7 #include <linux/list.h>
8 #include <linux/kernel.h>
9 #include <linux/bitops.h>
10 #include <sys/utsname.h>
16 #include "trace-event.h"
27 static bool no_buildid_cache = false;
29 static u32 header_argc;
30 static const char **header_argv;
34 * must be a numerical value to let the endianness
35 * determine the memory layout. That way we are able
36 * to detect endianness when reading the perf.data file
39 * we check for legacy (PERFFILE) format.
41 static const char *__perf_magic1 = "PERFFILE";
42 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
43 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
45 #define PERF_MAGIC __perf_magic2
47 struct perf_file_attr {
48 struct perf_event_attr attr;
49 struct perf_file_section ids;
52 void perf_header__set_feat(struct perf_header *header, int feat)
54 set_bit(feat, header->adds_features);
57 void perf_header__clear_feat(struct perf_header *header, int feat)
59 clear_bit(feat, header->adds_features);
62 bool perf_header__has_feat(const struct perf_header *header, int feat)
64 return test_bit(feat, header->adds_features);
67 static int do_write(int fd, const void *buf, size_t size)
70 int ret = write(fd, buf, size);
84 static int write_padded(int fd, const void *bf, size_t count,
87 static const char zero_buf[NAME_ALIGN];
88 int err = do_write(fd, bf, count);
91 err = do_write(fd, zero_buf, count_aligned - count);
96 static int do_write_string(int fd, const char *str)
101 olen = strlen(str) + 1;
102 len = PERF_ALIGN(olen, NAME_ALIGN);
104 /* write len, incl. \0 */
105 ret = do_write(fd, &len, sizeof(len));
109 return write_padded(fd, str, olen, len);
112 static char *do_read_string(int fd, struct perf_header *ph)
118 sz = readn(fd, &len, sizeof(len));
119 if (sz < (ssize_t)sizeof(len))
129 ret = readn(fd, buf, len);
130 if (ret == (ssize_t)len) {
132 * strings are padded by zeroes
133 * thus the actual strlen of buf
134 * may be less than len
144 perf_header__set_cmdline(int argc, const char **argv)
149 * If header_argv has already been set, do not override it.
150 * This allows a command to set the cmdline, parse args and
151 * then call another builtin function that implements a
152 * command -- e.g, cmd_kvm calling cmd_record.
157 header_argc = (u32)argc;
159 /* do not include NULL termination */
160 header_argv = calloc(argc, sizeof(char *));
165 * must copy argv contents because it gets moved
166 * around during option parsing
168 for (i = 0; i < argc ; i++)
169 header_argv[i] = argv[i];
174 #define dsos__for_each_with_build_id(pos, head) \
175 list_for_each_entry(pos, head, node) \
176 if (!pos->has_build_id) \
180 static int write_buildid(const char *name, size_t name_len, u8 *build_id,
181 pid_t pid, u16 misc, int fd)
184 struct build_id_event b;
188 len = PERF_ALIGN(len, NAME_ALIGN);
190 memset(&b, 0, sizeof(b));
191 memcpy(&b.build_id, build_id, BUILD_ID_SIZE);
193 b.header.misc = misc;
194 b.header.size = sizeof(b) + len;
196 err = do_write(fd, &b, sizeof(b));
200 return write_padded(fd, name, name_len + 1, len);
203 static int __dsos__hit_all(struct list_head *head)
207 list_for_each_entry(pos, head, node)
213 static int machine__hit_all_dsos(struct machine *machine)
217 err = __dsos__hit_all(&machine->kernel_dsos.head);
221 return __dsos__hit_all(&machine->user_dsos.head);
224 int dsos__hit_all(struct perf_session *session)
229 err = machine__hit_all_dsos(&session->machines.host);
233 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
234 struct machine *pos = rb_entry(nd, struct machine, rb_node);
236 err = machine__hit_all_dsos(pos);
244 static int __dsos__write_buildid_table(struct list_head *head,
245 struct machine *machine,
246 pid_t pid, u16 misc, int fd)
251 dsos__for_each_with_build_id(pos, head) {
259 if (dso__is_vdso(pos)) {
260 name = pos->short_name;
261 name_len = pos->short_name_len + 1;
262 } else if (dso__is_kcore(pos)) {
263 machine__mmap_name(machine, nm, sizeof(nm));
265 name_len = strlen(nm) + 1;
267 name = pos->long_name;
268 name_len = pos->long_name_len + 1;
271 err = write_buildid(name, name_len, pos->build_id,
280 static int machine__write_buildid_table(struct machine *machine, int fd)
283 u16 kmisc = PERF_RECORD_MISC_KERNEL,
284 umisc = PERF_RECORD_MISC_USER;
286 if (!machine__is_host(machine)) {
287 kmisc = PERF_RECORD_MISC_GUEST_KERNEL;
288 umisc = PERF_RECORD_MISC_GUEST_USER;
291 err = __dsos__write_buildid_table(&machine->kernel_dsos.head, machine,
292 machine->pid, kmisc, fd);
294 err = __dsos__write_buildid_table(&machine->user_dsos.head,
295 machine, machine->pid, umisc,
300 static int dsos__write_buildid_table(struct perf_header *header, int fd)
302 struct perf_session *session = container_of(header,
303 struct perf_session, header);
305 int err = machine__write_buildid_table(&session->machines.host, fd);
310 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
311 struct machine *pos = rb_entry(nd, struct machine, rb_node);
312 err = machine__write_buildid_table(pos, fd);
319 int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
320 const char *name, bool is_kallsyms, bool is_vdso)
322 const size_t size = PATH_MAX;
323 char *realname, *filename = zalloc(size),
324 *linkname = zalloc(size), *targetname;
326 bool slash = is_kallsyms || is_vdso;
329 if (symbol_conf.kptr_restrict) {
330 pr_debug("Not caching a kptr_restrict'ed /proc/kallsyms\n");
334 realname = (char *) name;
336 realname = realpath(name, NULL);
338 if (realname == NULL || filename == NULL || linkname == NULL)
341 len = scnprintf(filename, size, "%s%s%s",
342 debugdir, slash ? "/" : "",
343 is_vdso ? DSO__NAME_VDSO : realname);
344 if (mkdir_p(filename, 0755))
347 snprintf(filename + len, size - len, "/%s", sbuild_id);
349 if (access(filename, F_OK)) {
351 if (copyfile("/proc/kallsyms", filename))
353 } else if (link(realname, filename) && copyfile(name, filename))
357 len = scnprintf(linkname, size, "%s/.build-id/%.2s",
358 debugdir, sbuild_id);
360 if (access(linkname, X_OK) && mkdir_p(linkname, 0755))
363 snprintf(linkname + len, size - len, "/%s", sbuild_id + 2);
364 targetname = filename + strlen(debugdir) - 5;
365 memcpy(targetname, "../..", 5);
367 if (symlink(targetname, linkname) == 0)
377 static int build_id_cache__add_b(const u8 *build_id, size_t build_id_size,
378 const char *name, const char *debugdir,
379 bool is_kallsyms, bool is_vdso)
381 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
383 build_id__sprintf(build_id, build_id_size, sbuild_id);
385 return build_id_cache__add_s(sbuild_id, debugdir, name,
386 is_kallsyms, is_vdso);
389 int build_id_cache__remove_s(const char *sbuild_id, const char *debugdir)
391 const size_t size = PATH_MAX;
