7 #include <linux/list.h>
8 #include <linux/kernel.h>
9 #include <linux/bitops.h>
10 #include <sys/utsname.h>
16 #include "trace-event.h"
27 static bool no_buildid_cache = false;
29 static u32 header_argc;
30 static const char **header_argv;
34 * must be a numerical value to let the endianness
35 * determine the memory layout. That way we are able
36 * to detect endianness when reading the perf.data file
39 * we check for legacy (PERFFILE) format.
41 static const char *__perf_magic1 = "PERFFILE";
42 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
43 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
45 #define PERF_MAGIC __perf_magic2
47 struct perf_file_attr {
48 struct perf_event_attr attr;
49 struct perf_file_section ids;
52 void perf_header__set_feat(struct perf_header *header, int feat)
54 set_bit(feat, header->adds_features);
57 void perf_header__clear_feat(struct perf_header *header, int feat)
59 clear_bit(feat, header->adds_features);
62 bool perf_header__has_feat(const struct perf_header *header, int feat)
64 return test_bit(feat, header->adds_features);
67 static int do_write(int fd, const void *buf, size_t size)
70 int ret = write(fd, buf, size);
84 static int write_padded(int fd, const void *bf, size_t count,
87 static const char zero_buf[NAME_ALIGN];
88 int err = do_write(fd, bf, count);
91 err = do_write(fd, zero_buf, count_aligned - count);
96 static int do_write_string(int fd, const char *str)
101 olen = strlen(str) + 1;
102 len = PERF_ALIGN(olen, NAME_ALIGN);
104 /* write len, incl. \0 */
105 ret = do_write(fd, &len, sizeof(len));
109 return write_padded(fd, str, olen, len);
112 static char *do_read_string(int fd, struct perf_header *ph)
118 sz = readn(fd, &len, sizeof(len));
119 if (sz < (ssize_t)sizeof(len))
129 ret = readn(fd, buf, len);
130 if (ret == (ssize_t)len) {
132 * strings are padded by zeroes
133 * thus the actual strlen of buf
134 * may be less than len
144 perf_header__set_cmdline(int argc, const char **argv)
149 * If header_argv has already been set, do not override it.
150 * This allows a command to set the cmdline, parse args and
151 * then call another builtin function that implements a
152 * command -- e.g, cmd_kvm calling cmd_record.
157 header_argc = (u32)argc;
159 /* do not include NULL termination */
160 header_argv = calloc(argc, sizeof(char *));
165 * must copy argv contents because it gets moved
166 * around during option parsing
168 for (i = 0; i < argc ; i++)
169 header_argv[i] = argv[i];
174 #define dsos__for_each_with_build_id(pos, head) \
175 list_for_each_entry(pos, head, node) \
176 if (!pos->has_build_id) \
180 static int write_buildid(char *name, size_t name_len, u8 *build_id,
181 pid_t pid, u16 misc, int fd)
184 struct build_id_event b;
188 len = PERF_ALIGN(len, NAME_ALIGN);
190 memset(&b, 0, sizeof(b));
191 memcpy(&b.build_id, build_id, BUILD_ID_SIZE);
193 b.header.misc = misc;
194 b.header.size = sizeof(b) + len;
196 err = do_write(fd, &b, sizeof(b));
200 return write_padded(fd, name, name_len + 1, len);
203 static int __dsos__write_buildid_table(struct list_head *head,
204 struct machine *machine,
205 pid_t pid, u16 misc, int fd)
210 dsos__for_each_with_build_id(pos, head) {
218 if (is_vdso_map(pos->short_name)) {
219 name = (char *) VDSO__MAP_NAME;
220 name_len = sizeof(VDSO__MAP_NAME) + 1;
221 } else if (dso__is_kcore(pos)) {
222 machine__mmap_name(machine, nm, sizeof(nm));
224 name_len = strlen(nm) + 1;
226 name = pos->long_name;
227 name_len = pos->long_name_len + 1;
230 err = write_buildid(name, name_len, pos->build_id,
239 static int machine__write_buildid_table(struct machine *machine, int fd)
242 u16 kmisc = PERF_RECORD_MISC_KERNEL,
243 umisc = PERF_RECORD_MISC_USER;
245 if (!machine__is_host(machine)) {
246 kmisc = PERF_RECORD_MISC_GUEST_KERNEL;
247 umisc = PERF_RECORD_MISC_GUEST_USER;
250 err = __dsos__write_buildid_table(&machine->kernel_dsos, machine,
251 machine->pid, kmisc, fd);
253 err = __dsos__write_buildid_table(&machine->user_dsos, machine,
254 machine->pid, umisc, fd);
258 static int dsos__write_buildid_table(struct perf_header *header, int fd)
260 struct perf_session *session = container_of(header,
261 struct perf_session, header);
263 int err = machine__write_buildid_table(&session->machines.host, fd);
268 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
269 struct machine *pos = rb_entry(nd, struct machine, rb_node);
270 err = machine__write_buildid_table(pos, fd);
277 int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
278 const char *name, bool is_kallsyms, bool is_vdso)
280 const size_t size = PATH_MAX;
281 char *realname, *filename = zalloc(size),
282 *linkname = zalloc(size), *targetname;
284 bool slash = is_kallsyms || is_vdso;
287 if (symbol_conf.kptr_restrict) {
288 pr_debug("Not caching a kptr_restrict'ed /proc/kallsyms\n");
292 realname = (char *) name;
294 realname = realpath(name, NULL);
296 if (realname == NULL || filename == NULL || linkname == NULL)
299 len = scnprintf(filename, size, "%s%s%s",
300 debugdir, slash ? "/" : "",
301 is_vdso ? VDSO__MAP_NAME : realname);
302 if (mkdir_p(filename, 0755))
305 snprintf(filename + len, size - len, "/%s", sbuild_id);
307 if (access(filename, F_OK)) {
309 if (copyfile("/proc/kallsyms", filename))
311 } else if (link(realname, filename) && copyfile(name, filename))
315 len = scnprintf(linkname, size, "%s/.build-id/%.2s",
316 debugdir, sbuild_id);
318 if (access(linkname, X_OK) && mkdir_p(linkname, 0755))
321 snprintf(linkname + len, size - len, "/%s", sbuild_id + 2);
322 targetname = filename + strlen(debugdir) - 5;
323 memcpy(targetname, "../..", 5);
325 if (symlink(targetname, linkname) == 0)
335 static int build_id_cache__add_b(const u8 *build_id, size_t build_id_size,
336 const char *name, const char *debugdir,
337 bool is_kallsyms, bool is_vdso)
339 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
341 build_id__sprintf(build_id, build_id_size, sbuild_id);
343 return build_id_cache__add_s(sbuild_id, debugdir, name,
344 is_kallsyms, is_vdso);
347 int build_id_cache__remove_s(const char *sbuild_id, const char *debugdir)
349 const size_t size = PATH_MAX;
350 char *filename = zalloc(size),
351 *linkname = zalloc(size);
354 if (filename == NULL || linkname == NULL)
357 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
358 debugdir, sbuild_id, sbuild_id + 2);
360 if (access(linkname, F_OK))
363 if (readlink(linkname, filename, size - 1) < 0)
366 if (unlink(linkname))
370 * Since the link is relative, we must make it absolute:
