Merge branch 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/evalenti/linux...
[linux/fpc-iii.git] / tools / perf / util / header.c
blob90680ec9f8b89fa3cc0964df15585fbe92246a46
1 #include "util.h"
2 #include <sys/types.h>
3 #include <byteswap.h>
4 #include <unistd.h>
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <linux/list.h>
8 #include <linux/kernel.h>
9 #include <linux/bitops.h>
10 #include <sys/utsname.h>
12 #include "evlist.h"
13 #include "evsel.h"
14 #include "header.h"
15 #include "../perf.h"
16 #include "trace-event.h"
17 #include "session.h"
18 #include "symbol.h"
19 #include "debug.h"
20 #include "cpumap.h"
21 #include "pmu.h"
22 #include "vdso.h"
23 #include "strbuf.h"
24 #include "build-id.h"
25 #include "data.h"
26 #include <api/fs/fs.h>
27 #include "asm/bug.h"
30 * magic2 = "PERFILE2"
31 * must be a numerical value to let the endianness
32 * determine the memory layout. That way we are able
33 * to detect endianness when reading the perf.data file
34 * back.
36 * we check for legacy (PERFFILE) format.
38 static const char *__perf_magic1 = "PERFFILE";
39 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
40 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
42 #define PERF_MAGIC __perf_magic2
44 struct perf_file_attr {
45 struct perf_event_attr attr;
46 struct perf_file_section ids;
49 void perf_header__set_feat(struct perf_header *header, int feat)
51 set_bit(feat, header->adds_features);
54 void perf_header__clear_feat(struct perf_header *header, int feat)
56 clear_bit(feat, header->adds_features);
59 bool perf_header__has_feat(const struct perf_header *header, int feat)
61 return test_bit(feat, header->adds_features);
64 static int do_write(int fd, const void *buf, size_t size)
66 while (size) {
67 int ret = write(fd, buf, size);
69 if (ret < 0)
70 return -errno;
72 size -= ret;
73 buf += ret;
76 return 0;
79 int write_padded(int fd, const void *bf, size_t count, size_t count_aligned)
81 static const char zero_buf[NAME_ALIGN];
82 int err = do_write(fd, bf, count);
84 if (!err)
85 err = do_write(fd, zero_buf, count_aligned - count);
87 return err;
90 #define string_size(str) \
91 (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
93 static int do_write_string(int fd, const char *str)
95 u32 len, olen;
96 int ret;
98 olen = strlen(str) + 1;
99 len = PERF_ALIGN(olen, NAME_ALIGN);
101 /* write len, incl. \0 */
102 ret = do_write(fd, &len, sizeof(len));
103 if (ret < 0)
104 return ret;
106 return write_padded(fd, str, olen, len);
109 static char *do_read_string(int fd, struct perf_header *ph)
111 ssize_t sz, ret;
112 u32 len;
113 char *buf;
115 sz = readn(fd, &len, sizeof(len));
116 if (sz < (ssize_t)sizeof(len))
117 return NULL;
119 if (ph->needs_swap)
120 len = bswap_32(len);
122 buf = malloc(len);
123 if (!buf)
124 return NULL;
126 ret = readn(fd, buf, len);
127 if (ret == (ssize_t)len) {
129 * strings are padded by zeroes
130 * thus the actual strlen of buf
131 * may be less than len
133 return buf;
136 free(buf);
137 return NULL;
140 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
141 struct perf_evlist *evlist)
143 return read_tracing_data(fd, &evlist->entries);
147 static int write_build_id(int fd, struct perf_header *h,
148 struct perf_evlist *evlist __maybe_unused)
150 struct perf_session *session;
151 int err;
153 session = container_of(h, struct perf_session, header);
155 if (!perf_session__read_build_ids(session, true))
156 return -1;
158 err = perf_session__write_buildid_table(session, fd);
159 if (err < 0) {
160 pr_debug("failed to write buildid table\n");
161 return err;
163 perf_session__cache_build_ids(session);
165 return 0;
168 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
169 struct perf_evlist *evlist __maybe_unused)
171 struct utsname uts;
172 int ret;
174 ret = uname(&uts);
175 if (ret < 0)
176 return -1;
178 return do_write_string(fd, uts.nodename);
181 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
182 struct perf_evlist *evlist __maybe_unused)
184 struct utsname uts;
185 int ret;
187 ret = uname(&uts);
188 if (ret < 0)
189 return -1;
191 return do_write_string(fd, uts.release);
194 static int write_arch(int fd, struct perf_header *h __maybe_unused,
195 struct perf_evlist *evlist __maybe_unused)
197 struct utsname uts;
198 int ret;
200 ret = uname(&uts);
201 if (ret < 0)
202 return -1;
204 return do_write_string(fd, uts.machine);
207 static int write_version(int fd, struct perf_header *h __maybe_unused,
208 struct perf_evlist *evlist __maybe_unused)
210 return do_write_string(fd, perf_version_string);
213 static int __write_cpudesc(int fd, const char *cpuinfo_proc)
215 FILE *file;
216 char *buf = NULL;
217 char *s, *p;
218 const char *search = cpuinfo_proc;
219 size_t len = 0;
220 int ret = -1;
222 if (!search)
223 return -1;
225 file = fopen("/proc/cpuinfo", "r");
226 if (!file)
227 return -1;
229 while (getline(&buf, &len, file) > 0) {
230 ret = strncmp(buf, search, strlen(search));
231 if (!ret)
232 break;
235 if (ret) {
236 ret = -1;
237 goto done;
240 s = buf;
242 p = strchr(buf, ':');
243 if (p && *(p+1) == ' ' && *(p+2))
244 s = p + 2;
245 p = strchr(s, '\n');
246 if (p)
247 *p = '\0';
249 /* squash extra space characters (branding string) */
250 p = s;
251 while (*p) {
252 if (isspace(*p)) {
253 char *r = p + 1;
254 char *q = r;
255 *p = ' ';
256 while (*q && isspace(*q))
257 q++;
258 if (q != (p+1))
259 while ((*r++ = *q++));
261 p++;
263 ret = do_write_string(fd, s);
264 done:
265 free(buf);
266 fclose(file);
267 return ret;
270 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
271 struct perf_evlist *evlist __maybe_unused)
273 #ifndef CPUINFO_PROC
274 #define CPUINFO_PROC {"model name", }
275 #endif
276 const char *cpuinfo_procs[] = CPUINFO_PROC;
277 unsigned int i;
279 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
280 int ret;
281 ret = __write_cpudesc(fd, cpuinfo_procs[i]);
282 if (ret >= 0)
283 return ret;
285 return -1;
289 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
290 struct perf_evlist *evlist __maybe_unused)
292 long nr;
293 u32 nrc, nra;
294 int ret;
296 nr = sysconf(_SC_NPROCESSORS_CONF);
297 if (nr < 0)
298 return -1;
300 nrc = (u32)(nr & UINT_MAX);
302 nr = sysconf(_SC_NPROCESSORS_ONLN);
303 if (nr < 0)
304 return -1;
306 nra = (u32)(nr & UINT_MAX);
308 ret = do_write(fd, &nrc, sizeof(nrc));
309 if (ret < 0)
310 return ret;
312 return do_write(fd, &nra, sizeof(nra));
315 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
316 struct perf_evlist *evlist)
318 struct perf_evsel *evsel;
319 u32 nre, nri, sz;
320 int ret;
322 nre = evlist->nr_entries;
325 * write number of events
327 ret = do_write(fd, &nre, sizeof(nre));
328 if (ret < 0)
329 return ret;
332 * size of perf_event_attr struct
334 sz = (u32)sizeof(evsel->attr);
335 ret = do_write(fd, &sz, sizeof(sz));
336 if (ret < 0)
337 return ret;
339 evlist__for_each(evlist, evsel) {
340 ret = do_write(fd, &evsel->attr, sz);
341 if (ret < 0)
342 return ret;
344 * write number of unique id per event
345 * there is one id per instance of an event
347 * copy into an nri to be independent of the
348 * type of ids,
350 nri = evsel->ids;
351 ret = do_write(fd, &nri, sizeof(nri));
352 if (ret < 0)
353 return ret;
356 * write event string as passed on cmdline
358 ret = do_write_string(fd, perf_evsel__name(evsel));
359 if (ret < 0)
360 return ret;
362 * write unique ids for this event
364 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
365 if (ret < 0)
366 return ret;
368 return 0;
371 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
372 struct perf_evlist *evlist __maybe_unused)
374 char buf[MAXPATHLEN];
375 char proc[32];
376 u32 n;
377 int i, ret;
380 * actual atual path to perf binary
382 sprintf(proc, "/proc/%d/exe", getpid());
383 ret = readlink(proc, buf, sizeof(buf));
384 if (ret <= 0)
385 return -1;
387 /* readlink() does not add null termination */
388 buf[ret] = '\0';
390 /* account for binary path */
391 n = perf_env.nr_cmdline + 1;
393 ret = do_write(fd, &n, sizeof(n));
394 if (ret < 0)
395 return ret;
397 ret = do_write_string(fd, buf);
398 if (ret < 0)
399 return ret;
401 for (i = 0 ; i < perf_env.nr_cmdline; i++) {
402 ret = do_write_string(fd, perf_env.cmdline_argv[i]);
403 if (ret < 0)
404 return ret;
406 return 0;
409 #define CORE_SIB_FMT \
410 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
411 #define THRD_SIB_FMT \
412 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
414 struct cpu_topo {
415 u32 cpu_nr;
416 u32 core_sib;
417 u32 thread_sib;
418 char **core_siblings;
419 char **thread_siblings;
422 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
424 FILE *fp;
425 char filename[MAXPATHLEN];
426 char *buf = NULL, *p;
427 size_t len = 0;
428 ssize_t sret;
429 u32 i = 0;
430 int ret = -1;
432 sprintf(filename, CORE_SIB_FMT, cpu);
433 fp = fopen(filename, "r");
434 if (!fp)
435 goto try_threads;
437 sret = getline(&buf, &len, fp);
438 fclose(fp);
439 if (sret <= 0)
440 goto try_threads;
442 p = strchr(buf, '\n');
443 if (p)
444 *p = '\0';
446 for (i = 0; i < tp->core_sib; i++) {
447 if (!strcmp(buf, tp->core_siblings[i]))
448 break;
450 if (i == tp->core_sib) {
451 tp->core_siblings[i] = buf;
452 tp->core_sib++;
453 buf = NULL;
454 len = 0;
456 ret = 0;
458 try_threads:
459 sprintf(filename, THRD_SIB_FMT, cpu);
460 fp = fopen(filename, "r");
461 if (!fp)
462 goto done;
464 if (getline(&buf, &len, fp) <= 0)
465 goto done;
467 p = strchr(buf, '\n');
468 if (p)
469 *p = '\0';
471 for (i = 0; i < tp->thread_sib; i++) {
472 if (!strcmp(buf, tp->thread_siblings[i]))
473 break;
475 if (i == tp->thread_sib) {
476 tp->thread_siblings[i] = buf;
477 tp->thread_sib++;
478 buf = NULL;
480 ret = 0;
481 done:
482 if(fp)
483 fclose(fp);
484 free(buf);
485 return ret;
488 static void free_cpu_topo(struct cpu_topo *tp)
490 u32 i;
492 if (!tp)
493 return;
495 for (i = 0 ; i < tp->core_sib; i++)
496 zfree(&tp->core_siblings[i]);
498 for (i = 0 ; i < tp->thread_sib; i++)
499 zfree(&tp->thread_siblings[i]);
501 free(tp);
504 static struct cpu_topo *build_cpu_topology(void)
