1 // SPDX-License-Identifier: GPL-2.0
11 #include <linux/compiler.h>
12 #include <linux/list.h>
13 #include <linux/kernel.h>
14 #include <linux/bitops.h>
15 #include <linux/string.h>
16 #include <linux/stringify.h>
17 #include <linux/zalloc.h>
19 #include <sys/utsname.h>
20 #include <linux/time64.h>
22 #include <bpf/libbpf.h>
23 #include <perf/cpumap.h>
28 #include "util/evsel_fprintf.h"
31 #include "trace-event.h"
41 #include <api/fs/fs.h>
44 #include "time-utils.h"
46 #include "util/util.h" // perf_exe()
48 #include "bpf-event.h"
50 #include <linux/ctype.h>
51 #include <internal/lib.h>
55 * must be a numerical value to let the endianness
56 * determine the memory layout. That way we are able
57 * to detect endianness when reading the perf.data file
60 * we check for legacy (PERFFILE) format.
62 static const char *__perf_magic1 = "PERFFILE";
63 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
64 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
66 #define PERF_MAGIC __perf_magic2
68 const char perf_version_string[] = PERF_VERSION;
70 struct perf_file_attr {
71 struct perf_event_attr attr;
72 struct perf_file_section ids;
75 void perf_header__set_feat(struct perf_header *header, int feat)
77 set_bit(feat, header->adds_features);
80 void perf_header__clear_feat(struct perf_header *header, int feat)
82 clear_bit(feat, header->adds_features);
85 bool perf_header__has_feat(const struct perf_header *header, int feat)
87 return test_bit(feat, header->adds_features);
90 static int __do_write_fd(struct feat_fd *ff, const void *buf, size_t size)
92 ssize_t ret = writen(ff->fd, buf, size);
94 if (ret != (ssize_t)size)
95 return ret < 0 ? (int)ret : -1;
99 static int __do_write_buf(struct feat_fd *ff, const void *buf, size_t size)
101 /* struct perf_event_header::size is u16 */
102 const size_t max_size = 0xffff - sizeof(struct perf_event_header);
103 size_t new_size = ff->size;
106 if (size + ff->offset > max_size)
109 while (size > (new_size - ff->offset))
111 new_size = min(max_size, new_size);
113 if (ff->size < new_size) {
114 addr = realloc(ff->buf, new_size);
121 memcpy(ff->buf + ff->offset, buf, size);
127 /* Return: 0 if succeded, -ERR if failed. */
128 int do_write(struct feat_fd *ff, const void *buf, size_t size)
131 return __do_write_fd(ff, buf, size);
132 return __do_write_buf(ff, buf, size);
135 /* Return: 0 if succeded, -ERR if failed. */
136 static int do_write_bitmap(struct feat_fd *ff, unsigned long *set, u64 size)
138 u64 *p = (u64 *) set;
141 ret = do_write(ff, &size, sizeof(size));
145 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
146 ret = do_write(ff, p + i, sizeof(*p));
154 /* Return: 0 if succeded, -ERR if failed. */
155 int write_padded(struct feat_fd *ff, const void *bf,
156 size_t count, size_t count_aligned)
158 static const char zero_buf[NAME_ALIGN];
159 int err = do_write(ff, bf, count);
162 err = do_write(ff, zero_buf, count_aligned - count);
167 #define string_size(str) \
168 (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
170 /* Return: 0 if succeded, -ERR if failed. */
171 static int do_write_string(struct feat_fd *ff, const char *str)
176 olen = strlen(str) + 1;
177 len = PERF_ALIGN(olen, NAME_ALIGN);
179 /* write len, incl. \0 */
180 ret = do_write(ff, &len, sizeof(len));
184 return write_padded(ff, str, olen, len);
187 static int __do_read_fd(struct feat_fd *ff, void *addr, ssize_t size)
189 ssize_t ret = readn(ff->fd, addr, size);
192 return ret < 0 ? (int)ret : -1;
196 static int __do_read_buf(struct feat_fd *ff, void *addr, ssize_t size)
198 if (size > (ssize_t)ff->size - ff->offset)
201 memcpy(addr, ff->buf + ff->offset, size);
208 static int __do_read(struct feat_fd *ff, void *addr, ssize_t size)
211 return __do_read_fd(ff, addr, size);
212 return __do_read_buf(ff, addr, size);
215 static int do_read_u32(struct feat_fd *ff, u32 *addr)
219 ret = __do_read(ff, addr, sizeof(*addr));
223 if (ff->ph->needs_swap)
224 *addr = bswap_32(*addr);
228 static int do_read_u64(struct feat_fd *ff, u64 *addr)
232 ret = __do_read(ff, addr, sizeof(*addr));
236 if (ff->ph->needs_swap)
237 *addr = bswap_64(*addr);
241 static char *do_read_string(struct feat_fd *ff)
246 if (do_read_u32(ff, &len))
253 if (!__do_read(ff, buf, len)) {
255 * strings are padded by zeroes
256 * thus the actual strlen of buf
257 * may be less than len
266 /* Return: 0 if succeded, -ERR if failed. */
267 static int do_read_bitmap(struct feat_fd *ff, unsigned long **pset, u64 *psize)
273 ret = do_read_u64(ff, &size);
277 set = bitmap_alloc(size);
283 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
284 ret = do_read_u64(ff, p + i);
296 static int write_tracing_data(struct feat_fd *ff,
297 struct evlist *evlist)
299 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
302 return read_tracing_data(ff->fd, &evlist->core.entries);
305 static int write_build_id(struct feat_fd *ff,
306 struct evlist *evlist __maybe_unused)
308 struct perf_session *session;
311 session = container_of(ff->ph, struct perf_session, header);
313 if (!perf_session__read_build_ids(session, true))
316 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
319 err = perf_session__write_buildid_table(session, ff);
321 pr_debug("failed to write buildid table\n");
324 perf_session__cache_build_ids(session);
329 static int write_hostname(struct feat_fd *ff,
330 struct evlist *evlist __maybe_unused)
339 return do_write_string(ff, uts.nodename);
342 static int write_osrelease(struct feat_fd *ff,
343 struct evlist *evlist __maybe_unused)
352 return do_write_string(ff, uts.release);
355 static int write_arch(struct feat_fd *ff,
356 struct evlist *evlist __maybe_unused)
365 return do_write_string(ff, uts.machine);
368 static int write_version(struct feat_fd *ff,
369 struct evlist *evlist __maybe_unused)
371 return do_write_string(ff, perf_version_string);
374 static int __write_cpudesc(struct feat_fd *ff, const char *cpuinfo_proc)
379 const char *search = cpuinfo_proc;
386 file = fopen("/proc/cpuinfo", "r");
390 while (getline(&buf, &len, file) > 0) {
391 ret = strncmp(buf, search, strlen(search));
403 p = strchr(buf, ':');
404 if (p && *(p+1) == ' ' && *(p+2))
410 /* squash extra space characters (branding string) */
415 char *q = skip_spaces(r);
418 while ((*r++ = *q++));
422 ret = do_write_string(ff, s);
429 static int write_cpudesc(struct feat_fd *ff,
430 struct evlist *evlist __maybe_unused)
432 #if defined(__powerpc__) || defined(__hppa__) || defined(__sparc__)
433 #define CPUINFO_PROC { "cpu", }
434 #elif defined(__s390__)
435 #define CPUINFO_PROC { "vendor_id", }
436 #elif defined(__sh__)
437 #define CPUINFO_PROC { "cpu type", }
438 #elif defined(__alpha__) || defined(__mips__)
439 #define CPUINFO_PROC { "cpu model", }
440 #elif defined(__arm__)
441 #define CPUINFO_PROC { "model name", "Processor", }
442 #elif defined(__arc__)
443 #define CPUINFO_PROC { "Processor", }
444 #elif defined(__xtensa__)
445 #define CPUINFO_PROC { "core ID", }
447 #define CPUINFO_PROC { "model name", }
449 const char *cpuinfo_procs[] = CPUINFO_PROC;
453 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
455 ret = __write_cpudesc(ff, cpuinfo_procs[i]);
463 static int write_nrcpus(struct feat_fd *ff,
464 struct evlist *evlist __maybe_unused)
470 nrc = cpu__max_present_cpu();
472 nr = sysconf(_SC_NPROCESSORS_ONLN);
476 nra = (u32)(nr & UINT_MAX);
478 ret = do_write(ff, &nrc, sizeof(nrc));
482 return do_write(ff, &nra, sizeof(nra));
485 static int write_event_desc(struct feat_fd *ff,
486 struct evlist *evlist)
492 nre = evlist->core.nr_entries;
495 * write number of events
497 ret = do_write(ff, &nre, sizeof(nre));
502 * size of perf_event_attr struct
504 sz = (u32)sizeof(evsel->core.attr);
505 ret = do_write(ff, &sz, sizeof(sz));
509 evlist__for_each_entry(evlist, evsel) {
510 ret = do_write(ff, &evsel->core.attr, sz);
514 * write number of unique id per event
515 * there is one id per instance of an event
517 * copy into an nri to be independent of the
520 nri = evsel->core.ids;
521 ret = do_write(ff, &nri, sizeof(nri));
526 * write event string as passed on cmdline
528 ret = do_write_string(ff, perf_evsel__name(evsel));
532 * write unique ids for this event
534 ret = do_write(ff, evsel->core.id, evsel->core.ids * sizeof(u64));
541 static int write_cmdline(struct feat_fd *ff,
542 struct evlist *evlist __maybe_unused)
544 char pbuf[MAXPATHLEN], *buf;
547 /* actual path to perf binary */
548 buf = perf_exe(pbuf, MAXPATHLEN);
550 /* account for binary path */
551 n = perf_env.nr_cmdline + 1;
553 ret = do_write(ff, &n, sizeof(n));
557 ret = do_write_string(ff, buf);
561 for (i = 0 ; i < perf_env.nr_cmdline; i++) {
562 ret = do_write_string(ff, perf_env.cmdline_argv[i]);
570 static int write_cpu_topology(struct feat_fd *ff,
571 struct evlist *evlist __maybe_unused)
573 struct cpu_topology *tp;
577 tp = cpu_topology__new();
581 ret = do_write(ff, &tp->core_sib, sizeof(tp->core_sib));
585 for (i = 0; i < tp->core_sib; i++) {
586 ret = do_write_string(ff, tp->core_siblings[i]);
590 ret = do_write(ff, &tp->thread_sib, sizeof(tp->thread_sib));
594 for (i = 0; i < tp->thread_sib; i++) {
595 ret = do_write_string(ff, tp->thread_siblings[i]);
600 ret = perf_env__read_cpu_topology_map(&perf_env);
604 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
605 ret = do_write(ff, &perf_env.cpu[j].core_id,
606 sizeof(perf_env.cpu[j].core_id));
609 ret = do_write(ff, &perf_env.cpu[j].socket_id,
610 sizeof(perf_env.cpu[j].socket_id));
618 ret = do_write(ff, &tp->die_sib, sizeof(tp->die_sib));
622 for (i = 0; i < tp->die_sib; i++) {
623 ret = do_write_string(ff, tp->die_siblings[i]);
628 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
629 ret = do_write(ff, &perf_env.cpu[j].die_id,
630 sizeof(perf_env.cpu[j].die_id));
636 cpu_topology__delete(tp);
642 static int write_total_mem(struct feat_fd *ff,
643 struct evlist *evlist __maybe_unused)
651 fp = fopen("/proc/meminfo", "r");
655 while (getline(&buf, &len, fp) > 0) {
656 ret = strncmp(buf, "MemTotal:", 9);
661 n = sscanf(buf, "%*s %"PRIu64, &mem);
663 ret = do_write(ff, &mem, sizeof(mem));
671 static int write_numa_topology(struct feat_fd *ff,
672 struct evlist *evlist __maybe_unused)
674 struct numa_topology *tp;
678 tp = numa_topology__new();
682 ret = do_write(ff, &tp->nr, sizeof(u32));
686 for (i = 0; i < tp->nr; i++) {
687 struct numa_topology_node *n = &tp->nodes[i];
689 ret = do_write(ff, &n->node, sizeof(u32));
693 ret = do_write(ff, &n->mem_total, sizeof(u64));
697 ret = do_write(ff, &n->mem_free, sizeof(u64));
701 ret = do_write_string(ff, n->cpus);
709 numa_topology__delete(tp);
716 * struct pmu_mappings {
725 static int write_pmu_mappings(struct feat_fd *ff,
726 struct evlist *evlist __maybe_unused)
728 struct perf_pmu *pmu = NULL;
733 * Do a first pass to count number of pmu to avoid lseek so this
734 * works in pipe mode as well.