372 snprintf(linkname, size, "%s/.build-id/%.2s/%s",
373 debugdir, sbuild_id, filename);
375 if (unlink(linkname))
385 static int dso__cache_build_id(struct dso *dso, struct machine *machine,
386 const char *debugdir)
388 bool is_kallsyms = dso->kernel && dso->long_name[0] != '/';
389 bool is_vdso = is_vdso_map(dso->short_name);
390 char *name = dso->long_name;
393 if (dso__is_kcore(dso)) {
395 machine__mmap_name(machine, nm, sizeof(nm));
398 return build_id_cache__add_b(dso->build_id, sizeof(dso->build_id), name,
399 debugdir, is_kallsyms, is_vdso);
402 static int __dsos__cache_build_ids(struct list_head *head,
403 struct machine *machine, const char *debugdir)
408 dsos__for_each_with_build_id(pos, head)
409 if (dso__cache_build_id(pos, machine, debugdir))
415 static int machine__cache_build_ids(struct machine *machine, const char *debugdir)
417 int ret = __dsos__cache_build_ids(&machine->kernel_dsos, machine,
419 ret |= __dsos__cache_build_ids(&machine->user_dsos, machine, debugdir);
423 static int perf_session__cache_build_ids(struct perf_session *session)
427 char debugdir[PATH_MAX];
429 snprintf(debugdir, sizeof(debugdir), "%s", buildid_dir);
431 if (mkdir(debugdir, 0755) != 0 && errno != EEXIST)
434 ret = machine__cache_build_ids(&session->machines.host, debugdir);
436 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
437 struct machine *pos = rb_entry(nd, struct machine, rb_node);
438 ret |= machine__cache_build_ids(pos, debugdir);
443 static bool machine__read_build_ids(struct machine *machine, bool with_hits)
445 bool ret = __dsos__read_build_ids(&machine->kernel_dsos, with_hits);
446 ret |= __dsos__read_build_ids(&machine->user_dsos, with_hits);
450 static bool perf_session__read_build_ids(struct perf_session *session, bool with_hits)
453 bool ret = machine__read_build_ids(&session->machines.host, with_hits);
455 for (nd = rb_first(&session->machines.guests); nd; nd = rb_next(nd)) {
456 struct machine *pos = rb_entry(nd, struct machine, rb_node);
457 ret |= machine__read_build_ids(pos, with_hits);
463 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
464 struct perf_evlist *evlist)
466 return read_tracing_data(fd, &evlist->entries);
470 static int write_build_id(int fd, struct perf_header *h,
471 struct perf_evlist *evlist __maybe_unused)
473 struct perf_session *session;
476 session = container_of(h, struct perf_session, header);
478 if (!perf_session__read_build_ids(session, true))
481 err = dsos__write_buildid_table(h, fd);
483 pr_debug("failed to write buildid table\n");
486 if (!no_buildid_cache)
487 perf_session__cache_build_ids(session);
492 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
493 struct perf_evlist *evlist __maybe_unused)
502 return do_write_string(fd, uts.nodename);
505 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
506 struct perf_evlist *evlist __maybe_unused)
515 return do_write_string(fd, uts.release);
518 static int write_arch(int fd, struct perf_header *h __maybe_unused,
519 struct perf_evlist *evlist __maybe_unused)
528 return do_write_string(fd, uts.machine);
531 static int write_version(int fd, struct perf_header *h __maybe_unused,
532 struct perf_evlist *evlist __maybe_unused)
534 return do_write_string(fd, perf_version_string);
537 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
538 struct perf_evlist *evlist __maybe_unused)
541 #define CPUINFO_PROC NULL
546 const char *search = CPUINFO_PROC;
553 file = fopen("/proc/cpuinfo", "r");
557 while (getline(&buf, &len, file) > 0) {
558 ret = strncmp(buf, search, strlen(search));
568 p = strchr(buf, ':');
569 if (p && *(p+1) == ' ' && *(p+2))
575 /* squash extra space characters (branding string) */
582 while (*q && isspace(*q))
585 while ((*r++ = *q++));
589 ret = do_write_string(fd, s);
596 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
597 struct perf_evlist *evlist __maybe_unused)
603 nr = sysconf(_SC_NPROCESSORS_CONF);
607 nrc = (u32)(nr & UINT_MAX);
609 nr = sysconf(_SC_NPROCESSORS_ONLN);
613 nra = (u32)(nr & UINT_MAX);
615 ret = do_write(fd, &nrc, sizeof(nrc));
619 return do_write(fd, &nra, sizeof(nra));
622 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
623 struct perf_evlist *evlist)
625 struct perf_evsel *evsel;
629 nre = evlist->nr_entries;
632 * write number of events
634 ret = do_write(fd, &nre, sizeof(nre));
639 * size of perf_event_attr struct
641 sz = (u32)sizeof(evsel->attr);
642 ret = do_write(fd, &sz, sizeof(sz));
646 list_for_each_entry(evsel, &evlist->entries, node) {
648 ret = do_write(fd, &evsel->attr, sz);
652 * write number of unique id per event
653 * there is one id per instance of an event
655 * copy into an nri to be independent of the
659 ret = do_write(fd, &nri, sizeof(nri));
664 * write event string as passed on cmdline
666 ret = do_write_string(fd, perf_evsel__name(evsel));
670 * write unique ids for this event
672 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
679 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
680 struct perf_evlist *evlist __maybe_unused)
682 char buf[MAXPATHLEN];
688 * actual atual path to perf binary
690 sprintf(proc, "/proc/%d/exe", getpid());
691 ret = readlink(proc, buf, sizeof(buf));
695 /* readlink() does not add null termination */
698 /* account for binary path */
701 ret = do_write(fd, &n, sizeof(n));
705 ret = do_write_string(fd, buf);
709 for (i = 0 ; i < header_argc; i++) {
710 ret = do_write_string(fd, header_argv[i]);
717 #define CORE_SIB_FMT \
718 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
719 #define THRD_SIB_FMT \
720 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
725 char **core_siblings;
726 char **thread_siblings;
729 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
732 char filename[MAXPATHLEN];
733 char *buf = NULL, *p;
739 sprintf(filename, CORE_SIB_FMT, cpu);
740 fp = fopen(filename, "r");
744 sret = getline(&buf, &len, fp);
749 p = strchr(buf, '\n');
753 for (i = 0; i < tp->core_sib; i++) {
754 if (!strcmp(buf, tp->core_siblings[i]))
757 if (i == tp->core_sib) {
758 tp->core_siblings[i] = buf;
766 sprintf(filename, THRD_SIB_FMT, cpu);
767 fp = fopen(filename, "r");
771 if (getline(&buf, &len, fp) <= 0)
774 p = strchr(buf, '\n');
778 for (i = 0; i < tp->thread_sib; i++) {
779 if (!strcmp(buf, tp->thread_siblings[i]))
782 if (i == tp->thread_sib) {
783 tp->thread_siblings[i] = buf;
795 static void free_cpu_topo(struct cpu_topo *tp)
802 for (i = 0 ; i < tp->core_sib; i++)
803 free(tp->core_siblings[i]);
