1 #include <sys/sysmacros.h>
13 #include <linux/stringify.h>
27 #include "../builtin.h"
29 #include "sane_ctype.h"
32 struct perf_data_file *output;
33 struct perf_session *session;
34 struct machine *machine;
35 union jr_entry *entry;
40 bool needs_bswap; /* handles cross-endianess */
41 bool use_arch_timestamp;
44 uint64_t unwinding_size;
45 uint64_t unwinding_mapped_size;
46 uint64_t eh_frame_hdr_size;
47 size_t nr_debug_entries;
48 uint32_t code_load_count;
50 struct rb_root code_root;
54 struct debug_line_info {
57 /* The filename format is unspecified, absolute path, relative etc. */
58 char const filename[0];
62 struct perf_tool tool;
63 struct perf_data_file output;
64 struct perf_data_file input;
68 #define hmax(a, b) ((a) > (b) ? (a) : (b))
69 #define get_jit_tool(t) (container_of(tool, struct jit_tool, tool))
72 jit_emit_elf(char *filename,
80 uint32_t unwinding_header_size,
81 uint32_t unwinding_size)
86 fprintf(stderr, "write ELF image %s\n", filename);
88 fd = open(filename, O_CREAT|O_TRUNC|O_WRONLY, 0644);
90 pr_warning("cannot create jit ELF %s: %s\n", filename, strerror(errno));
94 ret = jit_write_elf(fd, code_addr, sym, (const void *)code, csize, debug, nr_debug_entries,
95 unwinding, unwinding_header_size, unwinding_size);
106 jit_close(struct jit_buf_desc *jd)
116 jit_validate_events(struct perf_session *session)
118 struct perf_evsel *evsel;
121 * check that all events use CLOCK_MONOTONIC
123 evlist__for_each_entry(session->evlist, evsel) {
124 if (evsel->attr.use_clockid == 0 || evsel->attr.clockid != CLOCK_MONOTONIC)
131 jit_open(struct jit_buf_desc *jd, const char *name)
133 struct jitheader header;
134 struct jr_prefix *prefix;
136 void *n, *buf = NULL;
137 int ret, retval = -1;
139 jd->in = fopen(name, "r");
143 bsz = hmax(sizeof(header), sizeof(*prefix));
150 * protect from writer modifying the file while we are reading it
154 ret = fread(buf, sizeof(header), 1, jd->in);
158 memcpy(&header, buf, sizeof(header));
160 if (header.magic != JITHEADER_MAGIC) {
161 if (header.magic != JITHEADER_MAGIC_SW)
163 jd->needs_bswap = true;
166 if (jd->needs_bswap) {
167 header.version = bswap_32(header.version);
168 header.total_size = bswap_32(header.total_size);
169 header.pid = bswap_32(header.pid);
170 header.elf_mach = bswap_32(header.elf_mach);
171 header.timestamp = bswap_64(header.timestamp);
172 header.flags = bswap_64(header.flags);
175 jd->use_arch_timestamp = header.flags & JITDUMP_FLAGS_ARCH_TIMESTAMP;
178 pr_debug("version=%u\nhdr.size=%u\nts=0x%llx\npid=%d\nelf_mach=%d\nuse_arch_timestamp=%d\n",
181 (unsigned long long)header.timestamp,
184 jd->use_arch_timestamp);
186 if (header.version > JITHEADER_VERSION) {
187 pr_err("wrong jitdump version %u, expected " __stringify(JITHEADER_VERSION),
192 if (header.flags & JITDUMP_FLAGS_RESERVED) {
193 pr_err("jitdump file contains invalid or unsupported flags 0x%llx\n",
194 (unsigned long long)header.flags & JITDUMP_FLAGS_RESERVED);
198 if (jd->use_arch_timestamp && !jd->session->time_conv.time_mult) {
199 pr_err("jitdump file uses arch timestamps but there is no timestamp conversion\n");
204 * validate event is using the correct clockid
206 if (!jd->use_arch_timestamp && jit_validate_events(jd->session)) {
207 pr_err("error, jitted code must be sampled with perf record -k 1\n");
211 bs = header.total_size - sizeof(header);
214 n = realloc(buf, bs);
219 /* read extra we do not know about */
