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,
927 nr = ph->env.nr_cmdline;
929 fprintf(fp, "# cmdline : ");
931 for (i = 0; i < nr; i++)
932 fprintf(fp, "%s ", ph->env.cmdline_argv[i]);
936 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
942 nr = ph->env.nr_sibling_cores;
943 str = ph->env.sibling_cores;
945 for (i = 0; i < nr; i++) {
946 fprintf(fp, "# sibling cores : %s\n", str);
947 str += strlen(str) + 1;
950 nr = ph->env.nr_sibling_threads;
951 str = ph->env.sibling_threads;
953 for (i = 0; i < nr; i++) {
954 fprintf(fp, "# sibling threads : %s\n", str);
955 str += strlen(str) + 1;
959 static void free_event_desc(struct perf_evsel *events)
961 struct perf_evsel *evsel;
966 for (evsel = events; evsel->attr.size; evsel++) {
974 static struct perf_evsel *
975 read_event_desc(struct perf_header *ph, int fd)
977 struct perf_evsel *evsel, *events = NULL;
980 u32 nre, sz, nr, i, j;
984 /* number of events */
985 ret = readn(fd, &nre, sizeof(nre));
986 if (ret != (ssize_t)sizeof(nre))
992 ret = readn(fd, &sz, sizeof(sz));
993 if (ret != (ssize_t)sizeof(sz))
999 /* buffer to hold on file attr struct */
1004 /* the last event terminates with evsel->attr.size == 0: */
1005 events = calloc(nre + 1, sizeof(*events));
1009 msz = sizeof(evsel->attr);
1013 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1017 * must read entire on-file attr struct to
1018 * sync up with layout.
1020 ret = readn(fd, buf, sz);
1021 if (ret != (ssize_t)sz)
1025 perf_event__attr_swap(buf);
1027 memcpy(&evsel->attr, buf, msz);
1029 ret = readn(fd, &nr, sizeof(nr));
1030 if (ret != (ssize_t)sizeof(nr))
1033 if (ph->needs_swap) {
1035 evsel->needs_swap = true;
1038 evsel->name = do_read_string(fd, ph);
1043 id = calloc(nr, sizeof(*id));
1049 for (j = 0 ; j < nr; j++) {
1050 ret = readn(fd, id, sizeof(*id));
1051 if (ret != (ssize_t)sizeof(*id))
1054 *id = bswap_64(*id);
1062 free_event_desc(events);
1067 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1068 void *priv __attribute__((unused)))
1070 return fprintf(fp, ", %s = %s", name, val);
1073 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1075 struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1080 fprintf(fp, "# event desc: not available or unable to read\n");
1084 for (evsel = events; evsel->attr.size; evsel++) {
1085 fprintf(fp, "# event : name = %s, ", evsel->name);
1088 fprintf(fp, ", id = {");
1089 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1092 fprintf(fp, " %"PRIu64, *id);
1097 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1102 free_event_desc(events);
1105 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1108 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1111 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1116 uint64_t mem_total, mem_free;
1119 nr = ph->env.nr_numa_nodes;
1120 str = ph->env.numa_nodes;
1122 for (i = 0; i < nr; i++) {
1124 c = strtoul(str, &tmp, 0);
1129 mem_total = strtoull(str, &tmp, 0);
1134 mem_free = strtoull(str, &tmp, 0);
1138 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1139 " free = %"PRIu64" kB\n",
1140 c, mem_total, mem_free);
1143 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1145 str += strlen(str) + 1;
1149 fprintf(fp, "# numa topology : not available\n");
1152 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1154 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1157 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1158 int fd __maybe_unused, FILE *fp)
1160 fprintf(fp, "# contains samples with branch stack\n");
1163 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1164 int fd __maybe_unused, FILE *fp)
1166 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1169 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1172 const char *delimiter = "# pmu mappings: ";
1177 pmu_num = ph->env.nr_pmu_mappings;
1179 fprintf(fp, "# pmu mappings: not available\n");
1183 str = ph->env.pmu_mappings;
1186 type = strtoul(str, &tmp, 0);
1191 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1194 str += strlen(str) + 1;
1203 fprintf(fp, "# pmu mappings: unable to read\n");
1206 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1209 struct perf_session *session;
1210 struct perf_evsel *evsel;
1213 session = container_of(ph, struct perf_session, header);
1215 evlist__for_each(session->evlist, evsel) {
1216 if (perf_evsel__is_group_leader(evsel) &&
1217 evsel->nr_members > 1) {
1218 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1219 perf_evsel__name(evsel));
1221 nr = evsel->nr_members - 1;
1223 fprintf(fp, ",%s", perf_evsel__name(evsel));
1231 static int __event_process_build_id(struct build_id_event *bev,
1233 struct perf_session *session)
