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 u32 header_argc;
29 static const char **header_argv;
33 * must be a numerical value to let the endianness
34 * determine the memory layout. That way we are able
35 * to detect endianness when reading the perf.data file
38 * we check for legacy (PERFFILE) format.
40 static const char *__perf_magic1 = "PERFFILE";
41 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
42 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
44 #define PERF_MAGIC __perf_magic2
46 struct perf_file_attr {
47 struct perf_event_attr attr;
48 struct perf_file_section ids;
51 void perf_header__set_feat(struct perf_header *header, int feat)
53 set_bit(feat, header->adds_features);
56 void perf_header__clear_feat(struct perf_header *header, int feat)
58 clear_bit(feat, header->adds_features);
61 bool perf_header__has_feat(const struct perf_header *header, int feat)
63 return test_bit(feat, header->adds_features);
66 static int do_write(int fd, const void *buf, size_t size)
69 int ret = write(fd, buf, size);
83 static int write_padded(int fd, const void *bf, size_t count,
86 static const char zero_buf[NAME_ALIGN];
87 int err = do_write(fd, bf, count);
90 err = do_write(fd, zero_buf, count_aligned - count);
95 static int do_write_string(int fd, const char *str)
100 olen = strlen(str) + 1;
101 len = PERF_ALIGN(olen, NAME_ALIGN);
103 /* write len, incl. \0 */
104 ret = do_write(fd, &len, sizeof(len));
108 return write_padded(fd, str, olen, len);
111 static char *do_read_string(int fd, struct perf_header *ph)
117 sz = readn(fd, &len, sizeof(len));
118 if (sz < (ssize_t)sizeof(len))
128 ret = readn(fd, buf, len);
129 if (ret == (ssize_t)len) {
131 * strings are padded by zeroes
132 * thus the actual strlen of buf
133 * may be less than len
143 perf_header__set_cmdline(int argc, const char **argv)
148 * If header_argv has already been set, do not override it.
149 * This allows a command to set the cmdline, parse args and
150 * then call another builtin function that implements a
151 * command -- e.g, cmd_kvm calling cmd_record.
156 header_argc = (u32)argc;
158 /* do not include NULL termination */
159 header_argv = calloc(argc, sizeof(char *));
164 * must copy argv contents because it gets moved
165 * around during option parsing
167 for (i = 0; i < argc ; i++)
168 header_argv[i] = argv[i];
173 #define dsos__for_each_with_build_id(pos, head) \
174 list_for_each_entry(pos, head, node) \
175 if (!pos->has_build_id) \
179 static int write_buildid(char *name, size_t name_len, u8 *build_id,
180 pid_t pid, u16 misc, int fd)
183 struct build_id_event b;
187 len = PERF_ALIGN(len, NAME_ALIGN);
189 memset(&b, 0, sizeof(b));
190 memcpy(&b.build_id, build_id, BUILD_ID_SIZE);
192 b.header.misc = misc;
193 b.header.size = sizeof(b) + len;
195 err = do_write(fd, &b, sizeof(b));
199 return write_padded(fd, name, name_len + 1, len);
202 static int __dsos__write_buildid_table(struct list_head *head, pid_t pid,
207 dsos__for_each_with_build_id(pos, head) {
215 if (is_vdso_map(pos->short_name)) {
216 name = (char *) VDSO__MAP_NAME;
217 name_len = sizeof(VDSO__MAP_NAME) + 1;
219 name = pos->long_name;
220 name_len = pos->long_name_len + 1;
223 err = write_buildid(name, name_len, pos->build_id,
232 static int machine__write_buildid_table(struct machine *machine, int fd)
235 u16 kmisc = PERF_RECORD_MISC_KERNEL,
236 umisc = PERF_RECORD_MISC_USER;
238 if (!machine__is_host(machine)) {
239 kmisc = PERF_RECORD_MISC_GUEST_KERNEL;
240 umisc = PERF_RECORD_MISC_GUEST_USER;
243 err = __dsos__write_buildid_table(&machine->kernel_dsos, machine->pid,
246 err = __dsos__write_buildid_table(&machine->user_dsos,
247 machine->pid, umisc, fd);
251 static int dsos__write_buildid_table(struct perf_header *header, int fd)
253 struct perf_session *session = container_of(header,
254 struct perf_session, header);
256 int err = machine__write_buildid_table(&session->machines.host, fd);
261 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
262 struct machine *pos = rb_entry(nd, struct machine, rb_node);
263 err = machine__write_buildid_table(pos, fd);
270 int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
271 const char *name, bool is_kallsyms, bool is_vdso)
273 const size_t size = PATH_MAX;
274 char *realname, *filename = zalloc(size),
275 *linkname = zalloc(size), *targetname;
277 bool slash = is_kallsyms || is_vdso;
280 if (symbol_conf.kptr_restrict) {
281 pr_debug("Not caching a kptr_restrict'ed /proc/kallsyms\n");
285 realname = (char *) name;
287 realname = realpath(name, NULL);
289 if (realname == NULL || filename == NULL || linkname == NULL)
292 len = scnprintf(filename, size, "%s%s%s",
293 debugdir, slash ? "/" : "",
294 is_vdso ? VDSO__MAP_NAME : realname);
295 if (mkdir_p(filename, 0755))
298 snprintf(filename + len, size - len, "/%s", sbuild_id);
300 if (access(filename, F_OK)) {
302 if (copyfile("/proc/kallsyms", filename))
304 } else if (link(realname, filename) && copyfile(name, filename))
308 len = scnprintf(linkname, size, "%s/.build-id/%.2s",
309 debugdir, sbuild_id);
311 if (access(linkname, X_OK) && mkdir_p(linkname, 0755))
314 snprintf(linkname + len, size - len, "/%s", sbuild_id + 2);
315 targetname = filename + strlen(debugdir) - 5;
316 memcpy(targetname, "../..", 5);
318 if (symlink(targetname, linkname) == 0)
328 static int build_id_cache__add_b(const u8 *build_id, size_t build_id_size,
329 const char *name, const char *debugdir,
330 bool is_kallsyms, bool is_vdso)
332 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
334 build_id__sprintf(build_id, build_id_size, sbuild_id);
336 return build_id_cache__add_s(sbuild_id, debugdir, name,
337 is_kallsyms, is_vdso);
340 int build_id_cache__remove_s(const char *sbuild_id, const char *debugdir)
342 const size_t size = PATH_MAX;
343 char *filename = zalloc(size),
344 *linkname = zalloc(size);
347 if (filename == NULL || linkname == NULL)
350 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
351 debugdir, sbuild_id, sbuild_id + 2);
353 if (access(linkname, F_OK))
356 if (readlink(linkname, filename, size - 1) < 0)
359 if (unlink(linkname))
363 * Since the link is relative, we must make it absolute:
365 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
366 debugdir, sbuild_id, filename);
368 if (unlink(linkname))