220 ret = fread(buf, bs - bsz, 1, jd->in);
225 * keep dirname for generating files and mmap records
227 strcpy(jd->dir, name);
237 static union jr_entry *
238 jit_get_next_entry(struct jit_buf_desc *jd)
240 struct jr_prefix *prefix;
249 if (jd->buf == NULL) {
250 size_t sz = getpagesize();
251 if (sz < sizeof(*prefix))
252 sz = sizeof(*prefix);
254 jd->buf = malloc(sz);
264 * file is still locked at this point
266 ret = fread(prefix, sizeof(*prefix), 1, jd->in);
270 if (jd->needs_bswap) {
271 prefix->id = bswap_32(prefix->id);
272 prefix->total_size = bswap_32(prefix->total_size);
273 prefix->timestamp = bswap_64(prefix->timestamp);
276 size = prefix->total_size;
279 if (bs < sizeof(*prefix))
282 if (id >= JIT_CODE_MAX) {
283 pr_warning("next_entry: unknown record type %d, skipping\n", id);
285 if (bs > jd->bufsize) {
287 n = realloc(jd->buf, bs);
294 addr = ((void *)jd->buf) + sizeof(*prefix);
296 ret = fread(addr, bs - sizeof(*prefix), 1, jd->in);
300 jr = (union jr_entry *)jd->buf;
303 case JIT_CODE_DEBUG_INFO:
304 if (jd->needs_bswap) {
306 jr->info.code_addr = bswap_64(jr->info.code_addr);
307 jr->info.nr_entry = bswap_64(jr->info.nr_entry);
308 for (n = 0 ; n < jr->info.nr_entry; n++) {
309 jr->info.entries[n].addr = bswap_64(jr->info.entries[n].addr);
310 jr->info.entries[n].lineno = bswap_32(jr->info.entries[n].lineno);
311 jr->info.entries[n].discrim = bswap_32(jr->info.entries[n].discrim);
315 case JIT_CODE_UNWINDING_INFO:
316 if (jd->needs_bswap) {
317 jr->unwinding.unwinding_size = bswap_64(jr->unwinding.unwinding_size);
318 jr->unwinding.eh_frame_hdr_size = bswap_64(jr->unwinding.eh_frame_hdr_size);
319 jr->unwinding.mapped_size = bswap_64(jr->unwinding.mapped_size);
325 if (jd->needs_bswap) {
326 jr->load.pid = bswap_32(jr->load.pid);
327 jr->load.tid = bswap_32(jr->load.tid);
328 jr->load.vma = bswap_64(jr->load.vma);
329 jr->load.code_addr = bswap_64(jr->load.code_addr);
330 jr->load.code_size = bswap_64(jr->load.code_size);
331 jr->load.code_index= bswap_64(jr->load.code_index);
333 jd->code_load_count++;
336 if (jd->needs_bswap) {
337 jr->move.pid = bswap_32(jr->move.pid);
338 jr->move.tid = bswap_32(jr->move.tid);
339 jr->move.vma = bswap_64(jr->move.vma);
340 jr->move.old_code_addr = bswap_64(jr->move.old_code_addr);
341 jr->move.new_code_addr = bswap_64(jr->move.new_code_addr);
342 jr->move.code_size = bswap_64(jr->move.code_size);
343 jr->move.code_index = bswap_64(jr->move.code_index);
348 /* skip unknown record (we have read them) */
355 jit_inject_event(struct jit_buf_desc *jd, union perf_event *event)
359 size = perf_data_file__write(jd->output, event, event->header.size);
363 jd->bytes_written += size;
367 static uint64_t convert_timestamp(struct jit_buf_desc *jd, uint64_t timestamp)
369 struct perf_tsc_conversion tc;
371 if (!jd->use_arch_timestamp)
374 tc.time_shift = jd->session->time_conv.time_shift;
375 tc.time_mult = jd->session->time_conv.time_mult;
376 tc.time_zero = jd->session->time_conv.time_zero;
381 return tsc_to_perf_time(timestamp, &tc);
384 static int jit_repipe_code_load(struct jit_buf_desc *jd, union jr_entry *jr)
386 struct perf_sample sample;
387 union perf_event *event;
388 struct perf_tool *tool = jd->session->tool;
397 int ret, csize, usize;
406 csize = jr->load.code_size;
407 usize = jd->unwinding_mapped_size;
408 addr = jr->load.code_addr;
409 sym = (void *)((unsigned long)jr + sizeof(jr->load));