1236 struct machine *machine;
1239 enum dso_kernel_type dso_type;
1241 machine = perf_session__findnew_machine(session, bev->pid);
1245 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1248 case PERF_RECORD_MISC_KERNEL:
1249 dso_type = DSO_TYPE_KERNEL;
1251 case PERF_RECORD_MISC_GUEST_KERNEL:
1252 dso_type = DSO_TYPE_GUEST_KERNEL;
1254 case PERF_RECORD_MISC_USER:
1255 case PERF_RECORD_MISC_GUEST_USER:
1256 dso_type = DSO_TYPE_USER;
1262 dso = machine__findnew_dso(machine, filename);
1264 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1266 dso__set_build_id(dso, &bev->build_id);
1268 if (!is_kernel_module(filename, cpumode))
1269 dso->kernel = dso_type;
1271 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1273 pr_debug("build id event received for %s: %s\n",
1274 dso->long_name, sbuild_id);
1283 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1284 int input, u64 offset, u64 size)
1286 struct perf_session *session = container_of(header, struct perf_session, header);
1288 struct perf_event_header header;
1289 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1292 struct build_id_event bev;
1293 char filename[PATH_MAX];
1294 u64 limit = offset + size;
1296 while (offset < limit) {
1299 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1302 if (header->needs_swap)
1303 perf_event_header__bswap(&old_bev.header);
1305 len = old_bev.header.size - sizeof(old_bev);
1306 if (readn(input, filename, len) != len)
1309 bev.header = old_bev.header;
1312 * As the pid is the missing value, we need to fill
1313 * it properly. The header.misc value give us nice hint.
1315 bev.pid = HOST_KERNEL_ID;
1316 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1317 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1318 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1320 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1321 __event_process_build_id(&bev, filename, session);
1323 offset += bev.header.size;
1329 static int perf_header__read_build_ids(struct perf_header *header,
1330 int input, u64 offset, u64 size)
1332 struct perf_session *session = container_of(header, struct perf_session, header);
1333 struct build_id_event bev;
1334 char filename[PATH_MAX];
1335 u64 limit = offset + size, orig_offset = offset;
1338 while (offset < limit) {
1341 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1344 if (header->needs_swap)
1345 perf_event_header__bswap(&bev.header);
1347 len = bev.header.size - sizeof(bev);
1348 if (readn(input, filename, len) != len)
1351 * The a1645ce1 changeset:
1353 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1355 * Added a field to struct build_id_event that broke the file
1358 * Since the kernel build-id is the first entry, process the
1359 * table using the old format if the well known
1360 * '[kernel.kallsyms]' string for the kernel build-id has the
1361 * first 4 characters chopped off (where the pid_t sits).
1363 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1364 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1366 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1369 __event_process_build_id(&bev, filename, session);
1371 offset += bev.header.size;
1378 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1379 struct perf_header *ph __maybe_unused,
1382 ssize_t ret = trace_report(fd, data, false);
1383 return ret < 0 ? -1 : 0;
1386 static int process_build_id(struct perf_file_section *section,
1387 struct perf_header *ph, int fd,
1388 void *data __maybe_unused)
1390 if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1391 pr_debug("Failed to read buildids, continuing...\n");
1395 static int process_hostname(struct perf_file_section *section __maybe_unused,
1396 struct perf_header *ph, int fd,
1397 void *data __maybe_unused)
1399 ph->env.hostname = do_read_string(fd, ph);
1400 return ph->env.hostname ? 0 : -ENOMEM;
1403 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1404 struct perf_header *ph, int fd,
1405 void *data __maybe_unused)
1407 ph->env.os_release = do_read_string(fd, ph);
1408 return ph->env.os_release ? 0 : -ENOMEM;
1411 static int process_version(struct perf_file_section *section __maybe_unused,
1412 struct perf_header *ph, int fd,
1413 void *data __maybe_unused)
1415 ph->env.version = do_read_string(fd, ph);
1416 return ph->env.version ? 0 : -ENOMEM;
1419 static int process_arch(struct perf_file_section *section __maybe_unused,
1420 struct perf_header *ph, int fd,
1421 void *data __maybe_unused)
1423 ph->env.arch = do_read_string(fd, ph);
1424 return ph->env.arch ? 0 : -ENOMEM;
1427 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1428 struct perf_header *ph, int fd,
1429 void *data __maybe_unused)