805 for (i = 0 ; i < tp->thread_sib; i++)
806 free(tp->thread_siblings[i]);
811 static struct cpu_topo *build_cpu_topology(void)
820 ncpus = sysconf(_SC_NPROCESSORS_CONF);
824 nr = (u32)(ncpus & UINT_MAX);
826 sz = nr * sizeof(char *);
828 addr = calloc(1, sizeof(*tp) + 2 * sz);
835 tp->core_siblings = addr;
837 tp->thread_siblings = addr;
839 for (i = 0; i < nr; i++) {
840 ret = build_cpu_topo(tp, i);
851 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
852 struct perf_evlist *evlist __maybe_unused)
858 tp = build_cpu_topology();
862 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
866 for (i = 0; i < tp->core_sib; i++) {
867 ret = do_write_string(fd, tp->core_siblings[i]);
871 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
875 for (i = 0; i < tp->thread_sib; i++) {
876 ret = do_write_string(fd, tp->thread_siblings[i]);
887 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
888 struct perf_evlist *evlist __maybe_unused)
896 fp = fopen("/proc/meminfo", "r");
900 while (getline(&buf, &len, fp) > 0) {
901 ret = strncmp(buf, "MemTotal:", 9);
906 n = sscanf(buf, "%*s %"PRIu64, &mem);
908 ret = do_write(fd, &mem, sizeof(mem));
915 static int write_topo_node(int fd, int node)
917 char str[MAXPATHLEN];
919 char *buf = NULL, *p;
922 u64 mem_total, mem_free, mem;
925 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
926 fp = fopen(str, "r");
930 while (getline(&buf, &len, fp) > 0) {
931 /* skip over invalid lines */
932 if (!strchr(buf, ':'))
934 if (sscanf(buf, "%*s %*d %s %"PRIu64, field, &mem) != 2)
936 if (!strcmp(field, "MemTotal:"))
938 if (!strcmp(field, "MemFree:"))
945 ret = do_write(fd, &mem_total, sizeof(u64));
949 ret = do_write(fd, &mem_free, sizeof(u64));
954 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
956 fp = fopen(str, "r");
960 if (getline(&buf, &len, fp) <= 0)
963 p = strchr(buf, '\n');
967 ret = do_write_string(fd, buf);
975 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
976 struct perf_evlist *evlist __maybe_unused)
981 struct cpu_map *node_map = NULL;
986 fp = fopen("/sys/devices/system/node/online", "r");
990 if (getline(&buf, &len, fp) <= 0)
993 c = strchr(buf, '\n');
997 node_map = cpu_map__new(buf);
1001 nr = (u32)node_map->nr;
1003 ret = do_write(fd, &nr, sizeof(nr));
1007 for (i = 0; i < nr; i++) {
1008 j = (u32)node_map->map[i];
1009 ret = do_write(fd, &j, sizeof(j));
1013 ret = write_topo_node(fd, i);
1027 * struct pmu_mappings {
1036 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
1037 struct perf_evlist *evlist __maybe_unused)
1039 struct perf_pmu *pmu = NULL;
1040 off_t offset = lseek(fd, 0, SEEK_CUR);
1044 /* write real pmu_num later */
1045 ret = do_write(fd, &pmu_num, sizeof(pmu_num));
1049 while ((pmu = perf_pmu__scan(pmu))) {
1054 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
1058 ret = do_write_string(fd, pmu->name);
1063 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
1065 lseek(fd, offset, SEEK_SET);
1075 * struct group_descs {
1077 * struct group_desc {
1084 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
1085 struct perf_evlist *evlist)
1087 u32 nr_groups = evlist->nr_groups;
1088 struct perf_evsel *evsel;
1091 ret = do_write(fd, &nr_groups, sizeof(nr_groups));
1095 list_for_each_entry(evsel, &evlist->entries, node) {
1096 if (perf_evsel__is_group_leader(evsel) &&
1097 evsel->nr_members > 1) {
1098 const char *name = evsel->group_name ?: "{anon_group}";
1099 u32 leader_idx = evsel->idx;
1100 u32 nr_members = evsel->nr_members;
1102 ret = do_write_string(fd, name);
1106 ret = do_write(fd, &leader_idx, sizeof(leader_idx));
1110 ret = do_write(fd, &nr_members, sizeof(nr_members));
1119 * default get_cpuid(): nothing gets recorded
1120 * actual implementation must be in arch/$(ARCH)/util/header.c
1122 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
1123 size_t sz __maybe_unused)
1128 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
1129 struct perf_evlist *evlist __maybe_unused)
1134 ret = get_cpuid(buffer, sizeof(buffer));
1140 return do_write_string(fd, buffer);
1143 static int write_branch_stack(int fd __maybe_unused,
1144 struct perf_header *h __maybe_unused,
1145 struct perf_evlist *evlist __maybe_unused)
1150 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
1153 fprintf(fp, "# hostname : %s\n", ph->env.hostname);
1156 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
1159 fprintf(fp, "# os release : %s\n", ph->env.os_release);
1162 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1164 fprintf(fp, "# arch : %s\n", ph->env.arch);
1167 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
1170 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
1173 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
1176 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
1177 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
1180 static void print_version(struct perf_header *ph, int fd __maybe_unused,
1183 fprintf(fp, "# perf version : %s\n", ph->env.version);
1186 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
1192 nr = ph->env.nr_cmdline;
1193 str = ph->env.cmdline;
1195 fprintf(fp, "# cmdline : ");
1197 for (i = 0; i < nr; i++) {
1198 fprintf(fp, "%s ", str);
1199 str += strlen(str) + 1;
1204 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
1210 nr = ph->env.nr_sibling_cores;
1211 str = ph->env.sibling_cores;
1213 for (i = 0; i < nr; i++) {
1214 fprintf(fp, "# sibling cores : %s\n", str);
1215 str += strlen(str) + 1;
1218 nr = ph->env.nr_sibling_threads;
1219 str = ph->env.sibling_threads;
1221 for (i = 0; i < nr; i++) {
1222 fprintf(fp, "# sibling threads : %s\n", str);
1223 str += strlen(str) + 1;
1227 static void free_event_desc(struct perf_evsel *events)
1229 struct perf_evsel *evsel;
1234 for (evsel = events; evsel->attr.size; evsel++) {
1244 static struct perf_evsel *
1245 read_event_desc(struct perf_header *ph, int fd)
1247 struct perf_evsel *evsel, *events = NULL;
1250 u32 nre, sz, nr, i, j;
1254 /* number of events */
1255 ret = readn(fd, &nre, sizeof(nre));
1256 if (ret != (ssize_t)sizeof(nre))
1260 nre = bswap_32(nre);
1262 ret = readn(fd, &sz, sizeof(sz));
1263 if (ret != (ssize_t)sizeof(sz))
1269 /* buffer to hold on file attr struct */
1274 /* the last event terminates with evsel->attr.size == 0: */
1275 events = calloc(nre + 1, sizeof(*events));
1279 msz = sizeof(evsel->attr);
1283 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1287 * must read entire on-file attr struct to
1288 * sync up with layout.
