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"
26 static bool no_buildid_cache = false;
28 static int trace_event_count;
29 static struct perf_trace_event_type *trace_events;
31 static u32 header_argc;
32 static const char **header_argv;
34 int perf_header__push_event(u64 id, const char *name)
36 struct perf_trace_event_type *nevents;
38 if (strlen(name) > MAX_EVENT_NAME)
39 pr_warning("Event %s will be truncated\n", name);
41 nevents = realloc(trace_events, (trace_event_count + 1) * sizeof(*trace_events));
44 trace_events = nevents;
46 memset(&trace_events[trace_event_count], 0, sizeof(struct perf_trace_event_type));
47 trace_events[trace_event_count].event_id = id;
48 strncpy(trace_events[trace_event_count].name, name, MAX_EVENT_NAME - 1);
53 char *perf_header__find_event(u64 id)
56 for (i = 0 ; i < trace_event_count; i++) {
57 if (trace_events[i].event_id == id)
58 return trace_events[i].name;
65 * must be a numerical value to let the endianness
66 * determine the memory layout. That way we are able
67 * to detect endianness when reading the perf.data file
70 * we check for legacy (PERFFILE) format.
72 static const char *__perf_magic1 = "PERFFILE";
73 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
74 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
76 #define PERF_MAGIC __perf_magic2
78 struct perf_file_attr {
79 struct perf_event_attr attr;
80 struct perf_file_section ids;
83 void perf_header__set_feat(struct perf_header *header, int feat)
85 set_bit(feat, header->adds_features);
88 void perf_header__clear_feat(struct perf_header *header, int feat)
90 clear_bit(feat, header->adds_features);
93 bool perf_header__has_feat(const struct perf_header *header, int feat)
95 return test_bit(feat, header->adds_features);
98 static int do_write(int fd, const void *buf, size_t size)
101 int ret = write(fd, buf, size);
113 #define NAME_ALIGN 64
115 static int write_padded(int fd, const void *bf, size_t count,
116 size_t count_aligned)
118 static const char zero_buf[NAME_ALIGN];
119 int err = do_write(fd, bf, count);
122 err = do_write(fd, zero_buf, count_aligned - count);
127 static int do_write_string(int fd, const char *str)
132 olen = strlen(str) + 1;
133 len = PERF_ALIGN(olen, NAME_ALIGN);
135 /* write len, incl. \0 */
136 ret = do_write(fd, &len, sizeof(len));
140 return write_padded(fd, str, olen, len);
143 static char *do_read_string(int fd, struct perf_header *ph)
149 sz = readn(fd, &len, sizeof(len));
150 if (sz < (ssize_t)sizeof(len))
160 ret = readn(fd, buf, len);
161 if (ret == (ssize_t)len) {
163 * strings are padded by zeroes
164 * thus the actual strlen of buf
165 * may be less than len
175 perf_header__set_cmdline(int argc, const char **argv)
180 * If header_argv has already been set, do not override it.
181 * This allows a command to set the cmdline, parse args and
182 * then call another builtin function that implements a
183 * command -- e.g, cmd_kvm calling cmd_record.
188 header_argc = (u32)argc;
190 /* do not include NULL termination */
191 header_argv = calloc(argc, sizeof(char *));
196 * must copy argv contents because it gets moved
197 * around during option parsing
199 for (i = 0; i < argc ; i++)
200 header_argv[i] = argv[i];
205 #define dsos__for_each_with_build_id(pos, head) \
206 list_for_each_entry(pos, head, node) \
207 if (!pos->has_build_id) \
211 static int write_buildid(char *name, size_t name_len, u8 *build_id,
212 pid_t pid, u16 misc, int fd)
215 struct build_id_event b;
219 len = PERF_ALIGN(len, NAME_ALIGN);
221 memset(&b, 0, sizeof(b));
222 memcpy(&b.build_id, build_id, BUILD_ID_SIZE);
224 b.header.misc = misc;
225 b.header.size = sizeof(b) + len;
227 err = do_write(fd, &b, sizeof(b));
231 return write_padded(fd, name, name_len + 1, len);
234 static int __dsos__write_buildid_table(struct list_head *head, pid_t pid,
239 dsos__for_each_with_build_id(pos, head) {
247 if (is_vdso_map(pos->short_name)) {
248 name = (char *) VDSO__MAP_NAME;
249 name_len = sizeof(VDSO__MAP_NAME) + 1;
251 name = pos->long_name;
252 name_len = pos->long_name_len + 1;
255 err = write_buildid(name, name_len, pos->build_id,
264 static int machine__write_buildid_table(struct machine *machine, int fd)
267 u16 kmisc = PERF_RECORD_MISC_KERNEL,
268 umisc = PERF_RECORD_MISC_USER;
270 if (!machine__is_host(machine)) {
271 kmisc = PERF_RECORD_MISC_GUEST_KERNEL;
272 umisc = PERF_RECORD_MISC_GUEST_USER;
275 err = __dsos__write_buildid_table(&machine->kernel_dsos, machine->pid,
278 err = __dsos__write_buildid_table(&machine->user_dsos,
279 machine->pid, umisc, fd);
283 static int dsos__write_buildid_table(struct perf_header *header, int fd)
285 struct perf_session *session = container_of(header,
286 struct perf_session, header);
288 int err = machine__write_buildid_table(&session->machines.host, fd);
293 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
294 struct machine *pos = rb_entry(nd, struct machine, rb_node);
295 err = machine__write_buildid_table(pos, fd);
302 int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
303 const char *name, bool is_kallsyms, bool is_vdso)
305 const size_t size = PATH_MAX;
306 char *realname, *filename = zalloc(size),
307 *linkname = zalloc(size), *targetname;
309 bool slash = is_kallsyms || is_vdso;
312 if (symbol_conf.kptr_restrict) {
313 pr_debug("Not caching a kptr_restrict'ed /proc/kallsyms\n");
317 realname = (char *) name;
319 realname = realpath(name, NULL);
321 if (realname == NULL || filename == NULL || linkname == NULL)
324 len = scnprintf(filename, size, "%s%s%s",
325 debugdir, slash ? "/" : "",
326 is_vdso ? VDSO__MAP_NAME : realname);
327 if (mkdir_p(filename, 0755))
330 snprintf(filename + len, size - len, "/%s", sbuild_id);
332 if (access(filename, F_OK)) {
334 if (copyfile("/proc/kallsyms", filename))
336 } else if (link(realname, filename) && copyfile(name, filename))
340 len = scnprintf(linkname, size, "%s/.build-id/%.2s",
341 debugdir, sbuild_id);
343 if (access(linkname, X_OK) && mkdir_p(linkname, 0755))
346 snprintf(linkname + len, size - len, "/%s", sbuild_id + 2);
347 targetname = filename + strlen(debugdir) - 5;
348 memcpy(targetname, "../..", 5);
350 if (symlink(targetname, linkname) == 0)
360 static int build_id_cache__add_b(const u8 *build_id, size_t build_id_size,
361 const char *name, const char *debugdir,
362 bool is_kallsyms, bool is_vdso)
364 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
366 build_id__sprintf(build_id, build_id_size, sbuild_id);
368 return build_id_cache__add_s(sbuild_id, debugdir, name,
369 is_kallsyms, is_vdso);
372 int build_id_cache__remove_s(const char *sbuild_id, const char *debugdir)