1434 ret = readn(fd, &nr, sizeof(nr));
1435 if (ret != sizeof(nr))
1441 ph->env.nr_cpus_avail = nr;
1443 ret = readn(fd, &nr, sizeof(nr));
1444 if (ret != sizeof(nr))
1450 ph->env.nr_cpus_online = nr;
1454 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1455 struct perf_header *ph, int fd,
1456 void *data __maybe_unused)
1458 ph->env.cpu_desc = do_read_string(fd, ph);
1459 return ph->env.cpu_desc ? 0 : -ENOMEM;
1462 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1463 struct perf_header *ph, int fd,
1464 void *data __maybe_unused)
1466 ph->env.cpuid = do_read_string(fd, ph);
1467 return ph->env.cpuid ? 0 : -ENOMEM;
1470 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1471 struct perf_header *ph, int fd,
1472 void *data __maybe_unused)
1477 ret = readn(fd, &mem, sizeof(mem));
1478 if (ret != sizeof(mem))
1482 mem = bswap_64(mem);
1484 ph->env.total_mem = mem;
1488 static struct perf_evsel *
1489 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1491 struct perf_evsel *evsel;
1493 evlist__for_each(evlist, evsel) {
1494 if (evsel->idx == idx)
1502 perf_evlist__set_event_name(struct perf_evlist *evlist,
1503 struct perf_evsel *event)
1505 struct perf_evsel *evsel;
1510 evsel = perf_evlist__find_by_index(evlist, event->idx);
1517 evsel->name = strdup(event->name);
1521 process_event_desc(struct perf_file_section *section __maybe_unused,
1522 struct perf_header *header, int fd,
1523 void *data __maybe_unused)
1525 struct perf_session *session;
1526 struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1531 session = container_of(header, struct perf_session, header);
1532 for (evsel = events; evsel->attr.size; evsel++)
1533 perf_evlist__set_event_name(session->evlist, evsel);
1535 free_event_desc(events);
1540 static int process_cmdline(struct perf_file_section *section,
1541 struct perf_header *ph, int fd,
1542 void *data __maybe_unused)
1545 char *str, *cmdline = NULL, **argv = NULL;
1548 ret = readn(fd, &nr, sizeof(nr));
1549 if (ret != sizeof(nr))
1555 ph->env.nr_cmdline = nr;
1557 cmdline = zalloc(section->size + nr + 1);
1561 argv = zalloc(sizeof(char *) * (nr + 1));
1565 for (i = 0; i < nr; i++) {
1566 str = do_read_string(fd, ph);
1570 argv[i] = cmdline + len;
1571 memcpy(argv[i], str, strlen(str) + 1);
1572 len += strlen(str) + 1;
1575 ph->env.cmdline = cmdline;
1576 ph->env.cmdline_argv = (const char **) argv;
1585 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1586 struct perf_header *ph, int fd,
1587 void *data __maybe_unused)
1594 ret = readn(fd, &nr, sizeof(nr));
1595 if (ret != sizeof(nr))
1601 ph->env.nr_sibling_cores = nr;
1602 strbuf_init(&sb, 128);
1604 for (i = 0; i < nr; i++) {
1605 str = do_read_string(fd, ph);
1609 /* include a NULL character at the end */
1610 strbuf_add(&sb, str, strlen(str) + 1);
1613 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1615 ret = readn(fd, &nr, sizeof(nr));
1616 if (ret != sizeof(nr))
1622 ph->env.nr_sibling_threads = nr;
1624 for (i = 0; i < nr; i++) {
1625 str = do_read_string(fd, ph);
1629 /* include a NULL character at the end */
1630 strbuf_add(&sb, str, strlen(str) + 1);
1633 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1637 strbuf_release(&sb);
1641 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1642 struct perf_header *ph, int fd,
1643 void *data __maybe_unused)
1648 uint64_t mem_total, mem_free;
1652 ret = readn(fd, &nr, sizeof(nr));
1653 if (ret != sizeof(nr))
1659 ph->env.nr_numa_nodes = nr;
1660 strbuf_init(&sb, 256);
1662 for (i = 0; i < nr; i++) {
1664 ret = readn(fd, &node, sizeof(node));
1665 if (ret != sizeof(node))
1668 ret = readn(fd, &mem_total, sizeof(u64));
1669 if (ret != sizeof(u64))
1672 ret = readn(fd, &mem_free, sizeof(u64));
1673 if (ret != sizeof(u64))
1676 if (ph->needs_swap) {
1677 node = bswap_32(node);
1678 mem_total = bswap_64(mem_total);
1679 mem_free = bswap_64(mem_free);
1682 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1683 node, mem_total, mem_free);
1685 str = do_read_string(fd, ph);
1689 /* include a NULL character at the end */
1690 strbuf_add(&sb, str, strlen(str) + 1);
1693 ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1697 strbuf_release(&sb);
1701 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1702 struct perf_header *ph, int fd,
1703 void *data __maybe_unused)
1711 ret = readn(fd, &pmu_num, sizeof(pmu_num));
1712 if (ret != sizeof(pmu_num))
1716 pmu_num = bswap_32(pmu_num);
1719 pr_debug("pmu mappings not available\n");
1723 ph->env.nr_pmu_mappings = pmu_num;
1724 strbuf_init(&sb, 128);
1727 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1730 type = bswap_32(type);
1732 name = do_read_string(fd, ph);
