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"
51 #include <linux/ctype.h>
52 #include <internal/lib.h>
56 * must be a numerical value to let the endianness
57 * determine the memory layout. That way we are able
58 * to detect endianness when reading the perf.data file
61 * we check for legacy (PERFFILE) format.
63 static const char *__perf_magic1 = "PERFFILE";
64 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
65 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
67 #define PERF_MAGIC __perf_magic2
69 const char perf_version_string[] = PERF_VERSION;
71 struct perf_file_attr {
72 struct perf_event_attr attr;
73 struct perf_file_section ids;
76 void perf_header__set_feat(struct perf_header *header, int feat)
78 set_bit(feat, header->adds_features);
81 void perf_header__clear_feat(struct perf_header *header, int feat)
83 clear_bit(feat, header->adds_features);
86 bool perf_header__has_feat(const struct perf_header *header, int feat)
88 return test_bit(feat, header->adds_features);
91 static int __do_write_fd(struct feat_fd *ff, const void *buf, size_t size)
93 ssize_t ret = writen(ff->fd, buf, size);
95 if (ret != (ssize_t)size)
96 return ret < 0 ? (int)ret : -1;
100 static int __do_write_buf(struct feat_fd *ff, const void *buf, size_t size)
102 /* struct perf_event_header::size is u16 */
103 const size_t max_size = 0xffff - sizeof(struct perf_event_header);
104 size_t new_size = ff->size;
107 if (size + ff->offset > max_size)
110 while (size > (new_size - ff->offset))
112 new_size = min(max_size, new_size);
114 if (ff->size < new_size) {
115 addr = realloc(ff->buf, new_size);
122 memcpy(ff->buf + ff->offset, buf, size);
128 /* Return: 0 if succeded, -ERR if failed. */
129 int do_write(struct feat_fd *ff, const void *buf, size_t size)
132 return __do_write_fd(ff, buf, size);
133 return __do_write_buf(ff, buf, size);
136 /* Return: 0 if succeded, -ERR if failed. */
137 static int do_write_bitmap(struct feat_fd *ff, unsigned long *set, u64 size)
139 u64 *p = (u64 *) set;
142 ret = do_write(ff, &size, sizeof(size));
146 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
147 ret = do_write(ff, p + i, sizeof(*p));
155 /* Return: 0 if succeded, -ERR if failed. */
156 int write_padded(struct feat_fd *ff, const void *bf,
157 size_t count, size_t count_aligned)
159 static const char zero_buf[NAME_ALIGN];
160 int err = do_write(ff, bf, count);
163 err = do_write(ff, zero_buf, count_aligned - count);
168 #define string_size(str) \
169 (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
171 /* Return: 0 if succeded, -ERR if failed. */
172 static int do_write_string(struct feat_fd *ff, const char *str)
177 olen = strlen(str) + 1;
178 len = PERF_ALIGN(olen, NAME_ALIGN);
180 /* write len, incl. \0 */
181 ret = do_write(ff, &len, sizeof(len));
185 return write_padded(ff, str, olen, len);
188 static int __do_read_fd(struct feat_fd *ff, void *addr, ssize_t size)
190 ssize_t ret = readn(ff->fd, addr, size);
193 return ret < 0 ? (int)ret : -1;
197 static int __do_read_buf(struct feat_fd *ff, void *addr, ssize_t size)
199 if (size > (ssize_t)ff->size - ff->offset)
202 memcpy(addr, ff->buf + ff->offset, size);
209 static int __do_read(struct feat_fd *ff, void *addr, ssize_t size)
212 return __do_read_fd(ff, addr, size);
213 return __do_read_buf(ff, addr, size);
216 static int do_read_u32(struct feat_fd *ff, u32 *addr)
220 ret = __do_read(ff, addr, sizeof(*addr));
224 if (ff->ph->needs_swap)
225 *addr = bswap_32(*addr);
229 static int do_read_u64(struct feat_fd *ff, u64 *addr)
233 ret = __do_read(ff, addr, sizeof(*addr));
237 if (ff->ph->needs_swap)
238 *addr = bswap_64(*addr);
242 static char *do_read_string(struct feat_fd *ff)
247 if (do_read_u32(ff, &len))
254 if (!__do_read(ff, buf, len)) {
256 * strings are padded by zeroes
257 * thus the actual strlen of buf
258 * may be less than len
267 /* Return: 0 if succeded, -ERR if failed. */
268 static int do_read_bitmap(struct feat_fd *ff, unsigned long **pset, u64 *psize)
274 ret = do_read_u64(ff, &size);
278 set = bitmap_alloc(size);
284 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
285 ret = do_read_u64(ff, p + i);
297 static int write_tracing_data(struct feat_fd *ff,
298 struct evlist *evlist)
300 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
303 return read_tracing_data(ff->fd, &evlist->core.entries);
306 static int write_build_id(struct feat_fd *ff,
307 struct evlist *evlist __maybe_unused)
309 struct perf_session *session;
312 session = container_of(ff->ph, struct perf_session, header);
314 if (!perf_session__read_build_ids(session, true))
317 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
320 err = perf_session__write_buildid_table(session, ff);
322 pr_debug("failed to write buildid table\n");
325 perf_session__cache_build_ids(session);
330 static int write_hostname(struct feat_fd *ff,
331 struct evlist *evlist __maybe_unused)
340 return do_write_string(ff, uts.nodename);
343 static int write_osrelease(struct feat_fd *ff,
344 struct evlist *evlist __maybe_unused)
353 return do_write_string(ff, uts.release);
356 static int write_arch(struct feat_fd *ff,
357 struct evlist *evlist __maybe_unused)
366 return do_write_string(ff, uts.machine);
369 static int write_version(struct feat_fd *ff,
370 struct evlist *evlist __maybe_unused)
372 return do_write_string(ff, perf_version_string);
375 static int __write_cpudesc(struct feat_fd *ff, const char *cpuinfo_proc)
380 const char *search = cpuinfo_proc;
387 file = fopen("/proc/cpuinfo", "r");
391 while (getline(&buf, &len, file) > 0) {
392 ret = strncmp(buf, search, strlen(search));
404 p = strchr(buf, ':');
405 if (p && *(p+1) == ' ' && *(p+2))
411 /* squash extra space characters (branding string) */
416 char *q = skip_spaces(r);
419 while ((*r++ = *q++));
423 ret = do_write_string(ff, s);
430 static int write_cpudesc(struct feat_fd *ff,
431 struct evlist *evlist __maybe_unused)
433 #if defined(__powerpc__) || defined(__hppa__) || defined(__sparc__)
434 #define CPUINFO_PROC { "cpu", }
435 #elif defined(__s390__)
436 #define CPUINFO_PROC { "vendor_id", }
437 #elif defined(__sh__)
438 #define CPUINFO_PROC { "cpu type", }
439 #elif defined(__alpha__) || defined(__mips__)
440 #define CPUINFO_PROC { "cpu model", }
441 #elif defined(__arm__)
442 #define CPUINFO_PROC { "model name", "Processor", }
443 #elif defined(__arc__)
444 #define CPUINFO_PROC { "Processor", }
445 #elif defined(__xtensa__)
446 #define CPUINFO_PROC { "core ID", }
448 #define CPUINFO_PROC { "model name", }
450 const char *cpuinfo_procs[] = CPUINFO_PROC;
454 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
456 ret = __write_cpudesc(ff, cpuinfo_procs[i]);
464 static int write_nrcpus(struct feat_fd *ff,
465 struct evlist *evlist __maybe_unused)
471 nrc = cpu__max_present_cpu();
473 nr = sysconf(_SC_NPROCESSORS_ONLN);
477 nra = (u32)(nr & UINT_MAX);
479 ret = do_write(ff, &nrc, sizeof(nrc));
483 return do_write(ff, &nra, sizeof(nra));
486 static int write_event_desc(struct feat_fd *ff,
487 struct evlist *evlist)
493 nre = evlist->core.nr_entries;
496 * write number of events
498 ret = do_write(ff, &nre, sizeof(nre));
503 * size of perf_event_attr struct
505 sz = (u32)sizeof(evsel->core.attr);
506 ret = do_write(ff, &sz, sizeof(sz));
510 evlist__for_each_entry(evlist, evsel) {
511 ret = do_write(ff, &evsel->core.attr, sz);
515 * write number of unique id per event
516 * there is one id per instance of an event
518 * copy into an nri to be independent of the
521 nri = evsel->core.ids;
522 ret = do_write(ff, &nri, sizeof(nri));
527 * write event string as passed on cmdline
529 ret = do_write_string(ff, evsel__name(evsel));
533 * write unique ids for this event
535 ret = do_write(ff, evsel->core.id, evsel->core.ids * sizeof(u64));
542 static int write_cmdline(struct feat_fd *ff,
543 struct evlist *evlist __maybe_unused)
545 char pbuf[MAXPATHLEN], *buf;
548 /* actual path to perf binary */
549 buf = perf_exe(pbuf, MAXPATHLEN);
551 /* account for binary path */
552 n = perf_env.nr_cmdline + 1;
554 ret = do_write(ff, &n, sizeof(n));
558 ret = do_write_string(ff, buf);
562 for (i = 0 ; i < perf_env.nr_cmdline; i++) {
563 ret = do_write_string(ff, perf_env.cmdline_argv[i]);
571 static int write_cpu_topology(struct feat_fd *ff,
572 struct evlist *evlist __maybe_unused)
574 struct cpu_topology *tp;
578 tp = cpu_topology__new();
582 ret = do_write(ff, &tp->core_sib, sizeof(tp->core_sib));
586 for (i = 0; i < tp->core_sib; i++) {
587 ret = do_write_string(ff, tp->core_siblings[i]);
591 ret = do_write(ff, &tp->thread_sib, sizeof(tp->thread_sib));
595 for (i = 0; i < tp->thread_sib; i++) {
596 ret = do_write_string(ff, tp->thread_siblings[i]);
601 ret = perf_env__read_cpu_topology_map(&perf_env);
605 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
606 ret = do_write(ff, &perf_env.cpu[j].core_id,
607 sizeof(perf_env.cpu[j].core_id));
610 ret = do_write(ff, &perf_env.cpu[j].socket_id,
611 sizeof(perf_env.cpu[j].socket_id));
619 ret = do_write(ff, &tp->die_sib, sizeof(tp->die_sib));
623 for (i = 0; i < tp->die_sib; i++) {
624 ret = do_write_string(ff, tp->die_siblings[i]);
629 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
630 ret = do_write(ff, &perf_env.cpu[j].die_id,
631 sizeof(perf_env.cpu[j].die_id));
637 cpu_topology__delete(tp);
643 static int write_total_mem(struct feat_fd *ff,
644 struct evlist *evlist __maybe_unused)
652 fp = fopen("/proc/meminfo", "r");
656 while (getline(&buf, &len, fp) > 0) {
657 ret = strncmp(buf, "MemTotal:", 9);
662 n = sscanf(buf, "%*s %"PRIu64, &mem);
664 ret = do_write(ff, &mem, sizeof(mem));
672 static int write_numa_topology(struct feat_fd *ff,
673 struct evlist *evlist __maybe_unused)
675 struct numa_topology *tp;
679 tp = numa_topology__new();
683 ret = do_write(ff, &tp->nr, sizeof(u32));
687 for (i = 0; i < tp->nr; i++) {
688 struct numa_topology_node *n = &tp->nodes[i];
690 ret = do_write(ff, &n->node, sizeof(u32));
694 ret = do_write(ff, &n->mem_total, sizeof(u64));
698 ret = do_write(ff, &n->mem_free, sizeof(u64));
702 ret = do_write_string(ff, n->cpus);
710 numa_topology__delete(tp);
717 * struct pmu_mappings {
726 static int write_pmu_mappings(struct feat_fd *ff,
727 struct evlist *evlist __maybe_unused)
729 struct perf_pmu *pmu = NULL;
734 * Do a first pass to count number of pmu to avoid lseek so this
735 * works in pipe mode as well.