392 char *filename = zalloc(size),
393 *linkname = zalloc(size);
396 if (filename == NULL || linkname == NULL)
399 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
400 debugdir, sbuild_id, sbuild_id + 2);
402 if (access(linkname, F_OK))
405 if (readlink(linkname, filename, size - 1) < 0)
408 if (unlink(linkname))
412 * Since the link is relative, we must make it absolute:
414 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
415 debugdir, sbuild_id, filename);
417 if (unlink(linkname))
427 static int dso__cache_build_id(struct dso *dso, struct machine *machine,
428 const char *debugdir)
430 bool is_kallsyms = dso->kernel && dso->long_name[0] != '/';
431 bool is_vdso = dso__is_vdso(dso);
432 const char *name = dso->long_name;
435 if (dso__is_kcore(dso)) {
437 machine__mmap_name(machine, nm, sizeof(nm));
440 return build_id_cache__add_b(dso->build_id, sizeof(dso->build_id), name,
441 debugdir, is_kallsyms, is_vdso);
444 static int __dsos__cache_build_ids(struct list_head *head,
445 struct machine *machine, const char *debugdir)
450 dsos__for_each_with_build_id(pos, head)
451 if (dso__cache_build_id(pos, machine, debugdir))
457 static int machine__cache_build_ids(struct machine *machine, const char *debugdir)
459 int ret = __dsos__cache_build_ids(&machine->kernel_dsos.head, machine,
461 ret |= __dsos__cache_build_ids(&machine->user_dsos.head, machine,
466 static int perf_session__cache_build_ids(struct perf_session *session)
470 char debugdir[PATH_MAX];
472 snprintf(debugdir, sizeof(debugdir), "%s", buildid_dir);
474 if (mkdir(debugdir, 0755) != 0 && errno != EEXIST)
477 ret = machine__cache_build_ids(&session->machines.host, debugdir);
479 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
480 struct machine *pos = rb_entry(nd, struct machine, rb_node);
481 ret |= machine__cache_build_ids(pos, debugdir);
486 static bool machine__read_build_ids(struct machine *machine, bool with_hits)
490 ret = __dsos__read_build_ids(&machine->kernel_dsos.head, with_hits);
491 ret |= __dsos__read_build_ids(&machine->user_dsos.head, with_hits);
495 static bool perf_session__read_build_ids(struct perf_session *session, bool with_hits)
498 bool ret = machine__read_build_ids(&session->machines.host, with_hits);
500 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
501 struct machine *pos = rb_entry(nd, struct machine, rb_node);
502 ret |= machine__read_build_ids(pos, with_hits);
508 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
509 struct perf_evlist *evlist)
511 return read_tracing_data(fd, &evlist->entries);
515 static int write_build_id(int fd, struct perf_header *h,
516 struct perf_evlist *evlist __maybe_unused)
518 struct perf_session *session;
521 session = container_of(h, struct perf_session, header);
523 if (!perf_session__read_build_ids(session, true))
526 err = dsos__write_buildid_table(h, fd);
528 pr_debug("failed to write buildid table\n");
531 if (!no_buildid_cache)
532 perf_session__cache_build_ids(session);
537 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
538 struct perf_evlist *evlist __maybe_unused)
547 return do_write_string(fd, uts.nodename);
550 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
551 struct perf_evlist *evlist __maybe_unused)
560 return do_write_string(fd, uts.release);
563 static int write_arch(int fd, struct perf_header *h __maybe_unused,
564 struct perf_evlist *evlist __maybe_unused)
573 return do_write_string(fd, uts.machine);
576 static int write_version(int fd, struct perf_header *h __maybe_unused,
577 struct perf_evlist *evlist __maybe_unused)
579 return do_write_string(fd, perf_version_string);
582 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
583 struct perf_evlist *evlist __maybe_unused)
586 #define CPUINFO_PROC NULL
591 const char *search = CPUINFO_PROC;
598 file = fopen("/proc/cpuinfo", "r");
602 while (getline(&buf, &len, file) > 0) {
603 ret = strncmp(buf, search, strlen(search));
613 p = strchr(buf, ':');
614 if (p && *(p+1) == ' ' && *(p+2))
620 /* squash extra space characters (branding string) */
627 while (*q && isspace(*q))
630 while ((*r++ = *q++));
634 ret = do_write_string(fd, s);
641 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
642 struct perf_evlist *evlist __maybe_unused)
648 nr = sysconf(_SC_NPROCESSORS_CONF);
652 nrc = (u32)(nr & UINT_MAX);
654 nr = sysconf(_SC_NPROCESSORS_ONLN);
658 nra = (u32)(nr & UINT_MAX);
660 ret = do_write(fd, &nrc, sizeof(nrc));
664 return do_write(fd, &nra, sizeof(nra));
667 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
668 struct perf_evlist *evlist)
670 struct perf_evsel *evsel;
674 nre = evlist->nr_entries;
677 * write number of events
679 ret = do_write(fd, &nre, sizeof(nre));
684 * size of perf_event_attr struct
686 sz = (u32)sizeof(evsel->attr);
687 ret = do_write(fd, &sz, sizeof(sz));
691 evlist__for_each(evlist, evsel) {
692 ret = do_write(fd, &evsel->attr, sz);
696 * write number of unique id per event
697 * there is one id per instance of an event
699 * copy into an nri to be independent of the
703 ret = do_write(fd, &nri, sizeof(nri));
708 * write event string as passed on cmdline
710 ret = do_write_string(fd, perf_evsel__name(evsel));
714 * write unique ids for this event
716 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
723 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
724 struct perf_evlist *evlist __maybe_unused)
726 char buf[MAXPATHLEN];
732 * actual atual path to perf binary
734 sprintf(proc, "/proc/%d/exe", getpid());
735 ret = readlink(proc, buf, sizeof(buf));
739 /* readlink() does not add null termination */
742 /* account for binary path */
745 ret = do_write(fd, &n, sizeof(n));
749 ret = do_write_string(fd, buf);
753 for (i = 0 ; i < header_argc; i++) {
754 ret = do_write_string(fd, header_argv[i]);
761 #define CORE_SIB_FMT \
762 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
763 #define THRD_SIB_FMT \
764 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
769 char **core_siblings;
770 char **thread_siblings;
773 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
776 char filename[MAXPATHLEN];
777 char *buf = NULL, *p;
783 sprintf(filename, CORE_SIB_FMT, cpu);
784 fp = fopen(filename, "r");
788 sret = getline(&buf, &len, fp);
793 p = strchr(buf, '\n');
797 for (i = 0; i < tp->core_sib; i++) {
798 if (!strcmp(buf, tp->core_siblings[i]))
801 if (i == tp->core_sib) {
802 tp->core_siblings[i] = buf;
810 sprintf(filename, THRD_SIB_FMT, cpu);
811 fp = fopen(filename, "r");
815 if (getline(&buf, &len, fp) <= 0)
818 p = strchr(buf, '\n');
822 for (i = 0; i < tp->thread_sib; i++) {
823 if (!strcmp(buf, tp->thread_siblings[i]))
826 if (i == tp->thread_sib) {