1736 strbuf_addf(&sb, "%u:%s", type, name);
1737 /* include a NULL character at the end */
1738 strbuf_add(&sb, "", 1);
1743 ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1747 strbuf_release(&sb);
1751 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1752 struct perf_header *ph, int fd,
1753 void *data __maybe_unused)
1756 u32 i, nr, nr_groups;
1757 struct perf_session *session;
1758 struct perf_evsel *evsel, *leader = NULL;
1765 if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1769 nr_groups = bswap_32(nr_groups);
1771 ph->env.nr_groups = nr_groups;
1773 pr_debug("group desc not available\n");
1777 desc = calloc(nr_groups, sizeof(*desc));
1781 for (i = 0; i < nr_groups; i++) {
1782 desc[i].name = do_read_string(fd, ph);
1786 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1789 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1792 if (ph->needs_swap) {
1793 desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1794 desc[i].nr_members = bswap_32(desc[i].nr_members);
1799 * Rebuild group relationship based on the group_desc
1801 session = container_of(ph, struct perf_session, header);
1802 session->evlist->nr_groups = nr_groups;
1805 evlist__for_each(session->evlist, evsel) {
1806 if (evsel->idx == (int) desc[i].leader_idx) {
1807 evsel->leader = evsel;
1808 /* {anon_group} is a dummy name */
1809 if (strcmp(desc[i].name, "{anon_group}")) {
1810 evsel->group_name = desc[i].name;
1811 desc[i].name = NULL;
1813 evsel->nr_members = desc[i].nr_members;
1815 if (i >= nr_groups || nr > 0) {
1816 pr_debug("invalid group desc\n");
1821 nr = evsel->nr_members - 1;
1824 /* This is a group member */
1825 evsel->leader = leader;
1831 if (i != nr_groups || nr != 0) {
1832 pr_debug("invalid group desc\n");
1838 for (i = 0; i < nr_groups; i++)
1839 zfree(&desc[i].name);
1845 static int process_auxtrace(struct perf_file_section *section,
1846 struct perf_header *ph, int fd,
1847 void *data __maybe_unused)
1849 struct perf_session *session;
1852 session = container_of(ph, struct perf_session, header);
1854 err = auxtrace_index__process(fd, section->size, session,
1857 pr_err("Failed to process auxtrace index\n");
1861 struct feature_ops {
1862 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1863 void (*print)(struct perf_header *h, int fd, FILE *fp);
1864 int (*process)(struct perf_file_section *section,
1865 struct perf_header *h, int fd, void *data);
1870 #define FEAT_OPA(n, func) \
1871 [n] = { .name = #n, .write = write_##func, .print = print_##func }
1872 #define FEAT_OPP(n, func) \
1873 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1874 .process = process_##func }
1875 #define FEAT_OPF(n, func) \
1876 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1877 .process = process_##func, .full_only = true }
1879 /* feature_ops not implemented: */
1880 #define print_tracing_data NULL
1881 #define print_build_id NULL
1883 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1884 FEAT_OPP(HEADER_TRACING_DATA, tracing_data),
1885 FEAT_OPP(HEADER_BUILD_ID, build_id),
1886 FEAT_OPP(HEADER_HOSTNAME, hostname),
1887 FEAT_OPP(HEADER_OSRELEASE, osrelease),
1888 FEAT_OPP(HEADER_VERSION, version),
1889 FEAT_OPP(HEADER_ARCH, arch),
1890 FEAT_OPP(HEADER_NRCPUS, nrcpus),
1891 FEAT_OPP(HEADER_CPUDESC, cpudesc),
1892 FEAT_OPP(HEADER_CPUID, cpuid),
1893 FEAT_OPP(HEADER_TOTAL_MEM, total_mem),
1894 FEAT_OPP(HEADER_EVENT_DESC, event_desc),
1895 FEAT_OPP(HEADER_CMDLINE, cmdline),
1896 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology),
1897 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology),
1898 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack),
1899 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings),
1900 FEAT_OPP(HEADER_GROUP_DESC, group_desc),
1901 FEAT_OPP(HEADER_AUXTRACE, auxtrace),
1904 struct header_print_data {
1906 bool full; /* extended list of headers */
1909 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1910 struct perf_header *ph,
1911 int feat, int fd, void *data)
1913 struct header_print_data *hd = data;
1915 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1916 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1917 "%d, continuing...\n", section->offset, feat);
1920 if (feat >= HEADER_LAST_FEATURE) {
1921 pr_warning("unknown feature %d\n", feat);
1924 if (!feat_ops[feat].print)
1927 if (!feat_ops[feat].full_only || hd->full)
1928 feat_ops[feat].print(ph, fd, hd->fp);
1930 fprintf(hd->fp, "# %s info available, use -I to display\n",
1931 feat_ops[feat].name);
1936 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1938 struct header_print_data hd;
1939 struct perf_header *header = &session->header;
1940 int fd = perf_data_file__fd(session->file);