378 static int dso__cache_build_id(struct dso *dso, const char *debugdir)
380 bool is_kallsyms = dso->kernel && dso->long_name[0] != '/';
381 bool is_vdso = is_vdso_map(dso->short_name);
383 return build_id_cache__add_b(dso->build_id, sizeof(dso->build_id),
384 dso->long_name, debugdir,
385 is_kallsyms, is_vdso);
388 static int __dsos__cache_build_ids(struct list_head *head, const char *debugdir)
393 dsos__for_each_with_build_id(pos, head)
394 if (dso__cache_build_id(pos, debugdir))
400 static int machine__cache_build_ids(struct machine *machine, const char *debugdir)
402 int ret = __dsos__cache_build_ids(&machine->kernel_dsos, debugdir);
403 ret |= __dsos__cache_build_ids(&machine->user_dsos, debugdir);
407 static int perf_session__cache_build_ids(struct perf_session *session)
411 char debugdir[PATH_MAX];
413 snprintf(debugdir, sizeof(debugdir), "%s", buildid_dir);
415 if (mkdir(debugdir, 0755) != 0 && errno != EEXIST)
418 ret = machine__cache_build_ids(&session->machines.host, debugdir);
420 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
421 struct machine *pos = rb_entry(nd, struct machine, rb_node);
422 ret |= machine__cache_build_ids(pos, debugdir);
427 static bool machine__read_build_ids(struct machine *machine, bool with_hits)
429 bool ret = __dsos__read_build_ids(&machine->kernel_dsos, with_hits);
430 ret |= __dsos__read_build_ids(&machine->user_dsos, with_hits);
434 static bool perf_session__read_build_ids(struct perf_session *session, bool with_hits)
437 bool ret = machine__read_build_ids(&session->machines.host, with_hits);
439 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
440 struct machine *pos = rb_entry(nd, struct machine, rb_node);
441 ret |= machine__read_build_ids(pos, with_hits);
447 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
448 struct perf_evlist *evlist)
450 return read_tracing_data(fd, &evlist->entries);
454 static int write_build_id(int fd, struct perf_header *h,
455 struct perf_evlist *evlist __maybe_unused)
457 struct perf_session *session;
460 session = container_of(h, struct perf_session, header);
462 if (!perf_session__read_build_ids(session, true))
465 err = dsos__write_buildid_table(h, fd);
467 pr_debug("failed to write buildid table\n");
470 if (!no_buildid_cache)
471 perf_session__cache_build_ids(session);
476 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
477 struct perf_evlist *evlist __maybe_unused)
486 return do_write_string(fd, uts.nodename);
489 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
490 struct perf_evlist *evlist __maybe_unused)
499 return do_write_string(fd, uts.release);
502 static int write_arch(int fd, struct perf_header *h __maybe_unused,
503 struct perf_evlist *evlist __maybe_unused)
512 return do_write_string(fd, uts.machine);
515 static int write_version(int fd, struct perf_header *h __maybe_unused,
516 struct perf_evlist *evlist __maybe_unused)
518 return do_write_string(fd, perf_version_string);
521 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
522 struct perf_evlist *evlist __maybe_unused)
525 #define CPUINFO_PROC NULL
530 const char *search = CPUINFO_PROC;
537 file = fopen("/proc/cpuinfo", "r");
541 while (getline(&buf, &len, file) > 0) {
542 ret = strncmp(buf, search, strlen(search));
552 p = strchr(buf, ':');
553 if (p && *(p+1) == ' ' && *(p+2))
559 /* squash extra space characters (branding string) */
566 while (*q && isspace(*q))
569 while ((*r++ = *q++));
573 ret = do_write_string(fd, s);
580 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
581 struct perf_evlist *evlist __maybe_unused)
587 nr = sysconf(_SC_NPROCESSORS_CONF);
591 nrc = (u32)(nr & UINT_MAX);
593 nr = sysconf(_SC_NPROCESSORS_ONLN);
597 nra = (u32)(nr & UINT_MAX);
599 ret = do_write(fd, &nrc, sizeof(nrc));
603 return do_write(fd, &nra, sizeof(nra));
606 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
607 struct perf_evlist *evlist)
609 struct perf_evsel *evsel;
613 nre = evlist->nr_entries;
616 * write number of events
618 ret = do_write(fd, &nre, sizeof(nre));
623 * size of perf_event_attr struct
625 sz = (u32)sizeof(evsel->attr);
626 ret = do_write(fd, &sz, sizeof(sz));
630 list_for_each_entry(evsel, &evlist->entries, node) {
632 ret = do_write(fd, &evsel->attr, sz);
636 * write number of unique id per event
637 * there is one id per instance of an event
639 * copy into an nri to be independent of the
643 ret = do_write(fd, &nri, sizeof(nri));
648 * write event string as passed on cmdline
650 ret = do_write_string(fd, perf_evsel__name(evsel));
654 * write unique ids for this event
656 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
663 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
664 struct perf_evlist *evlist __maybe_unused)
666 char buf[MAXPATHLEN];
672 * actual atual path to perf binary
674 sprintf(proc, "/proc/%d/exe", getpid());
675 ret = readlink(proc, buf, sizeof(buf));
679 /* readlink() does not add null termination */
682 /* account for binary path */
685 ret = do_write(fd, &n, sizeof(n));
689 ret = do_write_string(fd, buf);
693 for (i = 0 ; i < header_argc; i++) {
694 ret = do_write_string(fd, header_argv[i]);
701 #define CORE_SIB_FMT \
702 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
703 #define THRD_SIB_FMT \
704 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
709 char **core_siblings;
710 char **thread_siblings;
713 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
716 char filename[MAXPATHLEN];
717 char *buf = NULL, *p;
722 sprintf(filename, CORE_SIB_FMT, cpu);
723 fp = fopen(filename, "r");
727 if (getline(&buf, &len, fp) <= 0)
732 p = strchr(buf, '\n');
736 for (i = 0; i < tp->core_sib; i++) {
737 if (!strcmp(buf, tp->core_siblings[i]))
740 if (i == tp->core_sib) {
741 tp->core_siblings[i] = buf;
747 sprintf(filename, THRD_SIB_FMT, cpu);
748 fp = fopen(filename, "r");
752 if (getline(&buf, &len, fp) <= 0)
755 p = strchr(buf, '\n');
759 for (i = 0; i < tp->thread_sib; i++) {
760 if (!strcmp(buf, tp->thread_siblings[i]))
763 if (i == tp->thread_sib) {
764 tp->thread_siblings[i] = buf;
776 static void free_cpu_topo(struct cpu_topo *tp)
783 for (i = 0 ; i < tp->core_sib; i++)
784 free(tp->core_siblings[i]);
786 for (i = 0 ; i < tp->thread_sib; i++)
787 free(tp->thread_siblings[i]);
792 static struct cpu_topo *build_cpu_topology(void)
801 ncpus = sysconf(_SC_NPROCESSORS_CONF);
