample_period; 1476 } 1477 } 1478 1479 if (etm->sample_branches) { 1480 bool generate_sample = false; 1481 1482 /* Generate sample for tracing on packet */ 1483 if (tidq->prev_packet->sample_type == CS_ETM_DISCONTINUITY) 1484 generate_sample = true; 1485 1486 /* Generate sample for branch taken packet */ 1487 if (tidq->prev_packet->sample_type == CS_ETM_RANGE && 1488 tidq->prev_packet->last_instr_taken_branch) 1489 generate_sample = true; 1490 1491 if (generate_sample) { 1492 ret = cs_etm__synth_branch_sample(etmq, tidq); 1493 if (ret) 1494 return ret; 1495 } 1496 } 1497 1498 cs_etm__packet_swap(etm, tidq); 1499 1500 return 0; 1501 } 1502 1503 static int cs_etm__exception(struct cs_etm_traceid_queue *tidq) 1504 { 1505 /* 1506 * When the exception packet is inserted, whether the last instruction 1507 * in previous range packet is taken branch or not, we need to force 1508 * to set 'prev_packet->last_instr_taken_branch' to true. This ensures 1509 * to generate branch sample for the instruction range before the 1510 * exception is trapped to kernel or before the exception returning. 1511 * 1512 * The exception packet includes the dummy address values, so don't 1513 * swap PACKET with PREV_PACKET. This keeps PREV_PACKET to be useful 1514 * for generating instruction and branch samples. 1515 */ 1516 if (tidq->prev_packet->sample_type == CS_ETM_RANGE) 1517 tidq->prev_packet->last_instr_taken_branch = true; 1518 1519 return 0; 1520 } 1521 1522 static int cs_etm__flush(struct cs_etm_queue *etmq, 1523 struct cs_etm_traceid_queue *tidq) 1524 { 1525 int err = 0; 1526 struct cs_etm_auxtrace *etm = etmq->etm; 1527 1528 /* Handle start tracing packet */ 1529 if (tidq->prev_packet->sample_type == CS_ETM_EMPTY) 1530 goto swap_packet; 1531 1532 if (etmq->etm->synth_opts.last_branch && 1533 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1534 u64 addr; 1535 1536 /* Prepare last branches for instruction sample */ 1537 cs_etm__copy_last_branch_rb(etmq, tidq); 1538 1539 /* 1540 * Generate a last branch event for the branches left in the 1541 * circular buffer at the end of the trace. 1542 * 1543 * Use the address of the end of the last reported execution 1544 * range 1545 */ 1546 addr = cs_etm__last_executed_instr(tidq->prev_packet); 1547 1548 err = cs_etm__synth_instruction_sample( 1549 etmq, tidq, addr, 1550 tidq->period_instructions); 1551 if (err) 1552 return err; 1553 1554 tidq->period_instructions = 0; 1555 1556 } 1557 1558 if (etm->sample_branches && 1559 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1560 err = cs_etm__synth_branch_sample(etmq, tidq); 1561 if (err) 1562 return err; 1563 } 1564 1565 swap_packet: 1566 cs_etm__packet_swap(etm, tidq); 1567 1568 /* Reset last branches after flush the trace */ 1569 if (etm->synth_opts.last_branch) 1570 cs_etm__reset_last_branch_rb(tidq); 1571 1572 return err; 1573 } 1574 1575 static int cs_etm__end_block(struct cs_etm_queue *etmq, 1576 struct cs_etm_traceid_queue *tidq) 1577 { 1578 int err; 1579 1580 /* 1581 * It has no new packet coming and 'etmq->packet' contains the stale 1582 * packet which was set at the previous time with packets swapping; 1583 * so skip to generate branch sample to avoid stale packet. 1584 * 1585 * For this case only flush branch stack and generate a last branch 1586 * event for the branches left in the circular buffer at the end of 1587 * the trace. 1588 */ 1589 if (etmq->etm->synth_opts.last_branch && 1590 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1591 u64 addr; 1592 1593 /* Prepare last branches for instruction sample */ 1594 cs_etm__copy_last_branch_rb(etmq, tidq); 1595 1596 /* 1597 * Use the address of the end of the last reported execution 1598 * range. 1599 */ 1600 addr = cs_etm__last_executed_instr(tidq->prev_packet); 1601 1602 err = cs_etm__synth_instruction_sample( 1603 etmq, tidq, addr, 1604 tidq->period_instructions); 1605 if (err) 1606 return err; 1607 1608 tidq->period_instructions = 0; 1609 } 1610 1611 return 0; 1612 } 1613 /* 1614 * cs_etm__get_data_block: Fetch a block from the auxtrace_buffer queue 1615 * if need be. 1616 * Returns: < 0 if error 1617 * = 0 if no more auxtrace_buffer to read 1618 * > 0 if the current buffer isn't empty yet 1619 */ 1620 static int cs_etm__get_data_block(struct cs_etm_queue *etmq) 1621 { 1622 int ret; 1623 1624 if (!etmq->buf_len) { 1625 ret = cs_etm__get_trace(etmq); 1626 if (ret <= 0) 1627 return ret; 1628 /* 1629 * We cannot assume consecutive blocks in the data file 1630 * are contiguous, reset the decoder to force re-sync. 1631 */ 1632 ret = cs_etm_decoder__reset(etmq->decoder); 1633 if (ret) 1634 return ret; 1635 } 1636 1637 return etmq->buf_len; 1638 } 1639 1640 static bool cs_etm__is_svc_instr(struct cs_etm_queue *etmq, u8 trace_chan_id, 1641 struct cs_etm_packet *packet, 1642 u64 end_addr) 1643 { 1644 /* Initialise to keep compiler happy */ 1645 u16 instr16 = 0; 1646 u32 instr32 = 0; 1647 u64 addr; 1648 1649 switch (packet->isa) { 1650 case CS_ETM_ISA_T32: 1651 /* 1652 * The SVC of T32 is defined in ARM DDI 0487D.a, F5.1.247: 1653 * 1654 * b'15 b'8 1655 * +-----------------+--------+ 1656 * | 1 1 0 1 1 1 1 1 | imm8 | 1657 * +-----------------+--------+ 1658 * 1659 * According to the specification, it only defines SVC for T32 1660 * with 16 bits instruction and has no definition for 32bits; 1661 * so below only read 2 bytes as instruction size for T32. 1662 */ 1663 addr = end_addr - 2; 1664 cs_etm__mem_access(etmq, trace_chan_id, addr, 1665 sizeof(instr16), (u8 *)&instr16); 1666 if ((instr16 & 0xFF00) == 0xDF00) 1667 return true; 1668 1669 break; 1670 case CS_ETM_ISA_A32: 1671 /* 1672 * The SVC of A32 is defined in ARM DDI 0487D.a, F5.1.247: 1673 * 1674 * b'31 b'28 b'27 b'24 1675 * +---------+---------+-------------------------+ 1676 * | !1111 | 1 1 1 1 | imm24 | 1677 * +---------+---------+-------------------------+ 1678 */ 1679 addr = end_addr - 4; 1680 cs_etm__mem_access(etmq, trace_chan_id, addr, 1681 sizeof(instr32), (u8 *)&instr32); 1682 if ((instr32 & 0x0F000000) == 0x0F000000 && 1683 (instr32 & 0xF0000000) != 0xF0000000) 1684 return true; 1685 1686 break; 1687 case CS_ETM_ISA_A64: 1688 /* 1689 * The SVC of A64 is defined in ARM DDI 0487D.a, C6.2.294: 1690 * 1691 * b'31 b'21 b'4 b'0 1692 * +-----------------------+---------+-----------+ 1693 * | 1 1 0 1 0 1 0 0 0 0 0 | imm16 | 0 0 0 0 1 | 1694 * +-----------------------+---------+-----------+ 1695 */ 1696 addr = end_addr - 4; 1697 cs_etm__mem_access(etmq, trace_chan_id, addr, 1698 sizeof(instr32), (u8 *)&instr32); 1699 if ((instr32 & 0xFFE0001F) == 0xd4000001) 1700 return true; 1701 1702 break; 1703 case CS_ETM_ISA_UNKNOWN: 1704 default: 1705 break; 1706 } 1707 1708 return false; 1709 } 1710 1711 static bool cs_etm__is_syscall(struct cs_etm_queue *etmq, 1712 struct cs_etm_traceid_queue *tidq, u64 magic) 1713 { 1714 u8 trace_chan_id = tidq->trace_chan_id; 1715 struct cs_etm_packet *packet = tidq->packet; 1716 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1717 1718 if (magic == __perf_cs_etmv3_magic) 1719 if (packet->exception_number == CS_ETMV3_EXC_SVC) 1720 return true; 1721 1722 /* 1723 * ETMv4 exception type CS_ETMV4_EXC_CALL covers SVC, SMC and 1724 * HVC cases; need to check if it's SVC instruction based on 1725 * packet address. 