1944 perf_header__process_sections(header, fd, &hd,
1945 perf_file_section__fprintf_info);
1949 static int do_write_feat(int fd, struct perf_header *h, int type,
1950 struct perf_file_section **p,
1951 struct perf_evlist *evlist)
1956 if (perf_header__has_feat(h, type)) {
1957 if (!feat_ops[type].write)
1960 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1962 err = feat_ops[type].write(fd, h, evlist);
1964 pr_debug("failed to write feature %d\n", type);
1966 /* undo anything written */
1967 lseek(fd, (*p)->offset, SEEK_SET);
1971 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1977 static int perf_header__adds_write(struct perf_header *header,
1978 struct perf_evlist *evlist, int fd)
1981 struct perf_file_section *feat_sec, *p;
1987 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1991 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1992 if (feat_sec == NULL)
1995 sec_size = sizeof(*feat_sec) * nr_sections;
1997 sec_start = header->feat_offset;
1998 lseek(fd, sec_start + sec_size, SEEK_SET);
2000 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2001 if (do_write_feat(fd, header, feat, &p, evlist))
2002 perf_header__clear_feat(header, feat);
2005 lseek(fd, sec_start, SEEK_SET);
2007 * may write more than needed due to dropped feature, but
2008 * this is okay, reader will skip the mising entries
2010 err = do_write(fd, feat_sec, sec_size);
2012 pr_debug("failed to write feature section\n");
2017 int perf_header__write_pipe(int fd)
2019 struct perf_pipe_file_header f_header;
2022 f_header = (struct perf_pipe_file_header){
2023 .magic = PERF_MAGIC,
2024 .size = sizeof(f_header),
2027 err = do_write(fd, &f_header, sizeof(f_header));
2029 pr_debug("failed to write perf pipe header\n");
2036 int perf_session__write_header(struct perf_session *session,
2037 struct perf_evlist *evlist,
2038 int fd, bool at_exit)
2040 struct perf_file_header f_header;
2041 struct perf_file_attr f_attr;
2042 struct perf_header *header = &session->header;
2043 struct perf_evsel *evsel;
2047 lseek(fd, sizeof(f_header), SEEK_SET);
2049 evlist__for_each(session->evlist, evsel) {
2050 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2051 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2053 pr_debug("failed to write perf header\n");
2058 attr_offset = lseek(fd, 0, SEEK_CUR);
2060 evlist__for_each(evlist, evsel) {
2061 f_attr = (struct perf_file_attr){
2062 .attr = evsel->attr,
2064 .offset = evsel->id_offset,
2065 .size = evsel->ids * sizeof(u64),
2068 err = do_write(fd, &f_attr, sizeof(f_attr));
2070 pr_debug("failed to write perf header attribute\n");
2075 if (!header->data_offset)
2076 header->data_offset = lseek(fd, 0, SEEK_CUR);
2077 header->feat_offset = header->data_offset + header->data_size;
2080 err = perf_header__adds_write(header, evlist, fd);
2085 f_header = (struct perf_file_header){
2086 .magic = PERF_MAGIC,
2087 .size = sizeof(f_header),
2088 .attr_size = sizeof(f_attr),
2090 .offset = attr_offset,
2091 .size = evlist->nr_entries * sizeof(f_attr),
2094 .offset = header->data_offset,
2095 .size = header->data_size,
2097 /* event_types is ignored, store zeros */
2100 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2102 lseek(fd, 0, SEEK_SET);
2103 err = do_write(fd, &f_header, sizeof(f_header));
2105 pr_debug("failed to write perf header\n");
2108 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2113 static int perf_header__getbuffer64(struct perf_header *header,
2114 int fd, void *buf, size_t size)
2116 if (readn(fd, buf, size) <= 0)
2119 if (header->needs_swap)
2120 mem_bswap_64(buf, size);
2125 int perf_header__process_sections(struct perf_header *header, int fd,
2127 int (*process)(struct perf_file_section *section,
2128 struct perf_header *ph,
2129 int feat, int fd, void *data))
2131 struct perf_file_section *feat_sec, *sec;
2137 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2141 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2145 sec_size = sizeof(*feat_sec) * nr_sections;
2147 lseek(fd, header->feat_offset, SEEK_SET);
2149 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2153 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2154 err = process(sec++, header, feat, fd, data);
2164 static const int attr_file_abi_sizes[] = {
2165 [0] = PERF_ATTR_SIZE_VER0,
2166 [1] = PERF_ATTR_SIZE_VER1,
2167 [2] = PERF_ATTR_SIZE_VER2,
2168 [3] = PERF_ATTR_SIZE_VER3,
2169 [4] = PERF_ATTR_SIZE_VER4,
2174 * In the legacy file format, the magic number is not used to encode endianness.
2175 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2176 * on ABI revisions, we need to try all combinations for all endianness to
2177 * detect the endianness.