374 const size_t size = PATH_MAX;
375 char *filename = zalloc(size),
376 *linkname = zalloc(size);
379 if (filename == NULL || linkname == NULL)
382 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
383 debugdir, sbuild_id, sbuild_id + 2);
385 if (access(linkname, F_OK))
388 if (readlink(linkname, filename, size - 1) < 0)
391 if (unlink(linkname))
395 * Since the link is relative, we must make it absolute:
397 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
398 debugdir, sbuild_id, filename);
400 if (unlink(linkname))
410 static int dso__cache_build_id(struct dso *dso, const char *debugdir)
412 bool is_kallsyms = dso->kernel && dso->long_name[0] != '/';
413 bool is_vdso = is_vdso_map(dso->short_name);
415 return build_id_cache__add_b(dso->build_id, sizeof(dso->build_id),
416 dso->long_name, debugdir,
417 is_kallsyms, is_vdso);
420 static int __dsos__cache_build_ids(struct list_head *head, const char *debugdir)
425 dsos__for_each_with_build_id(pos, head)
426 if (dso__cache_build_id(pos, debugdir))
432 static int machine__cache_build_ids(struct machine *machine, const char *debugdir)
434 int ret = __dsos__cache_build_ids(&machine->kernel_dsos, debugdir);
435 ret |= __dsos__cache_build_ids(&machine->user_dsos, debugdir);
439 static int perf_session__cache_build_ids(struct perf_session *session)
443 char debugdir[PATH_MAX];
445 snprintf(debugdir, sizeof(debugdir), "%s", buildid_dir);
447 if (mkdir(debugdir, 0755) != 0 && errno != EEXIST)
450 ret = machine__cache_build_ids(&session->machines.host, debugdir);
452 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
453 struct machine *pos = rb_entry(nd, struct machine, rb_node);
454 ret |= machine__cache_build_ids(pos, debugdir);
459 static bool machine__read_build_ids(struct machine *machine, bool with_hits)
461 bool ret = __dsos__read_build_ids(&machine->kernel_dsos, with_hits);
462 ret |= __dsos__read_build_ids(&machine->user_dsos, with_hits);
466 static bool perf_session__read_build_ids(struct perf_session *session, bool with_hits)
469 bool ret = machine__read_build_ids(&session->machines.host, with_hits);
471 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
472 struct machine *pos = rb_entry(nd, struct machine, rb_node);
473 ret |= machine__read_build_ids(pos, with_hits);
479 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
480 struct perf_evlist *evlist)
482 return read_tracing_data(fd, &evlist->entries);
486 static int write_build_id(int fd, struct perf_header *h,
487 struct perf_evlist *evlist __maybe_unused)
489 struct perf_session *session;
492 session = container_of(h, struct perf_session, header);
494 if (!perf_session__read_build_ids(session, true))
497 err = dsos__write_buildid_table(h, fd);
499 pr_debug("failed to write buildid table\n");
502 if (!no_buildid_cache)
503 perf_session__cache_build_ids(session);
508 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
509 struct perf_evlist *evlist __maybe_unused)
518 return do_write_string(fd, uts.nodename);
521 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
522 struct perf_evlist *evlist __maybe_unused)
531 return do_write_string(fd, uts.release);
534 static int write_arch(int fd, struct perf_header *h __maybe_unused,
535 struct perf_evlist *evlist __maybe_unused)
544 return do_write_string(fd, uts.machine);
547 static int write_version(int fd, struct perf_header *h __maybe_unused,
548 struct perf_evlist *evlist __maybe_unused)
550 return do_write_string(fd, perf_version_string);
553 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
554 struct perf_evlist *evlist __maybe_unused)
557 #define CPUINFO_PROC NULL
562 const char *search = CPUINFO_PROC;
569 file = fopen("/proc/cpuinfo", "r");
573 while (getline(&buf, &len, file) > 0) {
574 ret = strncmp(buf, search, strlen(search));
584 p = strchr(buf, ':');
585 if (p && *(p+1) == ' ' && *(p+2))
591 /* squash extra space characters (branding string) */
598 while (*q && isspace(*q))
601 while ((*r++ = *q++));
605 ret = do_write_string(fd, s);
612 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
613 struct perf_evlist *evlist __maybe_unused)
619 nr = sysconf(_SC_NPROCESSORS_CONF);
623 nrc = (u32)(nr & UINT_MAX);
625 nr = sysconf(_SC_NPROCESSORS_ONLN);
629 nra = (u32)(nr & UINT_MAX);
631 ret = do_write(fd, &nrc, sizeof(nrc));
635 return do_write(fd, &nra, sizeof(nra));
638 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
639 struct perf_evlist *evlist)
641 struct perf_evsel *evsel;
645 nre = evlist->nr_entries;
648 * write number of events
650 ret = do_write(fd, &nre, sizeof(nre));
655 * size of perf_event_attr struct
657 sz = (u32)sizeof(evsel->attr);
658 ret = do_write(fd, &sz, sizeof(sz));
662 list_for_each_entry(evsel, &evlist->entries, node) {
664 ret = do_write(fd, &evsel->attr, sz);
668 * write number of unique id per event
669 * there is one id per instance of an event
671 * copy into an nri to be independent of the
675 ret = do_write(fd, &nri, sizeof(nri));
680 * write event string as passed on cmdline
682 ret = do_write_string(fd, perf_evsel__name(evsel));
686 * write unique ids for this event
688 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
695 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
696 struct perf_evlist *evlist __maybe_unused)
698 char buf[MAXPATHLEN];
704 * actual atual path to perf binary
706 sprintf(proc, "/proc/%d/exe", getpid());
707 ret = readlink(proc, buf, sizeof(buf));
711 /* readlink() does not add null termination */
714 /* account for binary path */
717 ret = do_write(fd, &n, sizeof(n));
721 ret = do_write_string(fd, buf);
725 for (i = 0 ; i < header_argc; i++) {
726 ret = do_write_string(fd, header_argv[i]);
733 #define CORE_SIB_FMT \
734 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
735 #define THRD_SIB_FMT \
736 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
741 char **core_siblings;
742 char **thread_siblings;
745 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
748 char filename[MAXPATHLEN];
749 char *buf = NULL, *p;
754 sprintf(filename, CORE_SIB_FMT, cpu);
755 fp = fopen(filename, "r");
759 if (getline(&buf, &len, fp) <= 0)
764 p = strchr(buf, '\n');
768 for (i = 0; i < tp->core_sib; i++) {
769 if (!strcmp(buf, tp->core_siblings[i]))
772 if (i == tp->core_sib) {
773 tp->core_siblings[i] = buf;
779 sprintf(filename, THRD_SIB_FMT, cpu);
780 fp = fopen(filename, "r");
784 if (getline(&buf, &len, fp) <= 0)
787 p = strchr(buf, '\n');
791 for (i = 0; i < tp->thread_sib; i++) {
792 if (!strcmp(buf, tp->thread_siblings[i]))
795 if (i == tp->thread_sib) {
796 tp->thread_siblings[i] = buf;