737 while ((pmu = perf_pmu__scan(pmu))) {
743 ret = do_write(ff, &pmu_num, sizeof(pmu_num));
747 while ((pmu = perf_pmu__scan(pmu))) {
751 ret = do_write(ff, &pmu->type, sizeof(pmu->type));
755 ret = do_write_string(ff, pmu->name);
766 * struct group_descs {
768 * struct group_desc {
775 static int write_group_desc(struct feat_fd *ff,
776 struct evlist *evlist)
778 u32 nr_groups = evlist->nr_groups;
782 ret = do_write(ff, &nr_groups, sizeof(nr_groups));
786 evlist__for_each_entry(evlist, evsel) {
787 if (evsel__is_group_leader(evsel) && 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.clock.clockid_res_ns,
896 sizeof(ff->ph->env.clock.clockid_res_ns));
899 static int write_clock_data(struct feat_fd *ff,
900 struct evlist *evlist __maybe_unused)
909 ret = do_write(ff, &data32, sizeof(data32));
914 data32 = ff->ph->env.clock.clockid;
916 ret = do_write(ff, &data32, sizeof(data32));
921 data64 = &ff->ph->env.clock.tod_ns;
923 ret = do_write(ff, data64, sizeof(*data64));
927 /* clockid ref time */
928 data64 = &ff->ph->env.clock.clockid_ns;
930 return do_write(ff, data64, sizeof(*data64));
933 static int write_dir_format(struct feat_fd *ff,
934 struct evlist *evlist __maybe_unused)
936 struct perf_session *session;
937 struct perf_data *data;
939 session = container_of(ff->ph, struct perf_session, header);
940 data = session->data;
942 if (WARN_ON(!perf_data__is_dir(data)))
945 return do_write(ff, &data->dir.version, sizeof(data->dir.version));
948 #ifdef HAVE_LIBBPF_SUPPORT
949 static int write_bpf_prog_info(struct feat_fd *ff,
950 struct evlist *evlist __maybe_unused)
952 struct perf_env *env = &ff->ph->env;
953 struct rb_root *root;
954 struct rb_node *next;
957 down_read(&env->bpf_progs.lock);
959 ret = do_write(ff, &env->bpf_progs.infos_cnt,
960 sizeof(env->bpf_progs.infos_cnt));
964 root = &env->bpf_progs.infos;
965 next = rb_first(root);
967 struct bpf_prog_info_node *node;
970 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
971 next = rb_next(&node->rb_node);
972 len = sizeof(struct bpf_prog_info_linear) +
973 node->info_linear->data_len;
975 /* before writing to file, translate address to offset */
976 bpf_program__bpil_addr_to_offs(node->info_linear);
977 ret = do_write(ff, node->info_linear, len);
979 * translate back to address even when do_write() fails,
980 * so that this function never changes the data.
982 bpf_program__bpil_offs_to_addr(node->info_linear);
987 up_read(&env->bpf_progs.lock);
990 #else // HAVE_LIBBPF_SUPPORT
991 static int write_bpf_prog_info(struct feat_fd *ff __maybe_unused,
992 struct evlist *evlist __maybe_unused)
996 #endif // HAVE_LIBBPF_SUPPORT
998 static int write_bpf_btf(struct feat_fd *ff,
999 struct evlist *evlist __maybe_unused)
1001 struct perf_env *env = &ff->ph->env;
1002 struct rb_root *root;
1003 struct rb_node *next;
1006 down_read(&env->bpf_progs.lock);
1008 ret = do_write(ff, &env->bpf_progs.btfs_cnt,
1009 sizeof(env->bpf_progs.btfs_cnt));
1014 root = &env->bpf_progs.btfs;
1015 next = rb_first(root);
1017 struct btf_node *node;
1019 node = rb_entry(next, struct btf_node, rb_node);
1020 next = rb_next(&node->rb_node);
1021 ret = do_write(ff, &node->id,
1022 sizeof(u32) * 2 + node->data_size);
1027 up_read(&env->bpf_progs.lock);
1031 static int cpu_cache_level__sort(const void *a, const void *b)
1033 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
1034 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
1036 return cache_a->level - cache_b->level;
1039 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
1041 if (a->level != b->level)
1044 if (a->line_size != b->line_size)
1047 if (a->sets != b->sets)
1050 if (a->ways != b->ways)
1053 if (strcmp(a->type, b->type))
1056 if (strcmp(a->size, b->size))
1059 if (strcmp(a->map, b->map))
1065 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
1067 char path[PATH_MAX], file[PATH_MAX];
1071 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
1072 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
1074 if (stat(file, &st))
1077 scnprintf(file, PATH_MAX, "%s/level", path);
1078 if (sysfs__read_int(file, (int *) &cache->level))
1081 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
1082 if (sysfs__read_int(file, (int *) &cache->line_size))
1085 scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
1086 if (sysfs__read_int(file, (int *) &cache->sets))
1089 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
1090 if (sysfs__read_int(file, (int *) &cache->ways))
1093 scnprintf(file, PATH_MAX, "%s/type", path);
1094 if (sysfs__read_str(file, &cache->type, &len))
1097 cache->type[len] = 0;
1098 cache->type = strim(cache->type);
1100 scnprintf(file, PATH_MAX, "%s/size", path);
1101 if (sysfs__read_str(file, &cache->size, &len)) {
1102 zfree(&cache->type);
1106 cache->size[len] = 0;
1107 cache->size = strim(cache->size);
1109 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
1110 if (sysfs__read_str(file, &cache->map, &len)) {
1111 zfree(&cache->size);
1112 zfree(&cache->type);
1116 cache->map[len] = 0;
1117 cache->map = strim(cache->map);
1121 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
1123 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
1126 #define MAX_CACHE_LVL 4
1128 static int build_caches(struct cpu_cache_level caches[], u32 *cntp)
1134 nr = cpu__max_cpu();
1136 for (cpu = 0; cpu < nr; cpu++) {
1137 for (level = 0; level < MAX_CACHE_LVL; level++) {
1138 struct cpu_cache_level c;
1141 err = cpu_cache_level__read(&c, cpu, level);
1148 for (i = 0; i < cnt; i++) {
1149 if (cpu_cache_level__cmp(&c, &caches[i]))
1156 cpu_cache_level__free(&c);
1163 static int write_cache(struct feat_fd *ff,
1164 struct evlist *evlist __maybe_unused)
1166 u32 max_caches = cpu__max_cpu() * MAX_CACHE_LVL;
1167 struct cpu_cache_level caches[max_caches];
1168 u32 cnt = 0, i, version = 1;
1171 ret = build_caches(caches, &cnt);
1175 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1177 ret = do_write(ff, &version, sizeof(u32));
1181 ret = do_write(ff, &cnt, sizeof(u32));
1185 for (i = 0; i < cnt; i++) {
1186 struct cpu_cache_level *c = &caches[i];
1189 ret = do_write(ff, &c->v, sizeof(u32)); \
1200 ret = do_write_string(ff, (const char *) c->v); \
1211 for (i = 0; i < cnt; i++)
1212 cpu_cache_level__free(&caches[i]);
1216 static int write_stat(struct feat_fd *ff __maybe_unused,
1217 struct evlist *evlist __maybe_unused)
1222 static int write_sample_time(struct feat_fd *ff,
1223 struct evlist *evlist)
1227 ret = do_write(ff, &evlist->first_sample_time,
1228 sizeof(evlist->first_sample_time));
1232 return do_write(ff, &evlist->last_sample_time,
1233 sizeof(evlist->last_sample_time));
1237 static int memory_node__read(struct memory_node *n, unsigned long idx)
1239 unsigned int phys, size = 0;
1240 char path[PATH_MAX];
1244 #define for_each_memory(mem, dir) \
1245 while ((ent = readdir(dir))) \
1246 if (strcmp(ent->d_name, ".") && \
1247 strcmp(ent->d_name, "..") && \
1248 sscanf(ent->d_name, "memory%u", &mem) == 1)
1250 scnprintf(path, PATH_MAX,
1251 "%s/devices/system/node/node%lu",
1252 sysfs__mountpoint(), idx);
1254 dir = opendir(path);
1256 pr_warning("failed: cant' open memory sysfs data\n");
1260 for_each_memory(phys, dir) {
1261 size = max(phys, size);
1266 n->set = bitmap_alloc(size);
1277 for_each_memory(phys, dir) {
1278 set_bit(phys, n->set);
1285 static int memory_node__sort(const void *a, const void *b)
1287 const struct memory_node *na = a;
1288 const struct memory_node *nb = b;
1290 return na->node - nb->node;
1293 static int build_mem_topology(struct memory_node *nodes, u64 size, u64 *cntp)
1295 char path[PATH_MAX];
1301 scnprintf(path, PATH_MAX, "%s/devices/system/node/",
1302 sysfs__mountpoint());
1304 dir = opendir(path);
1306 pr_debug2("%s: could't read %s, does this arch have topology information?\n",
1311 while (!ret && (ent = readdir(dir))) {
1315 if (!strcmp(ent->d_name, ".") ||
1316 !strcmp(ent->d_name, ".."))
