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 int write_cpu_pmu_caps(struct feat_fd *ff,
1399 struct evlist *evlist __maybe_unused)
1401 struct perf_pmu *cpu_pmu = perf_pmu__find("cpu");
1402 struct perf_pmu_caps *caps = NULL;
1409 nr_caps = perf_pmu__caps_parse(cpu_pmu);
1413 ret = do_write(ff, &nr_caps, sizeof(nr_caps));
1417 list_for_each_entry(caps, &cpu_pmu->caps, list) {
1418 ret = do_write_string(ff, caps->name);
1422 ret = do_write_string(ff, caps->value);
1430 static void print_hostname(struct feat_fd *ff, FILE *fp)
1432 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1435 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1437 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1440 static void print_arch(struct feat_fd *ff, FILE *fp)
1442 fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1445 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1447 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1450 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1452 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1453 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1456 static void print_version(struct feat_fd *ff, FILE *fp)
1458 fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1461 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1465 nr = ff->ph->env.nr_cmdline;
1467 fprintf(fp, "# cmdline : ");
1469 for (i = 0; i < nr; i++) {
1470 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]);
1472 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1476 char *quote = strchr(argv_i, '\'');
1480 fprintf(fp, "%s\\\'", argv_i);
1483 fprintf(fp, "%s ", argv_i);
1490 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1492 struct perf_header *ph = ff->ph;
1493 int cpu_nr = ph->env.nr_cpus_avail;
1497 nr = ph->env.nr_sibling_cores;
1498 str = ph->env.sibling_cores;
1500 for (i = 0; i < nr; i++) {
1501 fprintf(fp, "# sibling sockets : %s\n", str);
1502 str += strlen(str) + 1;
1505 if (ph->env.nr_sibling_dies) {
1506 nr = ph->env.nr_sibling_dies;
1507 str = ph->env.sibling_dies;
1509 for (i = 0; i < nr; i++) {
1510 fprintf(fp, "# sibling dies : %s\n", str);
1511 str += strlen(str) + 1;
1515 nr = ph->env.nr_sibling_threads;
1516 str = ph->env.sibling_threads;
1518 for (i = 0; i < nr; i++) {
1519 fprintf(fp, "# sibling threads : %s\n", str);
1520 str += strlen(str) + 1;
1523 if (ph->env.nr_sibling_dies) {
1524 if (ph->env.cpu != NULL) {
1525 for (i = 0; i < cpu_nr; i++)
1526 fprintf(fp, "# CPU %d: Core ID %d, "
1527 "Die ID %d, Socket ID %d\n",
1528 i, ph->env.cpu[i].core_id,
1529 ph->env.cpu[i].die_id,
1530 ph->env.cpu[i].socket_id);
1532 fprintf(fp, "# Core ID, Die ID and Socket ID "
1533 "information is not available\n");
1535 if (ph->env.cpu != NULL) {
1536 for (i = 0; i < cpu_nr; i++)
1537 fprintf(fp, "# CPU %d: Core ID %d, "
1539 i, ph->env.cpu[i].core_id,
1540 ph->env.cpu[i].socket_id);
1542 fprintf(fp, "# Core ID and Socket ID "
1543 "information is not available\n");
1547 static void print_clockid(struct feat_fd *ff, FILE *fp)
1549 fprintf(fp, "# clockid frequency: %"PRIu64" MHz\n",
1550 ff->ph->env.clockid_res_ns * 1000);
1553 static void print_dir_format(struct feat_fd *ff, FILE *fp)
1555 struct perf_session *session;
1556 struct perf_data *data;
1558 session = container_of(ff->ph, struct perf_session, header);
1559 data = session->data;
1561 fprintf(fp, "# directory data version : %"PRIu64"\n", data->dir.version);
1564 static void print_bpf_prog_info(struct feat_fd *ff, FILE *fp)
1566 struct perf_env *env = &ff->ph->env;
1567 struct rb_root *root;
1568 struct rb_node *next;
1570 down_read(&env->bpf_progs.lock);
1572 root = &env->bpf_progs.infos;
1573 next = rb_first(root);
1576 struct bpf_prog_info_node *node;
1578 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
1579 next = rb_next(&node->rb_node);
1581 bpf_event__print_bpf_prog_info(&node->info_linear->info,
1585 up_read(&env->bpf_progs.lock);
1588 static void print_bpf_btf(struct feat_fd *ff, FILE *fp)
1590 struct perf_env *env = &ff->ph->env;
1591 struct rb_root *root;
1592 struct rb_node *next;
1594 down_read(&env->bpf_progs.lock);
1596 root = &env->bpf_progs.btfs;
1597 next = rb_first(root);
1600 struct btf_node *node;
1602 node = rb_entry(next, struct btf_node, rb_node);
1603 next = rb_next(&node->rb_node);
1604 fprintf(fp, "# btf info of id %u\n", node->id);
1607 up_read(&env->bpf_progs.lock);
1610 static void free_event_desc(struct evsel *events)
1612 struct evsel *evsel;
1617 for (evsel = events; evsel->core.attr.size; evsel++) {
1618 zfree(&evsel->name);
1619 zfree(&evsel->core.id);
1625 static bool perf_attr_check(struct perf_event_attr *attr)
1627 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3) {
1628 pr_warning("Reserved bits are set unexpectedly. "
1629 "Please update perf tool.\n");
1633 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) {
1634 pr_warning("Unknown sample type (0x%llx) is detected. "
1635 "Please update perf tool.\n",
1640 if (attr->read_format & ~(PERF_FORMAT_MAX-1)) {
1641 pr_warning("Unknown read format (0x%llx) is detected. "
1642 "Please update perf tool.\n",
1647 if ((attr->sample_type & PERF_SAMPLE_BRANCH_STACK) &&
1648 (attr->branch_sample_type & ~(PERF_SAMPLE_BRANCH_MAX-1))) {
1649 pr_warning("Unknown branch sample type (0x%llx) is detected. "
1650 "Please update perf tool.\n",
1651 attr->branch_sample_type);
1659 static struct evsel *read_event_desc(struct feat_fd *ff)
1661 struct evsel *evsel, *events = NULL;
1664 u32 nre, sz, nr, i, j;
1667 /* number of events */
1668 if (do_read_u32(ff, &nre))
1671 if (do_read_u32(ff, &sz))
1674 /* buffer to hold on file attr struct */
1679 /* the last event terminates with evsel->core.attr.size == 0: */
1680 events = calloc(nre + 1, sizeof(*events));
1684 msz = sizeof(evsel->core.attr);
1688 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1692 * must read entire on-file attr struct to
1693 * sync up with layout.
