1 // SPDX-License-Identifier: GPL-2.0
12 #include <linux/compiler.h>
13 #include <linux/list.h>
14 #include <linux/kernel.h>
15 #include <linux/bitops.h>
16 #include <linux/stringify.h>
18 #include <sys/utsname.h>
19 #include <linux/time64.h>
21 #include <bpf/libbpf.h>
28 #include "trace-event.h"
38 #include <api/fs/fs.h>
41 #include "time-utils.h"
44 #include "bpf-event.h"
46 #include "sane_ctype.h"
50 * must be a numerical value to let the endianness
51 * determine the memory layout. That way we are able
52 * to detect endianness when reading the perf.data file
55 * we check for legacy (PERFFILE) format.
57 static const char *__perf_magic1 = "PERFFILE";
58 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
59 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
61 #define PERF_MAGIC __perf_magic2
63 const char perf_version_string[] = PERF_VERSION;
65 struct perf_file_attr {
66 struct perf_event_attr attr;
67 struct perf_file_section ids;
71 struct perf_header *ph;
73 void *buf; /* Either buf != NULL or fd >= 0 */
76 struct perf_evsel *events;
79 void perf_header__set_feat(struct perf_header *header, int feat)
81 set_bit(feat, header->adds_features);
84 void perf_header__clear_feat(struct perf_header *header, int feat)
86 clear_bit(feat, header->adds_features);
89 bool perf_header__has_feat(const struct perf_header *header, int feat)
91 return test_bit(feat, header->adds_features);
94 static int __do_write_fd(struct feat_fd *ff, const void *buf, size_t size)
96 ssize_t ret = writen(ff->fd, buf, size);
98 if (ret != (ssize_t)size)
99 return ret < 0 ? (int)ret : -1;
103 static int __do_write_buf(struct feat_fd *ff, const void *buf, size_t size)
105 /* struct perf_event_header::size is u16 */
106 const size_t max_size = 0xffff - sizeof(struct perf_event_header);
107 size_t new_size = ff->size;
110 if (size + ff->offset > max_size)
113 while (size > (new_size - ff->offset))
115 new_size = min(max_size, new_size);
117 if (ff->size < new_size) {
118 addr = realloc(ff->buf, new_size);
125 memcpy(ff->buf + ff->offset, buf, size);
131 /* Return: 0 if succeded, -ERR if failed. */
132 int do_write(struct feat_fd *ff, const void *buf, size_t size)
135 return __do_write_fd(ff, buf, size);
136 return __do_write_buf(ff, buf, size);
139 /* Return: 0 if succeded, -ERR if failed. */
140 static int do_write_bitmap(struct feat_fd *ff, unsigned long *set, u64 size)
142 u64 *p = (u64 *) set;
145 ret = do_write(ff, &size, sizeof(size));
149 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
150 ret = do_write(ff, p + i, sizeof(*p));
158 /* Return: 0 if succeded, -ERR if failed. */
159 int write_padded(struct feat_fd *ff, const void *bf,
160 size_t count, size_t count_aligned)
162 static const char zero_buf[NAME_ALIGN];
163 int err = do_write(ff, bf, count);
166 err = do_write(ff, zero_buf, count_aligned - count);
171 #define string_size(str) \
172 (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
174 /* Return: 0 if succeded, -ERR if failed. */
175 static int do_write_string(struct feat_fd *ff, const char *str)
180 olen = strlen(str) + 1;
181 len = PERF_ALIGN(olen, NAME_ALIGN);
183 /* write len, incl. \0 */
184 ret = do_write(ff, &len, sizeof(len));
188 return write_padded(ff, str, olen, len);
191 static int __do_read_fd(struct feat_fd *ff, void *addr, ssize_t size)
193 ssize_t ret = readn(ff->fd, addr, size);
196 return ret < 0 ? (int)ret : -1;
200 static int __do_read_buf(struct feat_fd *ff, void *addr, ssize_t size)
202 if (size > (ssize_t)ff->size - ff->offset)
205 memcpy(addr, ff->buf + ff->offset, size);
212 static int __do_read(struct feat_fd *ff, void *addr, ssize_t size)
215 return __do_read_fd(ff, addr, size);
216 return __do_read_buf(ff, addr, size);
219 static int do_read_u32(struct feat_fd *ff, u32 *addr)
223 ret = __do_read(ff, addr, sizeof(*addr));
227 if (ff->ph->needs_swap)
228 *addr = bswap_32(*addr);
232 static int do_read_u64(struct feat_fd *ff, u64 *addr)
236 ret = __do_read(ff, addr, sizeof(*addr));
240 if (ff->ph->needs_swap)
241 *addr = bswap_64(*addr);
245 static char *do_read_string(struct feat_fd *ff)
250 if (do_read_u32(ff, &len))
257 if (!__do_read(ff, buf, len)) {
259 * strings are padded by zeroes
260 * thus the actual strlen of buf
261 * may be less than len
270 /* Return: 0 if succeded, -ERR if failed. */
271 static int do_read_bitmap(struct feat_fd *ff, unsigned long **pset, u64 *psize)
277 ret = do_read_u64(ff, &size);
281 set = bitmap_alloc(size);
287 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
288 ret = do_read_u64(ff, p + i);
300 static int write_tracing_data(struct feat_fd *ff,
301 struct perf_evlist *evlist)
303 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
306 return read_tracing_data(ff->fd, &evlist->entries);
309 static int write_build_id(struct feat_fd *ff,
310 struct perf_evlist *evlist __maybe_unused)
312 struct perf_session *session;
315 session = container_of(ff->ph, struct perf_session, header);
317 if (!perf_session__read_build_ids(session, true))
320 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
323 err = perf_session__write_buildid_table(session, ff);
325 pr_debug("failed to write buildid table\n");
328 perf_session__cache_build_ids(session);
333 static int write_hostname(struct feat_fd *ff,
334 struct perf_evlist *evlist __maybe_unused)
343 return do_write_string(ff, uts.nodename);
346 static int write_osrelease(struct feat_fd *ff,
347 struct perf_evlist *evlist __maybe_unused)
356 return do_write_string(ff, uts.release);
359 static int write_arch(struct feat_fd *ff,
360 struct perf_evlist *evlist __maybe_unused)
369 return do_write_string(ff, uts.machine);
372 static int write_version(struct feat_fd *ff,
373 struct perf_evlist *evlist __maybe_unused)
375 return do_write_string(ff, perf_version_string);
378 static int __write_cpudesc(struct feat_fd *ff, const char *cpuinfo_proc)
383 const char *search = cpuinfo_proc;
390 file = fopen("/proc/cpuinfo", "r");
394 while (getline(&buf, &len, file) > 0) {
395 ret = strncmp(buf, search, strlen(search));
407 p = strchr(buf, ':');
408 if (p && *(p+1) == ' ' && *(p+2))
414 /* squash extra space characters (branding string) */
421 while (*q && isspace(*q))
424 while ((*r++ = *q++));
428 ret = do_write_string(ff, s);
435 static int write_cpudesc(struct feat_fd *ff,
436 struct perf_evlist *evlist __maybe_unused)
438 const char *cpuinfo_procs[] = CPUINFO_PROC;
441 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
443 ret = __write_cpudesc(ff, cpuinfo_procs[i]);
451 static int write_nrcpus(struct feat_fd *ff,
452 struct perf_evlist *evlist __maybe_unused)
458 nrc = cpu__max_present_cpu();
460 nr = sysconf(_SC_NPROCESSORS_ONLN);
464 nra = (u32)(nr & UINT_MAX);
466 ret = do_write(ff, &nrc, sizeof(nrc));
470 return do_write(ff, &nra, sizeof(nra));
473 static int write_event_desc(struct feat_fd *ff,
474 struct perf_evlist *evlist)
476 struct perf_evsel *evsel;
480 nre = evlist->nr_entries;
483 * write number of events
485 ret = do_write(ff, &nre, sizeof(nre));
490 * size of perf_event_attr struct
492 sz = (u32)sizeof(evsel->attr);
493 ret = do_write(ff, &sz, sizeof(sz));
497 evlist__for_each_entry(evlist, evsel) {
498 ret = do_write(ff, &evsel->attr, sz);
502 * write number of unique id per event
503 * there is one id per instance of an event
505 * copy into an nri to be independent of the
509 ret = do_write(ff, &nri, sizeof(nri));
514 * write event string as passed on cmdline
516 ret = do_write_string(ff, perf_evsel__name(evsel));
520 * write unique ids for this event
522 ret = do_write(ff, evsel->id, evsel->ids * sizeof(u64));
529 static int write_cmdline(struct feat_fd *ff,
530 struct perf_evlist *evlist __maybe_unused)
532 char pbuf[MAXPATHLEN], *buf;
535 /* actual path to perf binary */
536 buf = perf_exe(pbuf, MAXPATHLEN);
538 /* account for binary path */
539 n = perf_env.nr_cmdline + 1;
541 ret = do_write(ff, &n, sizeof(n));
545 ret = do_write_string(ff, buf);
549 for (i = 0 ; i < perf_env.nr_cmdline; i++) {
550 ret = do_write_string(ff, perf_env.cmdline_argv[i]);
558 static int write_cpu_topology(struct feat_fd *ff,
559 struct perf_evlist *evlist __maybe_unused)
561 struct cpu_topology *tp;
565 tp = cpu_topology__new();
569 ret = do_write(ff, &tp->core_sib, sizeof(tp->core_sib));
573 for (i = 0; i < tp->core_sib; i++) {
574 ret = do_write_string(ff, tp->core_siblings[i]);
578 ret = do_write(ff, &tp->thread_sib, sizeof(tp->thread_sib));
582 for (i = 0; i < tp->thread_sib; i++) {
583 ret = do_write_string(ff, tp->thread_siblings[i]);
588 ret = perf_env__read_cpu_topology_map(&perf_env);
592 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
593 ret = do_write(ff, &perf_env.cpu[j].core_id,
594 sizeof(perf_env.cpu[j].core_id));
597 ret = do_write(ff, &perf_env.cpu[j].socket_id,
598 sizeof(perf_env.cpu[j].socket_id));
603 cpu_topology__delete(tp);
609 static int write_total_mem(struct feat_fd *ff,
610 struct perf_evlist *evlist __maybe_unused)
618 fp = fopen("/proc/meminfo", "r");
622 while (getline(&buf, &len, fp) > 0) {
623 ret = strncmp(buf, "MemTotal:", 9);
628 n = sscanf(buf, "%*s %"PRIu64, &mem);
630 ret = do_write(ff, &mem, sizeof(mem));
638 static int write_numa_topology(struct feat_fd *ff,
639 struct perf_evlist *evlist __maybe_unused)
641 struct numa_topology *tp;
645 tp = numa_topology__new();
649 ret = do_write(ff, &tp->nr, sizeof(u32));
653 for (i = 0; i < tp->nr; i++) {
654 struct numa_topology_node *n = &tp->nodes[i];
656 ret = do_write(ff, &n->node, sizeof(u32));
660 ret = do_write(ff, &n->mem_total, sizeof(u64));
664 ret = do_write(ff, &n->mem_free, sizeof(u64));
668 ret = do_write_string(ff, n->cpus);
676 numa_topology__delete(tp);
683 * struct pmu_mappings {
692 static int write_pmu_mappings(struct feat_fd *ff,
693 struct perf_evlist *evlist __maybe_unused)
695 struct perf_pmu *pmu = NULL;
700 * Do a first pass to count number of pmu to avoid lseek so this
701 * works in pipe mode as well.