805 nr = (u32)(ncpus & UINT_MAX);
807 sz = nr * sizeof(char *);
809 addr = calloc(1, sizeof(*tp) + 2 * sz);
816 tp->core_siblings = addr;
818 tp->thread_siblings = addr;
820 for (i = 0; i < nr; i++) {
821 ret = build_cpu_topo(tp, i);
832 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
833 struct perf_evlist *evlist __maybe_unused)
839 tp = build_cpu_topology();
843 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
847 for (i = 0; i < tp->core_sib; i++) {
848 ret = do_write_string(fd, tp->core_siblings[i]);
852 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
856 for (i = 0; i < tp->thread_sib; i++) {
857 ret = do_write_string(fd, tp->thread_siblings[i]);
868 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
869 struct perf_evlist *evlist __maybe_unused)
877 fp = fopen("/proc/meminfo", "r");
881 while (getline(&buf, &len, fp) > 0) {
882 ret = strncmp(buf, "MemTotal:", 9);
887 n = sscanf(buf, "%*s %"PRIu64, &mem);
889 ret = do_write(fd, &mem, sizeof(mem));
896 static int write_topo_node(int fd, int node)
898 char str[MAXPATHLEN];
900 char *buf = NULL, *p;
903 u64 mem_total, mem_free, mem;
906 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
907 fp = fopen(str, "r");
911 while (getline(&buf, &len, fp) > 0) {
912 /* skip over invalid lines */
913 if (!strchr(buf, ':'))
915 if (sscanf(buf, "%*s %*d %s %"PRIu64, field, &mem) != 2)
917 if (!strcmp(field, "MemTotal:"))
919 if (!strcmp(field, "MemFree:"))
926 ret = do_write(fd, &mem_total, sizeof(u64));
930 ret = do_write(fd, &mem_free, sizeof(u64));
935 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
937 fp = fopen(str, "r");
941 if (getline(&buf, &len, fp) <= 0)
944 p = strchr(buf, '\n');
948 ret = do_write_string(fd, buf);
956 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
957 struct perf_evlist *evlist __maybe_unused)
962 struct cpu_map *node_map = NULL;
967 fp = fopen("/sys/devices/system/node/online", "r");
971 if (getline(&buf, &len, fp) <= 0)
974 c = strchr(buf, '\n');
978 node_map = cpu_map__new(buf);
982 nr = (u32)node_map->nr;
984 ret = do_write(fd, &nr, sizeof(nr));
988 for (i = 0; i < nr; i++) {
989 j = (u32)node_map->map[i];
990 ret = do_write(fd, &j, sizeof(j));
994 ret = write_topo_node(fd, i);
1008 * struct pmu_mappings {
1017 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
1018 struct perf_evlist *evlist __maybe_unused)
1020 struct perf_pmu *pmu = NULL;
1021 off_t offset = lseek(fd, 0, SEEK_CUR);
1025 /* write real pmu_num later */
1026 ret = do_write(fd, &pmu_num, sizeof(pmu_num));
1030 while ((pmu = perf_pmu__scan(pmu))) {
1035 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
1039 ret = do_write_string(fd, pmu->name);
1044 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
1046 lseek(fd, offset, SEEK_SET);
1056 * struct group_descs {
1058 * struct group_desc {
1065 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
1066 struct perf_evlist *evlist)
1068 u32 nr_groups = evlist->nr_groups;
1069 struct perf_evsel *evsel;
1072 ret = do_write(fd, &nr_groups, sizeof(nr_groups));
1076 list_for_each_entry(evsel, &evlist->entries, node) {
1077 if (perf_evsel__is_group_leader(evsel) &&
1078 evsel->nr_members > 1) {
1079 const char *name = evsel->group_name ?: "{anon_group}";
1080 u32 leader_idx = evsel->idx;
1081 u32 nr_members = evsel->nr_members;
1083 ret = do_write_string(fd, name);
1087 ret = do_write(fd, &leader_idx, sizeof(leader_idx));
1091 ret = do_write(fd, &nr_members, sizeof(nr_members));
1100 * default get_cpuid(): nothing gets recorded
1101 * actual implementation must be in arch/$(ARCH)/util/header.c
1103 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
1104 size_t sz __maybe_unused)
1109 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
1110 struct perf_evlist *evlist __maybe_unused)
1115 ret = get_cpuid(buffer, sizeof(buffer));
1121 return do_write_string(fd, buffer);
1124 static int write_branch_stack(int fd __maybe_unused,
1125 struct perf_header *h __maybe_unused,
1126 struct perf_evlist *evlist __maybe_unused)
1131 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
1134 fprintf(fp, "# hostname : %s\n", ph->env.hostname);
1137 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
1140 fprintf(fp, "# os release : %s\n", ph->env.os_release);
1143 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1145 fprintf(fp, "# arch : %s\n", ph->env.arch);
1148 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
1151 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
1154 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
1157 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
1158 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
1161 static void print_version(struct perf_header *ph, int fd __maybe_unused,
1164 fprintf(fp, "# perf version : %s\n", ph->env.version);
1167 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
1173 nr = ph->env.nr_cmdline;
1174 str = ph->env.cmdline;
1176 fprintf(fp, "# cmdline : ");
1178 for (i = 0; i < nr; i++) {
1179 fprintf(fp, "%s ", str);
1180 str += strlen(str) + 1;
1185 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
1191 nr = ph->env.nr_sibling_cores;
1192 str = ph->env.sibling_cores;
1194 for (i = 0; i < nr; i++) {
1195 fprintf(fp, "# sibling cores : %s\n", str);
1196 str += strlen(str) + 1;
1199 nr = ph->env.nr_sibling_threads;
1200 str = ph->env.sibling_threads;
1202 for (i = 0; i < nr; i++) {
1203 fprintf(fp, "# sibling threads : %s\n", str);
1204 str += strlen(str) + 1;
1208 static void free_event_desc(struct perf_evsel *events)
1210 struct perf_evsel *evsel;
1215 for (evsel = events; evsel->attr.size; evsel++) {
1225 static struct perf_evsel *
1226 read_event_desc(struct perf_header *ph, int fd)
1228 struct perf_evsel *evsel, *events = NULL;
1231 u32 nre, sz, nr, i, j;
1235 /* number of events */
1236 ret = readn(fd, &nre, sizeof(nre));
1237 if (ret != (ssize_t)sizeof(nre))
1241 nre = bswap_32(nre);
1243 ret = readn(fd, &sz, sizeof(sz));
1244 if (ret != (ssize_t)sizeof(sz))
1250 /* buffer to hold on file attr struct */
1255 /* the last event terminates with evsel->attr.size == 0: */
1256 events = calloc(nre + 1, sizeof(*events));
1260 msz = sizeof(evsel->attr);
1264 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1268 * must read entire on-file attr struct to
1269 * sync up with layout.
