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 u32 header_argc;
28 static const char **header_argv;
32 * must be a numerical value to let the endianness
33 * determine the memory layout. That way we are able
34 * to detect endianness when reading the perf.data file
37 * we check for legacy (PERFFILE) format.
39 static const char *__perf_magic1 = "PERFFILE";
40 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
41 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
43 #define PERF_MAGIC __perf_magic2
45 struct perf_file_attr {
46 struct perf_event_attr attr;
47 struct perf_file_section ids;
50 void perf_header__set_feat(struct perf_header *header, int feat)
52 set_bit(feat, header->adds_features);
55 void perf_header__clear_feat(struct perf_header *header, int feat)
57 clear_bit(feat, header->adds_features);
60 bool perf_header__has_feat(const struct perf_header *header, int feat)
62 return test_bit(feat, header->adds_features);
65 static int do_write(int fd, const void *buf, size_t size)
68 int ret = write(fd, buf, size);
80 int write_padded(int fd, const void *bf, size_t count, size_t count_aligned)
82 static const char zero_buf[NAME_ALIGN];
83 int err = do_write(fd, bf, count);
86 err = do_write(fd, zero_buf, count_aligned - count);
91 static int do_write_string(int fd, const char *str)
96 olen = strlen(str) + 1;
97 len = PERF_ALIGN(olen, NAME_ALIGN);
99 /* write len, incl. \0 */
100 ret = do_write(fd, &len, sizeof(len));
104 return write_padded(fd, str, olen, len);
107 static char *do_read_string(int fd, struct perf_header *ph)
113 sz = readn(fd, &len, sizeof(len));
114 if (sz < (ssize_t)sizeof(len))
124 ret = readn(fd, buf, len);
125 if (ret == (ssize_t)len) {
127 * strings are padded by zeroes
128 * thus the actual strlen of buf
129 * may be less than len
139 perf_header__set_cmdline(int argc, const char **argv)
144 * If header_argv has already been set, do not override it.
145 * This allows a command to set the cmdline, parse args and
146 * then call another builtin function that implements a
147 * command -- e.g, cmd_kvm calling cmd_record.
152 header_argc = (u32)argc;
154 /* do not include NULL termination */
155 header_argv = calloc(argc, sizeof(char *));
160 * must copy argv contents because it gets moved
161 * around during option parsing
163 for (i = 0; i < argc ; i++)
164 header_argv[i] = argv[i];
169 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
170 struct perf_evlist *evlist)
172 return read_tracing_data(fd, &evlist->entries);
176 static int write_build_id(int fd, struct perf_header *h,
177 struct perf_evlist *evlist __maybe_unused)
179 struct perf_session *session;
182 session = container_of(h, struct perf_session, header);
184 if (!perf_session__read_build_ids(session, true))
187 err = perf_session__write_buildid_table(session, fd);
189 pr_debug("failed to write buildid table\n");
192 perf_session__cache_build_ids(session);
197 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
198 struct perf_evlist *evlist __maybe_unused)
207 return do_write_string(fd, uts.nodename);
210 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
211 struct perf_evlist *evlist __maybe_unused)
220 return do_write_string(fd, uts.release);
223 static int write_arch(int fd, struct perf_header *h __maybe_unused,
224 struct perf_evlist *evlist __maybe_unused)
233 return do_write_string(fd, uts.machine);
236 static int write_version(int fd, struct perf_header *h __maybe_unused,
237 struct perf_evlist *evlist __maybe_unused)
239 return do_write_string(fd, perf_version_string);
242 static int __write_cpudesc(int fd, const char *cpuinfo_proc)
247 const char *search = cpuinfo_proc;
254 file = fopen("/proc/cpuinfo", "r");
258 while (getline(&buf, &len, file) > 0) {
259 ret = strncmp(buf, search, strlen(search));
271 p = strchr(buf, ':');
272 if (p && *(p+1) == ' ' && *(p+2))
278 /* squash extra space characters (branding string) */
285 while (*q && isspace(*q))
288 while ((*r++ = *q++));
292 ret = do_write_string(fd, s);
299 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
300 struct perf_evlist *evlist __maybe_unused)
303 #define CPUINFO_PROC {"model name", }
305 const char *cpuinfo_procs[] = CPUINFO_PROC;
308 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
310 ret = __write_cpudesc(fd, cpuinfo_procs[i]);
318 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
319 struct perf_evlist *evlist __maybe_unused)
325 nr = sysconf(_SC_NPROCESSORS_CONF);
329 nrc = (u32)(nr & UINT_MAX);
331 nr = sysconf(_SC_NPROCESSORS_ONLN);
335 nra = (u32)(nr & UINT_MAX);
337 ret = do_write(fd, &nrc, sizeof(nrc));
341 return do_write(fd, &nra, sizeof(nra));
344 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
345 struct perf_evlist *evlist)
347 struct perf_evsel *evsel;
351 nre = evlist->nr_entries;
354 * write number of events
356 ret = do_write(fd, &nre, sizeof(nre));
361 * size of perf_event_attr struct
363 sz = (u32)sizeof(evsel->attr);
364 ret = do_write(fd, &sz, sizeof(sz));
368 evlist__for_each(evlist, evsel) {
369 ret = do_write(fd, &evsel->attr, sz);
373 * write number of unique id per event
374 * there is one id per instance of an event
376 * copy into an nri to be independent of the
380 ret = do_write(fd, &nri, sizeof(nri));
385 * write event string as passed on cmdline
387 ret = do_write_string(fd, perf_evsel__name(evsel));
391 * write unique ids for this event
393 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
400 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
401 struct perf_evlist *evlist __maybe_unused)
403 char buf[MAXPATHLEN];
409 * actual atual path to perf binary
411 sprintf(proc, "/proc/%d/exe", getpid());
412 ret = readlink(proc, buf, sizeof(buf));
416 /* readlink() does not add null termination */
419 /* account for binary path */
422 ret = do_write(fd, &n, sizeof(n));
426 ret = do_write_string(fd, buf);
430 for (i = 0 ; i < header_argc; i++) {
431 ret = do_write_string(fd, header_argv[i]);
438 #define CORE_SIB_FMT \
439 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
440 #define THRD_SIB_FMT \
441 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
446 char **core_siblings;
447 char **thread_siblings;
450 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
453 char filename[MAXPATHLEN];
454 char *buf = NULL, *p;
460 sprintf(filename, CORE_SIB_FMT, cpu);
461 fp = fopen(filename, "r");
465 sret = getline(&buf, &len, fp);
470 p = strchr(buf, '\n');
474 for (i = 0; i < tp->core_sib; i++) {
475 if (!strcmp(buf, tp->core_siblings[i]))
478 if (i == tp->core_sib) {
479 tp->core_siblings[i] = buf;
487 sprintf(filename, THRD_SIB_FMT, cpu);
488 fp = fopen(filename, "r");
492 if (getline(&buf, &len, fp) <= 0)
495 p = strchr(buf, '\n');
499 for (i = 0; i < tp->thread_sib; i++) {
500 if (!strcmp(buf, tp->thread_siblings[i]))
503 if (i == tp->thread_sib) {