1695 if (__do_read(ff, buf, sz))
1698 if (ff->ph->needs_swap)
1699 perf_event__attr_swap(buf);
1701 memcpy(&evsel->core.attr, buf, msz);
1703 if (!perf_attr_check(&evsel->core.attr))
1706 if (do_read_u32(ff, &nr))
1709 if (ff->ph->needs_swap)
1710 evsel->needs_swap = true;
1712 evsel->name = do_read_string(ff);
1719 id = calloc(nr, sizeof(*id));
1722 evsel->core.ids = nr;
1723 evsel->core.id = id;
1725 for (j = 0 ; j < nr; j++) {
1726 if (do_read_u64(ff, id))
1735 free_event_desc(events);
1740 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1741 void *priv __maybe_unused)
1743 return fprintf(fp, ", %s = %s", name, val);
1746 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1748 struct evsel *evsel, *events;
1753 events = ff->events;
1755 events = read_event_desc(ff);
1758 fprintf(fp, "# event desc: not available or unable to read\n");
1762 for (evsel = events; evsel->core.attr.size; evsel++) {
1763 fprintf(fp, "# event : name = %s, ", evsel->name);
1765 if (evsel->core.ids) {
1766 fprintf(fp, ", id = {");
1767 for (j = 0, id = evsel->core.id; j < evsel->core.ids; j++, id++) {
1770 fprintf(fp, " %"PRIu64, *id);
1775 perf_event_attr__fprintf(fp, &evsel->core.attr, __desc_attr__fprintf, NULL);
1780 free_event_desc(events);
1784 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1786 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1789 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1792 struct numa_node *n;
1794 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1795 n = &ff->ph->env.numa_nodes[i];
1797 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1798 " free = %"PRIu64" kB\n",
1799 n->node, n->mem_total, n->mem_free);
1801 fprintf(fp, "# node%u cpu list : ", n->node);
1802 cpu_map__fprintf(n->map, fp);
1806 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1808 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1811 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1813 fprintf(fp, "# contains samples with branch stack\n");
1816 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1818 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1821 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1823 fprintf(fp, "# contains stat data\n");
1826 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1830 fprintf(fp, "# CPU cache info:\n");
1831 for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1833 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1837 static void print_compressed(struct feat_fd *ff, FILE *fp)
1839 fprintf(fp, "# compressed : %s, level = %d, ratio = %d\n",
1840 ff->ph->env.comp_type == PERF_COMP_ZSTD ? "Zstd" : "Unknown",
1841 ff->ph->env.comp_level, ff->ph->env.comp_ratio);
1844 static void print_cpu_pmu_caps(struct feat_fd *ff, FILE *fp)
1846 const char *delimiter = "# cpu pmu capabilities: ";
1847 u32 nr_caps = ff->ph->env.nr_cpu_pmu_caps;
1851 fprintf(fp, "# cpu pmu capabilities: not available\n");
1855 str = ff->ph->env.cpu_pmu_caps;
1857 fprintf(fp, "%s%s", delimiter, str);
1859 str += strlen(str) + 1;
1865 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1867 const char *delimiter = "# pmu mappings: ";
1872 pmu_num = ff->ph->env.nr_pmu_mappings;
1874 fprintf(fp, "# pmu mappings: not available\n");
1878 str = ff->ph->env.pmu_mappings;
1881 type = strtoul(str, &tmp, 0);
1886 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1889 str += strlen(str) + 1;
1898 fprintf(fp, "# pmu mappings: unable to read\n");
1901 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1903 struct perf_session *session;
1904 struct evsel *evsel;
1907 session = container_of(ff->ph, struct perf_session, header);
1909 evlist__for_each_entry(session->evlist, evsel) {
1910 if (perf_evsel__is_group_leader(evsel) &&
1911 evsel->core.nr_members > 1) {
1912 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1913 perf_evsel__name(evsel));
1915 nr = evsel->core.nr_members - 1;
1917 fprintf(fp, ",%s", perf_evsel__name(evsel));
1925 static void print_sample_time(struct feat_fd *ff, FILE *fp)
1927 struct perf_session *session;
1931 session = container_of(ff->ph, struct perf_session, header);
1933 timestamp__scnprintf_usec(session->evlist->first_sample_time,
1934 time_buf, sizeof(time_buf));
1935 fprintf(fp, "# time of first sample : %s\n", time_buf);
1937 timestamp__scnprintf_usec(session->evlist->last_sample_time,
1938 time_buf, sizeof(time_buf));
1939 fprintf(fp, "# time of last sample : %s\n", time_buf);
1941 d = (double)(session->evlist->last_sample_time -
1942 session->evlist->first_sample_time) / NSEC_PER_MSEC;
1944 fprintf(fp, "# sample duration : %10.3f ms\n", d);
1947 static void memory_node__fprintf(struct memory_node *n,
1948 unsigned long long bsize, FILE *fp)
1950 char buf_map[100], buf_size[50];
1951 unsigned long long size;
1953 size = bsize * bitmap_weight(n->set, n->size);
1954 unit_number__scnprintf(buf_size, 50, size);
1956 bitmap_scnprintf(n->set, n->size, buf_map, 100);
1957 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map);
1960 static void print_mem_topology(struct feat_fd *ff, FILE *fp)
1962 struct memory_node *nodes;
1965 nodes = ff->ph->env.memory_nodes;
1966 nr = ff->ph->env.nr_memory_nodes;
1968 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n",
1969 nr, ff->ph->env.memory_bsize);
1971 for (i = 0; i < nr; i++) {
1972 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp);
1976 static int __event_process_build_id(struct perf_record_header_build_id *bev,
1978 struct perf_session *session)
1981 struct machine *machine;
1984 enum dso_kernel_type dso_type;
1986 machine = perf_session__findnew_machine(session, bev->pid);
1990 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1993 case PERF_RECORD_MISC_KERNEL:
1994 dso_type = DSO_TYPE_KERNEL;
1996 case PERF_RECORD_MISC_GUEST_KERNEL:
1997 dso_type = DSO_TYPE_GUEST_KERNEL;
1999 case PERF_RECORD_MISC_USER:
2000 case PERF_RECORD_MISC_GUEST_USER:
2001 dso_type = DSO_TYPE_USER;
2007 dso = machine__findnew_dso(machine, filename);
2009 char sbuild_id[SBUILD_ID_SIZE];
2011 dso__set_build_id(dso, &bev->build_id);
2013 if (dso_type != DSO_TYPE_USER) {
2014 struct kmod_path m = { .name = NULL, };
2016 if (!kmod_path__parse_name(&m, filename) && m.kmod)
2017 dso__set_module_info(dso, &m, machine);
2019 dso->kernel = dso_type;
2024 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
2026 pr_debug("build id event received for %s: %s\n",
2027 dso->long_name, sbuild_id);
2036 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
2037 int input, u64 offset, u64 size)
2039 struct perf_session *session = container_of(header, struct perf_session, header);
2041 struct perf_event_header header;
2042 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
2045 struct perf_record_header_build_id bev;
2046 char filename[PATH_MAX];
2047 u64 limit = offset + size;
2049 while (offset < limit) {
2052 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
2055 if (header->needs_swap)
2056 perf_event_header__bswap(&old_bev.header);
2058 len = old_bev.header.size - sizeof(old_bev);
2059 if (readn(input, filename, len) != len)
2062 bev.header = old_bev.header;
2065 * As the pid is the missing value, we need to fill
2066 * it properly. The header.misc value give us nice hint.