1726 */ 1727 if (magic == __perf_cs_etmv4_magic) { 1728 if (packet->exception_number == CS_ETMV4_EXC_CALL && 1729 cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet, 1730 prev_packet->end_addr)) 1731 return true; 1732 } 1733 1734 return false; 1735 } 1736 1737 static bool cs_etm__is_async_exception(struct cs_etm_traceid_queue *tidq, 1738 u64 magic) 1739 { 1740 struct cs_etm_packet *packet = tidq->packet; 1741 1742 if (magic == __perf_cs_etmv3_magic) 1743 if (packet->exception_number == CS_ETMV3_EXC_DEBUG_HALT || 1744 packet->exception_number == CS_ETMV3_EXC_ASYNC_DATA_ABORT || 1745 packet->exception_number == CS_ETMV3_EXC_PE_RESET || 1746 packet->exception_number == CS_ETMV3_EXC_IRQ || 1747 packet->exception_number == CS_ETMV3_EXC_FIQ) 1748 return true; 1749 1750 if (magic == __perf_cs_etmv4_magic) 1751 if (packet->exception_number == CS_ETMV4_EXC_RESET || 1752 packet->exception_number == CS_ETMV4_EXC_DEBUG_HALT || 1753 packet->exception_number == CS_ETMV4_EXC_SYSTEM_ERROR || 1754 packet->exception_number == CS_ETMV4_EXC_INST_DEBUG || 1755 packet->exception_number == CS_ETMV4_EXC_DATA_DEBUG || 1756 packet->exception_number == CS_ETMV4_EXC_IRQ || 1757 packet->exception_number == CS_ETMV4_EXC_FIQ) 1758 return true; 1759 1760 return false; 1761 } 1762 1763 static bool cs_etm__is_sync_exception(struct cs_etm_queue *etmq, 1764 struct cs_etm_traceid_queue *tidq, 1765 u64 magic) 1766 { 1767 u8 trace_chan_id = tidq->trace_chan_id; 1768 struct cs_etm_packet *packet = tidq->packet; 1769 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1770 1771 if (magic == __perf_cs_etmv3_magic) 1772 if (packet->exception_number == CS_ETMV3_EXC_SMC || 1773 packet->exception_number == CS_ETMV3_EXC_HYP || 1774 packet->exception_number == CS_ETMV3_EXC_JAZELLE_THUMBEE || 1775 packet->exception_number == CS_ETMV3_EXC_UNDEFINED_INSTR || 1776 packet->exception_number == CS_ETMV3_EXC_PREFETCH_ABORT || 1777 packet->exception_number == CS_ETMV3_EXC_DATA_FAULT || 1778 packet->exception_number == CS_ETMV3_EXC_GENERIC) 1779 return true; 1780 1781 if (magic == __perf_cs_etmv4_magic) { 1782 if (packet->exception_number == CS_ETMV4_EXC_TRAP || 1783 packet->exception_number == CS_ETMV4_EXC_ALIGNMENT || 1784 packet->exception_number == CS_ETMV4_EXC_INST_FAULT || 1785 packet->exception_number == CS_ETMV4_EXC_DATA_FAULT) 1786 return true; 1787 1788 /* 1789 * For CS_ETMV4_EXC_CALL, except SVC other instructions 1790 * (SMC, HVC) are taken as sync exceptions. 1791 */ 1792 if (packet->exception_number == CS_ETMV4_EXC_CALL && 1793 !cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet, 1794 prev_packet->end_addr)) 1795 return true; 1796 1797 /* 1798 * ETMv4 has 5 bits for exception number; if the numbers 1799 * are in the range ( CS_ETMV4_EXC_FIQ, CS_ETMV4_EXC_END ] 1800 * they are implementation defined exceptions. 1801 * 1802 * For this case, simply take it as sync exception. 1803 */ 1804 if (packet->exception_number > CS_ETMV4_EXC_FIQ && 1805 packet->exception_number <= CS_ETMV4_EXC_END) 1806 return true; 1807 } 1808 1809 return false; 1810 } 1811 1812 static int cs_etm__set_sample_flags(struct cs_etm_queue *etmq, 1813 struct cs_etm_traceid_queue *tidq) 1814 { 1815 struct cs_etm_packet *packet = tidq->packet; 1816 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1817 u8 trace_chan_id = tidq->trace_chan_id; 1818 u64 magic; 1819 int ret; 1820 1821 switch (packet->sample_type) { 1822 case CS_ETM_RANGE: 1823 /* 1824 * Immediate branch instruction without neither link nor 1825 * return flag, it's normal branch instruction within 1826 * the function. 1827 */ 1828 if (packet->last_instr_type == OCSD_INSTR_BR && 1829 packet->last_instr_subtype == OCSD_S_INSTR_NONE) { 1830 packet->flags = PERF_IP_FLAG_BRANCH; 1831 1832 if (packet->last_instr_cond) 1833 packet->flags |= PERF_IP_FLAG_CONDITIONAL; 1834 } 1835 1836 /* 1837 * Immediate branch instruction with link (e.g. BL), this is 1838 * branch instruction for function call. 1839 */ 1840 if (packet->last_instr_type == OCSD_INSTR_BR && 1841 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK) 1842 packet->flags = PERF_IP_FLAG_BRANCH | 1843 PERF_IP_FLAG_CALL; 1844 1845 /* 1846 * Indirect branch instruction with link (e.g. BLR), this is 1847 * branch instruction for function call. 1848 */ 1849 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1850 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK) 1851 packet->flags = PERF_IP_FLAG_BRANCH | 1852 PERF_IP_FLAG_CALL; 1853 1854 /* 1855 * Indirect branch instruction with subtype of 1856 * OCSD_S_INSTR_V7_IMPLIED_RET, this is explicit hint for 1857 * function return for A32/T32. 1858 */ 1859 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1860 packet->last_instr_subtype == OCSD_S_INSTR_V7_IMPLIED_RET) 1861 packet->flags = PERF_IP_FLAG_BRANCH | 1862 PERF_IP_FLAG_RETURN; 1863 1864 /* 1865 * Indirect branch instruction without link (e.g. BR), usually 1866 * this is used for function return, especially for functions 1867 * within dynamic link lib. 1868 */ 1869 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1870 packet->last_instr_subtype == OCSD_S_INSTR_NONE) 1871 packet->flags = PERF_IP_FLAG_BRANCH | 1872 PERF_IP_FLAG_RETURN; 1873 1874 /* Return instruction for function return. */ 1875 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1876 packet->last_instr_subtype == OCSD_S_INSTR_V8_RET) 1877 packet->flags = PERF_IP_FLAG_BRANCH | 1878 PERF_IP_FLAG_RETURN; 1879 1880 /* 1881 * Decoder might insert a discontinuity in the middle of 1882 * instruction packets, fixup prev_packet with flag 1883 * PERF_IP_FLAG_TRACE_BEGIN to indicate restarting trace. 