504 tp->thread_siblings[i] = buf;
516 static void free_cpu_topo(struct cpu_topo *tp)
523 for (i = 0 ; i < tp->core_sib; i++)
524 zfree(&tp->core_siblings[i]);
526 for (i = 0 ; i < tp->thread_sib; i++)
527 zfree(&tp->thread_siblings[i]);
532 static struct cpu_topo *build_cpu_topology(void)
541 ncpus = sysconf(_SC_NPROCESSORS_CONF);
545 nr = (u32)(ncpus & UINT_MAX);
547 sz = nr * sizeof(char *);
549 addr = calloc(1, sizeof(*tp) + 2 * sz);
556 tp->core_siblings = addr;
558 tp->thread_siblings = addr;
560 for (i = 0; i < nr; i++) {
561 ret = build_cpu_topo(tp, i);
572 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
573 struct perf_evlist *evlist __maybe_unused)
579 tp = build_cpu_topology();
583 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
587 for (i = 0; i < tp->core_sib; i++) {
588 ret = do_write_string(fd, tp->core_siblings[i]);
592 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
596 for (i = 0; i < tp->thread_sib; i++) {
597 ret = do_write_string(fd, tp->thread_siblings[i]);
608 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
609 struct perf_evlist *evlist __maybe_unused)
617 fp = fopen("/proc/meminfo", "r");
621 while (getline(&buf, &len, fp) > 0) {
622 ret = strncmp(buf, "MemTotal:", 9);
627 n = sscanf(buf, "%*s %"PRIu64, &mem);
629 ret = do_write(fd, &mem, sizeof(mem));
637 static int write_topo_node(int fd, int node)
639 char str[MAXPATHLEN];
641 char *buf = NULL, *p;
644 u64 mem_total, mem_free, mem;
647 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
648 fp = fopen(str, "r");
652 while (getline(&buf, &len, fp) > 0) {
653 /* skip over invalid lines */
654 if (!strchr(buf, ':'))
656 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
658 if (!strcmp(field, "MemTotal:"))
660 if (!strcmp(field, "MemFree:"))
667 ret = do_write(fd, &mem_total, sizeof(u64));
671 ret = do_write(fd, &mem_free, sizeof(u64));
676 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
678 fp = fopen(str, "r");
682 if (getline(&buf, &len, fp) <= 0)
685 p = strchr(buf, '\n');
689 ret = do_write_string(fd, buf);
697 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
698 struct perf_evlist *evlist __maybe_unused)
703 struct cpu_map *node_map = NULL;
708 fp = fopen("/sys/devices/system/node/online", "r");
712 if (getline(&buf, &len, fp) <= 0)
715 c = strchr(buf, '\n');
719 node_map = cpu_map__new(buf);
723 nr = (u32)node_map->nr;
725 ret = do_write(fd, &nr, sizeof(nr));
729 for (i = 0; i < nr; i++) {
730 j = (u32)node_map->map[i];
731 ret = do_write(fd, &j, sizeof(j));
735 ret = write_topo_node(fd, i);
749 * struct pmu_mappings {
758 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
759 struct perf_evlist *evlist __maybe_unused)
761 struct perf_pmu *pmu = NULL;
762 off_t offset = lseek(fd, 0, SEEK_CUR);
766 /* write real pmu_num later */
767 ret = do_write(fd, &pmu_num, sizeof(pmu_num));
771 while ((pmu = perf_pmu__scan(pmu))) {
776 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
780 ret = do_write_string(fd, pmu->name);
785 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
787 lseek(fd, offset, SEEK_SET);
797 * struct group_descs {
799 * struct group_desc {
806 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
807 struct perf_evlist *evlist)
809 u32 nr_groups = evlist->nr_groups;
810 struct perf_evsel *evsel;
813 ret = do_write(fd, &nr_groups, sizeof(nr_groups));
817 evlist__for_each(evlist, evsel) {
818 if (perf_evsel__is_group_leader(evsel) &&
819 evsel->nr_members > 1) {
820 const char *name = evsel->group_name ?: "{anon_group}";
821 u32 leader_idx = evsel->idx;
822 u32 nr_members = evsel->nr_members;
824 ret = do_write_string(fd, name);
828 ret = do_write(fd, &leader_idx, sizeof(leader_idx));
832 ret = do_write(fd, &nr_members, sizeof(nr_members));
841 * default get_cpuid(): nothing gets recorded
842 * actual implementation must be in arch/$(ARCH)/util/header.c
844 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
845 size_t sz __maybe_unused)
850 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
851 struct perf_evlist *evlist __maybe_unused)
856 ret = get_cpuid(buffer, sizeof(buffer));
862 return do_write_string(fd, buffer);
865 static int write_branch_stack(int fd __maybe_unused,
866 struct perf_header *h __maybe_unused,
867 struct perf_evlist *evlist __maybe_unused)
872 static int write_auxtrace(int fd, struct perf_header *h,
873 struct perf_evlist *evlist __maybe_unused)
875 struct perf_session *session;
878 session = container_of(h, struct perf_session, header);
880 err = auxtrace_index__write(fd, &session->auxtrace_index);
882 pr_err("Failed to write auxtrace index\n");
886 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
889 fprintf(fp, "# hostname : %s\n", ph->env.hostname);
892 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
895 fprintf(fp, "# os release : %s\n", ph->env.os_release);
898 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
900 fprintf(fp, "# arch : %s\n", ph->env.arch);
903 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
906 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
909 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
912 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
913 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
916 static void print_version(struct perf_header *ph, int fd __maybe_unused,
919 fprintf(fp, "# perf version : %s\n", ph->env.version);
922 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
928 nr = ph->env.nr_cmdline;
929 str = ph->env.cmdline;
931 fprintf(fp, "# cmdline : ");
933 for (i = 0; i < nr; i++) {
934 fprintf(fp, "%s ", str);
935 str += strlen(str) + 1;
940 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
946 nr = ph->env.nr_sibling_cores;
947 str = ph->env.sibling_cores;
949 for (i = 0; i < nr; i++) {
950 fprintf(fp, "# sibling cores : %s\n", str);
951 str += strlen(str) + 1;
954 nr = ph->env.nr_sibling_threads;
955 str = ph->env.sibling_threads;
957 for (i = 0; i < nr; i++) {
958 fprintf(fp, "# sibling threads : %s\n", str);
959 str += strlen(str) + 1;
963 static void free_event_desc(struct perf_evsel *events)
965 struct perf_evsel *evsel;
970 for (evsel = events; evsel->attr.size; evsel++) {
978 static struct perf_evsel *
979 read_event_desc(struct perf_header *ph, int fd)
981 struct perf_evsel *evsel, *events = NULL;
984 u32 nre, sz, nr, i, j;
988 /* number of events */
989 ret = readn(fd, &nre, sizeof(nre));
990 if (ret != (ssize_t)sizeof(nre))
996 ret = readn(fd, &sz, sizeof(sz));
997 if (ret != (ssize_t)sizeof(sz))
1003 /* buffer to hold on file attr struct */
1008 /* the last event terminates with evsel->attr.size == 0: */
1009 events = calloc(nre + 1, sizeof(*events));
1013 msz = sizeof(evsel->attr);
1017 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1021 * must read entire on-file attr struct to
1022 * sync up with layout.