1319 r = sscanf(ent->d_name, "node%u", &idx);
1323 if (WARN_ONCE(cnt >= size,
1324 "failed to write MEM_TOPOLOGY, way too many nodes\n")) {
1329 ret = memory_node__read(&nodes[cnt++], idx);
1336 qsort(nodes, cnt, sizeof(nodes[0]), memory_node__sort);
1341 #define MAX_MEMORY_NODES 2000
1344 * The MEM_TOPOLOGY holds physical memory map for every
1345 * node in system. The format of data is as follows:
1347 * 0 - version | for future changes
1348 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes
1349 * 16 - count | number of nodes
1351 * For each node we store map of physical indexes for
1354 * 32 - node id | node index
1355 * 40 - size | size of bitmap
1356 * 48 - bitmap | bitmap of memory indexes that belongs to node
1358 static int write_mem_topology(struct feat_fd *ff __maybe_unused,
1359 struct evlist *evlist __maybe_unused)
1361 static struct memory_node nodes[MAX_MEMORY_NODES];
1362 u64 bsize, version = 1, i, nr;
1365 ret = sysfs__read_xll("devices/system/memory/block_size_bytes",
1366 (unsigned long long *) &bsize);
1370 ret = build_mem_topology(&nodes[0], MAX_MEMORY_NODES, &nr);
1374 ret = do_write(ff, &version, sizeof(version));
1378 ret = do_write(ff, &bsize, sizeof(bsize));
1382 ret = do_write(ff, &nr, sizeof(nr));
1386 for (i = 0; i < nr; i++) {
1387 struct memory_node *n = &nodes[i];
1390 ret = do_write(ff, &n->v, sizeof(n->v)); \
1399 ret = do_write_bitmap(ff, n->set, n->size);
1408 static int write_compressed(struct feat_fd *ff __maybe_unused,
1409 struct evlist *evlist __maybe_unused)
1413 ret = do_write(ff, &(ff->ph->env.comp_ver), sizeof(ff->ph->env.comp_ver));
1417 ret = do_write(ff, &(ff->ph->env.comp_type), sizeof(ff->ph->env.comp_type));
1421 ret = do_write(ff, &(ff->ph->env.comp_level), sizeof(ff->ph->env.comp_level));
1425 ret = do_write(ff, &(ff->ph->env.comp_ratio), sizeof(ff->ph->env.comp_ratio));
1429 return do_write(ff, &(ff->ph->env.comp_mmap_len), sizeof(ff->ph->env.comp_mmap_len));
1432 static int write_cpu_pmu_caps(struct feat_fd *ff,
1433 struct evlist *evlist __maybe_unused)
1435 struct perf_pmu *cpu_pmu = perf_pmu__find("cpu");
1436 struct perf_pmu_caps *caps = NULL;
1443 nr_caps = perf_pmu__caps_parse(cpu_pmu);
1447 ret = do_write(ff, &nr_caps, sizeof(nr_caps));
1451 list_for_each_entry(caps, &cpu_pmu->caps, list) {
1452 ret = do_write_string(ff, caps->name);
1456 ret = do_write_string(ff, caps->value);
1464 static void print_hostname(struct feat_fd *ff, FILE *fp)
1466 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1469 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1471 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1474 static void print_arch(struct feat_fd *ff, FILE *fp)
1476 fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1479 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1481 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1484 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1486 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1487 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1490 static void print_version(struct feat_fd *ff, FILE *fp)
1492 fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1495 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1499 nr = ff->ph->env.nr_cmdline;
1501 fprintf(fp, "# cmdline : ");
1503 for (i = 0; i < nr; i++) {
1504 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]);
1506 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1510 char *quote = strchr(argv_i, '\'');
1514 fprintf(fp, "%s\\\'", argv_i);
1517 fprintf(fp, "%s ", argv_i);
1524 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1526 struct perf_header *ph = ff->ph;
1527 int cpu_nr = ph->env.nr_cpus_avail;
1531 nr = ph->env.nr_sibling_cores;
1532 str = ph->env.sibling_cores;
1534 for (i = 0; i < nr; i++) {
1535 fprintf(fp, "# sibling sockets : %s\n", str);
1536 str += strlen(str) + 1;
1539 if (ph->env.nr_sibling_dies) {
1540 nr = ph->env.nr_sibling_dies;
1541 str = ph->env.sibling_dies;
1543 for (i = 0; i < nr; i++) {
1544 fprintf(fp, "# sibling dies : %s\n", str);
1545 str += strlen(str) + 1;
1549 nr = ph->env.nr_sibling_threads;
1550 str = ph->env.sibling_threads;
1552 for (i = 0; i < nr; i++) {
1553 fprintf(fp, "# sibling threads : %s\n", str);
1554 str += strlen(str) + 1;
1557 if (ph->env.nr_sibling_dies) {
1558 if (ph->env.cpu != NULL) {
1559 for (i = 0; i < cpu_nr; i++)
1560 fprintf(fp, "# CPU %d: Core ID %d, "
1561 "Die ID %d, Socket ID %d\n",
1562 i, ph->env.cpu[i].core_id,
1563 ph->env.cpu[i].die_id,
1564 ph->env.cpu[i].socket_id);
1566 fprintf(fp, "# Core ID, Die ID and Socket ID "
1567 "information is not available\n");
1569 if (ph->env.cpu != NULL) {
1570 for (i = 0; i < cpu_nr; i++)
1571 fprintf(fp, "# CPU %d: Core ID %d, "
1573 i, ph->env.cpu[i].core_id,
1574 ph->env.cpu[i].socket_id);
1576 fprintf(fp, "# Core ID and Socket ID "
1577 "information is not available\n");
1581 static void print_clockid(struct feat_fd *ff, FILE *fp)
1583 fprintf(fp, "# clockid frequency: %"PRIu64" MHz\n",
1584 ff->ph->env.clock.clockid_res_ns * 1000);
1587 static void print_clock_data(struct feat_fd *ff, FILE *fp)
1589 struct timespec clockid_ns;
1590 char tstr[64], date[64];
1591 struct timeval tod_ns;
1596 if (!ff->ph->env.clock.enabled) {
1597 fprintf(fp, "# reference time disabled\n");
1601 /* Compute TOD time. */
1602 ref = ff->ph->env.clock.tod_ns;
1603 tod_ns.tv_sec = ref / NSEC_PER_SEC;
1604 ref -= tod_ns.tv_sec * NSEC_PER_SEC;
1605 tod_ns.tv_usec = ref / NSEC_PER_USEC;
1607 /* Compute clockid time. */
1608 ref = ff->ph->env.clock.clockid_ns;
1609 clockid_ns.tv_sec = ref / NSEC_PER_SEC;
1610 ref -= clockid_ns.tv_sec * NSEC_PER_SEC;
1611 clockid_ns.tv_nsec = ref;
1613 clockid = ff->ph->env.clock.clockid;
1615 if (localtime_r(&tod_ns.tv_sec, <ime) == NULL)
1616 snprintf(tstr, sizeof(tstr), "<error>");
1618 strftime(date, sizeof(date), "%F %T", <ime);
1619 scnprintf(tstr, sizeof(tstr), "%s.%06d",
1620 date, (int) tod_ns.tv_usec);
1623 fprintf(fp, "# clockid: %s (%u)\n", clockid_name(clockid), clockid);
1624 fprintf(fp, "# reference time: %s = %ld.%06d (TOD) = %ld.%09ld (%s)\n",
1625 tstr, tod_ns.tv_sec, (int) tod_ns.tv_usec,
1626 clockid_ns.tv_sec, clockid_ns.tv_nsec,
1627 clockid_name(clockid));
1630 static void print_dir_format(struct feat_fd *ff, FILE *fp)
1632 struct perf_session *session;
1633 struct perf_data *data;
1635 session = container_of(ff->ph, struct perf_session, header);
1636 data = session->data;
1638 fprintf(fp, "# directory data version : %"PRIu64"\n", data->dir.version);
1641 static void print_bpf_prog_info(struct feat_fd *ff, FILE *fp)
1643 struct perf_env *env = &ff->ph->env;
1644 struct rb_root *root;
1645 struct rb_node *next;
1647 down_read(&env->bpf_progs.lock);
1649 root = &env->bpf_progs.infos;
1650 next = rb_first(root);
1653 struct bpf_prog_info_node *node;
1655 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
1656 next = rb_next(&node->rb_node);
1658 bpf_event__print_bpf_prog_info(&node->info_linear->info,
1662 up_read(&env->bpf_progs.lock);
1665 static void print_bpf_btf(struct feat_fd *ff, FILE *fp)
1667 struct perf_env *env = &ff->ph->env;
1668 struct rb_root *root;
1669 struct rb_node *next;
1671 down_read(&env->bpf_progs.lock);
1673 root = &env->bpf_progs.btfs;
1674 next = rb_first(root);
1677 struct btf_node *node;
1679 node = rb_entry(next, struct btf_node, rb_node);
1680 next = rb_next(&node->rb_node);
1681 fprintf(fp, "# btf info of id %u\n", node->id);
1684 up_read(&env->bpf_progs.lock);
1687 static void free_event_desc(struct evsel *events)
1689 struct evsel *evsel;
1694 for (evsel = events; evsel->core.attr.size; evsel++) {
1695 zfree(&evsel->name);
1696 zfree(&evsel->core.id);
1702 static bool perf_attr_check(struct perf_event_attr *attr)
1704 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3) {
1705 pr_warning("Reserved bits are set unexpectedly. "
1706 "Please update perf tool.\n");
1710 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) {
1711 pr_warning("Unknown sample type (0x%llx) is detected. "
1712 "Please update perf tool.\n",
1717 if (attr->read_format & ~(PERF_FORMAT_MAX-1)) {
1718 pr_warning("Unknown read format (0x%llx) is detected. "
1719 "Please update perf tool.\n",
1724 if ((attr->sample_type & PERF_SAMPLE_BRANCH_STACK) &&
1725 (attr->branch_sample_type & ~(PERF_SAMPLE_BRANCH_MAX-1))) {
1726 pr_warning("Unknown branch sample type (0x%llx) is detected. "
1727 "Please update perf tool.\n",
1728 attr->branch_sample_type);
1736 static struct evsel *read_event_desc(struct feat_fd *ff)
1738 struct evsel *evsel, *events = NULL;
1741 u32 nre, sz, nr, i, j;
1744 /* number of events */
1745 if (do_read_u32(ff, &nre))
1748 if (do_read_u32(ff, &sz))
1751 /* buffer to hold on file attr struct */
1756 /* the last event terminates with evsel->core.attr.size == 0: */
1757 events = calloc(nre + 1, sizeof(*events));
1761 msz = sizeof(evsel->core.attr);
1765 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1769 * must read entire on-file attr struct to
1770 * sync up with layout.