2068 bev.pid = HOST_KERNEL_ID;
2069 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
2070 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
2071 bev.pid = DEFAULT_GUEST_KERNEL_ID;
2073 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
2074 __event_process_build_id(&bev, filename, session);
2076 offset += bev.header.size;
2082 static int perf_header__read_build_ids(struct perf_header *header,
2083 int input, u64 offset, u64 size)
2085 struct perf_session *session = container_of(header, struct perf_session, header);
2086 struct perf_record_header_build_id bev;
2087 char filename[PATH_MAX];
2088 u64 limit = offset + size, orig_offset = offset;
2091 while (offset < limit) {
2094 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
2097 if (header->needs_swap)
2098 perf_event_header__bswap(&bev.header);
2100 len = bev.header.size - sizeof(bev);
2101 if (readn(input, filename, len) != len)
2104 * The a1645ce1 changeset:
2106 * "perf: 'perf kvm' tool for monitoring guest performance from host"
2108 * Added a field to struct perf_record_header_build_id that broke the file
2111 * Since the kernel build-id is the first entry, process the
2112 * table using the old format if the well known
2113 * '[kernel.kallsyms]' string for the kernel build-id has the
2114 * first 4 characters chopped off (where the pid_t sits).
2116 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
2117 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
2119 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
2122 __event_process_build_id(&bev, filename, session);
2124 offset += bev.header.size;
2131 /* Macro for features that simply need to read and store a string. */
2132 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
2133 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
2135 ff->ph->env.__feat_env = do_read_string(ff); \
2136 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
2139 FEAT_PROCESS_STR_FUN(hostname, hostname);
2140 FEAT_PROCESS_STR_FUN(osrelease, os_release);
2141 FEAT_PROCESS_STR_FUN(version, version);
2142 FEAT_PROCESS_STR_FUN(arch, arch);
2143 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
2144 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
2146 static int process_tracing_data(struct feat_fd *ff, void *data)
2148 ssize_t ret = trace_report(ff->fd, data, false);
2150 return ret < 0 ? -1 : 0;
2153 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
2155 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
2156 pr_debug("Failed to read buildids, continuing...\n");
2160 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
2163 u32 nr_cpus_avail, nr_cpus_online;
2165 ret = do_read_u32(ff, &nr_cpus_avail);
2169 ret = do_read_u32(ff, &nr_cpus_online);
2172 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
2173 ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
2177 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
2182 ret = do_read_u64(ff, &total_mem);
2185 ff->ph->env.total_mem = (unsigned long long)total_mem;
2189 static struct evsel *
2190 perf_evlist__find_by_index(struct evlist *evlist, int idx)
2192 struct evsel *evsel;
2194 evlist__for_each_entry(evlist, evsel) {
2195 if (evsel->idx == idx)
2203 perf_evlist__set_event_name(struct evlist *evlist,
2204 struct evsel *event)
2206 struct evsel *evsel;
2211 evsel = perf_evlist__find_by_index(evlist, event->idx);
2218 evsel->name = strdup(event->name);
2222 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
2224 struct perf_session *session;
2225 struct evsel *evsel, *events = read_event_desc(ff);
2230 session = container_of(ff->ph, struct perf_session, header);
2232 if (session->data->is_pipe) {
2233 /* Save events for reading later by print_event_desc,
2234 * since they can't be read again in pipe mode. */
2235 ff->events = events;
2238 for (evsel = events; evsel->core.attr.size; evsel++)
2239 perf_evlist__set_event_name(session->evlist, evsel);
2241 if (!session->data->is_pipe)
2242 free_event_desc(events);
2247 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
2249 char *str, *cmdline = NULL, **argv = NULL;
2252 if (do_read_u32(ff, &nr))
2255 ff->ph->env.nr_cmdline = nr;
2257 cmdline = zalloc(ff->size + nr + 1);
2261 argv = zalloc(sizeof(char *) * (nr + 1));
2265 for (i = 0; i < nr; i++) {
2266 str = do_read_string(ff);
2270 argv[i] = cmdline + len;
2271 memcpy(argv[i], str, strlen(str) + 1);
2272 len += strlen(str) + 1;
2275 ff->ph->env.cmdline = cmdline;
2276 ff->ph->env.cmdline_argv = (const char **) argv;
2285 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
2290 int cpu_nr = ff->ph->env.nr_cpus_avail;
2292 struct perf_header *ph = ff->ph;
2293 bool do_core_id_test = true;
2295 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
2299 if (do_read_u32(ff, &nr))
2302 ph->env.nr_sibling_cores = nr;
2303 size += sizeof(u32);
2304 if (strbuf_init(&sb, 128) < 0)
2307 for (i = 0; i < nr; i++) {
2308 str = do_read_string(ff);
2312 /* include a NULL character at the end */
2313 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2315 size += string_size(str);
2318 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
2320 if (do_read_u32(ff, &nr))
2323 ph->env.nr_sibling_threads = nr;
2324 size += sizeof(u32);
2326 for (i = 0; i < nr; i++) {
2327 str = do_read_string(ff);
2331 /* include a NULL character at the end */
2332 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2334 size += string_size(str);
2337 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
2340 * The header may be from old perf,
2341 * which doesn't include core id and socket id information.
2343 if (ff->size <= size) {
2344 zfree(&ph->env.cpu);
2348 /* On s390 the socket_id number is not related to the numbers of cpus.
2349 * The socket_id number might be higher than the numbers of cpus.
2350 * This depends on the configuration.
2351 * AArch64 is the same.
2353 if (ph->env.arch && (!strncmp(ph->env.arch, "s390", 4)
2354 || !strncmp(ph->env.arch, "aarch64", 7)))
2355 do_core_id_test = false;
2357 for (i = 0; i < (u32)cpu_nr; i++) {
2358 if (do_read_u32(ff, &nr))
2361 ph->env.cpu[i].core_id = nr;
2362 size += sizeof(u32);
2364 if (do_read_u32(ff, &nr))
2367 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
2368 pr_debug("socket_id number is too big."
2369 "You may need to upgrade the perf tool.\n");
2373 ph->env.cpu[i].socket_id = nr;
2374 size += sizeof(u32);
2378 * The header may be from old perf,
2379 * which doesn't include die information.