827 tp->thread_siblings[i] = buf;
839 static void free_cpu_topo(struct cpu_topo *tp)
846 for (i = 0 ; i < tp->core_sib; i++)
847 zfree(&tp->core_siblings[i]);
849 for (i = 0 ; i < tp->thread_sib; i++)
850 zfree(&tp->thread_siblings[i]);
855 static struct cpu_topo *build_cpu_topology(void)
864 ncpus = sysconf(_SC_NPROCESSORS_CONF);
868 nr = (u32)(ncpus & UINT_MAX);
870 sz = nr * sizeof(char *);
872 addr = calloc(1, sizeof(*tp) + 2 * sz);
879 tp->core_siblings = addr;
881 tp->thread_siblings = addr;
883 for (i = 0; i < nr; i++) {
884 ret = build_cpu_topo(tp, i);
895 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
896 struct perf_evlist *evlist __maybe_unused)
902 tp = build_cpu_topology();
906 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
910 for (i = 0; i < tp->core_sib; i++) {
911 ret = do_write_string(fd, tp->core_siblings[i]);
915 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
919 for (i = 0; i < tp->thread_sib; i++) {
920 ret = do_write_string(fd, tp->thread_siblings[i]);
931 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
932 struct perf_evlist *evlist __maybe_unused)
940 fp = fopen("/proc/meminfo", "r");
944 while (getline(&buf, &len, fp) > 0) {
945 ret = strncmp(buf, "MemTotal:", 9);
950 n = sscanf(buf, "%*s %"PRIu64, &mem);
952 ret = do_write(fd, &mem, sizeof(mem));
959 static int write_topo_node(int fd, int node)
961 char str[MAXPATHLEN];
963 char *buf = NULL, *p;
966 u64 mem_total, mem_free, mem;
969 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
970 fp = fopen(str, "r");
974 while (getline(&buf, &len, fp) > 0) {
975 /* skip over invalid lines */
976 if (!strchr(buf, ':'))
978 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
980 if (!strcmp(field, "MemTotal:"))
982 if (!strcmp(field, "MemFree:"))
989 ret = do_write(fd, &mem_total, sizeof(u64));
993 ret = do_write(fd, &mem_free, sizeof(u64));
998 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
1000 fp = fopen(str, "r");
1004 if (getline(&buf, &len, fp) <= 0)
1007 p = strchr(buf, '\n');
1011 ret = do_write_string(fd, buf);
1019 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
1020 struct perf_evlist *evlist __maybe_unused)
1025 struct cpu_map *node_map = NULL;
1030 fp = fopen("/sys/devices/system/node/online", "r");
1034 if (getline(&buf, &len, fp) <= 0)
1037 c = strchr(buf, '\n');
1041 node_map = cpu_map__new(buf);
1045 nr = (u32)node_map->nr;
1047 ret = do_write(fd, &nr, sizeof(nr));
1051 for (i = 0; i < nr; i++) {
1052 j = (u32)node_map->map[i];
1053 ret = do_write(fd, &j, sizeof(j));
1057 ret = write_topo_node(fd, i);
1071 * struct pmu_mappings {
1080 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
1081 struct perf_evlist *evlist __maybe_unused)
1083 struct perf_pmu *pmu = NULL;
1084 off_t offset = lseek(fd, 0, SEEK_CUR);
1088 /* write real pmu_num later */
1089 ret = do_write(fd, &pmu_num, sizeof(pmu_num));
1093 while ((pmu = perf_pmu__scan(pmu))) {
1098 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
1102 ret = do_write_string(fd, pmu->name);
1107 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
1109 lseek(fd, offset, SEEK_SET);
1119 * struct group_descs {
1121 * struct group_desc {
1128 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
1129 struct perf_evlist *evlist)
1131 u32 nr_groups = evlist->nr_groups;
1132 struct perf_evsel *evsel;
1135 ret = do_write(fd, &nr_groups, sizeof(nr_groups));
1139 evlist__for_each(evlist, evsel) {
1140 if (perf_evsel__is_group_leader(evsel) &&
1141 evsel->nr_members > 1) {
1142 const char *name = evsel->group_name ?: "{anon_group}";
1143 u32 leader_idx = evsel->idx;
1144 u32 nr_members = evsel->nr_members;
1146 ret = do_write_string(fd, name);
1150 ret = do_write(fd, &leader_idx, sizeof(leader_idx));
1154 ret = do_write(fd, &nr_members, sizeof(nr_members));
1163 * default get_cpuid(): nothing gets recorded
1164 * actual implementation must be in arch/$(ARCH)/util/header.c
1166 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
1167 size_t sz __maybe_unused)
1172 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
1173 struct perf_evlist *evlist __maybe_unused)
1178 ret = get_cpuid(buffer, sizeof(buffer));
1184 return do_write_string(fd, buffer);
1187 static int write_branch_stack(int fd __maybe_unused,
1188 struct perf_header *h __maybe_unused,
1189 struct perf_evlist *evlist __maybe_unused)
1194 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
1197 fprintf(fp, "# hostname : %s\n", ph->env.hostname);
1200 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
1203 fprintf(fp, "# os release : %s\n", ph->env.os_release);
1206 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1208 fprintf(fp, "# arch : %s\n", ph->env.arch);
1211 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
1214 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
1217 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
1220 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
1221 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
1224 static void print_version(struct perf_header *ph, int fd __maybe_unused,
1227 fprintf(fp, "# perf version : %s\n", ph->env.version);
1230 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
1236 nr = ph->env.nr_cmdline;
1237 str = ph->env.cmdline;
1239 fprintf(fp, "# cmdline : ");
1241 for (i = 0; i < nr; i++) {
1242 fprintf(fp, "%s ", str);
1243 str += strlen(str) + 1;
1248 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
1254 nr = ph->env.nr_sibling_cores;
1255 str = ph->env.sibling_cores;
1257 for (i = 0; i < nr; i++) {
1258 fprintf(fp, "# sibling cores : %s\n", str);
1259 str += strlen(str) + 1;
1262 nr = ph->env.nr_sibling_threads;
1263 str = ph->env.sibling_threads;
1265 for (i = 0; i < nr; i++) {
1266 fprintf(fp, "# sibling threads : %s\n", str);
1267 str += strlen(str) + 1;
1271 static void free_event_desc(struct perf_evsel *events)
1273 struct perf_evsel *evsel;
1278 for (evsel = events; evsel->attr.size; evsel++) {
1279 zfree(&evsel->name);
1286 static struct perf_evsel *
1287 read_event_desc(struct perf_header *ph, int fd)
1289 struct perf_evsel *evsel, *events = NULL;
1292 u32 nre, sz, nr, i, j;
1296 /* number of events */
1297 ret = readn(fd, &nre, sizeof(nre));
1298 if (ret != (ssize_t)sizeof(nre))
1302 nre = bswap_32(nre);
1304 ret = readn(fd, &sz, sizeof(sz));
1305 if (ret != (ssize_t)sizeof(sz))
1311 /* buffer to hold on file attr struct */
1316 /* the last event terminates with evsel->attr.size == 0: */
1317 events = calloc(nre + 1, sizeof(*events));
1321 msz = sizeof(evsel->attr);
1325 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1329 * must read entire on-file attr struct to
1330 * sync up with layout.