1024 ret = readn(fd, buf, sz);
1025 if (ret != (ssize_t)sz)
1029 perf_event__attr_swap(buf);
1031 memcpy(&evsel->attr, buf, msz);
1033 ret = readn(fd, &nr, sizeof(nr));
1034 if (ret != (ssize_t)sizeof(nr))
1037 if (ph->needs_swap) {
1039 evsel->needs_swap = true;
1042 evsel->name = do_read_string(fd, ph);
1047 id = calloc(nr, sizeof(*id));
1053 for (j = 0 ; j < nr; j++) {
1054 ret = readn(fd, id, sizeof(*id));
1055 if (ret != (ssize_t)sizeof(*id))
1058 *id = bswap_64(*id);
1067 free_event_desc(events);
1072 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1073 void *priv __attribute__((unused)))
1075 return fprintf(fp, ", %s = %s", name, val);
1078 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1080 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1085 fprintf(fp, "# event desc: not available or unable to read\n");
1089 for (evsel = events; evsel->attr.size; evsel++) {
1090 fprintf(fp, "# event : name = %s, ", evsel->name);
1093 fprintf(fp, ", id = {");
1094 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1097 fprintf(fp, " %"PRIu64, *id);
1102 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1107 free_event_desc(events);
1110 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1113 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1116 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1121 uint64_t mem_total, mem_free;
1124 nr = ph->env.nr_numa_nodes;
1125 str = ph->env.numa_nodes;
1127 for (i = 0; i < nr; i++) {
1129 c = strtoul(str, &tmp, 0);
1134 mem_total = strtoull(str, &tmp, 0);
1139 mem_free = strtoull(str, &tmp, 0);
1143 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1144 " free = %"PRIu64" kB\n",
1145 c, mem_total, mem_free);
1148 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1150 str += strlen(str) + 1;
1154 fprintf(fp, "# numa topology : not available\n");
1157 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1159 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1162 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1163 int fd __maybe_unused, FILE *fp)
1165 fprintf(fp, "# contains samples with branch stack\n");
1168 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1169 int fd __maybe_unused, FILE *fp)
1171 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1174 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1177 const char *delimiter = "# pmu mappings: ";
1182 pmu_num = ph->env.nr_pmu_mappings;
1184 fprintf(fp, "# pmu mappings: not available\n");
1188 str = ph->env.pmu_mappings;
1191 type = strtoul(str, &tmp, 0);
1196 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1199 str += strlen(str) + 1;
1208 fprintf(fp, "# pmu mappings: unable to read\n");
1211 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1214 struct perf_session *session;
1215 struct perf_evsel *evsel;
1218 session = container_of(ph, struct perf_session, header);
1220 evlist__for_each(session->evlist, evsel) {
1221 if (perf_evsel__is_group_leader(evsel) &&
1222 evsel->nr_members > 1) {
1223 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1224 perf_evsel__name(evsel));
1226 nr = evsel->nr_members - 1;
1228 fprintf(fp, ",%s", perf_evsel__name(evsel));
1236 static int __event_process_build_id(struct build_id_event *bev,
1238 struct perf_session *session)
1241 struct machine *machine;
1244 enum dso_kernel_type dso_type;
1246 machine = perf_session__findnew_machine(session, bev->pid);
1250 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1253 case PERF_RECORD_MISC_KERNEL:
1254 dso_type = DSO_TYPE_KERNEL;
1256 case PERF_RECORD_MISC_GUEST_KERNEL:
1257 dso_type = DSO_TYPE_GUEST_KERNEL;
1259 case PERF_RECORD_MISC_USER:
1260 case PERF_RECORD_MISC_GUEST_USER:
1261 dso_type = DSO_TYPE_USER;
1267 dso = machine__findnew_dso(machine, filename);
1269 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1271 dso__set_build_id(dso, &bev->build_id);
1273 if (!is_kernel_module(filename, cpumode))
1274 dso->kernel = dso_type;
1276 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1278 pr_debug("build id event received for %s: %s\n",
1279 dso->long_name, sbuild_id);
1288 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1289 int input, u64 offset, u64 size)
1291 struct perf_session *session = container_of(header, struct perf_session, header);
1293 struct perf_event_header header;
1294 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1297 struct build_id_event bev;
1298 char filename[PATH_MAX];
1299 u64 limit = offset + size;
1301 while (offset < limit) {
1304 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1307 if (header->needs_swap)
1308 perf_event_header__bswap(&old_bev.header);
1310 len = old_bev.header.size - sizeof(old_bev);
1311 if (readn(input, filename, len) != len)
1314 bev.header = old_bev.header;
1317 * As the pid is the missing value, we need to fill
1318 * it properly. The header.misc value give us nice hint.