1772 if (__do_read(ff, buf, sz))
1775 if (ff->ph->needs_swap)
1776 perf_event__attr_swap(buf);
1778 memcpy(&evsel->core.attr, buf, msz);
1780 if (!perf_attr_check(&evsel->core.attr))
1783 if (do_read_u32(ff, &nr))
1786 if (ff->ph->needs_swap)
1787 evsel->needs_swap = true;
1789 evsel->name = do_read_string(ff);
1796 id = calloc(nr, sizeof(*id));
1799 evsel->core.ids = nr;
1800 evsel->core.id = id;
1802 for (j = 0 ; j < nr; j++) {
1803 if (do_read_u64(ff, id))
1812 free_event_desc(events);
1817 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1818 void *priv __maybe_unused)
1820 return fprintf(fp, ", %s = %s", name, val);
1823 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1825 struct evsel *evsel, *events;
1830 events = ff->events;
1832 events = read_event_desc(ff);
1835 fprintf(fp, "# event desc: not available or unable to read\n");
1839 for (evsel = events; evsel->core.attr.size; evsel++) {
1840 fprintf(fp, "# event : name = %s, ", evsel->name);
1842 if (evsel->core.ids) {
1843 fprintf(fp, ", id = {");
1844 for (j = 0, id = evsel->core.id; j < evsel->core.ids; j++, id++) {
1847 fprintf(fp, " %"PRIu64, *id);
1852 perf_event_attr__fprintf(fp, &evsel->core.attr, __desc_attr__fprintf, NULL);
1857 free_event_desc(events);
1861 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1863 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1866 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1869 struct numa_node *n;
1871 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1872 n = &ff->ph->env.numa_nodes[i];
1874 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1875 " free = %"PRIu64" kB\n",
1876 n->node, n->mem_total, n->mem_free);
1878 fprintf(fp, "# node%u cpu list : ", n->node);
1879 cpu_map__fprintf(n->map, fp);
1883 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1885 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1888 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1890 fprintf(fp, "# contains samples with branch stack\n");
1893 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1895 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1898 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1900 fprintf(fp, "# contains stat data\n");
1903 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1907 fprintf(fp, "# CPU cache info:\n");
1908 for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1910 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1914 static void print_compressed(struct feat_fd *ff, FILE *fp)
1916 fprintf(fp, "# compressed : %s, level = %d, ratio = %d\n",
1917 ff->ph->env.comp_type == PERF_COMP_ZSTD ? "Zstd" : "Unknown",
1918 ff->ph->env.comp_level, ff->ph->env.comp_ratio);
1921 static void print_cpu_pmu_caps(struct feat_fd *ff, FILE *fp)
1923 const char *delimiter = "# cpu pmu capabilities: ";
1924 u32 nr_caps = ff->ph->env.nr_cpu_pmu_caps;
1928 fprintf(fp, "# cpu pmu capabilities: not available\n");
1932 str = ff->ph->env.cpu_pmu_caps;
1934 fprintf(fp, "%s%s", delimiter, str);
1936 str += strlen(str) + 1;
1942 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1944 const char *delimiter = "# pmu mappings: ";
1949 pmu_num = ff->ph->env.nr_pmu_mappings;
1951 fprintf(fp, "# pmu mappings: not available\n");
1955 str = ff->ph->env.pmu_mappings;
1958 type = strtoul(str, &tmp, 0);
1963 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1966 str += strlen(str) + 1;
1975 fprintf(fp, "# pmu mappings: unable to read\n");
1978 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1980 struct perf_session *session;
1981 struct evsel *evsel;
1984 session = container_of(ff->ph, struct perf_session, header);
1986 evlist__for_each_entry(session->evlist, evsel) {
1987 if (evsel__is_group_leader(evsel) && evsel->core.nr_members > 1) {
1988 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "", evsel__name(evsel));
1990 nr = evsel->core.nr_members - 1;
1992 fprintf(fp, ",%s", evsel__name(evsel));
2000 static void print_sample_time(struct feat_fd *ff, FILE *fp)
2002 struct perf_session *session;
2006 session = container_of(ff->ph, struct perf_session, header);
2008 timestamp__scnprintf_usec(session->evlist->first_sample_time,
2009 time_buf, sizeof(time_buf));
2010 fprintf(fp, "# time of first sample : %s\n", time_buf);
2012 timestamp__scnprintf_usec(session->evlist->last_sample_time,
2013 time_buf, sizeof(time_buf));
2014 fprintf(fp, "# time of last sample : %s\n", time_buf);
2016 d = (double)(session->evlist->last_sample_time -
2017 session->evlist->first_sample_time) / NSEC_PER_MSEC;
2019 fprintf(fp, "# sample duration : %10.3f ms\n", d);
2022 static void memory_node__fprintf(struct memory_node *n,
2023 unsigned long long bsize, FILE *fp)
2025 char buf_map[100], buf_size[50];
2026 unsigned long long size;
2028 size = bsize * bitmap_weight(n->set, n->size);
2029 unit_number__scnprintf(buf_size, 50, size);
2031 bitmap_scnprintf(n->set, n->size, buf_map, 100);
2032 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map);
2035 static void print_mem_topology(struct feat_fd *ff, FILE *fp)
2037 struct memory_node *nodes;
2040 nodes = ff->ph->env.memory_nodes;
2041 nr = ff->ph->env.nr_memory_nodes;
2043 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n",
2044 nr, ff->ph->env.memory_bsize);
2046 for (i = 0; i < nr; i++) {
2047 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp);
2051 static int __event_process_build_id(struct perf_record_header_build_id *bev,
2053 struct perf_session *session)
2056 struct machine *machine;
2059 enum dso_space_type dso_space;
2061 machine = perf_session__findnew_machine(session, bev->pid);
2065 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2068 case PERF_RECORD_MISC_KERNEL:
2069 dso_space = DSO_SPACE__KERNEL;
2071 case PERF_RECORD_MISC_GUEST_KERNEL:
2072 dso_space = DSO_SPACE__KERNEL_GUEST;
2074 case PERF_RECORD_MISC_USER:
2075 case PERF_RECORD_MISC_GUEST_USER:
2076 dso_space = DSO_SPACE__USER;
2082 dso = machine__findnew_dso(machine, filename);
2084 char sbuild_id[SBUILD_ID_SIZE];
2086 dso__set_build_id(dso, &bev->build_id);
2088 if (dso_space != DSO_SPACE__USER) {
2089 struct kmod_path m = { .name = NULL, };
2091 if (!kmod_path__parse_name(&m, filename) && m.kmod)
2092 dso__set_module_info(dso, &m, machine);
2094 dso->kernel = dso_space;
2098 build_id__sprintf(dso->bid.data, sizeof(dso->bid.data),
2100 pr_debug("build id event received for %s: %s\n",
2101 dso->long_name, sbuild_id);
2110 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
2111 int input, u64 offset, u64 size)
2113 struct perf_session *session = container_of(header, struct perf_session, header);
2115 struct perf_event_header header;
2116 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
2119 struct perf_record_header_build_id bev;
2120 char filename[PATH_MAX];
2121 u64 limit = offset + size;
2123 while (offset < limit) {
2126 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
2129 if (header->needs_swap)
2130 perf_event_header__bswap(&old_bev.header);
2132 len = old_bev.header.size - sizeof(old_bev);
2133 if (readn(input, filename, len) != len)
2136 bev.header = old_bev.header;
2139 * As the pid is the missing value, we need to fill
2140 * it properly. The header.misc value give us nice hint.
2142 bev.pid = HOST_KERNEL_ID;
2143 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
2144 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
2145 bev.pid = DEFAULT_GUEST_KERNEL_ID;
2147 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
2148 __event_process_build_id(&bev, filename, session);
2150 offset += bev.header.size;
2156 static int perf_header__read_build_ids(struct perf_header *header,
2157 int input, u64 offset, u64 size)
2159 struct perf_session *session = container_of(header, struct perf_session, header);
2160 struct perf_record_header_build_id bev;
2161 char filename[PATH_MAX];
2162 u64 limit = offset + size, orig_offset = offset;
2165 while (offset < limit) {
2168 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
2171 if (header->needs_swap)
2172 perf_event_header__bswap(&bev.header);
2174 len = bev.header.size - sizeof(bev);
2175 if (readn(input, filename, len) != len)
2178 * The a1645ce1 changeset:
2180 * "perf: 'perf kvm' tool for monitoring guest performance from host"
2182 * Added a field to struct perf_record_header_build_id that broke the file
2185 * Since the kernel build-id is the first entry, process the
2186 * table using the old format if the well known
2187 * '[kernel.kallsyms]' string for the kernel build-id has the
2188 * first 4 characters chopped off (where the pid_t sits).