808 static void free_cpu_topo(struct cpu_topo *tp)
815 for (i = 0 ; i < tp->core_sib; i++)
816 free(tp->core_siblings[i]);
818 for (i = 0 ; i < tp->thread_sib; i++)
819 free(tp->thread_siblings[i]);
824 static struct cpu_topo *build_cpu_topology(void)
833 ncpus = sysconf(_SC_NPROCESSORS_CONF);
837 nr = (u32)(ncpus & UINT_MAX);
839 sz = nr * sizeof(char *);
841 addr = calloc(1, sizeof(*tp) + 2 * sz);
848 tp->core_siblings = addr;
850 tp->thread_siblings = addr;
852 for (i = 0; i < nr; i++) {
853 ret = build_cpu_topo(tp, i);
864 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
865 struct perf_evlist *evlist __maybe_unused)
871 tp = build_cpu_topology();
875 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
879 for (i = 0; i < tp->core_sib; i++) {
880 ret = do_write_string(fd, tp->core_siblings[i]);
884 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
888 for (i = 0; i < tp->thread_sib; i++) {
889 ret = do_write_string(fd, tp->thread_siblings[i]);
900 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
901 struct perf_evlist *evlist __maybe_unused)
909 fp = fopen("/proc/meminfo", "r");
913 while (getline(&buf, &len, fp) > 0) {
914 ret = strncmp(buf, "MemTotal:", 9);
919 n = sscanf(buf, "%*s %"PRIu64, &mem);
921 ret = do_write(fd, &mem, sizeof(mem));
928 static int write_topo_node(int fd, int node)
930 char str[MAXPATHLEN];
932 char *buf = NULL, *p;
935 u64 mem_total, mem_free, mem;
938 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
939 fp = fopen(str, "r");
943 while (getline(&buf, &len, fp) > 0) {
944 /* skip over invalid lines */
945 if (!strchr(buf, ':'))
947 if (sscanf(buf, "%*s %*d %s %"PRIu64, field, &mem) != 2)
949 if (!strcmp(field, "MemTotal:"))
951 if (!strcmp(field, "MemFree:"))
958 ret = do_write(fd, &mem_total, sizeof(u64));
962 ret = do_write(fd, &mem_free, sizeof(u64));
967 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
969 fp = fopen(str, "r");
973 if (getline(&buf, &len, fp) <= 0)
976 p = strchr(buf, '\n');
980 ret = do_write_string(fd, buf);
988 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
989 struct perf_evlist *evlist __maybe_unused)
994 struct cpu_map *node_map = NULL;
999 fp = fopen("/sys/devices/system/node/online", "r");
1003 if (getline(&buf, &len, fp) <= 0)
1006 c = strchr(buf, '\n');
1010 node_map = cpu_map__new(buf);
1014 nr = (u32)node_map->nr;
1016 ret = do_write(fd, &nr, sizeof(nr));
1020 for (i = 0; i < nr; i++) {
1021 j = (u32)node_map->map[i];
1022 ret = do_write(fd, &j, sizeof(j));
1026 ret = write_topo_node(fd, i);
1040 * struct pmu_mappings {
1049 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
1050 struct perf_evlist *evlist __maybe_unused)
1052 struct perf_pmu *pmu = NULL;
1053 off_t offset = lseek(fd, 0, SEEK_CUR);
1057 /* write real pmu_num later */
1058 ret = do_write(fd, &pmu_num, sizeof(pmu_num));
1062 while ((pmu = perf_pmu__scan(pmu))) {
1067 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
1071 ret = do_write_string(fd, pmu->name);
1076 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
1078 lseek(fd, offset, SEEK_SET);
1088 * struct group_descs {
1090 * struct group_desc {
1097 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
1098 struct perf_evlist *evlist)
1100 u32 nr_groups = evlist->nr_groups;
1101 struct perf_evsel *evsel;
1104 ret = do_write(fd, &nr_groups, sizeof(nr_groups));
1108 list_for_each_entry(evsel, &evlist->entries, node) {
1109 if (perf_evsel__is_group_leader(evsel) &&
1110 evsel->nr_members > 1) {
1111 const char *name = evsel->group_name ?: "{anon_group}";
1112 u32 leader_idx = evsel->idx;
1113 u32 nr_members = evsel->nr_members;
1115 ret = do_write_string(fd, name);
1119 ret = do_write(fd, &leader_idx, sizeof(leader_idx));
1123 ret = do_write(fd, &nr_members, sizeof(nr_members));
1132 * default get_cpuid(): nothing gets recorded
1133 * actual implementation must be in arch/$(ARCH)/util/header.c
1135 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
1136 size_t sz __maybe_unused)
1141 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
1142 struct perf_evlist *evlist __maybe_unused)
1147 ret = get_cpuid(buffer, sizeof(buffer));
1153 return do_write_string(fd, buffer);
1156 static int write_branch_stack(int fd __maybe_unused,
1157 struct perf_header *h __maybe_unused,
1158 struct perf_evlist *evlist __maybe_unused)
1163 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
1166 fprintf(fp, "# hostname : %s\n", ph->env.hostname);
1169 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
1172 fprintf(fp, "# os release : %s\n", ph->env.os_release);
1175 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1177 fprintf(fp, "# arch : %s\n", ph->env.arch);
1180 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
1183 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
1186 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
1189 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
1190 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
1193 static void print_version(struct perf_header *ph, int fd __maybe_unused,
1196 fprintf(fp, "# perf version : %s\n", ph->env.version);
1199 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
1205 nr = ph->env.nr_cmdline;
1206 str = ph->env.cmdline;
1208 fprintf(fp, "# cmdline : ");
1210 for (i = 0; i < nr; i++) {
1211 fprintf(fp, "%s ", str);
1212 str += strlen(str) + 1;
1217 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
1223 nr = ph->env.nr_sibling_cores;
1224 str = ph->env.sibling_cores;
1226 for (i = 0; i < nr; i++) {
1227 fprintf(fp, "# sibling cores : %s\n", str);
1228 str += strlen(str) + 1;
1231 nr = ph->env.nr_sibling_threads;
1232 str = ph->env.sibling_threads;
1234 for (i = 0; i < nr; i++) {
1235 fprintf(fp, "# sibling threads : %s\n", str);
1236 str += strlen(str) + 1;
1240 static void free_event_desc(struct perf_evsel *events)
1242 struct perf_evsel *evsel;
1247 for (evsel = events; evsel->attr.size; evsel++) {
1257 static struct perf_evsel *
1258 read_event_desc(struct perf_header *ph, int fd)
1260 struct perf_evsel *evsel, *events = NULL;
1263 u32 nre, sz, nr, i, j;
1267 /* number of events */
1268 ret = readn(fd, &nre, sizeof(nre));
1269 if (ret != (ssize_t)sizeof(nre))
1273 nre = bswap_32(nre);
1275 ret = readn(fd, &sz, sizeof(sz));
1276 if (ret != (ssize_t)sizeof(sz))
1282 /* buffer to hold on file attr struct */
1287 /* the last event terminates with evsel->attr.size == 0: */
1288 events = calloc(nre + 1, sizeof(*events));
1292 msz = sizeof(evsel->attr);
1296 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1300 * must read entire on-file attr struct to
1301 * sync up with layout.