703 while ((pmu = perf_pmu__scan(pmu))) {
709 ret = do_write(ff, &pmu_num, sizeof(pmu_num));
713 while ((pmu = perf_pmu__scan(pmu))) {
717 ret = do_write(ff, &pmu->type, sizeof(pmu->type));
721 ret = do_write_string(ff, pmu->name);
732 * struct group_descs {
734 * struct group_desc {
741 static int write_group_desc(struct feat_fd *ff,
742 struct perf_evlist *evlist)
744 u32 nr_groups = evlist->nr_groups;
745 struct perf_evsel *evsel;
748 ret = do_write(ff, &nr_groups, sizeof(nr_groups));
752 evlist__for_each_entry(evlist, evsel) {
753 if (perf_evsel__is_group_leader(evsel) &&
754 evsel->nr_members > 1) {
755 const char *name = evsel->group_name ?: "{anon_group}";
756 u32 leader_idx = evsel->idx;
757 u32 nr_members = evsel->nr_members;
759 ret = do_write_string(ff, name);
763 ret = do_write(ff, &leader_idx, sizeof(leader_idx));
767 ret = do_write(ff, &nr_members, sizeof(nr_members));
776 * Return the CPU id as a raw string.
778 * Each architecture should provide a more precise id string that
779 * can be use to match the architecture's "mapfile".
781 char * __weak get_cpuid_str(struct perf_pmu *pmu __maybe_unused)
786 /* Return zero when the cpuid from the mapfile.csv matches the
787 * cpuid string generated on this platform.
788 * Otherwise return non-zero.
790 int __weak strcmp_cpuid_str(const char *mapcpuid, const char *cpuid)
793 regmatch_t pmatch[1];
796 if (regcomp(&re, mapcpuid, REG_EXTENDED) != 0) {
797 /* Warn unable to generate match particular string. */
798 pr_info("Invalid regular expression %s\n", mapcpuid);
802 match = !regexec(&re, cpuid, 1, pmatch, 0);
805 size_t match_len = (pmatch[0].rm_eo - pmatch[0].rm_so);
807 /* Verify the entire string matched. */
808 if (match_len == strlen(cpuid))
815 * default get_cpuid(): nothing gets recorded
816 * actual implementation must be in arch/$(SRCARCH)/util/header.c
818 int __weak get_cpuid(char *buffer __maybe_unused, size_t sz __maybe_unused)
823 static int write_cpuid(struct feat_fd *ff,
824 struct perf_evlist *evlist __maybe_unused)
829 ret = get_cpuid(buffer, sizeof(buffer));
833 return do_write_string(ff, buffer);
836 static int write_branch_stack(struct feat_fd *ff __maybe_unused,
837 struct perf_evlist *evlist __maybe_unused)
842 static int write_auxtrace(struct feat_fd *ff,
843 struct perf_evlist *evlist __maybe_unused)
845 struct perf_session *session;
848 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
851 session = container_of(ff->ph, struct perf_session, header);
853 err = auxtrace_index__write(ff->fd, &session->auxtrace_index);
855 pr_err("Failed to write auxtrace index\n");
859 static int write_clockid(struct feat_fd *ff,
860 struct perf_evlist *evlist __maybe_unused)
862 return do_write(ff, &ff->ph->env.clockid_res_ns,
863 sizeof(ff->ph->env.clockid_res_ns));
866 static int write_dir_format(struct feat_fd *ff,
867 struct perf_evlist *evlist __maybe_unused)
869 struct perf_session *session;
870 struct perf_data *data;
872 session = container_of(ff->ph, struct perf_session, header);
873 data = session->data;
875 if (WARN_ON(!perf_data__is_dir(data)))
878 return do_write(ff, &data->dir.version, sizeof(data->dir.version));
881 #ifdef HAVE_LIBBPF_SUPPORT
882 static int write_bpf_prog_info(struct feat_fd *ff,
883 struct perf_evlist *evlist __maybe_unused)
885 struct perf_env *env = &ff->ph->env;
886 struct rb_root *root;
887 struct rb_node *next;
890 down_read(&env->bpf_progs.lock);
892 ret = do_write(ff, &env->bpf_progs.infos_cnt,
893 sizeof(env->bpf_progs.infos_cnt));
897 root = &env->bpf_progs.infos;
898 next = rb_first(root);
900 struct bpf_prog_info_node *node;
903 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
904 next = rb_next(&node->rb_node);
905 len = sizeof(struct bpf_prog_info_linear) +
906 node->info_linear->data_len;
908 /* before writing to file, translate address to offset */
909 bpf_program__bpil_addr_to_offs(node->info_linear);
910 ret = do_write(ff, node->info_linear, len);
912 * translate back to address even when do_write() fails,
913 * so that this function never changes the data.
915 bpf_program__bpil_offs_to_addr(node->info_linear);
920 up_read(&env->bpf_progs.lock);
923 #else // HAVE_LIBBPF_SUPPORT
924 static int write_bpf_prog_info(struct feat_fd *ff __maybe_unused,
925 struct perf_evlist *evlist __maybe_unused)
929 #endif // HAVE_LIBBPF_SUPPORT
931 static int write_bpf_btf(struct feat_fd *ff,
932 struct perf_evlist *evlist __maybe_unused)
934 struct perf_env *env = &ff->ph->env;
935 struct rb_root *root;
936 struct rb_node *next;
939 down_read(&env->bpf_progs.lock);
941 ret = do_write(ff, &env->bpf_progs.btfs_cnt,
942 sizeof(env->bpf_progs.btfs_cnt));
947 root = &env->bpf_progs.btfs;
948 next = rb_first(root);
950 struct btf_node *node;
952 node = rb_entry(next, struct btf_node, rb_node);
953 next = rb_next(&node->rb_node);
954 ret = do_write(ff, &node->id,
955 sizeof(u32) * 2 + node->data_size);
960 up_read(&env->bpf_progs.lock);
964 static int cpu_cache_level__sort(const void *a, const void *b)
966 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
967 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
969 return cache_a->level - cache_b->level;
972 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
974 if (a->level != b->level)
977 if (a->line_size != b->line_size)
980 if (a->sets != b->sets)
983 if (a->ways != b->ways)
986 if (strcmp(a->type, b->type))
989 if (strcmp(a->size, b->size))
992 if (strcmp(a->map, b->map))
998 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
1000 char path[PATH_MAX], file[PATH_MAX];
1004 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
1005 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
1007 if (stat(file, &st))
1010 scnprintf(file, PATH_MAX, "%s/level", path);
1011 if (sysfs__read_int(file, (int *) &cache->level))
1014 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
1015 if (sysfs__read_int(file, (int *) &cache->line_size))
1018 scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
1019 if (sysfs__read_int(file, (int *) &cache->sets))
1022 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
1023 if (sysfs__read_int(file, (int *) &cache->ways))
1026 scnprintf(file, PATH_MAX, "%s/type", path);
1027 if (sysfs__read_str(file, &cache->type, &len))
1030 cache->type[len] = 0;
1031 cache->type = rtrim(cache->type);
1033 scnprintf(file, PATH_MAX, "%s/size", path);
1034 if (sysfs__read_str(file, &cache->size, &len)) {
1039 cache->size[len] = 0;
1040 cache->size = rtrim(cache->size);
1042 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
1043 if (sysfs__read_str(file, &cache->map, &len)) {
1049 cache->map[len] = 0;
1050 cache->map = rtrim(cache->map);
1054 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
1056 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
1059 static int build_caches(struct cpu_cache_level caches[], u32 size, u32 *cntp)
1066 ncpus = sysconf(_SC_NPROCESSORS_CONF);
1070 nr = (u32)(ncpus & UINT_MAX);
1072 for (cpu = 0; cpu < nr; cpu++) {
1073 for (level = 0; level < 10; level++) {
1074 struct cpu_cache_level c;
1077 err = cpu_cache_level__read(&c, cpu, level);
1084 for (i = 0; i < cnt; i++) {
1085 if (cpu_cache_level__cmp(&c, &caches[i]))
1092 cpu_cache_level__free(&c);
1094 if (WARN_ONCE(cnt == size, "way too many cpu caches.."))