1332 ret = readn(fd, buf, sz);
1333 if (ret != (ssize_t)sz)
1337 perf_event__attr_swap(buf);
1339 memcpy(&evsel->attr, buf, msz);
1341 ret = readn(fd, &nr, sizeof(nr));
1342 if (ret != (ssize_t)sizeof(nr))
1345 if (ph->needs_swap) {
1347 evsel->needs_swap = true;
1350 evsel->name = do_read_string(fd, ph);
1355 id = calloc(nr, sizeof(*id));
1361 for (j = 0 ; j < nr; j++) {
1362 ret = readn(fd, id, sizeof(*id));
1363 if (ret != (ssize_t)sizeof(*id))
1366 *id = bswap_64(*id);
1375 free_event_desc(events);
1380 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1382 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1387 fprintf(fp, "# event desc: not available or unable to read\n");
1391 for (evsel = events; evsel->attr.size; evsel++) {
1392 fprintf(fp, "# event : name = %s, ", evsel->name);
1394 fprintf(fp, "type = %d, config = 0x%"PRIx64
1395 ", config1 = 0x%"PRIx64", config2 = 0x%"PRIx64,
1397 (u64)evsel->attr.config,
1398 (u64)evsel->attr.config1,
1399 (u64)evsel->attr.config2);
1401 fprintf(fp, ", excl_usr = %d, excl_kern = %d",
1402 evsel->attr.exclude_user,
1403 evsel->attr.exclude_kernel);
1405 fprintf(fp, ", excl_host = %d, excl_guest = %d",
1406 evsel->attr.exclude_host,
1407 evsel->attr.exclude_guest);
1409 fprintf(fp, ", precise_ip = %d", evsel->attr.precise_ip);
1411 fprintf(fp, ", attr_mmap2 = %d", evsel->attr.mmap2);
1412 fprintf(fp, ", attr_mmap = %d", evsel->attr.mmap);
1413 fprintf(fp, ", attr_mmap_data = %d", evsel->attr.mmap_data);
1415 fprintf(fp, ", id = {");
1416 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1419 fprintf(fp, " %"PRIu64, *id);
1427 free_event_desc(events);
1430 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1433 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1436 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1441 uint64_t mem_total, mem_free;
1444 nr = ph->env.nr_numa_nodes;
1445 str = ph->env.numa_nodes;
1447 for (i = 0; i < nr; i++) {
1449 c = strtoul(str, &tmp, 0);
1454 mem_total = strtoull(str, &tmp, 0);
1459 mem_free = strtoull(str, &tmp, 0);
1463 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1464 " free = %"PRIu64" kB\n",
1465 c, mem_total, mem_free);
1468 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1470 str += strlen(str) + 1;
1474 fprintf(fp, "# numa topology : not available\n");
1477 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1479 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1482 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1483 int fd __maybe_unused, FILE *fp)
1485 fprintf(fp, "# contains samples with branch stack\n");
1488 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1491 const char *delimiter = "# pmu mappings: ";
1496 pmu_num = ph->env.nr_pmu_mappings;
1498 fprintf(fp, "# pmu mappings: not available\n");
1502 str = ph->env.pmu_mappings;
1505 type = strtoul(str, &tmp, 0);
1510 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1513 str += strlen(str) + 1;
1522 fprintf(fp, "# pmu mappings: unable to read\n");
1525 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1528 struct perf_session *session;
1529 struct perf_evsel *evsel;
1532 session = container_of(ph, struct perf_session, header);
1534 evlist__for_each(session->evlist, evsel) {
1535 if (perf_evsel__is_group_leader(evsel) &&
1536 evsel->nr_members > 1) {
1537 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1538 perf_evsel__name(evsel));
1540 nr = evsel->nr_members - 1;
1542 fprintf(fp, ",%s", perf_evsel__name(evsel));
1550 static int __event_process_build_id(struct build_id_event *bev,
1552 struct perf_session *session)
1556 struct machine *machine;
1559 enum dso_kernel_type dso_type;
1561 machine = perf_session__findnew_machine(session, bev->pid);
1565 misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1568 case PERF_RECORD_MISC_KERNEL:
1569 dso_type = DSO_TYPE_KERNEL;
1570 dsos = &machine->kernel_dsos;
1572 case PERF_RECORD_MISC_GUEST_KERNEL:
1573 dso_type = DSO_TYPE_GUEST_KERNEL;
1574 dsos = &machine->kernel_dsos;
1576 case PERF_RECORD_MISC_USER:
1577 case PERF_RECORD_MISC_GUEST_USER:
1578 dso_type = DSO_TYPE_USER;
1579 dsos = &machine->user_dsos;
1585 dso = __dsos__findnew(dsos, filename);
1587 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1589 dso__set_build_id(dso, &bev->build_id);
1591 if (filename[0] == '[')
1592 dso->kernel = dso_type;
1594 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1596 pr_debug("build id event received for %s: %s\n",
1597 dso->long_name, sbuild_id);
1605 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1606 int input, u64 offset, u64 size)
1608 struct perf_session *session = container_of(header, struct perf_session, header);
1610 struct perf_event_header header;
1611 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1614 struct build_id_event bev;
1615 char filename[PATH_MAX];
1616 u64 limit = offset + size;
1618 while (offset < limit) {
1621 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1624 if (header->needs_swap)
1625 perf_event_header__bswap(&old_bev.header);
1627 len = old_bev.header.size - sizeof(old_bev);
1628 if (readn(input, filename, len) != len)
1631 bev.header = old_bev.header;
1634 * As the pid is the missing value, we need to fill
1635 * it properly. The header.misc value give us nice hint.
1637 bev.pid = HOST_KERNEL_ID;
1638 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1639 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1640 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1642 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1643 __event_process_build_id(&bev, filename, session);
1645 offset += bev.header.size;
1651 static int perf_header__read_build_ids(struct perf_header *header,
1652 int input, u64 offset, u64 size)
1654 struct perf_session *session = container_of(header, struct perf_session, header);
1655 struct build_id_event bev;
1656 char filename[PATH_MAX];
1657 u64 limit = offset + size, orig_offset = offset;
1660 while (offset < limit) {
1663 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1666 if (header->needs_swap)
1667 perf_event_header__bswap(&bev.header);
1669 len = bev.header.size - sizeof(bev);
1670 if (readn(input, filename, len) != len)
1673 * The a1645ce1 changeset:
1675 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1677 * Added a field to struct build_id_event that broke the file
1680 * Since the kernel build-id is the first entry, process the
1681 * table using the old format if the well known
1682 * '[kernel.kallsyms]' string for the kernel build-id has the
1683 * first 4 characters chopped off (where the pid_t sits).