506 struct cpu_topo *tp;
507 void *addr;
508 u32 nr, i;
509 size_t sz;
510 long ncpus;
511 int ret = -1;
513 ncpus = sysconf(_SC_NPROCESSORS_CONF);
514 if (ncpus < 0)
515 return NULL;
517 nr = (u32)(ncpus & UINT_MAX);
519 sz = nr * sizeof(char *);
521 addr = calloc(1, sizeof(*tp) + 2 * sz);
522 if (!addr)
523 return NULL;
525 tp = addr;
526 tp->cpu_nr = nr;
527 addr += sizeof(*tp);
528 tp->core_siblings = addr;
529 addr += sz;
530 tp->thread_siblings = addr;
532 for (i = 0; i < nr; i++) {
533 ret = build_cpu_topo(tp, i);
534 if (ret < 0)
535 break;
537 if (ret) {
538 free_cpu_topo(tp);
539 tp = NULL;
541 return tp;
544 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
545 struct perf_evlist *evlist __maybe_unused)
547 struct cpu_topo *tp;
548 u32 i;
549 int ret, j;
551 tp = build_cpu_topology();
552 if (!tp)
553 return -1;
555 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
556 if (ret < 0)
557 goto done;
559 for (i = 0; i < tp->core_sib; i++) {
560 ret = do_write_string(fd, tp->core_siblings[i]);
561 if (ret < 0)
562 goto done;
564 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
565 if (ret < 0)
566 goto done;
568 for (i = 0; i < tp->thread_sib; i++) {
569 ret = do_write_string(fd, tp->thread_siblings[i]);
570 if (ret < 0)
571 break;
574 ret = perf_env__read_cpu_topology_map(&perf_env);
575 if (ret < 0)
576 goto done;
578 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
579 ret = do_write(fd, &perf_env.cpu[j].core_id,
580 sizeof(perf_env.cpu[j].core_id));
581 if (ret < 0)
582 return ret;
583 ret = do_write(fd, &perf_env.cpu[j].socket_id,
584 sizeof(perf_env.cpu[j].socket_id));
585 if (ret < 0)
586 return ret;
588 done:
589 free_cpu_topo(tp);
590 return ret;
595 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
596 struct perf_evlist *evlist __maybe_unused)
598 char *buf = NULL;
599 FILE *fp;
600 size_t len = 0;
601 int ret = -1, n;
602 uint64_t mem;
604 fp = fopen("/proc/meminfo", "r");
605 if (!fp)
606 return -1;
608 while (getline(&buf, &len, fp) > 0) {
609 ret = strncmp(buf, "MemTotal:", 9);
610 if (!ret)
611 break;
613 if (!ret) {
614 n = sscanf(buf, "%*s %"PRIu64, &mem);
615 if (n == 1)
616 ret = do_write(fd, &mem, sizeof(mem));
617 } else
618 ret = -1;
619 free(buf);
620 fclose(fp);
621 return ret;
624 static int write_topo_node(int fd, int node)
626 char str[MAXPATHLEN];
627 char field[32];
628 char *buf = NULL, *p;
629 size_t len = 0;
630 FILE *fp;
631 u64 mem_total, mem_free, mem;
632 int ret = -1;
634 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
635 fp = fopen(str, "r");
636 if (!fp)
637 return -1;
639 while (getline(&buf, &len, fp) > 0) {
640 /* skip over invalid lines */
641 if (!strchr(buf, ':'))
642 continue;
643 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
644 goto done;
645 if (!strcmp(field, "MemTotal:"))
646 mem_total = mem;
647 if (!strcmp(field, "MemFree:"))
648 mem_free = mem;
651 fclose(fp);
652 fp = NULL;
654 ret = do_write(fd, &mem_total, sizeof(u64));
655 if (ret)
656 goto done;
658 ret = do_write(fd, &mem_free, sizeof(u64));
659 if (ret)
660 goto done;
662 ret = -1;
663 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
665 fp = fopen(str, "r");
666 if (!fp)
667 goto done;
669 if (getline(&buf, &len, fp) <= 0)
670 goto done;
672 p = strchr(buf, '\n');
673 if (p)
674 *p = '\0';
676 ret = do_write_string(fd, buf);
677 done:
678 free(buf);
679 if (fp)
680 fclose(fp);
681 return ret;
684 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
685 struct perf_evlist *evlist __maybe_unused)
687 char *buf = NULL;
688 size_t len = 0;
689 FILE *fp;
690 struct cpu_map *node_map = NULL;
691 char *c;
692 u32 nr, i, j;
693 int ret = -1;
695 fp = fopen("/sys/devices/system/node/online", "r");
696 if (!fp)
697 return -1;
699 if (getline(&buf, &len, fp) <= 0)
700 goto done;
702 c = strchr(buf, '\n');
703 if (c)
704 *c = '\0';
706 node_map = cpu_map__new(buf);
707 if (!node_map)
708 goto done;
710 nr = (u32)node_map->nr;
712 ret = do_write(fd, &nr, sizeof(nr));
713 if (ret < 0)
714 goto done;
716 for (i = 0; i < nr; i++) {
717 j = (u32)node_map->map[i];
718 ret = do_write(fd, &j, sizeof(j));
719 if (ret < 0)
720 break;
722 ret = write_topo_node(fd, i);
723 if (ret < 0)
724 break;
726 done:
727 free(buf);
728 fclose(fp);
729 cpu_map__put(node_map);
730 return ret;
734 * File format:
736 * struct pmu_mappings {
737 * u32 pmu_num;
738 * struct pmu_map {
739 * u32 type;
740 * char name[];
741 * }[pmu_num];
742 * };
745 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
746 struct perf_evlist *evlist __maybe_unused)
748 struct perf_pmu *pmu = NULL;
749 off_t offset = lseek(fd, 0, SEEK_CUR);
750 __u32 pmu_num = 0;
751 int ret;
753 /* write real pmu_num later */
754 ret = do_write(fd, &pmu_num, sizeof(pmu_num));
755 if (ret < 0)
756 return ret;
758 while ((pmu = perf_pmu__scan(pmu))) {
759 if (!pmu->name)
760 continue;
761 pmu_num++;
763 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
764 if (ret < 0)
765 return ret;
767 ret = do_write_string(fd, pmu->name);
768 if (ret < 0)
769 return ret;
772 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
773 /* discard all */
774 lseek(fd, offset, SEEK_SET);
775 return -1;
778 return 0;
782 * File format:
784 * struct group_descs {
785 * u32 nr_groups;
786 * struct group_desc {
787 * char name[];
788 * u32 leader_idx;
789 * u32 nr_members;
790 * }[nr_groups];
791 * };
793 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
794 struct perf_evlist *evlist)
796 u32 nr_groups = evlist->nr_groups;
797 struct perf_evsel *evsel;
798 int ret;
800 ret = do_write(fd, &nr_groups, sizeof(nr_groups));
801 if (ret < 0)
802 return ret;
804 evlist__for_each(evlist, evsel) {
805 if (perf_evsel__is_group_leader(evsel) &&
806 evsel->nr_members > 1) {
807 const char *name = evsel->group_name ?: "{anon_group}";
808 u32 leader_idx = evsel->idx;
809 u32 nr_members = evsel->nr_members;
811 ret = do_write_string(fd, name);
812 if (ret < 0)
813 return ret;
815 ret = do_write(fd, &leader_idx, sizeof(leader_idx));
816 if (ret < 0)
817 return ret;
819 ret = do_write(fd, &nr_members, sizeof(nr_members));
820 if (ret < 0)
821 return ret;
824 return 0;
828 * default get_cpuid(): nothing gets recorded
829 * actual implementation must be in arch/$(ARCH)/util/header.c
831 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
832 size_t sz __maybe_unused)
834 return -1;
837 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
838 struct perf_evlist *evlist __maybe_unused)
840 char buffer[64];
841 int ret;
843 ret = get_cpuid(buffer, sizeof(buffer));
844 if (!ret)
845 goto write_it;
847 return -1;
848 write_it:
849 return do_write_string(fd, buffer);
852 static int write_branch_stack(int fd __maybe_unused,
853 struct perf_header *h __maybe_unused,
854 struct perf_evlist *evlist __maybe_unused)
856 return 0;
859 static int write_auxtrace(int fd, struct perf_header *h,
860 struct perf_evlist *evlist __maybe_unused)
862 struct perf_session *session;
863 int err;
865 session = container_of(h, struct perf_session, header);
867 err = auxtrace_index__write(fd, &session->auxtrace_index);
868 if (err < 0)
869 pr_err("Failed to write auxtrace index\n");
870 return err;
873 static int cpu_cache_level__sort(const void *a, const void *b)
875 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
876 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
878 return cache_a->level - cache_b->level;
881 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
883 if (a->level != b->level)
884 return false;
886 if (a->line_size != b->line_size)
887 return false;
889 if (a->sets != b->sets)
890 return false;
892 if (a->ways != b->ways)
893 return false;
895 if (strcmp(a->type, b->type))
896 return false;
898 if (strcmp(a->size, b->size))
899 return false;
901 if (strcmp(a->map, b->map))
902 return false;
904 return true;
907 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
909 char path[PATH_MAX], file[PATH_MAX];
910 struct stat st;
911 size_t len;
913 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
914 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
916 if (stat(file, &st))
917 return 1;
919 scnprintf(file, PATH_MAX, "%s/level", path);
920 if (sysfs__read_int(file, (int *) &cache->level))
921 return -1;
923 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
924 if (sysfs__read_int(file, (int *) &cache->line_size))
925 return -1;
927 scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
928 if (sysfs__read_int(file, (int *) &cache->sets))
929 return -1;
931 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
932 if (sysfs__read_int(file, (int *) &cache->ways))
933 return -1;
935 scnprintf(file, PATH_MAX, "%s/type", path);
936 if (sysfs__read_str(file, &cache->type, &len))
937 return -1;
939 cache->type[len] = 0;
940 cache->type = rtrim(cache->type);
942 scnprintf(file, PATH_MAX, "%s/size", path);
943 if (sysfs__read_str(file, &cache->size, &len)) {
944 free(cache->type);
945 return -1;
948 cache->size[len] = 0;
949 cache->size = rtrim(cache->size);
951 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
952 if (sysfs__read_str(file, &cache->map, &len)) {
953 free(cache->map);
954 free(cache->type);
955 return -1;
958 cache->map[len] = 0;
959 cache->map = rtrim(cache->map);
960 return 0;
963 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
965 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
968 static int build_caches(struct cpu_cache_level caches[], u32 size, u32 *cntp)
970 u32 i, cnt = 0;
971 long ncpus;
972 u32 nr, cpu;
973 u16 level;
975 ncpus = sysconf(_SC_NPROCESSORS_CONF);
976 if (ncpus < 0)
977 return -1;
979 nr = (u32)(ncpus & UINT_MAX);
981 for (cpu = 0; cpu < nr; cpu++) {
982 for (level = 0; level < 10; level++) {
983 struct cpu_cache_level c;
984 int err;
986 err = cpu_cache_level__read(&c, cpu, level);
987 if (err < 0)
988 return err;
990 if (err == 1)
991 break;
993 for (i = 0; i < cnt; i++) {
994 if (cpu_cache_level__cmp(&c, &caches[i]))
995 break;
998 if (i == cnt)
999 caches[cnt++] = c;
1000 else
1001 cpu_cache_level__free(&c);
1003 if (WARN_ONCE(cnt == size, "way too many cpu caches.."))