1303 ret = readn(fd, buf, sz);
1304 if (ret != (ssize_t)sz)
1308 perf_event__attr_swap(buf);
1310 memcpy(&evsel->attr, buf, msz);
1312 ret = readn(fd, &nr, sizeof(nr));
1313 if (ret != (ssize_t)sizeof(nr))
1316 if (ph->needs_swap) {
1318 evsel->needs_swap = true;
1321 evsel->name = do_read_string(fd, ph);
1326 id = calloc(nr, sizeof(*id));
1332 for (j = 0 ; j < nr; j++) {
1333 ret = readn(fd, id, sizeof(*id));
1334 if (ret != (ssize_t)sizeof(*id))
1337 *id = bswap_64(*id);
1347 free_event_desc(events);
1352 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1354 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1359 fprintf(fp, "# event desc: not available or unable to read\n");
1363 for (evsel = events; evsel->attr.size; evsel++) {
1364 fprintf(fp, "# event : name = %s, ", evsel->name);
1366 fprintf(fp, "type = %d, config = 0x%"PRIx64
1367 ", config1 = 0x%"PRIx64", config2 = 0x%"PRIx64,
1369 (u64)evsel->attr.config,
1370 (u64)evsel->attr.config1,
1371 (u64)evsel->attr.config2);
1373 fprintf(fp, ", excl_usr = %d, excl_kern = %d",
1374 evsel->attr.exclude_user,
1375 evsel->attr.exclude_kernel);
1377 fprintf(fp, ", excl_host = %d, excl_guest = %d",
1378 evsel->attr.exclude_host,
1379 evsel->attr.exclude_guest);
1381 fprintf(fp, ", precise_ip = %d", evsel->attr.precise_ip);
1384 fprintf(fp, ", id = {");
1385 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1388 fprintf(fp, " %"PRIu64, *id);
1396 free_event_desc(events);
1399 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1402 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1405 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1410 uint64_t mem_total, mem_free;
1413 nr = ph->env.nr_numa_nodes;
1414 str = ph->env.numa_nodes;
1416 for (i = 0; i < nr; i++) {
1418 c = strtoul(str, &tmp, 0);
1423 mem_total = strtoull(str, &tmp, 0);
1428 mem_free = strtoull(str, &tmp, 0);
1432 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1433 " free = %"PRIu64" kB\n",
1434 c, mem_total, mem_free);
1437 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1439 str += strlen(str) + 1;
1443 fprintf(fp, "# numa topology : not available\n");
1446 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1448 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1451 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1452 int fd __maybe_unused, FILE *fp)
1454 fprintf(fp, "# contains samples with branch stack\n");
1457 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1460 const char *delimiter = "# pmu mappings: ";
1465 pmu_num = ph->env.nr_pmu_mappings;
1467 fprintf(fp, "# pmu mappings: not available\n");
1471 str = ph->env.pmu_mappings;
1474 type = strtoul(str, &tmp, 0);
1479 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1482 str += strlen(str) + 1;
1491 fprintf(fp, "# pmu mappings: unable to read\n");
1494 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1497 struct perf_session *session;
1498 struct perf_evsel *evsel;
1501 session = container_of(ph, struct perf_session, header);
1503 list_for_each_entry(evsel, &session->evlist->entries, node) {
1504 if (perf_evsel__is_group_leader(evsel) &&
1505 evsel->nr_members > 1) {
1506 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1507 perf_evsel__name(evsel));
1509 nr = evsel->nr_members - 1;
1511 fprintf(fp, ",%s", perf_evsel__name(evsel));
1519 static int __event_process_build_id(struct build_id_event *bev,
1521 struct perf_session *session)
1524 struct list_head *head;
1525 struct machine *machine;
1528 enum dso_kernel_type dso_type;
1530 machine = perf_session__findnew_machine(session, bev->pid);
1534 misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1537 case PERF_RECORD_MISC_KERNEL:
1538 dso_type = DSO_TYPE_KERNEL;
1539 head = &machine->kernel_dsos;
1541 case PERF_RECORD_MISC_GUEST_KERNEL:
1542 dso_type = DSO_TYPE_GUEST_KERNEL;
1543 head = &machine->kernel_dsos;
1545 case PERF_RECORD_MISC_USER:
1546 case PERF_RECORD_MISC_GUEST_USER:
1547 dso_type = DSO_TYPE_USER;
1548 head = &machine->user_dsos;
1554 dso = __dsos__findnew(head, filename);
1556 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1558 dso__set_build_id(dso, &bev->build_id);
1560 if (filename[0] == '[')
1561 dso->kernel = dso_type;
1563 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1565 pr_debug("build id event received for %s: %s\n",
1566 dso->long_name, sbuild_id);
1574 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1575 int input, u64 offset, u64 size)
1577 struct perf_session *session = container_of(header, struct perf_session, header);
1579 struct perf_event_header header;
1580 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1583 struct build_id_event bev;
1584 char filename[PATH_MAX];
1585 u64 limit = offset + size;
1587 while (offset < limit) {
1590 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1593 if (header->needs_swap)
1594 perf_event_header__bswap(&old_bev.header);
1596 len = old_bev.header.size - sizeof(old_bev);
1597 if (readn(input, filename, len) != len)
1600 bev.header = old_bev.header;
1603 * As the pid is the missing value, we need to fill
1604 * it properly. The header.misc value give us nice hint.
1606 bev.pid = HOST_KERNEL_ID;
1607 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1608 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1609 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1611 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1612 __event_process_build_id(&bev, filename, session);
1614 offset += bev.header.size;
1620 static int perf_header__read_build_ids(struct perf_header *header,
1621 int input, u64 offset, u64 size)
1623 struct perf_session *session = container_of(header, struct perf_session, header);
1624 struct build_id_event bev;
1625 char filename[PATH_MAX];
1626 u64 limit = offset + size, orig_offset = offset;
1629 while (offset < limit) {
1632 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1635 if (header->needs_swap)
1636 perf_event_header__bswap(&bev.header);
1638 len = bev.header.size - sizeof(bev);
1639 if (readn(input, filename, len) != len)
1642 * The a1645ce1 changeset:
1644 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1646 * Added a field to struct build_id_event that broke the file
1649 * Since the kernel build-id is the first entry, process the
1650 * table using the old format if the well known
1651 * '[kernel.kallsyms]' string for the kernel build-id has the
1652 * first 4 characters chopped off (where the pid_t sits).
