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);
1066 free_event_desc(events);
1071 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1072 void *priv __attribute__((unused)))
1074 return fprintf(fp, ", %s = %s", name, val);
1077 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1079 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1084 fprintf(fp, "# event desc: not available or unable to read\n");
1088 for (evsel = events; evsel->attr.size; evsel++) {
1089 fprintf(fp, "# event : name = %s, ", evsel->name);
1092 fprintf(fp, ", id = {");
1093 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1096 fprintf(fp, " %"PRIu64, *id);
1101 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1106 free_event_desc(events);
1109 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1112 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1115 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1120 uint64_t mem_total, mem_free;
1123 nr = ph->env.nr_numa_nodes;
1124 str = ph->env.numa_nodes;
1126 for (i = 0; i < nr; i++) {
1128 c = strtoul(str, &tmp, 0);
1133 mem_total = strtoull(str, &tmp, 0);
1138 mem_free = strtoull(str, &tmp, 0);
1142 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1143 " free = %"PRIu64" kB\n",
1144 c, mem_total, mem_free);
1147 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1149 str += strlen(str) + 1;
1153 fprintf(fp, "# numa topology : not available\n");
1156 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1158 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1161 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1162 int fd __maybe_unused, FILE *fp)
1164 fprintf(fp, "# contains samples with branch stack\n");
1167 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1168 int fd __maybe_unused, FILE *fp)
1170 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1173 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1176 const char *delimiter = "# pmu mappings: ";
1181 pmu_num = ph->env.nr_pmu_mappings;
1183 fprintf(fp, "# pmu mappings: not available\n");
1187 str = ph->env.pmu_mappings;
1190 type = strtoul(str, &tmp, 0);
1195 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1198 str += strlen(str) + 1;
1207 fprintf(fp, "# pmu mappings: unable to read\n");
1210 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1213 struct perf_session *session;
1214 struct perf_evsel *evsel;
1217 session = container_of(ph, struct perf_session, header);
1219 evlist__for_each(session->evlist, evsel) {
1220 if (perf_evsel__is_group_leader(evsel) &&
1221 evsel->nr_members > 1) {
1222 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1223 perf_evsel__name(evsel));
1225 nr = evsel->nr_members - 1;
1227 fprintf(fp, ",%s", perf_evsel__name(evsel));
1235 static int __event_process_build_id(struct build_id_event *bev,
1237 struct perf_session *session)
1240 struct machine *machine;
1243 enum dso_kernel_type dso_type;
1245 machine = perf_session__findnew_machine(session, bev->pid);
1249 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1252 case PERF_RECORD_MISC_KERNEL:
1253 dso_type = DSO_TYPE_KERNEL;
1255 case PERF_RECORD_MISC_GUEST_KERNEL:
1256 dso_type = DSO_TYPE_GUEST_KERNEL;
1258 case PERF_RECORD_MISC_USER:
1259 case PERF_RECORD_MISC_GUEST_USER:
1260 dso_type = DSO_TYPE_USER;
1266 dso = machine__findnew_dso(machine, filename);
1268 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1270 dso__set_build_id(dso, &bev->build_id);
1272 if (!is_kernel_module(filename, cpumode))
1273 dso->kernel = dso_type;
1275 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1277 pr_debug("build id event received for %s: %s\n",
1278 dso->long_name, sbuild_id);
1287 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1288 int input, u64 offset, u64 size)
1290 struct perf_session *session = container_of(header, struct perf_session, header);
1292 struct perf_event_header header;
1293 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1296 struct build_id_event bev;
1297 char filename[PATH_MAX];
1298 u64 limit = offset + size;
1300 while (offset < limit) {
1303 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1306 if (header->needs_swap)
1307 perf_event_header__bswap(&old_bev.header);
1309 len = old_bev.header.size - sizeof(old_bev);
1310 if (readn(input, filename, len) != len)
1313 bev.header = old_bev.header;
1316 * As the pid is the missing value, we need to fill
1317 * it properly. The header.misc value give us nice hint.