1004 goto out;
1007 out:
1008 *cntp = cnt;
1009 return 0;
1012 #define MAX_CACHES 2000
1014 static int write_cache(int fd, struct perf_header *h __maybe_unused,
1015 struct perf_evlist *evlist __maybe_unused)
1017 struct cpu_cache_level caches[MAX_CACHES];
1018 u32 cnt = 0, i, version = 1;
1019 int ret;
1021 ret = build_caches(caches, MAX_CACHES, &cnt);
1022 if (ret)
1023 goto out;
1025 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1027 ret = do_write(fd, &version, sizeof(u32));
1028 if (ret < 0)
1029 goto out;
1031 ret = do_write(fd, &cnt, sizeof(u32));
1032 if (ret < 0)
1033 goto out;
1035 for (i = 0; i < cnt; i++) {
1036 struct cpu_cache_level *c = &caches[i];
1038 #define _W(v) \
1039 ret = do_write(fd, &c->v, sizeof(u32)); \
1040 if (ret < 0) \
1041 goto out;
1043 _W(level)
1044 _W(line_size)
1045 _W(sets)
1046 _W(ways)
1047 #undef _W
1049 #define _W(v) \
1050 ret = do_write_string(fd, (const char *) c->v); \
1051 if (ret < 0) \
1052 goto out;
1054 _W(type)
1055 _W(size)
1056 _W(map)
1057 #undef _W
1060 out:
1061 for (i = 0; i < cnt; i++)
1062 cpu_cache_level__free(&caches[i]);
1063 return ret;
1066 static int write_stat(int fd __maybe_unused,
1067 struct perf_header *h __maybe_unused,
1068 struct perf_evlist *evlist __maybe_unused)
1070 return 0;
1073 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
1074 FILE *fp)
1076 fprintf(fp, "# hostname : %s\n", ph->env.hostname);
1079 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
1080 FILE *fp)
1082 fprintf(fp, "# os release : %s\n", ph->env.os_release);
1085 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1087 fprintf(fp, "# arch : %s\n", ph->env.arch);
1090 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
1091 FILE *fp)
1093 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
1096 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
1097 FILE *fp)
1099 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
1100 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
1103 static void print_version(struct perf_header *ph, int fd __maybe_unused,
1104 FILE *fp)
1106 fprintf(fp, "# perf version : %s\n", ph->env.version);
1109 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
1110 FILE *fp)
1112 int nr, i;
1114 nr = ph->env.nr_cmdline;
1116 fprintf(fp, "# cmdline : ");
1118 for (i = 0; i < nr; i++)
1119 fprintf(fp, "%s ", ph->env.cmdline_argv[i]);
1120 fputc('\n', fp);
1123 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
1124 FILE *fp)
1126 int nr, i;
1127 char *str;
1128 int cpu_nr = ph->env.nr_cpus_online;
1130 nr = ph->env.nr_sibling_cores;
1131 str = ph->env.sibling_cores;
1133 for (i = 0; i < nr; i++) {
1134 fprintf(fp, "# sibling cores : %s\n", str);
1135 str += strlen(str) + 1;
1138 nr = ph->env.nr_sibling_threads;
1139 str = ph->env.sibling_threads;
1141 for (i = 0; i < nr; i++) {
1142 fprintf(fp, "# sibling threads : %s\n", str);
1143 str += strlen(str) + 1;
1146 if (ph->env.cpu != NULL) {
1147 for (i = 0; i < cpu_nr; i++)
1148 fprintf(fp, "# CPU %d: Core ID %d, Socket ID %d\n", i,
1149 ph->env.cpu[i].core_id, ph->env.cpu[i].socket_id);
1150 } else
1151 fprintf(fp, "# Core ID and Socket ID information is not available\n");
1154 static void free_event_desc(struct perf_evsel *events)
1156 struct perf_evsel *evsel;
1158 if (!events)
1159 return;
1161 for (evsel = events; evsel->attr.size; evsel++) {
1162 zfree(&evsel->name);
1163 zfree(&evsel->id);
1166 free(events);
1169 static struct perf_evsel *
1170 read_event_desc(struct perf_header *ph, int fd)
1172 struct perf_evsel *evsel, *events = NULL;
1173 u64 *id;
1174 void *buf = NULL;
1175 u32 nre, sz, nr, i, j;
1176 ssize_t ret;
1177 size_t msz;
1179 /* number of events */
1180 ret = readn(fd, &nre, sizeof(nre));
1181 if (ret != (ssize_t)sizeof(nre))
1182 goto error;
1184 if (ph->needs_swap)
1185 nre = bswap_32(nre);
1187 ret = readn(fd, &sz, sizeof(sz));
1188 if (ret != (ssize_t)sizeof(sz))
1189 goto error;
1191 if (ph->needs_swap)
1192 sz = bswap_32(sz);
1194 /* buffer to hold on file attr struct */
1195 buf = malloc(sz);
1196 if (!buf)
1197 goto error;
1199 /* the last event terminates with evsel->attr.size == 0: */
1200 events = calloc(nre + 1, sizeof(*events));
1201 if (!events)
1202 goto error;
1204 msz = sizeof(evsel->attr);
1205 if (sz < msz)
1206 msz = sz;
1208 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1209 evsel->idx = i;
1212 * must read entire on-file attr struct to
1213 * sync up with layout.
1215 ret = readn(fd, buf, sz);
1216 if (ret != (ssize_t)sz)
1217 goto error;
1219 if (ph->needs_swap)
1220 perf_event__attr_swap(buf);
1222 memcpy(&evsel->attr, buf, msz);
1224 ret = readn(fd, &nr, sizeof(nr));
1225 if (ret != (ssize_t)sizeof(nr))
1226 goto error;
1228 if (ph->needs_swap) {
1229 nr = bswap_32(nr);
1230 evsel->needs_swap = true;
1233 evsel->name = do_read_string(fd, ph);
1235 if (!nr)
1236 continue;
1238 id = calloc(nr, sizeof(*id));
1239 if (!id)
1240 goto error;
1241 evsel->ids = nr;
1242 evsel->id = id;
1244 for (j = 0 ; j < nr; j++) {
1245 ret = readn(fd, id, sizeof(*id));
1246 if (ret != (ssize_t)sizeof(*id))
1247 goto error;
1248 if (ph->needs_swap)
1249 *id = bswap_64(*id);
1250 id++;
1253 out:
1254 free(buf);
1255 return events;
1256 error:
1257 free_event_desc(events);
1258 events = NULL;
1259 goto out;
1262 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1263 void *priv __attribute__((unused)))
1265 return fprintf(fp, ", %s = %s", name, val);
1268 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1270 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1271 u32 j;
1272 u64 *id;
1274 if (!events) {
1275 fprintf(fp, "# event desc: not available or unable to read\n");
1276 return;
1279 for (evsel = events; evsel->attr.size; evsel++) {
1280 fprintf(fp, "# event : name = %s, ", evsel->name);
1282 if (evsel->ids) {
1283 fprintf(fp, ", id = {");
1284 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1285 if (j)
1286 fputc(',', fp);
1287 fprintf(fp, " %"PRIu64, *id);
1289 fprintf(fp, " }");
1292 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1294 fputc('\n', fp);
1297 free_event_desc(events);
1300 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1301 FILE *fp)
1303 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1306 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1307 FILE *fp)
1309 u32 nr, c, i;
1310 char *str, *tmp;
1311 uint64_t mem_total, mem_free;
1313 /* nr nodes */
1314 nr = ph->env.nr_numa_nodes;
1315 str = ph->env.numa_nodes;
1317 for (i = 0; i < nr; i++) {
1318 /* node number */
1319 c = strtoul(str, &tmp, 0);
1320 if (*tmp != ':')
1321 goto error;
1323 str = tmp + 1;
1324 mem_total = strtoull(str, &tmp, 0);
1325 if (*tmp != ':')
1326 goto error;
1328 str = tmp + 1;
1329 mem_free = strtoull(str, &tmp, 0);
1330 if (*tmp != ':')
1331 goto error;
1333 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1334 " free = %"PRIu64" kB\n",
1335 c, mem_total, mem_free);
1337 str = tmp + 1;
1338 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1340 str += strlen(str) + 1;
1342 return;
1343 error:
1344 fprintf(fp, "# numa topology : not available\n");
1347 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1349 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1352 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1353 int fd __maybe_unused, FILE *fp)
1355 fprintf(fp, "# contains samples with branch stack\n");
1358 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1359 int fd __maybe_unused, FILE *fp)
1361 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1364 static void print_stat(struct perf_header *ph __maybe_unused,
1365 int fd __maybe_unused, FILE *fp)
1367 fprintf(fp, "# contains stat data\n");
1370 static void print_cache(struct perf_header *ph __maybe_unused,
1371 int fd __maybe_unused, FILE *fp __maybe_unused)
1373 int i;
1375 fprintf(fp, "# CPU cache info:\n");
1376 for (i = 0; i < ph->env.caches_cnt; i++) {
1377 fprintf(fp, "# ");
1378 cpu_cache_level__fprintf(fp, &ph->env.caches[i]);
1382 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1383 FILE *fp)
1385 const char *delimiter = "# pmu mappings: ";
1386 char *str, *tmp;
1387 u32 pmu_num;
1388 u32 type;
1390 pmu_num = ph->env.nr_pmu_mappings;
1391 if (!pmu_num) {
1392 fprintf(fp, "# pmu mappings: not available\n");
1393 return;
1396 str = ph->env.pmu_mappings;
1398 while (pmu_num) {
1399 type = strtoul(str, &tmp, 0);
1400 if (*tmp != ':')
1401 goto error;
1403 str = tmp + 1;
1404 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1406 delimiter = ", ";
1407 str += strlen(str) + 1;
1408 pmu_num--;
1411 fprintf(fp, "\n");
1413 if (!pmu_num)
1414 return;
1415 error:
1416 fprintf(fp, "# pmu mappings: unable to read\n");
1419 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1420 FILE *fp)
1422 struct perf_session *session;
1423 struct perf_evsel *evsel;
1424 u32 nr = 0;
1426 session = container_of(ph, struct perf_session, header);
1428 evlist__for_each(session->evlist, evsel) {
1429 if (perf_evsel__is_group_leader(evsel) &&
1430 evsel->nr_members > 1) {
1431 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1432 perf_evsel__name(evsel));
1434 nr = evsel->nr_members - 1;
1435 } else if (nr) {
1436 fprintf(fp, ",%s", perf_evsel__name(evsel));
1438 if (--nr == 0)
1439 fprintf(fp, "}\n");
1444 static int __event_process_build_id(struct build_id_event *bev,
1445 char *filename,
1446 struct perf_session *session)
1448 int err = -1;
1449 struct machine *machine;
1450 u16 cpumode;
1451 struct dso *dso;
1452 enum dso_kernel_type dso_type;
1454 machine = perf_session__findnew_machine(session, bev->pid);
1455 if (!machine)
1456 goto out;
1458 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1460 switch (cpumode) {