736 while ((pmu = perf_pmu__scan(pmu))) {
742 ret = do_write(ff, &pmu_num, sizeof(pmu_num));
746 while ((pmu = perf_pmu__scan(pmu))) {
750 ret = do_write(ff, &pmu->type, sizeof(pmu->type));
754 ret = do_write_string(ff, pmu->name);
765 * struct group_descs {
767 * struct group_desc {
774 static int write_group_desc(struct feat_fd *ff,
775 struct evlist *evlist)
777 u32 nr_groups = evlist->nr_groups;
781 ret = do_write(ff, &nr_groups, sizeof(nr_groups));
785 evlist__for_each_entry(evlist, evsel) {
786 if (perf_evsel__is_group_leader(evsel) &&
787 evsel->core.nr_members > 1) {
788 const char *name = evsel->group_name ?: "{anon_group}";
789 u32 leader_idx = evsel->idx;
790 u32 nr_members = evsel->core.nr_members;
792 ret = do_write_string(ff, name);
796 ret = do_write(ff, &leader_idx, sizeof(leader_idx));
800 ret = do_write(ff, &nr_members, sizeof(nr_members));
809 * Return the CPU id as a raw string.
811 * Each architecture should provide a more precise id string that
812 * can be use to match the architecture's "mapfile".
814 char * __weak get_cpuid_str(struct perf_pmu *pmu __maybe_unused)
819 /* Return zero when the cpuid from the mapfile.csv matches the
820 * cpuid string generated on this platform.
821 * Otherwise return non-zero.
823 int __weak strcmp_cpuid_str(const char *mapcpuid, const char *cpuid)
826 regmatch_t pmatch[1];
829 if (regcomp(&re, mapcpuid, REG_EXTENDED) != 0) {
830 /* Warn unable to generate match particular string. */
831 pr_info("Invalid regular expression %s\n", mapcpuid);
835 match = !regexec(&re, cpuid, 1, pmatch, 0);
838 size_t match_len = (pmatch[0].rm_eo - pmatch[0].rm_so);
840 /* Verify the entire string matched. */
841 if (match_len == strlen(cpuid))
848 * default get_cpuid(): nothing gets recorded
849 * actual implementation must be in arch/$(SRCARCH)/util/header.c
851 int __weak get_cpuid(char *buffer __maybe_unused, size_t sz __maybe_unused)
853 return ENOSYS; /* Not implemented */
856 static int write_cpuid(struct feat_fd *ff,
857 struct evlist *evlist __maybe_unused)
862 ret = get_cpuid(buffer, sizeof(buffer));
866 return do_write_string(ff, buffer);
869 static int write_branch_stack(struct feat_fd *ff __maybe_unused,
870 struct evlist *evlist __maybe_unused)
875 static int write_auxtrace(struct feat_fd *ff,
876 struct evlist *evlist __maybe_unused)
878 struct perf_session *session;
881 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
884 session = container_of(ff->ph, struct perf_session, header);
886 err = auxtrace_index__write(ff->fd, &session->auxtrace_index);
888 pr_err("Failed to write auxtrace index\n");
892 static int write_clockid(struct feat_fd *ff,
893 struct evlist *evlist __maybe_unused)
895 return do_write(ff, &ff->ph->env.clockid_res_ns,
896 sizeof(ff->ph->env.clockid_res_ns));
899 static int write_dir_format(struct feat_fd *ff,
900 struct evlist *evlist __maybe_unused)
902 struct perf_session *session;
903 struct perf_data *data;
905 session = container_of(ff->ph, struct perf_session, header);
906 data = session->data;
908 if (WARN_ON(!perf_data__is_dir(data)))
911 return do_write(ff, &data->dir.version, sizeof(data->dir.version));
914 #ifdef HAVE_LIBBPF_SUPPORT
915 static int write_bpf_prog_info(struct feat_fd *ff,
916 struct evlist *evlist __maybe_unused)
918 struct perf_env *env = &ff->ph->env;
919 struct rb_root *root;
920 struct rb_node *next;
923 down_read(&env->bpf_progs.lock);
925 ret = do_write(ff, &env->bpf_progs.infos_cnt,
926 sizeof(env->bpf_progs.infos_cnt));
930 root = &env->bpf_progs.infos;
931 next = rb_first(root);
933 struct bpf_prog_info_node *node;
936 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
937 next = rb_next(&node->rb_node);
938 len = sizeof(struct bpf_prog_info_linear) +
939 node->info_linear->data_len;
941 /* before writing to file, translate address to offset */
942 bpf_program__bpil_addr_to_offs(node->info_linear);
943 ret = do_write(ff, node->info_linear, len);
945 * translate back to address even when do_write() fails,
946 * so that this function never changes the data.