2381 if (ff->size <= size)
2384 if (do_read_u32(ff, &nr))
2387 ph->env.nr_sibling_dies = nr;
2388 size += sizeof(u32);
2390 for (i = 0; i < nr; i++) {
2391 str = do_read_string(ff);
2395 /* include a NULL character at the end */
2396 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2398 size += string_size(str);
2401 ph->env.sibling_dies = strbuf_detach(&sb, NULL);
2403 for (i = 0; i < (u32)cpu_nr; i++) {
2404 if (do_read_u32(ff, &nr))
2407 ph->env.cpu[i].die_id = nr;
2413 strbuf_release(&sb);
2415 zfree(&ph->env.cpu);
2419 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
2421 struct numa_node *nodes, *n;
2426 if (do_read_u32(ff, &nr))
2429 nodes = zalloc(sizeof(*nodes) * nr);
2433 for (i = 0; i < nr; i++) {
2437 if (do_read_u32(ff, &n->node))
2440 if (do_read_u64(ff, &n->mem_total))
2443 if (do_read_u64(ff, &n->mem_free))
2446 str = do_read_string(ff);
2450 n->map = perf_cpu_map__new(str);
2456 ff->ph->env.nr_numa_nodes = nr;
2457 ff->ph->env.numa_nodes = nodes;
2465 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
2472 if (do_read_u32(ff, &pmu_num))
2476 pr_debug("pmu mappings not available\n");
2480 ff->ph->env.nr_pmu_mappings = pmu_num;
2481 if (strbuf_init(&sb, 128) < 0)
2485 if (do_read_u32(ff, &type))
2488 name = do_read_string(ff);
2492 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
2494 /* include a NULL character at the end */
2495 if (strbuf_add(&sb, "", 1) < 0)
2498 if (!strcmp(name, "msr"))
2499 ff->ph->env.msr_pmu_type = type;
2504 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2508 strbuf_release(&sb);
2512 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2515 u32 i, nr, nr_groups;
2516 struct perf_session *session;
2517 struct evsel *evsel, *leader = NULL;
2524 if (do_read_u32(ff, &nr_groups))
2527 ff->ph->env.nr_groups = nr_groups;
2529 pr_debug("group desc not available\n");
2533 desc = calloc(nr_groups, sizeof(*desc));
2537 for (i = 0; i < nr_groups; i++) {
2538 desc[i].name = do_read_string(ff);
2542 if (do_read_u32(ff, &desc[i].leader_idx))
2545 if (do_read_u32(ff, &desc[i].nr_members))
2550 * Rebuild group relationship based on the group_desc
2552 session = container_of(ff->ph, struct perf_session, header);
2553 session->evlist->nr_groups = nr_groups;
2556 evlist__for_each_entry(session->evlist, evsel) {
2557 if (evsel->idx == (int) desc[i].leader_idx) {
2558 evsel->leader = evsel;
2559 /* {anon_group} is a dummy name */
2560 if (strcmp(desc[i].name, "{anon_group}")) {
2561 evsel->group_name = desc[i].name;
2562 desc[i].name = NULL;
2564 evsel->core.nr_members = desc[i].nr_members;
2566 if (i >= nr_groups || nr > 0) {
2567 pr_debug("invalid group desc\n");
2572 nr = evsel->core.nr_members - 1;
2575 /* This is a group member */
2576 evsel->leader = leader;
2582 if (i != nr_groups || nr != 0) {
2583 pr_debug("invalid group desc\n");
2589 for (i = 0; i < nr_groups; i++)
2590 zfree(&desc[i].name);
2596 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2598 struct perf_session *session;
2601 session = container_of(ff->ph, struct perf_session, header);
2603 err = auxtrace_index__process(ff->fd, ff->size, session,
2604 ff->ph->needs_swap);
2606 pr_err("Failed to process auxtrace index\n");
2610 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2612 struct cpu_cache_level *caches;
2613 u32 cnt, i, version;
2615 if (do_read_u32(ff, &version))
2621 if (do_read_u32(ff, &cnt))
2624 caches = zalloc(sizeof(*caches) * cnt);
2628 for (i = 0; i < cnt; i++) {
2629 struct cpu_cache_level c;
2632 if (do_read_u32(ff, &c.v))\
2633 goto out_free_caches; \
2642 c.v = do_read_string(ff); \
2644 goto out_free_caches;
2654 ff->ph->env.caches = caches;
2655 ff->ph->env.caches_cnt = cnt;
2662 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused)
2664 struct perf_session *session;
2665 u64 first_sample_time, last_sample_time;
2668 session = container_of(ff->ph, struct perf_session, header);
2670 ret = do_read_u64(ff, &first_sample_time);
2674 ret = do_read_u64(ff, &last_sample_time);
2678 session->evlist->first_sample_time = first_sample_time;
2679 session->evlist->last_sample_time = last_sample_time;
2683 static int process_mem_topology(struct feat_fd *ff,
2684 void *data __maybe_unused)
2686 struct memory_node *nodes;
2687 u64 version, i, nr, bsize;
2690 if (do_read_u64(ff, &version))
2696 if (do_read_u64(ff, &bsize))
2699 if (do_read_u64(ff, &nr))
2702 nodes = zalloc(sizeof(*nodes) * nr);
2706 for (i = 0; i < nr; i++) {
2707 struct memory_node n;
2710 if (do_read_u64(ff, &n.v)) \
2718 if (do_read_bitmap(ff, &n.set, &n.size))
2724 ff->ph->env.memory_bsize = bsize;
2725 ff->ph->env.memory_nodes = nodes;
2726 ff->ph->env.nr_memory_nodes = nr;
2735 static int process_clockid(struct feat_fd *ff,
2736 void *data __maybe_unused)
2738 if (do_read_u64(ff, &ff->ph->env.clockid_res_ns))
2744 static int process_dir_format(struct feat_fd *ff,
2745 void *_data __maybe_unused)
2747 struct perf_session *session;
2748 struct perf_data *data;
2750 session = container_of(ff->ph, struct perf_session, header);
2751 data = session->data;
2753 if (WARN_ON(!perf_data__is_dir(data)))
2756 return do_read_u64(ff, &data->dir.version);
2759 #ifdef HAVE_LIBBPF_SUPPORT
2760 static int process_bpf_prog_info(struct feat_fd *ff, void *data __maybe_unused)
2762 struct bpf_prog_info_linear *info_linear;
2763 struct bpf_prog_info_node *info_node;
2764 struct perf_env *env = &ff->ph->env;
2768 if (ff->ph->needs_swap) {
2769 pr_warning("interpreting bpf_prog_info from systems with endianity is not yet supported\n");
2773 if (do_read_u32(ff, &count))
2776 down_write(&env->bpf_progs.lock);
2778 for (i = 0; i < count; ++i) {
2779 u32 info_len, data_len;
2783 if (do_read_u32(ff, &info_len))
2785 if (do_read_u32(ff, &data_len))
2788 if (info_len > sizeof(struct bpf_prog_info)) {
2789 pr_warning("detected invalid bpf_prog_info\n");
2793 info_linear = malloc(sizeof(struct bpf_prog_info_linear) +
2797 info_linear->info_len = sizeof(struct bpf_prog_info);
2798 info_linear->data_len = data_len;
2799 if (do_read_u64(ff, (u64 *)(&info_linear->arrays)))
2801 if (__do_read(ff, &info_linear->info, info_len))
2803 if (info_len < sizeof(struct bpf_prog_info))
2804 memset(((void *)(&info_linear->info)) + info_len, 0,
2805 sizeof(struct bpf_prog_info) - info_len);
2807 if (__do_read(ff, info_linear->data, data_len))
2810 info_node = malloc(sizeof(struct bpf_prog_info_node));
2814 /* after reading from file, translate offset to address */
2815 bpf_program__bpil_offs_to_addr(info_linear);
2816 info_node->info_linear = info_linear;
2817 perf_env__insert_bpf_prog_info(env, info_node);
2820 up_write(&env->bpf_progs.lock);
2825 up_write(&env->bpf_progs.lock);