1461 case PERF_RECORD_MISC_KERNEL:
1462 dso_type = DSO_TYPE_KERNEL;
1463 break;
1464 case PERF_RECORD_MISC_GUEST_KERNEL:
1465 dso_type = DSO_TYPE_GUEST_KERNEL;
1466 break;
1467 case PERF_RECORD_MISC_USER:
1468 case PERF_RECORD_MISC_GUEST_USER:
1469 dso_type = DSO_TYPE_USER;
1470 break;
1471 default:
1472 goto out;
1475 dso = machine__findnew_dso(machine, filename);
1476 if (dso != NULL) {
1477 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1479 dso__set_build_id(dso, &bev->build_id);
1481 if (!is_kernel_module(filename, cpumode))
1482 dso->kernel = dso_type;
1484 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1485 sbuild_id);
1486 pr_debug("build id event received for %s: %s\n",
1487 dso->long_name, sbuild_id);
1488 dso__put(dso);
1491 err = 0;
1492 out:
1493 return err;
1496 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1497 int input, u64 offset, u64 size)
1499 struct perf_session *session = container_of(header, struct perf_session, header);
1500 struct {
1501 struct perf_event_header header;
1502 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1503 char filename[0];
1504 } old_bev;
1505 struct build_id_event bev;
1506 char filename[PATH_MAX];
1507 u64 limit = offset + size;
1509 while (offset < limit) {
1510 ssize_t len;
1512 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1513 return -1;
1515 if (header->needs_swap)
1516 perf_event_header__bswap(&old_bev.header);
1518 len = old_bev.header.size - sizeof(old_bev);
1519 if (readn(input, filename, len) != len)
1520 return -1;
1522 bev.header = old_bev.header;
1525 * As the pid is the missing value, we need to fill
1526 * it properly. The header.misc value give us nice hint.
1528 bev.pid = HOST_KERNEL_ID;
1529 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1530 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1531 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1533 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1534 __event_process_build_id(&bev, filename, session);
1536 offset += bev.header.size;
1539 return 0;
1542 static int perf_header__read_build_ids(struct perf_header *header,
1543 int input, u64 offset, u64 size)
1545 struct perf_session *session = container_of(header, struct perf_session, header);
1546 struct build_id_event bev;
1547 char filename[PATH_MAX];
1548 u64 limit = offset + size, orig_offset = offset;
1549 int err = -1;
1551 while (offset < limit) {
1552 ssize_t len;
1554 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1555 goto out;
1557 if (header->needs_swap)
1558 perf_event_header__bswap(&bev.header);
1560 len = bev.header.size - sizeof(bev);
1561 if (readn(input, filename, len) != len)
1562 goto out;
1564 * The a1645ce1 changeset:
1566 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1568 * Added a field to struct build_id_event that broke the file
1569 * format.
1571 * Since the kernel build-id is the first entry, process the
1572 * table using the old format if the well known
1573 * '[kernel.kallsyms]' string for the kernel build-id has the
1574 * first 4 characters chopped off (where the pid_t sits).
1576 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1577 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1578 return -1;
1579 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1582 __event_process_build_id(&bev, filename, session);
1584 offset += bev.header.size;
1586 err = 0;
1587 out:
1588 return err;
1591 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1592 struct perf_header *ph __maybe_unused,
1593 int fd, void *data)
1595 ssize_t ret = trace_report(fd, data, false);
1596 return ret < 0 ? -1 : 0;
1599 static int process_build_id(struct perf_file_section *section,
1600 struct perf_header *ph, int fd,
1601 void *data __maybe_unused)
1603 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1604 pr_debug("Failed to read buildids, continuing...\n");
1605 return 0;
1608 static int process_hostname(struct perf_file_section *section __maybe_unused,
1609 struct perf_header *ph, int fd,
1610 void *data __maybe_unused)
1612 ph->env.hostname = do_read_string(fd, ph);
1613 return ph->env.hostname ? 0 : -ENOMEM;
1616 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1617 struct perf_header *ph, int fd,
1618 void *data __maybe_unused)
1620 ph->env.os_release = do_read_string(fd, ph);
1621 return ph->env.os_release ? 0 : -ENOMEM;
1624 static int process_version(struct perf_file_section *section __maybe_unused,
1625 struct perf_header *ph, int fd,
1626 void *data __maybe_unused)
1628 ph->env.version = do_read_string(fd, ph);
1629 return ph->env.version ? 0 : -ENOMEM;
1632 static int process_arch(struct perf_file_section *section __maybe_unused,
1633 struct perf_header *ph, int fd,
1634 void *data __maybe_unused)
1636 ph->env.arch = do_read_string(fd, ph);
1637 return ph->env.arch ? 0 : -ENOMEM;
1640 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1641 struct perf_header *ph, int fd,
1642 void *data __maybe_unused)
1644 ssize_t ret;
1645 u32 nr;
1647 ret = readn(fd, &nr, sizeof(nr));
1648 if (ret != sizeof(nr))
1649 return -1;
1651 if (ph->needs_swap)
1652 nr = bswap_32(nr);
1654 ph->env.nr_cpus_avail = nr;
1656 ret = readn(fd, &nr, sizeof(nr));
1657 if (ret != sizeof(nr))
1658 return -1;
1660 if (ph->needs_swap)
1661 nr = bswap_32(nr);
1663 ph->env.nr_cpus_online = nr;
1664 return 0;
1667 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1668 struct perf_header *ph, int fd,
1669 void *data __maybe_unused)
1671 ph->env.cpu_desc = do_read_string(fd, ph);
1672 return ph->env.cpu_desc ? 0 : -ENOMEM;
1675 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1676 struct perf_header *ph, int fd,
1677 void *data __maybe_unused)
1679 ph->env.cpuid = do_read_string(fd, ph);
1680 return ph->env.cpuid ? 0 : -ENOMEM;
1683 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1684 struct perf_header *ph, int fd,
1685 void *data __maybe_unused)
1687 uint64_t mem;
1688 ssize_t ret;
1690 ret = readn(fd, &mem, sizeof(mem));
1691 if (ret != sizeof(mem))
1692 return -1;
1694 if (ph->needs_swap)
1695 mem = bswap_64(mem);
1697 ph->env.total_mem = mem;
1698 return 0;
1701 static struct perf_evsel *
1702 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1704 struct perf_evsel *evsel;
1706 evlist__for_each(evlist, evsel) {
1707 if (evsel->idx == idx)
1708 return evsel;
1711 return NULL;
1714 static void
1715 perf_evlist__set_event_name(struct perf_evlist *evlist,
1716 struct perf_evsel *event)
1718 struct perf_evsel *evsel;
1720 if (!event->name)
1721 return;
1723 evsel = perf_evlist__find_by_index(evlist, event->idx);
1724 if (!evsel)
1725 return;
1727 if (evsel->name)
1728 return;
1730 evsel->name = strdup(event->name);
1733 static int
1734 process_event_desc(struct perf_file_section *section __maybe_unused,
1735 struct perf_header *header, int fd,
1736 void *data __maybe_unused)
1738 struct perf_session *session;
1739 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1741 if (!events)
1742 return 0;
1744 session = container_of(header, struct perf_session, header);
1745 for (evsel = events; evsel->attr.size; evsel++)
1746 perf_evlist__set_event_name(session->evlist, evsel);
1748 free_event_desc(events);
1750 return 0;
1753 static int process_cmdline(struct perf_file_section *section,
1754 struct perf_header *ph, int fd,
1755 void *data __maybe_unused)
1757 ssize_t ret;
1758 char *str, *cmdline = NULL, **argv = NULL;
1759 u32 nr, i, len = 0;
1761 ret = readn(fd, &nr, sizeof(nr));
1762 if (ret != sizeof(nr))
1763 return -1;
1765 if (ph->needs_swap)
1766 nr = bswap_32(nr);
1768 ph->env.nr_cmdline = nr;
1770 cmdline = zalloc(section->size + nr + 1);
1771 if (!cmdline)
1772 return -1;
1774 argv = zalloc(sizeof(char *) * (nr + 1));
1775 if (!argv)
1776 goto error;
1778 for (i = 0; i < nr; i++) {
1779 str = do_read_string(fd, ph);
1780 if (!str)
1781 goto error;
1783 argv[i] = cmdline + len;
1784 memcpy(argv[i], str, strlen(str) + 1);
1785 len += strlen(str) + 1;
1786 free(str);
1788 ph->env.cmdline = cmdline;
1789 ph->env.cmdline_argv = (const char **) argv;
1790 return 0;
1792 error:
1793 free(argv);
1794 free(cmdline);
1795 return -1;
1798 static int process_cpu_topology(struct perf_file_section *section,
1799 struct perf_header *ph, int fd,
1800 void *data __maybe_unused)
1802 ssize_t ret;
1803 u32 nr, i;
1804 char *str;
1805 struct strbuf sb;
1806 int cpu_nr = ph->env.nr_cpus_online;
1807 u64 size = 0;
1809 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
1810 if (!ph->env.cpu)
1811 return -1;
1813 ret = readn(fd, &nr, sizeof(nr));
1814 if (ret != sizeof(nr))
1815 goto free_cpu;
1817 if (ph->needs_swap)
1818 nr = bswap_32(nr);
1820 ph->env.nr_sibling_cores = nr;
1821 size += sizeof(u32);
1822 strbuf_init(&sb, 128);
1824 for (i = 0; i < nr; i++) {
1825 str = do_read_string(fd, ph);
1826 if (!str)
1827 goto error;
1829 /* include a NULL character at the end */
1830 strbuf_add(&sb, str, strlen(str) + 1);
1831 size += string_size(str);
1832 free(str);
1834 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1836 ret = readn(fd, &nr, sizeof(nr));
1837 if (ret != sizeof(nr))
1838 return -1;
1840 if (ph->needs_swap)
1841 nr = bswap_32(nr);
1843 ph->env.nr_sibling_threads = nr;
1844 size += sizeof(u32);
1846 for (i = 0; i < nr; i++) {
1847 str = do_read_string(fd, ph);
1848 if (!str)
1849 goto error;
1851 /* include a NULL character at the end */
1852 strbuf_add(&sb, str, strlen(str) + 1);
1853 size += string_size(str);
1854 free(str);
1856 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1859 * The header may be from old perf,
1860 * which doesn't include core id and socket id information.