2179 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2181 uint64_t ref_size, attr_size;
2184 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2185 ref_size = attr_file_abi_sizes[i]
2186 + sizeof(struct perf_file_section);
2187 if (hdr_sz != ref_size) {
2188 attr_size = bswap_64(hdr_sz);
2189 if (attr_size != ref_size)
2192 ph->needs_swap = true;
2194 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2199 /* could not determine endianness */
2203 #define PERF_PIPE_HDR_VER0 16
2205 static const size_t attr_pipe_abi_sizes[] = {
2206 [0] = PERF_PIPE_HDR_VER0,
2211 * In the legacy pipe format, there is an implicit assumption that endiannesss
2212 * between host recording the samples, and host parsing the samples is the
2213 * same. This is not always the case given that the pipe output may always be
2214 * redirected into a file and analyzed on a different machine with possibly a
2215 * different endianness and perf_event ABI revsions in the perf tool itself.
2217 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2222 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2223 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2224 attr_size = bswap_64(hdr_sz);
2225 if (attr_size != hdr_sz)
2228 ph->needs_swap = true;
2230 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2236 bool is_perf_magic(u64 magic)
2238 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2239 || magic == __perf_magic2
2240 || magic == __perf_magic2_sw)
2246 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2247 bool is_pipe, struct perf_header *ph)
2251 /* check for legacy format */
2252 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2254 ph->version = PERF_HEADER_VERSION_1;
2255 pr_debug("legacy perf.data format\n");
2257 return try_all_pipe_abis(hdr_sz, ph);
2259 return try_all_file_abis(hdr_sz, ph);
2262 * the new magic number serves two purposes:
2263 * - unique number to identify actual perf.data files
2264 * - encode endianness of file
2266 ph->version = PERF_HEADER_VERSION_2;
2268 /* check magic number with one endianness */
2269 if (magic == __perf_magic2)
2272 /* check magic number with opposite endianness */
2273 if (magic != __perf_magic2_sw)
2276 ph->needs_swap = true;
2281 int perf_file_header__read(struct perf_file_header *header,
2282 struct perf_header *ph, int fd)
2286 lseek(fd, 0, SEEK_SET);
2288 ret = readn(fd, header, sizeof(*header));
2292 if (check_magic_endian(header->magic,
2293 header->attr_size, false, ph) < 0) {
2294 pr_debug("magic/endian check failed\n");
2298 if (ph->needs_swap) {
2299 mem_bswap_64(header, offsetof(struct perf_file_header,
2303 if (header->size != sizeof(*header)) {
2304 /* Support the previous format */
2305 if (header->size == offsetof(typeof(*header), adds_features))
2306 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2309 } else if (ph->needs_swap) {
2311 * feature bitmap is declared as an array of unsigned longs --
2312 * not good since its size can differ between the host that
2313 * generated the data file and the host analyzing the file.
2315 * We need to handle endianness, but we don't know the size of
2316 * the unsigned long where the file was generated. Take a best
2317 * guess at determining it: try 64-bit swap first (ie., file
2318 * created on a 64-bit host), and check if the hostname feature
2319 * bit is set (this feature bit is forced on as of fbe96f2).
2320 * If the bit is not, undo the 64-bit swap and try a 32-bit
2321 * swap. If the hostname bit is still not set (e.g., older data
2322 * file), punt and fallback to the original behavior --
2323 * clearing all feature bits and setting buildid.
2325 mem_bswap_64(&header->adds_features,
2326 BITS_TO_U64(HEADER_FEAT_BITS));
2328 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2330 mem_bswap_64(&header->adds_features,
2331 BITS_TO_U64(HEADER_FEAT_BITS));
2334 mem_bswap_32(&header->adds_features,
2335 BITS_TO_U32(HEADER_FEAT_BITS));
2338 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2339 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2340 set_bit(HEADER_BUILD_ID, header->adds_features);
2344 memcpy(&ph->adds_features, &header->adds_features,
2345 sizeof(ph->adds_features));
2347 ph->data_offset = header->data.offset;
2348 ph->data_size = header->data.size;
2349 ph->feat_offset = header->data.offset + header->data.size;
2353 static int perf_file_section__process(struct perf_file_section *section,
2354 struct perf_header *ph,
2355 int feat, int fd, void *data)