2828 #else // HAVE_LIBBPF_SUPPORT
2829 static int process_bpf_prog_info(struct feat_fd *ff __maybe_unused, void *data __maybe_unused)
2833 #endif // HAVE_LIBBPF_SUPPORT
2835 static int process_bpf_btf(struct feat_fd *ff, void *data __maybe_unused)
2837 struct perf_env *env = &ff->ph->env;
2838 struct btf_node *node = NULL;
2842 if (ff->ph->needs_swap) {
2843 pr_warning("interpreting btf from systems with endianity is not yet supported\n");
2847 if (do_read_u32(ff, &count))
2850 down_write(&env->bpf_progs.lock);
2852 for (i = 0; i < count; ++i) {
2855 if (do_read_u32(ff, &id))
2857 if (do_read_u32(ff, &data_size))
2860 node = malloc(sizeof(struct btf_node) + data_size);
2865 node->data_size = data_size;
2867 if (__do_read(ff, node->data, data_size))
2870 perf_env__insert_btf(env, node);
2876 up_write(&env->bpf_progs.lock);
2881 static int process_compressed(struct feat_fd *ff,
2882 void *data __maybe_unused)
2884 if (do_read_u32(ff, &(ff->ph->env.comp_ver)))
2887 if (do_read_u32(ff, &(ff->ph->env.comp_type)))
2890 if (do_read_u32(ff, &(ff->ph->env.comp_level)))
2893 if (do_read_u32(ff, &(ff->ph->env.comp_ratio)))
2896 if (do_read_u32(ff, &(ff->ph->env.comp_mmap_len)))
2902 static int process_cpu_pmu_caps(struct feat_fd *ff,
2903 void *data __maybe_unused)
2909 if (do_read_u32(ff, &nr_caps))
2913 pr_debug("cpu pmu capabilities not available\n");
2917 ff->ph->env.nr_cpu_pmu_caps = nr_caps;
2919 if (strbuf_init(&sb, 128) < 0)
2923 name = do_read_string(ff);
2927 value = do_read_string(ff);
2931 if (strbuf_addf(&sb, "%s=%s", name, value) < 0)
2934 /* include a NULL character at the end */
2935 if (strbuf_add(&sb, "", 1) < 0)
2938 if (!strcmp(name, "branches"))
2939 ff->ph->env.max_branches = atoi(value);
2944 ff->ph->env.cpu_pmu_caps = strbuf_detach(&sb, NULL);
2952 strbuf_release(&sb);
2956 #define FEAT_OPR(n, func, __full_only) \
2958 .name = __stringify(n), \
2959 .write = write_##func, \
2960 .print = print_##func, \
2961 .full_only = __full_only, \
2962 .process = process_##func, \
2963 .synthesize = true \
2966 #define FEAT_OPN(n, func, __full_only) \
2968 .name = __stringify(n), \
2969 .write = write_##func, \
2970 .print = print_##func, \
2971 .full_only = __full_only, \
2972 .process = process_##func \
2975 /* feature_ops not implemented: */
2976 #define print_tracing_data NULL
2977 #define print_build_id NULL
2979 #define process_branch_stack NULL
2980 #define process_stat NULL
2982 // Only used in util/synthetic-events.c
2983 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
2985 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2986 FEAT_OPN(TRACING_DATA, tracing_data, false),
2987 FEAT_OPN(BUILD_ID, build_id, false),
2988 FEAT_OPR(HOSTNAME, hostname, false),
2989 FEAT_OPR(OSRELEASE, osrelease, false),
2990 FEAT_OPR(VERSION, version, false),
2991 FEAT_OPR(ARCH, arch, false),
2992 FEAT_OPR(NRCPUS, nrcpus, false),
2993 FEAT_OPR(CPUDESC, cpudesc, false),
2994 FEAT_OPR(CPUID, cpuid, false),
2995 FEAT_OPR(TOTAL_MEM, total_mem, false),
2996 FEAT_OPR(EVENT_DESC, event_desc, false),
2997 FEAT_OPR(CMDLINE, cmdline, false),
2998 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true),
2999 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true),
3000 FEAT_OPN(BRANCH_STACK, branch_stack, false),
3001 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false),
3002 FEAT_OPR(GROUP_DESC, group_desc, false),
3003 FEAT_OPN(AUXTRACE, auxtrace, false),
3004 FEAT_OPN(STAT, stat, false),
3005 FEAT_OPN(CACHE, cache, true),
3006 FEAT_OPR(SAMPLE_TIME, sample_time, false),
3007 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true),
3008 FEAT_OPR(CLOCKID, clockid, false),
3009 FEAT_OPN(DIR_FORMAT, dir_format, false),
3010 FEAT_OPR(BPF_PROG_INFO, bpf_prog_info, false),
3011 FEAT_OPR(BPF_BTF, bpf_btf, false),
3012 FEAT_OPR(COMPRESSED, compressed, false),
3013 FEAT_OPR(CPU_PMU_CAPS, cpu_pmu_caps, false),
3016 struct header_print_data {
3018 bool full; /* extended list of headers */
3021 static int perf_file_section__fprintf_info(struct perf_file_section *section,
3022 struct perf_header *ph,
3023 int feat, int fd, void *data)
3025 struct header_print_data *hd = data;
3028 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3029 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3030 "%d, continuing...\n", section->offset, feat);
3033 if (feat >= HEADER_LAST_FEATURE) {
3034 pr_warning("unknown feature %d\n", feat);
3037 if (!feat_ops[feat].print)
3040 ff = (struct feat_fd) {
3045 if (!feat_ops[feat].full_only || hd->full)
3046 feat_ops[feat].print(&ff, hd->fp);
3048 fprintf(hd->fp, "# %s info available, use -I to display\n",
3049 feat_ops[feat].name);
3054 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
3056 struct header_print_data hd;
3057 struct perf_header *header = &session->header;
3058 int fd = perf_data__fd(session->data);
3066 ret = fstat(fd, &st);
3070 stctime = st.st_mtime;
3071 fprintf(fp, "# captured on : %s", ctime(&stctime));
3073 fprintf(fp, "# header version : %u\n", header->version);
3074 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset);
3075 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size);
3076 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset);
3078 perf_header__process_sections(header, fd, &hd,
3079 perf_file_section__fprintf_info);
3081 if (session->data->is_pipe)
3084 fprintf(fp, "# missing features: ");
3085 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
3087 fprintf(fp, "%s ", feat_ops[bit].name);
3094 static int do_write_feat(struct feat_fd *ff, int type,
3095 struct perf_file_section **p,
3096 struct evlist *evlist)
3101 if (perf_header__has_feat(ff->ph, type)) {
3102 if (!feat_ops[type].write)
3105 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
3108 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
3110 err = feat_ops[type].write(ff, evlist);
3112 pr_debug("failed to write feature %s\n", feat_ops[type].name);
3114 /* undo anything written */
3115 lseek(ff->fd, (*p)->offset, SEEK_SET);
3119 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
3125 static int perf_header__adds_write(struct perf_header *header,
3126 struct evlist *evlist, int fd)
3130 struct perf_file_section *feat_sec, *p;
3136 ff = (struct feat_fd){
3141 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3145 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
3146 if (feat_sec == NULL)
3149 sec_size = sizeof(*feat_sec) * nr_sections;
3151 sec_start = header->feat_offset;
3152 lseek(fd, sec_start + sec_size, SEEK_SET);
3154 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3155 if (do_write_feat(&ff, feat, &p, evlist))
3156 perf_header__clear_feat(header, feat);