1319 bev.pid = HOST_KERNEL_ID;
1320 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1321 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1322 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1324 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1325 __event_process_build_id(&bev, filename, session);
1327 offset += bev.header.size;
1333 static int perf_header__read_build_ids(struct perf_header *header,
1334 int input, u64 offset, u64 size)
1336 struct perf_session *session = container_of(header, struct perf_session, header);
1337 struct build_id_event bev;
1338 char filename[PATH_MAX];
1339 u64 limit = offset + size, orig_offset = offset;
1342 while (offset < limit) {
1345 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1348 if (header->needs_swap)
1349 perf_event_header__bswap(&bev.header);
1351 len = bev.header.size - sizeof(bev);
1352 if (readn(input, filename, len) != len)
1355 * The a1645ce1 changeset:
1357 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1359 * Added a field to struct build_id_event that broke the file
1362 * Since the kernel build-id is the first entry, process the
1363 * table using the old format if the well known
1364 * '[kernel.kallsyms]' string for the kernel build-id has the
1365 * first 4 characters chopped off (where the pid_t sits).
1367 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1368 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1370 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1373 __event_process_build_id(&bev, filename, session);
1375 offset += bev.header.size;
1382 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1383 struct perf_header *ph __maybe_unused,
1386 ssize_t ret = trace_report(fd, data, false);
1387 return ret < 0 ? -1 : 0;
1390 static int process_build_id(struct perf_file_section *section,
1391 struct perf_header *ph, int fd,
1392 void *data __maybe_unused)
1394 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1395 pr_debug("Failed to read buildids, continuing...\n");
1399 static int process_hostname(struct perf_file_section *section __maybe_unused,
1400 struct perf_header *ph, int fd,
1401 void *data __maybe_unused)
1403 ph->env.hostname = do_read_string(fd, ph);
1404 return ph->env.hostname ? 0 : -ENOMEM;
1407 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1408 struct perf_header *ph, int fd,
1409 void *data __maybe_unused)
1411 ph->env.os_release = do_read_string(fd, ph);
1412 return ph->env.os_release ? 0 : -ENOMEM;
1415 static int process_version(struct perf_file_section *section __maybe_unused,
1416 struct perf_header *ph, int fd,
1417 void *data __maybe_unused)
1419 ph->env.version = do_read_string(fd, ph);
1420 return ph->env.version ? 0 : -ENOMEM;
1423 static int process_arch(struct perf_file_section *section __maybe_unused,
1424 struct perf_header *ph, int fd,
1425 void *data __maybe_unused)
1427 ph->env.arch = do_read_string(fd, ph);
1428 return ph->env.arch ? 0 : -ENOMEM;
1431 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1432 struct perf_header *ph, int fd,
1433 void *data __maybe_unused)
1438 ret = readn(fd, &nr, sizeof(nr));
1439 if (ret != sizeof(nr))
1445 ph->env.nr_cpus_online = nr;
1447 ret = readn(fd, &nr, sizeof(nr));
1448 if (ret != sizeof(nr))
1454 ph->env.nr_cpus_avail = nr;
1458 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1459 struct perf_header *ph, int fd,
1460 void *data __maybe_unused)
1462 ph->env.cpu_desc = do_read_string(fd, ph);
1463 return ph->env.cpu_desc ? 0 : -ENOMEM;
1466 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1467 struct perf_header *ph, int fd,
1468 void *data __maybe_unused)
1470 ph->env.cpuid = do_read_string(fd, ph);
1471 return ph->env.cpuid ? 0 : -ENOMEM;
1474 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1475 struct perf_header *ph, int fd,
1476 void *data __maybe_unused)
1481 ret = readn(fd, &mem, sizeof(mem));
1482 if (ret != sizeof(mem))
1486 mem = bswap_64(mem);
1488 ph->env.total_mem = mem;
1492 static struct perf_evsel *
1493 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1495 struct perf_evsel *evsel;
1497 evlist__for_each(evlist, evsel) {
1498 if (evsel->idx == idx)
1506 perf_evlist__set_event_name(struct perf_evlist *evlist,
1507 struct perf_evsel *event)
1509 struct perf_evsel *evsel;
1514 evsel = perf_evlist__find_by_index(evlist, event->idx);
1521 evsel->name = strdup(event->name);
1525 process_event_desc(struct perf_file_section *section __maybe_unused,
1526 struct perf_header *header, int fd,