1685 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1686 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1688 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1691 __event_process_build_id(&bev, filename, session);
1693 offset += bev.header.size;
1700 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1701 struct perf_header *ph __maybe_unused,
1704 ssize_t ret = trace_report(fd, data, false);
1705 return ret < 0 ? -1 : 0;
1708 static int process_build_id(struct perf_file_section *section,
1709 struct perf_header *ph, int fd,
1710 void *data __maybe_unused)
1712 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1713 pr_debug("Failed to read buildids, continuing...\n");
1717 static int process_hostname(struct perf_file_section *section __maybe_unused,
1718 struct perf_header *ph, int fd,
1719 void *data __maybe_unused)
1721 ph->env.hostname = do_read_string(fd, ph);
1722 return ph->env.hostname ? 0 : -ENOMEM;
1725 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1726 struct perf_header *ph, int fd,
1727 void *data __maybe_unused)
1729 ph->env.os_release = do_read_string(fd, ph);
1730 return ph->env.os_release ? 0 : -ENOMEM;
1733 static int process_version(struct perf_file_section *section __maybe_unused,
1734 struct perf_header *ph, int fd,
1735 void *data __maybe_unused)
1737 ph->env.version = do_read_string(fd, ph);
1738 return ph->env.version ? 0 : -ENOMEM;
1741 static int process_arch(struct perf_file_section *section __maybe_unused,
1742 struct perf_header *ph, int fd,
1743 void *data __maybe_unused)
1745 ph->env.arch = do_read_string(fd, ph);
1746 return ph->env.arch ? 0 : -ENOMEM;
1749 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1750 struct perf_header *ph, int fd,
1751 void *data __maybe_unused)
1756 ret = readn(fd, &nr, sizeof(nr));
1757 if (ret != sizeof(nr))
1763 ph->env.nr_cpus_online = nr;
1765 ret = readn(fd, &nr, sizeof(nr));
1766 if (ret != sizeof(nr))
1772 ph->env.nr_cpus_avail = nr;
1776 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1777 struct perf_header *ph, int fd,
1778 void *data __maybe_unused)
1780 ph->env.cpu_desc = do_read_string(fd, ph);
1781 return ph->env.cpu_desc ? 0 : -ENOMEM;
1784 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1785 struct perf_header *ph, int fd,
1786 void *data __maybe_unused)
1788 ph->env.cpuid = do_read_string(fd, ph);
1789 return ph->env.cpuid ? 0 : -ENOMEM;
1792 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1793 struct perf_header *ph, int fd,
1794 void *data __maybe_unused)
1799 ret = readn(fd, &mem, sizeof(mem));
1800 if (ret != sizeof(mem))
1804 mem = bswap_64(mem);
1806 ph->env.total_mem = mem;
1810 static struct perf_evsel *
1811 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1813 struct perf_evsel *evsel;
1815 evlist__for_each(evlist, evsel) {
1816 if (evsel->idx == idx)
1824 perf_evlist__set_event_name(struct perf_evlist *evlist,
1825 struct perf_evsel *event)
1827 struct perf_evsel *evsel;
1832 evsel = perf_evlist__find_by_index(evlist, event->idx);
1839 evsel->name = strdup(event->name);
1843 process_event_desc(struct perf_file_section *section __maybe_unused,
1844 struct perf_header *header, int fd,
1845 void *data __maybe_unused)
1847 struct perf_session *session;
1848 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1853 session = container_of(header, struct perf_session, header);
1854 for (evsel = events; evsel->attr.size; evsel++)
1855 perf_evlist__set_event_name(session->evlist, evsel);
1857 free_event_desc(events);
1862 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1863 struct perf_header *ph, int fd,
1864 void *data __maybe_unused)
1871 ret = readn(fd, &nr, sizeof(nr));
1872 if (ret != sizeof(nr))
1878 ph->env.nr_cmdline = nr;
1879 strbuf_init(&sb, 128);
1881 for (i = 0; i < nr; i++) {
1882 str = do_read_string(fd, ph);
1886 /* include a NULL character at the end */
1887 strbuf_add(&sb, str, strlen(str) + 1);
1890 ph->env.cmdline = strbuf_detach(&sb, NULL);
1894 strbuf_release(&sb);
1898 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1899 struct perf_header *ph, int fd,
1900 void *data __maybe_unused)
1907 ret = readn(fd, &nr, sizeof(nr));
1908 if (ret != sizeof(nr))
1914 ph->env.nr_sibling_cores = nr;
1915 strbuf_init(&sb, 128);
1917 for (i = 0; i < nr; i++) {
1918 str = do_read_string(fd, ph);
1922 /* include a NULL character at the end */
1923 strbuf_add(&sb, str, strlen(str) + 1);
1926 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1928 ret = readn(fd, &nr, sizeof(nr));
1929 if (ret != sizeof(nr))
1935 ph->env.nr_sibling_threads = nr;
1937 for (i = 0; i < nr; i++) {
1938 str = do_read_string(fd, ph);
1942 /* include a NULL character at the end */
1943 strbuf_add(&sb, str, strlen(str) + 1);
1946 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1950 strbuf_release(&sb);
1954 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1955 struct perf_header *ph, int fd,
1956 void *data __maybe_unused)
1961 uint64_t mem_total, mem_free;
1965 ret = readn(fd, &nr, sizeof(nr));
1966 if (ret != sizeof(nr))
1972 ph->env.nr_numa_nodes = nr;
1973 strbuf_init(&sb, 256);
1975 for (i = 0; i < nr; i++) {
1977 ret = readn(fd, &node, sizeof(node));
1978 if (ret != sizeof(node))
1981 ret = readn(fd, &mem_total, sizeof(u64));
1982 if (ret != sizeof(u64))
1985 ret = readn(fd, &mem_free, sizeof(u64));
1986 if (ret != sizeof(u64))
1989 if (ph->needs_swap) {
1990 node = bswap_32(node);
1991 mem_total = bswap_64(mem_total);
1992 mem_free = bswap_64(mem_free);
1995 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1996 node, mem_total, mem_free);
1998 str = do_read_string(fd, ph);
2002 /* include a NULL character at the end */
2003 strbuf_add(&sb, str, strlen(str) + 1);
2006 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
2010 strbuf_release(&sb);
2014 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
2015 struct perf_header *ph, int fd,
2016 void *data __maybe_unused)
2024 ret = readn(fd, &pmu_num, sizeof(pmu_num));
2025 if (ret != sizeof(pmu_num))
2029 pmu_num = bswap_32(pmu_num);
2032 pr_debug("pmu mappings not available\n");
2036 ph->env.nr_pmu_mappings = pmu_num;
2037 strbuf_init(&sb, 128);
2040 if (readn(fd, &type, sizeof(type)) != sizeof(type))
2043 type = bswap_32(type);
2045 name = do_read_string(fd, ph);
2049 strbuf_addf(&sb, "%u:%s", type, name);
2050 /* include a NULL character at the end */
2051 strbuf_add(&sb, "", 1);
2056 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2060 strbuf_release(&sb);
2064 static int process_group_desc(struct perf_file_section *section __maybe_unused,
2065 struct perf_header *ph, int fd,
2066 void *data __maybe_unused)
2069 u32 i, nr, nr_groups;
2070 struct perf_session *session;
2071 struct perf_evsel *evsel, *leader = NULL;
2078 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
2082 nr_groups = bswap_32(nr_groups);
2084 ph->env.nr_groups = nr_groups;
2086 pr_debug("group desc not available\n");
2090 desc = calloc(nr_groups, sizeof(*desc));
2094 for (i = 0; i < nr_groups; i++) {
2095 desc[i].name = do_read_string(fd, ph);
2099 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
2102 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
2105 if (ph->needs_swap) {
2106 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
2107 desc[i].nr_members = bswap_32(desc[i].nr_members);
2112 * Rebuild group relationship based on the group_desc
2114 session = container_of(ph, struct perf_session, header);
2115 session->evlist->nr_groups = nr_groups;
2118 evlist__for_each(session->evlist, evsel) {
2119 if (evsel->idx == (int) desc[i].leader_idx) {