948 bpf_program__bpil_offs_to_addr(node->info_linear);
953 up_read(&env->bpf_progs.lock);
956 #else // HAVE_LIBBPF_SUPPORT
957 static int write_bpf_prog_info(struct feat_fd *ff __maybe_unused,
958 struct evlist *evlist __maybe_unused)
962 #endif // HAVE_LIBBPF_SUPPORT
964 static int write_bpf_btf(struct feat_fd *ff,
965 struct evlist *evlist __maybe_unused)
967 struct perf_env *env = &ff->ph->env;
968 struct rb_root *root;
969 struct rb_node *next;
972 down_read(&env->bpf_progs.lock);
974 ret = do_write(ff, &env->bpf_progs.btfs_cnt,
975 sizeof(env->bpf_progs.btfs_cnt));
980 root = &env->bpf_progs.btfs;
981 next = rb_first(root);
983 struct btf_node *node;
985 node = rb_entry(next, struct btf_node, rb_node);
986 next = rb_next(&node->rb_node);
987 ret = do_write(ff, &node->id,
988 sizeof(u32) * 2 + node->data_size);
993 up_read(&env->bpf_progs.lock);
997 static int cpu_cache_level__sort(const void *a, const void *b)
999 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
1000 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
1002 return cache_a->level - cache_b->level;
1005 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
1007 if (a->level != b->level)
1010 if (a->line_size != b->line_size)
1013 if (a->sets != b->sets)
1016 if (a->ways != b->ways)
1019 if (strcmp(a->type, b->type))
1022 if (strcmp(a->size, b->size))
1025 if (strcmp(a->map, b->map))
1031 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
1033 char path[PATH_MAX], file[PATH_MAX];
1037 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
1038 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
1040 if (stat(file, &st))
1043 scnprintf(file, PATH_MAX, "%s/level", path);
1044 if (sysfs__read_int(file, (int *) &cache->level))
1047 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
1048 if (sysfs__read_int(file, (int *) &cache->line_size))
1051 scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
1052 if (sysfs__read_int(file, (int *) &cache->sets))
1055 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
1056 if (sysfs__read_int(file, (int *) &cache->ways))
1059 scnprintf(file, PATH_MAX, "%s/type", path);
1060 if (sysfs__read_str(file, &cache->type, &len))
1063 cache->type[len] = 0;
1064 cache->type = strim(cache->type);
1066 scnprintf(file, PATH_MAX, "%s/size", path);
1067 if (sysfs__read_str(file, &cache->size, &len)) {
1068 zfree(&cache->type);
1072 cache->size[len] = 0;
1073 cache->size = strim(cache->size);
1075 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
1076 if (sysfs__read_str(file, &cache->map, &len)) {
1077 zfree(&cache->size);
1078 zfree(&cache->type);
1082 cache->map[len] = 0;
1083 cache->map = strim(cache->map);
1087 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
1089 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
1092 #define MAX_CACHE_LVL 4
1094 static int build_caches(struct cpu_cache_level caches[], u32 *cntp)
1100 nr = cpu__max_cpu();
1102 for (cpu = 0; cpu < nr; cpu++) {
1103 for (level = 0; level < MAX_CACHE_LVL; level++) {
1104 struct cpu_cache_level c;
1107 err = cpu_cache_level__read(&c, cpu, level);
1114 for (i = 0; i < cnt; i++) {
1115 if (cpu_cache_level__cmp(&c, &caches[i]))
1122 cpu_cache_level__free(&c);
1129 static int write_cache(struct feat_fd *ff,
1130 struct evlist *evlist __maybe_unused)
1132 u32 max_caches = cpu__max_cpu() * MAX_CACHE_LVL;
1133 struct cpu_cache_level caches[max_caches];
1134 u32 cnt = 0, i, version = 1;
1137 ret = build_caches(caches, &cnt);
1141 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1143 ret = do_write(ff, &version, sizeof(u32));
1147 ret = do_write(ff, &cnt, sizeof(u32));
1151 for (i = 0; i < cnt; i++) {
1152 struct cpu_cache_level *c = &caches[i];
1155 ret = do_write(ff, &c->v, sizeof(u32)); \
1166 ret = do_write_string(ff, (const char *) c->v); \
1177 for (i = 0; i < cnt; i++)
1178 cpu_cache_level__free(&caches[i]);
1182 static int write_stat(struct feat_fd *ff __maybe_unused,
1183 struct evlist *evlist __maybe_unused)
1188 static int write_sample_time(struct feat_fd *ff,
1189 struct evlist *evlist)
1193 ret = do_write(ff, &evlist->first_sample_time,
1194 sizeof(evlist->first_sample_time));
1198 return do_write(ff, &evlist->last_sample_time,
1199 sizeof(evlist->last_sample_time));
1203 static int memory_node__read(struct memory_node *n, unsigned long idx)
1205 unsigned int phys, size = 0;
1206 char path[PATH_MAX];
1210 #define for_each_memory(mem, dir) \
1211 while ((ent = readdir(dir))) \
1212 if (strcmp(ent->d_name, ".") && \
1213 strcmp(ent->d_name, "..") && \
1214 sscanf(ent->d_name, "memory%u", &mem) == 1)
1216 scnprintf(path, PATH_MAX,
1217 "%s/devices/system/node/node%lu",
1218 sysfs__mountpoint(), idx);
1220 dir = opendir(path);
1222 pr_warning("failed: cant' open memory sysfs data\n");
1226 for_each_memory(phys, dir) {
1227 size = max(phys, size);
1232 n->set = bitmap_alloc(size);
1243 for_each_memory(phys, dir) {
1244 set_bit(phys, n->set);
1251 static int memory_node__sort(const void *a, const void *b)
1253 const struct memory_node *na = a;
1254 const struct memory_node *nb = b;
1256 return na->node - nb->node;
1259 static int build_mem_topology(struct memory_node *nodes, u64 size, u64 *cntp)
1261 char path[PATH_MAX];
1267 scnprintf(path, PATH_MAX, "%s/devices/system/node/",
1268 sysfs__mountpoint());
1270 dir = opendir(path);
1272 pr_debug2("%s: could't read %s, does this arch have topology information?\n",
1277 while (!ret && (ent = readdir(dir))) {
1281 if (!strcmp(ent->d_name, ".") ||
1282 !strcmp(ent->d_name, ".."))
1285 r = sscanf(ent->d_name, "node%u", &idx);
1289 if (WARN_ONCE(cnt >= size,
1290 "failed to write MEM_TOPOLOGY, way too many nodes\n")) {
1295 ret = memory_node__read(&nodes[cnt++], idx);
1302 qsort(nodes, cnt, sizeof(nodes[0]), memory_node__sort);
1307 #define MAX_MEMORY_NODES 2000
1310 * The MEM_TOPOLOGY holds physical memory map for every
1311 * node in system. The format of data is as follows:
1313 * 0 - version | for future changes
1314 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes
1315 * 16 - count | number of nodes
1317 * For each node we store map of physical indexes for
1320 * 32 - node id | node index
1321 * 40 - size | size of bitmap
1322 * 48 - bitmap | bitmap of memory indexes that belongs to node
1324 static int write_mem_topology(struct feat_fd *ff __maybe_unused,
1325 struct evlist *evlist __maybe_unused)
1327 static struct memory_node nodes[MAX_MEMORY_NODES];
1328 u64 bsize, version = 1, i, nr;
1331 ret = sysfs__read_xll("devices/system/memory/block_size_bytes",
1332 (unsigned long long *) &bsize);
1336 ret = build_mem_topology(&nodes[0], MAX_MEMORY_NODES, &nr);
1340 ret = do_write(ff, &version, sizeof(version));
1344 ret = do_write(ff, &bsize, sizeof(bsize));
1348 ret = do_write(ff, &nr, sizeof(nr));
1352 for (i = 0; i < nr; i++) {
1353 struct memory_node *n = &nodes[i];
1356 ret = do_write(ff, &n->v, sizeof(n->v)); \
1365 ret = do_write_bitmap(ff, n->set, n->size);
1374 static int write_compressed(struct feat_fd *ff __maybe_unused,
1375 struct evlist *evlist __maybe_unused)
1379 ret = do_write(ff, &(ff->ph->env.comp_ver), sizeof(ff->ph->env.comp_ver));
1383 ret = do_write(ff, &(ff->ph->env.comp_type), sizeof(ff->ph->env.comp_type));
1387 ret = do_write(ff, &(ff->ph->env.comp_level), sizeof(ff->ph->env.comp_level));
1391 ret = do_write(ff, &(ff->ph->env.comp_ratio), sizeof(ff->ph->env.comp_ratio));
1395 return do_write(ff, &(ff->ph->env.comp_mmap_len), sizeof(ff->ph->env.comp_mmap_len));
1398 static void print_hostname(struct feat_fd *ff, FILE *fp)
1400 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1403 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1405 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1408 static void print_arch(struct feat_fd *ff, FILE *fp)
1410 fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1413 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1415 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1418 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1420 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1421 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1424 static void print_version(struct feat_fd *ff, FILE *fp)
1426 fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1429 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1433 nr = ff->ph->env.nr_cmdline;
1435 fprintf(fp, "# cmdline : ");
1437 for (i = 0; i < nr; i++) {
1438 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]);
1440 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1444 char *quote = strchr(argv_i, '\'');
1448 fprintf(fp, "%s\\\'", argv_i);
1451 fprintf(fp, "%s ", argv_i);
1458 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1460 struct perf_header *ph = ff->ph;
1461 int cpu_nr = ph->env.nr_cpus_avail;
1465 nr = ph->env.nr_sibling_cores;
1466 str = ph->env.sibling_cores;
1468 for (i = 0; i < nr; i++) {
1469 fprintf(fp, "# sibling sockets : %s\n", str);
1470 str += strlen(str) + 1;
1473 if (ph->env.nr_sibling_dies) {
1474 nr = ph->env.nr_sibling_dies;
1475 str = ph->env.sibling_dies;
1477 for (i = 0; i < nr; i++) {
1478 fprintf(fp, "# sibling dies : %s\n", str);
1479 str += strlen(str) + 1;
1483 nr = ph->env.nr_sibling_threads;
1484 str = ph->env.sibling_threads;
1486 for (i = 0; i < nr; i++) {
1487 fprintf(fp, "# sibling threads : %s\n", str);
1488 str += strlen(str) + 1;
1491 if (ph->env.nr_sibling_dies) {
1492 if (ph->env.cpu != NULL) {
1493 for (i = 0; i < cpu_nr; i++)
1494 fprintf(fp, "# CPU %d: Core ID %d, "
1495 "Die ID %d, Socket ID %d\n",
1496 i, ph->env.cpu[i].core_id,
1497 ph->env.cpu[i].die_id,
1498 ph->env.cpu[i].socket_id);
1500 fprintf(fp, "# Core ID, Die ID and Socket ID "
1501 "information is not available\n");
1503 if (ph->env.cpu != NULL) {
1504 for (i = 0; i < cpu_nr; i++)
1505 fprintf(fp, "# CPU %d: Core ID %d, "
1507 i, ph->env.cpu[i].core_id,
1508 ph->env.cpu[i].socket_id);
1510 fprintf(fp, "# Core ID and Socket ID "
1511 "information is not available\n");
1515 static void print_clockid(struct feat_fd *ff, FILE *fp)
1517 fprintf(fp, "# clockid frequency: %"PRIu64" MHz\n",
1518 ff->ph->env.clockid_res_ns * 1000);
1521 static void print_dir_format(struct feat_fd *ff, FILE *fp)
1523 struct perf_session *session;
1524 struct perf_data *data;
1526 session = container_of(ff->ph, struct perf_session, header);
1527 data = session->data;
1529 fprintf(fp, "# directory data version : %"PRIu64"\n", data->dir.version);
1532 static void print_bpf_prog_info(struct feat_fd *ff, FILE *fp)
1534 struct perf_env *env = &ff->ph->env;
1535 struct rb_root *root;
1536 struct rb_node *next;
1538 down_read(&env->bpf_progs.lock);
1540 root = &env->bpf_progs.infos;
1541 next = rb_first(root);
1544 struct bpf_prog_info_node *node;
1546 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
1547 next = rb_next(&node->rb_node);
1549 bpf_event__print_bpf_prog_info(&node->info_linear->info,
1553 up_read(&env->bpf_progs.lock);
1556 static void print_bpf_btf(struct feat_fd *ff, FILE *fp)
1558 struct perf_env *env = &ff->ph->env;
1559 struct rb_root *root;
1560 struct rb_node *next;
1562 down_read(&env->bpf_progs.lock);
1564 root = &env->bpf_progs.btfs;
1565 next = rb_first(root);
1568 struct btf_node *node;
1570 node = rb_entry(next, struct btf_node, rb_node);
1571 next = rb_next(&node->rb_node);
1572 fprintf(fp, "# btf info of id %u\n", node->id);
1575 up_read(&env->bpf_progs.lock);
1578 static void free_event_desc(struct evsel *events)
1580 struct evsel *evsel;
1585 for (evsel = events; evsel->core.attr.size; evsel++) {
1586 zfree(&evsel->name);
1587 zfree(&evsel->core.id);
1593 static struct evsel *read_event_desc(struct feat_fd *ff)
1595 struct evsel *evsel, *events = NULL;
1598 u32 nre, sz, nr, i, j;
1601 /* number of events */
1602 if (do_read_u32(ff, &nre))
1605 if (do_read_u32(ff, &sz))
1608 /* buffer to hold on file attr struct */
1613 /* the last event terminates with evsel->core.attr.size == 0: */
1614 events = calloc(nre + 1, sizeof(*events));
1618 msz = sizeof(evsel->core.attr);
1622 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1626 * must read entire on-file attr struct to
1627 * sync up with layout.