2190 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
2191 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
2193 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
2196 __event_process_build_id(&bev, filename, session);
2198 offset += bev.header.size;
2205 /* Macro for features that simply need to read and store a string. */
2206 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
2207 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
2209 ff->ph->env.__feat_env = do_read_string(ff); \
2210 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
2213 FEAT_PROCESS_STR_FUN(hostname, hostname);
2214 FEAT_PROCESS_STR_FUN(osrelease, os_release);
2215 FEAT_PROCESS_STR_FUN(version, version);
2216 FEAT_PROCESS_STR_FUN(arch, arch);
2217 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
2218 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
2220 static int process_tracing_data(struct feat_fd *ff, void *data)
2222 ssize_t ret = trace_report(ff->fd, data, false);
2224 return ret < 0 ? -1 : 0;
2227 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
2229 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
2230 pr_debug("Failed to read buildids, continuing...\n");
2234 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
2237 u32 nr_cpus_avail, nr_cpus_online;
2239 ret = do_read_u32(ff, &nr_cpus_avail);
2243 ret = do_read_u32(ff, &nr_cpus_online);
2246 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
2247 ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
2251 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
2256 ret = do_read_u64(ff, &total_mem);
2259 ff->ph->env.total_mem = (unsigned long long)total_mem;
2263 static struct evsel *
2264 perf_evlist__find_by_index(struct evlist *evlist, int idx)
2266 struct evsel *evsel;
2268 evlist__for_each_entry(evlist, evsel) {
2269 if (evsel->idx == idx)
2277 perf_evlist__set_event_name(struct evlist *evlist,
2278 struct evsel *event)
2280 struct evsel *evsel;
2285 evsel = perf_evlist__find_by_index(evlist, event->idx);
2292 evsel->name = strdup(event->name);
2296 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
2298 struct perf_session *session;
2299 struct evsel *evsel, *events = read_event_desc(ff);
2304 session = container_of(ff->ph, struct perf_session, header);
2306 if (session->data->is_pipe) {
2307 /* Save events for reading later by print_event_desc,
2308 * since they can't be read again in pipe mode. */
2309 ff->events = events;
2312 for (evsel = events; evsel->core.attr.size; evsel++)
2313 perf_evlist__set_event_name(session->evlist, evsel);
2315 if (!session->data->is_pipe)
2316 free_event_desc(events);
2321 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
2323 char *str, *cmdline = NULL, **argv = NULL;
2326 if (do_read_u32(ff, &nr))
2329 ff->ph->env.nr_cmdline = nr;
2331 cmdline = zalloc(ff->size + nr + 1);
2335 argv = zalloc(sizeof(char *) * (nr + 1));
2339 for (i = 0; i < nr; i++) {
2340 str = do_read_string(ff);
2344 argv[i] = cmdline + len;
2345 memcpy(argv[i], str, strlen(str) + 1);
2346 len += strlen(str) + 1;
2349 ff->ph->env.cmdline = cmdline;
2350 ff->ph->env.cmdline_argv = (const char **) argv;
2359 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
2364 int cpu_nr = ff->ph->env.nr_cpus_avail;
2366 struct perf_header *ph = ff->ph;
2367 bool do_core_id_test = true;
2369 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
2373 if (do_read_u32(ff, &nr))
2376 ph->env.nr_sibling_cores = nr;
2377 size += sizeof(u32);
2378 if (strbuf_init(&sb, 128) < 0)
2381 for (i = 0; i < nr; i++) {
2382 str = do_read_string(ff);
2386 /* include a NULL character at the end */
2387 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2389 size += string_size(str);
2392 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
2394 if (do_read_u32(ff, &nr))
2397 ph->env.nr_sibling_threads = nr;
2398 size += sizeof(u32);
2400 for (i = 0; i < nr; i++) {
2401 str = do_read_string(ff);
2405 /* include a NULL character at the end */
2406 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2408 size += string_size(str);
2411 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
2414 * The header may be from old perf,
2415 * which doesn't include core id and socket id information.
2417 if (ff->size <= size) {
2418 zfree(&ph->env.cpu);
2422 /* On s390 the socket_id number is not related to the numbers of cpus.
2423 * The socket_id number might be higher than the numbers of cpus.
2424 * This depends on the configuration.
2425 * AArch64 is the same.
2427 if (ph->env.arch && (!strncmp(ph->env.arch, "s390", 4)
2428 || !strncmp(ph->env.arch, "aarch64", 7)))
2429 do_core_id_test = false;
2431 for (i = 0; i < (u32)cpu_nr; i++) {
2432 if (do_read_u32(ff, &nr))
2435 ph->env.cpu[i].core_id = nr;
2436 size += sizeof(u32);
2438 if (do_read_u32(ff, &nr))
2441 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
2442 pr_debug("socket_id number is too big."
2443 "You may need to upgrade the perf tool.\n");
2447 ph->env.cpu[i].socket_id = nr;
2448 size += sizeof(u32);
2452 * The header may be from old perf,
2453 * which doesn't include die information.
2455 if (ff->size <= size)
2458 if (do_read_u32(ff, &nr))
2461 ph->env.nr_sibling_dies = nr;
2462 size += sizeof(u32);
2464 for (i = 0; i < nr; i++) {
2465 str = do_read_string(ff);
2469 /* include a NULL character at the end */
2470 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2472 size += string_size(str);
2475 ph->env.sibling_dies = strbuf_detach(&sb, NULL);
2477 for (i = 0; i < (u32)cpu_nr; i++) {
2478 if (do_read_u32(ff, &nr))
2481 ph->env.cpu[i].die_id = nr;
2487 strbuf_release(&sb);
2489 zfree(&ph->env.cpu);
2493 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
2495 struct numa_node *nodes, *n;
2500 if (do_read_u32(ff, &nr))
2503 nodes = zalloc(sizeof(*nodes) * nr);
2507 for (i = 0; i < nr; i++) {
2511 if (do_read_u32(ff, &n->node))
2514 if (do_read_u64(ff, &n->mem_total))
2517 if (do_read_u64(ff, &n->mem_free))
2520 str = do_read_string(ff);
2524 n->map = perf_cpu_map__new(str);
2530 ff->ph->env.nr_numa_nodes = nr;
2531 ff->ph->env.numa_nodes = nodes;
2539 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
2546 if (do_read_u32(ff, &pmu_num))
2550 pr_debug("pmu mappings not available\n");
2554 ff->ph->env.nr_pmu_mappings = pmu_num;
2555 if (strbuf_init(&sb, 128) < 0)
2559 if (do_read_u32(ff, &type))
2562 name = do_read_string(ff);
2566 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
2568 /* include a NULL character at the end */
2569 if (strbuf_add(&sb, "", 1) < 0)
2572 if (!strcmp(name, "msr"))
2573 ff->ph->env.msr_pmu_type = type;
2578 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2582 strbuf_release(&sb);
2586 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2589 u32 i, nr, nr_groups;
2590 struct perf_session *session;
2591 struct evsel *evsel, *leader = NULL;
2598 if (do_read_u32(ff, &nr_groups))
2601 ff->ph->env.nr_groups = nr_groups;
2603 pr_debug("group desc not available\n");
2607 desc = calloc(nr_groups, sizeof(*desc));
2611 for (i = 0; i < nr_groups; i++) {
2612 desc[i].name = do_read_string(ff);
2616 if (do_read_u32(ff, &desc[i].leader_idx))
2619 if (do_read_u32(ff, &desc[i].nr_members))
2624 * Rebuild group relationship based on the group_desc
2626 session = container_of(ff->ph, struct perf_session, header);
2627 session->evlist->nr_groups = nr_groups;
2630 evlist__for_each_entry(session->evlist, evsel) {
2631 if (evsel->idx == (int) desc[i].leader_idx) {
2632 evsel->leader = evsel;
2633 /* {anon_group} is a dummy name */
2634 if (strcmp(desc[i].name, "{anon_group}")) {
2635 evsel->group_name = desc[i].name;
2636 desc[i].name = NULL;
2638 evsel->core.nr_members = desc[i].nr_members;
2640 if (i >= nr_groups || nr > 0) {
2641 pr_debug("invalid group desc\n");
2646 nr = evsel->core.nr_members - 1;
2649 /* This is a group member */
2650 evsel->leader = leader;
2656 if (i != nr_groups || nr != 0) {
2657 pr_debug("invalid group desc\n");
2663 for (i = 0; i < nr_groups; i++)
2664 zfree(&desc[i].name);
2670 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2672 struct perf_session *session;
2675 session = container_of(ff->ph, struct perf_session, header);
2677 err = auxtrace_index__process(ff->fd, ff->size, session,
2678 ff->ph->needs_swap);
2680 pr_err("Failed to process auxtrace index\n");
2684 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2686 struct cpu_cache_level *caches;
2687 u32 cnt, i, version;
2689 if (do_read_u32(ff, &version))
2695 if (do_read_u32(ff, &cnt))
2698 caches = zalloc(sizeof(*caches) * cnt);
2702 for (i = 0; i < cnt; i++) {
2703 struct cpu_cache_level c;
2706 if (do_read_u32(ff, &c.v))\
2707 goto out_free_caches; \
2716 c.v = do_read_string(ff); \
2718 goto out_free_caches;
2728 ff->ph->env.caches = caches;
2729 ff->ph->env.caches_cnt = cnt;
2736 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused)
2738 struct perf_session *session;
2739 u64 first_sample_time, last_sample_time;
2742 session = container_of(ff->ph, struct perf_session, header);
2744 ret = do_read_u64(ff, &first_sample_time);
2748 ret = do_read_u64(ff, &last_sample_time);
2752 session->evlist->first_sample_time = first_sample_time;
2753 session->evlist->last_sample_time = last_sample_time;
2757 static int process_mem_topology(struct feat_fd *ff,
2758 void *data __maybe_unused)
2760 struct memory_node *nodes;
2761 u64 version, i, nr, bsize;
2764 if (do_read_u64(ff, &version))
2770 if (do_read_u64(ff, &bsize))
2773 if (do_read_u64(ff, &nr))
2776 nodes = zalloc(sizeof(*nodes) * nr);
2780 for (i = 0; i < nr; i++) {
2781 struct memory_node n;
2784 if (do_read_u64(ff, &n.v)) \
2792 if (do_read_bitmap(ff, &n.set, &n.size))
2798 ff->ph->env.memory_bsize = bsize;
2799 ff->ph->env.memory_nodes = nodes;
2800 ff->ph->env.nr_memory_nodes = nr;
2809 static int process_clockid(struct feat_fd *ff,
2810 void *data __maybe_unused)
2812 if (do_read_u64(ff, &ff->ph->env.clock.clockid_res_ns))
2818 static int process_clock_data(struct feat_fd *ff,
2819 void *_data __maybe_unused)
2825 if (do_read_u32(ff, &data32))
2832 if (do_read_u32(ff, &data32))
2835 ff->ph->env.clock.clockid = data32;
2838 if (do_read_u64(ff, &data64))
2841 ff->ph->env.clock.tod_ns = data64;