1103 #define MAX_CACHES 2000
1105 static int write_cache(struct feat_fd *ff,
1106 struct perf_evlist *evlist __maybe_unused)
1108 struct cpu_cache_level caches[MAX_CACHES];
1109 u32 cnt = 0, i, version = 1;
1112 ret = build_caches(caches, MAX_CACHES, &cnt);
1116 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1118 ret = do_write(ff, &version, sizeof(u32));
1122 ret = do_write(ff, &cnt, sizeof(u32));
1126 for (i = 0; i < cnt; i++) {
1127 struct cpu_cache_level *c = &caches[i];
1130 ret = do_write(ff, &c->v, sizeof(u32)); \
1141 ret = do_write_string(ff, (const char *) c->v); \
1152 for (i = 0; i < cnt; i++)
1153 cpu_cache_level__free(&caches[i]);
1157 static int write_stat(struct feat_fd *ff __maybe_unused,
1158 struct perf_evlist *evlist __maybe_unused)
1163 static int write_sample_time(struct feat_fd *ff,
1164 struct perf_evlist *evlist)
1168 ret = do_write(ff, &evlist->first_sample_time,
1169 sizeof(evlist->first_sample_time));
1173 return do_write(ff, &evlist->last_sample_time,
1174 sizeof(evlist->last_sample_time));
1178 static int memory_node__read(struct memory_node *n, unsigned long idx)
1180 unsigned int phys, size = 0;
1181 char path[PATH_MAX];
1185 #define for_each_memory(mem, dir) \
1186 while ((ent = readdir(dir))) \
1187 if (strcmp(ent->d_name, ".") && \
1188 strcmp(ent->d_name, "..") && \
1189 sscanf(ent->d_name, "memory%u", &mem) == 1)
1191 scnprintf(path, PATH_MAX,
1192 "%s/devices/system/node/node%lu",
1193 sysfs__mountpoint(), idx);
1195 dir = opendir(path);
1197 pr_warning("failed: cant' open memory sysfs data\n");
1201 for_each_memory(phys, dir) {
1202 size = max(phys, size);
1207 n->set = bitmap_alloc(size);
1218 for_each_memory(phys, dir) {
1219 set_bit(phys, n->set);
1226 static int memory_node__sort(const void *a, const void *b)
1228 const struct memory_node *na = a;
1229 const struct memory_node *nb = b;
1231 return na->node - nb->node;
1234 static int build_mem_topology(struct memory_node *nodes, u64 size, u64 *cntp)
1236 char path[PATH_MAX];
1242 scnprintf(path, PATH_MAX, "%s/devices/system/node/",
1243 sysfs__mountpoint());
1245 dir = opendir(path);
1247 pr_debug2("%s: could't read %s, does this arch have topology information?\n",
1252 while (!ret && (ent = readdir(dir))) {
1256 if (!strcmp(ent->d_name, ".") ||
1257 !strcmp(ent->d_name, ".."))
1260 r = sscanf(ent->d_name, "node%u", &idx);
1264 if (WARN_ONCE(cnt >= size,
1265 "failed to write MEM_TOPOLOGY, way too many nodes\n"))
1268 ret = memory_node__read(&nodes[cnt++], idx);
1275 qsort(nodes, cnt, sizeof(nodes[0]), memory_node__sort);
1280 #define MAX_MEMORY_NODES 2000
1283 * The MEM_TOPOLOGY holds physical memory map for every
1284 * node in system. The format of data is as follows:
1286 * 0 - version | for future changes
1287 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes
1288 * 16 - count | number of nodes
1290 * For each node we store map of physical indexes for
1293 * 32 - node id | node index
1294 * 40 - size | size of bitmap
1295 * 48 - bitmap | bitmap of memory indexes that belongs to node
1297 static int write_mem_topology(struct feat_fd *ff __maybe_unused,
1298 struct perf_evlist *evlist __maybe_unused)
1300 static struct memory_node nodes[MAX_MEMORY_NODES];
1301 u64 bsize, version = 1, i, nr;
1304 ret = sysfs__read_xll("devices/system/memory/block_size_bytes",
1305 (unsigned long long *) &bsize);
1309 ret = build_mem_topology(&nodes[0], MAX_MEMORY_NODES, &nr);
1313 ret = do_write(ff, &version, sizeof(version));
1317 ret = do_write(ff, &bsize, sizeof(bsize));
1321 ret = do_write(ff, &nr, sizeof(nr));
1325 for (i = 0; i < nr; i++) {
1326 struct memory_node *n = &nodes[i];
1329 ret = do_write(ff, &n->v, sizeof(n->v)); \
1338 ret = do_write_bitmap(ff, n->set, n->size);
1347 static int write_compressed(struct feat_fd *ff __maybe_unused,
1348 struct perf_evlist *evlist __maybe_unused)
1352 ret = do_write(ff, &(ff->ph->env.comp_ver), sizeof(ff->ph->env.comp_ver));
1356 ret = do_write(ff, &(ff->ph->env.comp_type), sizeof(ff->ph->env.comp_type));
1360 ret = do_write(ff, &(ff->ph->env.comp_level), sizeof(ff->ph->env.comp_level));
1364 ret = do_write(ff, &(ff->ph->env.comp_ratio), sizeof(ff->ph->env.comp_ratio));
1368 return do_write(ff, &(ff->ph->env.comp_mmap_len), sizeof(ff->ph->env.comp_mmap_len));
1371 static void print_hostname(struct feat_fd *ff, FILE *fp)
1373 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1376 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1378 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1381 static void print_arch(struct feat_fd *ff, FILE *fp)
1383 fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1386 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1388 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1391 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1393 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1394 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1397 static void print_version(struct feat_fd *ff, FILE *fp)
1399 fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1402 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1406 nr = ff->ph->env.nr_cmdline;
1408 fprintf(fp, "# cmdline : ");
1410 for (i = 0; i < nr; i++) {
1411 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]);
1413 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1417 char *quote = strchr(argv_i, '\'');
1421 fprintf(fp, "%s\\\'", argv_i);
1424 fprintf(fp, "%s ", argv_i);
1431 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1433 struct perf_header *ph = ff->ph;
1434 int cpu_nr = ph->env.nr_cpus_avail;
1438 nr = ph->env.nr_sibling_cores;
1439 str = ph->env.sibling_cores;
1441 for (i = 0; i < nr; i++) {
1442 fprintf(fp, "# sibling cores : %s\n", str);
1443 str += strlen(str) + 1;
1446 nr = ph->env.nr_sibling_threads;
1447 str = ph->env.sibling_threads;
1449 for (i = 0; i < nr; i++) {
1450 fprintf(fp, "# sibling threads : %s\n", str);
1451 str += strlen(str) + 1;
1454 if (ph->env.cpu != NULL) {
1455 for (i = 0; i < cpu_nr; i++)
1456 fprintf(fp, "# CPU %d: Core ID %d, Socket ID %d\n", i,
1457 ph->env.cpu[i].core_id, ph->env.cpu[i].socket_id);
1459 fprintf(fp, "# Core ID and Socket ID information is not available\n");
1462 static void print_clockid(struct feat_fd *ff, FILE *fp)
1464 fprintf(fp, "# clockid frequency: %"PRIu64" MHz\n",
1465 ff->ph->env.clockid_res_ns * 1000);
1468 static void print_dir_format(struct feat_fd *ff, FILE *fp)
1470 struct perf_session *session;
1471 struct perf_data *data;
1473 session = container_of(ff->ph, struct perf_session, header);
1474 data = session->data;
1476 fprintf(fp, "# directory data version : %"PRIu64"\n", data->dir.version);
1479 static void print_bpf_prog_info(struct feat_fd *ff, FILE *fp)
1481 struct perf_env *env = &ff->ph->env;
1482 struct rb_root *root;
1483 struct rb_node *next;
1485 down_read(&env->bpf_progs.lock);
1487 root = &env->bpf_progs.infos;
1488 next = rb_first(root);
1491 struct bpf_prog_info_node *node;
1493 node = rb_entry(next, struct bpf_prog_info_node, rb_node);
1494 next = rb_next(&node->rb_node);
1496 bpf_event__print_bpf_prog_info(&node->info_linear->info,
1500 up_read(&env->bpf_progs.lock);
1503 static void print_bpf_btf(struct feat_fd *ff, FILE *fp)
1505 struct perf_env *env = &ff->ph->env;
1506 struct rb_root *root;
1507 struct rb_node *next;
1509 down_read(&env->bpf_progs.lock);
1511 root = &env->bpf_progs.btfs;
1512 next = rb_first(root);
1515 struct btf_node *node;
1517 node = rb_entry(next, struct btf_node, rb_node);
1518 next = rb_next(&node->rb_node);
1519 fprintf(fp, "# btf info of id %u\n", node->id);
1522 up_read(&env->bpf_progs.lock);
1525 static void free_event_desc(struct perf_evsel *events)
1527 struct perf_evsel *evsel;
1532 for (evsel = events; evsel->attr.size; evsel++) {
1533 zfree(&evsel->name);
1540 static struct perf_evsel *read_event_desc(struct feat_fd *ff)
1542 struct perf_evsel *evsel, *events = NULL;
1545 u32 nre, sz, nr, i, j;
1548 /* number of events */
1549 if (do_read_u32(ff, &nre))
1552 if (do_read_u32(ff, &sz))
1555 /* buffer to hold on file attr struct */
1560 /* the last event terminates with evsel->attr.size == 0: */
1561 events = calloc(nre + 1, sizeof(*events));
1565 msz = sizeof(evsel->attr);
1569 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1573 * must read entire on-file attr struct to
1574 * sync up with layout.