1527 void *data __maybe_unused)
1529 struct perf_session *session;
1530 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1535 session = container_of(header, struct perf_session, header);
1536 for (evsel = events; evsel->attr.size; evsel++)
1537 perf_evlist__set_event_name(session->evlist, evsel);
1539 free_event_desc(events);
1544 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1545 struct perf_header *ph, int fd,
1546 void *data __maybe_unused)
1553 ret = readn(fd, &nr, sizeof(nr));
1554 if (ret != sizeof(nr))
1560 ph->env.nr_cmdline = nr;
1561 strbuf_init(&sb, 128);
1563 for (i = 0; i < nr; i++) {
1564 str = do_read_string(fd, ph);
1568 /* include a NULL character at the end */
1569 strbuf_add(&sb, str, strlen(str) + 1);
1572 ph->env.cmdline = strbuf_detach(&sb, NULL);
1576 strbuf_release(&sb);
1580 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1581 struct perf_header *ph, int fd,
1582 void *data __maybe_unused)
1589 ret = readn(fd, &nr, sizeof(nr));
1590 if (ret != sizeof(nr))
1596 ph->env.nr_sibling_cores = nr;
1597 strbuf_init(&sb, 128);
1599 for (i = 0; i < nr; i++) {
1600 str = do_read_string(fd, ph);
1604 /* include a NULL character at the end */
1605 strbuf_add(&sb, str, strlen(str) + 1);
1608 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1610 ret = readn(fd, &nr, sizeof(nr));
1611 if (ret != sizeof(nr))
1617 ph->env.nr_sibling_threads = nr;
1619 for (i = 0; i < nr; i++) {
1620 str = do_read_string(fd, ph);
1624 /* include a NULL character at the end */
1625 strbuf_add(&sb, str, strlen(str) + 1);
1628 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1632 strbuf_release(&sb);
1636 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1637 struct perf_header *ph, int fd,
1638 void *data __maybe_unused)
1643 uint64_t mem_total, mem_free;
1647 ret = readn(fd, &nr, sizeof(nr));
1648 if (ret != sizeof(nr))
1654 ph->env.nr_numa_nodes = nr;
1655 strbuf_init(&sb, 256);
1657 for (i = 0; i < nr; i++) {
1659 ret = readn(fd, &node, sizeof(node));
1660 if (ret != sizeof(node))
1663 ret = readn(fd, &mem_total, sizeof(u64));
1664 if (ret != sizeof(u64))
1667 ret = readn(fd, &mem_free, sizeof(u64));
1668 if (ret != sizeof(u64))
1671 if (ph->needs_swap) {
1672 node = bswap_32(node);
1673 mem_total = bswap_64(mem_total);
1674 mem_free = bswap_64(mem_free);
1677 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1678 node, mem_total, mem_free);
1680 str = do_read_string(fd, ph);
1684 /* include a NULL character at the end */
1685 strbuf_add(&sb, str, strlen(str) + 1);
1688 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1692 strbuf_release(&sb);
1696 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1697 struct perf_header *ph, int fd,
1698 void *data __maybe_unused)
1706 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1707 if (ret != sizeof(pmu_num))
1711 pmu_num = bswap_32(pmu_num);
1714 pr_debug("pmu mappings not available\n");
1718 ph->env.nr_pmu_mappings = pmu_num;
1719 strbuf_init(&sb, 128);
1722 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1725 type = bswap_32(type);
1727 name = do_read_string(fd, ph);
1731 strbuf_addf(&sb, "%u:%s", type, name);
1732 /* include a NULL character at the end */
1733 strbuf_add(&sb, "", 1);
1738 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1742 strbuf_release(&sb);
1746 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1747 struct perf_header *ph, int fd,
1748 void *data __maybe_unused)
1751 u32 i, nr, nr_groups;
1752 struct perf_session *session;
1753 struct perf_evsel *evsel, *leader = NULL;
1760 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1764 nr_groups = bswap_32(nr_groups);
1766 ph->env.nr_groups = nr_groups;
1768 pr_debug("group desc not available\n");
1772 desc = calloc(nr_groups, sizeof(*desc));
1776 for (i = 0; i < nr_groups; i++) {
1777 desc[i].name = do_read_string(fd, ph);
1781 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1784 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1787 if (ph->needs_swap) {
1788 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1789 desc[i].nr_members = bswap_32(desc[i].nr_members);
1794 * Rebuild group relationship based on the group_desc
1796 session = container_of(ph, struct perf_session, header);
1797 session->evlist->nr_groups = nr_groups;
1800 evlist__for_each(session->evlist, evsel) {
1801 if (evsel->idx == (int) desc[i].leader_idx) {
1802 evsel->leader = evsel;
1803 /* {anon_group} is a dummy name */
1804 if (strcmp(desc[i].name, "{anon_group}")) {