1884 */ 1885 if (prev_packet->sample_type == CS_ETM_DISCONTINUITY) 1886 prev_packet->flags |= PERF_IP_FLAG_BRANCH | 1887 PERF_IP_FLAG_TRACE_BEGIN; 1888 1889 /* 1890 * If the previous packet is an exception return packet 1891 * and the return address just follows SVC instruction, 1892 * it needs to calibrate the previous packet sample flags 1893 * as PERF_IP_FLAG_SYSCALLRET. 1894 */ 1895 if (prev_packet->flags == (PERF_IP_FLAG_BRANCH | 1896 PERF_IP_FLAG_RETURN | 1897 PERF_IP_FLAG_INTERRUPT) && 1898 cs_etm__is_svc_instr(etmq, trace_chan_id, 1899 packet, packet->start_addr)) 1900 prev_packet->flags = PERF_IP_FLAG_BRANCH | 1901 PERF_IP_FLAG_RETURN | 1902 PERF_IP_FLAG_SYSCALLRET; 1903 break; 1904 case CS_ETM_DISCONTINUITY: 1905 /* 1906 * The trace is discontinuous, if the previous packet is 1907 * instruction packet, set flag PERF_IP_FLAG_TRACE_END 1908 * for previous packet. 1909 */ 1910 if (prev_packet->sample_type == CS_ETM_RANGE) 1911 prev_packet->flags |= PERF_IP_FLAG_BRANCH | 1912 PERF_IP_FLAG_TRACE_END; 1913 break; 1914 case CS_ETM_EXCEPTION: 1915 ret = cs_etm__get_magic(packet->trace_chan_id, &magic); 1916 if (ret) 1917 return ret; 1918 1919 /* The exception is for system call. */ 1920 if (cs_etm__is_syscall(etmq, tidq, magic)) 1921 packet->flags = PERF_IP_FLAG_BRANCH | 1922 PERF_IP_FLAG_CALL | 1923 PERF_IP_FLAG_SYSCALLRET; 1924 /* 1925 * The exceptions are triggered by external signals from bus, 1926 * interrupt controller, debug module, PE reset or halt. 1927 */ 1928 else if (cs_etm__is_async_exception(tidq, magic)) 1929 packet->flags = PERF_IP_FLAG_BRANCH | 1930 PERF_IP_FLAG_CALL | 1931 PERF_IP_FLAG_ASYNC | 1932 PERF_IP_FLAG_INTERRUPT; 1933 /* 1934 * Otherwise, exception is caused by trap, instruction & 1935 * data fault, or alignment errors. 1936 */ 1937 else if (cs_etm__is_sync_exception(etmq, tidq, magic)) 1938 packet->flags = PERF_IP_FLAG_BRANCH | 1939 PERF_IP_FLAG_CALL | 1940 PERF_IP_FLAG_INTERRUPT; 1941 1942 /* 1943 * When the exception packet is inserted, since exception 1944 * packet is not used standalone for generating samples 1945 * and it's affiliation to the previous instruction range 1946 * packet; so set previous range packet flags to tell perf 1947 * it is an exception taken branch. 1948 */ 1949 if (prev_packet->sample_type == CS_ETM_RANGE) 1950 prev_packet->flags = packet->flags; 1951 break; 1952 case CS_ETM_EXCEPTION_RET: 1953 /* 1954 * When the exception return packet is inserted, since 1955 * exception return packet is not used standalone for 1956 * generating samples and it's affiliation to the previous 1957 * instruction range packet; so set previous range packet 1958 * flags to tell perf it is an exception return branch. 1959 * 1960 * The exception return can be for either system call or 1961 * other exception types; unfortunately the packet doesn't 1962 * contain exception type related info so we cannot decide 1963 * the exception type purely based on exception return packet. 1964 * If we record the exception number from exception packet and 1965 * reuse it for exception return packet, this is not reliable 1966 * due the trace can be discontinuity or the interrupt can 1967 * be nested, thus the recorded exception number cannot be 1968 * used for exception return packet for these two cases. 1969 * 1970 * For exception return packet, we only need to distinguish the 1971 * packet is for system call or for other types. Thus the 1972 * decision can be deferred when receive the next packet which 1973 * contains the return address, based on the return address we 1974 * can read out the previous instruction and check if it's a 1975 * system call instruction and then calibrate the sample flag 1976 * as needed. 1977 */ 1978 if (prev_packet->sample_type == CS_ETM_RANGE) 1979 prev_packet->flags = PERF_IP_FLAG_BRANCH | 1980 PERF_IP_FLAG_RETURN | 1981 PERF_IP_FLAG_INTERRUPT; 1982 break; 1983 case CS_ETM_EMPTY: 1984 default: 1985 break; 1986 } 1987 1988 return 0; 1989 } 1990 1991 static int cs_etm__decode_data_block(struct cs_etm_queue *etmq) 1992 { 1993 int ret = 0; 1994 size_t processed = 0; 1995 1996 /* 1997 * Packets are decoded and added to the decoder's packet queue 1998 * until the decoder packet processing callback has requested that 1999 * processing stops or there is nothing left in the buffer. Normal 2000 * operations that stop processing are a timestamp packet or a full 2001 * decoder buffer queue. 2002 */ 2003 ret = cs_etm_decoder__process_data_block(etmq->decoder, 2004 etmq->offset, 2005 &etmq->buf[etmq->buf_used], 2006 etmq->buf_len, 2007 &processed); 2008 if (ret) 2009 goto out; 2010 2011 etmq->offset += processed; 2012 etmq->buf_used += processed; 2013 etmq->buf_len -= processed; 2014 2015 out: 2016 return ret; 2017 } 2018 2019 static int cs_etm__process_traceid_queue(struct cs_etm_queue *etmq, 2020 struct cs_etm_traceid_queue *tidq) 2021 { 2022 int ret; 2023 struct cs_etm_packet_queue *packet_queue; 2024 2025 packet_queue = &tidq->packet_queue; 2026 2027 /* Process each packet in this chunk */ 2028 while (1) { 2029 ret = cs_etm_decoder__get_packet(packet_queue, 2030 tidq->packet); 2031 if (ret <= 0) 2032 /* 2033 * Stop processing this chunk on 2034 * end of data or error 2035 */ 2036 break; 2037 2038 /* 2039 * Since packet addresses are swapped in packet 2040 * handling within below switch() statements, 2041 * thus setting sample flags must be called 2042 * prior to switch() statement to use address 2043 * information before packets swapping. 2044 */ 2045 ret = cs_etm__set_sample_flags(etmq, tidq); 2046 if (ret < 0) 2047 break; 2048 2049 switch (tidq->packet->sample_type) { 2050 case CS_ETM_RANGE: 2051 /* 2052 * If the packet contains an instruction 2053 * range, generate instruction sequence 2054 * events. 2055 */ 2056 cs_etm__sample(etmq, tidq); 2057 break; 2058 case CS_ETM_EXCEPTION: 2059 case CS_ETM_EXCEPTION_RET: 2060 /* 2061 * If the exception packet is coming, 2062 * make sure the previous instruction 2063 * range packet to be handled properly. 2064 */ 2065 cs_etm__exception(tidq); 2066 break; 2067 case CS_ETM_DISCONTINUITY: 2068 /* 2069 * Discontinuity in trace, flush 2070 * previous branch stack 2071 */ 2072 cs_etm__flush(etmq, tidq); 2073 break; 2074 case CS_ETM_EMPTY: 2075 /* 2076 * Should not receive empty packet, 2077 * report error. 