1576 if (__do_read(ff, buf, sz))
1579 if (ff->ph->needs_swap)
1580 perf_event__attr_swap(buf);
1582 memcpy(&evsel->attr, buf, msz);
1584 if (do_read_u32(ff, &nr))
1587 if (ff->ph->needs_swap)
1588 evsel->needs_swap = true;
1590 evsel->name = do_read_string(ff);
1597 id = calloc(nr, sizeof(*id));
1603 for (j = 0 ; j < nr; j++) {
1604 if (do_read_u64(ff, id))
1613 free_event_desc(events);
1618 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1619 void *priv __maybe_unused)
1621 return fprintf(fp, ", %s = %s", name, val);
1624 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1626 struct perf_evsel *evsel, *events;
1631 events = ff->events;
1633 events = read_event_desc(ff);
1636 fprintf(fp, "# event desc: not available or unable to read\n");
1640 for (evsel = events; evsel->attr.size; evsel++) {
1641 fprintf(fp, "# event : name = %s, ", evsel->name);
1644 fprintf(fp, ", id = {");
1645 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1648 fprintf(fp, " %"PRIu64, *id);
1653 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1658 free_event_desc(events);
1662 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1664 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1667 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1670 struct numa_node *n;
1672 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1673 n = &ff->ph->env.numa_nodes[i];
1675 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1676 " free = %"PRIu64" kB\n",
1677 n->node, n->mem_total, n->mem_free);
1679 fprintf(fp, "# node%u cpu list : ", n->node);
1680 cpu_map__fprintf(n->map, fp);
1684 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1686 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1689 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1691 fprintf(fp, "# contains samples with branch stack\n");
1694 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1696 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1699 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1701 fprintf(fp, "# contains stat data\n");
1704 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1708 fprintf(fp, "# CPU cache info:\n");
1709 for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1711 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1715 static void print_compressed(struct feat_fd *ff, FILE *fp)
1717 fprintf(fp, "# compressed : %s, level = %d, ratio = %d\n",
1718 ff->ph->env.comp_type == PERF_COMP_ZSTD ? "Zstd" : "Unknown",
1719 ff->ph->env.comp_level, ff->ph->env.comp_ratio);
1722 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1724 const char *delimiter = "# pmu mappings: ";
1729 pmu_num = ff->ph->env.nr_pmu_mappings;
1731 fprintf(fp, "# pmu mappings: not available\n");
1735 str = ff->ph->env.pmu_mappings;
1738 type = strtoul(str, &tmp, 0);
1743 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1746 str += strlen(str) + 1;
1755 fprintf(fp, "# pmu mappings: unable to read\n");
1758 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1760 struct perf_session *session;
1761 struct perf_evsel *evsel;
1764 session = container_of(ff->ph, struct perf_session, header);
1766 evlist__for_each_entry(session->evlist, evsel) {
1767 if (perf_evsel__is_group_leader(evsel) &&
1768 evsel->nr_members > 1) {
1769 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1770 perf_evsel__name(evsel));
1772 nr = evsel->nr_members - 1;
1774 fprintf(fp, ",%s", perf_evsel__name(evsel));
1782 static void print_sample_time(struct feat_fd *ff, FILE *fp)
1784 struct perf_session *session;
1788 session = container_of(ff->ph, struct perf_session, header);
1790 timestamp__scnprintf_usec(session->evlist->first_sample_time,
1791 time_buf, sizeof(time_buf));
1792 fprintf(fp, "# time of first sample : %s\n", time_buf);
1794 timestamp__scnprintf_usec(session->evlist->last_sample_time,
1795 time_buf, sizeof(time_buf));
1796 fprintf(fp, "# time of last sample : %s\n", time_buf);
1798 d = (double)(session->evlist->last_sample_time -
1799 session->evlist->first_sample_time) / NSEC_PER_MSEC;
1801 fprintf(fp, "# sample duration : %10.3f ms\n", d);
1804 static void memory_node__fprintf(struct memory_node *n,
1805 unsigned long long bsize, FILE *fp)
1807 char buf_map[100], buf_size[50];
1808 unsigned long long size;
1810 size = bsize * bitmap_weight(n->set, n->size);
1811 unit_number__scnprintf(buf_size, 50, size);
1813 bitmap_scnprintf(n->set, n->size, buf_map, 100);
1814 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map);
1817 static void print_mem_topology(struct feat_fd *ff, FILE *fp)
1819 struct memory_node *nodes;
1822 nodes = ff->ph->env.memory_nodes;
1823 nr = ff->ph->env.nr_memory_nodes;
1825 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n",
1826 nr, ff->ph->env.memory_bsize);
1828 for (i = 0; i < nr; i++) {
1829 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp);
1833 static int __event_process_build_id(struct build_id_event *bev,
1835 struct perf_session *session)
1838 struct machine *machine;
1841 enum dso_kernel_type dso_type;
1843 machine = perf_session__findnew_machine(session, bev->pid);
1847 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1850 case PERF_RECORD_MISC_KERNEL:
1851 dso_type = DSO_TYPE_KERNEL;
1853 case PERF_RECORD_MISC_GUEST_KERNEL:
1854 dso_type = DSO_TYPE_GUEST_KERNEL;
1856 case PERF_RECORD_MISC_USER:
1857 case PERF_RECORD_MISC_GUEST_USER:
1858 dso_type = DSO_TYPE_USER;
1864 dso = machine__findnew_dso(machine, filename);
1866 char sbuild_id[SBUILD_ID_SIZE];
1868 dso__set_build_id(dso, &bev->build_id);
1870 if (dso_type != DSO_TYPE_USER) {
1871 struct kmod_path m = { .name = NULL, };
1873 if (!kmod_path__parse_name(&m, filename) && m.kmod)
1874 dso__set_module_info(dso, &m, machine);
1876 dso->kernel = dso_type;
1881 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1883 pr_debug("build id event received for %s: %s\n",
1884 dso->long_name, sbuild_id);
1893 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1894 int input, u64 offset, u64 size)
1896 struct perf_session *session = container_of(header, struct perf_session, header);
1898 struct perf_event_header header;
1899 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1902 struct build_id_event bev;
1903 char filename[PATH_MAX];
1904 u64 limit = offset + size;
1906 while (offset < limit) {
1909 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1912 if (header->needs_swap)
1913 perf_event_header__bswap(&old_bev.header);
1915 len = old_bev.header.size - sizeof(old_bev);
1916 if (readn(input, filename, len) != len)
1919 bev.header = old_bev.header;
1922 * As the pid is the missing value, we need to fill
1923 * it properly. The header.misc value give us nice hint.
1925 bev.pid = HOST_KERNEL_ID;
1926 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1927 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1928 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1930 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1931 __event_process_build_id(&bev, filename, session);
1933 offset += bev.header.size;
1939 static int perf_header__read_build_ids(struct perf_header *header,
1940 int input, u64 offset, u64 size)
1942 struct perf_session *session = container_of(header, struct perf_session, header);
1943 struct build_id_event bev;
1944 char filename[PATH_MAX];
1945 u64 limit = offset + size, orig_offset = offset;
1948 while (offset < limit) {
1951 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1954 if (header->needs_swap)
1955 perf_event_header__bswap(&bev.header);
1957 len = bev.header.size - sizeof(bev);
1958 if (readn(input, filename, len) != len)
1961 * The a1645ce1 changeset:
1963 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1965 * Added a field to struct build_id_event that broke the file
1968 * Since the kernel build-id is the first entry, process the
1969 * table using the old format if the well known
1970 * '[kernel.kallsyms]' string for the kernel build-id has the
1971 * first 4 characters chopped off (where the pid_t sits).
1973 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1974 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1976 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1979 __event_process_build_id(&bev, filename, session);
1981 offset += bev.header.size;
1988 /* Macro for features that simply need to read and store a string. */
1989 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
1990 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
1992 ff->ph->env.__feat_env = do_read_string(ff); \
1993 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
1996 FEAT_PROCESS_STR_FUN(hostname, hostname);
1997 FEAT_PROCESS_STR_FUN(osrelease, os_release);
1998 FEAT_PROCESS_STR_FUN(version, version);
1999 FEAT_PROCESS_STR_FUN(arch, arch);
2000 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
2001 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
2003 static int process_tracing_data(struct feat_fd *ff, void *data)
2005 ssize_t ret = trace_report(ff->fd, data, false);
2007 return ret < 0 ? -1 : 0;
2010 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
2012 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
2013 pr_debug("Failed to read buildids, continuing...\n");
2017 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
2020 u32 nr_cpus_avail, nr_cpus_online;
2022 ret = do_read_u32(ff, &nr_cpus_avail);
2026 ret = do_read_u32(ff, &nr_cpus_online);
2029 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
2030 ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
2034 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
2039 ret = do_read_u64(ff, &total_mem);
2042 ff->ph->env.total_mem = (unsigned long long)total_mem;
2046 static struct perf_evsel *
2047 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
2049 struct perf_evsel *evsel;
2051 evlist__for_each_entry(evlist, evsel) {
2052 if (evsel->idx == idx)
2060 perf_evlist__set_event_name(struct perf_evlist *evlist,
2061 struct perf_evsel *event)
2063 struct perf_evsel *evsel;
2068 evsel = perf_evlist__find_by_index(evlist, event->idx);
2075 evsel->name = strdup(event->name);
2079 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
2081 struct perf_session *session;
2082 struct perf_evsel *evsel, *events = read_event_desc(ff);
2087 session = container_of(ff->ph, struct perf_session, header);
2089 if (session->data->is_pipe) {
2090 /* Save events for reading later by print_event_desc,
2091 * since they can't be read again in pipe mode. */
2092 ff->events = events;
2095 for (evsel = events; evsel->attr.size; evsel++)
2096 perf_evlist__set_event_name(session->evlist, evsel);
2098 if (!session->data->is_pipe)
2099 free_event_desc(events);
2104 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
2106 char *str, *cmdline = NULL, **argv = NULL;
2109 if (do_read_u32(ff, &nr))
2112 ff->ph->env.nr_cmdline = nr;
2114 cmdline = zalloc(ff->size + nr + 1);
2118 argv = zalloc(sizeof(char *) * (nr + 1));
2122 for (i = 0; i < nr; i++) {
2123 str = do_read_string(ff);
2127 argv[i] = cmdline + len;
2128 memcpy(argv[i], str, strlen(str) + 1);
2129 len += strlen(str) + 1;
2132 ff->ph->env.cmdline = cmdline;
2133 ff->ph->env.cmdline_argv = (const char **) argv;
2142 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
2147 int cpu_nr = ff->ph->env.nr_cpus_avail;
2149 struct perf_header *ph = ff->ph;
2150 bool do_core_id_test = true;
2152 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
2156 if (do_read_u32(ff, &nr))
2159 ph->env.nr_sibling_cores = nr;
2160 size += sizeof(u32);
2161 if (strbuf_init(&sb, 128) < 0)
2164 for (i = 0; i < nr; i++) {
2165 str = do_read_string(ff);
2169 /* include a NULL character at the end */
2170 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2172 size += string_size(str);
2175 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
2177 if (do_read_u32(ff, &nr))
2180 ph->env.nr_sibling_threads = nr;
2181 size += sizeof(u32);
2183 for (i = 0; i < nr; i++) {
2184 str = do_read_string(ff);
2188 /* include a NULL character at the end */
2189 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2191 size += string_size(str);
2194 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
2197 * The header may be from old perf,
2198 * which doesn't include core id and socket id information.
2200 if (ff->size <= size) {
2201 zfree(&ph->env.cpu);
2205 /* On s390 the socket_id number is not related to the numbers of cpus.
2206 * The socket_id number might be higher than the numbers of cpus.
2207 * This depends on the configuration.
2209 if (ph->env.arch && !strncmp(ph->env.arch, "s390", 4))
2210 do_core_id_test = false;
2212 for (i = 0; i < (u32)cpu_nr; i++) {
2213 if (do_read_u32(ff, &nr))
2216 ph->env.cpu[i].core_id = nr;
2218 if (do_read_u32(ff, &nr))
2221 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
2222 pr_debug("socket_id number is too big."