1805 evsel->group_name = desc[i].name;
1806 desc[i].name = NULL;
1808 evsel->nr_members = desc[i].nr_members;
1810 if (i >= nr_groups || nr > 0) {
1811 pr_debug("invalid group desc\n");
1816 nr = evsel->nr_members - 1;
1819 /* This is a group member */
1820 evsel->leader = leader;
1826 if (i != nr_groups || nr != 0) {
1827 pr_debug("invalid group desc\n");
1833 for (i = 0; i < nr_groups; i++)
1834 zfree(&desc[i].name);
1840 static int process_auxtrace(struct perf_file_section *section,
1841 struct perf_header *ph, int fd,
1842 void *data __maybe_unused)
1844 struct perf_session *session;
1847 session = container_of(ph, struct perf_session, header);
1849 err = auxtrace_index__process(fd, section->size, session,
1852 pr_err("Failed to process auxtrace index\n");
1856 struct feature_ops {
1857 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1858 void (*print)(struct perf_header *h, int fd, FILE *fp);
1859 int (*process)(struct perf_file_section *section,
1860 struct perf_header *h, int fd, void *data);
1865 #define FEAT_OPA(n, func) \
1866 [n] = { .name = #n, .write = write_##func, .print = print_##func }
1867 #define FEAT_OPP(n, func) \
1868 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1869 .process = process_##func }
1870 #define FEAT_OPF(n, func) \
1871 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1872 .process = process_##func, .full_only = true }
1874 /* feature_ops not implemented: */
1875 #define print_tracing_data NULL
1876 #define print_build_id NULL
1878 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1879 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
1880 FEAT_OPP(HEADER_BUILD_ID, build_id),
1881 FEAT_OPP(HEADER_HOSTNAME, hostname),
1882 FEAT_OPP(HEADER_OSRELEASE, osrelease),
1883 FEAT_OPP(HEADER_VERSION, version),
1884 FEAT_OPP(HEADER_ARCH, arch),
1885 FEAT_OPP(HEADER_NRCPUS, nrcpus),
1886 FEAT_OPP(HEADER_CPUDESC, cpudesc),
1887 FEAT_OPP(HEADER_CPUID, cpuid),
1888 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
1889 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
1890 FEAT_OPP(HEADER_CMDLINE, cmdline),
1891 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
1892 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
1893 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
1894 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
1895 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
1896 FEAT_OPP(HEADER_AUXTRACE, auxtrace),
1899 struct header_print_data {
1901 bool full; /* extended list of headers */
1904 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1905 struct perf_header *ph,
1906 int feat, int fd, void *data)
1908 struct header_print_data *hd = data;
1910 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1911 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1912 "%d, continuing...\n", section->offset, feat);
1915 if (feat >= HEADER_LAST_FEATURE) {
1916 pr_warning("unknown feature %d\n", feat);
1919 if (!feat_ops[feat].print)
1922 if (!feat_ops[feat].full_only || hd->full)
1923 feat_ops[feat].print(ph, fd, hd->fp);
1925 fprintf(hd->fp, "# %s info available, use -I to display\n",
1926 feat_ops[feat].name);
1931 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1933 struct header_print_data hd;
1934 struct perf_header *header = &session->header;
1935 int fd = perf_data_file__fd(session->file);
1939 perf_header__process_sections(header, fd, &hd,
1940 perf_file_section__fprintf_info);
1944 static int do_write_feat(int fd, struct perf_header *h, int type,
1945 struct perf_file_section **p,
1946 struct perf_evlist *evlist)
1951 if (perf_header__has_feat(h, type)) {
1952 if (!feat_ops[type].write)
1955 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1957 err = feat_ops[type].write(fd, h, evlist);
1959 pr_debug("failed to write feature %d\n", type);
1961 /* undo anything written */
1962 lseek(fd, (*p)->offset, SEEK_SET);
1966 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1972 static int perf_header__adds_write(struct perf_header *header,
1973 struct perf_evlist *evlist, int fd)
1976 struct perf_file_section *feat_sec, *p;
1982 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1986 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1987 if (feat_sec == NULL)
1990 sec_size = sizeof(*feat_sec) * nr_sections;
1992 sec_start = header->feat_offset;
1993 lseek(fd, sec_start + sec_size, SEEK_SET);
1995 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
1996 if (do_write_feat(fd, header, feat, &p, evlist))