410 code = (unsigned long)jr + jr->load.p.total_size - csize;
411 count = jr->load.code_index;
412 idr_size = jd->machine->id_hdr_size;
414 event = calloc(1, sizeof(*event) + idr_size);
418 filename = event->mmap2.filename;
419 size = snprintf(filename, PATH_MAX, "%s/jitted-%d-%u.so",
426 size = PERF_ALIGN(size, sizeof(u64));
427 uaddr = (uintptr_t)code;
428 ret = jit_emit_elf(filename, sym, addr, (const void *)uaddr, csize, jd->debug_data, jd->nr_debug_entries,
429 jd->unwinding_data, jd->eh_frame_hdr_size, jd->unwinding_size);
431 if (jd->debug_data && jd->nr_debug_entries) {
432 free(jd->debug_data);
433 jd->debug_data = NULL;
434 jd->nr_debug_entries = 0;
437 if (jd->unwinding_data && jd->eh_frame_hdr_size) {
438 free(jd->unwinding_data);
439 jd->unwinding_data = NULL;
440 jd->eh_frame_hdr_size = 0;
441 jd->unwinding_mapped_size = 0;
442 jd->unwinding_size = 0;
449 if (stat(filename, &st))
450 memset(&st, 0, sizeof(st));
452 event->mmap2.header.type = PERF_RECORD_MMAP2;
453 event->mmap2.header.misc = PERF_RECORD_MISC_USER;
454 event->mmap2.header.size = (sizeof(event->mmap2) -
455 (sizeof(event->mmap2.filename) - size) + idr_size);
457 event->mmap2.pgoff = GEN_ELF_TEXT_OFFSET;
458 event->mmap2.start = addr;
459 event->mmap2.len = usize ? ALIGN_8(csize) + usize : csize;
460 event->mmap2.pid = pid;
461 event->mmap2.tid = tid;
462 event->mmap2.ino = st.st_ino;
463 event->mmap2.maj = major(st.st_dev);
464 event->mmap2.min = minor(st.st_dev);
465 event->mmap2.prot = st.st_mode;
466 event->mmap2.flags = MAP_SHARED;
467 event->mmap2.ino_generation = 1;
469 id = (void *)((unsigned long)event + event->mmap.header.size - idr_size);
470 if (jd->sample_type & PERF_SAMPLE_TID) {
474 if (jd->sample_type & PERF_SAMPLE_TIME)
475 id->time = convert_timestamp(jd, jr->load.p.timestamp);
478 * create pseudo sample to induce dso hit increment
479 * use first address as sample address
481 memset(&sample, 0, sizeof(sample));
482 sample.cpumode = PERF_RECORD_MISC_USER;
485 sample.time = id->time;
488 ret = perf_event__process_mmap2(tool, event, &sample, jd->machine);
492 ret = jit_inject_event(jd, event);
494 * mark dso as use to generate buildid in the header
497 build_id__mark_dso_hit(tool, event, &sample, NULL, jd->machine);
502 static int jit_repipe_code_move(struct jit_buf_desc *jd, union jr_entry *jr)
504 struct perf_sample sample;
505 union perf_event *event;
506 struct perf_tool *tool = jd->session->tool;
521 usize = jd->unwinding_mapped_size;
522 idr_size = jd->machine->id_hdr_size;
525 * +16 to account for sample_id_all (hack)
527 event = calloc(1, sizeof(*event) + 16);
531 filename = event->mmap2.filename;
532 size = snprintf(filename, PATH_MAX, "%s/jitted-%d-%"PRIu64,
535 jr->move.code_index);
539 if (stat(filename, &st))
540 memset(&st, 0, sizeof(st));
542 size = PERF_ALIGN(size, sizeof(u64));
544 event->mmap2.header.type = PERF_RECORD_MMAP2;
545 event->mmap2.header.misc = PERF_RECORD_MISC_USER;
546 event->mmap2.header.size = (sizeof(event->mmap2) -
547 (sizeof(event->mmap2.filename) - size) + idr_size);
548 event->mmap2.pgoff = GEN_ELF_TEXT_OFFSET;
549 event->mmap2.start = jr->move.new_code_addr;
550 event->mmap2.len = usize ? ALIGN_8(jr->move.code_size) + usize
551 : jr->move.code_size;
552 event->mmap2.pid = pid;
553 event->mmap2.tid = tid;
554 event->mmap2.ino = st.st_ino;
555 event->mmap2.maj = major(st.st_dev);
556 event->mmap2.min = minor(st.st_dev);
557 event->mmap2.prot = st.st_mode;