1290 ret = readn(fd, buf, sz);
1291 if (ret != (ssize_t)sz)
1295 perf_event__attr_swap(buf);
1297 memcpy(&evsel->attr, buf, msz);
1299 ret = readn(fd, &nr, sizeof(nr));
1300 if (ret != (ssize_t)sizeof(nr))
1303 if (ph->needs_swap) {
1305 evsel->needs_swap = true;
1308 evsel->name = do_read_string(fd, ph);
1313 id = calloc(nr, sizeof(*id));
1319 for (j = 0 ; j < nr; j++) {
1320 ret = readn(fd, id, sizeof(*id));
1321 if (ret != (ssize_t)sizeof(*id))
1324 *id = bswap_64(*id);
1334 free_event_desc(events);
1339 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1341 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1346 fprintf(fp, "# event desc: not available or unable to read\n");
1350 for (evsel = events; evsel->attr.size; evsel++) {
1351 fprintf(fp, "# event : name = %s, ", evsel->name);
1353 fprintf(fp, "type = %d, config = 0x%"PRIx64
1354 ", config1 = 0x%"PRIx64", config2 = 0x%"PRIx64,
1356 (u64)evsel->attr.config,
1357 (u64)evsel->attr.config1,
1358 (u64)evsel->attr.config2);
1360 fprintf(fp, ", excl_usr = %d, excl_kern = %d",
1361 evsel->attr.exclude_user,
1362 evsel->attr.exclude_kernel);
1364 fprintf(fp, ", excl_host = %d, excl_guest = %d",
1365 evsel->attr.exclude_host,
1366 evsel->attr.exclude_guest);
1368 fprintf(fp, ", precise_ip = %d", evsel->attr.precise_ip);
1370 fprintf(fp, ", attr_mmap2 = %d", evsel->attr.mmap2);
1371 fprintf(fp, ", attr_mmap = %d", evsel->attr.mmap);
1372 fprintf(fp, ", attr_mmap_data = %d", evsel->attr.mmap_data);
1374 fprintf(fp, ", id = {");
1375 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1378 fprintf(fp, " %"PRIu64, *id);
1386 free_event_desc(events);
1389 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1392 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1395 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1400 uint64_t mem_total, mem_free;
1403 nr = ph->env.nr_numa_nodes;
1404 str = ph->env.numa_nodes;
1406 for (i = 0; i < nr; i++) {
1408 c = strtoul(str, &tmp, 0);
1413 mem_total = strtoull(str, &tmp, 0);
1418 mem_free = strtoull(str, &tmp, 0);
1422 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1423 " free = %"PRIu64" kB\n",
1424 c, mem_total, mem_free);
1427 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1429 str += strlen(str) + 1;
1433 fprintf(fp, "# numa topology : not available\n");
1436 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1438 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1441 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1442 int fd __maybe_unused, FILE *fp)
1444 fprintf(fp, "# contains samples with branch stack\n");
1447 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1450 const char *delimiter = "# pmu mappings: ";
1455 pmu_num = ph->env.nr_pmu_mappings;
1457 fprintf(fp, "# pmu mappings: not available\n");
1461 str = ph->env.pmu_mappings;
1464 type = strtoul(str, &tmp, 0);
1469 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1472 str += strlen(str) + 1;
1481 fprintf(fp, "# pmu mappings: unable to read\n");
1484 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1487 struct perf_session *session;
1488 struct perf_evsel *evsel;
1491 session = container_of(ph, struct perf_session, header);
1493 list_for_each_entry(evsel, &session->evlist->entries, node) {
1494 if (perf_evsel__is_group_leader(evsel) &&
1495 evsel->nr_members > 1) {
1496 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1497 perf_evsel__name(evsel));
1499 nr = evsel->nr_members - 1;
1501 fprintf(fp, ",%s", perf_evsel__name(evsel));
1509 static int __event_process_build_id(struct build_id_event *bev,
1511 struct perf_session *session)
1514 struct list_head *head;
1515 struct machine *machine;
1518 enum dso_kernel_type dso_type;
1520 machine = perf_session__findnew_machine(session, bev->pid);
1524 misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1527 case PERF_RECORD_MISC_KERNEL:
1528 dso_type = DSO_TYPE_KERNEL;
1529 head = &machine->kernel_dsos;
1531 case PERF_RECORD_MISC_GUEST_KERNEL:
1532 dso_type = DSO_TYPE_GUEST_KERNEL;
1533 head = &machine->kernel_dsos;
1535 case PERF_RECORD_MISC_USER:
1536 case PERF_RECORD_MISC_GUEST_USER:
1537 dso_type = DSO_TYPE_USER;
1538 head = &machine->user_dsos;
1544 dso = __dsos__findnew(head, filename);
1546 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1548 dso__set_build_id(dso, &bev->build_id);
1550 if (filename[0] == '[')
1551 dso->kernel = dso_type;
1553 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1555 pr_debug("build id event received for %s: %s\n",
1556 dso->long_name, sbuild_id);
1564 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1565 int input, u64 offset, u64 size)
1567 struct perf_session *session = container_of(header, struct perf_session, header);
1569 struct perf_event_header header;
1570 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1573 struct build_id_event bev;
1574 char filename[PATH_MAX];
1575 u64 limit = offset + size;
1577 while (offset < limit) {
1580 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1583 if (header->needs_swap)
1584 perf_event_header__bswap(&old_bev.header);
1586 len = old_bev.header.size - sizeof(old_bev);
1587 if (readn(input, filename, len) != len)
1590 bev.header = old_bev.header;
1593 * As the pid is the missing value, we need to fill
1594 * it properly. The header.misc value give us nice hint.
1596 bev.pid = HOST_KERNEL_ID;
1597 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1598 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1599 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1601 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1602 __event_process_build_id(&bev, filename, session);
1604 offset += bev.header.size;
1610 static int perf_header__read_build_ids(struct perf_header *header,
1611 int input, u64 offset, u64 size)
1613 struct perf_session *session = container_of(header, struct perf_session, header);
1614 struct build_id_event bev;
1615 char filename[PATH_MAX];
1616 u64 limit = offset + size, orig_offset = offset;
1619 while (offset < limit) {
1622 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1625 if (header->needs_swap)
1626 perf_event_header__bswap(&bev.header);
1628 len = bev.header.size - sizeof(bev);
1629 if (readn(input, filename, len) != len)
1632 * The a1645ce1 changeset:
1634 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1636 * Added a field to struct build_id_event that broke the file
1639 * Since the kernel build-id is the first entry, process the
1640 * table using the old format if the well known
1641 * '[kernel.kallsyms]' string for the kernel build-id has the
1642 * first 4 characters chopped off (where the pid_t sits).