1997 perf_header__clear_feat(header, feat);
2000 lseek(fd, sec_start, SEEK_SET);
2002 * may write more than needed due to dropped feature, but
2003 * this is okay, reader will skip the mising entries
2005 err = do_write(fd, feat_sec, sec_size);
2007 pr_debug("failed to write feature section\n");
2012 int perf_header__write_pipe(int fd)
2014 struct perf_pipe_file_header f_header;
2017 f_header = (struct perf_pipe_file_header){
2018 .magic = PERF_MAGIC,
2019 .size = sizeof(f_header),
2022 err = do_write(fd, &f_header, sizeof(f_header));
2024 pr_debug("failed to write perf pipe header\n");
2031 int perf_session__write_header(struct perf_session *session,
2032 struct perf_evlist *evlist,
2033 int fd, bool at_exit)
2035 struct perf_file_header f_header;
2036 struct perf_file_attr f_attr;
2037 struct perf_header *header = &session->header;
2038 struct perf_evsel *evsel;
2042 lseek(fd, sizeof(f_header), SEEK_SET);
2044 evlist__for_each(session->evlist, evsel) {
2045 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2046 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2048 pr_debug("failed to write perf header\n");
2053 attr_offset = lseek(fd, 0, SEEK_CUR);
2055 evlist__for_each(evlist, evsel) {
2056 f_attr = (struct perf_file_attr){
2057 .attr = evsel->attr,
2059 .offset = evsel->id_offset,
2060 .size = evsel->ids * sizeof(u64),
2063 err = do_write(fd, &f_attr, sizeof(f_attr));
2065 pr_debug("failed to write perf header attribute\n");
2070 if (!header->data_offset)
2071 header->data_offset = lseek(fd, 0, SEEK_CUR);
2072 header->feat_offset = header->data_offset + header->data_size;
2075 err = perf_header__adds_write(header, evlist, fd);
2080 f_header = (struct perf_file_header){
2081 .magic = PERF_MAGIC,
2082 .size = sizeof(f_header),
2083 .attr_size = sizeof(f_attr),
2085 .offset = attr_offset,
2086 .size = evlist->nr_entries * sizeof(f_attr),
2089 .offset = header->data_offset,
2090 .size = header->data_size,
2092 /* event_types is ignored, store zeros */
2095 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2097 lseek(fd, 0, SEEK_SET);
2098 err = do_write(fd, &f_header, sizeof(f_header));
2100 pr_debug("failed to write perf header\n");
2103 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2108 static int perf_header__getbuffer64(struct perf_header *header,
2109 int fd, void *buf, size_t size)
2111 if (readn(fd, buf, size) <= 0)
2114 if (header->needs_swap)
2115 mem_bswap_64(buf, size);
2120 int perf_header__process_sections(struct perf_header *header, int fd,
2122 int (*process)(struct perf_file_section *section,
2123 struct perf_header *ph,
2124 int feat, int fd, void *data))
2126 struct perf_file_section *feat_sec, *sec;
2132 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2136 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2140 sec_size = sizeof(*feat_sec) * nr_sections;
2142 lseek(fd, header->feat_offset, SEEK_SET);
2144 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2148 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2149 err = process(sec++, header, feat, fd, data);
2159 static const int attr_file_abi_sizes[] = {
2160 [0] = PERF_ATTR_SIZE_VER0,
2161 [1] = PERF_ATTR_SIZE_VER1,
2162 [2] = PERF_ATTR_SIZE_VER2,
2163 [3] = PERF_ATTR_SIZE_VER3,
2164 [4] = PERF_ATTR_SIZE_VER4,
2169 * In the legacy file format, the magic number is not used to encode endianness.
2170 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2171 * on ABI revisions, we need to try all combinations for all endianness to
2172 * detect the endianness.
2174 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2176 uint64_t ref_size, attr_size;
2179 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2180 ref_size = attr_file_abi_sizes[i]
2181 + sizeof(struct perf_file_section);
2182 if (hdr_sz != ref_size) {
2183 attr_size = bswap_64(hdr_sz);
2184 if (attr_size != ref_size)
2187 ph->needs_swap = true;
2189 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2194 /* could not determine endianness */
2198 #define PERF_PIPE_HDR_VER0 16
2200 static const size_t attr_pipe_abi_sizes[] = {
2201 [0] = PERF_PIPE_HDR_VER0,
2206 * In the legacy pipe format, there is an implicit assumption that endiannesss
2207 * between host recording the samples, and host parsing the samples is the
2208 * same. This is not always the case given that the pipe output may always be
2209 * redirected into a file and analyzed on a different machine with possibly a
2210 * different endianness and perf_event ABI revsions in the perf tool itself.