2120 evsel->leader = evsel;
2121 /* {anon_group} is a dummy name */
2122 if (strcmp(desc[i].name, "{anon_group}")) {
2123 evsel->group_name = desc[i].name;
2124 desc[i].name = NULL;
2126 evsel->nr_members = desc[i].nr_members;
2128 if (i >= nr_groups || nr > 0) {
2129 pr_debug("invalid group desc\n");
2134 nr = evsel->nr_members - 1;
2137 /* This is a group member */
2138 evsel->leader = leader;
2144 if (i != nr_groups || nr != 0) {
2145 pr_debug("invalid group desc\n");
2151 for (i = 0; i < nr_groups; i++)
2152 zfree(&desc[i].name);
2158 struct feature_ops {
2159 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
2160 void (*print)(struct perf_header *h, int fd, FILE *fp);
2161 int (*process)(struct perf_file_section *section,
2162 struct perf_header *h, int fd, void *data);
2167 #define FEAT_OPA(n, func) \
2168 [n] = { .name = #n, .write = write_##func, .print = print_##func }
2169 #define FEAT_OPP(n, func) \
2170 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2171 .process = process_##func }
2172 #define FEAT_OPF(n, func) \
2173 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2174 .process = process_##func, .full_only = true }
2176 /* feature_ops not implemented: */
2177 #define print_tracing_data NULL
2178 #define print_build_id NULL
2180 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2181 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
2182 FEAT_OPP(HEADER_BUILD_ID, build_id),
2183 FEAT_OPP(HEADER_HOSTNAME, hostname),
2184 FEAT_OPP(HEADER_OSRELEASE, osrelease),
2185 FEAT_OPP(HEADER_VERSION, version),
2186 FEAT_OPP(HEADER_ARCH, arch),
2187 FEAT_OPP(HEADER_NRCPUS, nrcpus),
2188 FEAT_OPP(HEADER_CPUDESC, cpudesc),
2189 FEAT_OPP(HEADER_CPUID, cpuid),
2190 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
2191 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
2192 FEAT_OPP(HEADER_CMDLINE, cmdline),
2193 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
2194 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
2195 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
2196 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
2197 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
2200 struct header_print_data {
2202 bool full; /* extended list of headers */
2205 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2206 struct perf_header *ph,
2207 int feat, int fd, void *data)
2209 struct header_print_data *hd = data;
2211 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2212 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2213 "%d, continuing...\n", section->offset, feat);
2216 if (feat >= HEADER_LAST_FEATURE) {
2217 pr_warning("unknown feature %d\n", feat);
2220 if (!feat_ops[feat].print)
2223 if (!feat_ops[feat].full_only || hd->full)
2224 feat_ops[feat].print(ph, fd, hd->fp);
2226 fprintf(hd->fp, "# %s info available, use -I to display\n",
2227 feat_ops[feat].name);
2232 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2234 struct header_print_data hd;
2235 struct perf_header *header = &session->header;
2236 int fd = perf_data_file__fd(session->file);
2240 perf_header__process_sections(header, fd, &hd,
2241 perf_file_section__fprintf_info);
2245 static int do_write_feat(int fd, struct perf_header *h, int type,
2246 struct perf_file_section **p,
2247 struct perf_evlist *evlist)
2252 if (perf_header__has_feat(h, type)) {
2253 if (!feat_ops[type].write)
2256 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2258 err = feat_ops[type].write(fd, h, evlist);
2260 pr_debug("failed to write feature %d\n", type);
2262 /* undo anything written */
2263 lseek(fd, (*p)->offset, SEEK_SET);
2267 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2273 static int perf_header__adds_write(struct perf_header *header,
2274 struct perf_evlist *evlist, int fd)
2277 struct perf_file_section *feat_sec, *p;
2283 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2287 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2288 if (feat_sec == NULL)
2291 sec_size = sizeof(*feat_sec) * nr_sections;
2293 sec_start = header->feat_offset;
2294 lseek(fd, sec_start + sec_size, SEEK_SET);
2296 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2297 if (do_write_feat(fd, header, feat, &p, evlist))
2298 perf_header__clear_feat(header, feat);
2301 lseek(fd, sec_start, SEEK_SET);
2303 * may write more than needed due to dropped feature, but
2304 * this is okay, reader will skip the mising entries
2306 err = do_write(fd, feat_sec, sec_size);
2308 pr_debug("failed to write feature section\n");
2313 int perf_header__write_pipe(int fd)
2315 struct perf_pipe_file_header f_header;
2318 f_header = (struct perf_pipe_file_header){
2319 .magic = PERF_MAGIC,
2320 .size = sizeof(f_header),
2323 err = do_write(fd, &f_header, sizeof(f_header));
2325 pr_debug("failed to write perf pipe header\n");
2332 int perf_session__write_header(struct perf_session *session,
2333 struct perf_evlist *evlist,
2334 int fd, bool at_exit)
2336 struct perf_file_header f_header;
2337 struct perf_file_attr f_attr;
2338 struct perf_header *header = &session->header;
2339 struct perf_evsel *evsel;
2343 lseek(fd, sizeof(f_header), SEEK_SET);
2345 evlist__for_each(session->evlist, evsel) {
2346 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2347 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2349 pr_debug("failed to write perf header\n");
2354 attr_offset = lseek(fd, 0, SEEK_CUR);
2356 evlist__for_each(evlist, evsel) {
2357 f_attr = (struct perf_file_attr){
2358 .attr = evsel->attr,
2360 .offset = evsel->id_offset,
2361 .size = evsel->ids * sizeof(u64),
2364 err = do_write(fd, &f_attr, sizeof(f_attr));
2366 pr_debug("failed to write perf header attribute\n");
2371 if (!header->data_offset)
2372 header->data_offset = lseek(fd, 0, SEEK_CUR);
2373 header->feat_offset = header->data_offset + header->data_size;
2376 err = perf_header__adds_write(header, evlist, fd);
2381 f_header = (struct perf_file_header){
2382 .magic = PERF_MAGIC,
2383 .size = sizeof(f_header),
2384 .attr_size = sizeof(f_attr),
2386 .offset = attr_offset,
2387 .size = evlist->nr_entries * sizeof(f_attr),
2390 .offset = header->data_offset,
2391 .size = header->data_size,
2393 /* event_types is ignored, store zeros */
2396 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2398 lseek(fd, 0, SEEK_SET);
2399 err = do_write(fd, &f_header, sizeof(f_header));
2401 pr_debug("failed to write perf header\n");
2404 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2409 static int perf_header__getbuffer64(struct perf_header *header,
2410 int fd, void *buf, size_t size)
2412 if (readn(fd, buf, size) <= 0)
2415 if (header->needs_swap)
2416 mem_bswap_64(buf, size);
2421 int perf_header__process_sections(struct perf_header *header, int fd,
2423 int (*process)(struct perf_file_section *section,
2424 struct perf_header *ph,
2425 int feat, int fd, void *data))
2427 struct perf_file_section *feat_sec, *sec;
2433 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2437 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2441 sec_size = sizeof(*feat_sec) * nr_sections;
2443 lseek(fd, header->feat_offset, SEEK_SET);
2445 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2449 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2450 err = process(sec++, header, feat, fd, data);
2460 static const int attr_file_abi_sizes[] = {
2461 [0] = PERF_ATTR_SIZE_VER0,
2462 [1] = PERF_ATTR_SIZE_VER1,
2463 [2] = PERF_ATTR_SIZE_VER2,
2464 [3] = PERF_ATTR_SIZE_VER3,
2469 * In the legacy file format, the magic number is not used to encode endianness.