2843 /* clockid ref time */
2844 if (do_read_u64(ff, &data64))
2847 ff->ph->env.clock.clockid_ns = data64;
2848 ff->ph->env.clock.enabled = true;
2852 static int process_dir_format(struct feat_fd *ff,
2853 void *_data __maybe_unused)
2855 struct perf_session *session;
2856 struct perf_data *data;
2858 session = container_of(ff->ph, struct perf_session, header);
2859 data = session->data;
2861 if (WARN_ON(!perf_data__is_dir(data)))
2864 return do_read_u64(ff, &data->dir.version);
2867 #ifdef HAVE_LIBBPF_SUPPORT
2868 static int process_bpf_prog_info(struct feat_fd *ff, void *data __maybe_unused)
2870 struct bpf_prog_info_linear *info_linear;
2871 struct bpf_prog_info_node *info_node;
2872 struct perf_env *env = &ff->ph->env;
2876 if (ff->ph->needs_swap) {
2877 pr_warning("interpreting bpf_prog_info from systems with endianity is not yet supported\n");
2881 if (do_read_u32(ff, &count))
2884 down_write(&env->bpf_progs.lock);
2886 for (i = 0; i < count; ++i) {
2887 u32 info_len, data_len;
2891 if (do_read_u32(ff, &info_len))
2893 if (do_read_u32(ff, &data_len))
2896 if (info_len > sizeof(struct bpf_prog_info)) {
2897 pr_warning("detected invalid bpf_prog_info\n");
2901 info_linear = malloc(sizeof(struct bpf_prog_info_linear) +
2905 info_linear->info_len = sizeof(struct bpf_prog_info);
2906 info_linear->data_len = data_len;
2907 if (do_read_u64(ff, (u64 *)(&info_linear->arrays)))
2909 if (__do_read(ff, &info_linear->info, info_len))
2911 if (info_len < sizeof(struct bpf_prog_info))
2912 memset(((void *)(&info_linear->info)) + info_len, 0,
2913 sizeof(struct bpf_prog_info) - info_len);
2915 if (__do_read(ff, info_linear->data, data_len))
2918 info_node = malloc(sizeof(struct bpf_prog_info_node));
2922 /* after reading from file, translate offset to address */
2923 bpf_program__bpil_offs_to_addr(info_linear);
2924 info_node->info_linear = info_linear;
2925 perf_env__insert_bpf_prog_info(env, info_node);
2928 up_write(&env->bpf_progs.lock);
2933 up_write(&env->bpf_progs.lock);
2936 #else // HAVE_LIBBPF_SUPPORT
2937 static int process_bpf_prog_info(struct feat_fd *ff __maybe_unused, void *data __maybe_unused)
2941 #endif // HAVE_LIBBPF_SUPPORT
2943 static int process_bpf_btf(struct feat_fd *ff, void *data __maybe_unused)
2945 struct perf_env *env = &ff->ph->env;
2946 struct btf_node *node = NULL;
2950 if (ff->ph->needs_swap) {
2951 pr_warning("interpreting btf from systems with endianity is not yet supported\n");
2955 if (do_read_u32(ff, &count))
2958 down_write(&env->bpf_progs.lock);
2960 for (i = 0; i < count; ++i) {
2963 if (do_read_u32(ff, &id))
2965 if (do_read_u32(ff, &data_size))
2968 node = malloc(sizeof(struct btf_node) + data_size);
2973 node->data_size = data_size;
2975 if (__do_read(ff, node->data, data_size))
2978 perf_env__insert_btf(env, node);
2984 up_write(&env->bpf_progs.lock);
2989 static int process_compressed(struct feat_fd *ff,
2990 void *data __maybe_unused)
2992 if (do_read_u32(ff, &(ff->ph->env.comp_ver)))
2995 if (do_read_u32(ff, &(ff->ph->env.comp_type)))
2998 if (do_read_u32(ff, &(ff->ph->env.comp_level)))
3001 if (do_read_u32(ff, &(ff->ph->env.comp_ratio)))
3004 if (do_read_u32(ff, &(ff->ph->env.comp_mmap_len)))
3010 static int process_cpu_pmu_caps(struct feat_fd *ff,
3011 void *data __maybe_unused)
3017 if (do_read_u32(ff, &nr_caps))
3021 pr_debug("cpu pmu capabilities not available\n");
3025 ff->ph->env.nr_cpu_pmu_caps = nr_caps;
3027 if (strbuf_init(&sb, 128) < 0)
3031 name = do_read_string(ff);
3035 value = do_read_string(ff);
3039 if (strbuf_addf(&sb, "%s=%s", name, value) < 0)
3042 /* include a NULL character at the end */
3043 if (strbuf_add(&sb, "", 1) < 0)
3046 if (!strcmp(name, "branches"))
3047 ff->ph->env.max_branches = atoi(value);
3052 ff->ph->env.cpu_pmu_caps = strbuf_detach(&sb, NULL);
3060 strbuf_release(&sb);
3064 #define FEAT_OPR(n, func, __full_only) \
3066 .name = __stringify(n), \
3067 .write = write_##func, \
3068 .print = print_##func, \
3069 .full_only = __full_only, \
3070 .process = process_##func, \
3071 .synthesize = true \
3074 #define FEAT_OPN(n, func, __full_only) \
3076 .name = __stringify(n), \
3077 .write = write_##func, \
3078 .print = print_##func, \
3079 .full_only = __full_only, \
3080 .process = process_##func \
3083 /* feature_ops not implemented: */
3084 #define print_tracing_data NULL
3085 #define print_build_id NULL
3087 #define process_branch_stack NULL
3088 #define process_stat NULL
3090 // Only used in util/synthetic-events.c
3091 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
3093 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE] = {
3094 FEAT_OPN(TRACING_DATA, tracing_data, false),
3095 FEAT_OPN(BUILD_ID, build_id, false),
3096 FEAT_OPR(HOSTNAME, hostname, false),
3097 FEAT_OPR(OSRELEASE, osrelease, false),
3098 FEAT_OPR(VERSION, version, false),
3099 FEAT_OPR(ARCH, arch, false),
3100 FEAT_OPR(NRCPUS, nrcpus, false),
3101 FEAT_OPR(CPUDESC, cpudesc, false),
3102 FEAT_OPR(CPUID, cpuid, false),
3103 FEAT_OPR(TOTAL_MEM, total_mem, false),
3104 FEAT_OPR(EVENT_DESC, event_desc, false),
3105 FEAT_OPR(CMDLINE, cmdline, false),
3106 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true),
3107 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true),
3108 FEAT_OPN(BRANCH_STACK, branch_stack, false),
3109 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false),
3110 FEAT_OPR(GROUP_DESC, group_desc, false),
3111 FEAT_OPN(AUXTRACE, auxtrace, false),
3112 FEAT_OPN(STAT, stat, false),
3113 FEAT_OPN(CACHE, cache, true),
3114 FEAT_OPR(SAMPLE_TIME, sample_time, false),
3115 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true),
3116 FEAT_OPR(CLOCKID, clockid, false),
3117 FEAT_OPN(DIR_FORMAT, dir_format, false),
3118 FEAT_OPR(BPF_PROG_INFO, bpf_prog_info, false),
3119 FEAT_OPR(BPF_BTF, bpf_btf, false),
3120 FEAT_OPR(COMPRESSED, compressed, false),
3121 FEAT_OPR(CPU_PMU_CAPS, cpu_pmu_caps, false),
3122 FEAT_OPR(CLOCK_DATA, clock_data, false),
3125 struct header_print_data {
3127 bool full; /* extended list of headers */
3130 static int perf_file_section__fprintf_info(struct perf_file_section *section,
3131 struct perf_header *ph,
3132 int feat, int fd, void *data)
3134 struct header_print_data *hd = data;
3137 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3138 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3139 "%d, continuing...\n", section->offset, feat);
3142 if (feat >= HEADER_LAST_FEATURE) {
3143 pr_warning("unknown feature %d\n", feat);
3146 if (!feat_ops[feat].print)
3149 ff = (struct feat_fd) {
3154 if (!feat_ops[feat].full_only || hd->full)
3155 feat_ops[feat].print(&ff, hd->fp);
3157 fprintf(hd->fp, "# %s info available, use -I to display\n",
3158 feat_ops[feat].name);
3163 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
3165 struct header_print_data hd;
3166 struct perf_header *header = &session->header;
3167 int fd = perf_data__fd(session->data);
3175 ret = fstat(fd, &st);
3179 stctime = st.st_mtime;
3180 fprintf(fp, "# captured on : %s", ctime(&stctime));
3182 fprintf(fp, "# header version : %u\n", header->version);
3183 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset);
3184 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size);
3185 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset);
3187 perf_header__process_sections(header, fd, &hd,
3188 perf_file_section__fprintf_info);
3190 if (session->data->is_pipe)
3193 fprintf(fp, "# missing features: ");
3194 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
3196 fprintf(fp, "%s ", feat_ops[bit].name);
3203 static int do_write_feat(struct feat_fd *ff, int type,
3204 struct perf_file_section **p,
3205 struct evlist *evlist)
3210 if (perf_header__has_feat(ff->ph, type)) {
3211 if (!feat_ops[type].write)
3214 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
3217 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
3219 err = feat_ops[type].write(ff, evlist);
3221 pr_debug("failed to write feature %s\n", feat_ops[type].name);
3223 /* undo anything written */
3224 lseek(ff->fd, (*p)->offset, SEEK_SET);
3228 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
3234 static int perf_header__adds_write(struct perf_header *header,
3235 struct evlist *evlist, int fd)
3239 struct perf_file_section *feat_sec, *p;
3245 ff = (struct feat_fd){
3250 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3254 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
3255 if (feat_sec == NULL)
3258 sec_size = sizeof(*feat_sec) * nr_sections;
3260 sec_start = header->feat_offset;
3261 lseek(fd, sec_start + sec_size, SEEK_SET);
3263 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3264 if (do_write_feat(&ff, feat, &p, evlist))
3265 perf_header__clear_feat(header, feat);
3268 lseek(fd, sec_start, SEEK_SET);
3270 * may write more than needed due to dropped feature, but
3271 * this is okay, reader will skip the missing entries
3273 err = do_write(&ff, feat_sec, sec_size);
3275 pr_debug("failed to write feature section\n");
3280 int perf_header__write_pipe(int fd)
3282 struct perf_pipe_file_header f_header;
3286 ff = (struct feat_fd){ .fd = fd };
3288 f_header = (struct perf_pipe_file_header){
3289 .magic = PERF_MAGIC,
3290 .size = sizeof(f_header),
3293 err = do_write(&ff, &f_header, sizeof(f_header));
3295 pr_debug("failed to write perf pipe header\n");
3302 int perf_session__write_header(struct perf_session *session,
3303 struct evlist *evlist,
3304 int fd, bool at_exit)
3306 struct perf_file_header f_header;
3307 struct perf_file_attr f_attr;
3308 struct perf_header *header = &session->header;
3309 struct evsel *evsel;
3314 ff = (struct feat_fd){ .fd = fd};
3315 lseek(fd, sizeof(f_header), SEEK_SET);
3317 evlist__for_each_entry(session->evlist, evsel) {
3318 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
3319 err = do_write(&ff, evsel->core.id, evsel->core.ids * sizeof(u64));
3321 pr_debug("failed to write perf header\n");
3326 attr_offset = lseek(ff.fd, 0, SEEK_CUR);
3328 evlist__for_each_entry(evlist, evsel) {
3329 f_attr = (struct perf_file_attr){
3330 .attr = evsel->core.attr,
3332 .offset = evsel->id_offset,
3333 .size = evsel->core.ids * sizeof(u64),
3336 err = do_write(&ff, &f_attr, sizeof(f_attr));