2078 */ 2079 pr_err("CS ETM Trace: empty packet\n"); 2080 return -EINVAL; 2081 default: 2082 break; 2083 } 2084 } 2085 2086 return ret; 2087 } 2088 2089 static void cs_etm__clear_all_traceid_queues(struct cs_etm_queue *etmq) 2090 { 2091 int idx; 2092 struct int_node *inode; 2093 struct cs_etm_traceid_queue *tidq; 2094 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 2095 2096 intlist__for_each_entry(inode, traceid_queues_list) { 2097 idx = (int)(intptr_t)inode->priv; 2098 tidq = etmq->traceid_queues[idx]; 2099 2100 /* Ignore return value */ 2101 cs_etm__process_traceid_queue(etmq, tidq); 2102 2103 /* 2104 * Generate an instruction sample with the remaining 2105 * branchstack entries. 2106 */ 2107 cs_etm__flush(etmq, tidq); 2108 } 2109 } 2110 2111 static int cs_etm__run_decoder(struct cs_etm_queue *etmq) 2112 { 2113 int err = 0; 2114 struct cs_etm_traceid_queue *tidq; 2115 2116 tidq = cs_etm__etmq_get_traceid_queue(etmq, CS_ETM_PER_THREAD_TRACEID); 2117 if (!tidq) 2118 return -EINVAL; 2119 2120 /* Go through each buffer in the queue and decode them one by one */ 2121 while (1) { 2122 err = cs_etm__get_data_block(etmq); 2123 if (err <= 0) 2124 return err; 2125 2126 /* Run trace decoder until buffer consumed or end of trace */ 2127 do { 2128 err = cs_etm__decode_data_block(etmq); 2129 if (err) 2130 return err; 2131 2132 /* 2133 * Process each packet in this chunk, nothing to do if 2134 * an error occurs other than hoping the next one will 2135 * be better. 2136 */ 2137 err = cs_etm__process_traceid_queue(etmq, tidq); 2138 2139 } while (etmq->buf_len); 2140 2141 if (err == 0) 2142 /* Flush any remaining branch stack entries */ 2143 err = cs_etm__end_block(etmq, tidq); 2144 } 2145 2146 return err; 2147 } 2148 2149 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm, 2150 pid_t tid) 2151 { 2152 unsigned int i; 2153 struct auxtrace_queues *queues = &etm->queues; 2154 2155 for (i = 0; i < queues->nr_queues; i++) { 2156 struct auxtrace_queue *queue = &etm->queues.queue_array[i]; 2157 struct cs_etm_queue *etmq = queue->priv; 2158 struct cs_etm_traceid_queue *tidq; 2159 2160 if (!etmq) 2161 continue; 2162 2163 tidq = cs_etm__etmq_get_traceid_queue(etmq, 2164 CS_ETM_PER_THREAD_TRACEID); 2165 2166 if (!tidq) 2167 continue; 2168 2169 if ((tid == -1) || (tidq->tid == tid)) { 2170 cs_etm__set_pid_tid_cpu(etm, tidq); 2171 cs_etm__run_decoder(etmq); 2172 } 2173 } 2174 2175 return 0; 2176 } 2177 2178 static int cs_etm__process_queues(struct cs_etm_auxtrace *etm) 2179 { 2180 int ret = 0; 2181 unsigned int cs_queue_nr, queue_nr, i; 2182 u8 trace_chan_id; 2183 u64 cs_timestamp; 2184 struct auxtrace_queue *queue; 2185 struct cs_etm_queue *etmq; 2186 struct cs_etm_traceid_queue *tidq; 2187 2188 /* 2189 * Pre-populate the heap with one entry from each queue so that we can 2190 * start processing in time order across all queues. 2191 */ 2192 for (i = 0; i < etm->queues.nr_queues; i++) { 2193 etmq = etm->queues.queue_array[i].priv; 2194 if (!etmq) 2195 continue; 2196 2197 ret = cs_etm__queue_first_cs_timestamp(etm, etmq, i); 2198 if (ret) 2199 return ret; 2200 } 2201 2202 while (1) { 2203 if (!etm->heap.heap_cnt) 2204 goto out; 2205 2206 /* Take the entry at the top of the min heap */ 2207 cs_queue_nr = etm->heap.heap_array[0].queue_nr; 2208 queue_nr = TO_QUEUE_NR(cs_queue_nr); 2209 trace_chan_id = TO_TRACE_CHAN_ID(cs_queue_nr); 2210 queue = &etm->queues.queue_array[queue_nr]; 2211 etmq = queue->priv; 2212 2213 /* 2214 * Remove the top entry from the heap since we are about 2215 * to process it. 2216 */ 2217 auxtrace_heap__pop(&etm->heap); 2218 2219 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 2220 if (!tidq) { 2221 /* 2222 * No traceID queue has been allocated for this traceID, 2223 * which means something somewhere went very wrong. No 2224 * other choice than simply exit. 2225 */ 2226 ret = -EINVAL; 2227 goto out; 2228 } 2229 2230 /* 2231 * Packets associated with this timestamp are already in 2232 * the etmq's traceID queue, so process them. 2233 */ 2234 ret = cs_etm__process_traceid_queue(etmq, tidq); 2235 if (ret < 0) 2236 goto out; 2237 2238 /* 2239 * Packets for this timestamp have been processed, time to 2240 * move on to the next timestamp, fetching a new auxtrace_buffer 2241 * if need be. 2242 */ 2243 refetch: 2244 ret = cs_etm__get_data_block(etmq); 2245 if (ret < 0) 2246 goto out; 2247 2248 /* 2249 * No more auxtrace_buffers to process in this etmq, simply 2250 * move on to another entry in the auxtrace_heap. 2251 */ 2252 if (!ret) 2253 continue; 2254 2255 ret = cs_etm__decode_data_block(etmq); 2256 if (ret) 2257 goto out; 2258 2259 cs_timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id); 2260 2261 if (!cs_timestamp) { 2262 /* 2263 * Function cs_etm__decode_data_block() returns when 2264 * there is no more traces to decode in the current 2265 * auxtrace_buffer OR when a timestamp has been 2266 * encountered on any of the traceID queues. Since we 2267 * did not get a timestamp, there is no more traces to 2268 * process in this auxtrace_buffer. As such empty and 2269 * flush all traceID queues. 2270 */ 2271 cs_etm__clear_all_traceid_queues(etmq); 2272 2273 /* Fetch another auxtrace_buffer for this etmq */ 2274 goto refetch; 2275 } 2276 2277 /* 2278 * Add to the min heap the timestamp for packets that have 2279 * just been decoded. They will be processed and synthesized 2280 * during the next call to cs_etm__process_traceid_queue() for 2281 * this queue/traceID. 2282 */ 2283 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id); 2284 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, cs_timestamp); 2285 } 2286 2287 out: 2288 return ret; 2289 } 2290 2291 static int cs_etm__process_itrace_start(struct cs_etm_auxtrace *etm, 2292 union perf_event *event) 2293 { 2294 struct thread *th; 2295 2296 if (etm->timeless_decoding) 2297 return 0; 2298 2299 /* 2300 * Add the tid/pid to the log so that we can get a match when 2301 * we get a contextID from the decoder. 2302 */ 2303 th = machine__findnew_thread(etm->machine, 2304 event->itrace_start.pid, 2305 event->itrace_start.tid); 2306 if (!th) 2307 return -ENOMEM; 2308 2309 thread__put(th); 2310 2311 return 0; 2312 } 2313 2314 static int cs_etm__process_switch_cpu_wide(struct cs_etm_auxtrace *etm, 2315 union perf_event *event) 2316 { 2317 struct thread *th; 2318 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT; 2319 2320 /* 2321 * Context switch in per-thread mode are irrelevant since perf 2322 * will start/stop tracing as the process is scheduled. 