1654 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1655 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1657 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1660 __event_process_build_id(&bev, filename, session);
1662 offset += bev.header.size;
1669 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1670 struct perf_header *ph __maybe_unused,
1673 ssize_t ret = trace_report(fd, data, false);
1674 return ret < 0 ? -1 : 0;
1677 static int process_build_id(struct perf_file_section *section,
1678 struct perf_header *ph, int fd,
1679 void *data __maybe_unused)
1681 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1682 pr_debug("Failed to read buildids, continuing...\n");
1686 static int process_hostname(struct perf_file_section *section __maybe_unused,
1687 struct perf_header *ph, int fd,
1688 void *data __maybe_unused)
1690 ph->env.hostname = do_read_string(fd, ph);
1691 return ph->env.hostname ? 0 : -ENOMEM;
1694 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1695 struct perf_header *ph, int fd,
1696 void *data __maybe_unused)
1698 ph->env.os_release = do_read_string(fd, ph);
1699 return ph->env.os_release ? 0 : -ENOMEM;
1702 static int process_version(struct perf_file_section *section __maybe_unused,
1703 struct perf_header *ph, int fd,
1704 void *data __maybe_unused)
1706 ph->env.version = do_read_string(fd, ph);
1707 return ph->env.version ? 0 : -ENOMEM;
1710 static int process_arch(struct perf_file_section *section __maybe_unused,
1711 struct perf_header *ph, int fd,
1712 void *data __maybe_unused)
1714 ph->env.arch = do_read_string(fd, ph);
1715 return ph->env.arch ? 0 : -ENOMEM;
1718 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1719 struct perf_header *ph, int fd,
1720 void *data __maybe_unused)
1725 ret = readn(fd, &nr, sizeof(nr));
1726 if (ret != sizeof(nr))
1732 ph->env.nr_cpus_online = nr;
1734 ret = readn(fd, &nr, sizeof(nr));
1735 if (ret != sizeof(nr))
1741 ph->env.nr_cpus_avail = nr;
1745 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1746 struct perf_header *ph, int fd,
1747 void *data __maybe_unused)
1749 ph->env.cpu_desc = do_read_string(fd, ph);
1750 return ph->env.cpu_desc ? 0 : -ENOMEM;
1753 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1754 struct perf_header *ph, int fd,
1755 void *data __maybe_unused)
1757 ph->env.cpuid = do_read_string(fd, ph);
1758 return ph->env.cpuid ? 0 : -ENOMEM;
1761 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1762 struct perf_header *ph, int fd,
1763 void *data __maybe_unused)
1768 ret = readn(fd, &mem, sizeof(mem));
1769 if (ret != sizeof(mem))
1773 mem = bswap_64(mem);
1775 ph->env.total_mem = mem;
1779 static struct perf_evsel *
1780 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1782 struct perf_evsel *evsel;
1784 list_for_each_entry(evsel, &evlist->entries, node) {
1785 if (evsel->idx == idx)
1793 perf_evlist__set_event_name(struct perf_evlist *evlist,
1794 struct perf_evsel *event)
1796 struct perf_evsel *evsel;
1801 evsel = perf_evlist__find_by_index(evlist, event->idx);
1808 evsel->name = strdup(event->name);
1812 process_event_desc(struct perf_file_section *section __maybe_unused,
1813 struct perf_header *header, int fd,
1814 void *data __maybe_unused)
1816 struct perf_session *session;
1817 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1822 session = container_of(header, struct perf_session, header);
1823 for (evsel = events; evsel->attr.size; evsel++)
1824 perf_evlist__set_event_name(session->evlist, evsel);
1826 free_event_desc(events);
1831 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1832 struct perf_header *ph, int fd,
1833 void *data __maybe_unused)
1840 ret = readn(fd, &nr, sizeof(nr));
1841 if (ret != sizeof(nr))
1847 ph->env.nr_cmdline = nr;
1848 strbuf_init(&sb, 128);
1850 for (i = 0; i < nr; i++) {
1851 str = do_read_string(fd, ph);
1855 /* include a NULL character at the end */
1856 strbuf_add(&sb, str, strlen(str) + 1);
1859 ph->env.cmdline = strbuf_detach(&sb, NULL);
1863 strbuf_release(&sb);
1867 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1868 struct perf_header *ph, int fd,
1869 void *data __maybe_unused)
1876 ret = readn(fd, &nr, sizeof(nr));
1877 if (ret != sizeof(nr))
1883 ph->env.nr_sibling_cores = nr;
1884 strbuf_init(&sb, 128);
1886 for (i = 0; i < nr; i++) {
1887 str = do_read_string(fd, ph);
1891 /* include a NULL character at the end */
1892 strbuf_add(&sb, str, strlen(str) + 1);
1895 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1897 ret = readn(fd, &nr, sizeof(nr));
1898 if (ret != sizeof(nr))
1904 ph->env.nr_sibling_threads = nr;
1906 for (i = 0; i < nr; i++) {
1907 str = do_read_string(fd, ph);
1911 /* include a NULL character at the end */
1912 strbuf_add(&sb, str, strlen(str) + 1);
1915 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1919 strbuf_release(&sb);
1923 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1924 struct perf_header *ph, int fd,
1925 void *data __maybe_unused)
1930 uint64_t mem_total, mem_free;
1934 ret = readn(fd, &nr, sizeof(nr));
1935 if (ret != sizeof(nr))
1941 ph->env.nr_numa_nodes = nr;
1942 strbuf_init(&sb, 256);
1944 for (i = 0; i < nr; i++) {
1946 ret = readn(fd, &node, sizeof(node));
1947 if (ret != sizeof(node))
1950 ret = readn(fd, &mem_total, sizeof(u64));
1951 if (ret != sizeof(u64))
1954 ret = readn(fd, &mem_free, sizeof(u64));
1955 if (ret != sizeof(u64))
1958 if (ph->needs_swap) {
1959 node = bswap_32(node);
1960 mem_total = bswap_64(mem_total);
1961 mem_free = bswap_64(mem_free);
1964 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1965 node, mem_total, mem_free);
1967 str = do_read_string(fd, ph);
1971 /* include a NULL character at the end */
1972 strbuf_add(&sb, str, strlen(str) + 1);
1975 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1979 strbuf_release(&sb);
1983 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1984 struct perf_header *ph, int fd,
1985 void *data __maybe_unused)
1993 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1994 if (ret != sizeof(pmu_num))
1998 pmu_num = bswap_32(pmu_num);
2001 pr_debug("pmu mappings not available\n");
2005 ph->env.nr_pmu_mappings = pmu_num;
2006 strbuf_init(&sb, 128);
2009 if (readn(fd, &type, sizeof(type)) != sizeof(type))
2012 type = bswap_32(type);
2014 name = do_read_string(fd, ph);
2018 strbuf_addf(&sb, "%u:%s", type, name);
2019 /* include a NULL character at the end */
2020 strbuf_add(&sb, "", 1);
2025 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2029 strbuf_release(&sb);
2033 static int process_group_desc(struct perf_file_section *section __maybe_unused,
2034 struct perf_header *ph, int fd,
2035 void *data __maybe_unused)
2038 u32 i, nr, nr_groups;
2039 struct perf_session *session;
2040 struct perf_evsel *evsel, *leader = NULL;
2047 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
2051 nr_groups = bswap_32(nr_groups);
2053 ph->env.nr_groups = nr_groups;
2055 pr_debug("group desc not available\n");
2059 desc = calloc(nr_groups, sizeof(*desc));
2063 for (i = 0; i < nr_groups; i++) {
2064 desc[i].name = do_read_string(fd, ph);
2068 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
2071 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
2074 if (ph->needs_swap) {
2075 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
2076 desc[i].nr_members = bswap_32(desc[i].nr_members);
2081 * Rebuild group relationship based on the group_desc
2083 session = container_of(ph, struct perf_session, header);
2084 session->evlist->nr_groups = nr_groups;
2087 list_for_each_entry(evsel, &session->evlist->entries, node) {
2088 if (evsel->idx == (int) desc[i].leader_idx) {
2089 evsel->leader = evsel;
2090 /* {anon_group} is a dummy name */
2091 if (strcmp(desc[i].name, "{anon_group}"))
2092 evsel->group_name = desc[i].name;
2093 evsel->nr_members = desc[i].nr_members;
2095 if (i >= nr_groups || nr > 0) {
2096 pr_debug("invalid group desc\n");
2101 nr = evsel->nr_members - 1;
2104 /* This is a group member */
2105 evsel->leader = leader;
2111 if (i != nr_groups || nr != 0) {
2112 pr_debug("invalid group desc\n");
2118 while ((int) --i >= 0)
2125 struct feature_ops {
2126 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
2127 void (*print)(struct perf_header *h, int fd, FILE *fp);
2128 int (*process)(struct perf_file_section *section,
2129 struct perf_header *h, int fd, void *data);