3338 pr_debug("failed to write perf header attribute\n");
3343 if (!header->data_offset)
3344 header->data_offset = lseek(fd, 0, SEEK_CUR);
3345 header->feat_offset = header->data_offset + header->data_size;
3348 err = perf_header__adds_write(header, evlist, fd);
3353 f_header = (struct perf_file_header){
3354 .magic = PERF_MAGIC,
3355 .size = sizeof(f_header),
3356 .attr_size = sizeof(f_attr),
3358 .offset = attr_offset,
3359 .size = evlist->core.nr_entries * sizeof(f_attr),
3362 .offset = header->data_offset,
3363 .size = header->data_size,
3365 /* event_types is ignored, store zeros */
3368 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
3370 lseek(fd, 0, SEEK_SET);
3371 err = do_write(&ff, &f_header, sizeof(f_header));
3373 pr_debug("failed to write perf header\n");
3376 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
3381 static int perf_header__getbuffer64(struct perf_header *header,
3382 int fd, void *buf, size_t size)
3384 if (readn(fd, buf, size) <= 0)
3387 if (header->needs_swap)
3388 mem_bswap_64(buf, size);
3393 int perf_header__process_sections(struct perf_header *header, int fd,
3395 int (*process)(struct perf_file_section *section,
3396 struct perf_header *ph,
3397 int feat, int fd, void *data))
3399 struct perf_file_section *feat_sec, *sec;
3405 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3409 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
3413 sec_size = sizeof(*feat_sec) * nr_sections;
3415 lseek(fd, header->feat_offset, SEEK_SET);
3417 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
3421 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
3422 err = process(sec++, header, feat, fd, data);
3432 static const int attr_file_abi_sizes[] = {
3433 [0] = PERF_ATTR_SIZE_VER0,
3434 [1] = PERF_ATTR_SIZE_VER1,
3435 [2] = PERF_ATTR_SIZE_VER2,
3436 [3] = PERF_ATTR_SIZE_VER3,
3437 [4] = PERF_ATTR_SIZE_VER4,
3442 * In the legacy file format, the magic number is not used to encode endianness.
3443 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
3444 * on ABI revisions, we need to try all combinations for all endianness to
3445 * detect the endianness.
3447 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
3449 uint64_t ref_size, attr_size;
3452 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
3453 ref_size = attr_file_abi_sizes[i]
3454 + sizeof(struct perf_file_section);
3455 if (hdr_sz != ref_size) {
3456 attr_size = bswap_64(hdr_sz);
3457 if (attr_size != ref_size)
3460 ph->needs_swap = true;
3462 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
3467 /* could not determine endianness */
3471 #define PERF_PIPE_HDR_VER0 16
3473 static const size_t attr_pipe_abi_sizes[] = {
3474 [0] = PERF_PIPE_HDR_VER0,
3479 * In the legacy pipe format, there is an implicit assumption that endiannesss
3480 * between host recording the samples, and host parsing the samples is the
3481 * same. This is not always the case given that the pipe output may always be
3482 * redirected into a file and analyzed on a different machine with possibly a
3483 * different endianness and perf_event ABI revsions in the perf tool itself.
3485 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
3490 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
3491 if (hdr_sz != attr_pipe_abi_sizes[i]) {
3492 attr_size = bswap_64(hdr_sz);
3493 if (attr_size != hdr_sz)
3496 ph->needs_swap = true;
3498 pr_debug("Pipe ABI%d perf.data file detected\n", i);
3504 bool is_perf_magic(u64 magic)
3506 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
3507 || magic == __perf_magic2
3508 || magic == __perf_magic2_sw)
3514 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
3515 bool is_pipe, struct perf_header *ph)
3519 /* check for legacy format */
3520 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
3522 ph->version = PERF_HEADER_VERSION_1;
3523 pr_debug("legacy perf.data format\n");
3525 return try_all_pipe_abis(hdr_sz, ph);
3527 return try_all_file_abis(hdr_sz, ph);
3530 * the new magic number serves two purposes:
3531 * - unique number to identify actual perf.data files
3532 * - encode endianness of file
3534 ph->version = PERF_HEADER_VERSION_2;
3536 /* check magic number with one endianness */
3537 if (magic == __perf_magic2)
3540 /* check magic number with opposite endianness */
3541 if (magic != __perf_magic2_sw)
3544 ph->needs_swap = true;
3549 int perf_file_header__read(struct perf_file_header *header,
3550 struct perf_header *ph, int fd)
3554 lseek(fd, 0, SEEK_SET);
3556 ret = readn(fd, header, sizeof(*header));
3560 if (check_magic_endian(header->magic,
3561 header->attr_size, false, ph) < 0) {
3562 pr_debug("magic/endian check failed\n");
3566 if (ph->needs_swap) {
3567 mem_bswap_64(header, offsetof(struct perf_file_header,
3571 if (header->size != sizeof(*header)) {
3572 /* Support the previous format */
3573 if (header->size == offsetof(typeof(*header), adds_features))
3574 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3577 } else if (ph->needs_swap) {
3579 * feature bitmap is declared as an array of unsigned longs --
3580 * not good since its size can differ between the host that
3581 * generated the data file and the host analyzing the file.
3583 * We need to handle endianness, but we don't know the size of
3584 * the unsigned long where the file was generated. Take a best
3585 * guess at determining it: try 64-bit swap first (ie., file
3586 * created on a 64-bit host), and check if the hostname feature
3587 * bit is set (this feature bit is forced on as of fbe96f2).
3588 * If the bit is not, undo the 64-bit swap and try a 32-bit
3589 * swap. If the hostname bit is still not set (e.g., older data
3590 * file), punt and fallback to the original behavior --
3591 * clearing all feature bits and setting buildid.
3593 mem_bswap_64(&header->adds_features,
3594 BITS_TO_U64(HEADER_FEAT_BITS));
3596 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3598 mem_bswap_64(&header->adds_features,
3599 BITS_TO_U64(HEADER_FEAT_BITS));
3602 mem_bswap_32(&header->adds_features,
3603 BITS_TO_U32(HEADER_FEAT_BITS));
3606 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3607 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3608 set_bit(HEADER_BUILD_ID, header->adds_features);
3612 memcpy(&ph->adds_features, &header->adds_features,
3613 sizeof(ph->adds_features));
3615 ph->data_offset = header->data.offset;
3616 ph->data_size = header->data.size;
3617 ph->feat_offset = header->data.offset + header->data.size;
3621 static int perf_file_section__process(struct perf_file_section *section,
3622 struct perf_header *ph,
3623 int feat, int fd, void *data)
3625 struct feat_fd fdd = {
3628 .size = section->size,
3629 .offset = section->offset,
3632 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3633 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3634 "%d, continuing...\n", section->offset, feat);
3638 if (feat >= HEADER_LAST_FEATURE) {
3639 pr_debug("unknown feature %d, continuing...\n", feat);
3643 if (!feat_ops[feat].process)
3646 return feat_ops[feat].process(&fdd, data);
3649 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
3650 struct perf_header *ph, int fd,
3653 struct feat_fd ff = {
3654 .fd = STDOUT_FILENO,
3659 ret = readn(fd, header, sizeof(*header));
3663 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
3664 pr_debug("endian/magic failed\n");
3669 header->size = bswap_64(header->size);
3671 if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
3677 static int perf_header__read_pipe(struct perf_session *session)
3679 struct perf_header *header = &session->header;
3680 struct perf_pipe_file_header f_header;
3682 if (perf_file_header__read_pipe(&f_header, header,
3683 perf_data__fd(session->data),
3684 session->repipe) < 0) {
3685 pr_debug("incompatible file format\n");
3689 return f_header.size == sizeof(f_header) ? 0 : -1;
3692 static int read_attr(int fd, struct perf_header *ph,
3693 struct perf_file_attr *f_attr)
3695 struct perf_event_attr *attr = &f_attr->attr;
3697 size_t our_sz = sizeof(f_attr->attr);
3700 memset(f_attr, 0, sizeof(*f_attr));
3702 /* read minimal guaranteed structure */
3703 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
3705 pr_debug("cannot read %d bytes of header attr\n",
3706 PERF_ATTR_SIZE_VER0);
3710 /* on file perf_event_attr size */
3718 sz = PERF_ATTR_SIZE_VER0;
3719 } else if (sz > our_sz) {
3720 pr_debug("file uses a more recent and unsupported ABI"
3721 " (%zu bytes extra)\n", sz - our_sz);
3724 /* what we have not yet read and that we know about */
3725 left = sz - PERF_ATTR_SIZE_VER0;
3728 ptr += PERF_ATTR_SIZE_VER0;
3730 ret = readn(fd, ptr, left);
3732 /* read perf_file_section, ids are read in caller */
3733 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
3735 return ret <= 0 ? -1 : 0;
3738 static int perf_evsel__prepare_tracepoint_event(struct evsel *evsel,
3739 struct tep_handle *pevent)
3741 struct tep_event *event;
3744 /* already prepared */
3745 if (evsel->tp_format)
3748 if (pevent == NULL) {
3749 pr_debug("broken or missing trace data\n");
3753 event = tep_find_event(pevent, evsel->core.attr.config);
3754 if (event == NULL) {
3755 pr_debug("cannot find event format for %d\n", (int)evsel->core.attr.config);
3760 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
3761 evsel->name = strdup(bf);
3762 if (evsel->name == NULL)
3766 evsel->tp_format = event;
3770 static int perf_evlist__prepare_tracepoint_events(struct evlist *evlist,
3771 struct tep_handle *pevent)
3775 evlist__for_each_entry(evlist, pos) {
3776 if (pos->core.attr.type == PERF_TYPE_TRACEPOINT &&
3777 perf_evsel__prepare_tracepoint_event(pos, pevent))
3784 int perf_session__read_header(struct perf_session *session)
3786 struct perf_data *data = session->data;
3787 struct perf_header *header = &session->header;
3788 struct perf_file_header f_header;
3789 struct perf_file_attr f_attr;
3791 int nr_attrs, nr_ids, i, j, err;
3792 int fd = perf_data__fd(data);
3794 session->evlist = evlist__new();
3795 if (session->evlist == NULL)
3798 session->evlist->env = &header->env;
3799 session->machines.host.env = &header->env;
3802 * We can read 'pipe' data event from regular file,
3803 * check for the pipe header regardless of source.