2323 */ 2324 if (etm->timeless_decoding) 2325 return 0; 2326 2327 /* 2328 * SWITCH_IN events carry the next process to be switched out while 2329 * SWITCH_OUT events carry the process to be switched in. As such 2330 * we don't care about IN events. 2331 */ 2332 if (!out) 2333 return 0; 2334 2335 /* 2336 * Add the tid/pid to the log so that we can get a match when 2337 * we get a contextID from the decoder. 2338 */ 2339 th = machine__findnew_thread(etm->machine, 2340 event->context_switch.next_prev_pid, 2341 event->context_switch.next_prev_tid); 2342 if (!th) 2343 return -ENOMEM; 2344 2345 thread__put(th); 2346 2347 return 0; 2348 } 2349 2350 static int cs_etm__process_event(struct perf_session *session, 2351 union perf_event *event, 2352 struct perf_sample *sample, 2353 struct perf_tool *tool) 2354 { 2355 u64 sample_kernel_timestamp; 2356 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2357 struct cs_etm_auxtrace, 2358 auxtrace); 2359 2360 if (dump_trace) 2361 return 0; 2362 2363 if (!tool->ordered_events) { 2364 pr_err("CoreSight ETM Trace requires ordered events\n"); 2365 return -EINVAL; 2366 } 2367 2368 if (sample->time && (sample->time != (u64) -1)) 2369 sample_kernel_timestamp = sample->time; 2370 else 2371 sample_kernel_timestamp = 0; 2372 2373 /* 2374 * Don't wait for cs_etm__flush_events() in per-thread/timeless mode to start the decode. We 2375 * need the tid of the PERF_RECORD_EXIT event to assign to the synthesised samples because 2376 * ETM_OPT_CTXTID is not enabled. 2377 */ 2378 if (etm->timeless_decoding && 2379 event->header.type == PERF_RECORD_EXIT) 2380 return cs_etm__process_timeless_queues(etm, 2381 event->fork.tid); 2382 2383 if (event->header.type == PERF_RECORD_ITRACE_START) 2384 return cs_etm__process_itrace_start(etm, event); 2385 else if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) 2386 return cs_etm__process_switch_cpu_wide(etm, event); 2387 2388 if (!etm->timeless_decoding && event->header.type == PERF_RECORD_AUX) { 2389 /* 2390 * Record the latest kernel timestamp available in the header 2391 * for samples so that synthesised samples occur from this point 2392 * onwards. 2393 */ 2394 etm->latest_kernel_timestamp = sample_kernel_timestamp; 2395 } 2396 2397 return 0; 2398 } 2399 2400 static void dump_queued_data(struct cs_etm_auxtrace *etm, 2401 struct perf_record_auxtrace *event) 2402 { 2403 struct auxtrace_buffer *buf; 2404 unsigned int i; 2405 /* 2406 * Find all buffers with same reference in the queues and dump them. 2407 * This is because the queues can contain multiple entries of the same 2408 * buffer that were split on aux records. 2409 */ 2410 for (i = 0; i < etm->queues.nr_queues; ++i) 2411 list_for_each_entry(buf, &etm->queues.queue_array[i].head, list) 2412 if (buf->reference == event->reference) 2413 cs_etm__dump_event(etm->queues.queue_array[i].priv, buf); 2414 } 2415 2416 static int cs_etm__process_auxtrace_event(struct perf_session *session, 2417 union perf_event *event, 2418 struct perf_tool *tool __maybe_unused) 2419 { 2420 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2421 struct cs_etm_auxtrace, 2422 auxtrace); 2423 if (!etm->data_queued) { 2424 struct auxtrace_buffer *buffer; 2425 off_t data_offset; 2426 int fd = perf_data__fd(session->data); 2427 bool is_pipe = perf_data__is_pipe(session->data); 2428 int err; 2429 int idx = event->auxtrace.idx; 2430 2431 if (is_pipe) 2432 data_offset = 0; 2433 else { 2434 data_offset = lseek(fd, 0, SEEK_CUR); 2435 if (data_offset == -1) 2436 return -errno; 2437 } 2438 2439 err = auxtrace_queues__add_event(&etm->queues, session, 2440 event, data_offset, &buffer); 2441 if (err) 2442 return err; 2443 2444 /* 2445 * Knowing if the trace is formatted or not requires a lookup of 2446 * the aux record so only works in non-piped mode where data is 2447 * queued in cs_etm__queue_aux_records(). Always assume 2448 * formatted in piped mode (true). 2449 */ 2450 err = cs_etm__setup_queue(etm, &etm->queues.queue_array[idx], 2451 idx, true); 2452 if (err) 2453 return err; 2454 2455 if (dump_trace) 2456 if (auxtrace_buffer__get_data(buffer, fd)) { 2457 cs_etm__dump_event(etm->queues.queue_array[idx].priv, buffer); 2458 auxtrace_buffer__put_data(buffer); 2459 } 2460 } else if (dump_trace) 2461 dump_queued_data(etm, &event->auxtrace); 2462 2463 return 0; 2464 } 2465 2466 static bool cs_etm__is_timeless_decoding(struct cs_etm_auxtrace *etm) 2467 { 2468 struct evsel *evsel; 2469 struct evlist *evlist = etm->session->evlist; 2470 bool timeless_decoding = true; 2471 2472 /* Override timeless mode with user input from --itrace=Z */ 2473 if (etm->synth_opts.timeless_decoding) 2474 return true; 2475 2476 /* 2477 * Circle through the list of event and complain if we find one 2478 * with the time bit set. 2479 */ 2480 evlist__for_each_entry(evlist, evsel) { 2481 if ((evsel->core.attr.sample_type & PERF_SAMPLE_TIME)) 2482 timeless_decoding = false; 2483 } 2484 2485 return timeless_decoding; 2486 } 2487 2488 static const char * const cs_etm_global_header_fmts[] = { 2489 [CS_HEADER_VERSION] = " Header version %llx\n", 2490 [CS_PMU_TYPE_CPUS] = " PMU type/num cpus %llx\n", 2491 [CS_ETM_SNAPSHOT] = " Snapshot %llx\n", 2492 }; 2493 2494 static const char * const cs_etm_priv_fmts[] = { 2495 [CS_ETM_MAGIC] = " Magic number %llx\n", 2496 [CS_ETM_CPU] = " CPU %lld\n", 2497 [CS_ETM_NR_TRC_PARAMS] = " NR_TRC_PARAMS %llx\n", 2498 [CS_ETM_ETMCR] = " ETMCR %llx\n", 2499 [CS_ETM_ETMTRACEIDR] = " ETMTRACEIDR %llx\n", 2500 [CS_ETM_ETMCCER] = " ETMCCER %llx\n", 2501 [CS_ETM_ETMIDR] = " ETMIDR %llx\n", 2502 }; 2503 2504 static const char * const cs_etmv4_priv_fmts[] = { 2505 [CS_ETM_MAGIC] = " Magic number %llx\n", 2506 [CS_ETM_CPU] = " CPU %lld\n", 2507 [CS_ETM_NR_TRC_PARAMS] = " NR_TRC_PARAMS %llx\n", 2508 [CS_ETMV4_TRCCONFIGR] = " TRCCONFIGR %llx\n", 2509 [CS_ETMV4_TRCTRACEIDR] = " TRCTRACEIDR %llx\n", 2510 [CS_ETMV4_TRCIDR0] = " TRCIDR0 %llx\n", 2511 [CS_ETMV4_TRCIDR1] = " TRCIDR1 %llx\n", 2512 [CS_ETMV4_TRCIDR2] = " TRCIDR2 %llx\n", 2513 [CS_ETMV4_TRCIDR8] = " TRCIDR8 %llx\n", 2514 [CS_ETMV4_TRCAUTHSTATUS] = " TRCAUTHSTATUS %llx\n", 2515 }; 2516 2517 static const char * const param_unk_fmt = 2518 " Unknown parameter [%d] %llx\n"; 2519 static const char * const magic_unk_fmt = 2520 " Magic