2134 #define FEAT_OPA(n, func) \
2135 [n] = { .name = #n, .write = write_##func, .print = print_##func }
2136 #define FEAT_OPP(n, func) \
2137 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2138 .process = process_##func }
2139 #define FEAT_OPF(n, func) \
2140 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2141 .process = process_##func, .full_only = true }
2143 /* feature_ops not implemented: */
2144 #define print_tracing_data NULL
2145 #define print_build_id NULL
2147 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2148 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
2149 FEAT_OPP(HEADER_BUILD_ID, build_id),
2150 FEAT_OPP(HEADER_HOSTNAME, hostname),
2151 FEAT_OPP(HEADER_OSRELEASE, osrelease),
2152 FEAT_OPP(HEADER_VERSION, version),
2153 FEAT_OPP(HEADER_ARCH, arch),
2154 FEAT_OPP(HEADER_NRCPUS, nrcpus),
2155 FEAT_OPP(HEADER_CPUDESC, cpudesc),
2156 FEAT_OPP(HEADER_CPUID, cpuid),
2157 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
2158 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
2159 FEAT_OPP(HEADER_CMDLINE, cmdline),
2160 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
2161 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
2162 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
2163 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
2164 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
2167 struct header_print_data {
2169 bool full; /* extended list of headers */
2172 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2173 struct perf_header *ph,
2174 int feat, int fd, void *data)
2176 struct header_print_data *hd = data;
2178 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2179 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2180 "%d, continuing...\n", section->offset, feat);
2183 if (feat >= HEADER_LAST_FEATURE) {
2184 pr_warning("unknown feature %d\n", feat);
2187 if (!feat_ops[feat].print)
2190 if (!feat_ops[feat].full_only || hd->full)
2191 feat_ops[feat].print(ph, fd, hd->fp);
2193 fprintf(hd->fp, "# %s info available, use -I to display\n",
2194 feat_ops[feat].name);
2199 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2201 struct header_print_data hd;
2202 struct perf_header *header = &session->header;
2203 int fd = session->fd;
2207 perf_header__process_sections(header, fd, &hd,
2208 perf_file_section__fprintf_info);
2212 static int do_write_feat(int fd, struct perf_header *h, int type,
2213 struct perf_file_section **p,
2214 struct perf_evlist *evlist)
2219 if (perf_header__has_feat(h, type)) {
2220 if (!feat_ops[type].write)
2223 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2225 err = feat_ops[type].write(fd, h, evlist);
2227 pr_debug("failed to write feature %d\n", type);
2229 /* undo anything written */
2230 lseek(fd, (*p)->offset, SEEK_SET);
2234 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2240 static int perf_header__adds_write(struct perf_header *header,
2241 struct perf_evlist *evlist, int fd)
2244 struct perf_file_section *feat_sec, *p;
2250 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2254 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2255 if (feat_sec == NULL)
2258 sec_size = sizeof(*feat_sec) * nr_sections;
2260 sec_start = header->data_offset + header->data_size;
2261 lseek(fd, sec_start + sec_size, SEEK_SET);
2263 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2264 if (do_write_feat(fd, header, feat, &p, evlist))
2265 perf_header__clear_feat(header, feat);
2268 lseek(fd, sec_start, SEEK_SET);
2270 * may write more than needed due to dropped feature, but
2271 * this is okay, reader will skip the mising entries
2273 err = do_write(fd, feat_sec, sec_size);
2275 pr_debug("failed to write feature section\n");
2280 int perf_header__write_pipe(int fd)
2282 struct perf_pipe_file_header f_header;
2285 f_header = (struct perf_pipe_file_header){
2286 .magic = PERF_MAGIC,
2287 .size = sizeof(f_header),
2290 err = do_write(fd, &f_header, sizeof(f_header));
2292 pr_debug("failed to write perf pipe header\n");
2299 int perf_session__write_header(struct perf_session *session,
2300 struct perf_evlist *evlist,
2301 int fd, bool at_exit)
2303 struct perf_file_header f_header;
2304 struct perf_file_attr f_attr;
2305 struct perf_header *header = &session->header;
2306 struct perf_evsel *evsel, *pair = NULL;
2309 lseek(fd, sizeof(f_header), SEEK_SET);
2311 if (session->evlist != evlist)
2312 pair = perf_evlist__first(session->evlist);
2314 list_for_each_entry(evsel, &evlist->entries, node) {
2315 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2316 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2319 pr_debug("failed to write perf header\n");
2322 if (session->evlist != evlist) {
2323 err = do_write(fd, pair->id, pair->ids * sizeof(u64));
2326 evsel->ids += pair->ids;
2327 pair = perf_evsel__next(pair);
2331 header->attr_offset = lseek(fd, 0, SEEK_CUR);
2333 list_for_each_entry(evsel, &evlist->entries, node) {
2334 f_attr = (struct perf_file_attr){
2335 .attr = evsel->attr,
2337 .offset = evsel->id_offset,
2338 .size = evsel->ids * sizeof(u64),
2341 err = do_write(fd, &f_attr, sizeof(f_attr));
2343 pr_debug("failed to write perf header attribute\n");
2348 header->event_offset = lseek(fd, 0, SEEK_CUR);
2349 header->event_size = trace_event_count * sizeof(struct perf_trace_event_type);
2351 err = do_write(fd, trace_events, header->event_size);
2353 pr_debug("failed to write perf header events\n");
2358 header->data_offset = lseek(fd, 0, SEEK_CUR);
2361 err = perf_header__adds_write(header, evlist, fd);
2366 f_header = (struct perf_file_header){
2367 .magic = PERF_MAGIC,
2368 .size = sizeof(f_header),
2369 .attr_size = sizeof(f_attr),
2371 .offset = header->attr_offset,
2372 .size = evlist->nr_entries * sizeof(f_attr),
2375 .offset = header->data_offset,
2376 .size = header->data_size,
2379 .offset = header->event_offset,
2380 .size = header->event_size,
2384 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2386 lseek(fd, 0, SEEK_SET);
2387 err = do_write(fd, &f_header, sizeof(f_header));
2389 pr_debug("failed to write perf header\n");
2392 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2398 static int perf_header__getbuffer64(struct perf_header *header,
2399 int fd, void *buf, size_t size)
2401 if (readn(fd, buf, size) <= 0)
2404 if (header->needs_swap)
2405 mem_bswap_64(buf, size);
2410 int perf_header__process_sections(struct perf_header *header, int fd,
2412 int (*process)(struct perf_file_section *section,
2413 struct perf_header *ph,
2414 int feat, int fd, void *data))
2416 struct perf_file_section *feat_sec, *sec;
2422 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2426 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2430 sec_size = sizeof(*feat_sec) * nr_sections;
2432 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2434 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2438 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2439 err = process(sec++, header, feat, fd, data);
2449 static const int attr_file_abi_sizes[] = {
2450 [0] = PERF_ATTR_SIZE_VER0,
2451 [1] = PERF_ATTR_SIZE_VER1,
2452 [2] = PERF_ATTR_SIZE_VER2,
2453 [3] = PERF_ATTR_SIZE_VER3,
2458 * In the legacy file format, the magic number is not used to encode endianness.
2459 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2460 * on ABI revisions, we need to try all combinations for all endianness to
2461 * detect the endianness.
2463 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2465 uint64_t ref_size, attr_size;
2468 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2469 ref_size = attr_file_abi_sizes[i]
2470 + sizeof(struct perf_file_section);
2471 if (hdr_sz != ref_size) {
2472 attr_size = bswap_64(hdr_sz);
2473 if (attr_size != ref_size)
2476 ph->needs_swap = true;
2478 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2483 /* could not determine endianness */
2487 #define PERF_PIPE_HDR_VER0 16
2489 static const size_t attr_pipe_abi_sizes[] = {
2490 [0] = PERF_PIPE_HDR_VER0,
2495 * In the legacy pipe format, there is an implicit assumption that endiannesss
2496 * between host recording the samples, and host parsing the samples is the
2497 * same. This is not always the case given that the pipe output may always be
2498 * redirected into a file and analyzed on a different machine with possibly a
2499 * different endianness and perf_event ABI revsions in the perf tool itself.