1644 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1645 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1647 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1650 __event_process_build_id(&bev, filename, session);
1652 offset += bev.header.size;
1659 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1660 struct perf_header *ph __maybe_unused,
1663 ssize_t ret = trace_report(fd, data, false);
1664 return ret < 0 ? -1 : 0;
1667 static int process_build_id(struct perf_file_section *section,
1668 struct perf_header *ph, int fd,
1669 void *data __maybe_unused)
1671 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1672 pr_debug("Failed to read buildids, continuing...\n");
1676 static int process_hostname(struct perf_file_section *section __maybe_unused,
1677 struct perf_header *ph, int fd,
1678 void *data __maybe_unused)
1680 ph->env.hostname = do_read_string(fd, ph);
1681 return ph->env.hostname ? 0 : -ENOMEM;
1684 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1685 struct perf_header *ph, int fd,
1686 void *data __maybe_unused)
1688 ph->env.os_release = do_read_string(fd, ph);
1689 return ph->env.os_release ? 0 : -ENOMEM;
1692 static int process_version(struct perf_file_section *section __maybe_unused,
1693 struct perf_header *ph, int fd,
1694 void *data __maybe_unused)
1696 ph->env.version = do_read_string(fd, ph);
1697 return ph->env.version ? 0 : -ENOMEM;
1700 static int process_arch(struct perf_file_section *section __maybe_unused,
1701 struct perf_header *ph, int fd,
1702 void *data __maybe_unused)
1704 ph->env.arch = do_read_string(fd, ph);
1705 return ph->env.arch ? 0 : -ENOMEM;
1708 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1709 struct perf_header *ph, int fd,
1710 void *data __maybe_unused)
1715 ret = readn(fd, &nr, sizeof(nr));
1716 if (ret != sizeof(nr))
1722 ph->env.nr_cpus_online = nr;
1724 ret = readn(fd, &nr, sizeof(nr));
1725 if (ret != sizeof(nr))
1731 ph->env.nr_cpus_avail = nr;
1735 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1736 struct perf_header *ph, int fd,
1737 void *data __maybe_unused)
1739 ph->env.cpu_desc = do_read_string(fd, ph);
1740 return ph->env.cpu_desc ? 0 : -ENOMEM;
1743 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1744 struct perf_header *ph, int fd,
1745 void *data __maybe_unused)
1747 ph->env.cpuid = do_read_string(fd, ph);
1748 return ph->env.cpuid ? 0 : -ENOMEM;
1751 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1752 struct perf_header *ph, int fd,
1753 void *data __maybe_unused)
1758 ret = readn(fd, &mem, sizeof(mem));
1759 if (ret != sizeof(mem))
1763 mem = bswap_64(mem);
1765 ph->env.total_mem = mem;
1769 static struct perf_evsel *
1770 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1772 struct perf_evsel *evsel;
1774 list_for_each_entry(evsel, &evlist->entries, node) {
1775 if (evsel->idx == idx)
1783 perf_evlist__set_event_name(struct perf_evlist *evlist,
1784 struct perf_evsel *event)
1786 struct perf_evsel *evsel;
1791 evsel = perf_evlist__find_by_index(evlist, event->idx);
1798 evsel->name = strdup(event->name);
1802 process_event_desc(struct perf_file_section *section __maybe_unused,
1803 struct perf_header *header, int fd,
1804 void *data __maybe_unused)
1806 struct perf_session *session;
1807 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1812 session = container_of(header, struct perf_session, header);
1813 for (evsel = events; evsel->attr.size; evsel++)
1814 perf_evlist__set_event_name(session->evlist, evsel);
1816 free_event_desc(events);
1821 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1822 struct perf_header *ph, int fd,
1823 void *data __maybe_unused)
1830 ret = readn(fd, &nr, sizeof(nr));
1831 if (ret != sizeof(nr))
1837 ph->env.nr_cmdline = nr;
1838 strbuf_init(&sb, 128);
1840 for (i = 0; i < nr; i++) {
1841 str = do_read_string(fd, ph);
1845 /* include a NULL character at the end */
1846 strbuf_add(&sb, str, strlen(str) + 1);
1849 ph->env.cmdline = strbuf_detach(&sb, NULL);
1853 strbuf_release(&sb);
1857 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1858 struct perf_header *ph, int fd,
1859 void *data __maybe_unused)
1866 ret = readn(fd, &nr, sizeof(nr));
1867 if (ret != sizeof(nr))
1873 ph->env.nr_sibling_cores = nr;
1874 strbuf_init(&sb, 128);
1876 for (i = 0; i < nr; i++) {
1877 str = do_read_string(fd, ph);
1881 /* include a NULL character at the end */
1882 strbuf_add(&sb, str, strlen(str) + 1);
1885 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1887 ret = readn(fd, &nr, sizeof(nr));
1888 if (ret != sizeof(nr))
1894 ph->env.nr_sibling_threads = nr;
1896 for (i = 0; i < nr; i++) {
1897 str = do_read_string(fd, ph);
1901 /* include a NULL character at the end */
1902 strbuf_add(&sb, str, strlen(str) + 1);
1905 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1909 strbuf_release(&sb);
1913 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1914 struct perf_header *ph, int fd,
1915 void *data __maybe_unused)
1920 uint64_t mem_total, mem_free;
1924 ret = readn(fd, &nr, sizeof(nr));
1925 if (ret != sizeof(nr))
1931 ph->env.nr_numa_nodes = nr;
1932 strbuf_init(&sb, 256);
1934 for (i = 0; i < nr; i++) {
1936 ret = readn(fd, &node, sizeof(node));
1937 if (ret != sizeof(node))
1940 ret = readn(fd, &mem_total, sizeof(u64));
1941 if (ret != sizeof(u64))
1944 ret = readn(fd, &mem_free, sizeof(u64));
1945 if (ret != sizeof(u64))
1948 if (ph->needs_swap) {
1949 node = bswap_32(node);
1950 mem_total = bswap_64(mem_total);
1951 mem_free = bswap_64(mem_free);
1954 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1955 node, mem_total, mem_free);
1957 str = do_read_string(fd, ph);
1961 /* include a NULL character at the end */
1962 strbuf_add(&sb, str, strlen(str) + 1);
1965 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1969 strbuf_release(&sb);
1973 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1974 struct perf_header *ph, int fd,
1975 void *data __maybe_unused)
1983 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1984 if (ret != sizeof(pmu_num))
1988 pmu_num = bswap_32(pmu_num);
1991 pr_debug("pmu mappings not available\n");
1995 ph->env.nr_pmu_mappings = pmu_num;
1996 strbuf_init(&sb, 128);
1999 if (readn(fd, &type, sizeof(type)) != sizeof(type))
2002 type = bswap_32(type);
2004 name = do_read_string(fd, ph);
2008 strbuf_addf(&sb, "%u:%s", type, name);
2009 /* include a NULL character at the end */
2010 strbuf_add(&sb, "", 1);
2015 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2019 strbuf_release(&sb);
2023 static int process_group_desc(struct perf_file_section *section __maybe_unused,
2024 struct perf_header *ph, int fd,
2025 void *data __maybe_unused)
2028 u32 i, nr, nr_groups;
2029 struct perf_session *session;
2030 struct perf_evsel *evsel, *leader = NULL;
2037 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
2041 nr_groups = bswap_32(nr_groups);
2043 ph->env.nr_groups = nr_groups;
2045 pr_debug("group desc not available\n");
2049 desc = calloc(nr_groups, sizeof(*desc));
2053 for (i = 0; i < nr_groups; i++) {
2054 desc[i].name = do_read_string(fd, ph);
2058 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
2061 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
2064 if (ph->needs_swap) {
2065 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
2066 desc[i].nr_members = bswap_32(desc[i].nr_members);
2071 * Rebuild group relationship based on the group_desc
2073 session = container_of(ph, struct perf_session, header);
2074 session->evlist->nr_groups = nr_groups;
2077 list_for_each_entry(evsel, &session->evlist->entries, node) {