1629 if (__do_read(ff, buf, sz))
1632 if (ff->ph->needs_swap)
1633 perf_event__attr_swap(buf);
1635 memcpy(&evsel->core.attr, buf, msz);
1637 if (do_read_u32(ff, &nr))
1640 if (ff->ph->needs_swap)
1641 evsel->needs_swap = true;
1643 evsel->name = do_read_string(ff);
1650 id = calloc(nr, sizeof(*id));
1653 evsel->core.ids = nr;
1654 evsel->core.id = id;
1656 for (j = 0 ; j < nr; j++) {
1657 if (do_read_u64(ff, id))
1666 free_event_desc(events);
1671 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1672 void *priv __maybe_unused)
1674 return fprintf(fp, ", %s = %s", name, val);
1677 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1679 struct evsel *evsel, *events;
1684 events = ff->events;
1686 events = read_event_desc(ff);
1689 fprintf(fp, "# event desc: not available or unable to read\n");
1693 for (evsel = events; evsel->core.attr.size; evsel++) {
1694 fprintf(fp, "# event : name = %s, ", evsel->name);
1696 if (evsel->core.ids) {
1697 fprintf(fp, ", id = {");
1698 for (j = 0, id = evsel->core.id; j < evsel->core.ids; j++, id++) {
1701 fprintf(fp, " %"PRIu64, *id);
1706 perf_event_attr__fprintf(fp, &evsel->core.attr, __desc_attr__fprintf, NULL);
1711 free_event_desc(events);
1715 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1717 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1720 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1723 struct numa_node *n;
1725 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1726 n = &ff->ph->env.numa_nodes[i];
1728 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1729 " free = %"PRIu64" kB\n",
1730 n->node, n->mem_total, n->mem_free);
1732 fprintf(fp, "# node%u cpu list : ", n->node);
1733 cpu_map__fprintf(n->map, fp);
1737 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1739 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1742 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1744 fprintf(fp, "# contains samples with branch stack\n");
1747 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1749 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1752 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1754 fprintf(fp, "# contains stat data\n");
1757 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1761 fprintf(fp, "# CPU cache info:\n");
1762 for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1764 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1768 static void print_compressed(struct feat_fd *ff, FILE *fp)
1770 fprintf(fp, "# compressed : %s, level = %d, ratio = %d\n",
1771 ff->ph->env.comp_type == PERF_COMP_ZSTD ? "Zstd" : "Unknown",
1772 ff->ph->env.comp_level, ff->ph->env.comp_ratio);
1775 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1777 const char *delimiter = "# pmu mappings: ";
1782 pmu_num = ff->ph->env.nr_pmu_mappings;
1784 fprintf(fp, "# pmu mappings: not available\n");
1788 str = ff->ph->env.pmu_mappings;
1791 type = strtoul(str, &tmp, 0);
1796 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1799 str += strlen(str) + 1;
1808 fprintf(fp, "# pmu mappings: unable to read\n");
1811 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1813 struct perf_session *session;
1814 struct evsel *evsel;
1817 session = container_of(ff->ph, struct perf_session, header);
1819 evlist__for_each_entry(session->evlist, evsel) {
1820 if (perf_evsel__is_group_leader(evsel) &&
1821 evsel->core.nr_members > 1) {
1822 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1823 perf_evsel__name(evsel));
1825 nr = evsel->core.nr_members - 1;
1827 fprintf(fp, ",%s", perf_evsel__name(evsel));
1835 static void print_sample_time(struct feat_fd *ff, FILE *fp)
1837 struct perf_session *session;
1841 session = container_of(ff->ph, struct perf_session, header);
1843 timestamp__scnprintf_usec(session->evlist->first_sample_time,
1844 time_buf, sizeof(time_buf));
1845 fprintf(fp, "# time of first sample : %s\n", time_buf);
1847 timestamp__scnprintf_usec(session->evlist->last_sample_time,
1848 time_buf, sizeof(time_buf));
1849 fprintf(fp, "# time of last sample : %s\n", time_buf);
1851 d = (double)(session->evlist->last_sample_time -
1852 session->evlist->first_sample_time) / NSEC_PER_MSEC;
1854 fprintf(fp, "# sample duration : %10.3f ms\n", d);
1857 static void memory_node__fprintf(struct memory_node *n,
1858 unsigned long long bsize, FILE *fp)
1860 char buf_map[100], buf_size[50];
1861 unsigned long long size;
1863 size = bsize * bitmap_weight(n->set, n->size);
1864 unit_number__scnprintf(buf_size, 50, size);
1866 bitmap_scnprintf(n->set, n->size, buf_map, 100);
1867 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map);
1870 static void print_mem_topology(struct feat_fd *ff, FILE *fp)
1872 struct memory_node *nodes;
1875 nodes = ff->ph->env.memory_nodes;
1876 nr = ff->ph->env.nr_memory_nodes;
1878 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n",
1879 nr, ff->ph->env.memory_bsize);
1881 for (i = 0; i < nr; i++) {
1882 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp);
1886 static int __event_process_build_id(struct perf_record_header_build_id *bev,
1888 struct perf_session *session)
1891 struct machine *machine;
1894 enum dso_kernel_type dso_type;
1896 machine = perf_session__findnew_machine(session, bev->pid);
1900 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1903 case PERF_RECORD_MISC_KERNEL:
1904 dso_type = DSO_TYPE_KERNEL;
1906 case PERF_RECORD_MISC_GUEST_KERNEL:
1907 dso_type = DSO_TYPE_GUEST_KERNEL;
1909 case PERF_RECORD_MISC_USER:
1910 case PERF_RECORD_MISC_GUEST_USER:
1911 dso_type = DSO_TYPE_USER;
1917 dso = machine__findnew_dso(machine, filename);
1919 char sbuild_id[SBUILD_ID_SIZE];
1921 dso__set_build_id(dso, &bev->build_id);
1923 if (dso_type != DSO_TYPE_USER) {
1924 struct kmod_path m = { .name = NULL, };
1926 if (!kmod_path__parse_name(&m, filename) && m.kmod)
1927 dso__set_module_info(dso, &m, machine);
1929 dso->kernel = dso_type;
1934 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1936 pr_debug("build id event received for %s: %s\n",
1937 dso->long_name, sbuild_id);
1946 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1947 int input, u64 offset, u64 size)
1949 struct perf_session *session = container_of(header, struct perf_session, header);
1951 struct perf_event_header header;
1952 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1955 struct perf_record_header_build_id bev;
1956 char filename[PATH_MAX];
1957 u64 limit = offset + size;
1959 while (offset < limit) {
1962 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1965 if (header->needs_swap)
1966 perf_event_header__bswap(&old_bev.header);
1968 len = old_bev.header.size - sizeof(old_bev);
1969 if (readn(input, filename, len) != len)
1972 bev.header = old_bev.header;
1975 * As the pid is the missing value, we need to fill
1976 * it properly. The header.misc value give us nice hint.
1978 bev.pid = HOST_KERNEL_ID;
1979 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1980 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1981 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1983 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1984 __event_process_build_id(&bev, filename, session);
1986 offset += bev.header.size;
1992 static int perf_header__read_build_ids(struct perf_header *header,
1993 int input, u64 offset, u64 size)
1995 struct perf_session *session = container_of(header, struct perf_session, header);
1996 struct perf_record_header_build_id bev;
1997 char filename[PATH_MAX];
1998 u64 limit = offset + size, orig_offset = offset;
2001 while (offset < limit) {
2004 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
2007 if (header->needs_swap)
2008 perf_event_header__bswap(&bev.header);
2010 len = bev.header.size - sizeof(bev);
2011 if (readn(input, filename, len) != len)
2014 * The a1645ce1 changeset:
2016 * "perf: 'perf kvm' tool for monitoring guest performance from host"
2018 * Added a field to struct perf_record_header_build_id that broke the file
2021 * Since the kernel build-id is the first entry, process the
2022 * table using the old format if the well known
2023 * '[kernel.kallsyms]' string for the kernel build-id has the
2024 * first 4 characters chopped off (where the pid_t sits).