1271 ret = readn(fd, buf, sz);
1272 if (ret != (ssize_t)sz)
1276 perf_event__attr_swap(buf);
1278 memcpy(&evsel->attr, buf, msz);
1280 ret = readn(fd, &nr, sizeof(nr));
1281 if (ret != (ssize_t)sizeof(nr))
1284 if (ph->needs_swap) {
1286 evsel->needs_swap = true;
1289 evsel->name = do_read_string(fd, ph);
1294 id = calloc(nr, sizeof(*id));
1300 for (j = 0 ; j < nr; j++) {
1301 ret = readn(fd, id, sizeof(*id));
1302 if (ret != (ssize_t)sizeof(*id))
1305 *id = bswap_64(*id);
1315 free_event_desc(events);
1320 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1322 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1327 fprintf(fp, "# event desc: not available or unable to read\n");
1331 for (evsel = events; evsel->attr.size; evsel++) {
1332 fprintf(fp, "# event : name = %s, ", evsel->name);
1334 fprintf(fp, "type = %d, config = 0x%"PRIx64
1335 ", config1 = 0x%"PRIx64", config2 = 0x%"PRIx64,
1337 (u64)evsel->attr.config,
1338 (u64)evsel->attr.config1,
1339 (u64)evsel->attr.config2);
1341 fprintf(fp, ", excl_usr = %d, excl_kern = %d",
1342 evsel->attr.exclude_user,
1343 evsel->attr.exclude_kernel);
1345 fprintf(fp, ", excl_host = %d, excl_guest = %d",
1346 evsel->attr.exclude_host,
1347 evsel->attr.exclude_guest);
1349 fprintf(fp, ", precise_ip = %d", evsel->attr.precise_ip);
1352 fprintf(fp, ", id = {");
1353 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1356 fprintf(fp, " %"PRIu64, *id);
1364 free_event_desc(events);
1367 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1370 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1373 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1378 uint64_t mem_total, mem_free;
1381 nr = ph->env.nr_numa_nodes;
1382 str = ph->env.numa_nodes;
1384 for (i = 0; i < nr; i++) {
1386 c = strtoul(str, &tmp, 0);
1391 mem_total = strtoull(str, &tmp, 0);
1396 mem_free = strtoull(str, &tmp, 0);
1400 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1401 " free = %"PRIu64" kB\n",
1402 c, mem_total, mem_free);
1405 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1407 str += strlen(str) + 1;
1411 fprintf(fp, "# numa topology : not available\n");
1414 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1416 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1419 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1420 int fd __maybe_unused, FILE *fp)
1422 fprintf(fp, "# contains samples with branch stack\n");
1425 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1428 const char *delimiter = "# pmu mappings: ";
1433 pmu_num = ph->env.nr_pmu_mappings;
1435 fprintf(fp, "# pmu mappings: not available\n");
1439 str = ph->env.pmu_mappings;
1442 type = strtoul(str, &tmp, 0);
1447 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1450 str += strlen(str) + 1;
1459 fprintf(fp, "# pmu mappings: unable to read\n");
1462 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1465 struct perf_session *session;
1466 struct perf_evsel *evsel;
1469 session = container_of(ph, struct perf_session, header);
1471 list_for_each_entry(evsel, &session->evlist->entries, node) {
1472 if (perf_evsel__is_group_leader(evsel) &&
1473 evsel->nr_members > 1) {
1474 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1475 perf_evsel__name(evsel));
1477 nr = evsel->nr_members - 1;
1479 fprintf(fp, ",%s", perf_evsel__name(evsel));
1487 static int __event_process_build_id(struct build_id_event *bev,
1489 struct perf_session *session)
1492 struct list_head *head;
1493 struct machine *machine;
1496 enum dso_kernel_type dso_type;
1498 machine = perf_session__findnew_machine(session, bev->pid);
1502 misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1505 case PERF_RECORD_MISC_KERNEL:
1506 dso_type = DSO_TYPE_KERNEL;
1507 head = &machine->kernel_dsos;
1509 case PERF_RECORD_MISC_GUEST_KERNEL:
1510 dso_type = DSO_TYPE_GUEST_KERNEL;
1511 head = &machine->kernel_dsos;
1513 case PERF_RECORD_MISC_USER:
1514 case PERF_RECORD_MISC_GUEST_USER:
1515 dso_type = DSO_TYPE_USER;
1516 head = &machine->user_dsos;
1522 dso = __dsos__findnew(head, filename);
1524 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1526 dso__set_build_id(dso, &bev->build_id);
1528 if (filename[0] == '[')
1529 dso->kernel = dso_type;
1531 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1533 pr_debug("build id event received for %s: %s\n",
1534 dso->long_name, sbuild_id);
1542 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1543 int input, u64 offset, u64 size)
1545 struct perf_session *session = container_of(header, struct perf_session, header);
1547 struct perf_event_header header;
1548 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1551 struct build_id_event bev;
1552 char filename[PATH_MAX];
1553 u64 limit = offset + size;
1555 while (offset < limit) {
1558 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1561 if (header->needs_swap)
1562 perf_event_header__bswap(&old_bev.header);
1564 len = old_bev.header.size - sizeof(old_bev);
1565 if (readn(input, filename, len) != len)
1568 bev.header = old_bev.header;
1571 * As the pid is the missing value, we need to fill
1572 * it properly. The header.misc value give us nice hint.
1574 bev.pid = HOST_KERNEL_ID;
1575 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1576 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1577 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1579 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1580 __event_process_build_id(&bev, filename, session);
1582 offset += bev.header.size;
1588 static int perf_header__read_build_ids(struct perf_header *header,
1589 int input, u64 offset, u64 size)
1591 struct perf_session *session = container_of(header, struct perf_session, header);
1592 struct build_id_event bev;
1593 char filename[PATH_MAX];
1594 u64 limit = offset + size, orig_offset = offset;
1597 while (offset < limit) {
1600 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1603 if (header->needs_swap)
1604 perf_event_header__bswap(&bev.header);
1606 len = bev.header.size - sizeof(bev);
1607 if (readn(input, filename, len) != len)
1610 * The a1645ce1 changeset:
1612 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1614 * Added a field to struct build_id_event that broke the file
1617 * Since the kernel build-id is the first entry, process the
1618 * table using the old format if the well known
1619 * '[kernel.kallsyms]' string for the kernel build-id has the
1620 * first 4 characters chopped off (where the pid_t sits).