2078 if (evsel->idx == (int) desc[i].leader_idx) {
2079 evsel->leader = evsel;
2080 /* {anon_group} is a dummy name */
2081 if (strcmp(desc[i].name, "{anon_group}"))
2082 evsel->group_name = desc[i].name;
2083 evsel->nr_members = desc[i].nr_members;
2085 if (i >= nr_groups || nr > 0) {
2086 pr_debug("invalid group desc\n");
2091 nr = evsel->nr_members - 1;
2094 /* This is a group member */
2095 evsel->leader = leader;
2101 if (i != nr_groups || nr != 0) {
2102 pr_debug("invalid group desc\n");
2108 while ((int) --i >= 0)
2115 struct feature_ops {
2116 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
2117 void (*print)(struct perf_header *h, int fd, FILE *fp);
2118 int (*process)(struct perf_file_section *section,
2119 struct perf_header *h, int fd, void *data);
2124 #define FEAT_OPA(n, func) \
2125 [n] = { .name = #n, .write = write_##func, .print = print_##func }
2126 #define FEAT_OPP(n, func) \
2127 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2128 .process = process_##func }
2129 #define FEAT_OPF(n, func) \
2130 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
2131 .process = process_##func, .full_only = true }
2133 /* feature_ops not implemented: */
2134 #define print_tracing_data NULL
2135 #define print_build_id NULL
2137 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2138 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
2139 FEAT_OPP(HEADER_BUILD_ID, build_id),
2140 FEAT_OPP(HEADER_HOSTNAME, hostname),
2141 FEAT_OPP(HEADER_OSRELEASE, osrelease),
2142 FEAT_OPP(HEADER_VERSION, version),
2143 FEAT_OPP(HEADER_ARCH, arch),
2144 FEAT_OPP(HEADER_NRCPUS, nrcpus),
2145 FEAT_OPP(HEADER_CPUDESC, cpudesc),
2146 FEAT_OPP(HEADER_CPUID, cpuid),
2147 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
2148 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
2149 FEAT_OPP(HEADER_CMDLINE, cmdline),
2150 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
2151 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
2152 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
2153 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
2154 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
2157 struct header_print_data {
2159 bool full; /* extended list of headers */
2162 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2163 struct perf_header *ph,
2164 int feat, int fd, void *data)
2166 struct header_print_data *hd = data;
2168 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2169 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2170 "%d, continuing...\n", section->offset, feat);
2173 if (feat >= HEADER_LAST_FEATURE) {
2174 pr_warning("unknown feature %d\n", feat);
2177 if (!feat_ops[feat].print)
2180 if (!feat_ops[feat].full_only || hd->full)
2181 feat_ops[feat].print(ph, fd, hd->fp);
2183 fprintf(hd->fp, "# %s info available, use -I to display\n",
2184 feat_ops[feat].name);
2189 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2191 struct header_print_data hd;
2192 struct perf_header *header = &session->header;
2193 int fd = perf_data_file__fd(session->file);
2197 perf_header__process_sections(header, fd, &hd,
2198 perf_file_section__fprintf_info);
2202 static int do_write_feat(int fd, struct perf_header *h, int type,
2203 struct perf_file_section **p,
2204 struct perf_evlist *evlist)
2209 if (perf_header__has_feat(h, type)) {
2210 if (!feat_ops[type].write)
2213 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2215 err = feat_ops[type].write(fd, h, evlist);
2217 pr_debug("failed to write feature %d\n", type);
2219 /* undo anything written */
2220 lseek(fd, (*p)->offset, SEEK_SET);
2224 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2230 static int perf_header__adds_write(struct perf_header *header,
2231 struct perf_evlist *evlist, int fd)
2234 struct perf_file_section *feat_sec, *p;
2240 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2244 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2245 if (feat_sec == NULL)
2248 sec_size = sizeof(*feat_sec) * nr_sections;
2250 sec_start = header->feat_offset;
2251 lseek(fd, sec_start + sec_size, SEEK_SET);
2253 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2254 if (do_write_feat(fd, header, feat, &p, evlist))
2255 perf_header__clear_feat(header, feat);
2258 lseek(fd, sec_start, SEEK_SET);
2260 * may write more than needed due to dropped feature, but
2261 * this is okay, reader will skip the mising entries
2263 err = do_write(fd, feat_sec, sec_size);
2265 pr_debug("failed to write feature section\n");
2270 int perf_header__write_pipe(int fd)
2272 struct perf_pipe_file_header f_header;
2275 f_header = (struct perf_pipe_file_header){
2276 .magic = PERF_MAGIC,
2277 .size = sizeof(f_header),
2280 err = do_write(fd, &f_header, sizeof(f_header));
2282 pr_debug("failed to write perf pipe header\n");
2289 int perf_session__write_header(struct perf_session *session,
2290 struct perf_evlist *evlist,
2291 int fd, bool at_exit)
2293 struct perf_file_header f_header;
2294 struct perf_file_attr f_attr;
2295 struct perf_header *header = &session->header;
2296 struct perf_evsel *evsel;
2300 lseek(fd, sizeof(f_header), SEEK_SET);
2302 list_for_each_entry(evsel, &evlist->entries, node) {
2303 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2304 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2306 pr_debug("failed to write perf header\n");
2311 attr_offset = lseek(fd, 0, SEEK_CUR);
2313 list_for_each_entry(evsel, &evlist->entries, node) {
2314 f_attr = (struct perf_file_attr){
2315 .attr = evsel->attr,
2317 .offset = evsel->id_offset,
2318 .size = evsel->ids * sizeof(u64),
2321 err = do_write(fd, &f_attr, sizeof(f_attr));
2323 pr_debug("failed to write perf header attribute\n");
2328 header->data_offset = lseek(fd, 0, SEEK_CUR);
2329 header->feat_offset = header->data_offset + header->data_size;
2332 err = perf_header__adds_write(header, evlist, fd);
2337 f_header = (struct perf_file_header){
2338 .magic = PERF_MAGIC,
2339 .size = sizeof(f_header),
2340 .attr_size = sizeof(f_attr),
2342 .offset = attr_offset,
2343 .size = evlist->nr_entries * sizeof(f_attr),
2346 .offset = header->data_offset,
2347 .size = header->data_size,
2349 /* event_types is ignored, store zeros */
2352 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2354 lseek(fd, 0, SEEK_SET);
2355 err = do_write(fd, &f_header, sizeof(f_header));
2357 pr_debug("failed to write perf header\n");
2360 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2365 static int perf_header__getbuffer64(struct perf_header *header,
2366 int fd, void *buf, size_t size)
2368 if (readn(fd, buf, size) <= 0)
2371 if (header->needs_swap)
2372 mem_bswap_64(buf, size);
2377 int perf_header__process_sections(struct perf_header *header, int fd,
2379 int (*process)(struct perf_file_section *section,
2380 struct perf_header *ph,
2381 int feat, int fd, void *data))
2383 struct perf_file_section *feat_sec, *sec;
2389 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2393 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2397 sec_size = sizeof(*feat_sec) * nr_sections;
2399 lseek(fd, header->feat_offset, SEEK_SET);
2401 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2405 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2406 err = process(sec++, header, feat, fd, data);
2416 static const int attr_file_abi_sizes[] = {
2417 [0] = PERF_ATTR_SIZE_VER0,
2418 [1] = PERF_ATTR_SIZE_VER1,
2419 [2] = PERF_ATTR_SIZE_VER2,
2420 [3] = PERF_ATTR_SIZE_VER3,
2425 * In the legacy file format, the magic number is not used to encode endianness.