1862 if (section->size <= size) {
1863 zfree(&ph->env.cpu);
1864 return 0;
1867 for (i = 0; i < (u32)cpu_nr; i++) {
1868 ret = readn(fd, &nr, sizeof(nr));
1869 if (ret != sizeof(nr))
1870 goto free_cpu;
1872 if (ph->needs_swap)
1873 nr = bswap_32(nr);
1875 ph->env.cpu[i].core_id = nr;
1877 ret = readn(fd, &nr, sizeof(nr));
1878 if (ret != sizeof(nr))
1879 goto free_cpu;
1881 if (ph->needs_swap)
1882 nr = bswap_32(nr);
1884 if (nr > (u32)cpu_nr) {
1885 pr_debug("socket_id number is too big."
1886 "You may need to upgrade the perf tool.\n");
1887 goto free_cpu;
1890 ph->env.cpu[i].socket_id = nr;
1893 return 0;
1895 error:
1896 strbuf_release(&sb);
1897 free_cpu:
1898 zfree(&ph->env.cpu);
1899 return -1;
1902 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1903 struct perf_header *ph, int fd,
1904 void *data __maybe_unused)
1906 ssize_t ret;
1907 u32 nr, node, i;
1908 char *str;
1909 uint64_t mem_total, mem_free;
1910 struct strbuf sb;
1912 /* nr nodes */
1913 ret = readn(fd, &nr, sizeof(nr));
1914 if (ret != sizeof(nr))
1915 goto error;
1917 if (ph->needs_swap)
1918 nr = bswap_32(nr);
1920 ph->env.nr_numa_nodes = nr;
1921 strbuf_init(&sb, 256);
1923 for (i = 0; i < nr; i++) {
1924 /* node number */
1925 ret = readn(fd, &node, sizeof(node));
1926 if (ret != sizeof(node))
1927 goto error;
1929 ret = readn(fd, &mem_total, sizeof(u64));
1930 if (ret != sizeof(u64))
1931 goto error;
1933 ret = readn(fd, &mem_free, sizeof(u64));
1934 if (ret != sizeof(u64))
1935 goto error;
1937 if (ph->needs_swap) {
1938 node = bswap_32(node);
1939 mem_total = bswap_64(mem_total);
1940 mem_free = bswap_64(mem_free);
1943 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1944 node, mem_total, mem_free);
1946 str = do_read_string(fd, ph);
1947 if (!str)
1948 goto error;
1950 /* include a NULL character at the end */
1951 strbuf_add(&sb, str, strlen(str) + 1);
1952 free(str);
1954 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1955 return 0;
1957 error:
1958 strbuf_release(&sb);
1959 return -1;
1962 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1963 struct perf_header *ph, int fd,
1964 void *data __maybe_unused)
1966 ssize_t ret;
1967 char *name;
1968 u32 pmu_num;
1969 u32 type;
1970 struct strbuf sb;
1972 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1973 if (ret != sizeof(pmu_num))
1974 return -1;
1976 if (ph->needs_swap)
1977 pmu_num = bswap_32(pmu_num);
1979 if (!pmu_num) {
1980 pr_debug("pmu mappings not available\n");
1981 return 0;
1984 ph->env.nr_pmu_mappings = pmu_num;
1985 strbuf_init(&sb, 128);
1987 while (pmu_num) {
1988 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1989 goto error;
1990 if (ph->needs_swap)
1991 type = bswap_32(type);
1993 name = do_read_string(fd, ph);
1994 if (!name)
1995 goto error;
1997 strbuf_addf(&sb, "%u:%s", type, name);
1998 /* include a NULL character at the end */
1999 strbuf_add(&sb, "", 1);
2001 if (!strcmp(name, "msr"))
2002 ph->env.msr_pmu_type = type;
2004 free(name);
2005 pmu_num--;
2007 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2008 return 0;
2010 error:
2011 strbuf_release(&sb);
2012 return -1;
2015 static int process_group_desc(struct perf_file_section *section __maybe_unused,
2016 struct perf_header *ph, int fd,
2017 void *data __maybe_unused)
2019 size_t ret = -1;
2020 u32 i, nr, nr_groups;
2021 struct perf_session *session;
2022 struct perf_evsel *evsel, *leader = NULL;
2023 struct group_desc {
2024 char *name;
2025 u32 leader_idx;
2026 u32 nr_members;
2027 } *desc;
2029 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
2030 return -1;
2032 if (ph->needs_swap)
2033 nr_groups = bswap_32(nr_groups);
2035 ph->env.nr_groups = nr_groups;
2036 if (!nr_groups) {
2037 pr_debug("group desc not available\n");
2038 return 0;
2041 desc = calloc(nr_groups, sizeof(*desc));
2042 if (!desc)
2043 return -1;
2045 for (i = 0; i < nr_groups; i++) {
2046 desc[i].name = do_read_string(fd, ph);
2047 if (!desc[i].name)
2048 goto out_free;
2050 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
2051 goto out_free;
2053 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
2054 goto out_free;
2056 if (ph->needs_swap) {
2057 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
2058 desc[i].nr_members = bswap_32(desc[i].nr_members);
2063 * Rebuild group relationship based on the group_desc
2065 session = container_of(ph, struct perf_session, header);
2066 session->evlist->nr_groups = nr_groups;
2068 i = nr = 0;
2069 evlist__for_each(session->evlist, evsel) {
2070 if (evsel->idx == (int) desc[i].leader_idx) {
2071 evsel->leader = evsel;
2072 /* {anon_group} is a dummy name */
2073 if (strcmp(desc[i].name, "{anon_group}")) {
2074 evsel->group_name = desc[i].name;
2075 desc[i].name = NULL;
2077 evsel->nr_members = desc[i].nr_members;
2079 if (i >= nr_groups || nr > 0) {
2080 pr_debug("invalid group desc\n");
2081 goto out_free;
2084 leader = evsel;
2085 nr = evsel->nr_members - 1;
2086 i++;
2087 } else if (nr) {
2088 /* This is a group member */
2089 evsel->leader = leader;
2091 nr--;
2095 if (i != nr_groups || nr != 0) {
2096 pr_debug("invalid group desc\n");
2097 goto out_free;
2100 ret = 0;
2101 out_free:
2102 for (i = 0; i < nr_groups; i++)
2103 zfree(&desc[i].name);
2104 free(desc);
2106 return ret;
2109 static int process_auxtrace(struct perf_file_section *section,
2110 struct perf_header *ph, int fd,
2111 void *data __maybe_unused)
2113 struct perf_session *session;
2114 int err;
2116 session = container_of(ph, struct perf_session, header);
2118 err = auxtrace_index__process(fd, section->size, session,
2119 ph->needs_swap);
2120 if (err < 0)
2121 pr_err("Failed to process auxtrace index\n");
2122 return err;
2125 static int process_cache(struct perf_file_section *section __maybe_unused,
2126 struct perf_header *ph __maybe_unused, int fd __maybe_unused,
2127 void *data __maybe_unused)
2129 struct cpu_cache_level *caches;
2130 u32 cnt, i, version;
2132 if (readn(fd, &version, sizeof(version)) != sizeof(version))
2133 return -1;
2135 if (ph->needs_swap)
2136 version = bswap_32(version);
2138 if (version != 1)
2139 return -1;
2141 if (readn(fd, &cnt, sizeof(cnt)) != sizeof(cnt))
2142 return -1;
2144 if (ph->needs_swap)
2145 cnt = bswap_32(cnt);
2147 caches = zalloc(sizeof(*caches) * cnt);
2148 if (!caches)
2149 return -1;
2151 for (i = 0; i < cnt; i++) {
2152 struct cpu_cache_level c;
2154 #define _R(v) \
2155 if (readn(fd, &c.v, sizeof(u32)) != sizeof(u32))\
2156 goto out_free_caches; \
2157 if (ph->needs_swap) \
2158 c.v = bswap_32(c.v); \
2160 _R(level)
2161 _R(line_size)
2162 _R(sets)
2163 _R(ways)
2164 #undef _R
2166 #define _R(v) \
2167 c.v = do_read_string(fd, ph); \
2168 if (!c.v) \
2169 goto out_free_caches;
2171 _R(type)
2172 _R(size)
2173 _R(map)
2174 #undef _R
2176 caches[i] = c;
2179 ph->env.caches = caches;
2180 ph->env.caches_cnt = cnt;
2181 return 0;
2182 out_free_caches:
2183 free(caches);
2184 return -1;
2187 struct feature_ops {
2188 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
2189 void (*print)(struct perf_header *h, int fd, FILE *fp);
2190 int (*process)(struct perf_file_section *section,
2191 struct perf_header *h, int fd, void *data);
2192 const char *name;
2193 bool full_only;
2196 #define FEAT_OPA(n, func) \
2197 [n] = { .name = #n, .write = write_##func, .print = print_##func }
2198 #define FEAT_OPP(n, func) \
2199 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2200 .process = process_##func }
2201 #define FEAT_OPF(n, func) \
2202 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2203 .process = process_##func, .full_only = true }
2205 /* feature_ops not implemented: */
2206 #define print_tracing_data NULL
2207 #define print_build_id NULL
2209 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2210 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
2211 FEAT_OPP(HEADER_BUILD_ID, build_id),
2212 FEAT_OPP(HEADER_HOSTNAME, hostname),
2213 FEAT_OPP(HEADER_OSRELEASE, osrelease),
2214 FEAT_OPP(HEADER_VERSION, version),
2215 FEAT_OPP(HEADER_ARCH, arch),
2216 FEAT_OPP(HEADER_NRCPUS, nrcpus),
2217 FEAT_OPP(HEADER_CPUDESC, cpudesc),
2218 FEAT_OPP(HEADER_CPUID, cpuid),
2219 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
2220 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