2026 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
2027 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
2029 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
2032 __event_process_build_id(&bev, filename, session);
2034 offset += bev.header.size;
2041 /* Macro for features that simply need to read and store a string. */
2042 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
2043 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
2045 ff->ph->env.__feat_env = do_read_string(ff); \
2046 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
2049 FEAT_PROCESS_STR_FUN(hostname, hostname);
2050 FEAT_PROCESS_STR_FUN(osrelease, os_release);
2051 FEAT_PROCESS_STR_FUN(version, version);
2052 FEAT_PROCESS_STR_FUN(arch, arch);
2053 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
2054 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
2056 static int process_tracing_data(struct feat_fd *ff, void *data)
2058 ssize_t ret = trace_report(ff->fd, data, false);
2060 return ret < 0 ? -1 : 0;
2063 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
2065 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
2066 pr_debug("Failed to read buildids, continuing...\n");
2070 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
2073 u32 nr_cpus_avail, nr_cpus_online;
2075 ret = do_read_u32(ff, &nr_cpus_avail);
2079 ret = do_read_u32(ff, &nr_cpus_online);
2082 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
2083 ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
2087 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
2092 ret = do_read_u64(ff, &total_mem);
2095 ff->ph->env.total_mem = (unsigned long long)total_mem;
2099 static struct evsel *
2100 perf_evlist__find_by_index(struct evlist *evlist, int idx)
2102 struct evsel *evsel;
2104 evlist__for_each_entry(evlist, evsel) {
2105 if (evsel->idx == idx)
2113 perf_evlist__set_event_name(struct evlist *evlist,
2114 struct evsel *event)
2116 struct evsel *evsel;
2121 evsel = perf_evlist__find_by_index(evlist, event->idx);
2128 evsel->name = strdup(event->name);
2132 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
2134 struct perf_session *session;
2135 struct evsel *evsel, *events = read_event_desc(ff);
2140 session = container_of(ff->ph, struct perf_session, header);
2142 if (session->data->is_pipe) {
2143 /* Save events for reading later by print_event_desc,
2144 * since they can't be read again in pipe mode. */
2145 ff->events = events;
2148 for (evsel = events; evsel->core.attr.size; evsel++)
2149 perf_evlist__set_event_name(session->evlist, evsel);
2151 if (!session->data->is_pipe)
2152 free_event_desc(events);
2157 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
2159 char *str, *cmdline = NULL, **argv = NULL;
2162 if (do_read_u32(ff, &nr))
2165 ff->ph->env.nr_cmdline = nr;
2167 cmdline = zalloc(ff->size + nr + 1);
2171 argv = zalloc(sizeof(char *) * (nr + 1));
2175 for (i = 0; i < nr; i++) {
2176 str = do_read_string(ff);
2180 argv[i] = cmdline + len;
2181 memcpy(argv[i], str, strlen(str) + 1);
2182 len += strlen(str) + 1;
2185 ff->ph->env.cmdline = cmdline;
2186 ff->ph->env.cmdline_argv = (const char **) argv;
2195 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
2200 int cpu_nr = ff->ph->env.nr_cpus_avail;
2202 struct perf_header *ph = ff->ph;
2203 bool do_core_id_test = true;
2205 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
2209 if (do_read_u32(ff, &nr))
2212 ph->env.nr_sibling_cores = nr;
2213 size += sizeof(u32);
2214 if (strbuf_init(&sb, 128) < 0)
2217 for (i = 0; i < nr; i++) {
2218 str = do_read_string(ff);
2222 /* include a NULL character at the end */
2223 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2225 size += string_size(str);
2228 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
2230 if (do_read_u32(ff, &nr))
2233 ph->env.nr_sibling_threads = nr;
2234 size += sizeof(u32);
2236 for (i = 0; i < nr; i++) {
2237 str = do_read_string(ff);
2241 /* include a NULL character at the end */
2242 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2244 size += string_size(str);
2247 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
2250 * The header may be from old perf,
2251 * which doesn't include core id and socket id information.
2253 if (ff->size <= size) {
2254 zfree(&ph->env.cpu);
2258 /* On s390 the socket_id number is not related to the numbers of cpus.
2259 * The socket_id number might be higher than the numbers of cpus.
2260 * This depends on the configuration.
2261 * AArch64 is the same.
2263 if (ph->env.arch && (!strncmp(ph->env.arch, "s390", 4)
2264 || !strncmp(ph->env.arch, "aarch64", 7)))
2265 do_core_id_test = false;
2267 for (i = 0; i < (u32)cpu_nr; i++) {
2268 if (do_read_u32(ff, &nr))
2271 ph->env.cpu[i].core_id = nr;
2272 size += sizeof(u32);
2274 if (do_read_u32(ff, &nr))
2277 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
2278 pr_debug("socket_id number is too big."
2279 "You may need to upgrade the perf tool.\n");
2283 ph->env.cpu[i].socket_id = nr;
2284 size += sizeof(u32);
2288 * The header may be from old perf,
2289 * which doesn't include die information.
2291 if (ff->size <= size)
2294 if (do_read_u32(ff, &nr))
2297 ph->env.nr_sibling_dies = nr;
2298 size += sizeof(u32);
2300 for (i = 0; i < nr; i++) {
2301 str = do_read_string(ff);
2305 /* include a NULL character at the end */
2306 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2308 size += string_size(str);
2311 ph->env.sibling_dies = strbuf_detach(&sb, NULL);
2313 for (i = 0; i < (u32)cpu_nr; i++) {
2314 if (do_read_u32(ff, &nr))
2317 ph->env.cpu[i].die_id = nr;
2323 strbuf_release(&sb);
2325 zfree(&ph->env.cpu);
2329 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
2331 struct numa_node *nodes, *n;
2336 if (do_read_u32(ff, &nr))
2339 nodes = zalloc(sizeof(*nodes) * nr);
2343 for (i = 0; i < nr; i++) {
2347 if (do_read_u32(ff, &n->node))
2350 if (do_read_u64(ff, &n->mem_total))
2353 if (do_read_u64(ff, &n->mem_free))
2356 str = do_read_string(ff);
2360 n->map = perf_cpu_map__new(str);
2366 ff->ph->env.nr_numa_nodes = nr;
2367 ff->ph->env.numa_nodes = nodes;
2375 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
2382 if (do_read_u32(ff, &pmu_num))
2386 pr_debug("pmu mappings not available\n");
2390 ff->ph->env.nr_pmu_mappings = pmu_num;
2391 if (strbuf_init(&sb, 128) < 0)
2395 if (do_read_u32(ff, &type))
2398 name = do_read_string(ff);
2402 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
2404 /* include a NULL character at the end */
2405 if (strbuf_add(&sb, "", 1) < 0)
2408 if (!strcmp(name, "msr"))
2409 ff->ph->env.msr_pmu_type = type;
2414 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2418 strbuf_release(&sb);
2422 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2425 u32 i, nr, nr_groups;
2426 struct perf_session *session;
2427 struct evsel *evsel, *leader = NULL;
2434 if (do_read_u32(ff, &nr_groups))
2437 ff->ph->env.nr_groups = nr_groups;
2439 pr_debug("group desc not available\n");
2443 desc = calloc(nr_groups, sizeof(*desc));
2447 for (i = 0; i < nr_groups; i++) {
2448 desc[i].name = do_read_string(ff);
2452 if (do_read_u32(ff, &desc[i].leader_idx))
2455 if (do_read_u32(ff, &desc[i].nr_members))
2460 * Rebuild group relationship based on the group_desc
2462 session = container_of(ff->ph, struct perf_session, header);
2463 session->evlist->nr_groups = nr_groups;
2466 evlist__for_each_entry(session->evlist, evsel) {
2467 if (evsel->idx == (int) desc[i].leader_idx) {
2468 evsel->leader = evsel;
2469 /* {anon_group} is a dummy name */
2470 if (strcmp(desc[i].name, "{anon_group}")) {
2471 evsel->group_name = desc[i].name;
2472 desc[i].name = NULL;
2474 evsel->core.nr_members = desc[i].nr_members;
2476 if (i >= nr_groups || nr > 0) {
2477 pr_debug("invalid group desc\n");
2482 nr = evsel->core.nr_members - 1;
2485 /* This is a group member */
2486 evsel->leader = leader;
2492 if (i != nr_groups || nr != 0) {
2493 pr_debug("invalid group desc\n");
2499 for (i = 0; i < nr_groups; i++)
2500 zfree(&desc[i].name);
2506 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2508 struct perf_session *session;
2511 session = container_of(ff->ph, struct perf_session, header);
2513 err = auxtrace_index__process(ff->fd, ff->size, session,
2514 ff->ph->needs_swap);
2516 pr_err("Failed to process auxtrace index\n");
2520 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2522 struct cpu_cache_level *caches;
2523 u32 cnt, i, version;
2525 if (do_read_u32(ff, &version))
2531 if (do_read_u32(ff, &cnt))
2534 caches = zalloc(sizeof(*caches) * cnt);
2538 for (i = 0; i < cnt; i++) {
2539 struct cpu_cache_level c;
2542 if (do_read_u32(ff, &c.v))\
2543 goto out_free_caches; \
2552 c.v = do_read_string(ff); \
2554 goto out_free_caches;
2564 ff->ph->env.caches = caches;
2565 ff->ph->env.caches_cnt = cnt;
2572 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused)
2574 struct perf_session *session;
2575 u64 first_sample_time, last_sample_time;
2578 session = container_of(ff->ph, struct perf_session, header);
2580 ret = do_read_u64(ff, &first_sample_time);
2584 ret = do_read_u64(ff, &last_sample_time);
2588 session->evlist->first_sample_time = first_sample_time;
2589 session->evlist->last_sample_time = last_sample_time;
2593 static int process_mem_topology(struct feat_fd *ff,
2594 void *data __maybe_unused)
2596 struct memory_node *nodes;
2597 u64 version, i, nr, bsize;
2600 if (do_read_u64(ff, &version))
2606 if (do_read_u64(ff, &bsize))
2609 if (do_read_u64(ff, &nr))
2612 nodes = zalloc(sizeof(*nodes) * nr);
2616 for (i = 0; i < nr; i++) {
2617 struct memory_node n;
2620 if (do_read_u64(ff, &n.v)) \
2628 if (do_read_bitmap(ff, &n.set, &n.size))
2634 ff->ph->env.memory_bsize = bsize;
2635 ff->ph->env.memory_nodes = nodes;
2636 ff->ph->env.nr_memory_nodes = nr;
2645 static int process_clockid(struct feat_fd *ff,