2223 "You may need to upgrade the perf tool.\n");
2227 ph->env.cpu[i].socket_id = nr;
2233 strbuf_release(&sb);
2235 zfree(&ph->env.cpu);
2239 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
2241 struct numa_node *nodes, *n;
2246 if (do_read_u32(ff, &nr))
2249 nodes = zalloc(sizeof(*nodes) * nr);
2253 for (i = 0; i < nr; i++) {
2257 if (do_read_u32(ff, &n->node))
2260 if (do_read_u64(ff, &n->mem_total))
2263 if (do_read_u64(ff, &n->mem_free))
2266 str = do_read_string(ff);
2270 n->map = cpu_map__new(str);
2276 ff->ph->env.nr_numa_nodes = nr;
2277 ff->ph->env.numa_nodes = nodes;
2285 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
2292 if (do_read_u32(ff, &pmu_num))
2296 pr_debug("pmu mappings not available\n");
2300 ff->ph->env.nr_pmu_mappings = pmu_num;
2301 if (strbuf_init(&sb, 128) < 0)
2305 if (do_read_u32(ff, &type))
2308 name = do_read_string(ff);
2312 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
2314 /* include a NULL character at the end */
2315 if (strbuf_add(&sb, "", 1) < 0)
2318 if (!strcmp(name, "msr"))
2319 ff->ph->env.msr_pmu_type = type;
2324 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2328 strbuf_release(&sb);
2332 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2335 u32 i, nr, nr_groups;
2336 struct perf_session *session;
2337 struct perf_evsel *evsel, *leader = NULL;
2344 if (do_read_u32(ff, &nr_groups))
2347 ff->ph->env.nr_groups = nr_groups;
2349 pr_debug("group desc not available\n");
2353 desc = calloc(nr_groups, sizeof(*desc));
2357 for (i = 0; i < nr_groups; i++) {
2358 desc[i].name = do_read_string(ff);
2362 if (do_read_u32(ff, &desc[i].leader_idx))
2365 if (do_read_u32(ff, &desc[i].nr_members))
2370 * Rebuild group relationship based on the group_desc
2372 session = container_of(ff->ph, struct perf_session, header);
2373 session->evlist->nr_groups = nr_groups;
2376 evlist__for_each_entry(session->evlist, evsel) {
2377 if (evsel->idx == (int) desc[i].leader_idx) {
2378 evsel->leader = evsel;
2379 /* {anon_group} is a dummy name */
2380 if (strcmp(desc[i].name, "{anon_group}")) {
2381 evsel->group_name = desc[i].name;
2382 desc[i].name = NULL;
2384 evsel->nr_members = desc[i].nr_members;
2386 if (i >= nr_groups || nr > 0) {
2387 pr_debug("invalid group desc\n");
2392 nr = evsel->nr_members - 1;
2395 /* This is a group member */
2396 evsel->leader = leader;
2402 if (i != nr_groups || nr != 0) {
2403 pr_debug("invalid group desc\n");
2409 for (i = 0; i < nr_groups; i++)
2410 zfree(&desc[i].name);
2416 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2418 struct perf_session *session;
2421 session = container_of(ff->ph, struct perf_session, header);
2423 err = auxtrace_index__process(ff->fd, ff->size, session,
2424 ff->ph->needs_swap);
2426 pr_err("Failed to process auxtrace index\n");
2430 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2432 struct cpu_cache_level *caches;
2433 u32 cnt, i, version;
2435 if (do_read_u32(ff, &version))
2441 if (do_read_u32(ff, &cnt))
2444 caches = zalloc(sizeof(*caches) * cnt);
2448 for (i = 0; i < cnt; i++) {
2449 struct cpu_cache_level c;
2452 if (do_read_u32(ff, &c.v))\
2453 goto out_free_caches; \
2462 c.v = do_read_string(ff); \
2464 goto out_free_caches;
2474 ff->ph->env.caches = caches;
2475 ff->ph->env.caches_cnt = cnt;
2482 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused)
2484 struct perf_session *session;
2485 u64 first_sample_time, last_sample_time;
2488 session = container_of(ff->ph, struct perf_session, header);
2490 ret = do_read_u64(ff, &first_sample_time);
2494 ret = do_read_u64(ff, &last_sample_time);
2498 session->evlist->first_sample_time = first_sample_time;
2499 session->evlist->last_sample_time = last_sample_time;
2503 static int process_mem_topology(struct feat_fd *ff,
2504 void *data __maybe_unused)
2506 struct memory_node *nodes;
2507 u64 version, i, nr, bsize;
2510 if (do_read_u64(ff, &version))
2516 if (do_read_u64(ff, &bsize))
2519 if (do_read_u64(ff, &nr))
2522 nodes = zalloc(sizeof(*nodes) * nr);
2526 for (i = 0; i < nr; i++) {
2527 struct memory_node n;
2530 if (do_read_u64(ff, &n.v)) \
2538 if (do_read_bitmap(ff, &n.set, &n.size))
2544 ff->ph->env.memory_bsize = bsize;
2545 ff->ph->env.memory_nodes = nodes;
2546 ff->ph->env.nr_memory_nodes = nr;
2555 static int process_clockid(struct feat_fd *ff,
2556 void *data __maybe_unused)
2558 if (do_read_u64(ff, &ff->ph->env.clockid_res_ns))
2564 static int process_dir_format(struct feat_fd *ff,
2565 void *_data __maybe_unused)
2567 struct perf_session *session;
2568 struct perf_data *data;
2570 session = container_of(ff->ph, struct perf_session, header);
2571 data = session->data;
2573 if (WARN_ON(!perf_data__is_dir(data)))
2576 return do_read_u64(ff, &data->dir.version);
2579 #ifdef HAVE_LIBBPF_SUPPORT
2580 static int process_bpf_prog_info(struct feat_fd *ff, void *data __maybe_unused)
2582 struct bpf_prog_info_linear *info_linear;
2583 struct bpf_prog_info_node *info_node;
2584 struct perf_env *env = &ff->ph->env;
2588 if (ff->ph->needs_swap) {
2589 pr_warning("interpreting bpf_prog_info from systems with endianity is not yet supported\n");
2593 if (do_read_u32(ff, &count))
2596 down_write(&env->bpf_progs.lock);
2598 for (i = 0; i < count; ++i) {
2599 u32 info_len, data_len;
2603 if (do_read_u32(ff, &info_len))
2605 if (do_read_u32(ff, &data_len))
2608 if (info_len > sizeof(struct bpf_prog_info)) {
2609 pr_warning("detected invalid bpf_prog_info\n");
2613 info_linear = malloc(sizeof(struct bpf_prog_info_linear) +
2617 info_linear->info_len = sizeof(struct bpf_prog_info);
2618 info_linear->data_len = data_len;
2619 if (do_read_u64(ff, (u64 *)(&info_linear->arrays)))
2621 if (__do_read(ff, &info_linear->info, info_len))
2623 if (info_len < sizeof(struct bpf_prog_info))
2624 memset(((void *)(&info_linear->info)) + info_len, 0,
2625 sizeof(struct bpf_prog_info) - info_len);
2627 if (__do_read(ff, info_linear->data, data_len))
2630 info_node = malloc(sizeof(struct bpf_prog_info_node));
2634 /* after reading from file, translate offset to address */
2635 bpf_program__bpil_offs_to_addr(info_linear);
2636 info_node->info_linear = info_linear;
2637 perf_env__insert_bpf_prog_info(env, info_node);
2640 up_write(&env->bpf_progs.lock);
2645 up_write(&env->bpf_progs.lock);
2648 #else // HAVE_LIBBPF_SUPPORT
2649 static int process_bpf_prog_info(struct feat_fd *ff __maybe_unused, void *data __maybe_unused)
2653 #endif // HAVE_LIBBPF_SUPPORT
2655 static int process_bpf_btf(struct feat_fd *ff, void *data __maybe_unused)
2657 struct perf_env *env = &ff->ph->env;
2658 struct btf_node *node = NULL;
2662 if (ff->ph->needs_swap) {
2663 pr_warning("interpreting btf from systems with endianity is not yet supported\n");
2667 if (do_read_u32(ff, &count))
2670 down_write(&env->bpf_progs.lock);
2672 for (i = 0; i < count; ++i) {
2675 if (do_read_u32(ff, &id))
2677 if (do_read_u32(ff, &data_size))
2680 node = malloc(sizeof(struct btf_node) + data_size);
2685 node->data_size = data_size;
2687 if (__do_read(ff, node->data, data_size))
2690 perf_env__insert_btf(env, node);
2696 up_write(&env->bpf_progs.lock);
2701 static int process_compressed(struct feat_fd *ff,
2702 void *data __maybe_unused)
2704 if (do_read_u32(ff, &(ff->ph->env.comp_ver)))
2707 if (do_read_u32(ff, &(ff->ph->env.comp_type)))
2710 if (do_read_u32(ff, &(ff->ph->env.comp_level)))
2713 if (do_read_u32(ff, &(ff->ph->env.comp_ratio)))
2716 if (do_read_u32(ff, &(ff->ph->env.comp_mmap_len)))
2722 struct feature_ops {
2723 int (*write)(struct feat_fd *ff, struct perf_evlist *evlist);
2724 void (*print)(struct feat_fd *ff, FILE *fp);
2725 int (*process)(struct feat_fd *ff, void *data);
2731 #define FEAT_OPR(n, func, __full_only) \
2733 .name = __stringify(n), \
2734 .write = write_##func, \
2735 .print = print_##func, \
2736 .full_only = __full_only, \
2737 .process = process_##func, \
2738 .synthesize = true \
2741 #define FEAT_OPN(n, func, __full_only) \
2743 .name = __stringify(n), \
2744 .write = write_##func, \
2745 .print = print_##func, \
2746 .full_only = __full_only, \
2747 .process = process_##func \
2750 /* feature_ops not implemented: */
2751 #define print_tracing_data NULL
2752 #define print_build_id NULL
2754 #define process_branch_stack NULL
2755 #define process_stat NULL
2758 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2759 FEAT_OPN(TRACING_DATA, tracing_data, false),
2760 FEAT_OPN(BUILD_ID, build_id, false),
2761 FEAT_OPR(HOSTNAME, hostname, false),
2762 FEAT_OPR(OSRELEASE, osrelease, false),
2763 FEAT_OPR(VERSION, version, false),
2764 FEAT_OPR(ARCH, arch, false),
2765 FEAT_OPR(NRCPUS, nrcpus, false),
2766 FEAT_OPR(CPUDESC, cpudesc, false),
2767 FEAT_OPR(CPUID, cpuid, false),
2768 FEAT_OPR(TOTAL_MEM, total_mem, false),
2769 FEAT_OPR(EVENT_DESC, event_desc, false),
2770 FEAT_OPR(CMDLINE, cmdline, false),
2771 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true),
2772 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true),
2773 FEAT_OPN(BRANCH_STACK, branch_stack, false),
2774 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false),
2775 FEAT_OPR(GROUP_DESC, group_desc, false),
2776 FEAT_OPN(AUXTRACE, auxtrace, false),
2777 FEAT_OPN(STAT, stat, false),
2778 FEAT_OPN(CACHE, cache, true),
2779 FEAT_OPR(SAMPLE_TIME, sample_time, false),
2780 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true),
2781 FEAT_OPR(CLOCKID, clockid, false),
2782 FEAT_OPN(DIR_FORMAT, dir_format, false),
2783 FEAT_OPR(BPF_PROG_INFO, bpf_prog_info, false),