2221 FEAT_OPP(HEADER_CMDLINE, cmdline),
2222 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
2223 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
2224 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
2225 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
2226 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
2227 FEAT_OPP(HEADER_AUXTRACE, auxtrace),
2228 FEAT_OPA(HEADER_STAT, stat),
2229 FEAT_OPF(HEADER_CACHE, cache),
2232 struct header_print_data {
2233 FILE *fp;
2234 bool full; /* extended list of headers */
2237 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2238 struct perf_header *ph,
2239 int feat, int fd, void *data)
2241 struct header_print_data *hd = data;
2243 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2244 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2245 "%d, continuing...\n", section->offset, feat);
2246 return 0;
2248 if (feat >= HEADER_LAST_FEATURE) {
2249 pr_warning("unknown feature %d\n", feat);
2250 return 0;
2252 if (!feat_ops[feat].print)
2253 return 0;
2255 if (!feat_ops[feat].full_only || hd->full)
2256 feat_ops[feat].print(ph, fd, hd->fp);
2257 else
2258 fprintf(hd->fp, "# %s info available, use -I to display\n",
2259 feat_ops[feat].name);
2261 return 0;
2264 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2266 struct header_print_data hd;
2267 struct perf_header *header = &session->header;
2268 int fd = perf_data_file__fd(session->file);
2269 hd.fp = fp;
2270 hd.full = full;
2272 perf_header__process_sections(header, fd, &hd,
2273 perf_file_section__fprintf_info);
2274 return 0;
2277 static int do_write_feat(int fd, struct perf_header *h, int type,
2278 struct perf_file_section **p,
2279 struct perf_evlist *evlist)
2281 int err;
2282 int ret = 0;
2284 if (perf_header__has_feat(h, type)) {
2285 if (!feat_ops[type].write)
2286 return -1;
2288 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2290 err = feat_ops[type].write(fd, h, evlist);
2291 if (err < 0) {
2292 pr_debug("failed to write feature %d\n", type);
2294 /* undo anything written */
2295 lseek(fd, (*p)->offset, SEEK_SET);
2297 return -1;
2299 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2300 (*p)++;
2302 return ret;
2305 static int perf_header__adds_write(struct perf_header *header,
2306 struct perf_evlist *evlist, int fd)
2308 int nr_sections;
2309 struct perf_file_section *feat_sec, *p;
2310 int sec_size;
2311 u64 sec_start;
2312 int feat;
2313 int err;
2315 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2316 if (!nr_sections)
2317 return 0;
2319 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2320 if (feat_sec == NULL)
2321 return -ENOMEM;
2323 sec_size = sizeof(*feat_sec) * nr_sections;
2325 sec_start = header->feat_offset;
2326 lseek(fd, sec_start + sec_size, SEEK_SET);
2328 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2329 if (do_write_feat(fd, header, feat, &p, evlist))
2330 perf_header__clear_feat(header, feat);
2333 lseek(fd, sec_start, SEEK_SET);
2335 * may write more than needed due to dropped feature, but
2336 * this is okay, reader will skip the mising entries
2338 err = do_write(fd, feat_sec, sec_size);
2339 if (err < 0)
2340 pr_debug("failed to write feature section\n");
2341 free(feat_sec);
2342 return err;
2345 int perf_header__write_pipe(int fd)
2347 struct perf_pipe_file_header f_header;
2348 int err;
2350 f_header = (struct perf_pipe_file_header){
2351 .magic = PERF_MAGIC,
2352 .size = sizeof(f_header),
2355 err = do_write(fd, &f_header, sizeof(f_header));
2356 if (err < 0) {
2357 pr_debug("failed to write perf pipe header\n");
2358 return err;
2361 return 0;
2364 int perf_session__write_header(struct perf_session *session,
2365 struct perf_evlist *evlist,
2366 int fd, bool at_exit)
2368 struct perf_file_header f_header;
2369 struct perf_file_attr f_attr;
2370 struct perf_header *header = &session->header;
2371 struct perf_evsel *evsel;
2372 u64 attr_offset;
2373 int err;
2375 lseek(fd, sizeof(f_header), SEEK_SET);
2377 evlist__for_each(session->evlist, evsel) {
2378 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2379 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2380 if (err < 0) {
2381 pr_debug("failed to write perf header\n");
2382 return err;
2386 attr_offset = lseek(fd, 0, SEEK_CUR);
2388 evlist__for_each(evlist, evsel) {
2389 f_attr = (struct perf_file_attr){
2390 .attr = evsel->attr,
2391 .ids = {
2392 .offset = evsel->id_offset,
2393 .size = evsel->ids * sizeof(u64),
2396 err = do_write(fd, &f_attr, sizeof(f_attr));
2397 if (err < 0) {
2398 pr_debug("failed to write perf header attribute\n");
2399 return err;
2403 if (!header->data_offset)
2404 header->data_offset = lseek(fd, 0, SEEK_CUR);
2405 header->feat_offset = header->data_offset + header->data_size;
2407 if (at_exit) {
2408 err = perf_header__adds_write(header, evlist, fd);
2409 if (err < 0)
2410 return err;
2413 f_header = (struct perf_file_header){
2414 .magic = PERF_MAGIC,
2415 .size = sizeof(f_header),
2416 .attr_size = sizeof(f_attr),
2417 .attrs = {
2418 .offset = attr_offset,
2419 .size = evlist->nr_entries * sizeof(f_attr),
2421 .data = {
2422 .offset = header->data_offset,
2423 .size = header->data_size,
2425 /* event_types is ignored, store zeros */
2428 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2430 lseek(fd, 0, SEEK_SET);
2431 err = do_write(fd, &f_header, sizeof(f_header));
2432 if (err < 0) {
2433 pr_debug("failed to write perf header\n");
2434 return err;
2436 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2438 return 0;
2441 static int perf_header__getbuffer64(struct perf_header *header,
2442 int fd, void *buf, size_t size)
2444 if (readn(fd, buf, size) <= 0)
2445 return -1;
2447 if (header->needs_swap)
2448 mem_bswap_64(buf, size);
2450 return 0;
2453 int perf_header__process_sections(struct perf_header *header, int fd,
2454 void *data,
2455 int (*process)(struct perf_file_section *section,
2456 struct perf_header *ph,
2457 int feat, int fd, void *data))
2459 struct perf_file_section *feat_sec, *sec;
2460 int nr_sections;
2461 int sec_size;
2462 int feat;
2463 int err;
2465 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2466 if (!nr_sections)
2467 return 0;
2469 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2470 if (!feat_sec)
2471 return -1;
2473 sec_size = sizeof(*feat_sec) * nr_sections;
2475 lseek(fd, header->feat_offset, SEEK_SET);
2477 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2478 if (err < 0)
2479 goto out_free;
2481 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2482 err = process(sec++, header, feat, fd, data);
2483 if (err < 0)
2484 goto out_free;
2486 err = 0;
2487 out_free:
2488 free(feat_sec);
2489 return err;
2492 static const int attr_file_abi_sizes[] = {
2493 [0] = PERF_ATTR_SIZE_VER0,
2494 [1] = PERF_ATTR_SIZE_VER1,
2495 [2] = PERF_ATTR_SIZE_VER2,
2496 [3] = PERF_ATTR_SIZE_VER3,
2497 [4] = PERF_ATTR_SIZE_VER4,
2502 * In the legacy file format, the magic number is not used to encode endianness.
2503 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2504 * on ABI revisions, we need to try all combinations for all endianness to
2505 * detect the endianness.
2507 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2509 uint64_t ref_size, attr_size;
2510 int i;
2512 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2513 ref_size = attr_file_abi_sizes[i]
2514 + sizeof(struct perf_file_section);
2515 if (hdr_sz != ref_size) {
2516 attr_size = bswap_64(hdr_sz);
2517 if (attr_size != ref_size)
2518 continue;
2520 ph->needs_swap = true;
2522 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2524 ph->needs_swap);
2525 return 0;
2527 /* could not determine endianness */
2528 return -1;
2531 #define PERF_PIPE_HDR_VER0 16
2533 static const size_t attr_pipe_abi_sizes[] = {
2534 [0] = PERF_PIPE_HDR_VER0,
2539 * In the legacy pipe format, there is an implicit assumption that endiannesss
2540 * between host recording the samples, and host parsing the samples is the
2541 * same. This is not always the case given that the pipe output may always be
2542 * redirected into a file and analyzed on a different machine with possibly a
2543 * different endianness and perf_event ABI revsions in the perf tool itself.