1622 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1623 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1625 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1628 __event_process_build_id(&bev, filename, session);
1630 offset += bev.header.size;
1637 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1638 struct perf_header *ph __maybe_unused,
1641 ssize_t ret = trace_report(fd, data, false);
1642 return ret < 0 ? -1 : 0;
1645 static int process_build_id(struct perf_file_section *section,
1646 struct perf_header *ph, int fd,
1647 void *data __maybe_unused)
1649 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1650 pr_debug("Failed to read buildids, continuing...\n");
1654 static int process_hostname(struct perf_file_section *section __maybe_unused,
1655 struct perf_header *ph, int fd,
1656 void *data __maybe_unused)
1658 ph->env.hostname = do_read_string(fd, ph);
1659 return ph->env.hostname ? 0 : -ENOMEM;
1662 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1663 struct perf_header *ph, int fd,
1664 void *data __maybe_unused)
1666 ph->env.os_release = do_read_string(fd, ph);
1667 return ph->env.os_release ? 0 : -ENOMEM;
1670 static int process_version(struct perf_file_section *section __maybe_unused,
1671 struct perf_header *ph, int fd,
1672 void *data __maybe_unused)
1674 ph->env.version = do_read_string(fd, ph);
1675 return ph->env.version ? 0 : -ENOMEM;
1678 static int process_arch(struct perf_file_section *section __maybe_unused,
1679 struct perf_header *ph, int fd,
1680 void *data __maybe_unused)
1682 ph->env.arch = do_read_string(fd, ph);
1683 return ph->env.arch ? 0 : -ENOMEM;
1686 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1687 struct perf_header *ph, int fd,
1688 void *data __maybe_unused)
1693 ret = readn(fd, &nr, sizeof(nr));
1694 if (ret != sizeof(nr))
1700 ph->env.nr_cpus_online = nr;
1702 ret = readn(fd, &nr, sizeof(nr));
1703 if (ret != sizeof(nr))
1709 ph->env.nr_cpus_avail = nr;
1713 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1714 struct perf_header *ph, int fd,
1715 void *data __maybe_unused)
1717 ph->env.cpu_desc = do_read_string(fd, ph);
1718 return ph->env.cpu_desc ? 0 : -ENOMEM;
1721 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1722 struct perf_header *ph, int fd,
1723 void *data __maybe_unused)
1725 ph->env.cpuid = do_read_string(fd, ph);
1726 return ph->env.cpuid ? 0 : -ENOMEM;
1729 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1730 struct perf_header *ph, int fd,
1731 void *data __maybe_unused)
1736 ret = readn(fd, &mem, sizeof(mem));
1737 if (ret != sizeof(mem))
1741 mem = bswap_64(mem);
1743 ph->env.total_mem = mem;
1747 static struct perf_evsel *
1748 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1750 struct perf_evsel *evsel;
1752 list_for_each_entry(evsel, &evlist->entries, node) {
1753 if (evsel->idx == idx)
1761 perf_evlist__set_event_name(struct perf_evlist *evlist,
1762 struct perf_evsel *event)
1764 struct perf_evsel *evsel;
1769 evsel = perf_evlist__find_by_index(evlist, event->idx);
1776 evsel->name = strdup(event->name);
1780 process_event_desc(struct perf_file_section *section __maybe_unused,
1781 struct perf_header *header, int fd,
1782 void *data __maybe_unused)
1784 struct perf_session *session;
1785 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1790 session = container_of(header, struct perf_session, header);
1791 for (evsel = events; evsel->attr.size; evsel++)
1792 perf_evlist__set_event_name(session->evlist, evsel);
1794 free_event_desc(events);
1799 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1800 struct perf_header *ph, int fd,
1801 void *data __maybe_unused)
1808 ret = readn(fd, &nr, sizeof(nr));
1809 if (ret != sizeof(nr))
1815 ph->env.nr_cmdline = nr;
1816 strbuf_init(&sb, 128);
1818 for (i = 0; i < nr; i++) {
1819 str = do_read_string(fd, ph);
1823 /* include a NULL character at the end */
1824 strbuf_add(&sb, str, strlen(str) + 1);
1827 ph->env.cmdline = strbuf_detach(&sb, NULL);
1831 strbuf_release(&sb);
1835 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1836 struct perf_header *ph, int fd,
1837 void *data __maybe_unused)
1844 ret = readn(fd, &nr, sizeof(nr));
1845 if (ret != sizeof(nr))
1851 ph->env.nr_sibling_cores = nr;
1852 strbuf_init(&sb, 128);
1854 for (i = 0; i < nr; i++) {
1855 str = do_read_string(fd, ph);
1859 /* include a NULL character at the end */
1860 strbuf_add(&sb, str, strlen(str) + 1);
1863 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1865 ret = readn(fd, &nr, sizeof(nr));
1866 if (ret != sizeof(nr))
1872 ph->env.nr_sibling_threads = nr;
1874 for (i = 0; i < nr; i++) {
1875 str = do_read_string(fd, ph);
1879 /* include a NULL character at the end */
1880 strbuf_add(&sb, str, strlen(str) + 1);
1883 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1887 strbuf_release(&sb);
1891 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1892 struct perf_header *ph, int fd,
1893 void *data __maybe_unused)
1898 uint64_t mem_total, mem_free;
1902 ret = readn(fd, &nr, sizeof(nr));
1903 if (ret != sizeof(nr))
1909 ph->env.nr_numa_nodes = nr;
1910 strbuf_init(&sb, 256);
1912 for (i = 0; i < nr; i++) {
1914 ret = readn(fd, &node, sizeof(node));
1915 if (ret != sizeof(node))
1918 ret = readn(fd, &mem_total, sizeof(u64));
1919 if (ret != sizeof(u64))
1922 ret = readn(fd, &mem_free, sizeof(u64));
1923 if (ret != sizeof(u64))
1926 if (ph->needs_swap) {
1927 node = bswap_32(node);
1928 mem_total = bswap_64(mem_total);
1929 mem_free = bswap_64(mem_free);
1932 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1933 node, mem_total, mem_free);
1935 str = do_read_string(fd, ph);
1939 /* include a NULL character at the end */
1940 strbuf_add(&sb, str, strlen(str) + 1);
1943 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1947 strbuf_release(&sb);
1951 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1952 struct perf_header *ph, int fd,
1953 void *data __maybe_unused)
1961 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1962 if (ret != sizeof(pmu_num))
1966 pmu_num = bswap_32(pmu_num);
1969 pr_debug("pmu mappings not available\n");
1973 ph->env.nr_pmu_mappings = pmu_num;
1974 strbuf_init(&sb, 128);
1977 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1980 type = bswap_32(type);
1982 name = do_read_string(fd, ph);
1986 strbuf_addf(&sb, "%u:%s", type, name);
1987 /* include a NULL character at the end */
1988 strbuf_add(&sb, "", 1);
1993 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1997 strbuf_release(&sb);
2001 static int process_group_desc(struct perf_file_section *section __maybe_unused,
2002 struct perf_header *ph, int fd,
2003 void *data __maybe_unused)
2006 u32 i, nr, nr_groups;
2007 struct perf_session *session;
2008 struct perf_evsel *evsel, *leader = NULL;
2015 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
2019 nr_groups = bswap_32(nr_groups);
2021 ph->env.nr_groups = nr_groups;
2023 pr_debug("group desc not available\n");
2027 desc = calloc(nr_groups, sizeof(*desc));
2031 for (i = 0; i < nr_groups; i++) {
2032 desc[i].name = do_read_string(fd, ph);
2036 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
2039 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
2042 if (ph->needs_swap) {
2043 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
2044 desc[i].nr_members = bswap_32(desc[i].nr_members);
2049 * Rebuild group relationship based on the group_desc
2051 session = container_of(ph, struct perf_session, header);
2052 session->evlist->nr_groups = nr_groups;