2784 FEAT_OPR(BPF_BTF, bpf_btf, false),
2785 FEAT_OPR(COMPRESSED, compressed, false),
2788 struct header_print_data {
2790 bool full; /* extended list of headers */
2793 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2794 struct perf_header *ph,
2795 int feat, int fd, void *data)
2797 struct header_print_data *hd = data;
2800 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2801 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2802 "%d, continuing...\n", section->offset, feat);
2805 if (feat >= HEADER_LAST_FEATURE) {
2806 pr_warning("unknown feature %d\n", feat);
2809 if (!feat_ops[feat].print)
2812 ff = (struct feat_fd) {
2817 if (!feat_ops[feat].full_only || hd->full)
2818 feat_ops[feat].print(&ff, hd->fp);
2820 fprintf(hd->fp, "# %s info available, use -I to display\n",
2821 feat_ops[feat].name);
2826 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2828 struct header_print_data hd;
2829 struct perf_header *header = &session->header;
2830 int fd = perf_data__fd(session->data);
2838 ret = fstat(fd, &st);
2842 stctime = st.st_ctime;
2843 fprintf(fp, "# captured on : %s", ctime(&stctime));
2845 fprintf(fp, "# header version : %u\n", header->version);
2846 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset);
2847 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size);
2848 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset);
2850 perf_header__process_sections(header, fd, &hd,
2851 perf_file_section__fprintf_info);
2853 if (session->data->is_pipe)
2856 fprintf(fp, "# missing features: ");
2857 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
2859 fprintf(fp, "%s ", feat_ops[bit].name);
2866 static int do_write_feat(struct feat_fd *ff, int type,
2867 struct perf_file_section **p,
2868 struct perf_evlist *evlist)
2873 if (perf_header__has_feat(ff->ph, type)) {
2874 if (!feat_ops[type].write)
2877 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
2880 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
2882 err = feat_ops[type].write(ff, evlist);
2884 pr_debug("failed to write feature %s\n", feat_ops[type].name);
2886 /* undo anything written */
2887 lseek(ff->fd, (*p)->offset, SEEK_SET);
2891 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
2897 static int perf_header__adds_write(struct perf_header *header,
2898 struct perf_evlist *evlist, int fd)
2902 struct perf_file_section *feat_sec, *p;
2908 ff = (struct feat_fd){
2913 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2917 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2918 if (feat_sec == NULL)
2921 sec_size = sizeof(*feat_sec) * nr_sections;
2923 sec_start = header->feat_offset;
2924 lseek(fd, sec_start + sec_size, SEEK_SET);
2926 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2927 if (do_write_feat(&ff, feat, &p, evlist))
2928 perf_header__clear_feat(header, feat);
2931 lseek(fd, sec_start, SEEK_SET);
2933 * may write more than needed due to dropped feature, but
2934 * this is okay, reader will skip the missing entries
2936 err = do_write(&ff, feat_sec, sec_size);
2938 pr_debug("failed to write feature section\n");
2943 int perf_header__write_pipe(int fd)
2945 struct perf_pipe_file_header f_header;
2949 ff = (struct feat_fd){ .fd = fd };
2951 f_header = (struct perf_pipe_file_header){
2952 .magic = PERF_MAGIC,
2953 .size = sizeof(f_header),
2956 err = do_write(&ff, &f_header, sizeof(f_header));
2958 pr_debug("failed to write perf pipe header\n");
2965 int perf_session__write_header(struct perf_session *session,
2966 struct perf_evlist *evlist,
2967 int fd, bool at_exit)
2969 struct perf_file_header f_header;
2970 struct perf_file_attr f_attr;
2971 struct perf_header *header = &session->header;
2972 struct perf_evsel *evsel;
2977 ff = (struct feat_fd){ .fd = fd};
2978 lseek(fd, sizeof(f_header), SEEK_SET);
2980 evlist__for_each_entry(session->evlist, evsel) {
2981 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2982 err = do_write(&ff, evsel->id, evsel->ids * sizeof(u64));
2984 pr_debug("failed to write perf header\n");
2989 attr_offset = lseek(ff.fd, 0, SEEK_CUR);
2991 evlist__for_each_entry(evlist, evsel) {
2992 f_attr = (struct perf_file_attr){
2993 .attr = evsel->attr,
2995 .offset = evsel->id_offset,
2996 .size = evsel->ids * sizeof(u64),
2999 err = do_write(&ff, &f_attr, sizeof(f_attr));
3001 pr_debug("failed to write perf header attribute\n");
3006 if (!header->data_offset)
3007 header->data_offset = lseek(fd, 0, SEEK_CUR);
3008 header->feat_offset = header->data_offset + header->data_size;
3011 err = perf_header__adds_write(header, evlist, fd);
3016 f_header = (struct perf_file_header){
3017 .magic = PERF_MAGIC,
3018 .size = sizeof(f_header),
3019 .attr_size = sizeof(f_attr),
3021 .offset = attr_offset,
3022 .size = evlist->nr_entries * sizeof(f_attr),
3025 .offset = header->data_offset,
3026 .size = header->data_size,
3028 /* event_types is ignored, store zeros */
3031 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
3033 lseek(fd, 0, SEEK_SET);
3034 err = do_write(&ff, &f_header, sizeof(f_header));
3036 pr_debug("failed to write perf header\n");
3039 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
3044 static int perf_header__getbuffer64(struct perf_header *header,
3045 int fd, void *buf, size_t size)
3047 if (readn(fd, buf, size) <= 0)
3050 if (header->needs_swap)
3051 mem_bswap_64(buf, size);
3056 int perf_header__process_sections(struct perf_header *header, int fd,
3058 int (*process)(struct perf_file_section *section,
3059 struct perf_header *ph,
3060 int feat, int fd, void *data))
3062 struct perf_file_section *feat_sec, *sec;
3068 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
3072 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
3076 sec_size = sizeof(*feat_sec) * nr_sections;
3078 lseek(fd, header->feat_offset, SEEK_SET);
3080 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
3084 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
3085 err = process(sec++, header, feat, fd, data);
3095 static const int attr_file_abi_sizes[] = {
3096 [0] = PERF_ATTR_SIZE_VER0,
3097 [1] = PERF_ATTR_SIZE_VER1,
3098 [2] = PERF_ATTR_SIZE_VER2,
3099 [3] = PERF_ATTR_SIZE_VER3,
3100 [4] = PERF_ATTR_SIZE_VER4,
3105 * In the legacy file format, the magic number is not used to encode endianness.
3106 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
3107 * on ABI revisions, we need to try all combinations for all endianness to
3108 * detect the endianness.
3110 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
3112 uint64_t ref_size, attr_size;
3115 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
3116 ref_size = attr_file_abi_sizes[i]
3117 + sizeof(struct perf_file_section);
3118 if (hdr_sz != ref_size) {
3119 attr_size = bswap_64(hdr_sz);
3120 if (attr_size != ref_size)
3123 ph->needs_swap = true;
3125 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
3130 /* could not determine endianness */
3134 #define PERF_PIPE_HDR_VER0 16
3136 static const size_t attr_pipe_abi_sizes[] = {
3137 [0] = PERF_PIPE_HDR_VER0,
3142 * In the legacy pipe format, there is an implicit assumption that endiannesss
3143 * between host recording the samples, and host parsing the samples is the
3144 * same. This is not always the case given that the pipe output may always be
3145 * redirected into a file and analyzed on a different machine with possibly a
3146 * different endianness and perf_event ABI revsions in the perf tool itself.
3148 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
3153 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
3154 if (hdr_sz != attr_pipe_abi_sizes[i]) {
3155 attr_size = bswap_64(hdr_sz);
3156 if (attr_size != hdr_sz)
3159 ph->needs_swap = true;
3161 pr_debug("Pipe ABI%d perf.data file detected\n", i);
3167 bool is_perf_magic(u64 magic)
3169 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
3170 || magic == __perf_magic2
3171 || magic == __perf_magic2_sw)
3177 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
3178 bool is_pipe, struct perf_header *ph)
3182 /* check for legacy format */
3183 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
3185 ph->version = PERF_HEADER_VERSION_1;
3186 pr_debug("legacy perf.data format\n");
3188 return try_all_pipe_abis(hdr_sz, ph);
3190 return try_all_file_abis(hdr_sz, ph);
3193 * the new magic number serves two purposes:
3194 * - unique number to identify actual perf.data files
3195 * - encode endianness of file
3197 ph->version = PERF_HEADER_VERSION_2;
3199 /* check magic number with one endianness */
3200 if (magic == __perf_magic2)
3203 /* check magic number with opposite endianness */
3204 if (magic != __perf_magic2_sw)
3207 ph->needs_swap = true;
3212 int perf_file_header__read(struct perf_file_header *header,
3213 struct perf_header *ph, int fd)
3217 lseek(fd, 0, SEEK_SET);
3219 ret = readn(fd, header, sizeof(*header));
3223 if (check_magic_endian(header->magic,
3224 header->attr_size, false, ph) < 0) {
3225 pr_debug("magic/endian check failed\n");
3229 if (ph->needs_swap) {
3230 mem_bswap_64(header, offsetof(struct perf_file_header,
3234 if (header->size != sizeof(*header)) {
3235 /* Support the previous format */
3236 if (header->size == offsetof(typeof(*header), adds_features))
3237 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3240 } else if (ph->needs_swap) {
3242 * feature bitmap is declared as an array of unsigned longs --
3243 * not good since its size can differ between the host that
3244 * generated the data file and the host analyzing the file.