2212 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2217 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2218 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2219 attr_size = bswap_64(hdr_sz);
2220 if (attr_size != hdr_sz)
2223 ph->needs_swap = true;
2225 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2231 bool is_perf_magic(u64 magic)
2233 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2234 || magic == __perf_magic2
2235 || magic == __perf_magic2_sw)
2241 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2242 bool is_pipe, struct perf_header *ph)
2246 /* check for legacy format */
2247 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2249 ph->version = PERF_HEADER_VERSION_1;
2250 pr_debug("legacy perf.data format\n");
2252 return try_all_pipe_abis(hdr_sz, ph);
2254 return try_all_file_abis(hdr_sz, ph);
2257 * the new magic number serves two purposes:
2258 * - unique number to identify actual perf.data files
2259 * - encode endianness of file
2261 ph->version = PERF_HEADER_VERSION_2;
2263 /* check magic number with one endianness */
2264 if (magic == __perf_magic2)
2267 /* check magic number with opposite endianness */
2268 if (magic != __perf_magic2_sw)
2271 ph->needs_swap = true;
2276 int perf_file_header__read(struct perf_file_header *header,
2277 struct perf_header *ph, int fd)
2281 lseek(fd, 0, SEEK_SET);
2283 ret = readn(fd, header, sizeof(*header));
2287 if (check_magic_endian(header->magic,
2288 header->attr_size, false, ph) < 0) {
2289 pr_debug("magic/endian check failed\n");
2293 if (ph->needs_swap) {
2294 mem_bswap_64(header, offsetof(struct perf_file_header,
2298 if (header->size != sizeof(*header)) {
2299 /* Support the previous format */
2300 if (header->size == offsetof(typeof(*header), adds_features))
2301 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2304 } else if (ph->needs_swap) {
2306 * feature bitmap is declared as an array of unsigned longs --
2307 * not good since its size can differ between the host that
2308 * generated the data file and the host analyzing the file.
2310 * We need to handle endianness, but we don't know the size of
2311 * the unsigned long where the file was generated. Take a best
2312 * guess at determining it: try 64-bit swap first (ie., file
2313 * created on a 64-bit host), and check if the hostname feature
2314 * bit is set (this feature bit is forced on as of fbe96f2).
2315 * If the bit is not, undo the 64-bit swap and try a 32-bit
2316 * swap. If the hostname bit is still not set (e.g., older data
2317 * file), punt and fallback to the original behavior --
2318 * clearing all feature bits and setting buildid.
2320 mem_bswap_64(&header->adds_features,
2321 BITS_TO_U64(HEADER_FEAT_BITS));
2323 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2325 mem_bswap_64(&header->adds_features,
2326 BITS_TO_U64(HEADER_FEAT_BITS));
2329 mem_bswap_32(&header->adds_features,
2330 BITS_TO_U32(HEADER_FEAT_BITS));
2333 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2334 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2335 set_bit(HEADER_BUILD_ID, header->adds_features);
2339 memcpy(&ph->adds_features, &header->adds_features,
2340 sizeof(ph->adds_features));
2342 ph->data_offset = header->data.offset;
2343 ph->data_size = header->data.size;
2344 ph->feat_offset = header->data.offset + header->data.size;
2348 static int perf_file_section__process(struct perf_file_section *section,
2349 struct perf_header *ph,
2350 int feat, int fd, void *data)
2352 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2353 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2354 "%d, continuing...\n", section->offset, feat);
2358 if (feat >= HEADER_LAST_FEATURE) {
2359 pr_debug("unknown feature %d, continuing...\n", feat);
2363 if (!feat_ops[feat].process)
2366 return feat_ops[feat].process(section, ph, fd, data);
2369 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2370 struct perf_header *ph, int fd,
2375 ret = readn(fd, header, sizeof(*header));