2055 list_for_each_entry(evsel, &session->evlist->entries, node) {
2056 if (evsel->idx == (int) desc[i].leader_idx) {
2057 evsel->leader = evsel;
2058 /* {anon_group} is a dummy name */
2059 if (strcmp(desc[i].name, "{anon_group}"))
2060 evsel->group_name = desc[i].name;
2061 evsel->nr_members = desc[i].nr_members;
2063 if (i >= nr_groups || nr > 0) {
2064 pr_debug("invalid group desc\n");
2069 nr = evsel->nr_members - 1;
2072 /* This is a group member */
2073 evsel->leader = leader;
2079 if (i != nr_groups || nr != 0) {
2080 pr_debug("invalid group desc\n");
2086 while ((int) --i >= 0)
2093 struct feature_ops {
2094 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
2095 void (*print)(struct perf_header *h, int fd, FILE *fp);
2096 int (*process)(struct perf_file_section *section,
2097 struct perf_header *h, int fd, void *data);
2102 #define FEAT_OPA(n, func) \
2103 [n] = { .name = #n, .write = write_##func, .print = print_##func }
2104 #define FEAT_OPP(n, func) \
2105 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2106 .process = process_##func }
2107 #define FEAT_OPF(n, func) \
2108 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2109 .process = process_##func, .full_only = true }
2111 /* feature_ops not implemented: */
2112 #define print_tracing_data NULL
2113 #define print_build_id NULL
2115 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2116 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
2117 FEAT_OPP(HEADER_BUILD_ID, build_id),
2118 FEAT_OPP(HEADER_HOSTNAME, hostname),
2119 FEAT_OPP(HEADER_OSRELEASE, osrelease),
2120 FEAT_OPP(HEADER_VERSION, version),
2121 FEAT_OPP(HEADER_ARCH, arch),
2122 FEAT_OPP(HEADER_NRCPUS, nrcpus),
2123 FEAT_OPP(HEADER_CPUDESC, cpudesc),
2124 FEAT_OPP(HEADER_CPUID, cpuid),
2125 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
2126 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
2127 FEAT_OPP(HEADER_CMDLINE, cmdline),
2128 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
2129 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
2130 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
2131 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
2132 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
2135 struct header_print_data {
2137 bool full; /* extended list of headers */
2140 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2141 struct perf_header *ph,
2142 int feat, int fd, void *data)
2144 struct header_print_data *hd = data;
2146 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2147 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2148 "%d, continuing...\n", section->offset, feat);
2151 if (feat >= HEADER_LAST_FEATURE) {
2152 pr_warning("unknown feature %d\n", feat);
2155 if (!feat_ops[feat].print)
2158 if (!feat_ops[feat].full_only || hd->full)
2159 feat_ops[feat].print(ph, fd, hd->fp);
2161 fprintf(hd->fp, "# %s info available, use -I to display\n",
2162 feat_ops[feat].name);
2167 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2169 struct header_print_data hd;
2170 struct perf_header *header = &session->header;
2171 int fd = session->fd;
2175 perf_header__process_sections(header, fd, &hd,
2176 perf_file_section__fprintf_info);
2180 static int do_write_feat(int fd, struct perf_header *h, int type,
2181 struct perf_file_section **p,
2182 struct perf_evlist *evlist)
2187 if (perf_header__has_feat(h, type)) {
2188 if (!feat_ops[type].write)
2191 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2193 err = feat_ops[type].write(fd, h, evlist);
2195 pr_debug("failed to write feature %d\n", type);
2197 /* undo anything written */
2198 lseek(fd, (*p)->offset, SEEK_SET);
2202 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2208 static int perf_header__adds_write(struct perf_header *header,
2209 struct perf_evlist *evlist, int fd)
2212 struct perf_file_section *feat_sec, *p;
2218 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2222 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2223 if (feat_sec == NULL)
2226 sec_size = sizeof(*feat_sec) * nr_sections;
2228 sec_start = header->data_offset + header->data_size;
2229 lseek(fd, sec_start + sec_size, SEEK_SET);
2231 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2232 if (do_write_feat(fd, header, feat, &p, evlist))
2233 perf_header__clear_feat(header, feat);
2236 lseek(fd, sec_start, SEEK_SET);
2238 * may write more than needed due to dropped feature, but
2239 * this is okay, reader will skip the mising entries
2241 err = do_write(fd, feat_sec, sec_size);
2243 pr_debug("failed to write feature section\n");
2248 int perf_header__write_pipe(int fd)
2250 struct perf_pipe_file_header f_header;
2253 f_header = (struct perf_pipe_file_header){
2254 .magic = PERF_MAGIC,
2255 .size = sizeof(f_header),
2258 err = do_write(fd, &f_header, sizeof(f_header));
2260 pr_debug("failed to write perf pipe header\n");
2267 int perf_session__write_header(struct perf_session *session,
2268 struct perf_evlist *evlist,
2269 int fd, bool at_exit)
2271 struct perf_file_header f_header;
2272 struct perf_file_attr f_attr;
2273 struct perf_header *header = &session->header;
2274 struct perf_evsel *evsel;
2278 lseek(fd, sizeof(f_header), SEEK_SET);
2280 list_for_each_entry(evsel, &evlist->entries, node) {
2281 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2282 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2284 pr_debug("failed to write perf header\n");
2289 attr_offset = lseek(fd, 0, SEEK_CUR);
2291 list_for_each_entry(evsel, &evlist->entries, node) {
2292 f_attr = (struct perf_file_attr){
2293 .attr = evsel->attr,
2295 .offset = evsel->id_offset,
2296 .size = evsel->ids * sizeof(u64),
2299 err = do_write(fd, &f_attr, sizeof(f_attr));
2301 pr_debug("failed to write perf header attribute\n");
2306 header->data_offset = lseek(fd, 0, SEEK_CUR);
2309 err = perf_header__adds_write(header, evlist, fd);
2314 f_header = (struct perf_file_header){
2315 .magic = PERF_MAGIC,
2316 .size = sizeof(f_header),
2317 .attr_size = sizeof(f_attr),
2319 .offset = attr_offset,
2320 .size = evlist->nr_entries * sizeof(f_attr),
2323 .offset = header->data_offset,
2324 .size = header->data_size,
2326 /* event_types is ignored, store zeros */
2329 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2331 lseek(fd, 0, SEEK_SET);
2332 err = do_write(fd, &f_header, sizeof(f_header));
2334 pr_debug("failed to write perf header\n");
2337 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2342 static int perf_header__getbuffer64(struct perf_header *header,
2343 int fd, void *buf, size_t size)
2345 if (readn(fd, buf, size) <= 0)
2348 if (header->needs_swap)
2349 mem_bswap_64(buf, size);
2354 int perf_header__process_sections(struct perf_header *header, int fd,
2356 int (*process)(struct perf_file_section *section,
2357 struct perf_header *ph,
2358 int feat, int fd, void *data))
2360 struct perf_file_section *feat_sec, *sec;
2366 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2370 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2374 sec_size = sizeof(*feat_sec) * nr_sections;
2376 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2378 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2382 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2383 err = process(sec++, header, feat, fd, data);
2393 static const int attr_file_abi_sizes[] = {
2394 [0] = PERF_ATTR_SIZE_VER0,
2395 [1] = PERF_ATTR_SIZE_VER1,
2396 [2] = PERF_ATTR_SIZE_VER2,
2397 [3] = PERF_ATTR_SIZE_VER3,
2402 * In the legacy file format, the magic number is not used to encode endianness.