2357 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2358 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2359 "%d, continuing...\n", section->offset, feat);
2363 if (feat >= HEADER_LAST_FEATURE) {
2364 pr_debug("unknown feature %d, continuing...\n", feat);
2368 if (!feat_ops[feat].process)
2371 return feat_ops[feat].process(section, ph, fd, data);
2374 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2375 struct perf_header *ph, int fd,
2380 ret = readn(fd, header, sizeof(*header));
2384 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2385 pr_debug("endian/magic failed\n");
2390 header->size = bswap_64(header->size);
2392 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2398 static int perf_header__read_pipe(struct perf_session *session)
2400 struct perf_header *header = &session->header;
2401 struct perf_pipe_file_header f_header;
2403 if (perf_file_header__read_pipe(&f_header, header,
2404 perf_data_file__fd(session->file),
2405 session->repipe) < 0) {
2406 pr_debug("incompatible file format\n");
2413 static int read_attr(int fd, struct perf_header *ph,
2414 struct perf_file_attr *f_attr)
2416 struct perf_event_attr *attr = &f_attr->attr;
2418 size_t our_sz = sizeof(f_attr->attr);
2421 memset(f_attr, 0, sizeof(*f_attr));
2423 /* read minimal guaranteed structure */
2424 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2426 pr_debug("cannot read %d bytes of header attr\n",
2427 PERF_ATTR_SIZE_VER0);
2431 /* on file perf_event_attr size */
2439 sz = PERF_ATTR_SIZE_VER0;
2440 } else if (sz > our_sz) {
2441 pr_debug("file uses a more recent and unsupported ABI"
2442 " (%zu bytes extra)\n", sz - our_sz);
2445 /* what we have not yet read and that we know about */
2446 left = sz - PERF_ATTR_SIZE_VER0;
2449 ptr += PERF_ATTR_SIZE_VER0;
2451 ret = readn(fd, ptr, left);
2453 /* read perf_file_section, ids are read in caller */
2454 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2456 return ret <= 0 ? -1 : 0;
2459 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2460 struct pevent *pevent)
2462 struct event_format *event;
2465 /* already prepared */
2466 if (evsel->tp_format)
2469 if (pevent == NULL) {
2470 pr_debug("broken or missing trace data\n");
2474 event = pevent_find_event(pevent, evsel->attr.config);
2479 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2480 evsel->name = strdup(bf);
2481 if (evsel->name == NULL)
2485 evsel->tp_format = event;
2489 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2490 struct pevent *pevent)
2492 struct perf_evsel *pos;
2494 evlist__for_each(evlist, pos) {
2495 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2496 perf_evsel__prepare_tracepoint_event(pos, pevent))
2503 int perf_session__read_header(struct perf_session *session)
2505 struct perf_data_file *file = session->file;
2506 struct perf_header *header = &session->header;
2507 struct perf_file_header f_header;
2508 struct perf_file_attr f_attr;
2510 int nr_attrs, nr_ids, i, j;
2511 int fd = perf_data_file__fd(file);
2513 session->evlist = perf_evlist__new();
2514 if (session->evlist == NULL)
2517 session->evlist->env = &header->env;
2518 if (perf_data_file__is_pipe(file))
2519 return perf_header__read_pipe(session);
2521 if (perf_file_header__read(&f_header, header, fd) < 0)
2525 * Sanity check that perf.data was written cleanly; data size is
2526 * initialized to 0 and updated only if the on_exit function is run.
2527 * If data size is still 0 then the file contains only partial
2528 * information. Just warn user and process it as much as it can.
2530 if (f_header.data.size == 0) {
2531 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2532 "Was the 'perf record' command properly terminated?\n",
2536 nr_attrs = f_header.attrs.size / f_header.attr_size;
2537 lseek(fd, f_header.attrs.offset, SEEK_SET);
2539 for (i = 0; i < nr_attrs; i++) {
2540 struct perf_evsel *evsel;
2543 if (read_attr(fd, header, &f_attr) < 0)
2546 if (header->needs_swap) {
2547 f_attr.ids.size = bswap_64(f_attr.ids.size);
2548 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2549 perf_event__attr_swap(&f_attr.attr);
2552 tmp = lseek(fd, 0, SEEK_CUR);
2553 evsel = perf_evsel__new(&f_attr.attr);
2556 goto out_delete_evlist;
2558 evsel->needs_swap = header->needs_swap;
2560 * Do it before so that if perf_evsel__alloc_id fails, this
2561 * entry gets purged too at perf_evlist__delete().
2563 perf_evlist__add(session->evlist, evsel);
2565 nr_ids = f_attr.ids.size / sizeof(u64);
2567 * We don't have the cpu and thread maps on the header, so
2568 * for allocating the perf_sample_id table we fake 1 cpu and
2569 * hattr->ids threads.