2545 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2547 u64 attr_size;
2548 int i;
2550 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2551 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2552 attr_size = bswap_64(hdr_sz);
2553 if (attr_size != hdr_sz)
2554 continue;
2556 ph->needs_swap = true;
2558 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2559 return 0;
2561 return -1;
2564 bool is_perf_magic(u64 magic)
2566 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2567 || magic == __perf_magic2
2568 || magic == __perf_magic2_sw)
2569 return true;
2571 return false;
2574 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2575 bool is_pipe, struct perf_header *ph)
2577 int ret;
2579 /* check for legacy format */
2580 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2581 if (ret == 0) {
2582 ph->version = PERF_HEADER_VERSION_1;
2583 pr_debug("legacy perf.data format\n");
2584 if (is_pipe)
2585 return try_all_pipe_abis(hdr_sz, ph);
2587 return try_all_file_abis(hdr_sz, ph);
2590 * the new magic number serves two purposes:
2591 * - unique number to identify actual perf.data files
2592 * - encode endianness of file
2594 ph->version = PERF_HEADER_VERSION_2;
2596 /* check magic number with one endianness */
2597 if (magic == __perf_magic2)
2598 return 0;
2600 /* check magic number with opposite endianness */
2601 if (magic != __perf_magic2_sw)
2602 return -1;
2604 ph->needs_swap = true;
2606 return 0;
2609 int perf_file_header__read(struct perf_file_header *header,
2610 struct perf_header *ph, int fd)
2612 ssize_t ret;
2614 lseek(fd, 0, SEEK_SET);
2616 ret = readn(fd, header, sizeof(*header));
2617 if (ret <= 0)
2618 return -1;
2620 if (check_magic_endian(header->magic,
2621 header->attr_size, false, ph) < 0) {
2622 pr_debug("magic/endian check failed\n");
2623 return -1;
2626 if (ph->needs_swap) {
2627 mem_bswap_64(header, offsetof(struct perf_file_header,
2628 adds_features));
2631 if (header->size != sizeof(*header)) {
2632 /* Support the previous format */
2633 if (header->size == offsetof(typeof(*header), adds_features))
2634 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2635 else
2636 return -1;
2637 } else if (ph->needs_swap) {
2639 * feature bitmap is declared as an array of unsigned longs --
2640 * not good since its size can differ between the host that
2641 * generated the data file and the host analyzing the file.
2643 * We need to handle endianness, but we don't know the size of
2644 * the unsigned long where the file was generated. Take a best
2645 * guess at determining it: try 64-bit swap first (ie., file
2646 * created on a 64-bit host), and check if the hostname feature
2647 * bit is set (this feature bit is forced on as of fbe96f2).
2648 * If the bit is not, undo the 64-bit swap and try a 32-bit
2649 * swap. If the hostname bit is still not set (e.g., older data
2650 * file), punt and fallback to the original behavior --
2651 * clearing all feature bits and setting buildid.
2653 mem_bswap_64(&header->adds_features,
2654 BITS_TO_U64(HEADER_FEAT_BITS));
2656 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2657 /* unswap as u64 */
2658 mem_bswap_64(&header->adds_features,
2659 BITS_TO_U64(HEADER_FEAT_BITS));
2661 /* unswap as u32 */
2662 mem_bswap_32(&header->adds_features,
2663 BITS_TO_U32(HEADER_FEAT_BITS));
2666 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2667 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2668 set_bit(HEADER_BUILD_ID, header->adds_features);
2672 memcpy(&ph->adds_features, &header->adds_features,
2673 sizeof(ph->adds_features));
2675 ph->data_offset = header->data.offset;
2676 ph->data_size = header->data.size;
2677 ph->feat_offset = header->data.offset + header->data.size;
2678 return 0;
2681 static int perf_file_section__process(struct perf_file_section *section,
2682 struct perf_header *ph,
2683 int feat, int fd, void *data)
2685 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2686 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2687 "%d, continuing...\n", section->offset, feat);
2688 return 0;
2691 if (feat >= HEADER_LAST_FEATURE) {
2692 pr_debug("unknown feature %d, continuing...\n", feat);
2693 return 0;
2696 if (!feat_ops[feat].process)
2697 return 0;
2699 return feat_ops[feat].process(section, ph, fd, data);
2702 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2703 struct perf_header *ph, int fd,
2704 bool repipe)
2706 ssize_t ret;
2708 ret = readn(fd, header, sizeof(*header));
2709 if (ret <= 0)
2710 return -1;
2712 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2713 pr_debug("endian/magic failed\n");
2714 return -1;
2717 if (ph->needs_swap)
2718 header->size = bswap_64(header->size);
2720 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2721 return -1;
2723 return 0;
2726 static int perf_header__read_pipe(struct perf_session *session)
2728 struct perf_header *header = &session->header;
2729 struct perf_pipe_file_header f_header;
2731 if (perf_file_header__read_pipe(&f_header, header,
2732 perf_data_file__fd(session->file),
2733 session->repipe) < 0) {
2734 pr_debug("incompatible file format\n");
2735 return -EINVAL;
2738 return 0;
2741 static int read_attr(int fd, struct perf_header *ph,
2742 struct perf_file_attr *f_attr)
2744 struct perf_event_attr *attr = &f_attr->attr;
2745 size_t sz, left;
2746 size_t our_sz = sizeof(f_attr->attr);
2747 ssize_t ret;
2749 memset(f_attr, 0, sizeof(*f_attr));
2751 /* read minimal guaranteed structure */
2752 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2753 if (ret <= 0) {
2754 pr_debug("cannot read %d bytes of header attr\n",
2755 PERF_ATTR_SIZE_VER0);
2756 return -1;
2759 /* on file perf_event_attr size */
2760 sz = attr->size;
2762 if (ph->needs_swap)
2763 sz = bswap_32(sz);
2765 if (sz == 0) {
2766 /* assume ABI0 */
2767 sz = PERF_ATTR_SIZE_VER0;
2768 } else if (sz > our_sz) {
2769 pr_debug("file uses a more recent and unsupported ABI"
2770 " (%zu bytes extra)\n", sz - our_sz);
2771 return -1;
2773 /* what we have not yet read and that we know about */
2774 left = sz - PERF_ATTR_SIZE_VER0;
2775 if (left) {
2776 void *ptr = attr;
2777 ptr += PERF_ATTR_SIZE_VER0;
2779 ret = readn(fd, ptr, left);
2781 /* read perf_file_section, ids are read in caller */
2782 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2784 return ret <= 0 ? -1 : 0;
2787 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2788 struct pevent *pevent)
2790 struct event_format *event;
2791 char bf[128];
2793 /* already prepared */
2794 if (evsel->tp_format)
2795 return 0;
2797 if (pevent == NULL) {
2798 pr_debug("broken or missing trace data\n");
2799 return -1;
2802 event = pevent_find_event(pevent, evsel->attr.config);
2803 if (event == NULL)
2804 return -1;
2806 if (!evsel->name) {
2807 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2808 evsel->name = strdup(bf);
2809 if (evsel->name == NULL)
2810 return -1;
2813 evsel->tp_format = event;
2814 return 0;
2817 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2818 struct pevent *pevent)
2820 struct perf_evsel *pos;
2822 evlist__for_each(evlist, pos) {
2823 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2824 perf_evsel__prepare_tracepoint_event(pos, pevent))
2825 return -1;
2828 return 0;
2831 int perf_session__read_header(struct perf_session *session)
2833 struct perf_data_file *file = session->file;
2834 struct perf_header *header = &session->header;
2835 struct perf_file_header f_header;
2836 struct perf_file_attr f_attr;
2837 u64 f_id;
2838 int nr_attrs, nr_ids, i, j;
2839 int fd = perf_data_file__fd(file);
2841 session->evlist = perf_evlist__new();
2842 if (session->evlist == NULL)
2843 return -ENOMEM;
2845 session->evlist->env = &header->env;
2846 session->machines.host.env = &header->env;
2847 if (perf_data_file__is_pipe(file))
2848 return perf_header__read_pipe(session);
2850 if (perf_file_header__read(&f_header, header, fd) < 0)
2851 return -EINVAL;
2854 * Sanity check that perf.data was written cleanly; data size is
2855 * initialized to 0 and updated only if the on_exit function is run.
2856 * If data size is still 0 then the file contains only partial
2857 * information. Just warn user and process it as much as it can.
2859 if (f_header.data.size == 0) {
2860 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2861 "Was the 'perf record' command properly terminated?\n",
2862 file->path);
2865 nr_attrs = f_header.attrs.size / f_header.attr_size;
2866 lseek(fd, f_header.attrs.offset, SEEK_SET);
2868 for (i = 0; i < nr_attrs; i++) {
2869 struct perf_evsel *evsel;
2870 off_t tmp;
2872 if (read_attr(fd, header, &f_attr) < 0)
2873 goto out_errno;
2875 if (header->needs_swap) {
2876 f_attr.ids.size = bswap_64(f_attr.ids.size);
2877 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2878 perf_event__attr_swap(&f_attr.attr);
2881 tmp = lseek(fd, 0, SEEK_CUR);
2882 evsel = perf_evsel__new(&f_attr.attr);
2884 if (evsel == NULL)
2885 goto out_delete_evlist;
2887 evsel->needs_swap = header->needs_swap;
2889 * Do it before so that if perf_evsel__alloc_id fails, this
2890 * entry gets purged too at perf_evlist__delete().
2892 perf_evlist__add(session->evlist, evsel);
2894 nr_ids = f_attr.ids.size / sizeof(u64);
2896 * We don't have the cpu and thread maps on the header, so
2897 * for allocating the perf_sample_id table we fake 1 cpu and
2898 * hattr->ids threads.