2379 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2380 pr_debug("endian/magic failed\n");
2385 header->size = bswap_64(header->size);
2387 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2393 static int perf_header__read_pipe(struct perf_session *session)
2395 struct perf_header *header = &session->header;
2396 struct perf_pipe_file_header f_header;
2398 if (perf_file_header__read_pipe(&f_header, header,
2399 perf_data_file__fd(session->file),
2400 session->repipe) < 0) {
2401 pr_debug("incompatible file format\n");
2408 static int read_attr(int fd, struct perf_header *ph,
2409 struct perf_file_attr *f_attr)
2411 struct perf_event_attr *attr = &f_attr->attr;
2413 size_t our_sz = sizeof(f_attr->attr);
2416 memset(f_attr, 0, sizeof(*f_attr));
2418 /* read minimal guaranteed structure */
2419 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2421 pr_debug("cannot read %d bytes of header attr\n",
2422 PERF_ATTR_SIZE_VER0);
2426 /* on file perf_event_attr size */
2434 sz = PERF_ATTR_SIZE_VER0;
2435 } else if (sz > our_sz) {
2436 pr_debug("file uses a more recent and unsupported ABI"
2437 " (%zu bytes extra)\n", sz - our_sz);
2440 /* what we have not yet read and that we know about */
2441 left = sz - PERF_ATTR_SIZE_VER0;
2444 ptr += PERF_ATTR_SIZE_VER0;
2446 ret = readn(fd, ptr, left);
2448 /* read perf_file_section, ids are read in caller */
2449 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2451 return ret <= 0 ? -1 : 0;
2454 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2455 struct pevent *pevent)
2457 struct event_format *event;
2460 /* already prepared */
2461 if (evsel->tp_format)
2464 if (pevent == NULL) {
2465 pr_debug("broken or missing trace data\n");
2469 event = pevent_find_event(pevent, evsel->attr.config);
2474 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2475 evsel->name = strdup(bf);
2476 if (evsel->name == NULL)
2480 evsel->tp_format = event;
2484 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2485 struct pevent *pevent)
2487 struct perf_evsel *pos;
2489 evlist__for_each(evlist, pos) {
2490 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2491 perf_evsel__prepare_tracepoint_event(pos, pevent))
2498 int perf_session__read_header(struct perf_session *session)
2500 struct perf_data_file *file = session->file;
2501 struct perf_header *header = &session->header;
2502 struct perf_file_header f_header;
2503 struct perf_file_attr f_attr;
2505 int nr_attrs, nr_ids, i, j;
2506 int fd = perf_data_file__fd(file);
2508 session->evlist = perf_evlist__new();
2509 if (session->evlist == NULL)
2512 if (perf_data_file__is_pipe(file))
2513 return perf_header__read_pipe(session);
2515 if (perf_file_header__read(&f_header, header, fd) < 0)
2519 * Sanity check that perf.data was written cleanly; data size is
2520 * initialized to 0 and updated only if the on_exit function is run.
2521 * If data size is still 0 then the file contains only partial
2522 * information. Just warn user and process it as much as it can.
2524 if (f_header.data.size == 0) {
2525 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2526 "Was the 'perf record' command properly terminated?\n",
2530 nr_attrs = f_header.attrs.size / f_header.attr_size;
2531 lseek(fd, f_header.attrs.offset, SEEK_SET);
2533 for (i = 0; i < nr_attrs; i++) {
2534 struct perf_evsel *evsel;
2537 if (read_attr(fd, header, &f_attr) < 0)
2540 if (header->needs_swap) {
2541 f_attr.ids.size = bswap_64(f_attr.ids.size);
2542 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2543 perf_event__attr_swap(&f_attr.attr);
2546 tmp = lseek(fd, 0, SEEK_CUR);
2547 evsel = perf_evsel__new(&f_attr.attr);
2550 goto out_delete_evlist;
2552 evsel->needs_swap = header->needs_swap;
2554 * Do it before so that if perf_evsel__alloc_id fails, this
2555 * entry gets purged too at perf_evlist__delete().
2557 perf_evlist__add(session->evlist, evsel);
2559 nr_ids = f_attr.ids.size / sizeof(u64);
2561 * We don't have the cpu and thread maps on the header, so
2562 * for allocating the perf_sample_id table we fake 1 cpu and
2563 * hattr->ids threads.