2403 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2404 * on ABI revisions, we need to try all combinations for all endianness to
2405 * detect the endianness.
2407 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2409 uint64_t ref_size, attr_size;
2412 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2413 ref_size = attr_file_abi_sizes[i]
2414 + sizeof(struct perf_file_section);
2415 if (hdr_sz != ref_size) {
2416 attr_size = bswap_64(hdr_sz);
2417 if (attr_size != ref_size)
2420 ph->needs_swap = true;
2422 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2427 /* could not determine endianness */
2431 #define PERF_PIPE_HDR_VER0 16
2433 static const size_t attr_pipe_abi_sizes[] = {
2434 [0] = PERF_PIPE_HDR_VER0,
2439 * In the legacy pipe format, there is an implicit assumption that endiannesss
2440 * between host recording the samples, and host parsing the samples is the
2441 * same. This is not always the case given that the pipe output may always be
2442 * redirected into a file and analyzed on a different machine with possibly a
2443 * different endianness and perf_event ABI revsions in the perf tool itself.
2445 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2450 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2451 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2452 attr_size = bswap_64(hdr_sz);
2453 if (attr_size != hdr_sz)
2456 ph->needs_swap = true;
2458 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2464 bool is_perf_magic(u64 magic)
2466 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2467 || magic == __perf_magic2
2468 || magic == __perf_magic2_sw)
2474 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2475 bool is_pipe, struct perf_header *ph)
2479 /* check for legacy format */
2480 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2482 pr_debug("legacy perf.data format\n");
2484 return try_all_pipe_abis(hdr_sz, ph);
2486 return try_all_file_abis(hdr_sz, ph);
2489 * the new magic number serves two purposes:
2490 * - unique number to identify actual perf.data files
2491 * - encode endianness of file
2494 /* check magic number with one endianness */
2495 if (magic == __perf_magic2)
2498 /* check magic number with opposite endianness */
2499 if (magic != __perf_magic2_sw)
2502 ph->needs_swap = true;
2507 int perf_file_header__read(struct perf_file_header *header,
2508 struct perf_header *ph, int fd)
2512 lseek(fd, 0, SEEK_SET);
2514 ret = readn(fd, header, sizeof(*header));
2518 if (check_magic_endian(header->magic,
2519 header->attr_size, false, ph) < 0) {
2520 pr_debug("magic/endian check failed\n");
2524 if (ph->needs_swap) {
2525 mem_bswap_64(header, offsetof(struct perf_file_header,
2529 if (header->size != sizeof(*header)) {
2530 /* Support the previous format */
2531 if (header->size == offsetof(typeof(*header), adds_features))
2532 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2535 } else if (ph->needs_swap) {
2537 * feature bitmap is declared as an array of unsigned longs --
2538 * not good since its size can differ between the host that
2539 * generated the data file and the host analyzing the file.
2541 * We need to handle endianness, but we don't know the size of
2542 * the unsigned long where the file was generated. Take a best
2543 * guess at determining it: try 64-bit swap first (ie., file
2544 * created on a 64-bit host), and check if the hostname feature
2545 * bit is set (this feature bit is forced on as of fbe96f2).
2546 * If the bit is not, undo the 64-bit swap and try a 32-bit
2547 * swap. If the hostname bit is still not set (e.g., older data
2548 * file), punt and fallback to the original behavior --
2549 * clearing all feature bits and setting buildid.
2551 mem_bswap_64(&header->adds_features,
2552 BITS_TO_U64(HEADER_FEAT_BITS));
2554 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2556 mem_bswap_64(&header->adds_features,
2557 BITS_TO_U64(HEADER_FEAT_BITS));
2560 mem_bswap_32(&header->adds_features,
2561 BITS_TO_U32(HEADER_FEAT_BITS));
2564 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2565 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2566 set_bit(HEADER_BUILD_ID, header->adds_features);
2570 memcpy(&ph->adds_features, &header->adds_features,
2571 sizeof(ph->adds_features));
2573 ph->data_offset = header->data.offset;
2574 ph->data_size = header->data.size;
2578 static int perf_file_section__process(struct perf_file_section *section,
2579 struct perf_header *ph,
2580 int feat, int fd, void *data)
2582 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2583 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2584 "%d, continuing...\n", section->offset, feat);
2588 if (feat >= HEADER_LAST_FEATURE) {
2589 pr_debug("unknown feature %d, continuing...\n", feat);
2593 if (!feat_ops[feat].process)
2596 return feat_ops[feat].process(section, ph, fd, data);
2599 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2600 struct perf_header *ph, int fd,
2605 ret = readn(fd, header, sizeof(*header));
2609 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2610 pr_debug("endian/magic failed\n");
2615 header->size = bswap_64(header->size);
2617 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2623 static int perf_header__read_pipe(struct perf_session *session)
2625 struct perf_header *header = &session->header;
2626 struct perf_pipe_file_header f_header;
2628 if (perf_file_header__read_pipe(&f_header, header, session->fd,
2629 session->repipe) < 0) {
2630 pr_debug("incompatible file format\n");
2637 static int read_attr(int fd, struct perf_header *ph,
2638 struct perf_file_attr *f_attr)
2640 struct perf_event_attr *attr = &f_attr->attr;
2642 size_t our_sz = sizeof(f_attr->attr);
2645 memset(f_attr, 0, sizeof(*f_attr));
2647 /* read minimal guaranteed structure */
2648 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2650 pr_debug("cannot read %d bytes of header attr\n",
2651 PERF_ATTR_SIZE_VER0);
2655 /* on file perf_event_attr size */
2663 sz = PERF_ATTR_SIZE_VER0;
2664 } else if (sz > our_sz) {
2665 pr_debug("file uses a more recent and unsupported ABI"
2666 " (%zu bytes extra)\n", sz - our_sz);
2669 /* what we have not yet read and that we know about */
2670 left = sz - PERF_ATTR_SIZE_VER0;
2673 ptr += PERF_ATTR_SIZE_VER0;
2675 ret = readn(fd, ptr, left);
2677 /* read perf_file_section, ids are read in caller */
2678 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2680 return ret <= 0 ? -1 : 0;
2683 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2684 struct pevent *pevent)
2686 struct event_format *event;
2689 /* already prepared */
2690 if (evsel->tp_format)
2693 if (pevent == NULL) {
2694 pr_debug("broken or missing trace data\n");
2698 event = pevent_find_event(pevent, evsel->attr.config);
2703 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2704 evsel->name = strdup(bf);
2705 if (evsel->name == NULL)
2709 evsel->tp_format = event;
2713 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2714 struct pevent *pevent)
2716 struct perf_evsel *pos;
2718 list_for_each_entry(pos, &evlist->entries, node) {
2719 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2720 perf_evsel__prepare_tracepoint_event(pos, pevent))
2727 int perf_session__read_header(struct perf_session *session)
2729 struct perf_header *header = &session->header;
2730 struct perf_file_header f_header;
2731 struct perf_file_attr f_attr;
2733 int nr_attrs, nr_ids, i, j;
2734 int fd = session->fd;
2736 session->evlist = perf_evlist__new();
2737 if (session->evlist == NULL)
2740 if (session->fd_pipe)
2741 return perf_header__read_pipe(session);
2743 if (perf_file_header__read(&f_header, header, fd) < 0)
2746 nr_attrs = f_header.attrs.size / f_header.attr_size;
2747 lseek(fd, f_header.attrs.offset, SEEK_SET);
2749 for (i = 0; i < nr_attrs; i++) {
2750 struct perf_evsel *evsel;
2753 if (read_attr(fd, header, &f_attr) < 0)
2756 if (header->needs_swap)
2757 perf_event__attr_swap(&f_attr.attr);
2759 tmp = lseek(fd, 0, SEEK_CUR);
2760 evsel = perf_evsel__new(&f_attr.attr, i);
2763 goto out_delete_evlist;
2765 evsel->needs_swap = header->needs_swap;
2767 * Do it before so that if perf_evsel__alloc_id fails, this
2768 * entry gets purged too at perf_evlist__delete().