2900 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2901 goto out_delete_evlist;
2903 lseek(fd, f_attr.ids.offset, SEEK_SET);
2905 for (j = 0; j < nr_ids; j++) {
2906 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2907 goto out_errno;
2909 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2912 lseek(fd, tmp, SEEK_SET);
2915 symbol_conf.nr_events = nr_attrs;
2917 perf_header__process_sections(header, fd, &session->tevent,
2918 perf_file_section__process);
2920 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2921 session->tevent.pevent))
2922 goto out_delete_evlist;
2924 return 0;
2925 out_errno:
2926 return -errno;
2928 out_delete_evlist:
2929 perf_evlist__delete(session->evlist);
2930 session->evlist = NULL;
2931 return -ENOMEM;
2934 int perf_event__synthesize_attr(struct perf_tool *tool,
2935 struct perf_event_attr *attr, u32 ids, u64 *id,
2936 perf_event__handler_t process)
2938 union perf_event *ev;
2939 size_t size;
2940 int err;
2942 size = sizeof(struct perf_event_attr);
2943 size = PERF_ALIGN(size, sizeof(u64));
2944 size += sizeof(struct perf_event_header);
2945 size += ids * sizeof(u64);
2947 ev = malloc(size);
2949 if (ev == NULL)
2950 return -ENOMEM;
2952 ev->attr.attr = *attr;
2953 memcpy(ev->attr.id, id, ids * sizeof(u64));
2955 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2956 ev->attr.header.size = (u16)size;
2958 if (ev->attr.header.size == size)
2959 err = process(tool, ev, NULL, NULL);
2960 else
2961 err = -E2BIG;
2963 free(ev);
2965 return err;
2968 static struct event_update_event *
2969 event_update_event__new(size_t size, u64 type, u64 id)
2971 struct event_update_event *ev;
2973 size += sizeof(*ev);
2974 size = PERF_ALIGN(size, sizeof(u64));
2976 ev = zalloc(size);
2977 if (ev) {
2978 ev->header.type = PERF_RECORD_EVENT_UPDATE;
2979 ev->header.size = (u16)size;
2980 ev->type = type;
2981 ev->id = id;
2983 return ev;
2987 perf_event__synthesize_event_update_unit(struct perf_tool *tool,
2988 struct perf_evsel *evsel,
2989 perf_event__handler_t process)
2991 struct event_update_event *ev;
2992 size_t size = strlen(evsel->unit);
2993 int err;
2995 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->id[0]);
2996 if (ev == NULL)
2997 return -ENOMEM;
2999 strncpy(ev->data, evsel->unit, size);
3000 err = process(tool, (union perf_event *)ev, NULL, NULL);
3001 free(ev);
3002 return err;
3006 perf_event__synthesize_event_update_scale(struct perf_tool *tool,
3007 struct perf_evsel *evsel,
3008 perf_event__handler_t process)
3010 struct event_update_event *ev;
3011 struct event_update_event_scale *ev_data;
3012 int err;
3014 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->id[0]);
3015 if (ev == NULL)
3016 return -ENOMEM;
3018 ev_data = (struct event_update_event_scale *) ev->data;
3019 ev_data->scale = evsel->scale;
3020 err = process(tool, (union perf_event*) ev, NULL, NULL);
3021 free(ev);
3022 return err;
3026 perf_event__synthesize_event_update_name(struct perf_tool *tool,
3027 struct perf_evsel *evsel,
3028 perf_event__handler_t process)
3030 struct event_update_event *ev;
3031 size_t len = strlen(evsel->name);
3032 int err;
3034 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->id[0]);
3035 if (ev == NULL)
3036 return -ENOMEM;
3038 strncpy(ev->data, evsel->name, len);
3039 err = process(tool, (union perf_event*) ev, NULL, NULL);
3040 free(ev);
3041 return err;
3045 perf_event__synthesize_event_update_cpus(struct perf_tool *tool,
3046 struct perf_evsel *evsel,
3047 perf_event__handler_t process)
3049 size_t size = sizeof(struct event_update_event);
3050 struct event_update_event *ev;
3051 int max, err;
3052 u16 type;
3054 if (!evsel->own_cpus)
3055 return 0;
3057 ev = cpu_map_data__alloc(evsel->own_cpus, &size, &type, &max);
3058 if (!ev)
3059 return -ENOMEM;
3061 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3062 ev->header.size = (u16)size;
3063 ev->type = PERF_EVENT_UPDATE__CPUS;
3064 ev->id = evsel->id[0];
3066 cpu_map_data__synthesize((struct cpu_map_data *) ev->data,
3067 evsel->own_cpus,
3068 type, max);
3070 err = process(tool, (union perf_event*) ev, NULL, NULL);
3071 free(ev);
3072 return err;
3075 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3077 struct event_update_event *ev = &event->event_update;
3078 struct event_update_event_scale *ev_scale;
3079 struct event_update_event_cpus *ev_cpus;
3080 struct cpu_map *map;
3081 size_t ret;
3083 ret = fprintf(fp, "\n... id: %" PRIu64 "\n", ev->id);
3085 switch (ev->type) {
3086 case PERF_EVENT_UPDATE__SCALE:
3087 ev_scale = (struct event_update_event_scale *) ev->data;
3088 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3089 break;
3090 case PERF_EVENT_UPDATE__UNIT:
3091 ret += fprintf(fp, "... unit: %s\n", ev->data);
3092 break;
3093 case PERF_EVENT_UPDATE__NAME:
3094 ret += fprintf(fp, "... name: %s\n", ev->data);
3095 break;
3096 case PERF_EVENT_UPDATE__CPUS:
3097 ev_cpus = (struct event_update_event_cpus *) ev->data;
3098 ret += fprintf(fp, "... ");
3100 map = cpu_map__new_data(&ev_cpus->cpus);
3101 if (map)
3102 ret += cpu_map__fprintf(map, fp);
3103 else
3104 ret += fprintf(fp, "failed to get cpus\n");
3105 break;
3106 default:
3107 ret += fprintf(fp, "... unknown type\n");
3108 break;
3111 return ret;
3114 int perf_event__synthesize_attrs(struct perf_tool *tool,
3115 struct perf_session *session,
3116 perf_event__handler_t process)
3118 struct perf_evsel *evsel;
3119 int err = 0;
3121 evlist__for_each(session->evlist, evsel) {
3122 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
3123 evsel->id, process);
3124 if (err) {
3125 pr_debug("failed to create perf header attribute\n");
3126 return err;
3130 return err;
3133 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3134 union perf_event *event,
3135 struct perf_evlist **pevlist)
3137 u32 i, ids, n_ids;
3138 struct perf_evsel *evsel;
3139 struct perf_evlist *evlist = *pevlist;
3141 if (evlist == NULL) {
3142 *pevlist = evlist = perf_evlist__new();
3143 if (evlist == NULL)
3144 return -ENOMEM;
3147 evsel = perf_evsel__new(&event->attr.attr);
3148 if (evsel == NULL)
3149 return -ENOMEM;
3151 perf_evlist__add(evlist, evsel);
3153 ids = event->header.size;
3154 ids -= (void *)&event->attr.id - (void *)event;
3155 n_ids = ids / sizeof(u64);
3157 * We don't have the cpu and thread maps on the header, so
3158 * for allocating the perf_sample_id table we fake 1 cpu and
3159 * hattr->ids threads.
3161 if (perf_evsel__alloc_id(evsel, 1, n_ids))
3162 return -ENOMEM;
3164 for (i = 0; i < n_ids; i++) {
3165 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
3168 symbol_conf.nr_events = evlist->nr_entries;
3170 return 0;
3173 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3174 union perf_event *event,
3175 struct perf_evlist **pevlist)
3177 struct event_update_event *ev = &event->event_update;
3178 struct event_update_event_scale *ev_scale;
3179 struct event_update_event_cpus *ev_cpus;
3180 struct perf_evlist *evlist;
3181 struct perf_evsel *evsel;
3182 struct cpu_map *map;
3184 if (!pevlist || *pevlist == NULL)
3185 return -EINVAL;
3187 evlist = *pevlist;
3189 evsel = perf_evlist__id2evsel(evlist, ev->id);
3190 if (evsel == NULL)
3191 return -EINVAL;
3193 switch (ev->type) {
3194 case PERF_EVENT_UPDATE__UNIT:
3195 evsel->unit = strdup(ev->data);
3196 break;
3197 case PERF_EVENT_UPDATE__NAME:
3198 evsel->name = strdup(ev->data);
3199 break;
3200 case PERF_EVENT_UPDATE__SCALE:
3201 ev_scale = (struct event_update_event_scale *) ev->data;
3202 evsel->scale = ev_scale->scale;
3203 case PERF_EVENT_UPDATE__CPUS:
3204 ev_cpus = (struct event_update_event_cpus *) ev->data;
3206 map = cpu_map__new_data(&ev_cpus->cpus);
3207 if (map)
3208 evsel->own_cpus = map;
3209 else
3210 pr_err("failed to get event_update cpus\n");
3211 default:
3212 break;
3215 return 0;
3218 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
3219 struct perf_evlist *evlist,
3220 perf_event__handler_t process)
3222 union perf_event ev;
3223 struct tracing_data *tdata;
3224 ssize_t size = 0, aligned_size = 0, padding;
3225 int err __maybe_unused = 0;
3228 * We are going to store the size of the data followed
3229 * by the data contents. Since the fd descriptor is a pipe,
3230 * we cannot seek back to store the size of the data once
3231 * we know it. Instead we:
3233 * - write the tracing data to the temp file
3234 * - get/write the data size to pipe
3235 * - write the tracing data from the temp file
3236 * to the pipe
3238 tdata = tracing_data_get(&evlist->entries, fd, true);
3239 if (!tdata)
3240 return -1;
3242 memset(&ev, 0, sizeof(ev));
3244 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
3245 size = tdata->size;
3246 aligned_size = PERF_ALIGN(size, sizeof(u64));
3247 padding = aligned_size - size;
3248 ev.tracing_data.header.size = sizeof(ev.tracing_data);
3249 ev.tracing_data.size = aligned_size;
3251 process(tool, &ev, NULL, NULL);
3254 * The put function will copy all the tracing data
3255 * stored in temp file to the pipe.
3257 tracing_data_put(tdata);
3259 write_padded(fd, NULL, 0, padding);
3261 return aligned_size;
3264 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
3265 union perf_event *event,
3266 struct perf_session *session)
3268 ssize_t size_read, padding, size = event->tracing_data.size;
3269 int fd = perf_data_file__fd(session->file);
3270 off_t offset = lseek(fd, 0, SEEK_CUR);
3271 char buf[BUFSIZ];
3273 /* setup for reading amidst mmap */
3274 lseek(fd, offset + sizeof(struct tracing_data_event),
3275 SEEK_SET);
3277 size_read = trace_report(fd, &session->tevent,
3278 session->repipe);
3279 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3281 if (readn(fd, buf, padding) < 0) {
3282 pr_err("%s: reading input file", __func__);
3283 return -1;
3285 if (session->repipe) {
3286 int retw = write(STDOUT_FILENO, buf, padding);
3287 if (retw <= 0 || retw != padding) {
3288 pr_err("%s: repiping tracing data padding", __func__);
3289 return -1;
3293 if (size_read + padding != size) {
3294 pr_err("%s: tracing data size mismatch", __func__);
3295 return -1;
3298 perf_evlist__prepare_tracepoint_events(session->evlist,
3299 session->tevent.pevent);
3301 return size_read + padding;
3304 int perf_event__synthesize_build_id(struct perf_tool *tool,
3305 struct dso *pos, u16 misc,
3306 perf_event__handler_t process,
3307 struct machine *machine)
3309 union perf_event ev;
3310 size_t len;
3311 int err = 0;
3313 if (!pos->hit)
3314 return err;
3316 memset(&ev, 0, sizeof(ev));
3318 len = pos->long_name_len + 1;
3319 len = PERF_ALIGN(len, NAME_ALIGN);
3320 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3321 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3322 ev.build_id.header.misc = misc;
3323 ev.build_id.pid = machine->pid;
3324 ev.build_id.header.size = sizeof(ev.build_id) + len;
3325 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3327 err = process(tool, &ev, NULL, machine);
3329 return err;
3332 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3333 union perf_event *event,
3334 struct perf_session *session)
3336 __event_process_build_id(&event->build_id,
3337 event->build_id.filename,
3338 session);
3339 return 0;