2646 void *data __maybe_unused)
2648 if (do_read_u64(ff, &ff->ph->env.clockid_res_ns))
2654 static int process_dir_format(struct feat_fd *ff,
2655 void *_data __maybe_unused)
2657 struct perf_session *session;
2658 struct perf_data *data;
2660 session = container_of(ff->ph, struct perf_session, header);
2661 data = session->data;
2663 if (WARN_ON(!perf_data__is_dir(data)))
2666 return do_read_u64(ff, &data->dir.version);
2669 #ifdef HAVE_LIBBPF_SUPPORT
2670 static int process_bpf_prog_info(struct feat_fd *ff, void *data __maybe_unused)
2672 struct bpf_prog_info_linear *info_linear;
2673 struct bpf_prog_info_node *info_node;
2674 struct perf_env *env = &ff->ph->env;
2678 if (ff->ph->needs_swap) {
2679 pr_warning("interpreting bpf_prog_info from systems with endianity is not yet supported\n");
2683 if (do_read_u32(ff, &count))
2686 down_write(&env->bpf_progs.lock);
2688 for (i = 0; i < count; ++i) {
2689 u32 info_len, data_len;
2693 if (do_read_u32(ff, &info_len))
2695 if (do_read_u32(ff, &data_len))
2698 if (info_len > sizeof(struct bpf_prog_info)) {
2699 pr_warning("detected invalid bpf_prog_info\n");
2703 info_linear = malloc(sizeof(struct bpf_prog_info_linear) +
2707 info_linear->info_len = sizeof(struct bpf_prog_info);
2708 info_linear->data_len = data_len;
2709 if (do_read_u64(ff, (u64 *)(&info_linear->arrays)))
2711 if (__do_read(ff, &info_linear->info, info_len))
2713 if (info_len < sizeof(struct bpf_prog_info))
2714 memset(((void *)(&info_linear->info)) + info_len, 0,
2715 sizeof(struct bpf_prog_info) - info_len);
2717 if (__do_read(ff, info_linear->data, data_len))
2720 info_node = malloc(sizeof(struct bpf_prog_info_node));
2724 /* after reading from file, translate offset to address */
2725 bpf_program__bpil_offs_to_addr(info_linear);
2726 info_node->info_linear = info_linear;
2727 perf_env__insert_bpf_prog_info(env, info_node);
2730 up_write(&env->bpf_progs.lock);
2735 up_write(&env->bpf_progs.lock);
2738 #else // HAVE_LIBBPF_SUPPORT
2739 static int process_bpf_prog_info(struct feat_fd *ff __maybe_unused, void *data __maybe_unused)
2743 #endif // HAVE_LIBBPF_SUPPORT
2745 static int process_bpf_btf(struct feat_fd *ff, void *data __maybe_unused)
2747 struct perf_env *env = &ff->ph->env;
2748 struct btf_node *node = NULL;
2752 if (ff->ph->needs_swap) {
2753 pr_warning("interpreting btf from systems with endianity is not yet supported\n");
2757 if (do_read_u32(ff, &count))
2760 down_write(&env->bpf_progs.lock);
2762 for (i = 0; i < count; ++i) {
2765 if (do_read_u32(ff, &id))
2767 if (do_read_u32(ff, &data_size))
2770 node = malloc(sizeof(struct btf_node) + data_size);
2775 node->data_size = data_size;
2777 if (__do_read(ff, node->data, data_size))
2780 perf_env__insert_btf(env, node);
2786 up_write(&env->bpf_progs.lock);
2791 static int process_compressed(struct feat_fd *ff,
2792 void *data __maybe_unused)
2794 if (do_read_u32(ff, &(ff->ph->env.comp_ver)))
2797 if (do_read_u32(ff, &(ff->ph->env.comp_type)))
2800 if (do_read_u32(ff, &(ff->ph->env.comp_level)))
2803 if (do_read_u32(ff, &(ff->ph->env.comp_ratio)))
2806 if (do_read_u32(ff, &(ff->ph->env.comp_mmap_len)))
2812 #define FEAT_OPR(n, func, __full_only) \
2814 .name = __stringify(n), \
2815 .write = write_##func, \
2816 .print = print_##func, \
2817 .full_only = __full_only, \
2818 .process = process_##func, \
2819 .synthesize = true \
2822 #define FEAT_OPN(n, func, __full_only) \
2824 .name = __stringify(n), \
2825 .write = write_##func, \
2826 .print = print_##func, \
2827 .full_only = __full_only, \
2828 .process = process_##func \
2831 /* feature_ops not implemented: */
2832 #define print_tracing_data NULL
2833 #define print_build_id NULL
2835 #define process_branch_stack NULL
2836 #define process_stat NULL
2838 // Only used in util/synthetic-events.c
2839 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
2841 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2842 FEAT_OPN(TRACING_DATA, tracing_data, false),
2843 FEAT_OPN(BUILD_ID, build_id, false),
2844 FEAT_OPR(HOSTNAME, hostname, false),
2845 FEAT_OPR(OSRELEASE, osrelease, false),
2846 FEAT_OPR(VERSION, version, false),
2847 FEAT_OPR(ARCH, arch, false),
2848 FEAT_OPR(NRCPUS, nrcpus, false),
2849 FEAT_OPR(CPUDESC, cpudesc, false),
2850 FEAT_OPR(CPUID, cpuid, false),
2851 FEAT_OPR(TOTAL_MEM, total_mem, false),
2852 FEAT_OPR(EVENT_DESC, event_desc, false),
2853 FEAT_OPR(CMDLINE, cmdline, false),
2854 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true),
2855 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true),
2856 FEAT_OPN(BRANCH_STACK, branch_stack, false),
2857 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false),
2858 FEAT_OPR(GROUP_DESC, group_desc, false),
2859 FEAT_OPN(AUXTRACE, auxtrace, false),
2860 FEAT_OPN(STAT, stat, false),
2861 FEAT_OPN(CACHE, cache, true),
2862 FEAT_OPR(SAMPLE_TIME, sample_time, false),
2863 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true),
2864 FEAT_OPR(CLOCKID, clockid, false),
2865 FEAT_OPN(DIR_FORMAT, dir_format, false),
2866 FEAT_OPR(BPF_PROG_INFO, bpf_prog_info, false),
2867 FEAT_OPR(BPF_BTF, bpf_btf, false),
2868 FEAT_OPR(COMPRESSED, compressed, false),
2871 struct header_print_data {
2873 bool full; /* extended list of headers */
2876 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2877 struct perf_header *ph,
2878 int feat, int fd, void *data)
2880 struct header_print_data *hd = data;
2883 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2884 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2885 "%d, continuing...\n", section->offset, feat);
2888 if (feat >= HEADER_LAST_FEATURE) {
2889 pr_warning("unknown feature %d\n", feat);
2892 if (!feat_ops[feat].print)
2895 ff = (struct feat_fd) {
2900 if (!feat_ops[feat].full_only || hd->full)
2901 feat_ops[feat].print(&ff, hd->fp);
2903 fprintf(hd->fp, "# %s info available, use -I to display\n",
2904 feat_ops[feat].name);
2909 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2911 struct header_print_data hd;
2912 struct perf_header *header = &session->header;
2913 int fd = perf_data__fd(session->data);
2921 ret = fstat(fd, &st);
2925 stctime = st.st_ctime;
2926 fprintf(fp, "# captured on : %s", ctime(&stctime));
2928 fprintf(fp, "# header version : %u\n", header->version);
2929 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset);
2930 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size);
2931 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset);
2933 perf_header__process_sections(header, fd, &hd,
2934 perf_file_section__fprintf_info);
2936 if (session->data->is_pipe)
2939 fprintf(fp, "# missing features: ");
2940 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
2942 fprintf(fp, "%s ", feat_ops[bit].name);
2949 static int do_write_feat(struct feat_fd *ff, int type,
2950 struct perf_file_section **p,
2951 struct evlist *evlist)
2956 if (perf_header__has_feat(ff->ph, type)) {
2957 if (!feat_ops[type].write)
2960 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
2963 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
2965 err = feat_ops[type].write(ff, evlist);
2967 pr_debug("failed to write feature %s\n", feat_ops[type].name);
2969 /* undo anything written */
2970 lseek(ff->fd, (*p)->offset, SEEK_SET);
2974 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
2980 static int perf_header__adds_write(struct perf_header *header,
2981 struct evlist *evlist, int fd)
2985 struct perf_file_section *feat_sec, *p;
2991 ff = (struct feat_fd){
2996 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3000 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
3001 if (feat_sec == NULL)
3004 sec_size = sizeof(*feat_sec) * nr_sections;
3006 sec_start = header->feat_offset;
3007 lseek(fd, sec_start + sec_size, SEEK_SET);
3009 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3010 if (do_write_feat(&ff, feat, &p, evlist))
3011 perf_header__clear_feat(header, feat);
3014 lseek(fd, sec_start, SEEK_SET);
3016 * may write more than needed due to dropped feature, but
3017 * this is okay, reader will skip the missing entries
3019 err = do_write(&ff, feat_sec, sec_size);
3021 pr_debug("failed to write feature section\n");
3026 int perf_header__write_pipe(int fd)
3028 struct perf_pipe_file_header f_header;
3032 ff = (struct feat_fd){ .fd = fd };
3034 f_header = (struct perf_pipe_file_header){
3035 .magic = PERF_MAGIC,
3036 .size = sizeof(f_header),
3039 err = do_write(&ff, &f_header, sizeof(f_header));
3041 pr_debug("failed to write perf pipe header\n");
3048 int perf_session__write_header(struct perf_session *session,
3049 struct evlist *evlist,
3050 int fd, bool at_exit)
3052 struct perf_file_header f_header;
3053 struct perf_file_attr f_attr;
3054 struct perf_header *header = &session->header;
3055 struct evsel *evsel;
3060 ff = (struct feat_fd){ .fd = fd};
3061 lseek(fd, sizeof(f_header), SEEK_SET);
3063 evlist__for_each_entry(session->evlist, evsel) {
3064 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
3065 err = do_write(&ff, evsel->core.id, evsel->core.ids * sizeof(u64));
3067 pr_debug("failed to write perf header\n");
3072 attr_offset = lseek(ff.fd, 0, SEEK_CUR);
3074 evlist__for_each_entry(evlist, evsel) {
3075 f_attr = (struct perf_file_attr){
3076 .attr = evsel->core.attr,
3078 .offset = evsel->id_offset,
3079 .size = evsel->core.ids * sizeof(u64),
3082 err = do_write(&ff, &f_attr, sizeof(f_attr));
3084 pr_debug("failed to write perf header attribute\n");
3089 if (!header->data_offset)
3090 header->data_offset = lseek(fd, 0, SEEK_CUR);
3091 header->feat_offset = header->data_offset + header->data_size;
3094 err = perf_header__adds_write(header, evlist, fd);
3099 f_header = (struct perf_file_header){
3100 .magic = PERF_MAGIC,
3101 .size = sizeof(f_header),
3102 .attr_size = sizeof(f_attr),
3104 .offset = attr_offset,
3105 .size = evlist->core.nr_entries * sizeof(f_attr),