3805 err = perf_header__read_pipe(session);
3806 if (!err || (err && perf_data__is_pipe(data))) {
3807 data->is_pipe = true;
3811 if (perf_file_header__read(&f_header, header, fd) < 0)
3815 * Sanity check that perf.data was written cleanly; data size is
3816 * initialized to 0 and updated only if the on_exit function is run.
3817 * If data size is still 0 then the file contains only partial
3818 * information. Just warn user and process it as much as it can.
3820 if (f_header.data.size == 0) {
3821 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3822 "Was the 'perf record' command properly terminated?\n",
3826 if (f_header.attr_size == 0) {
3827 pr_err("ERROR: The %s file's attr size field is 0 which is unexpected.\n"
3828 "Was the 'perf record' command properly terminated?\n",
3833 nr_attrs = f_header.attrs.size / f_header.attr_size;
3834 lseek(fd, f_header.attrs.offset, SEEK_SET);
3836 for (i = 0; i < nr_attrs; i++) {
3837 struct evsel *evsel;
3840 if (read_attr(fd, header, &f_attr) < 0)
3843 if (header->needs_swap) {
3844 f_attr.ids.size = bswap_64(f_attr.ids.size);
3845 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
3846 perf_event__attr_swap(&f_attr.attr);
3849 tmp = lseek(fd, 0, SEEK_CUR);
3850 evsel = evsel__new(&f_attr.attr);
3853 goto out_delete_evlist;
3855 evsel->needs_swap = header->needs_swap;
3857 * Do it before so that if perf_evsel__alloc_id fails, this
3858 * entry gets purged too at evlist__delete().
3860 evlist__add(session->evlist, evsel);
3862 nr_ids = f_attr.ids.size / sizeof(u64);
3864 * We don't have the cpu and thread maps on the header, so
3865 * for allocating the perf_sample_id table we fake 1 cpu and
3866 * hattr->ids threads.
3868 if (perf_evsel__alloc_id(&evsel->core, 1, nr_ids))
3869 goto out_delete_evlist;
3871 lseek(fd, f_attr.ids.offset, SEEK_SET);
3873 for (j = 0; j < nr_ids; j++) {
3874 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
3877 perf_evlist__id_add(&session->evlist->core, &evsel->core, 0, j, f_id);
3880 lseek(fd, tmp, SEEK_SET);
3883 perf_header__process_sections(header, fd, &session->tevent,
3884 perf_file_section__process);
3886 if (perf_evlist__prepare_tracepoint_events(session->evlist,
3887 session->tevent.pevent))
3888 goto out_delete_evlist;
3895 evlist__delete(session->evlist);
3896 session->evlist = NULL;
3900 int perf_event__process_feature(struct perf_session *session,
3901 union perf_event *event)
3903 struct perf_tool *tool = session->tool;
3904 struct feat_fd ff = { .fd = 0 };
3905 struct perf_record_header_feature *fe = (struct perf_record_header_feature *)event;
3906 int type = fe->header.type;
3907 u64 feat = fe->feat_id;
3909 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3910 pr_warning("invalid record type %d in pipe-mode\n", type);
3913 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) {
3914 pr_warning("invalid record type %d in pipe-mode\n", type);
3918 if (!feat_ops[feat].process)
3921 ff.buf = (void *)fe->data;
3922 ff.size = event->header.size - sizeof(*fe);
3923 ff.ph = &session->header;
3925 if (feat_ops[feat].process(&ff, NULL))
3928 if (!feat_ops[feat].print || !tool->show_feat_hdr)
3931 if (!feat_ops[feat].full_only ||
3932 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3933 feat_ops[feat].print(&ff, stdout);
3935 fprintf(stdout, "# %s info available, use -I to display\n",
3936 feat_ops[feat].name);
3942 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3944 struct perf_record_event_update *ev = &event->event_update;
3945 struct perf_record_event_update_scale *ev_scale;
3946 struct perf_record_event_update_cpus *ev_cpus;
3947 struct perf_cpu_map *map;
3950 ret = fprintf(fp, "\n... id: %" PRI_lu64 "\n", ev->id);
3953 case PERF_EVENT_UPDATE__SCALE:
3954 ev_scale = (struct perf_record_event_update_scale *)ev->data;
3955 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3957 case PERF_EVENT_UPDATE__UNIT:
3958 ret += fprintf(fp, "... unit: %s\n", ev->data);
3960 case PERF_EVENT_UPDATE__NAME:
3961 ret += fprintf(fp, "... name: %s\n", ev->data);
3963 case PERF_EVENT_UPDATE__CPUS:
3964 ev_cpus = (struct perf_record_event_update_cpus *)ev->data;
3965 ret += fprintf(fp, "... ");
3967 map = cpu_map__new_data(&ev_cpus->cpus);
3969 ret += cpu_map__fprintf(map, fp);
3971 ret += fprintf(fp, "failed to get cpus\n");
3974 ret += fprintf(fp, "... unknown type\n");
3981 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3982 union perf_event *event,
3983 struct evlist **pevlist)
3986 struct evsel *evsel;
3987 struct evlist *evlist = *pevlist;
3989 if (evlist == NULL) {
3990 *pevlist = evlist = evlist__new();
3995 evsel = evsel__new(&event->attr.attr);
3999 evlist__add(evlist, evsel);
4001 ids = event->header.size;
4002 ids -= (void *)&event->attr.id - (void *)event;
4003 n_ids = ids / sizeof(u64);
4005 * We don't have the cpu and thread maps on the header, so
4006 * for allocating the perf_sample_id table we fake 1 cpu and
4007 * hattr->ids threads.
4009 if (perf_evsel__alloc_id(&evsel->core, 1, n_ids))
4012 for (i = 0; i < n_ids; i++) {
4013 perf_evlist__id_add(&evlist->core, &evsel->core, 0, i, event->attr.id[i]);
4019 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
4020 union perf_event *event,
4021 struct evlist **pevlist)
4023 struct perf_record_event_update *ev = &event->event_update;
4024 struct perf_record_event_update_scale *ev_scale;
4025 struct perf_record_event_update_cpus *ev_cpus;
4026 struct evlist *evlist;
4027 struct evsel *evsel;
4028 struct perf_cpu_map *map;
4030 if (!pevlist || *pevlist == NULL)
4035 evsel = perf_evlist__id2evsel(evlist, ev->id);
4040 case PERF_EVENT_UPDATE__UNIT:
4041 evsel->unit = strdup(ev->data);
4043 case PERF_EVENT_UPDATE__NAME:
4044 evsel->name = strdup(ev->data);
4046 case PERF_EVENT_UPDATE__SCALE:
4047 ev_scale = (struct perf_record_event_update_scale *)ev->data;
4048 evsel->scale = ev_scale->scale;
4050 case PERF_EVENT_UPDATE__CPUS:
4051 ev_cpus = (struct perf_record_event_update_cpus *)ev->data;
4053 map = cpu_map__new_data(&ev_cpus->cpus);
4055 evsel->core.own_cpus = map;
4057 pr_err("failed to get event_update cpus\n");
4065 int perf_event__process_tracing_data(struct perf_session *session,
4066 union perf_event *event)
4068 ssize_t size_read, padding, size = event->tracing_data.size;
4069 int fd = perf_data__fd(session->data);
4073 * The pipe fd is already in proper place and in any case
4074 * we can't move it, and we'd screw the case where we read
4075 * 'pipe' data from regular file. The trace_report reads
4076 * data from 'fd' so we need to set it directly behind the
4077 * event, where the tracing data starts.
4079 if (!perf_data__is_pipe(session->data)) {
4080 off_t offset = lseek(fd, 0, SEEK_CUR);
4082 /* setup for reading amidst mmap */
4083 lseek(fd, offset + sizeof(struct perf_record_header_tracing_data),
4087 size_read = trace_report(fd, &session->tevent,
4089 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
4091 if (readn(fd, buf, padding) < 0) {
4092 pr_err("%s: reading input file", __func__);
4095 if (session->repipe) {
4096 int retw = write(STDOUT_FILENO, buf, padding);
4097 if (retw <= 0 || retw != padding) {
4098 pr_err("%s: repiping tracing data padding", __func__);
4103 if (size_read + padding != size) {
4104 pr_err("%s: tracing data size mismatch", __func__);
4108 perf_evlist__prepare_tracepoint_events(session->evlist,
4109 session->tevent.pevent);
4111 return size_read + padding;
4114 int perf_event__process_build_id(struct perf_session *session,
4115 union perf_event *event)
4117 __event_process_build_id(&event->build_id,
4118 event->build_id.filename,