1320 bev.pid = HOST_KERNEL_ID;
1321 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1322 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1323 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1325 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1326 __event_process_build_id(&bev, filename, session);
1328 offset += bev.header.size;
1334 static int perf_header__read_build_ids(struct perf_header *header,
1335 int input, u64 offset, u64 size)
1337 struct perf_session *session = container_of(header, struct perf_session, header);
1338 struct build_id_event bev;
1339 char filename[PATH_MAX];
1340 u64 limit = offset + size, orig_offset = offset;
1343 while (offset < limit) {
1346 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1349 if (header->needs_swap)
1350 perf_event_header__bswap(&bev.header);
1352 len = bev.header.size - sizeof(bev);
1353 if (readn(input, filename, len) != len)
1356 * The a1645ce1 changeset:
1358 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1360 * Added a field to struct build_id_event that broke the file
1363 * Since the kernel build-id is the first entry, process the
1364 * table using the old format if the well known
1365 * '[kernel.kallsyms]' string for the kernel build-id has the
1366 * first 4 characters chopped off (where the pid_t sits).
1368 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1369 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1371 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1374 __event_process_build_id(&bev, filename, session);
1376 offset += bev.header.size;
1383 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1384 struct perf_header *ph __maybe_unused,
1387 ssize_t ret = trace_report(fd, data, false);
1388 return ret < 0 ? -1 : 0;
1391 static int process_build_id(struct perf_file_section *section,
1392 struct perf_header *ph, int fd,
1393 void *data __maybe_unused)
1395 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1396 pr_debug("Failed to read buildids, continuing...\n");
1400 static int process_hostname(struct perf_file_section *section __maybe_unused,
1401 struct perf_header *ph, int fd,
1402 void *data __maybe_unused)
1404 ph->env.hostname = do_read_string(fd, ph);
1405 return ph->env.hostname ? 0 : -ENOMEM;
1408 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1409 struct perf_header *ph, int fd,
1410 void *data __maybe_unused)
1412 ph->env.os_release = do_read_string(fd, ph);
1413 return ph->env.os_release ? 0 : -ENOMEM;
1416 static int process_version(struct perf_file_section *section __maybe_unused,
1417 struct perf_header *ph, int fd,
1418 void *data __maybe_unused)
1420 ph->env.version = do_read_string(fd, ph);
1421 return ph->env.version ? 0 : -ENOMEM;
1424 static int process_arch(struct perf_file_section *section __maybe_unused,
1425 struct perf_header *ph, int fd,
1426 void *data __maybe_unused)
1428 ph->env.arch = do_read_string(fd, ph);
1429 return ph->env.arch ? 0 : -ENOMEM;
1432 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1433 struct perf_header *ph, int fd,
1434 void *data __maybe_unused)
1439 ret = readn(fd, &nr, sizeof(nr));
1440 if (ret != sizeof(nr))
1446 ph->env.nr_cpus_online = nr;
1448 ret = readn(fd, &nr, sizeof(nr));
1449 if (ret != sizeof(nr))
1455 ph->env.nr_cpus_avail = nr;
1459 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1460 struct perf_header *ph, int fd,
1461 void *data __maybe_unused)
1463 ph->env.cpu_desc = do_read_string(fd, ph);
1464 return ph->env.cpu_desc ? 0 : -ENOMEM;
1467 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1468 struct perf_header *ph, int fd,
1469 void *data __maybe_unused)
1471 ph->env.cpuid = do_read_string(fd, ph);
1472 return ph->env.cpuid ? 0 : -ENOMEM;
1475 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1476 struct perf_header *ph, int fd,
1477 void *data __maybe_unused)
1482 ret = readn(fd, &mem, sizeof(mem));
1483 if (ret != sizeof(mem))
1487 mem = bswap_64(mem);
1489 ph->env.total_mem = mem;
1493 static struct perf_evsel *
1494 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1496 struct perf_evsel *evsel;
1498 evlist__for_each(evlist, evsel) {
1499 if (evsel->idx == idx)
1507 perf_evlist__set_event_name(struct perf_evlist *evlist,
1508 struct perf_evsel *event)
1510 struct perf_evsel *evsel;
1515 evsel = perf_evlist__find_by_index(evlist, event->idx);
1522 evsel->name = strdup(event->name);
1526 process_event_desc(struct perf_file_section *section __maybe_unused,
1527 struct perf_header *header, int fd,
1528 void *data __maybe_unused)
1530 struct perf_session *session;
1531 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1536 session = container_of(header, struct perf_session, header);
1537 for (evsel = events; evsel->attr.size; evsel++)
1538 perf_evlist__set_event_name(session->evlist, evsel);
1540 free_event_desc(events);
1545 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1546 struct perf_header *ph, int fd,
1547 void *data __maybe_unused)
1554 ret = readn(fd, &nr, sizeof(nr));
1555 if (ret != sizeof(nr))
1561 ph->env.nr_cmdline = nr;
1562 strbuf_init(&sb, 128);
1564 for (i = 0; i < nr; i++) {
1565 str = do_read_string(fd, ph);
1569 /* include a NULL character at the end */
1570 strbuf_add(&sb, str, strlen(str) + 1);
1573 ph->env.cmdline = strbuf_detach(&sb, NULL);
1577 strbuf_release(&sb);
1581 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1582 struct perf_header *ph, int fd,
1583 void *data __maybe_unused)
1590 ret = readn(fd, &nr, sizeof(nr));
1591 if (ret != sizeof(nr))
1597 ph->env.nr_sibling_cores = nr;
1598 strbuf_init(&sb, 128);
1600 for (i = 0; i < nr; i++) {
1601 str = do_read_string(fd, ph);
1605 /* include a NULL character at the end */
1606 strbuf_add(&sb, str, strlen(str) + 1);
1609 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1611 ret = readn(fd, &nr, sizeof(nr));
1612 if (ret != sizeof(nr))
1618 ph->env.nr_sibling_threads = nr;
1620 for (i = 0; i < nr; i++) {
1621 str = do_read_string(fd, ph);
1625 /* include a NULL character at the end */