2470 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2471 * on ABI revisions, we need to try all combinations for all endianness to
2472 * detect the endianness.
2474 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2476 uint64_t ref_size, attr_size;
2479 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2480 ref_size = attr_file_abi_sizes[i]
2481 + sizeof(struct perf_file_section);
2482 if (hdr_sz != ref_size) {
2483 attr_size = bswap_64(hdr_sz);
2484 if (attr_size != ref_size)
2487 ph->needs_swap = true;
2489 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2494 /* could not determine endianness */
2498 #define PERF_PIPE_HDR_VER0 16
2500 static const size_t attr_pipe_abi_sizes[] = {
2501 [0] = PERF_PIPE_HDR_VER0,
2506 * In the legacy pipe format, there is an implicit assumption that endiannesss
2507 * between host recording the samples, and host parsing the samples is the
2508 * same. This is not always the case given that the pipe output may always be
2509 * redirected into a file and analyzed on a different machine with possibly a
2510 * different endianness and perf_event ABI revsions in the perf tool itself.
2512 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2517 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2518 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2519 attr_size = bswap_64(hdr_sz);
2520 if (attr_size != hdr_sz)
2523 ph->needs_swap = true;
2525 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2531 bool is_perf_magic(u64 magic)
2533 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2534 || magic == __perf_magic2
2535 || magic == __perf_magic2_sw)
2541 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2542 bool is_pipe, struct perf_header *ph)
2546 /* check for legacy format */
2547 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2549 ph->version = PERF_HEADER_VERSION_1;
2550 pr_debug("legacy perf.data format\n");
2552 return try_all_pipe_abis(hdr_sz, ph);
2554 return try_all_file_abis(hdr_sz, ph);
2557 * the new magic number serves two purposes:
2558 * - unique number to identify actual perf.data files
2559 * - encode endianness of file
2562 /* check magic number with one endianness */
2563 if (magic == __perf_magic2)
2566 /* check magic number with opposite endianness */
2567 if (magic != __perf_magic2_sw)
2570 ph->needs_swap = true;
2571 ph->version = PERF_HEADER_VERSION_2;
2576 int perf_file_header__read(struct perf_file_header *header,
2577 struct perf_header *ph, int fd)
2581 lseek(fd, 0, SEEK_SET);
2583 ret = readn(fd, header, sizeof(*header));
2587 if (check_magic_endian(header->magic,
2588 header->attr_size, false, ph) < 0) {
2589 pr_debug("magic/endian check failed\n");
2593 if (ph->needs_swap) {
2594 mem_bswap_64(header, offsetof(struct perf_file_header,
2598 if (header->size != sizeof(*header)) {
2599 /* Support the previous format */
2600 if (header->size == offsetof(typeof(*header), adds_features))
2601 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2604 } else if (ph->needs_swap) {
2606 * feature bitmap is declared as an array of unsigned longs --
2607 * not good since its size can differ between the host that
2608 * generated the data file and the host analyzing the file.
2610 * We need to handle endianness, but we don't know the size of
2611 * the unsigned long where the file was generated. Take a best
2612 * guess at determining it: try 64-bit swap first (ie., file
2613 * created on a 64-bit host), and check if the hostname feature
2614 * bit is set (this feature bit is forced on as of fbe96f2).
2615 * If the bit is not, undo the 64-bit swap and try a 32-bit
2616 * swap. If the hostname bit is still not set (e.g., older data
2617 * file), punt and fallback to the original behavior --
2618 * clearing all feature bits and setting buildid.
2620 mem_bswap_64(&header->adds_features,
2621 BITS_TO_U64(HEADER_FEAT_BITS));
2623 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2625 mem_bswap_64(&header->adds_features,
2626 BITS_TO_U64(HEADER_FEAT_BITS));
2629 mem_bswap_32(&header->adds_features,
2630 BITS_TO_U32(HEADER_FEAT_BITS));
2633 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2634 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2635 set_bit(HEADER_BUILD_ID, header->adds_features);
2639 memcpy(&ph->adds_features, &header->adds_features,
2640 sizeof(ph->adds_features));
2642 ph->data_offset = header->data.offset;
2643 ph->data_size = header->data.size;
2644 ph->feat_offset = header->data.offset + header->data.size;
2648 static int perf_file_section__process(struct perf_file_section *section,
2649 struct perf_header *ph,
2650 int feat, int fd, void *data)
2652 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2653 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2654 "%d, continuing...\n", section->offset, feat);
2658 if (feat >= HEADER_LAST_FEATURE) {
2659 pr_debug("unknown feature %d, continuing...\n", feat);
2663 if (!feat_ops[feat].process)
2666 return feat_ops[feat].process(section, ph, fd, data);
2669 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2670 struct perf_header *ph, int fd,
2675 ret = readn(fd, header, sizeof(*header));
2679 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2680 pr_debug("endian/magic failed\n");
2685 header->size = bswap_64(header->size);
2687 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2693 static int perf_header__read_pipe(struct perf_session *session)
2695 struct perf_header *header = &session->header;
2696 struct perf_pipe_file_header f_header;
2698 if (perf_file_header__read_pipe(&f_header, header,
2699 perf_data_file__fd(session->file),
2700 session->repipe) < 0) {
2701 pr_debug("incompatible file format\n");
2708 static int read_attr(int fd, struct perf_header *ph,
2709 struct perf_file_attr *f_attr)
2711 struct perf_event_attr *attr = &f_attr->attr;
2713 size_t our_sz = sizeof(f_attr->attr);
2716 memset(f_attr, 0, sizeof(*f_attr));
2718 /* read minimal guaranteed structure */
2719 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2721 pr_debug("cannot read %d bytes of header attr\n",
2722 PERF_ATTR_SIZE_VER0);
2726 /* on file perf_event_attr size */
2734 sz = PERF_ATTR_SIZE_VER0;
2735 } else if (sz > our_sz) {
2736 pr_debug("file uses a more recent and unsupported ABI"
2737 " (%zu bytes extra)\n", sz - our_sz);
2740 /* what we have not yet read and that we know about */
2741 left = sz - PERF_ATTR_SIZE_VER0;
2744 ptr += PERF_ATTR_SIZE_VER0;
2746 ret = readn(fd, ptr, left);
2748 /* read perf_file_section, ids are read in caller */
2749 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2751 return ret <= 0 ? -1 : 0;
2754 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2755 struct pevent *pevent)
2757 struct event_format *event;
2760 /* already prepared */
2761 if (evsel->tp_format)
2764 if (pevent == NULL) {
2765 pr_debug("broken or missing trace data\n");
2769 event = pevent_find_event(pevent, evsel->attr.config);
2774 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2775 evsel->name = strdup(bf);
2776 if (evsel->name == NULL)
2780 evsel->tp_format = event;
2784 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2785 struct pevent *pevent)
2787 struct perf_evsel *pos;
2789 evlist__for_each(evlist, pos) {
2790 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2791 perf_evsel__prepare_tracepoint_event(pos, pevent))
2798 int perf_session__read_header(struct perf_session *session)
2800 struct perf_data_file *file = session->file;
2801 struct perf_header *header = &session->header;
2802 struct perf_file_header f_header;
2803 struct perf_file_attr f_attr;
2805 int nr_attrs, nr_ids, i, j;
2806 int fd = perf_data_file__fd(file);
2808 session->evlist = perf_evlist__new();
2809 if (session->evlist == NULL)
2812 if (perf_data_file__is_pipe(file))
2813 return perf_header__read_pipe(session);
2815 if (perf_file_header__read(&f_header, header, fd) < 0)
2819 * Sanity check that perf.data was written cleanly; data size is
2820 * initialized to 0 and updated only if the on_exit function is run.