2426 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2427 * on ABI revisions, we need to try all combinations for all endianness to
2428 * detect the endianness.
2430 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2432 uint64_t ref_size, attr_size;
2435 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2436 ref_size = attr_file_abi_sizes[i]
2437 + sizeof(struct perf_file_section);
2438 if (hdr_sz != ref_size) {
2439 attr_size = bswap_64(hdr_sz);
2440 if (attr_size != ref_size)
2443 ph->needs_swap = true;
2445 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2450 /* could not determine endianness */
2454 #define PERF_PIPE_HDR_VER0 16
2456 static const size_t attr_pipe_abi_sizes[] = {
2457 [0] = PERF_PIPE_HDR_VER0,
2462 * In the legacy pipe format, there is an implicit assumption that endiannesss
2463 * between host recording the samples, and host parsing the samples is the
2464 * same. This is not always the case given that the pipe output may always be
2465 * redirected into a file and analyzed on a different machine with possibly a
2466 * different endianness and perf_event ABI revsions in the perf tool itself.
2468 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2473 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2474 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2475 attr_size = bswap_64(hdr_sz);
2476 if (attr_size != hdr_sz)
2479 ph->needs_swap = true;
2481 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2487 bool is_perf_magic(u64 magic)
2489 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2490 || magic == __perf_magic2
2491 || magic == __perf_magic2_sw)
2497 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2498 bool is_pipe, struct perf_header *ph)
2502 /* check for legacy format */
2503 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2505 ph->version = PERF_HEADER_VERSION_1;
2506 pr_debug("legacy perf.data format\n");
2508 return try_all_pipe_abis(hdr_sz, ph);
2510 return try_all_file_abis(hdr_sz, ph);
2513 * the new magic number serves two purposes:
2514 * - unique number to identify actual perf.data files
2515 * - encode endianness of file
2518 /* check magic number with one endianness */
2519 if (magic == __perf_magic2)
2522 /* check magic number with opposite endianness */
2523 if (magic != __perf_magic2_sw)
2526 ph->needs_swap = true;
2527 ph->version = PERF_HEADER_VERSION_2;
2532 int perf_file_header__read(struct perf_file_header *header,
2533 struct perf_header *ph, int fd)
2537 lseek(fd, 0, SEEK_SET);
2539 ret = readn(fd, header, sizeof(*header));
2543 if (check_magic_endian(header->magic,
2544 header->attr_size, false, ph) < 0) {
2545 pr_debug("magic/endian check failed\n");
2549 if (ph->needs_swap) {
2550 mem_bswap_64(header, offsetof(struct perf_file_header,
2554 if (header->size != sizeof(*header)) {
2555 /* Support the previous format */
2556 if (header->size == offsetof(typeof(*header), adds_features))
2557 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2560 } else if (ph->needs_swap) {
2562 * feature bitmap is declared as an array of unsigned longs --
2563 * not good since its size can differ between the host that
2564 * generated the data file and the host analyzing the file.
2566 * We need to handle endianness, but we don't know the size of
2567 * the unsigned long where the file was generated. Take a best
2568 * guess at determining it: try 64-bit swap first (ie., file
2569 * created on a 64-bit host), and check if the hostname feature
2570 * bit is set (this feature bit is forced on as of fbe96f2).
2571 * If the bit is not, undo the 64-bit swap and try a 32-bit
2572 * swap. If the hostname bit is still not set (e.g., older data
2573 * file), punt and fallback to the original behavior --
2574 * clearing all feature bits and setting buildid.
2576 mem_bswap_64(&header->adds_features,
2577 BITS_TO_U64(HEADER_FEAT_BITS));
2579 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2581 mem_bswap_64(&header->adds_features,
2582 BITS_TO_U64(HEADER_FEAT_BITS));
2585 mem_bswap_32(&header->adds_features,
2586 BITS_TO_U32(HEADER_FEAT_BITS));
2589 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2590 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2591 set_bit(HEADER_BUILD_ID, header->adds_features);
2595 memcpy(&ph->adds_features, &header->adds_features,
2596 sizeof(ph->adds_features));
2598 ph->data_offset = header->data.offset;
2599 ph->data_size = header->data.size;
2600 ph->feat_offset = header->data.offset + header->data.size;
2604 static int perf_file_section__process(struct perf_file_section *section,
2605 struct perf_header *ph,
2606 int feat, int fd, void *data)
2608 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2609 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2610 "%d, continuing...\n", section->offset, feat);
2614 if (feat >= HEADER_LAST_FEATURE) {
2615 pr_debug("unknown feature %d, continuing...\n", feat);
2619 if (!feat_ops[feat].process)
2622 return feat_ops[feat].process(section, ph, fd, data);
2625 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2626 struct perf_header *ph, int fd,
2631 ret = readn(fd, header, sizeof(*header));
2635 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2636 pr_debug("endian/magic failed\n");
2641 header->size = bswap_64(header->size);
2643 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2649 static int perf_header__read_pipe(struct perf_session *session)
2651 struct perf_header *header = &session->header;
2652 struct perf_pipe_file_header f_header;
2654 if (perf_file_header__read_pipe(&f_header, header,
2655 perf_data_file__fd(session->file),
2656 session->repipe) < 0) {
2657 pr_debug("incompatible file format\n");
2664 static int read_attr(int fd, struct perf_header *ph,
2665 struct perf_file_attr *f_attr)
2667 struct perf_event_attr *attr = &f_attr->attr;
2669 size_t our_sz = sizeof(f_attr->attr);
2672 memset(f_attr, 0, sizeof(*f_attr));
2674 /* read minimal guaranteed structure */
2675 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2677 pr_debug("cannot read %d bytes of header attr\n",
2678 PERF_ATTR_SIZE_VER0);
2682 /* on file perf_event_attr size */
2690 sz = PERF_ATTR_SIZE_VER0;
2691 } else if (sz > our_sz) {
2692 pr_debug("file uses a more recent and unsupported ABI"
2693 " (%zu bytes extra)\n", sz - our_sz);
2696 /* what we have not yet read and that we know about */
2697 left = sz - PERF_ATTR_SIZE_VER0;
2700 ptr += PERF_ATTR_SIZE_VER0;
2702 ret = readn(fd, ptr, left);
2704 /* read perf_file_section, ids are read in caller */
2705 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2707 return ret <= 0 ? -1 : 0;
2710 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2711 struct pevent *pevent)
2713 struct event_format *event;
2716 /* already prepared */
2717 if (evsel->tp_format)
2720 if (pevent == NULL) {
2721 pr_debug("broken or missing trace data\n");
2725 event = pevent_find_event(pevent, evsel->attr.config);
2730 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2731 evsel->name = strdup(bf);
2732 if (evsel->name == NULL)
2736 evsel->tp_format = event;
2740 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2741 struct pevent *pevent)
2743 struct perf_evsel *pos;
2745 list_for_each_entry(pos, &evlist->entries, node) {
2746 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2747 perf_evsel__prepare_tracepoint_event(pos, pevent))
2754 int perf_session__read_header(struct perf_session *session)
2756 struct perf_data_file *file = session->file;
2757 struct perf_header *header = &session->header;
2758 struct perf_file_header f_header;
2759 struct perf_file_attr f_attr;
2761 int nr_attrs, nr_ids, i, j;
2762 int fd = perf_data_file__fd(file);
2764 session->evlist = perf_evlist__new();
2765 if (session->evlist == NULL)
2768 if (perf_data_file__is_pipe(file))
2769 return perf_header__read_pipe(session);
2771 if (perf_file_header__read(&f_header, header, fd) < 0)
2775 * Sanity check that perf.data was written cleanly; data size is
2776 * initialized to 0 and updated only if the on_exit function is run.