3108 .offset = header->data_offset,
3109 .size = header->data_size,
3111 /* event_types is ignored, store zeros */
3114 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
3116 lseek(fd, 0, SEEK_SET);
3117 err = do_write(&ff, &f_header, sizeof(f_header));
3119 pr_debug("failed to write perf header\n");
3122 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
3127 static int perf_header__getbuffer64(struct perf_header *header,
3128 int fd, void *buf, size_t size)
3130 if (readn(fd, buf, size) <= 0)
3133 if (header->needs_swap)
3134 mem_bswap_64(buf, size);
3139 int perf_header__process_sections(struct perf_header *header, int fd,
3141 int (*process)(struct perf_file_section *section,
3142 struct perf_header *ph,
3143 int feat, int fd, void *data))
3145 struct perf_file_section *feat_sec, *sec;
3151 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3155 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
3159 sec_size = sizeof(*feat_sec) * nr_sections;
3161 lseek(fd, header->feat_offset, SEEK_SET);
3163 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
3167 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
3168 err = process(sec++, header, feat, fd, data);
3178 static const int attr_file_abi_sizes[] = {
3179 [0] = PERF_ATTR_SIZE_VER0,
3180 [1] = PERF_ATTR_SIZE_VER1,
3181 [2] = PERF_ATTR_SIZE_VER2,
3182 [3] = PERF_ATTR_SIZE_VER3,
3183 [4] = PERF_ATTR_SIZE_VER4,
3188 * In the legacy file format, the magic number is not used to encode endianness.
3189 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
3190 * on ABI revisions, we need to try all combinations for all endianness to
3191 * detect the endianness.
3193 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
3195 uint64_t ref_size, attr_size;
3198 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
3199 ref_size = attr_file_abi_sizes[i]
3200 + sizeof(struct perf_file_section);
3201 if (hdr_sz != ref_size) {
3202 attr_size = bswap_64(hdr_sz);
3203 if (attr_size != ref_size)
3206 ph->needs_swap = true;
3208 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
3213 /* could not determine endianness */
3217 #define PERF_PIPE_HDR_VER0 16
3219 static const size_t attr_pipe_abi_sizes[] = {
3220 [0] = PERF_PIPE_HDR_VER0,
3225 * In the legacy pipe format, there is an implicit assumption that endiannesss
3226 * between host recording the samples, and host parsing the samples is the
3227 * same. This is not always the case given that the pipe output may always be
3228 * redirected into a file and analyzed on a different machine with possibly a
3229 * different endianness and perf_event ABI revsions in the perf tool itself.
3231 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
3236 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
3237 if (hdr_sz != attr_pipe_abi_sizes[i]) {
3238 attr_size = bswap_64(hdr_sz);
3239 if (attr_size != hdr_sz)
3242 ph->needs_swap = true;
3244 pr_debug("Pipe ABI%d perf.data file detected\n", i);
3250 bool is_perf_magic(u64 magic)
3252 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
3253 || magic == __perf_magic2
3254 || magic == __perf_magic2_sw)
3260 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
3261 bool is_pipe, struct perf_header *ph)
3265 /* check for legacy format */
3266 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
3268 ph->version = PERF_HEADER_VERSION_1;
3269 pr_debug("legacy perf.data format\n");
3271 return try_all_pipe_abis(hdr_sz, ph);
3273 return try_all_file_abis(hdr_sz, ph);
3276 * the new magic number serves two purposes:
3277 * - unique number to identify actual perf.data files
3278 * - encode endianness of file
3280 ph->version = PERF_HEADER_VERSION_2;
3282 /* check magic number with one endianness */
3283 if (magic == __perf_magic2)
3286 /* check magic number with opposite endianness */
3287 if (magic != __perf_magic2_sw)
3290 ph->needs_swap = true;
3295 int perf_file_header__read(struct perf_file_header *header,
3296 struct perf_header *ph, int fd)
3300 lseek(fd, 0, SEEK_SET);
3302 ret = readn(fd, header, sizeof(*header));
3306 if (check_magic_endian(header->magic,
3307 header->attr_size, false, ph) < 0) {
3308 pr_debug("magic/endian check failed\n");
3312 if (ph->needs_swap) {
3313 mem_bswap_64(header, offsetof(struct perf_file_header,
3317 if (header->size != sizeof(*header)) {
3318 /* Support the previous format */
3319 if (header->size == offsetof(typeof(*header), adds_features))
3320 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3323 } else if (ph->needs_swap) {
3325 * feature bitmap is declared as an array of unsigned longs --
3326 * not good since its size can differ between the host that
3327 * generated the data file and the host analyzing the file.
3329 * We need to handle endianness, but we don't know the size of
3330 * the unsigned long where the file was generated. Take a best
3331 * guess at determining it: try 64-bit swap first (ie., file
3332 * created on a 64-bit host), and check if the hostname feature
3333 * bit is set (this feature bit is forced on as of fbe96f2).
3334 * If the bit is not, undo the 64-bit swap and try a 32-bit
3335 * swap. If the hostname bit is still not set (e.g., older data
3336 * file), punt and fallback to the original behavior --
3337 * clearing all feature bits and setting buildid.
3339 mem_bswap_64(&header->adds_features,
3340 BITS_TO_U64(HEADER_FEAT_BITS));
3342 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3344 mem_bswap_64(&header->adds_features,
3345 BITS_TO_U64(HEADER_FEAT_BITS));
3348 mem_bswap_32(&header->adds_features,
3349 BITS_TO_U32(HEADER_FEAT_BITS));
3352 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3353 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3354 set_bit(HEADER_BUILD_ID, header->adds_features);
3358 memcpy(&ph->adds_features, &header->adds_features,
3359 sizeof(ph->adds_features));
3361 ph->data_offset = header->data.offset;
3362 ph->data_size = header->data.size;
3363 ph->feat_offset = header->data.offset + header->data.size;
3367 static int perf_file_section__process(struct perf_file_section *section,
3368 struct perf_header *ph,
3369 int feat, int fd, void *data)
3371 struct feat_fd fdd = {
3374 .size = section->size,
3375 .offset = section->offset,
3378 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3379 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3380 "%d, continuing...\n", section->offset, feat);
3384 if (feat >= HEADER_LAST_FEATURE) {
3385 pr_debug("unknown feature %d, continuing...\n", feat);
3389 if (!feat_ops[feat].process)
3392 return feat_ops[feat].process(&fdd, data);
3395 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
3396 struct perf_header *ph, int fd,
3399 struct feat_fd ff = {
3400 .fd = STDOUT_FILENO,
3405 ret = readn(fd, header, sizeof(*header));
3409 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
3410 pr_debug("endian/magic failed\n");
3415 header->size = bswap_64(header->size);
3417 if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
3423 static int perf_header__read_pipe(struct perf_session *session)
3425 struct perf_header *header = &session->header;
3426 struct perf_pipe_file_header f_header;
3428 if (perf_file_header__read_pipe(&f_header, header,
3429 perf_data__fd(session->data),
3430 session->repipe) < 0) {
3431 pr_debug("incompatible file format\n");
3438 static int read_attr(int fd, struct perf_header *ph,
3439 struct perf_file_attr *f_attr)
3441 struct perf_event_attr *attr = &f_attr->attr;
3443 size_t our_sz = sizeof(f_attr->attr);
3446 memset(f_attr, 0, sizeof(*f_attr));
3448 /* read minimal guaranteed structure */
3449 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
3451 pr_debug("cannot read %d bytes of header attr\n",
3452 PERF_ATTR_SIZE_VER0);
3456 /* on file perf_event_attr size */
3464 sz = PERF_ATTR_SIZE_VER0;
3465 } else if (sz > our_sz) {
3466 pr_debug("file uses a more recent and unsupported ABI"
3467 " (%zu bytes extra)\n", sz - our_sz);
3470 /* what we have not yet read and that we know about */
3471 left = sz - PERF_ATTR_SIZE_VER0;
3474 ptr += PERF_ATTR_SIZE_VER0;
3476 ret = readn(fd, ptr, left);
3478 /* read perf_file_section, ids are read in caller */
3479 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
3481 return ret <= 0 ? -1 : 0;
3484 static int perf_evsel__prepare_tracepoint_event(struct evsel *evsel,
3485 struct tep_handle *pevent)
3487 struct tep_event *event;
3490 /* already prepared */
3491 if (evsel->tp_format)
3494 if (pevent == NULL) {
3495 pr_debug("broken or missing trace data\n");
3499 event = tep_find_event(pevent, evsel->core.attr.config);
3500 if (event == NULL) {
3501 pr_debug("cannot find event format for %d\n", (int)evsel->core.attr.config);
3506 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
3507 evsel->name = strdup(bf);
3508 if (evsel->name == NULL)
3512 evsel->tp_format = event;
3516 static int perf_evlist__prepare_tracepoint_events(struct evlist *evlist,
3517 struct tep_handle *pevent)
3521 evlist__for_each_entry(evlist, pos) {
3522 if (pos->core.attr.type == PERF_TYPE_TRACEPOINT &&
3523 perf_evsel__prepare_tracepoint_event(pos, pevent))
3530 int perf_session__read_header(struct perf_session *session)
3532 struct perf_data *data = session->data;
3533 struct perf_header *header = &session->header;
3534 struct perf_file_header f_header;
3535 struct perf_file_attr f_attr;
3537 int nr_attrs, nr_ids, i, j;
3538 int fd = perf_data__fd(data);
3540 session->evlist = evlist__new();
3541 if (session->evlist == NULL)
3544 session->evlist->env = &header->env;
3545 session->machines.host.env = &header->env;
3546 if (perf_data__is_pipe(data))
3547 return perf_header__read_pipe(session);
3549 if (perf_file_header__read(&f_header, header, fd) < 0)
3553 * Sanity check that perf.data was written cleanly; data size is
3554 * initialized to 0 and updated only if the on_exit function is run.