1626 strbuf_add(&sb, str, strlen(str) + 1);
1629 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1633 strbuf_release(&sb);
1637 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1638 struct perf_header *ph, int fd,
1639 void *data __maybe_unused)
1644 uint64_t mem_total, mem_free;
1648 ret = readn(fd, &nr, sizeof(nr));
1649 if (ret != sizeof(nr))
1655 ph->env.nr_numa_nodes = nr;
1656 strbuf_init(&sb, 256);
1658 for (i = 0; i < nr; i++) {
1660 ret = readn(fd, &node, sizeof(node));
1661 if (ret != sizeof(node))
1664 ret = readn(fd, &mem_total, sizeof(u64));
1665 if (ret != sizeof(u64))
1668 ret = readn(fd, &mem_free, sizeof(u64));
1669 if (ret != sizeof(u64))
1672 if (ph->needs_swap) {
1673 node = bswap_32(node);
1674 mem_total = bswap_64(mem_total);
1675 mem_free = bswap_64(mem_free);
1678 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1679 node, mem_total, mem_free);
1681 str = do_read_string(fd, ph);
1685 /* include a NULL character at the end */
1686 strbuf_add(&sb, str, strlen(str) + 1);
1689 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1693 strbuf_release(&sb);
1697 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1698 struct perf_header *ph, int fd,
1699 void *data __maybe_unused)
1707 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1708 if (ret != sizeof(pmu_num))
1712 pmu_num = bswap_32(pmu_num);
1715 pr_debug("pmu mappings not available\n");
1719 ph->env.nr_pmu_mappings = pmu_num;
1720 strbuf_init(&sb, 128);
1723 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1726 type = bswap_32(type);
1728 name = do_read_string(fd, ph);
1732 strbuf_addf(&sb, "%u:%s", type, name);
1733 /* include a NULL character at the end */
1734 strbuf_add(&sb, "", 1);
1739 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1743 strbuf_release(&sb);
1747 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1748 struct perf_header *ph, int fd,
1749 void *data __maybe_unused)
1752 u32 i, nr, nr_groups;
1753 struct perf_session *session;
1754 struct perf_evsel *evsel, *leader = NULL;
1761 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1765 nr_groups = bswap_32(nr_groups);
1767 ph->env.nr_groups = nr_groups;
1769 pr_debug("group desc not available\n");
1773 desc = calloc(nr_groups, sizeof(*desc));
1777 for (i = 0; i < nr_groups; i++) {
1778 desc[i].name = do_read_string(fd, ph);
1782 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1785 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1788 if (ph->needs_swap) {
1789 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1790 desc[i].nr_members = bswap_32(desc[i].nr_members);
1795 * Rebuild group relationship based on the group_desc
1797 session = container_of(ph, struct perf_session, header);
1798 session->evlist->nr_groups = nr_groups;
1801 evlist__for_each(session->evlist, evsel) {
1802 if (evsel->idx == (int) desc[i].leader_idx) {
1803 evsel->leader = evsel;
1804 /* {anon_group} is a dummy name */
1805 if (strcmp(desc[i].name, "{anon_group}")) {
1806 evsel->group_name = desc[i].name;
1807 desc[i].name = NULL;
1809 evsel->nr_members = desc[i].nr_members;
1811 if (i >= nr_groups || nr > 0) {
1812 pr_debug("invalid group desc\n");
1817 nr = evsel->nr_members - 1;
1820 /* This is a group member */
1821 evsel->leader = leader;
1827 if (i != nr_groups || nr != 0) {
1828 pr_debug("invalid group desc\n");
1834 for (i = 0; i < nr_groups; i++)
1835 zfree(&desc[i].name);
1841 static int process_auxtrace(struct perf_file_section *section,
1842 struct perf_header *ph, int fd,
1843 void *data __maybe_unused)
1845 struct perf_session *session;
1848 session = container_of(ph, struct perf_session, header);
1850 err = auxtrace_index__process(fd, section->size, session,
1853 pr_err("Failed to process auxtrace index\n");
1857 struct feature_ops {
1858 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1859 void (*print)(struct perf_header *h, int fd, FILE *fp);
1860 int (*process)(struct perf_file_section *section,
1861 struct perf_header *h, int fd, void *data);
1866 #define FEAT_OPA(n, func) \
1867 [n] = { .name = #n, .write = write_##func, .print = print_##func }
1868 #define FEAT_OPP(n, func) \
1869 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1870 .process = process_##func }
1871 #define FEAT_OPF(n, func) \
1872 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1873 .process = process_##func, .full_only = true }
1875 /* feature_ops not implemented: */
1876 #define print_tracing_data NULL
1877 #define print_build_id NULL
1879 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1880 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
1881 FEAT_OPP(HEADER_BUILD_ID, build_id),
1882 FEAT_OPP(HEADER_HOSTNAME, hostname),
1883 FEAT_OPP(HEADER_OSRELEASE, osrelease),
1884 FEAT_OPP(HEADER_VERSION, version),
1885 FEAT_OPP(HEADER_ARCH, arch),
1886 FEAT_OPP(HEADER_NRCPUS, nrcpus),
1887 FEAT_OPP(HEADER_CPUDESC, cpudesc),
1888 FEAT_OPP(HEADER_CPUID, cpuid),
1889 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
1890 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
1891 FEAT_OPP(HEADER_CMDLINE, cmdline),
1892 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
1893 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
1894 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
1895 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
1896 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
1897 FEAT_OPP(HEADER_AUXTRACE, auxtrace),
1900 struct header_print_data {
1902 bool full; /* extended list of headers */
1905 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1906 struct perf_header *ph,
1907 int feat, int fd, void *data)
1909 struct header_print_data *hd = data;
1911 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1912 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1913 "%d, continuing...\n", section->offset, feat);
1916 if (feat >= HEADER_LAST_FEATURE) {
1917 pr_warning("unknown feature %d\n", feat);
1920 if (!feat_ops[feat].print)
1923 if (!feat_ops[feat].full_only || hd->full)