2565 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2566 goto out_delete_evlist;
2568 lseek(fd, f_attr.ids.offset, SEEK_SET);
2570 for (j = 0; j < nr_ids; j++) {
2571 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2574 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2577 lseek(fd, tmp, SEEK_SET);
2580 symbol_conf.nr_events = nr_attrs;
2582 perf_header__process_sections(header, fd, &session->tevent,
2583 perf_file_section__process);
2585 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2586 session->tevent.pevent))
2587 goto out_delete_evlist;
2594 perf_evlist__delete(session->evlist);
2595 session->evlist = NULL;
2599 int perf_event__synthesize_attr(struct perf_tool *tool,
2600 struct perf_event_attr *attr, u32 ids, u64 *id,
2601 perf_event__handler_t process)
2603 union perf_event *ev;
2607 size = sizeof(struct perf_event_attr);
2608 size = PERF_ALIGN(size, sizeof(u64));
2609 size += sizeof(struct perf_event_header);
2610 size += ids * sizeof(u64);
2617 ev->attr.attr = *attr;
2618 memcpy(ev->attr.id, id, ids * sizeof(u64));
2620 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2621 ev->attr.header.size = (u16)size;
2623 if (ev->attr.header.size == size)
2624 err = process(tool, ev, NULL, NULL);
2633 int perf_event__synthesize_attrs(struct perf_tool *tool,
2634 struct perf_session *session,
2635 perf_event__handler_t process)
2637 struct perf_evsel *evsel;
2640 evlist__for_each(session->evlist, evsel) {
2641 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2642 evsel->id, process);
2644 pr_debug("failed to create perf header attribute\n");
2652 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2653 union perf_event *event,
2654 struct perf_evlist **pevlist)
2657 struct perf_evsel *evsel;
2658 struct perf_evlist *evlist = *pevlist;
2660 if (evlist == NULL) {
2661 *pevlist = evlist = perf_evlist__new();
2666 evsel = perf_evsel__new(&event->attr.attr);
2670 perf_evlist__add(evlist, evsel);
2672 ids = event->header.size;
2673 ids -= (void *)&event->attr.id - (void *)event;
2674 n_ids = ids / sizeof(u64);
2676 * We don't have the cpu and thread maps on the header, so
2677 * for allocating the perf_sample_id table we fake 1 cpu and
2678 * hattr->ids threads.
2680 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2683 for (i = 0; i < n_ids; i++) {
2684 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2687 symbol_conf.nr_events = evlist->nr_entries;
2692 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2693 struct perf_evlist *evlist,
2694 perf_event__handler_t process)
2696 union perf_event ev;
2697 struct tracing_data *tdata;
2698 ssize_t size = 0, aligned_size = 0, padding;
2699 int err __maybe_unused = 0;
2702 * We are going to store the size of the data followed
2703 * by the data contents. Since the fd descriptor is a pipe,
2704 * we cannot seek back to store the size of the data once
2705 * we know it. Instead we:
2707 * - write the tracing data to the temp file
2708 * - get/write the data size to pipe
2709 * - write the tracing data from the temp file
2712 tdata = tracing_data_get(&evlist->entries, fd, true);
2716 memset(&ev, 0, sizeof(ev));
2718 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2720 aligned_size = PERF_ALIGN(size, sizeof(u64));
2721 padding = aligned_size - size;
2722 ev.tracing_data.header.size = sizeof(ev.tracing_data);
2723 ev.tracing_data.size = aligned_size;
2725 process(tool, &ev, NULL, NULL);
2728 * The put function will copy all the tracing data
2729 * stored in temp file to the pipe.
2731 tracing_data_put(tdata);
2733 write_padded(fd, NULL, 0, padding);
2735 return aligned_size;
2738 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2739 union perf_event *event,
2740 struct perf_session *session)
2742 ssize_t size_read, padding, size = event->tracing_data.size;
2743 int fd = perf_data_file__fd(session->file);
2744 off_t offset = lseek(fd, 0, SEEK_CUR);
2747 /* setup for reading amidst mmap */
2748 lseek(fd, offset + sizeof(struct tracing_data_event),
2751 size_read = trace_report(fd, &session->tevent,
2753 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2755 if (readn(fd, buf, padding) < 0) {
2756 pr_err("%s: reading input file", __func__);
2759 if (session->repipe) {
2760 int retw = write(STDOUT_FILENO, buf, padding);
2761 if (retw <= 0 || retw != padding) {
2762 pr_err("%s: repiping tracing data padding", __func__);
2767 if (size_read + padding != size) {
2768 pr_err("%s: tracing data size mismatch", __func__);
2772 perf_evlist__prepare_tracepoint_events(session->evlist,
2773 session->tevent.pevent);
2775 return size_read + padding;
2778 int perf_event__synthesize_build_id(struct perf_tool *tool,
2779 struct dso *pos, u16 misc,
2780 perf_event__handler_t process,
2781 struct machine *machine)
2783 union perf_event ev;
2790 memset(&ev, 0, sizeof(ev));
2792 len = pos->long_name_len + 1;
2793 len = PERF_ALIGN(len, NAME_ALIGN);
2794 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2795 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2796 ev.build_id.header.misc = misc;
2797 ev.build_id.pid = machine->pid;
2798 ev.build_id.header.size = sizeof(ev.build_id) + len;
2799 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2801 err = process(tool, &ev, NULL, machine);
2806 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2807 union perf_event *event,
2808 struct perf_session *session)
2810 __event_process_build_id(&event->build_id,
2811 event->build_id.filename,