number Unknown %llx\n"; 2521 2522 static int cs_etm__print_cpu_metadata_v0(__u64 *val, int *offset) 2523 { 2524 int i = *offset, j, nr_params = 0, fmt_offset; 2525 __u64 magic; 2526 2527 /* check magic value */ 2528 magic = val[i + CS_ETM_MAGIC]; 2529 if ((magic != __perf_cs_etmv3_magic) && 2530 (magic != __perf_cs_etmv4_magic)) { 2531 /* failure - note bad magic value */ 2532 fprintf(stdout, magic_unk_fmt, magic); 2533 return -EINVAL; 2534 } 2535 2536 /* print common header block */ 2537 fprintf(stdout, cs_etm_priv_fmts[CS_ETM_MAGIC], val[i++]); 2538 fprintf(stdout, cs_etm_priv_fmts[CS_ETM_CPU], val[i++]); 2539 2540 if (magic == __perf_cs_etmv3_magic) { 2541 nr_params = CS_ETM_NR_TRC_PARAMS_V0; 2542 fmt_offset = CS_ETM_ETMCR; 2543 /* after common block, offset format index past NR_PARAMS */ 2544 for (j = fmt_offset; j < nr_params + fmt_offset; j++, i++) 2545 fprintf(stdout, cs_etm_priv_fmts[j], val[i]); 2546 } else if (magic == __perf_cs_etmv4_magic) { 2547 nr_params = CS_ETMV4_NR_TRC_PARAMS_V0; 2548 fmt_offset = CS_ETMV4_TRCCONFIGR; 2549 /* after common block, offset format index past NR_PARAMS */ 2550 for (j = fmt_offset; j < nr_params + fmt_offset; j++, i++) 2551 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]); 2552 } 2553 *offset = i; 2554 return 0; 2555 } 2556 2557 static int cs_etm__print_cpu_metadata_v1(__u64 *val, int *offset) 2558 { 2559 int i = *offset, j, total_params = 0; 2560 __u64 magic; 2561 2562 magic = val[i + CS_ETM_MAGIC]; 2563 /* total params to print is NR_PARAMS + common block size for v1 */ 2564 total_params = val[i + CS_ETM_NR_TRC_PARAMS] + CS_ETM_COMMON_BLK_MAX_V1; 2565 2566 if (magic == __perf_cs_etmv3_magic) { 2567 for (j = 0; j < total_params; j++, i++) { 2568 /* if newer record - could be excess params */ 2569 if (j >= CS_ETM_PRIV_MAX) 2570 fprintf(stdout, param_unk_fmt, j, val[i]); 2571 else 2572 fprintf(stdout, cs_etm_priv_fmts[j], val[i]); 2573 } 2574 } else if (magic == __perf_cs_etmv4_magic) { 2575 for (j = 0; j < total_params; j++, i++) { 2576 /* if newer record - could be excess params */ 2577 if (j >= CS_ETMV4_PRIV_MAX) 2578 fprintf(stdout, param_unk_fmt, j, val[i]); 2579 else 2580 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]); 2581 } 2582 } else { 2583 /* failure - note bad magic value and error out */ 2584 fprintf(stdout, magic_unk_fmt, magic); 2585 return -EINVAL; 2586 } 2587 *offset = i; 2588 return 0; 2589 } 2590 2591 static void cs_etm__print_auxtrace_info(__u64 *val, int num) 2592 { 2593 int i, cpu = 0, version, err; 2594 2595 /* bail out early on bad header version */ 2596 version = val[0]; 2597 if (version > CS_HEADER_CURRENT_VERSION) { 2598 /* failure.. return */ 2599 fprintf(stdout, " Unknown Header Version = %x, ", version); 2600 fprintf(stdout, "Version supported <= %x\n", CS_HEADER_CURRENT_VERSION); 2601 return; 2602 } 2603 2604 for (i = 0; i < CS_HEADER_VERSION_MAX; i++) 2605 fprintf(stdout, cs_etm_global_header_fmts[i], val[i]); 2606 2607 for (i = CS_HEADER_VERSION_MAX; cpu < num; cpu++) { 2608 if (version == 0) 2609 err = cs_etm__print_cpu_metadata_v0(val, &i); 2610 else if (version == 1) 2611 err = cs_etm__print_cpu_metadata_v1(val, &i); 2612 if (err) 2613 return; 2614 } 2615 } 2616 2617 /* 2618 * Read a single cpu parameter block from the auxtrace_info priv block. 2619 * 2620 * For version 1 there is a per cpu nr_params entry. If we are handling 2621 * version 1 file, then there may be less, the same, or more params 2622 * indicated by this value than the compile time number we understand. 2623 * 2624 * For a version 0 info block, there are a fixed number, and we need to 2625 * fill out the nr_param value in the metadata we create. 2626 */ 2627 static u64 *cs_etm__create_meta_blk(u64 *buff_in, int *buff_in_offset, 2628 int out_blk_size, int nr_params_v0) 2629 { 2630 u64 *metadata = NULL; 2631 int hdr_version; 2632 int nr_in_params, nr_out_params, nr_cmn_params; 2633 int i, k; 2634 2635 metadata = zalloc(sizeof(*metadata) * out_blk_size); 2636 if (!metadata) 2637 return NULL; 2638 2639 /* read block current index & version */ 2640 i = *buff_in_offset; 2641 hdr_version = buff_in[CS_HEADER_VERSION]; 2642 2643 if (!hdr_version) { 2644 /* read version 0 info block into a version 1 metadata block */ 2645 nr_in_params = nr_params_v0; 2646 metadata[CS_ETM_MAGIC] = buff_in[i + CS_ETM_MAGIC]; 2647 metadata[CS_ETM_CPU] = buff_in[i + CS_ETM_CPU]; 2648 metadata[CS_ETM_NR_TRC_PARAMS] = nr_in_params; 2649 /* remaining block params at offset +1 from source */ 2650 for (k = CS_ETM_COMMON_BLK_MAX_V1 - 1; k < nr_in_params; k++) 2651 metadata[k + 1] = buff_in[i + k]; 2652 /* version 0 has 2 common params */ 2653 nr_cmn_params = 2; 2654 } else { 2655 /* read version 1 info block - input and output nr_params may differ */ 2656 /* version 1 has 3 common params */ 2657 nr_cmn_params = 3; 2658 nr_in_params = buff_in[i + CS_ETM_NR_TRC_PARAMS]; 2659 2660 /* if input has more params than output - skip excess */ 2661 nr_out_params = nr_in_params + nr_cmn_params; 2662 if (nr_out_params > out_blk_size) 2663 nr_out_params = out_blk_size; 2664 2665 for (k = CS_ETM_MAGIC; k < nr_out_params; k++) 2666 metadata[k] = buff_in[i + k]; 2667 2668 /* record the actual nr params we copied */ 2669 metadata[CS_ETM_NR_TRC_PARAMS] = nr_out_params - nr_cmn_params; 2670 } 2671 2672 /* adjust in offset by number of in params used */ 2673 i += nr_in_params + nr_cmn_params; 2674 *buff_in_offset = i; 2675 return metadata; 2676 } 2677 2678 /** 2679 * Puts a fragment of an auxtrace buffer into the auxtrace queues based 2680 * on the bounds of aux_event, if it matches with the buffer that's at 2681 * file_offset. 2682 * 2683 * Normally, whole auxtrace buffers would be added to the queue. But we 2684 * want to reset the decoder for every PERF_RECORD_AUX event, and the decoder 2685 * is reset across each buffer, so splitting the buffers up in advance has 2686 * the same effect. 2687 */ 2688 static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_offset, size_t sz, 2689 struct perf_record_aux *aux_event, struct perf_sample *sample) 2690 { 2691 int err; 2692 char buf[PERF_SAMPLE_MAX_SIZE]; 2693 union perf_event *auxtrace_event_union; 2694 struct perf_record_auxtrace *auxtrace_event; 2695 union perf_event auxtrace_fragment; 2696 __u64 aux_offset, aux_size; 2697 __u32 idx; 2698 bool formatted; 2699 2700 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2701 struct cs_etm_auxtrace, 2702 auxtrace); 2703 2704 /* 2705 * There should be a PERF_RECORD_AUXTRACE event at the file_offset that we got 2706 * from looping through the auxtrace index. 