1924 feat_ops[feat].print(ph, fd, hd->fp);
1926 fprintf(hd->fp, "# %s info available, use -I to display\n",
1927 feat_ops[feat].name);
1932 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1934 struct header_print_data hd;
1935 struct perf_header *header = &session->header;
1936 int fd = perf_data_file__fd(session->file);
1940 perf_header__process_sections(header, fd, &hd,
1941 perf_file_section__fprintf_info);
1945 static int do_write_feat(int fd, struct perf_header *h, int type,
1946 struct perf_file_section **p,
1947 struct perf_evlist *evlist)
1952 if (perf_header__has_feat(h, type)) {
1953 if (!feat_ops[type].write)
1956 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1958 err = feat_ops[type].write(fd, h, evlist);
1960 pr_debug("failed to write feature %d\n", type);
1962 /* undo anything written */
1963 lseek(fd, (*p)->offset, SEEK_SET);
1967 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1973 static int perf_header__adds_write(struct perf_header *header,
1974 struct perf_evlist *evlist, int fd)
1977 struct perf_file_section *feat_sec, *p;
1983 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1987 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1988 if (feat_sec == NULL)
1991 sec_size = sizeof(*feat_sec) * nr_sections;
1993 sec_start = header->feat_offset;
1994 lseek(fd, sec_start + sec_size, SEEK_SET);
1996 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
1997 if (do_write_feat(fd, header, feat, &p, evlist))
1998 perf_header__clear_feat(header, feat);
2001 lseek(fd, sec_start, SEEK_SET);
2003 * may write more than needed due to dropped feature, but
2004 * this is okay, reader will skip the mising entries
2006 err = do_write(fd, feat_sec, sec_size);
2008 pr_debug("failed to write feature section\n");
2013 int perf_header__write_pipe(int fd)
2015 struct perf_pipe_file_header f_header;
2018 f_header = (struct perf_pipe_file_header){
2019 .magic = PERF_MAGIC,
2020 .size = sizeof(f_header),
2023 err = do_write(fd, &f_header, sizeof(f_header));
2025 pr_debug("failed to write perf pipe header\n");
2032 int perf_session__write_header(struct perf_session *session,
2033 struct perf_evlist *evlist,
2034 int fd, bool at_exit)
2036 struct perf_file_header f_header;
2037 struct perf_file_attr f_attr;
2038 struct perf_header *header = &session->header;
2039 struct perf_evsel *evsel;
2043 lseek(fd, sizeof(f_header), SEEK_SET);
2045 evlist__for_each(session->evlist, evsel) {
2046 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2047 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2049 pr_debug("failed to write perf header\n");
2054 attr_offset = lseek(fd, 0, SEEK_CUR);
2056 evlist__for_each(evlist, evsel) {
2057 f_attr = (struct perf_file_attr){
2058 .attr = evsel->attr,
2060 .offset = evsel->id_offset,
2061 .size = evsel->ids * sizeof(u64),
2064 err = do_write(fd, &f_attr, sizeof(f_attr));
2066 pr_debug("failed to write perf header attribute\n");
2071 if (!header->data_offset)
2072 header->data_offset = lseek(fd, 0, SEEK_CUR);
2073 header->feat_offset = header->data_offset + header->data_size;
2076 err = perf_header__adds_write(header, evlist, fd);
2081 f_header = (struct perf_file_header){
2082 .magic = PERF_MAGIC,
2083 .size = sizeof(f_header),
2084 .attr_size = sizeof(f_attr),
2086 .offset = attr_offset,
2087 .size = evlist->nr_entries * sizeof(f_attr),
2090 .offset = header->data_offset,
2091 .size = header->data_size,
2093 /* event_types is ignored, store zeros */
2096 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2098 lseek(fd, 0, SEEK_SET);
2099 err = do_write(fd, &f_header, sizeof(f_header));
2101 pr_debug("failed to write perf header\n");
2104 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2109 static int perf_header__getbuffer64(struct perf_header *header,
2110 int fd, void *buf, size_t size)
2112 if (readn(fd, buf, size) <= 0)
2115 if (header->needs_swap)
2116 mem_bswap_64(buf, size);
2121 int perf_header__process_sections(struct perf_header *header, int fd,
2123 int (*process)(struct perf_file_section *section,
2124 struct perf_header *ph,
2125 int feat, int fd, void *data))
2127 struct perf_file_section *feat_sec, *sec;
2133 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2137 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2141 sec_size = sizeof(*feat_sec) * nr_sections;
2143 lseek(fd, header->feat_offset, SEEK_SET);
2145 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2149 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2150 err = process(sec++, header, feat, fd, data);
2160 static const int attr_file_abi_sizes[] = {
2161 [0] = PERF_ATTR_SIZE_VER0,
2162 [1] = PERF_ATTR_SIZE_VER1,
2163 [2] = PERF_ATTR_SIZE_VER2,
2164 [3] = PERF_ATTR_SIZE_VER3,
2165 [4] = PERF_ATTR_SIZE_VER4,
2170 * In the legacy file format, the magic number is not used to encode endianness.
2171 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2172 * on ABI revisions, we need to try all combinations for all endianness to
2173 * detect the endianness.
2175 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2177 uint64_t ref_size, attr_size;
2180 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2181 ref_size = attr_file_abi_sizes[i]
2182 + sizeof(struct perf_file_section);
2183 if (hdr_sz != ref_size) {
2184 attr_size = bswap_64(hdr_sz);
2185 if (attr_size != ref_size)
2188 ph->needs_swap = true;
2190 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2195 /* could not determine endianness */
2199 #define PERF_PIPE_HDR_VER0 16
2201 static const size_t attr_pipe_abi_sizes[] = {
2202 [0] = PERF_PIPE_HDR_VER0,
2207 * In the legacy pipe format, there is an implicit assumption that endiannesss
2208 * between host recording the samples, and host parsing the samples is the
2209 * same. This is not always the case given that the pipe output may always be
2210 * redirected into a file and analyzed on a different machine with possibly a
2211 * different endianness and perf_event ABI revsions in the perf tool itself.