3159 lseek(fd, sec_start, SEEK_SET);
3161 * may write more than needed due to dropped feature, but
3162 * this is okay, reader will skip the missing entries
3164 err = do_write(&ff, feat_sec, sec_size);
3166 pr_debug("failed to write feature section\n");
3171 int perf_header__write_pipe(int fd)
3173 struct perf_pipe_file_header f_header;
3177 ff = (struct feat_fd){ .fd = fd };
3179 f_header = (struct perf_pipe_file_header){
3180 .magic = PERF_MAGIC,
3181 .size = sizeof(f_header),
3184 err = do_write(&ff, &f_header, sizeof(f_header));
3186 pr_debug("failed to write perf pipe header\n");
3193 int perf_session__write_header(struct perf_session *session,
3194 struct evlist *evlist,
3195 int fd, bool at_exit)
3197 struct perf_file_header f_header;
3198 struct perf_file_attr f_attr;
3199 struct perf_header *header = &session->header;
3200 struct evsel *evsel;
3205 ff = (struct feat_fd){ .fd = fd};
3206 lseek(fd, sizeof(f_header), SEEK_SET);
3208 evlist__for_each_entry(session->evlist, evsel) {
3209 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
3210 err = do_write(&ff, evsel->core.id, evsel->core.ids * sizeof(u64));
3212 pr_debug("failed to write perf header\n");
3217 attr_offset = lseek(ff.fd, 0, SEEK_CUR);
3219 evlist__for_each_entry(evlist, evsel) {
3220 f_attr = (struct perf_file_attr){
3221 .attr = evsel->core.attr,
3223 .offset = evsel->id_offset,
3224 .size = evsel->core.ids * sizeof(u64),
3227 err = do_write(&ff, &f_attr, sizeof(f_attr));
3229 pr_debug("failed to write perf header attribute\n");
3234 if (!header->data_offset)
3235 header->data_offset = lseek(fd, 0, SEEK_CUR);
3236 header->feat_offset = header->data_offset + header->data_size;
3239 err = perf_header__adds_write(header, evlist, fd);
3244 f_header = (struct perf_file_header){
3245 .magic = PERF_MAGIC,
3246 .size = sizeof(f_header),
3247 .attr_size = sizeof(f_attr),
3249 .offset = attr_offset,
3250 .size = evlist->core.nr_entries * sizeof(f_attr),
3253 .offset = header->data_offset,
3254 .size = header->data_size,
3256 /* event_types is ignored, store zeros */
3259 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
3261 lseek(fd, 0, SEEK_SET);
3262 err = do_write(&ff, &f_header, sizeof(f_header));
3264 pr_debug("failed to write perf header\n");
3267 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
3272 static int perf_header__getbuffer64(struct perf_header *header,
3273 int fd, void *buf, size_t size)
3275 if (readn(fd, buf, size) <= 0)
3278 if (header->needs_swap)
3279 mem_bswap_64(buf, size);
3284 int perf_header__process_sections(struct perf_header *header, int fd,
3286 int (*process)(struct perf_file_section *section,
3287 struct perf_header *ph,
3288 int feat, int fd, void *data))
3290 struct perf_file_section *feat_sec, *sec;
3296 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3300 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
3304 sec_size = sizeof(*feat_sec) * nr_sections;
3306 lseek(fd, header->feat_offset, SEEK_SET);
3308 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
3312 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
3313 err = process(sec++, header, feat, fd, data);
3323 static const int attr_file_abi_sizes[] = {
3324 [0] = PERF_ATTR_SIZE_VER0,
3325 [1] = PERF_ATTR_SIZE_VER1,
3326 [2] = PERF_ATTR_SIZE_VER2,
3327 [3] = PERF_ATTR_SIZE_VER3,
3328 [4] = PERF_ATTR_SIZE_VER4,
3333 * In the legacy file format, the magic number is not used to encode endianness.
3334 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
3335 * on ABI revisions, we need to try all combinations for all endianness to
3336 * detect the endianness.
3338 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
3340 uint64_t ref_size, attr_size;
3343 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
3344 ref_size = attr_file_abi_sizes[i]
3345 + sizeof(struct perf_file_section);
3346 if (hdr_sz != ref_size) {
3347 attr_size = bswap_64(hdr_sz);
3348 if (attr_size != ref_size)
3351 ph->needs_swap = true;
3353 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
3358 /* could not determine endianness */
3362 #define PERF_PIPE_HDR_VER0 16
3364 static const size_t attr_pipe_abi_sizes[] = {
3365 [0] = PERF_PIPE_HDR_VER0,
3370 * In the legacy pipe format, there is an implicit assumption that endiannesss
3371 * between host recording the samples, and host parsing the samples is the
3372 * same. This is not always the case given that the pipe output may always be
3373 * redirected into a file and analyzed on a different machine with possibly a
3374 * different endianness and perf_event ABI revsions in the perf tool itself.
3376 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
3381 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
3382 if (hdr_sz != attr_pipe_abi_sizes[i]) {
3383 attr_size = bswap_64(hdr_sz);
3384 if (attr_size != hdr_sz)
3387 ph->needs_swap = true;
3389 pr_debug("Pipe ABI%d perf.data file detected\n", i);
3395 bool is_perf_magic(u64 magic)
3397 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
3398 || magic == __perf_magic2
3399 || magic == __perf_magic2_sw)
3405 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
3406 bool is_pipe, struct perf_header *ph)
3410 /* check for legacy format */
3411 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
3413 ph->version = PERF_HEADER_VERSION_1;
3414 pr_debug("legacy perf.data format\n");
3416 return try_all_pipe_abis(hdr_sz, ph);
3418 return try_all_file_abis(hdr_sz, ph);
3421 * the new magic number serves two purposes:
3422 * - unique number to identify actual perf.data files
3423 * - encode endianness of file
3425 ph->version = PERF_HEADER_VERSION_2;
3427 /* check magic number with one endianness */
3428 if (magic == __perf_magic2)
3431 /* check magic number with opposite endianness */
3432 if (magic != __perf_magic2_sw)
3435 ph->needs_swap = true;
3440 int perf_file_header__read(struct perf_file_header *header,
3441 struct perf_header *ph, int fd)
3445 lseek(fd, 0, SEEK_SET);
3447 ret = readn(fd, header, sizeof(*header));
3451 if (check_magic_endian(header->magic,
3452 header->attr_size, false, ph) < 0) {
3453 pr_debug("magic/endian check failed\n");
3457 if (ph->needs_swap) {
3458 mem_bswap_64(header, offsetof(struct perf_file_header,
3462 if (header->size != sizeof(*header)) {
3463 /* Support the previous format */
3464 if (header->size == offsetof(typeof(*header), adds_features))
3465 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3468 } else if (ph->needs_swap) {
3470 * feature bitmap is declared as an array of unsigned longs --
3471 * not good since its size can differ between the host that
3472 * generated the data file and the host analyzing the file.