2770 perf_evlist__add(session->evlist, evsel);
2772 nr_ids = f_attr.ids.size / sizeof(u64);
2774 * We don't have the cpu and thread maps on the header, so
2775 * for allocating the perf_sample_id table we fake 1 cpu and
2776 * hattr->ids threads.
2778 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2779 goto out_delete_evlist;
2781 lseek(fd, f_attr.ids.offset, SEEK_SET);
2783 for (j = 0; j < nr_ids; j++) {
2784 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2787 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2790 lseek(fd, tmp, SEEK_SET);
2793 symbol_conf.nr_events = nr_attrs;
2795 perf_header__process_sections(header, fd, &session->pevent,
2796 perf_file_section__process);
2798 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2800 goto out_delete_evlist;
2807 perf_evlist__delete(session->evlist);
2808 session->evlist = NULL;
2812 int perf_event__synthesize_attr(struct perf_tool *tool,
2813 struct perf_event_attr *attr, u32 ids, u64 *id,
2814 perf_event__handler_t process)
2816 union perf_event *ev;
2820 size = sizeof(struct perf_event_attr);
2821 size = PERF_ALIGN(size, sizeof(u64));
2822 size += sizeof(struct perf_event_header);
2823 size += ids * sizeof(u64);
2830 ev->attr.attr = *attr;
2831 memcpy(ev->attr.id, id, ids * sizeof(u64));
2833 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2834 ev->attr.header.size = (u16)size;
2836 if (ev->attr.header.size == size)
2837 err = process(tool, ev, NULL, NULL);
2846 int perf_event__synthesize_attrs(struct perf_tool *tool,
2847 struct perf_session *session,
2848 perf_event__handler_t process)
2850 struct perf_evsel *evsel;
2853 list_for_each_entry(evsel, &session->evlist->entries, node) {
2854 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2855 evsel->id, process);
2857 pr_debug("failed to create perf header attribute\n");
2865 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2866 union perf_event *event,
2867 struct perf_evlist **pevlist)
2870 struct perf_evsel *evsel;
2871 struct perf_evlist *evlist = *pevlist;
2873 if (evlist == NULL) {
2874 *pevlist = evlist = perf_evlist__new();
2879 evsel = perf_evsel__new(&event->attr.attr, evlist->nr_entries);
2883 perf_evlist__add(evlist, evsel);
2885 ids = event->header.size;
2886 ids -= (void *)&event->attr.id - (void *)event;
2887 n_ids = ids / sizeof(u64);
2889 * We don't have the cpu and thread maps on the header, so
2890 * for allocating the perf_sample_id table we fake 1 cpu and
2891 * hattr->ids threads.
2893 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2896 for (i = 0; i < n_ids; i++) {
2897 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2900 symbol_conf.nr_events = evlist->nr_entries;
2905 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2906 struct perf_evlist *evlist,
2907 perf_event__handler_t process)
2909 union perf_event ev;
2910 struct tracing_data *tdata;
2911 ssize_t size = 0, aligned_size = 0, padding;
2912 int err __maybe_unused = 0;
2915 * We are going to store the size of the data followed
2916 * by the data contents. Since the fd descriptor is a pipe,
2917 * we cannot seek back to store the size of the data once
2918 * we know it. Instead we:
2920 * - write the tracing data to the temp file
2921 * - get/write the data size to pipe
2922 * - write the tracing data from the temp file
2925 tdata = tracing_data_get(&evlist->entries, fd, true);
2929 memset(&ev, 0, sizeof(ev));
2931 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2933 aligned_size = PERF_ALIGN(size, sizeof(u64));
2934 padding = aligned_size - size;
2935 ev.tracing_data.header.size = sizeof(ev.tracing_data);
2936 ev.tracing_data.size = aligned_size;
2938 process(tool, &ev, NULL, NULL);
2941 * The put function will copy all the tracing data
2942 * stored in temp file to the pipe.
2944 tracing_data_put(tdata);
2946 write_padded(fd, NULL, 0, padding);
2948 return aligned_size;
2951 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2952 union perf_event *event,
2953 struct perf_session *session)
2955 ssize_t size_read, padding, size = event->tracing_data.size;
2956 off_t offset = lseek(session->fd, 0, SEEK_CUR);
2959 /* setup for reading amidst mmap */
2960 lseek(session->fd, offset + sizeof(struct tracing_data_event),
2963 size_read = trace_report(session->fd, &session->pevent,
2965 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2967 if (readn(session->fd, buf, padding) < 0) {
2968 pr_err("%s: reading input file", __func__);
2971 if (session->repipe) {
2972 int retw = write(STDOUT_FILENO, buf, padding);
2973 if (retw <= 0 || retw != padding) {
2974 pr_err("%s: repiping tracing data padding", __func__);
2979 if (size_read + padding != size) {
2980 pr_err("%s: tracing data size mismatch", __func__);
2984 perf_evlist__prepare_tracepoint_events(session->evlist,
2987 return size_read + padding;
2990 int perf_event__synthesize_build_id(struct perf_tool *tool,
2991 struct dso *pos, u16 misc,
2992 perf_event__handler_t process,
2993 struct machine *machine)
2995 union perf_event ev;
3002 memset(&ev, 0, sizeof(ev));
3004 len = pos->long_name_len + 1;
3005 len = PERF_ALIGN(len, NAME_ALIGN);
3006 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3007 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3008 ev.build_id.header.misc = misc;
3009 ev.build_id.pid = machine->pid;
3010 ev.build_id.header.size = sizeof(ev.build_id) + len;
3011 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3013 err = process(tool, &ev, NULL, machine);
3018 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3019 union perf_event *event,
3020 struct perf_session *session)
3022 __event_process_build_id(&event->build_id,
3023 event->build_id.filename,
3028 void disable_buildid_cache(void)
3030 no_buildid_cache = true;