3246 * We need to handle endianness, but we don't know the size of
3247 * the unsigned long where the file was generated. Take a best
3248 * guess at determining it: try 64-bit swap first (ie., file
3249 * created on a 64-bit host), and check if the hostname feature
3250 * bit is set (this feature bit is forced on as of fbe96f2).
3251 * If the bit is not, undo the 64-bit swap and try a 32-bit
3252 * swap. If the hostname bit is still not set (e.g., older data
3253 * file), punt and fallback to the original behavior --
3254 * clearing all feature bits and setting buildid.
3256 mem_bswap_64(&header->adds_features,
3257 BITS_TO_U64(HEADER_FEAT_BITS));
3259 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3261 mem_bswap_64(&header->adds_features,
3262 BITS_TO_U64(HEADER_FEAT_BITS));
3265 mem_bswap_32(&header->adds_features,
3266 BITS_TO_U32(HEADER_FEAT_BITS));
3269 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3270 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3271 set_bit(HEADER_BUILD_ID, header->adds_features);
3275 memcpy(&ph->adds_features, &header->adds_features,
3276 sizeof(ph->adds_features));
3278 ph->data_offset = header->data.offset;
3279 ph->data_size = header->data.size;
3280 ph->feat_offset = header->data.offset + header->data.size;
3284 static int perf_file_section__process(struct perf_file_section *section,
3285 struct perf_header *ph,
3286 int feat, int fd, void *data)
3288 struct feat_fd fdd = {
3291 .size = section->size,
3292 .offset = section->offset,
3295 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3296 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3297 "%d, continuing...\n", section->offset, feat);
3301 if (feat >= HEADER_LAST_FEATURE) {
3302 pr_debug("unknown feature %d, continuing...\n", feat);
3306 if (!feat_ops[feat].process)
3309 return feat_ops[feat].process(&fdd, data);
3312 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
3313 struct perf_header *ph, int fd,
3316 struct feat_fd ff = {
3317 .fd = STDOUT_FILENO,
3322 ret = readn(fd, header, sizeof(*header));
3326 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
3327 pr_debug("endian/magic failed\n");
3332 header->size = bswap_64(header->size);
3334 if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
3340 static int perf_header__read_pipe(struct perf_session *session)
3342 struct perf_header *header = &session->header;
3343 struct perf_pipe_file_header f_header;
3345 if (perf_file_header__read_pipe(&f_header, header,
3346 perf_data__fd(session->data),
3347 session->repipe) < 0) {
3348 pr_debug("incompatible file format\n");
3355 static int read_attr(int fd, struct perf_header *ph,
3356 struct perf_file_attr *f_attr)
3358 struct perf_event_attr *attr = &f_attr->attr;
3360 size_t our_sz = sizeof(f_attr->attr);
3363 memset(f_attr, 0, sizeof(*f_attr));
3365 /* read minimal guaranteed structure */
3366 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
3368 pr_debug("cannot read %d bytes of header attr\n",
3369 PERF_ATTR_SIZE_VER0);
3373 /* on file perf_event_attr size */
3381 sz = PERF_ATTR_SIZE_VER0;
3382 } else if (sz > our_sz) {
3383 pr_debug("file uses a more recent and unsupported ABI"
3384 " (%zu bytes extra)\n", sz - our_sz);
3387 /* what we have not yet read and that we know about */
3388 left = sz - PERF_ATTR_SIZE_VER0;
3391 ptr += PERF_ATTR_SIZE_VER0;
3393 ret = readn(fd, ptr, left);
3395 /* read perf_file_section, ids are read in caller */
3396 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
3398 return ret <= 0 ? -1 : 0;
3401 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
3402 struct tep_handle *pevent)
3404 struct tep_event *event;
3407 /* already prepared */
3408 if (evsel->tp_format)
3411 if (pevent == NULL) {
3412 pr_debug("broken or missing trace data\n");
3416 event = tep_find_event(pevent, evsel->attr.config);
3417 if (event == NULL) {
3418 pr_debug("cannot find event format for %d\n", (int)evsel->attr.config);
3423 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
3424 evsel->name = strdup(bf);
3425 if (evsel->name == NULL)
3429 evsel->tp_format = event;
3433 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
3434 struct tep_handle *pevent)
3436 struct perf_evsel *pos;
3438 evlist__for_each_entry(evlist, pos) {
3439 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
3440 perf_evsel__prepare_tracepoint_event(pos, pevent))
3447 int perf_session__read_header(struct perf_session *session)
3449 struct perf_data *data = session->data;
3450 struct perf_header *header = &session->header;
3451 struct perf_file_header f_header;
3452 struct perf_file_attr f_attr;
3454 int nr_attrs, nr_ids, i, j;
3455 int fd = perf_data__fd(data);
3457 session->evlist = perf_evlist__new();
3458 if (session->evlist == NULL)
3461 session->evlist->env = &header->env;
3462 session->machines.host.env = &header->env;
3463 if (perf_data__is_pipe(data))
3464 return perf_header__read_pipe(session);
3466 if (perf_file_header__read(&f_header, header, fd) < 0)
3470 * Sanity check that perf.data was written cleanly; data size is
3471 * initialized to 0 and updated only if the on_exit function is run.
3472 * If data size is still 0 then the file contains only partial
3473 * information. Just warn user and process it as much as it can.
3475 if (f_header.data.size == 0) {
3476 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3477 "Was the 'perf record' command properly terminated?\n",
3481 nr_attrs = f_header.attrs.size / f_header.attr_size;
3482 lseek(fd, f_header.attrs.offset, SEEK_SET);
3484 for (i = 0; i < nr_attrs; i++) {
3485 struct perf_evsel *evsel;
3488 if (read_attr(fd, header, &f_attr) < 0)
3491 if (header->needs_swap) {
3492 f_attr.ids.size = bswap_64(f_attr.ids.size);
3493 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
3494 perf_event__attr_swap(&f_attr.attr);
3497 tmp = lseek(fd, 0, SEEK_CUR);
3498 evsel = perf_evsel__new(&f_attr.attr);
3501 goto out_delete_evlist;
3503 evsel->needs_swap = header->needs_swap;
3505 * Do it before so that if perf_evsel__alloc_id fails, this
3506 * entry gets purged too at perf_evlist__delete().
3508 perf_evlist__add(session->evlist, evsel);
3510 nr_ids = f_attr.ids.size / sizeof(u64);
3512 * We don't have the cpu and thread maps on the header, so
3513 * for allocating the perf_sample_id table we fake 1 cpu and
3514 * hattr->ids threads.
3516 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
3517 goto out_delete_evlist;
3519 lseek(fd, f_attr.ids.offset, SEEK_SET);
3521 for (j = 0; j < nr_ids; j++) {
3522 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
3525 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
3528 lseek(fd, tmp, SEEK_SET);
3531 perf_header__process_sections(header, fd, &session->tevent,
3532 perf_file_section__process);
3534 if (perf_evlist__prepare_tracepoint_events(session->evlist,
3535 session->tevent.pevent))
3536 goto out_delete_evlist;
3543 perf_evlist__delete(session->evlist);
3544 session->evlist = NULL;
3548 int perf_event__synthesize_attr(struct perf_tool *tool,
3549 struct perf_event_attr *attr, u32 ids, u64 *id,
3550 perf_event__handler_t process)
3552 union perf_event *ev;
3556 size = sizeof(struct perf_event_attr);
3557 size = PERF_ALIGN(size, sizeof(u64));
3558 size += sizeof(struct perf_event_header);
3559 size += ids * sizeof(u64);
3566 ev->attr.attr = *attr;
3567 memcpy(ev->attr.id, id, ids * sizeof(u64));
3569 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
3570 ev->attr.header.size = (u16)size;
3572 if (ev->attr.header.size == size)
3573 err = process(tool, ev, NULL, NULL);
3582 int perf_event__synthesize_features(struct perf_tool *tool,
3583 struct perf_session *session,
3584 struct perf_evlist *evlist,
3585 perf_event__handler_t process)
3587 struct perf_header *header = &session->header;
3589 struct feature_event *fe;
3593 sz_hdr = sizeof(fe->header);
3594 sz = sizeof(union perf_event);
3595 /* get a nice alignment */
3596 sz = PERF_ALIGN(sz, page_size);
3598 memset(&ff, 0, sizeof(ff));
3600 ff.buf = malloc(sz);
3604 ff.size = sz - sz_hdr;
3606 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3607 if (!feat_ops[feat].synthesize) {
3608 pr_debug("No record header feature for header :%d\n", feat);
3612 ff.offset = sizeof(*fe);
3614 ret = feat_ops[feat].write(&ff, evlist);
3615 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
3616 pr_debug("Error writing feature\n");
3619 /* ff.buf may have changed due to realloc in do_write() */
3621 memset(fe, 0, sizeof(*fe));
3624 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3625 fe->header.size = ff.offset;
3627 ret = process(tool, ff.buf, NULL, NULL);
3634 /* Send HEADER_LAST_FEATURE mark. */
3636 fe->feat_id = HEADER_LAST_FEATURE;
3637 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3638 fe->header.size = sizeof(*fe);
3640 ret = process(tool, ff.buf, NULL, NULL);
3646 int perf_event__process_feature(struct perf_session *session,
3647 union perf_event *event)
3649 struct perf_tool *tool = session->tool;
3650 struct feat_fd ff = { .fd = 0 };
3651 struct feature_event *fe = (struct feature_event *)event;
3652 int type = fe->header.type;
3653 u64 feat = fe->feat_id;
3655 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3656 pr_warning("invalid record type %d in pipe-mode\n", type);
3659 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) {
3660 pr_warning("invalid record type %d in pipe-mode\n", type);
3664 if (!feat_ops[feat].process)
3667 ff.buf = (void *)fe->data;
3668 ff.size = event->header.size - sizeof(event->header);
3669 ff.ph = &session->header;
3671 if (feat_ops[feat].process(&ff, NULL))
3674 if (!feat_ops[feat].print || !tool->show_feat_hdr)
3677 if (!feat_ops[feat].full_only ||