2501 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2506 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2507 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2508 attr_size = bswap_64(hdr_sz);
2509 if (attr_size != hdr_sz)
2512 ph->needs_swap = true;
2514 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2520 bool is_perf_magic(u64 magic)
2522 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2523 || magic == __perf_magic2
2524 || magic == __perf_magic2_sw)
2530 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2531 bool is_pipe, struct perf_header *ph)
2535 /* check for legacy format */
2536 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2538 pr_debug("legacy perf.data format\n");
2540 return try_all_pipe_abis(hdr_sz, ph);
2542 return try_all_file_abis(hdr_sz, ph);
2545 * the new magic number serves two purposes:
2546 * - unique number to identify actual perf.data files
2547 * - encode endianness of file
2550 /* check magic number with one endianness */
2551 if (magic == __perf_magic2)
2554 /* check magic number with opposite endianness */
2555 if (magic != __perf_magic2_sw)
2558 ph->needs_swap = true;
2563 int perf_file_header__read(struct perf_file_header *header,
2564 struct perf_header *ph, int fd)
2568 lseek(fd, 0, SEEK_SET);
2570 ret = readn(fd, header, sizeof(*header));
2574 if (check_magic_endian(header->magic,
2575 header->attr_size, false, ph) < 0) {
2576 pr_debug("magic/endian check failed\n");
2580 if (ph->needs_swap) {
2581 mem_bswap_64(header, offsetof(struct perf_file_header,
2585 if (header->size != sizeof(*header)) {
2586 /* Support the previous format */
2587 if (header->size == offsetof(typeof(*header), adds_features))
2588 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2591 } else if (ph->needs_swap) {
2593 * feature bitmap is declared as an array of unsigned longs --
2594 * not good since its size can differ between the host that
2595 * generated the data file and the host analyzing the file.
2597 * We need to handle endianness, but we don't know the size of
2598 * the unsigned long where the file was generated. Take a best
2599 * guess at determining it: try 64-bit swap first (ie., file
2600 * created on a 64-bit host), and check if the hostname feature
2601 * bit is set (this feature bit is forced on as of fbe96f2).
2602 * If the bit is not, undo the 64-bit swap and try a 32-bit
2603 * swap. If the hostname bit is still not set (e.g., older data
2604 * file), punt and fallback to the original behavior --
2605 * clearing all feature bits and setting buildid.
2607 mem_bswap_64(&header->adds_features,
2608 BITS_TO_U64(HEADER_FEAT_BITS));
2610 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2612 mem_bswap_64(&header->adds_features,
2613 BITS_TO_U64(HEADER_FEAT_BITS));
2616 mem_bswap_32(&header->adds_features,
2617 BITS_TO_U32(HEADER_FEAT_BITS));
2620 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2621 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2622 set_bit(HEADER_BUILD_ID, header->adds_features);
2626 memcpy(&ph->adds_features, &header->adds_features,
2627 sizeof(ph->adds_features));
2629 ph->event_offset = header->event_types.offset;
2630 ph->event_size = header->event_types.size;
2631 ph->data_offset = header->data.offset;
2632 ph->data_size = header->data.size;
2636 static int perf_file_section__process(struct perf_file_section *section,
2637 struct perf_header *ph,
2638 int feat, int fd, void *data)
2640 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2641 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2642 "%d, continuing...\n", section->offset, feat);
2646 if (feat >= HEADER_LAST_FEATURE) {
2647 pr_debug("unknown feature %d, continuing...\n", feat);
2651 if (!feat_ops[feat].process)
2654 return feat_ops[feat].process(section, ph, fd, data);
2657 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2658 struct perf_header *ph, int fd,
2663 ret = readn(fd, header, sizeof(*header));
2667 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2668 pr_debug("endian/magic failed\n");
2673 header->size = bswap_64(header->size);
2675 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2681 static int perf_header__read_pipe(struct perf_session *session, int fd)
2683 struct perf_header *header = &session->header;
2684 struct perf_pipe_file_header f_header;
2686 if (perf_file_header__read_pipe(&f_header, header, fd,
2687 session->repipe) < 0) {
2688 pr_debug("incompatible file format\n");
2697 static int read_attr(int fd, struct perf_header *ph,
2698 struct perf_file_attr *f_attr)
2700 struct perf_event_attr *attr = &f_attr->attr;
2702 size_t our_sz = sizeof(f_attr->attr);
2705 memset(f_attr, 0, sizeof(*f_attr));
2707 /* read minimal guaranteed structure */
2708 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2710 pr_debug("cannot read %d bytes of header attr\n",
2711 PERF_ATTR_SIZE_VER0);
2715 /* on file perf_event_attr size */
2723 sz = PERF_ATTR_SIZE_VER0;
2724 } else if (sz > our_sz) {
2725 pr_debug("file uses a more recent and unsupported ABI"
2726 " (%zu bytes extra)\n", sz - our_sz);
2729 /* what we have not yet read and that we know about */
2730 left = sz - PERF_ATTR_SIZE_VER0;
2733 ptr += PERF_ATTR_SIZE_VER0;
2735 ret = readn(fd, ptr, left);
2737 /* read perf_file_section, ids are read in caller */
2738 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2740 return ret <= 0 ? -1 : 0;
2743 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2744 struct pevent *pevent)
2746 struct event_format *event;
2749 /* already prepared */
2750 if (evsel->tp_format)
2753 if (pevent == NULL) {
2754 pr_debug("broken or missing trace data\n");
2758 event = pevent_find_event(pevent, evsel->attr.config);
2763 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2764 evsel->name = strdup(bf);
2765 if (evsel->name == NULL)
2769 evsel->tp_format = event;
2773 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2774 struct pevent *pevent)
2776 struct perf_evsel *pos;
2778 list_for_each_entry(pos, &evlist->entries, node) {
2779 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2780 perf_evsel__prepare_tracepoint_event(pos, pevent))
2787 int perf_session__read_header(struct perf_session *session, int fd)
2789 struct perf_header *header = &session->header;
2790 struct perf_file_header f_header;
2791 struct perf_file_attr f_attr;
2793 int nr_attrs, nr_ids, i, j;
2795 session->evlist = perf_evlist__new();
2796 if (session->evlist == NULL)
2799 if (session->fd_pipe)
2800 return perf_header__read_pipe(session, fd);
2802 if (perf_file_header__read(&f_header, header, fd) < 0)
2805 nr_attrs = f_header.attrs.size / f_header.attr_size;
2806 lseek(fd, f_header.attrs.offset, SEEK_SET);
2808 for (i = 0; i < nr_attrs; i++) {
2809 struct perf_evsel *evsel;
2812 if (read_attr(fd, header, &f_attr) < 0)
2815 if (header->needs_swap)
2816 perf_event__attr_swap(&f_attr.attr);
2818 tmp = lseek(fd, 0, SEEK_CUR);
2819 evsel = perf_evsel__new(&f_attr.attr, i);
2822 goto out_delete_evlist;
2824 evsel->needs_swap = header->needs_swap;
2826 * Do it before so that if perf_evsel__alloc_id fails, this
2827 * entry gets purged too at perf_evlist__delete().
2829 perf_evlist__add(session->evlist, evsel);
2831 nr_ids = f_attr.ids.size / sizeof(u64);
2833 * We don't have the cpu and thread maps on the header, so
2834 * for allocating the perf_sample_id table we fake 1 cpu and
2835 * hattr->ids threads.