3555 * If data size is still 0 then the file contains only partial
3556 * information. Just warn user and process it as much as it can.
3558 if (f_header.data.size == 0) {
3559 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3560 "Was the 'perf record' command properly terminated?\n",
3564 if (f_header.attr_size == 0) {
3565 pr_err("ERROR: The %s file's attr size field is 0 which is unexpected.\n"
3566 "Was the 'perf record' command properly terminated?\n",
3571 nr_attrs = f_header.attrs.size / f_header.attr_size;
3572 lseek(fd, f_header.attrs.offset, SEEK_SET);
3574 for (i = 0; i < nr_attrs; i++) {
3575 struct evsel *evsel;
3578 if (read_attr(fd, header, &f_attr) < 0)
3581 if (header->needs_swap) {
3582 f_attr.ids.size = bswap_64(f_attr.ids.size);
3583 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
3584 perf_event__attr_swap(&f_attr.attr);
3587 tmp = lseek(fd, 0, SEEK_CUR);
3588 evsel = evsel__new(&f_attr.attr);
3591 goto out_delete_evlist;
3593 evsel->needs_swap = header->needs_swap;
3595 * Do it before so that if perf_evsel__alloc_id fails, this
3596 * entry gets purged too at evlist__delete().
3598 evlist__add(session->evlist, evsel);
3600 nr_ids = f_attr.ids.size / sizeof(u64);
3602 * We don't have the cpu and thread maps on the header, so
3603 * for allocating the perf_sample_id table we fake 1 cpu and
3604 * hattr->ids threads.
3606 if (perf_evsel__alloc_id(&evsel->core, 1, nr_ids))
3607 goto out_delete_evlist;
3609 lseek(fd, f_attr.ids.offset, SEEK_SET);
3611 for (j = 0; j < nr_ids; j++) {
3612 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
3615 perf_evlist__id_add(&session->evlist->core, &evsel->core, 0, j, f_id);
3618 lseek(fd, tmp, SEEK_SET);
3621 perf_header__process_sections(header, fd, &session->tevent,
3622 perf_file_section__process);
3624 if (perf_evlist__prepare_tracepoint_events(session->evlist,
3625 session->tevent.pevent))
3626 goto out_delete_evlist;
3633 evlist__delete(session->evlist);
3634 session->evlist = NULL;
3638 int perf_event__process_feature(struct perf_session *session,
3639 union perf_event *event)
3641 struct perf_tool *tool = session->tool;
3642 struct feat_fd ff = { .fd = 0 };
3643 struct perf_record_header_feature *fe = (struct perf_record_header_feature *)event;
3644 int type = fe->header.type;
3645 u64 feat = fe->feat_id;
3647 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3648 pr_warning("invalid record type %d in pipe-mode\n", type);
3651 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) {
3652 pr_warning("invalid record type %d in pipe-mode\n", type);
3656 if (!feat_ops[feat].process)
3659 ff.buf = (void *)fe->data;
3660 ff.size = event->header.size - sizeof(*fe);
3661 ff.ph = &session->header;
3663 if (feat_ops[feat].process(&ff, NULL))
3666 if (!feat_ops[feat].print || !tool->show_feat_hdr)
3669 if (!feat_ops[feat].full_only ||
3670 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3671 feat_ops[feat].print(&ff, stdout);
3673 fprintf(stdout, "# %s info available, use -I to display\n",
3674 feat_ops[feat].name);
3680 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3682 struct perf_record_event_update *ev = &event->event_update;
3683 struct perf_record_event_update_scale *ev_scale;
3684 struct perf_record_event_update_cpus *ev_cpus;
3685 struct perf_cpu_map *map;
3688 ret = fprintf(fp, "\n... id: %" PRI_lu64 "\n", ev->id);
3691 case PERF_EVENT_UPDATE__SCALE:
3692 ev_scale = (struct perf_record_event_update_scale *)ev->data;
3693 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3695 case PERF_EVENT_UPDATE__UNIT:
3696 ret += fprintf(fp, "... unit: %s\n", ev->data);
3698 case PERF_EVENT_UPDATE__NAME:
3699 ret += fprintf(fp, "... name: %s\n", ev->data);
3701 case PERF_EVENT_UPDATE__CPUS:
3702 ev_cpus = (struct perf_record_event_update_cpus *)ev->data;
3703 ret += fprintf(fp, "... ");
3705 map = cpu_map__new_data(&ev_cpus->cpus);
3707 ret += cpu_map__fprintf(map, fp);
3709 ret += fprintf(fp, "failed to get cpus\n");
3712 ret += fprintf(fp, "... unknown type\n");
3719 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3720 union perf_event *event,
3721 struct evlist **pevlist)
3724 struct evsel *evsel;
3725 struct evlist *evlist = *pevlist;
3727 if (evlist == NULL) {
3728 *pevlist = evlist = evlist__new();
3733 evsel = evsel__new(&event->attr.attr);
3737 evlist__add(evlist, evsel);
3739 ids = event->header.size;
3740 ids -= (void *)&event->attr.id - (void *)event;
3741 n_ids = ids / sizeof(u64);
3743 * We don't have the cpu and thread maps on the header, so
3744 * for allocating the perf_sample_id table we fake 1 cpu and
3745 * hattr->ids threads.
3747 if (perf_evsel__alloc_id(&evsel->core, 1, n_ids))
3750 for (i = 0; i < n_ids; i++) {
3751 perf_evlist__id_add(&evlist->core, &evsel->core, 0, i, event->attr.id[i]);
3757 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3758 union perf_event *event,
3759 struct evlist **pevlist)
3761 struct perf_record_event_update *ev = &event->event_update;
3762 struct perf_record_event_update_scale *ev_scale;
3763 struct perf_record_event_update_cpus *ev_cpus;
3764 struct evlist *evlist;
3765 struct evsel *evsel;
3766 struct perf_cpu_map *map;
3768 if (!pevlist || *pevlist == NULL)
3773 evsel = perf_evlist__id2evsel(evlist, ev->id);
3778 case PERF_EVENT_UPDATE__UNIT:
3779 evsel->unit = strdup(ev->data);
3781 case PERF_EVENT_UPDATE__NAME:
3782 evsel->name = strdup(ev->data);
3784 case PERF_EVENT_UPDATE__SCALE:
3785 ev_scale = (struct perf_record_event_update_scale *)ev->data;
3786 evsel->scale = ev_scale->scale;
3788 case PERF_EVENT_UPDATE__CPUS:
3789 ev_cpus = (struct perf_record_event_update_cpus *)ev->data;
3791 map = cpu_map__new_data(&ev_cpus->cpus);
3793 evsel->core.own_cpus = map;
3795 pr_err("failed to get event_update cpus\n");
3803 int perf_event__process_tracing_data(struct perf_session *session,
3804 union perf_event *event)
3806 ssize_t size_read, padding, size = event->tracing_data.size;
3807 int fd = perf_data__fd(session->data);
3808 off_t offset = lseek(fd, 0, SEEK_CUR);
3811 /* setup for reading amidst mmap */
3812 lseek(fd, offset + sizeof(struct perf_record_header_tracing_data),
3815 size_read = trace_report(fd, &session->tevent,
3817 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3819 if (readn(fd, buf, padding) < 0) {
3820 pr_err("%s: reading input file", __func__);
3823 if (session->repipe) {
3824 int retw = write(STDOUT_FILENO, buf, padding);
3825 if (retw <= 0 || retw != padding) {
3826 pr_err("%s: repiping tracing data padding", __func__);
3831 if (size_read + padding != size) {
3832 pr_err("%s: tracing data size mismatch", __func__);
3836 perf_evlist__prepare_tracepoint_events(session->evlist,
3837 session->tevent.pevent);
3839 return size_read + padding;
3842 int perf_event__process_build_id(struct perf_session *session,
3843 union perf_event *event)
3845 __event_process_build_id(&event->build_id,
3846 event->build_id.filename,