3678 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3679 feat_ops[feat].print(&ff, stdout);
3681 fprintf(stdout, "# %s info available, use -I to display\n",
3682 feat_ops[feat].name);
3688 static struct event_update_event *
3689 event_update_event__new(size_t size, u64 type, u64 id)
3691 struct event_update_event *ev;
3693 size += sizeof(*ev);
3694 size = PERF_ALIGN(size, sizeof(u64));
3698 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3699 ev->header.size = (u16)size;
3707 perf_event__synthesize_event_update_unit(struct perf_tool *tool,
3708 struct perf_evsel *evsel,
3709 perf_event__handler_t process)
3711 struct event_update_event *ev;
3712 size_t size = strlen(evsel->unit);
3715 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->id[0]);
3719 strlcpy(ev->data, evsel->unit, size + 1);
3720 err = process(tool, (union perf_event *)ev, NULL, NULL);
3726 perf_event__synthesize_event_update_scale(struct perf_tool *tool,
3727 struct perf_evsel *evsel,
3728 perf_event__handler_t process)
3730 struct event_update_event *ev;
3731 struct event_update_event_scale *ev_data;
3734 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->id[0]);
3738 ev_data = (struct event_update_event_scale *) ev->data;
3739 ev_data->scale = evsel->scale;
3740 err = process(tool, (union perf_event*) ev, NULL, NULL);
3746 perf_event__synthesize_event_update_name(struct perf_tool *tool,
3747 struct perf_evsel *evsel,
3748 perf_event__handler_t process)
3750 struct event_update_event *ev;
3751 size_t len = strlen(evsel->name);
3754 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->id[0]);
3758 strlcpy(ev->data, evsel->name, len + 1);
3759 err = process(tool, (union perf_event*) ev, NULL, NULL);
3765 perf_event__synthesize_event_update_cpus(struct perf_tool *tool,
3766 struct perf_evsel *evsel,
3767 perf_event__handler_t process)
3769 size_t size = sizeof(struct event_update_event);
3770 struct event_update_event *ev;
3774 if (!evsel->own_cpus)
3777 ev = cpu_map_data__alloc(evsel->own_cpus, &size, &type, &max);
3781 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3782 ev->header.size = (u16)size;
3783 ev->type = PERF_EVENT_UPDATE__CPUS;
3784 ev->id = evsel->id[0];
3786 cpu_map_data__synthesize((struct cpu_map_data *) ev->data,
3790 err = process(tool, (union perf_event*) ev, NULL, NULL);
3795 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3797 struct event_update_event *ev = &event->event_update;
3798 struct event_update_event_scale *ev_scale;
3799 struct event_update_event_cpus *ev_cpus;
3800 struct cpu_map *map;
3803 ret = fprintf(fp, "\n... id: %" PRIu64 "\n", ev->id);
3806 case PERF_EVENT_UPDATE__SCALE:
3807 ev_scale = (struct event_update_event_scale *) ev->data;
3808 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3810 case PERF_EVENT_UPDATE__UNIT:
3811 ret += fprintf(fp, "... unit: %s\n", ev->data);
3813 case PERF_EVENT_UPDATE__NAME:
3814 ret += fprintf(fp, "... name: %s\n", ev->data);
3816 case PERF_EVENT_UPDATE__CPUS:
3817 ev_cpus = (struct event_update_event_cpus *) ev->data;
3818 ret += fprintf(fp, "... ");
3820 map = cpu_map__new_data(&ev_cpus->cpus);
3822 ret += cpu_map__fprintf(map, fp);
3824 ret += fprintf(fp, "failed to get cpus\n");
3827 ret += fprintf(fp, "... unknown type\n");
3834 int perf_event__synthesize_attrs(struct perf_tool *tool,
3835 struct perf_evlist *evlist,
3836 perf_event__handler_t process)
3838 struct perf_evsel *evsel;
3841 evlist__for_each_entry(evlist, evsel) {
3842 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
3843 evsel->id, process);
3845 pr_debug("failed to create perf header attribute\n");
3853 static bool has_unit(struct perf_evsel *counter)
3855 return counter->unit && *counter->unit;
3858 static bool has_scale(struct perf_evsel *counter)
3860 return counter->scale != 1;
3863 int perf_event__synthesize_extra_attr(struct perf_tool *tool,
3864 struct perf_evlist *evsel_list,
3865 perf_event__handler_t process,
3868 struct perf_evsel *counter;
3872 * Synthesize other events stuff not carried within
3873 * attr event - unit, scale, name
3875 evlist__for_each_entry(evsel_list, counter) {
3876 if (!counter->supported)
3880 * Synthesize unit and scale only if it's defined.
3882 if (has_unit(counter)) {
3883 err = perf_event__synthesize_event_update_unit(tool, counter, process);
3885 pr_err("Couldn't synthesize evsel unit.\n");
3890 if (has_scale(counter)) {
3891 err = perf_event__synthesize_event_update_scale(tool, counter, process);
3893 pr_err("Couldn't synthesize evsel counter.\n");
3898 if (counter->own_cpus) {
3899 err = perf_event__synthesize_event_update_cpus(tool, counter, process);
3901 pr_err("Couldn't synthesize evsel cpus.\n");
3907 * Name is needed only for pipe output,
3908 * perf.data carries event names.
3911 err = perf_event__synthesize_event_update_name(tool, counter, process);
3913 pr_err("Couldn't synthesize evsel name.\n");
3921 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3922 union perf_event *event,
3923 struct perf_evlist **pevlist)
3926 struct perf_evsel *evsel;
3927 struct perf_evlist *evlist = *pevlist;
3929 if (evlist == NULL) {
3930 *pevlist = evlist = perf_evlist__new();
3935 evsel = perf_evsel__new(&event->attr.attr);
3939 perf_evlist__add(evlist, evsel);
3941 ids = event->header.size;
3942 ids -= (void *)&event->attr.id - (void *)event;
3943 n_ids = ids / sizeof(u64);
3945 * We don't have the cpu and thread maps on the header, so
3946 * for allocating the perf_sample_id table we fake 1 cpu and
3947 * hattr->ids threads.
3949 if (perf_evsel__alloc_id(evsel, 1, n_ids))
3952 for (i = 0; i < n_ids; i++) {
3953 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
3959 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3960 union perf_event *event,
3961 struct perf_evlist **pevlist)
3963 struct event_update_event *ev = &event->event_update;
3964 struct event_update_event_scale *ev_scale;
3965 struct event_update_event_cpus *ev_cpus;
3966 struct perf_evlist *evlist;
3967 struct perf_evsel *evsel;
3968 struct cpu_map *map;
3970 if (!pevlist || *pevlist == NULL)
3975 evsel = perf_evlist__id2evsel(evlist, ev->id);
3980 case PERF_EVENT_UPDATE__UNIT:
3981 evsel->unit = strdup(ev->data);
3983 case PERF_EVENT_UPDATE__NAME:
3984 evsel->name = strdup(ev->data);
3986 case PERF_EVENT_UPDATE__SCALE:
3987 ev_scale = (struct event_update_event_scale *) ev->data;
3988 evsel->scale = ev_scale->scale;
3990 case PERF_EVENT_UPDATE__CPUS:
3991 ev_cpus = (struct event_update_event_cpus *) ev->data;
3993 map = cpu_map__new_data(&ev_cpus->cpus);
3995 evsel->own_cpus = map;
3997 pr_err("failed to get event_update cpus\n");
4005 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
4006 struct perf_evlist *evlist,
4007 perf_event__handler_t process)
4009 union perf_event ev;
4010 struct tracing_data *tdata;
4011 ssize_t size = 0, aligned_size = 0, padding;
4013 int err __maybe_unused = 0;
4016 * We are going to store the size of the data followed
4017 * by the data contents. Since the fd descriptor is a pipe,
4018 * we cannot seek back to store the size of the data once
4019 * we know it. Instead we:
4021 * - write the tracing data to the temp file
4022 * - get/write the data size to pipe
4023 * - write the tracing data from the temp file
4026 tdata = tracing_data_get(&evlist->entries, fd, true);
4030 memset(&ev, 0, sizeof(ev));
4032 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
4034 aligned_size = PERF_ALIGN(size, sizeof(u64));
4035 padding = aligned_size - size;
4036 ev.tracing_data.header.size = sizeof(ev.tracing_data);
4037 ev.tracing_data.size = aligned_size;
4039 process(tool, &ev, NULL, NULL);
4042 * The put function will copy all the tracing data
4043 * stored in temp file to the pipe.
4045 tracing_data_put(tdata);
4047 ff = (struct feat_fd){ .fd = fd };
4048 if (write_padded(&ff, NULL, 0, padding))
4051 return aligned_size;
4054 int perf_event__process_tracing_data(struct perf_session *session,
4055 union perf_event *event)
4057 ssize_t size_read, padding, size = event->tracing_data.size;
4058 int fd = perf_data__fd(session->data);
4059 off_t offset = lseek(fd, 0, SEEK_CUR);
4062 /* setup for reading amidst mmap */
4063 lseek(fd, offset + sizeof(struct tracing_data_event),
4066 size_read = trace_report(fd, &session->tevent,
4068 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
4070 if (readn(fd, buf, padding) < 0) {
4071 pr_err("%s: reading input file", __func__);
4074 if (session->repipe) {
4075 int retw = write(STDOUT_FILENO, buf, padding);
4076 if (retw <= 0 || retw != padding) {
4077 pr_err("%s: repiping tracing data padding", __func__);
4082 if (size_read + padding != size) {
4083 pr_err("%s: tracing data size mismatch", __func__);
4087 perf_evlist__prepare_tracepoint_events(session->evlist,
4088 session->tevent.pevent);
4090 return size_read + padding;
4093 int perf_event__synthesize_build_id(struct perf_tool *tool,
4094 struct dso *pos, u16 misc,
4095 perf_event__handler_t process,
4096 struct machine *machine)
4098 union perf_event ev;
4105 memset(&ev, 0, sizeof(ev));
4107 len = pos->long_name_len + 1;
4108 len = PERF_ALIGN(len, NAME_ALIGN);
4109 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
4110 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
4111 ev.build_id.header.misc = misc;
4112 ev.build_id.pid = machine->pid;
4113 ev.build_id.header.size = sizeof(ev.build_id) + len;
4114 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
4116 err = process(tool, &ev, NULL, machine);
4121 int perf_event__process_build_id(struct perf_session *session,
4122 union perf_event *event)
4124 __event_process_build_id(&event->build_id,
4125 event->build_id.filename,