2821 * If data size is still 0 then the file contains only partial
2822 * information. Just warn user and process it as much as it can.
2824 if (f_header.data.size == 0) {
2825 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2826 "Was the 'perf record' command properly terminated?\n",
2830 nr_attrs = f_header.attrs.size / f_header.attr_size;
2831 lseek(fd, f_header.attrs.offset, SEEK_SET);
2833 for (i = 0; i < nr_attrs; i++) {
2834 struct perf_evsel *evsel;
2837 if (read_attr(fd, header, &f_attr) < 0)
2840 if (header->needs_swap)
2841 perf_event__attr_swap(&f_attr.attr);
2843 tmp = lseek(fd, 0, SEEK_CUR);
2844 evsel = perf_evsel__new(&f_attr.attr);
2847 goto out_delete_evlist;
2849 evsel->needs_swap = header->needs_swap;
2851 * Do it before so that if perf_evsel__alloc_id fails, this
2852 * entry gets purged too at perf_evlist__delete().
2854 perf_evlist__add(session->evlist, evsel);
2856 nr_ids = f_attr.ids.size / sizeof(u64);
2858 * We don't have the cpu and thread maps on the header, so
2859 * for allocating the perf_sample_id table we fake 1 cpu and
2860 * hattr->ids threads.
2862 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2863 goto out_delete_evlist;
2865 lseek(fd, f_attr.ids.offset, SEEK_SET);
2867 for (j = 0; j < nr_ids; j++) {
2868 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2871 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2874 lseek(fd, tmp, SEEK_SET);
2877 symbol_conf.nr_events = nr_attrs;
2879 perf_header__process_sections(header, fd, &session->tevent,
2880 perf_file_section__process);
2882 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2883 session->tevent.pevent))
2884 goto out_delete_evlist;
2891 perf_evlist__delete(session->evlist);
2892 session->evlist = NULL;
2896 int perf_event__synthesize_attr(struct perf_tool *tool,
2897 struct perf_event_attr *attr, u32 ids, u64 *id,
2898 perf_event__handler_t process)
2900 union perf_event *ev;
2904 size = sizeof(struct perf_event_attr);
2905 size = PERF_ALIGN(size, sizeof(u64));
2906 size += sizeof(struct perf_event_header);
2907 size += ids * sizeof(u64);
2914 ev->attr.attr = *attr;
2915 memcpy(ev->attr.id, id, ids * sizeof(u64));
2917 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2918 ev->attr.header.size = (u16)size;
2920 if (ev->attr.header.size == size)
2921 err = process(tool, ev, NULL, NULL);
2930 int perf_event__synthesize_attrs(struct perf_tool *tool,
2931 struct perf_session *session,
2932 perf_event__handler_t process)
2934 struct perf_evsel *evsel;
2937 evlist__for_each(session->evlist, evsel) {
2938 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2939 evsel->id, process);
2941 pr_debug("failed to create perf header attribute\n");
2949 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2950 union perf_event *event,
2951 struct perf_evlist **pevlist)
2954 struct perf_evsel *evsel;
2955 struct perf_evlist *evlist = *pevlist;
2957 if (evlist == NULL) {
2958 *pevlist = evlist = perf_evlist__new();
2963 evsel = perf_evsel__new(&event->attr.attr);
2967 perf_evlist__add(evlist, evsel);
2969 ids = event->header.size;
2970 ids -= (void *)&event->attr.id - (void *)event;
2971 n_ids = ids / sizeof(u64);
2973 * We don't have the cpu and thread maps on the header, so
2974 * for allocating the perf_sample_id table we fake 1 cpu and
2975 * hattr->ids threads.
2977 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2980 for (i = 0; i < n_ids; i++) {
2981 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2984 symbol_conf.nr_events = evlist->nr_entries;
2989 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2990 struct perf_evlist *evlist,
2991 perf_event__handler_t process)
2993 union perf_event ev;
2994 struct tracing_data *tdata;
2995 ssize_t size = 0, aligned_size = 0, padding;
2996 int err __maybe_unused = 0;
2999 * We are going to store the size of the data followed
3000 * by the data contents. Since the fd descriptor is a pipe,
3001 * we cannot seek back to store the size of the data once
3002 * we know it. Instead we:
3004 * - write the tracing data to the temp file
3005 * - get/write the data size to pipe
3006 * - write the tracing data from the temp file
3009 tdata = tracing_data_get(&evlist->entries, fd, true);
3013 memset(&ev, 0, sizeof(ev));
3015 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
3017 aligned_size = PERF_ALIGN(size, sizeof(u64));
3018 padding = aligned_size - size;
3019 ev.tracing_data.header.size = sizeof(ev.tracing_data);
3020 ev.tracing_data.size = aligned_size;
3022 process(tool, &ev, NULL, NULL);
3025 * The put function will copy all the tracing data
3026 * stored in temp file to the pipe.
3028 tracing_data_put(tdata);
3030 write_padded(fd, NULL, 0, padding);
3032 return aligned_size;
3035 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
3036 union perf_event *event,
3037 struct perf_session *session)
3039 ssize_t size_read, padding, size = event->tracing_data.size;
3040 int fd = perf_data_file__fd(session->file);
3041 off_t offset = lseek(fd, 0, SEEK_CUR);
3044 /* setup for reading amidst mmap */
3045 lseek(fd, offset + sizeof(struct tracing_data_event),
3048 size_read = trace_report(fd, &session->tevent,
3050 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3052 if (readn(fd, buf, padding) < 0) {
3053 pr_err("%s: reading input file", __func__);
3056 if (session->repipe) {
3057 int retw = write(STDOUT_FILENO, buf, padding);
3058 if (retw <= 0 || retw != padding) {
3059 pr_err("%s: repiping tracing data padding", __func__);
3064 if (size_read + padding != size) {
3065 pr_err("%s: tracing data size mismatch", __func__);
3069 perf_evlist__prepare_tracepoint_events(session->evlist,
3070 session->tevent.pevent);
3072 return size_read + padding;
3075 int perf_event__synthesize_build_id(struct perf_tool *tool,
3076 struct dso *pos, u16 misc,
3077 perf_event__handler_t process,
3078 struct machine *machine)
3080 union perf_event ev;
3087 memset(&ev, 0, sizeof(ev));
3089 len = pos->long_name_len + 1;
3090 len = PERF_ALIGN(len, NAME_ALIGN);
3091 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3092 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3093 ev.build_id.header.misc = misc;
3094 ev.build_id.pid = machine->pid;
3095 ev.build_id.header.size = sizeof(ev.build_id) + len;
3096 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3098 err = process(tool, &ev, NULL, machine);
3103 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3104 union perf_event *event,
3105 struct perf_session *session)
3107 __event_process_build_id(&event->build_id,
3108 event->build_id.filename,
3113 void disable_buildid_cache(void)
3115 no_buildid_cache = true;