558 event->mmap2.flags = MAP_SHARED;
559 event->mmap2.ino_generation = 1;
561 id = (void *)((unsigned long)event + event->mmap.header.size - idr_size);
562 if (jd->sample_type & PERF_SAMPLE_TID) {
566 if (jd->sample_type & PERF_SAMPLE_TIME)
567 id->time = convert_timestamp(jd, jr->load.p.timestamp);
570 * create pseudo sample to induce dso hit increment
571 * use first address as sample address
573 memset(&sample, 0, sizeof(sample));
574 sample.cpumode = PERF_RECORD_MISC_USER;
577 sample.time = id->time;
578 sample.ip = jr->move.new_code_addr;
580 ret = perf_event__process_mmap2(tool, event, &sample, jd->machine);
584 ret = jit_inject_event(jd, event);
586 build_id__mark_dso_hit(tool, event, &sample, NULL, jd->machine);
591 static int jit_repipe_debug_info(struct jit_buf_desc *jd, union jr_entry *jr)
599 sz = jr->prefix.total_size - sizeof(jr->info);
604 memcpy(data, &jr->info.entries, sz);
606 jd->debug_data = data;
609 * we must use nr_entry instead of size here because
610 * we cannot distinguish actual entry from padding otherwise
612 jd->nr_debug_entries = jr->info.nr_entry;
618 jit_repipe_unwinding_info(struct jit_buf_desc *jd, union jr_entry *jr)
620 void *unwinding_data;
621 uint32_t unwinding_data_size;
626 unwinding_data_size = jr->prefix.total_size - sizeof(jr->unwinding);
627 unwinding_data = malloc(unwinding_data_size);
631 memcpy(unwinding_data, &jr->unwinding.unwinding_data,
632 unwinding_data_size);
634 jd->eh_frame_hdr_size = jr->unwinding.eh_frame_hdr_size;
635 jd->unwinding_size = jr->unwinding.unwinding_size;
636 jd->unwinding_mapped_size = jr->unwinding.mapped_size;
637 jd->unwinding_data = unwinding_data;
643 jit_process_dump(struct jit_buf_desc *jd)
648 while ((jr = jit_get_next_entry(jd))) {
649 switch(jr->prefix.id) {
651 ret = jit_repipe_code_load(jd, jr);
654 ret = jit_repipe_code_move(jd, jr);
656 case JIT_CODE_DEBUG_INFO:
657 ret = jit_repipe_debug_info(jd, jr);
659 case JIT_CODE_UNWINDING_INFO:
660 ret = jit_repipe_unwinding_info(jd, jr);
671 jit_inject(struct jit_buf_desc *jd, char *path)
676 fprintf(stderr, "injecting: %s\n", path);
678 ret = jit_open(jd, path);
682 ret = jit_process_dump(jd);
687 fprintf(stderr, "injected: %s (%d)\n", path, ret);
693 * File must be with pattern .../jit-XXXX.dump
694 * where XXXX is the PID of the process which did the mmap()
695 * as captured in the RECORD_MMAP record
698 jit_detect(char *mmap_name, pid_t pid)
705 fprintf(stderr, "jit marker trying : %s\n", mmap_name);
709 p = strrchr(mmap_name, '/');
716 if (strncmp(p, "/jit-", 5))
725 * must be followed by a pid
730 pid2 = (int)strtol(p, &end, 10);
735 * pid does not match mmap pid
736 * pid==0 in system-wide mode (synthesized)
738 if (pid && pid2 != pid)
743 if (strcmp(end, ".dump"))
747 fprintf(stderr, "jit marker found: %s\n", mmap_name);
753 jit_process(struct perf_session *session,
754 struct perf_data_file *output,
755 struct machine *machine,
760 struct perf_evsel *first;
761 struct jit_buf_desc jd;
765 * first, detect marker mmap (i.e., the jitdump mmap)
767 if (jit_detect(filename, pid))
770 memset(&jd, 0, sizeof(jd));
772 jd.session = session;
774 jd.machine = machine;
777 * track sample_type to compute id_all layout
778 * perf sets the same sample type to all events as of now
780 first = perf_evlist__first(session->evlist);
781 jd.sample_type = first->attr.sample_type;
785 ret = jit_inject(&jd, filename);
787 *nbytes = jd.bytes_written;