2777 * If data size is still 0 then the file contains only partial
2778 * information. Just warn user and process it as much as it can.
2780 if (f_header.data.size == 0) {
2781 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2782 "Was the 'perf record' command properly terminated?\n",
2786 nr_attrs = f_header.attrs.size / f_header.attr_size;
2787 lseek(fd, f_header.attrs.offset, SEEK_SET);
2789 for (i = 0; i < nr_attrs; i++) {
2790 struct perf_evsel *evsel;
2793 if (read_attr(fd, header, &f_attr) < 0)
2796 if (header->needs_swap)
2797 perf_event__attr_swap(&f_attr.attr);
2799 tmp = lseek(fd, 0, SEEK_CUR);
2800 evsel = perf_evsel__new(&f_attr.attr, i);
2803 goto out_delete_evlist;
2805 evsel->needs_swap = header->needs_swap;
2807 * Do it before so that if perf_evsel__alloc_id fails, this
2808 * entry gets purged too at perf_evlist__delete().
2810 perf_evlist__add(session->evlist, evsel);
2812 nr_ids = f_attr.ids.size / sizeof(u64);
2814 * We don't have the cpu and thread maps on the header, so
2815 * for allocating the perf_sample_id table we fake 1 cpu and
2816 * hattr->ids threads.
2818 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2819 goto out_delete_evlist;
2821 lseek(fd, f_attr.ids.offset, SEEK_SET);
2823 for (j = 0; j < nr_ids; j++) {
2824 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2827 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2830 lseek(fd, tmp, SEEK_SET);
2833 symbol_conf.nr_events = nr_attrs;
2835 perf_header__process_sections(header, fd, &session->pevent,
2836 perf_file_section__process);
2838 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2840 goto out_delete_evlist;
2847 perf_evlist__delete(session->evlist);
2848 session->evlist = NULL;
2852 int perf_event__synthesize_attr(struct perf_tool *tool,
2853 struct perf_event_attr *attr, u32 ids, u64 *id,
2854 perf_event__handler_t process)
2856 union perf_event *ev;
2860 size = sizeof(struct perf_event_attr);
2861 size = PERF_ALIGN(size, sizeof(u64));
2862 size += sizeof(struct perf_event_header);
2863 size += ids * sizeof(u64);
2870 ev->attr.attr = *attr;
2871 memcpy(ev->attr.id, id, ids * sizeof(u64));
2873 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2874 ev->attr.header.size = (u16)size;
2876 if (ev->attr.header.size == size)
2877 err = process(tool, ev, NULL, NULL);
2886 int perf_event__synthesize_attrs(struct perf_tool *tool,
2887 struct perf_session *session,
2888 perf_event__handler_t process)
2890 struct perf_evsel *evsel;
2893 list_for_each_entry(evsel, &session->evlist->entries, node) {
2894 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2895 evsel->id, process);
2897 pr_debug("failed to create perf header attribute\n");
2905 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2906 union perf_event *event,
2907 struct perf_evlist **pevlist)
2910 struct perf_evsel *evsel;
2911 struct perf_evlist *evlist = *pevlist;
2913 if (evlist == NULL) {
2914 *pevlist = evlist = perf_evlist__new();
2919 evsel = perf_evsel__new(&event->attr.attr, evlist->nr_entries);
2923 perf_evlist__add(evlist, evsel);
2925 ids = event->header.size;
2926 ids -= (void *)&event->attr.id - (void *)event;
2927 n_ids = ids / sizeof(u64);
2929 * We don't have the cpu and thread maps on the header, so
2930 * for allocating the perf_sample_id table we fake 1 cpu and
2931 * hattr->ids threads.
2933 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2936 for (i = 0; i < n_ids; i++) {
2937 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2940 symbol_conf.nr_events = evlist->nr_entries;
2945 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2946 struct perf_evlist *evlist,
2947 perf_event__handler_t process)
2949 union perf_event ev;
2950 struct tracing_data *tdata;
2951 ssize_t size = 0, aligned_size = 0, padding;
2952 int err __maybe_unused = 0;
2955 * We are going to store the size of the data followed
2956 * by the data contents. Since the fd descriptor is a pipe,
2957 * we cannot seek back to store the size of the data once
2958 * we know it. Instead we:
2960 * - write the tracing data to the temp file
2961 * - get/write the data size to pipe
2962 * - write the tracing data from the temp file
2965 tdata = tracing_data_get(&evlist->entries, fd, true);
2969 memset(&ev, 0, sizeof(ev));
2971 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2973 aligned_size = PERF_ALIGN(size, sizeof(u64));
2974 padding = aligned_size - size;
2975 ev.tracing_data.header.size = sizeof(ev.tracing_data);
2976 ev.tracing_data.size = aligned_size;
2978 process(tool, &ev, NULL, NULL);
2981 * The put function will copy all the tracing data
2982 * stored in temp file to the pipe.
2984 tracing_data_put(tdata);
2986 write_padded(fd, NULL, 0, padding);
2988 return aligned_size;
2991 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2992 union perf_event *event,
2993 struct perf_session *session)
2995 ssize_t size_read, padding, size = event->tracing_data.size;
2996 int fd = perf_data_file__fd(session->file);
2997 off_t offset = lseek(fd, 0, SEEK_CUR);
3000 /* setup for reading amidst mmap */
3001 lseek(fd, offset + sizeof(struct tracing_data_event),
3004 size_read = trace_report(fd, &session->pevent,
3006 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3008 if (readn(fd, buf, padding) < 0) {
3009 pr_err("%s: reading input file", __func__);
3012 if (session->repipe) {
3013 int retw = write(STDOUT_FILENO, buf, padding);
3014 if (retw <= 0 || retw != padding) {
3015 pr_err("%s: repiping tracing data padding", __func__);
3020 if (size_read + padding != size) {
3021 pr_err("%s: tracing data size mismatch", __func__);
3025 perf_evlist__prepare_tracepoint_events(session->evlist,
3028 return size_read + padding;
3031 int perf_event__synthesize_build_id(struct perf_tool *tool,
3032 struct dso *pos, u16 misc,
3033 perf_event__handler_t process,
3034 struct machine *machine)
3036 union perf_event ev;
3043 memset(&ev, 0, sizeof(ev));
3045 len = pos->long_name_len + 1;
3046 len = PERF_ALIGN(len, NAME_ALIGN);
3047 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3048 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3049 ev.build_id.header.misc = misc;
3050 ev.build_id.pid = machine->pid;
3051 ev.build_id.header.size = sizeof(ev.build_id) + len;
3052 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3054 err = process(tool, &ev, NULL, machine);
3059 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3060 union perf_event *event,
3061 struct perf_session *session)
3063 __event_process_build_id(&event->build_id,
3064 event->build_id.filename,
3069 void disable_buildid_cache(void)
3071 no_buildid_cache = true;