2213 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2218 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2219 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2220 attr_size = bswap_64(hdr_sz);
2221 if (attr_size != hdr_sz)
2224 ph->needs_swap = true;
2226 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2232 bool is_perf_magic(u64 magic)
2234 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2235 || magic == __perf_magic2
2236 || magic == __perf_magic2_sw)
2242 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2243 bool is_pipe, struct perf_header *ph)
2247 /* check for legacy format */
2248 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2250 ph->version = PERF_HEADER_VERSION_1;
2251 pr_debug("legacy perf.data format\n");
2253 return try_all_pipe_abis(hdr_sz, ph);
2255 return try_all_file_abis(hdr_sz, ph);
2258 * the new magic number serves two purposes:
2259 * - unique number to identify actual perf.data files
2260 * - encode endianness of file
2262 ph->version = PERF_HEADER_VERSION_2;
2264 /* check magic number with one endianness */
2265 if (magic == __perf_magic2)
2268 /* check magic number with opposite endianness */
2269 if (magic != __perf_magic2_sw)
2272 ph->needs_swap = true;
2277 int perf_file_header__read(struct perf_file_header *header,
2278 struct perf_header *ph, int fd)
2282 lseek(fd, 0, SEEK_SET);
2284 ret = readn(fd, header, sizeof(*header));
2288 if (check_magic_endian(header->magic,
2289 header->attr_size, false, ph) < 0) {
2290 pr_debug("magic/endian check failed\n");
2294 if (ph->needs_swap) {
2295 mem_bswap_64(header, offsetof(struct perf_file_header,
2299 if (header->size != sizeof(*header)) {
2300 /* Support the previous format */
2301 if (header->size == offsetof(typeof(*header), adds_features))
2302 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2305 } else if (ph->needs_swap) {
2307 * feature bitmap is declared as an array of unsigned longs --
2308 * not good since its size can differ between the host that
2309 * generated the data file and the host analyzing the file.
2311 * We need to handle endianness, but we don't know the size of
2312 * the unsigned long where the file was generated. Take a best
2313 * guess at determining it: try 64-bit swap first (ie., file
2314 * created on a 64-bit host), and check if the hostname feature
2315 * bit is set (this feature bit is forced on as of fbe96f2).
2316 * If the bit is not, undo the 64-bit swap and try a 32-bit
2317 * swap. If the hostname bit is still not set (e.g., older data
2318 * file), punt and fallback to the original behavior --
2319 * clearing all feature bits and setting buildid.
2321 mem_bswap_64(&header->adds_features,
2322 BITS_TO_U64(HEADER_FEAT_BITS));
2324 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2326 mem_bswap_64(&header->adds_features,
2327 BITS_TO_U64(HEADER_FEAT_BITS));
2330 mem_bswap_32(&header->adds_features,
2331 BITS_TO_U32(HEADER_FEAT_BITS));
2334 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2335 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2336 set_bit(HEADER_BUILD_ID, header->adds_features);
2340 memcpy(&ph->adds_features, &header->adds_features,
2341 sizeof(ph->adds_features));
2343 ph->data_offset = header->data.offset;
2344 ph->data_size = header->data.size;
2345 ph->feat_offset = header->data.offset + header->data.size;
2349 static int perf_file_section__process(struct perf_file_section *section,
2350 struct perf_header *ph,
2351 int feat, int fd, void *data)
2353 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2354 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2355 "%d, continuing...\n", section->offset, feat);
2359 if (feat >= HEADER_LAST_FEATURE) {
2360 pr_debug("unknown feature %d, continuing...\n", feat);
2364 if (!feat_ops[feat].process)
2367 return feat_ops[feat].process(section, ph, fd, data);
2370 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2371 struct perf_header *ph, int fd,
2376 ret = readn(fd, header, sizeof(*header));
2380 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2381 pr_debug("endian/magic failed\n");
2386 header->size = bswap_64(header->size);
2388 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2394 static int perf_header__read_pipe(struct perf_session *session)
2396 struct perf_header *header = &session->header;
2397 struct perf_pipe_file_header f_header;
2399 if (perf_file_header__read_pipe(&f_header, header,
2400 perf_data_file__fd(session->file),
2401 session->repipe) < 0) {
2402 pr_debug("incompatible file format\n");
2409 static int read_attr(int fd, struct perf_header *ph,
2410 struct perf_file_attr *f_attr)
2412 struct perf_event_attr *attr = &f_attr->attr;
2414 size_t our_sz = sizeof(f_attr->attr);
2417 memset(f_attr, 0, sizeof(*f_attr));
2419 /* read minimal guaranteed structure */
2420 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2422 pr_debug("cannot read %d bytes of header attr\n",
2423 PERF_ATTR_SIZE_VER0);
2427 /* on file perf_event_attr size */
2435 sz = PERF_ATTR_SIZE_VER0;
2436 } else if (sz > our_sz) {
2437 pr_debug("file uses a more recent and unsupported ABI"
2438 " (%zu bytes extra)\n", sz - our_sz);
2441 /* what we have not yet read and that we know about */
2442 left = sz - PERF_ATTR_SIZE_VER0;
2445 ptr += PERF_ATTR_SIZE_VER0;
2447 ret = readn(fd, ptr, left);
2449 /* read perf_file_section, ids are read in caller */
2450 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2452 return ret <= 0 ? -1 : 0;
2455 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2456 struct pevent *pevent)
2458 struct event_format *event;
2461 /* already prepared */
2462 if (evsel->tp_format)
2465 if (pevent == NULL) {
2466 pr_debug("broken or missing trace data\n");
2470 event = pevent_find_event(pevent, evsel->attr.config);
2475 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2476 evsel->name = strdup(bf);
2477 if (evsel->name == NULL)
2481 evsel->tp_format = event;
2485 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2486 struct pevent *pevent)
2488 struct perf_evsel *pos;
2490 evlist__for_each(evlist, pos) {
2491 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2492 perf_evsel__prepare_tracepoint_event(pos, pevent))
2499 int perf_session__read_header(struct perf_session *session)
2501 struct perf_data_file *file = session->file;
2502 struct perf_header *header = &session->header;
2503 struct perf_file_header f_header;
2504 struct perf_file_attr f_attr;
2506 int nr_attrs, nr_ids, i, j;
2507 int fd = perf_data_file__fd(file);
2509 session->evlist = perf_evlist__new();
2510 if (session->evlist == NULL)
2513 if (perf_data_file__is_pipe(file))
2514 return perf_header__read_pipe(session);
2516 if (perf_file_header__read(&f_header, header, fd) < 0)
2520 * Sanity check that perf.data was written cleanly; data size is
2521 * initialized to 0 and updated only if the on_exit function is run.
2522 * If data size is still 0 then the file contains only partial
2523 * information. Just warn user and process it as much as it can.