2571 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2572 goto out_delete_evlist;
2574 lseek(fd, f_attr.ids.offset, SEEK_SET);
2576 for (j = 0; j < nr_ids; j++) {
2577 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2580 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2583 lseek(fd, tmp, SEEK_SET);
2586 symbol_conf.nr_events = nr_attrs;
2588 perf_header__process_sections(header, fd, &session->tevent,
2589 perf_file_section__process);
2591 if (perf_evlist__prepare_tracepoint_events(session->evlist,
2592 session->tevent.pevent))
2593 goto out_delete_evlist;
2600 perf_evlist__delete(session->evlist);
2601 session->evlist = NULL;
2605 int perf_event__synthesize_attr(struct perf_tool *tool,
2606 struct perf_event_attr *attr, u32 ids, u64 *id,
2607 perf_event__handler_t process)
2609 union perf_event *ev;
2613 size = sizeof(struct perf_event_attr);
2614 size = PERF_ALIGN(size, sizeof(u64));
2615 size += sizeof(struct perf_event_header);
2616 size += ids * sizeof(u64);
2623 ev->attr.attr = *attr;
2624 memcpy(ev->attr.id, id, ids * sizeof(u64));
2626 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2627 ev->attr.header.size = (u16)size;
2629 if (ev->attr.header.size == size)
2630 err = process(tool, ev, NULL, NULL);
2639 int perf_event__synthesize_attrs(struct perf_tool *tool,
2640 struct perf_session *session,
2641 perf_event__handler_t process)
2643 struct perf_evsel *evsel;
2646 evlist__for_each(session->evlist, evsel) {
2647 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2648 evsel->id, process);
2650 pr_debug("failed to create perf header attribute\n");
2658 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2659 union perf_event *event,
2660 struct perf_evlist **pevlist)
2663 struct perf_evsel *evsel;
2664 struct perf_evlist *evlist = *pevlist;
2666 if (evlist == NULL) {
2667 *pevlist = evlist = perf_evlist__new();
2672 evsel = perf_evsel__new(&event->attr.attr);
2676 perf_evlist__add(evlist, evsel);
2678 ids = event->header.size;
2679 ids -= (void *)&event->attr.id - (void *)event;
2680 n_ids = ids / sizeof(u64);
2682 * We don't have the cpu and thread maps on the header, so
2683 * for allocating the perf_sample_id table we fake 1 cpu and
2684 * hattr->ids threads.
2686 if (perf_evsel__alloc_id(evsel, 1, n_ids))
2689 for (i = 0; i < n_ids; i++) {
2690 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2693 symbol_conf.nr_events = evlist->nr_entries;
2698 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2699 struct perf_evlist *evlist,
2700 perf_event__handler_t process)
2702 union perf_event ev;
2703 struct tracing_data *tdata;
2704 ssize_t size = 0, aligned_size = 0, padding;
2705 int err __maybe_unused = 0;
2708 * We are going to store the size of the data followed
2709 * by the data contents. Since the fd descriptor is a pipe,
2710 * we cannot seek back to store the size of the data once
2711 * we know it. Instead we:
2713 * - write the tracing data to the temp file
2714 * - get/write the data size to pipe
2715 * - write the tracing data from the temp file
2718 tdata = tracing_data_get(&evlist->entries, fd, true);
2722 memset(&ev, 0, sizeof(ev));
2724 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2726 aligned_size = PERF_ALIGN(size, sizeof(u64));
2727 padding = aligned_size - size;
2728 ev.tracing_data.header.size = sizeof(ev.tracing_data);
2729 ev.tracing_data.size = aligned_size;
2731 process(tool, &ev, NULL, NULL);
2734 * The put function will copy all the tracing data
2735 * stored in temp file to the pipe.
2737 tracing_data_put(tdata);
2739 write_padded(fd, NULL, 0, padding);
2741 return aligned_size;
2744 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2745 union perf_event *event,
2746 struct perf_session *session)
2748 ssize_t size_read, padding, size = event->tracing_data.size;
2749 int fd = perf_data_file__fd(session->file);
2750 off_t offset = lseek(fd, 0, SEEK_CUR);
2753 /* setup for reading amidst mmap */
2754 lseek(fd, offset + sizeof(struct tracing_data_event),
2757 size_read = trace_report(fd, &session->tevent,
2759 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2761 if (readn(fd, buf, padding) < 0) {
2762 pr_err("%s: reading input file", __func__);
2765 if (session->repipe) {
2766 int retw = write(STDOUT_FILENO, buf, padding);
2767 if (retw <= 0 || retw != padding) {
2768 pr_err("%s: repiping tracing data padding", __func__);
2773 if (size_read + padding != size) {
2774 pr_err("%s: tracing data size mismatch", __func__);
2778 perf_evlist__prepare_tracepoint_events(session->evlist,
2779 session->tevent.pevent);
2781 return size_read + padding;
2784 int perf_event__synthesize_build_id(struct perf_tool *tool,
2785 struct dso *pos, u16 misc,
2786 perf_event__handler_t process,
2787 struct machine *machine)
2789 union perf_event ev;
2796 memset(&ev, 0, sizeof(ev));
2798 len = pos->long_name_len + 1;
2799 len = PERF_ALIGN(len, NAME_ALIGN);
2800 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2801 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2802 ev.build_id.header.misc = misc;
2803 ev.build_id.pid = machine->pid;
2804 ev.build_id.header.size = sizeof(ev.build_id) + len;
2805 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2807 err = process(tool, &ev, NULL, machine);
2812 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2813 union perf_event *event,
2814 struct perf_session *session)
2816 __event_process_build_id(&event->build_id,
2817 event->build_id.filename,