2837 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2838 goto out_delete_evlist;
2840 lseek(fd, f_attr.ids.offset, SEEK_SET);
2842 for (j = 0; j < nr_ids; j++) {
2843 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2846 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2849 lseek(fd, tmp, SEEK_SET);
2852 symbol_conf.nr_events = nr_attrs;
2854 if (f_header.event_types.size) {
2855 lseek(fd, f_header.event_types.offset, SEEK_SET);
2856 trace_events = malloc(f_header.event_types.size);
2857 if (trace_events == NULL)
2859 if (perf_header__getbuffer64(header, fd, trace_events,
2860 f_header.event_types.size))
2862 trace_event_count = f_header.event_types.size / sizeof(struct perf_trace_event_type);
2865 perf_header__process_sections(header, fd, &session->pevent,
2866 perf_file_section__process);
2868 lseek(fd, header->data_offset, SEEK_SET);
2870 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2872 goto out_delete_evlist;
2880 perf_evlist__delete(session->evlist);
2881 session->evlist = NULL;
2885 int perf_event__synthesize_attr(struct perf_tool *tool,
2886 struct perf_event_attr *attr, u32 ids, u64 *id,
2887 perf_event__handler_t process)
2889 union perf_event *ev;
2893 size = sizeof(struct perf_event_attr);
2894 size = PERF_ALIGN(size, sizeof(u64));
2895 size += sizeof(struct perf_event_header);
2896 size += ids * sizeof(u64);
2903 ev->attr.attr = *attr;
2904 memcpy(ev->attr.id, id, ids * sizeof(u64));
2906 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2907 ev->attr.header.size = (u16)size;
2909 if (ev->attr.header.size == size)
2910 err = process(tool, ev, NULL, NULL);
2919 int perf_event__synthesize_attrs(struct perf_tool *tool,
2920 struct perf_session *session,
2921 perf_event__handler_t process)
2923 struct perf_evsel *evsel;
2926 list_for_each_entry(evsel, &session->evlist->entries, node) {
2927 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2928 evsel->id, process);
2930 pr_debug("failed to create perf header attribute\n");
2938 int perf_event__process_attr(union perf_event *event,
2939 struct perf_evlist **pevlist)
2942 struct perf_evsel *evsel;
2943 struct perf_evlist *evlist = *pevlist;
2945 if (evlist == NULL) {
2946 *pevlist = evlist = perf_evlist__new();
2951 evsel = perf_evsel__new(&event->attr.attr, evlist->nr_entries);
2955 perf_evlist__add(evlist, evsel);
2957 ids = event->header.size;
2958 ids -= (void *)&event->attr.id - (void *)event;
2959 n_ids = ids / sizeof(u64);
2961 * We don't have the cpu and thread maps on the header, so
2962 * for allocating the perf_sample_id table we fake 1 cpu and
2963 * hattr->ids threads.
2965 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2968 for (i = 0; i < n_ids; i++) {
2969 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2975 int perf_event__synthesize_event_type(struct perf_tool *tool,
2976 u64 event_id, char *name,
2977 perf_event__handler_t process,
2978 struct machine *machine)
2980 union perf_event ev;
2984 memset(&ev, 0, sizeof(ev));
2986 ev.event_type.event_type.event_id = event_id;
2987 memset(ev.event_type.event_type.name, 0, MAX_EVENT_NAME);
2988 strncpy(ev.event_type.event_type.name, name, MAX_EVENT_NAME - 1);
2990 ev.event_type.header.type = PERF_RECORD_HEADER_EVENT_TYPE;
2991 size = strlen(ev.event_type.event_type.name);
2992 size = PERF_ALIGN(size, sizeof(u64));
2993 ev.event_type.header.size = sizeof(ev.event_type) -
2994 (sizeof(ev.event_type.event_type.name) - size);
2996 err = process(tool, &ev, NULL, machine);
3001 int perf_event__synthesize_event_types(struct perf_tool *tool,
3002 perf_event__handler_t process,
3003 struct machine *machine)
3005 struct perf_trace_event_type *type;
3008 for (i = 0; i < trace_event_count; i++) {
3009 type = &trace_events[i];
3011 err = perf_event__synthesize_event_type(tool, type->event_id,
3012 type->name, process,
3015 pr_debug("failed to create perf header event type\n");
3023 int perf_event__process_event_type(struct perf_tool *tool __maybe_unused,
3024 union perf_event *event)
3026 if (perf_header__push_event(event->event_type.event_type.event_id,
3027 event->event_type.event_type.name) < 0)
3033 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
3034 struct perf_evlist *evlist,
3035 perf_event__handler_t process)
3037 union perf_event ev;
3038 struct tracing_data *tdata;
3039 ssize_t size = 0, aligned_size = 0, padding;
3040 int err __maybe_unused = 0;
3043 * We are going to store the size of the data followed
3044 * by the data contents. Since the fd descriptor is a pipe,
3045 * we cannot seek back to store the size of the data once
3046 * we know it. Instead we:
3048 * - write the tracing data to the temp file
3049 * - get/write the data size to pipe
3050 * - write the tracing data from the temp file
3053 tdata = tracing_data_get(&evlist->entries, fd, true);
3057 memset(&ev, 0, sizeof(ev));
3059 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
3061 aligned_size = PERF_ALIGN(size, sizeof(u64));
3062 padding = aligned_size - size;
3063 ev.tracing_data.header.size = sizeof(ev.tracing_data);
3064 ev.tracing_data.size = aligned_size;
3066 process(tool, &ev, NULL, NULL);
3069 * The put function will copy all the tracing data
3070 * stored in temp file to the pipe.
3072 tracing_data_put(tdata);
3074 write_padded(fd, NULL, 0, padding);
3076 return aligned_size;
3079 int perf_event__process_tracing_data(union perf_event *event,
3080 struct perf_session *session)
3082 ssize_t size_read, padding, size = event->tracing_data.size;
3083 off_t offset = lseek(session->fd, 0, SEEK_CUR);
3086 /* setup for reading amidst mmap */
3087 lseek(session->fd, offset + sizeof(struct tracing_data_event),
3090 size_read = trace_report(session->fd, &session->pevent,
3092 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3094 if (readn(session->fd, buf, padding) < 0) {
3095 pr_err("%s: reading input file", __func__);
3098 if (session->repipe) {
3099 int retw = write(STDOUT_FILENO, buf, padding);
3100 if (retw <= 0 || retw != padding) {
3101 pr_err("%s: repiping tracing data padding", __func__);
3106 if (size_read + padding != size) {
3107 pr_err("%s: tracing data size mismatch", __func__);
3111 perf_evlist__prepare_tracepoint_events(session->evlist,
3114 return size_read + padding;
3117 int perf_event__synthesize_build_id(struct perf_tool *tool,
3118 struct dso *pos, u16 misc,
3119 perf_event__handler_t process,
3120 struct machine *machine)
3122 union perf_event ev;
3129 memset(&ev, 0, sizeof(ev));
3131 len = pos->long_name_len + 1;
3132 len = PERF_ALIGN(len, NAME_ALIGN);
3133 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3134 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3135 ev.build_id.header.misc = misc;
3136 ev.build_id.pid = machine->pid;
3137 ev.build_id.header.size = sizeof(ev.build_id) + len;
3138 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3140 err = process(tool, &ev, NULL, machine);
3145 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3146 union perf_event *event,
3147 struct perf_session *session)
3149 __event_process_build_id(&event->build_id,
3150 event->build_id.filename,
3155 void disable_buildid_cache(void)
3157 no_buildid_cache = true;