2525 if (f_header.data.size == 0) {
2526 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2527 "Was the 'perf record' command properly terminated?\n",
2531 nr_attrs = f_header.attrs.size / f_header.attr_size;
2532 lseek(fd, f_header.attrs.offset, SEEK_SET);
2534 for (i = 0; i < nr_attrs; i++) {
2535 struct perf_evsel *evsel;
2538 if (read_attr(fd, header, &f_attr) < 0)
2541 if (header->needs_swap) {
2542 f_attr.ids.size = bswap_64(f_attr.ids.size);
2543 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2544 perf_event__attr_swap(&f_attr.attr);
2547 tmp = lseek(fd, 0, SEEK_CUR);
2548 evsel = perf_evsel__new(&f_attr.attr);
2551 goto out_delete_evlist;
2553 evsel->needs_swap = header->needs_swap;
2555 * Do it before so that if perf_evsel__alloc_id fails, this
2556 * entry gets purged too at perf_evlist__delete().
2558 perf_evlist__add(session->evlist, evsel);
2560 nr_ids = f_attr.ids.size / sizeof(u64);
2562 * We don't have the cpu and thread maps on the header, so
2563 * for allocating the perf_sample_id table we fake 1 cpu and
2564 * hattr->ids threads.
2566 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2567 goto out_delete_evlist;
2569 lseek(fd, f_attr.ids.offset, SEEK_SET);
2571 for (j = 0; j < nr_ids; j++) {
2572 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2575 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2578 lseek(fd, tmp, SEEK_SET);
2581 symbol_conf.nr_events = nr_attrs;
2583 perf_header__process_sections(header, fd, &session->tevent,
2584 perf_file_section__process);
2586 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2587 session->tevent.pevent))
2588 goto out_delete_evlist;
2595 perf_evlist__delete(session->evlist);
2596 session->evlist = NULL;
2600 int perf_event__synthesize_attr(struct perf_tool *tool,
2601 struct perf_event_attr *attr, u32 ids, u64 *id,
2602 perf_event__handler_t process)
2604 union perf_event *ev;
2608 size = sizeof(struct perf_event_attr);
2609 size = PERF_ALIGN(size, sizeof(u64));
2610 size += sizeof(struct perf_event_header);
2611 size += ids * sizeof(u64);
2618 ev->attr.attr = *attr;
2619 memcpy(ev->attr.id, id, ids * sizeof(u64));
2621 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2622 ev->attr.header.size = (u16)size;
2624 if (ev->attr.header.size == size)
2625 err = process(tool, ev, NULL, NULL);
2634 int perf_event__synthesize_attrs(struct perf_tool *tool,
2635 struct perf_session *session,
2636 perf_event__handler_t process)
2638 struct perf_evsel *evsel;
2641 evlist__for_each(session->evlist, evsel) {
2642 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2643 evsel->id, process);
2645 pr_debug("failed to create perf header attribute\n");
2653 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2654 union perf_event *event,
2655 struct perf_evlist **pevlist)
2658 struct perf_evsel *evsel;
2659 struct perf_evlist *evlist = *pevlist;
2661 if (evlist == NULL) {
2662 *pevlist = evlist = perf_evlist__new();
2667 evsel = perf_evsel__new(&event->attr.attr);
2671 perf_evlist__add(evlist, evsel);
2673 ids = event->header.size;
2674 ids -= (void *)&event->attr.id - (void *)event;
2675 n_ids = ids / sizeof(u64);
2677 * We don't have the cpu and thread maps on the header, so
2678 * for allocating the perf_sample_id table we fake 1 cpu and
2679 * hattr->ids threads.
2681 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2684 for (i = 0; i < n_ids; i++) {
2685 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2688 symbol_conf.nr_events = evlist->nr_entries;
2693 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2694 struct perf_evlist *evlist,
2695 perf_event__handler_t process)
2697 union perf_event ev;
2698 struct tracing_data *tdata;
2699 ssize_t size = 0, aligned_size = 0, padding;
2700 int err __maybe_unused = 0;
2703 * We are going to store the size of the data followed
2704 * by the data contents. Since the fd descriptor is a pipe,
2705 * we cannot seek back to store the size of the data once
2706 * we know it. Instead we:
2708 * - write the tracing data to the temp file
2709 * - get/write the data size to pipe
2710 * - write the tracing data from the temp file
2713 tdata = tracing_data_get(&evlist->entries, fd, true);
2717 memset(&ev, 0, sizeof(ev));
2719 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2721 aligned_size = PERF_ALIGN(size, sizeof(u64));
2722 padding = aligned_size - size;
2723 ev.tracing_data.header.size = sizeof(ev.tracing_data);
2724 ev.tracing_data.size = aligned_size;
2726 process(tool, &ev, NULL, NULL);
2729 * The put function will copy all the tracing data
2730 * stored in temp file to the pipe.
2732 tracing_data_put(tdata);
2734 write_padded(fd, NULL, 0, padding);
2736 return aligned_size;
2739 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2740 union perf_event *event,
2741 struct perf_session *session)
2743 ssize_t size_read, padding, size = event->tracing_data.size;
2744 int fd = perf_data_file__fd(session->file);
2745 off_t offset = lseek(fd, 0, SEEK_CUR);
2748 /* setup for reading amidst mmap */
2749 lseek(fd, offset + sizeof(struct tracing_data_event),
2752 size_read = trace_report(fd, &session->tevent,
2754 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2756 if (readn(fd, buf, padding) < 0) {
2757 pr_err("%s: reading input file", __func__);
2760 if (session->repipe) {
2761 int retw = write(STDOUT_FILENO, buf, padding);
2762 if (retw <= 0 || retw != padding) {
2763 pr_err("%s: repiping tracing data padding", __func__);
2768 if (size_read + padding != size) {
2769 pr_err("%s: tracing data size mismatch", __func__);
2773 perf_evlist__prepare_tracepoint_events(session->evlist,
2774 session->tevent.pevent);
2776 return size_read + padding;
2779 int perf_event__synthesize_build_id(struct perf_tool *tool,
2780 struct dso *pos, u16 misc,
2781 perf_event__handler_t process,
2782 struct machine *machine)
2784 union perf_event ev;
2791 memset(&ev, 0, sizeof(ev));
2793 len = pos->long_name_len + 1;
2794 len = PERF_ALIGN(len, NAME_ALIGN);
2795 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2796 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2797 ev.build_id.header.misc = misc;
2798 ev.build_id.pid = machine->pid;
2799 ev.build_id.header.size = sizeof(ev.build_id) + len;
2800 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2802 err = process(tool, &ev, NULL, machine);
2807 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2808 union perf_event *event,
2809 struct perf_session *session)
2811 __event_process_build_id(&event->build_id,
2812 event->build_id.filename,