3474 * We need to handle endianness, but we don't know the size of
3475 * the unsigned long where the file was generated. Take a best
3476 * guess at determining it: try 64-bit swap first (ie., file
3477 * created on a 64-bit host), and check if the hostname feature
3478 * bit is set (this feature bit is forced on as of fbe96f2).
3479 * If the bit is not, undo the 64-bit swap and try a 32-bit
3480 * swap. If the hostname bit is still not set (e.g., older data
3481 * file), punt and fallback to the original behavior --
3482 * clearing all feature bits and setting buildid.
3484 mem_bswap_64(&header->adds_features,
3485 BITS_TO_U64(HEADER_FEAT_BITS));
3487 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3489 mem_bswap_64(&header->adds_features,
3490 BITS_TO_U64(HEADER_FEAT_BITS));
3493 mem_bswap_32(&header->adds_features,
3494 BITS_TO_U32(HEADER_FEAT_BITS));
3497 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3498 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3499 set_bit(HEADER_BUILD_ID, header->adds_features);
3503 memcpy(&ph->adds_features, &header->adds_features,
3504 sizeof(ph->adds_features));
3506 ph->data_offset = header->data.offset;
3507 ph->data_size = header->data.size;
3508 ph->feat_offset = header->data.offset + header->data.size;
3512 static int perf_file_section__process(struct perf_file_section *section,
3513 struct perf_header *ph,
3514 int feat, int fd, void *data)
3516 struct feat_fd fdd = {
3519 .size = section->size,
3520 .offset = section->offset,
3523 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3524 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3525 "%d, continuing...\n", section->offset, feat);
3529 if (feat >= HEADER_LAST_FEATURE) {
3530 pr_debug("unknown feature %d, continuing...\n", feat);
3534 if (!feat_ops[feat].process)
3537 return feat_ops[feat].process(&fdd, data);
3540 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
3541 struct perf_header *ph, int fd,
3544 struct feat_fd ff = {
3545 .fd = STDOUT_FILENO,
3550 ret = readn(fd, header, sizeof(*header));
3554 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
3555 pr_debug("endian/magic failed\n");
3560 header->size = bswap_64(header->size);
3562 if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
3568 static int perf_header__read_pipe(struct perf_session *session)
3570 struct perf_header *header = &session->header;
3571 struct perf_pipe_file_header f_header;
3573 if (perf_file_header__read_pipe(&f_header, header,
3574 perf_data__fd(session->data),
3575 session->repipe) < 0) {
3576 pr_debug("incompatible file format\n");
3583 static int read_attr(int fd, struct perf_header *ph,
3584 struct perf_file_attr *f_attr)
3586 struct perf_event_attr *attr = &f_attr->attr;
3588 size_t our_sz = sizeof(f_attr->attr);
3591 memset(f_attr, 0, sizeof(*f_attr));
3593 /* read minimal guaranteed structure */
3594 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
3596 pr_debug("cannot read %d bytes of header attr\n",
3597 PERF_ATTR_SIZE_VER0);
3601 /* on file perf_event_attr size */
3609 sz = PERF_ATTR_SIZE_VER0;
3610 } else if (sz > our_sz) {
3611 pr_debug("file uses a more recent and unsupported ABI"
3612 " (%zu bytes extra)\n", sz - our_sz);
3615 /* what we have not yet read and that we know about */
3616 left = sz - PERF_ATTR_SIZE_VER0;
3619 ptr += PERF_ATTR_SIZE_VER0;
3621 ret = readn(fd, ptr, left);
3623 /* read perf_file_section, ids are read in caller */
3624 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
3626 return ret <= 0 ? -1 : 0;
3629 static int perf_evsel__prepare_tracepoint_event(struct evsel *evsel,
3630 struct tep_handle *pevent)
3632 struct tep_event *event;
3635 /* already prepared */
3636 if (evsel->tp_format)
3639 if (pevent == NULL) {
3640 pr_debug("broken or missing trace data\n");
3644 event = tep_find_event(pevent, evsel->core.attr.config);
3645 if (event == NULL) {
3646 pr_debug("cannot find event format for %d\n", (int)evsel->core.attr.config);
3651 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
3652 evsel->name = strdup(bf);
3653 if (evsel->name == NULL)
3657 evsel->tp_format = event;
3661 static int perf_evlist__prepare_tracepoint_events(struct evlist *evlist,
3662 struct tep_handle *pevent)
3666 evlist__for_each_entry(evlist, pos) {
3667 if (pos->core.attr.type == PERF_TYPE_TRACEPOINT &&
3668 perf_evsel__prepare_tracepoint_event(pos, pevent))
3675 int perf_session__read_header(struct perf_session *session)
3677 struct perf_data *data = session->data;
3678 struct perf_header *header = &session->header;
3679 struct perf_file_header f_header;
3680 struct perf_file_attr f_attr;
3682 int nr_attrs, nr_ids, i, j;
3683 int fd = perf_data__fd(data);
3685 session->evlist = evlist__new();
3686 if (session->evlist == NULL)
3689 session->evlist->env = &header->env;
3690 session->machines.host.env = &header->env;
3691 if (perf_data__is_pipe(data))
3692 return perf_header__read_pipe(session);
3694 if (perf_file_header__read(&f_header, header, fd) < 0)
3698 * Sanity check that perf.data was written cleanly; data size is
3699 * initialized to 0 and updated only if the on_exit function is run.
3700 * If data size is still 0 then the file contains only partial
3701 * information. Just warn user and process it as much as it can.
3703 if (f_header.data.size == 0) {
3704 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3705 "Was the 'perf record' command properly terminated?\n",
3709 if (f_header.attr_size == 0) {
3710 pr_err("ERROR: The %s file's attr size field is 0 which is unexpected.\n"
3711 "Was the 'perf record' command properly terminated?\n",
3716 nr_attrs = f_header.attrs.size / f_header.attr_size;
3717 lseek(fd, f_header.attrs.offset, SEEK_SET);
3719 for (i = 0; i < nr_attrs; i++) {
3720 struct evsel *evsel;
3723 if (read_attr(fd, header, &f_attr) < 0)
3726 if (header->needs_swap) {
3727 f_attr.ids.size = bswap_64(f_attr.ids.size);
3728 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
3729 perf_event__attr_swap(&f_attr.attr);
3732 tmp = lseek(fd, 0, SEEK_CUR);
3733 evsel = evsel__new(&f_attr.attr);
3736 goto out_delete_evlist;
3738 evsel->needs_swap = header->needs_swap;
3740 * Do it before so that if perf_evsel__alloc_id fails, this
3741 * entry gets purged too at evlist__delete().