2707 */ 2708 err = perf_session__peek_event(session, file_offset, buf, 2709 PERF_SAMPLE_MAX_SIZE, &auxtrace_event_union, NULL); 2710 if (err) 2711 return err; 2712 auxtrace_event = &auxtrace_event_union->auxtrace; 2713 if (auxtrace_event->header.type != PERF_RECORD_AUXTRACE) 2714 return -EINVAL; 2715 2716 if (auxtrace_event->header.size < sizeof(struct perf_record_auxtrace) || 2717 auxtrace_event->header.size != sz) { 2718 return -EINVAL; 2719 } 2720 2721 /* 2722 * In per-thread mode, CPU is set to -1, but TID will be set instead. See 2723 * auxtrace_mmap_params__set_idx(). Return 'not found' if neither CPU nor TID match. 2724 */ 2725 if ((auxtrace_event->cpu == (__u32) -1 && auxtrace_event->tid != sample->tid) || 2726 auxtrace_event->cpu != sample->cpu) 2727 return 1; 2728 2729 if (aux_event->flags & PERF_AUX_FLAG_OVERWRITE) { 2730 /* 2731 * Clamp size in snapshot mode. The buffer size is clamped in 2732 * __auxtrace_mmap__read() for snapshots, so the aux record size doesn't reflect 2733 * the buffer size. 2734 */ 2735 aux_size = min(aux_event->aux_size, auxtrace_event->size); 2736 2737 /* 2738 * In this mode, the head also points to the end of the buffer so aux_offset 2739 * needs to have the size subtracted so it points to the beginning as in normal mode 2740 */ 2741 aux_offset = aux_event->aux_offset - aux_size; 2742 } else { 2743 aux_size = aux_event->aux_size; 2744 aux_offset = aux_event->aux_offset; 2745 } 2746 2747 if (aux_offset >= auxtrace_event->offset && 2748 aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) { 2749 /* 2750 * If this AUX event was inside this buffer somewhere, create a new auxtrace event 2751 * based on the sizes of the aux event, and queue that fragment. 2752 */ 2753 auxtrace_fragment.auxtrace = *auxtrace_event; 2754 auxtrace_fragment.auxtrace.size = aux_size; 2755 auxtrace_fragment.auxtrace.offset = aux_offset; 2756 file_offset += aux_offset - auxtrace_event->offset + auxtrace_event->header.size; 2757 2758 pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64 2759 " tid: %d cpu: %d\n", aux_size, aux_offset, sample->tid, sample->cpu); 2760 err = auxtrace_queues__add_event(&etm->queues, session, &auxtrace_fragment, 2761 file_offset, NULL); 2762 if (err) 2763 return err; 2764 2765 idx = auxtrace_event->idx; 2766 formatted = !(aux_event->flags & PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW); 2767 return cs_etm__setup_queue(etm, &etm->queues.queue_array[idx], 2768 idx, formatted); 2769 } 2770 2771 /* Wasn't inside this buffer, but there were no parse errors. 1 == 'not found' */ 2772 return 1; 2773 } 2774 2775 static int cs_etm__queue_aux_records_cb(struct perf_session *session, union perf_event *event, 2776 u64 offset __maybe_unused, void *data __maybe_unused) 2777 { 2778 struct perf_sample sample; 2779 int ret; 2780 struct auxtrace_index_entry *ent; 2781 struct auxtrace_index *auxtrace_index; 2782 struct evsel *evsel; 2783 size_t i; 2784 2785 /* Don't care about any other events, we're only queuing buffers for AUX events */ 2786 if (event->header.type != PERF_RECORD_AUX) 2787 return 0; 2788 2789 if (event->header.size < sizeof(struct perf_record_aux)) 2790 return -EINVAL; 2791 2792 /* Truncated Aux records can have 0 size and shouldn't result in anything being queued. */ 2793 if (!event->aux.aux_size) 2794 return 0; 2795 2796 /* 2797 * Parse the sample, we need the sample_id_all data that comes after the event so that the 2798 * CPU or PID can be matched to an AUXTRACE buffer's CPU or PID. 2799 */ 2800 evsel = evlist__event2evsel(session->evlist, event); 2801 if (!evsel) 2802 return -EINVAL; 2803 ret = evsel__parse_sample(evsel, event, &sample); 2804 if (ret) 2805 return ret; 2806 2807 /* 2808 * Loop through the auxtrace index to find the buffer that matches up with this aux event. 2809 */ 2810 list_for_each_entry(auxtrace_index, &session->auxtrace_index, list) { 2811 for (i = 0; i < auxtrace_index->nr; i++) { 2812 ent = &auxtrace_index->entries[i]; 2813 ret = cs_etm__queue_aux_fragment(session, ent->file_offset, 2814 ent->sz, &event->aux, &sample); 2815 /* 2816 * Stop search on error or successful values. Continue search on 2817 * 1 ('not found') 2818 */ 2819 if (ret != 1) 2820 return ret; 2821 } 2822 } 2823 2824 /* 2825 * Couldn't find the buffer corresponding to this aux record, something went wrong. Warn but 2826 * don't exit with an error because it will still be possible to decode other aux records. 2827 */ 2828 pr_err("CS ETM: Couldn't find auxtrace buffer for aux_offset: %#"PRI_lx64 2829 " tid: %d cpu: %d\n", event->aux.aux_offset, sample.tid, sample.cpu); 2830 return 0; 2831 } 2832 2833 static int cs_etm__queue_aux_records(struct perf_session *session) 2834 { 2835 struct auxtrace_index *index = list_first_entry_or_null(&session->auxtrace_index, 2836 struct auxtrace_index, list); 2837 if (index && index->nr > 0) 2838 return perf_session__peek_events(session, session->header.data_offset, 2839 session->header.data_size, 2840 cs_etm__queue_aux_records_cb, NULL); 2841 2842 /* 2843 * We would get here if there are no entries in the index (either no auxtrace 2844 * buffers or no index at all). Fail silently as there is the possibility of 2845 * queueing them in cs_etm__process_auxtrace_event() if etm->data_queued is still 2846 * false. 2847 * 2848 * In that scenario, buffers will not be split by AUX records. 