3743 evlist__add(session->evlist, evsel);
3745 nr_ids = f_attr.ids.size / sizeof(u64);
3747 * We don't have the cpu and thread maps on the header, so
3748 * for allocating the perf_sample_id table we fake 1 cpu and
3749 * hattr->ids threads.
3751 if (perf_evsel__alloc_id(&evsel->core, 1, nr_ids))
3752 goto out_delete_evlist;
3754 lseek(fd, f_attr.ids.offset, SEEK_SET);
3756 for (j = 0; j < nr_ids; j++) {
3757 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
3760 perf_evlist__id_add(&session->evlist->core, &evsel->core, 0, j, f_id);
3763 lseek(fd, tmp, SEEK_SET);
3766 perf_header__process_sections(header, fd, &session->tevent,
3767 perf_file_section__process);
3769 if (perf_evlist__prepare_tracepoint_events(session->evlist,
3770 session->tevent.pevent))
3771 goto out_delete_evlist;
3778 evlist__delete(session->evlist);
3779 session->evlist = NULL;
3783 int perf_event__process_feature(struct perf_session *session,
3784 union perf_event *event)
3786 struct perf_tool *tool = session->tool;
3787 struct feat_fd ff = { .fd = 0 };
3788 struct perf_record_header_feature *fe = (struct perf_record_header_feature *)event;
3789 int type = fe->header.type;
3790 u64 feat = fe->feat_id;
3792 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3793 pr_warning("invalid record type %d in pipe-mode\n", type);
3796 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) {
3797 pr_warning("invalid record type %d in pipe-mode\n", type);
3801 if (!feat_ops[feat].process)
3804 ff.buf = (void *)fe->data;
3805 ff.size = event->header.size - sizeof(*fe);
3806 ff.ph = &session->header;
3808 if (feat_ops[feat].process(&ff, NULL))
3811 if (!feat_ops[feat].print || !tool->show_feat_hdr)
3814 if (!feat_ops[feat].full_only ||
3815 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3816 feat_ops[feat].print(&ff, stdout);
3818 fprintf(stdout, "# %s info available, use -I to display\n",
3819 feat_ops[feat].name);
3825 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3827 struct perf_record_event_update *ev = &event->event_update;
3828 struct perf_record_event_update_scale *ev_scale;
3829 struct perf_record_event_update_cpus *ev_cpus;
3830 struct perf_cpu_map *map;
3833 ret = fprintf(fp, "\n... id: %" PRI_lu64 "\n", ev->id);
3836 case PERF_EVENT_UPDATE__SCALE:
3837 ev_scale = (struct perf_record_event_update_scale *)ev->data;
3838 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3840 case PERF_EVENT_UPDATE__UNIT:
3841 ret += fprintf(fp, "... unit: %s\n", ev->data);
3843 case PERF_EVENT_UPDATE__NAME:
3844 ret += fprintf(fp, "... name: %s\n", ev->data);
3846 case PERF_EVENT_UPDATE__CPUS:
3847 ev_cpus = (struct perf_record_event_update_cpus *)ev->data;
3848 ret += fprintf(fp, "... ");
3850 map = cpu_map__new_data(&ev_cpus->cpus);
3852 ret += cpu_map__fprintf(map, fp);
3854 ret += fprintf(fp, "failed to get cpus\n");
3857 ret += fprintf(fp, "... unknown type\n");
3864 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3865 union perf_event *event,
3866 struct evlist **pevlist)
3869 struct evsel *evsel;
3870 struct evlist *evlist = *pevlist;
3872 if (evlist == NULL) {
3873 *pevlist = evlist = evlist__new();
3878 evsel = evsel__new(&event->attr.attr);
3882 evlist__add(evlist, evsel);
3884 ids = event->header.size;
3885 ids -= (void *)&event->attr.id - (void *)event;
3886 n_ids = ids / sizeof(u64);
3888 * We don't have the cpu and thread maps on the header, so
3889 * for allocating the perf_sample_id table we fake 1 cpu and
3890 * hattr->ids threads.
3892 if (perf_evsel__alloc_id(&evsel->core, 1, n_ids))
3895 for (i = 0; i < n_ids; i++) {
3896 perf_evlist__id_add(&evlist->core, &evsel->core, 0, i, event->attr.id[i]);
3902 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3903 union perf_event *event,
3904 struct evlist **pevlist)
3906 struct perf_record_event_update *ev = &event->event_update;
3907 struct perf_record_event_update_scale *ev_scale;
3908 struct perf_record_event_update_cpus *ev_cpus;
3909 struct evlist *evlist;
3910 struct evsel *evsel;
3911 struct perf_cpu_map *map;
3913 if (!pevlist || *pevlist == NULL)
3918 evsel = perf_evlist__id2evsel(evlist, ev->id);
3923 case PERF_EVENT_UPDATE__UNIT:
3924 evsel->unit = strdup(ev->data);
3926 case PERF_EVENT_UPDATE__NAME:
3927 evsel->name = strdup(ev->data);
3929 case PERF_EVENT_UPDATE__SCALE:
3930 ev_scale = (struct perf_record_event_update_scale *)ev->data;
3931 evsel->scale = ev_scale->scale;
3933 case PERF_EVENT_UPDATE__CPUS:
3934 ev_cpus = (struct perf_record_event_update_cpus *)ev->data;
3936 map = cpu_map__new_data(&ev_cpus->cpus);
3938 evsel->core.own_cpus = map;
3940 pr_err("failed to get event_update cpus\n");
3948 int perf_event__process_tracing_data(struct perf_session *session,
3949 union perf_event *event)
3951 ssize_t size_read, padding, size = event->tracing_data.size;
3952 int fd = perf_data__fd(session->data);
3953 off_t offset = lseek(fd, 0, SEEK_CUR);
3956 /* setup for reading amidst mmap */
3957 lseek(fd, offset + sizeof(struct perf_record_header_tracing_data),
3960 size_read = trace_report(fd, &session->tevent,
3962 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3964 if (readn(fd, buf, padding) < 0) {
3965 pr_err("%s: reading input file", __func__);
3968 if (session->repipe) {
3969 int retw = write(STDOUT_FILENO, buf, padding);
3970 if (retw <= 0 || retw != padding) {
3971 pr_err("%s: repiping tracing data padding", __func__);
3976 if (size_read + padding != size) {
3977 pr_err("%s: tracing data size mismatch", __func__);
3981 perf_evlist__prepare_tracepoint_events(session->evlist,
3982 session->tevent.pevent);
3984 return size_read + padding;
3987 int perf_event__process_build_id(struct perf_session *session,
3988 union perf_event *event)
3990 __event_process_build_id(&event->build_id,
3991 event->build_id.filename,