2849 */ 2850 return 0; 2851 } 2852 2853 int cs_etm__process_auxtrace_info(union perf_event *event, 2854 struct perf_session *session) 2855 { 2856 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info; 2857 struct cs_etm_auxtrace *etm = NULL; 2858 struct int_node *inode; 2859 unsigned int pmu_type; 2860 int event_header_size = sizeof(struct perf_event_header); 2861 int info_header_size; 2862 int total_size = auxtrace_info->header.size; 2863 int priv_size = 0; 2864 int num_cpu, trcidr_idx; 2865 int err = 0; 2866 int i, j; 2867 u64 *ptr, *hdr = NULL; 2868 u64 **metadata = NULL; 2869 u64 hdr_version; 2870 2871 /* 2872 * sizeof(auxtrace_info_event::type) + 2873 * sizeof(auxtrace_info_event::reserved) == 8 2874 */ 2875 info_header_size = 8; 2876 2877 if (total_size < (event_header_size + info_header_size)) 2878 return -EINVAL; 2879 2880 priv_size = total_size - event_header_size - info_header_size; 2881 2882 /* First the global part */ 2883 ptr = (u64 *) auxtrace_info->priv; 2884 2885 /* Look for version of the header */ 2886 hdr_version = ptr[0]; 2887 if (hdr_version > CS_HEADER_CURRENT_VERSION) { 2888 /* print routine will print an error on bad version */ 2889 if (dump_trace) 2890 cs_etm__print_auxtrace_info(auxtrace_info->priv, 0); 2891 return -EINVAL; 2892 } 2893 2894 hdr = zalloc(sizeof(*hdr) * CS_HEADER_VERSION_MAX); 2895 if (!hdr) 2896 return -ENOMEM; 2897 2898 /* Extract header information - see cs-etm.h for format */ 2899 for (i = 0; i < CS_HEADER_VERSION_MAX; i++) 2900 hdr[i] = ptr[i]; 2901 num_cpu = hdr[CS_PMU_TYPE_CPUS] & 0xffffffff; 2902 pmu_type = (unsigned int) ((hdr[CS_PMU_TYPE_CPUS] >> 32) & 2903 0xffffffff); 2904 2905 /* 2906 * Create an RB tree for traceID-metadata tuple. Since the conversion 2907 * has to be made for each packet that gets decoded, optimizing access 2908 * in anything other than a sequential array is worth doing. 2909 */ 2910 traceid_list = intlist__new(NULL); 2911 if (!traceid_list) { 2912 err = -ENOMEM; 2913 goto err_free_hdr; 2914 } 2915 2916 metadata = zalloc(sizeof(*metadata) * num_cpu); 2917 if (!metadata) { 2918 err = -ENOMEM; 2919 goto err_free_traceid_list; 2920 } 2921 2922 /* 2923 * The metadata is stored in the auxtrace_info section and encodes 2924 * the configuration of the ARM embedded trace macrocell which is 2925 * required by the trace decoder to properly decode the trace due 2926 * to its highly compressed nature. 2927 */ 2928 for (j = 0; j < num_cpu; j++) { 2929 if (ptr[i] == __perf_cs_etmv3_magic) { 2930 metadata[j] = 2931 cs_etm__create_meta_blk(ptr, &i, 2932 CS_ETM_PRIV_MAX, 2933 CS_ETM_NR_TRC_PARAMS_V0); 2934 2935 /* The traceID is our handle */ 2936 trcidr_idx = CS_ETM_ETMTRACEIDR; 2937 2938 } else if (ptr[i] == __perf_cs_etmv4_magic) { 2939 metadata[j] = 2940 cs_etm__create_meta_blk(ptr, &i, 2941 CS_ETMV4_PRIV_MAX, 2942 CS_ETMV4_NR_TRC_PARAMS_V0); 2943 2944 /* The traceID is our handle */ 2945 trcidr_idx = CS_ETMV4_TRCTRACEIDR; 2946 } 2947 2948 if (!metadata[j]) { 2949 err = -ENOMEM; 2950 goto err_free_metadata; 2951 } 2952 2953 /* Get an RB node for this CPU */ 2954 inode = intlist__findnew(traceid_list, metadata[j][trcidr_idx]); 2955 2956 /* Something went wrong, no need to continue */ 2957 if (!inode) { 2958 err = -ENOMEM; 2959 goto err_free_metadata; 2960 } 2961 2962 /* 2963 * The node for that CPU should not be taken. 2964 * Back out if that's the case. 2965 */ 2966 if (inode->priv) { 2967 err = -EINVAL; 2968 goto err_free_metadata; 2969 } 2970 /* All good, associate the traceID with the metadata pointer */ 2971 inode->priv = metadata[j]; 2972 } 2973 2974 /* 2975 * Each of CS_HEADER_VERSION_MAX, CS_ETM_PRIV_MAX and 2976 * CS_ETMV4_PRIV_MAX mark how many double words are in the 2977 * global metadata, and each cpu's metadata respectively. 2978 * The following tests if the correct number of double words was 2979 * present in the auxtrace info section. 2980 */ 2981 if (i * 8 != priv_size) { 2982 err = -EINVAL; 2983 goto err_free_metadata; 2984 } 2985 2986 etm = zalloc(sizeof(*etm)); 2987 2988 if (!etm) { 2989 err = -ENOMEM; 2990 goto err_free_metadata; 2991 } 2992 2993 err = auxtrace_queues__init(&etm->queues); 2994 if (err) 2995 goto err_free_etm; 2996 2997 if (session->itrace_synth_opts->set) { 2998 etm->synth_opts = *session->itrace_synth_opts; 2999 } else { 3000 itrace_synth_opts__set_default(&etm->synth_opts, 3001 session->itrace_synth_opts->default_no_sample); 3002 etm->synth_opts.callchain = false; 3003 } 3004 3005 etm->session = session; 3006 etm->machine = &session->machines.host; 3007 3008 etm->num_cpu = num_cpu; 3009 etm->pmu_type = pmu_type; 3010 etm->snapshot_mode = (hdr[CS_ETM_SNAPSHOT] != 0); 3011 etm->metadata = metadata; 3012 etm->auxtrace_type = auxtrace_info->type; 3013 etm->timeless_decoding = cs_etm__is_timeless_decoding(etm); 3014 3015 etm->auxtrace.process_event = cs_etm__process_event; 3016 etm->auxtrace.process_auxtrace_event = cs_etm__process_auxtrace_event; 3017 etm->auxtrace.flush_events = cs_etm__flush_events; 3018 etm->auxtrace.free_events = cs_etm__free_events; 3019 etm->auxtrace.free = cs_etm__free; 3020 etm->auxtrace.evsel_is_auxtrace = cs_etm__evsel_is_auxtrace; 3021 session->auxtrace = &etm->auxtrace; 3022 3023 etm->unknown_thread = thread__new(999999999, 999999999); 3024 if (!etm->unknown_thread) { 3025 err = -ENOMEM; 3026 goto err_free_queues; 3027 } 3028 3029 /* 3030 * Initialize list node so that at thread__zput() we can avoid 3031 * segmentation fault at list_del_init(). 3032 */ 3033 INIT_LIST_HEAD(&etm->unknown_thread->node); 3034 3035 err = thread__set_comm(etm->unknown_thread, "unknown", 0); 3036 if (err) 3037 goto err_delete_thread; 3038 3039 if (thread__init_maps(etm->unknown_thread, etm->machine)) { 3040 err = -ENOMEM; 3041 goto err_delete_thread; 3042 } 3043 3044 if (dump_trace) { 3045 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu); 3046 } 3047 3048 err = cs_etm__synth_events(etm, session); 3049 if (err) 3050 goto err_delete_thread; 3051 3052 err = cs_etm__queue_aux_records(session); 3053 if (err) 3054 goto err_delete_thread; 3055 3056 etm->data_queued = etm->queues.populated; 3057 /* 3058 * Print warning in pipe mode, see cs_etm__process_auxtrace_event() and 3059 * cs_etm__queue_aux_fragment() for details relating to limitations. 3060 */ 3061 if (!etm->data_queued) 3062 pr_warning("CS ETM warning: Coresight decode and TRBE support requires random file access.\n" 3063 "Continuing with best effort decoding in piped mode.\n\n"); 3064 3065 return 0; 3066 3067 err_delete_thread: 3068 thread__zput(etm->unknown_thread); 3069 err_free_queues: 3070 auxtrace_queues__free(&etm->queues); 3071 session->auxtrace = NULL; 3072 err_free_etm: 3073 zfree(&etm); 3074 err_free_metadata: 3075 /* No need to check @metadata[j], free(NULL) is supported */ 3076 for (j = 0; j < num_cpu; j++) 3077 zfree(&metadata[j]); 3078 zfree(&metadata); 3079 err_free_traceid_list: 3080 intlist__delete(traceid_list); 3081 err_free_hdr: 3082 zfree(&hdr); 3083 /* 3084 * At this point, as a minimum we have valid header. Dump the rest of 3085 * the info section - the print routines will error out on structural 3086 * issues. 3087 */ 3088 if (dump_trace) 3089 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu); 3090 return err; 3091 } 3092