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>
27 #include "trace-event.h"
37 #include <api/fs/fs.h>
40 #include "time-utils.h"
43 #include "sane_ctype.h"
47 * must be a numerical value to let the endianness
48 * determine the memory layout. That way we are able
49 * to detect endianness when reading the perf.data file
52 * we check for legacy (PERFFILE) format.
54 static const char *__perf_magic1 = "PERFFILE";
55 static const u64 __perf_magic2 = 0x32454c4946524550ULL;
56 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
58 #define PERF_MAGIC __perf_magic2
60 const char perf_version_string[] = PERF_VERSION;
62 struct perf_file_attr {
63 struct perf_event_attr attr;
64 struct perf_file_section ids;
68 struct perf_header *ph;
70 void *buf; /* Either buf != NULL or fd >= 0 */
73 struct perf_evsel *events;
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 perf_evlist *evlist)
300 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
303 return read_tracing_data(ff->fd, &evlist->entries);
306 static int write_build_id(struct feat_fd *ff,
307 struct perf_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 perf_evlist *evlist __maybe_unused)
340 return do_write_string(ff, uts.nodename);
343 static int write_osrelease(struct feat_fd *ff,
344 struct perf_evlist *evlist __maybe_unused)
353 return do_write_string(ff, uts.release);
356 static int write_arch(struct feat_fd *ff,
357 struct perf_evlist *evlist __maybe_unused)
366 return do_write_string(ff, uts.machine);
369 static int write_version(struct feat_fd *ff,
370 struct perf_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) */
418 while (*q && isspace(*q))
421 while ((*r++ = *q++));
425 ret = do_write_string(ff, s);
432 static int write_cpudesc(struct feat_fd *ff,
433 struct perf_evlist *evlist __maybe_unused)
435 const char *cpuinfo_procs[] = CPUINFO_PROC;
438 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
440 ret = __write_cpudesc(ff, cpuinfo_procs[i]);
448 static int write_nrcpus(struct feat_fd *ff,
449 struct perf_evlist *evlist __maybe_unused)
455 nrc = cpu__max_present_cpu();
457 nr = sysconf(_SC_NPROCESSORS_ONLN);
461 nra = (u32)(nr & UINT_MAX);
463 ret = do_write(ff, &nrc, sizeof(nrc));
467 return do_write(ff, &nra, sizeof(nra));
470 static int write_event_desc(struct feat_fd *ff,
471 struct perf_evlist *evlist)
473 struct perf_evsel *evsel;
477 nre = evlist->nr_entries;
480 * write number of events
482 ret = do_write(ff, &nre, sizeof(nre));
487 * size of perf_event_attr struct
489 sz = (u32)sizeof(evsel->attr);
490 ret = do_write(ff, &sz, sizeof(sz));
494 evlist__for_each_entry(evlist, evsel) {
495 ret = do_write(ff, &evsel->attr, sz);
499 * write number of unique id per event
500 * there is one id per instance of an event
502 * copy into an nri to be independent of the
506 ret = do_write(ff, &nri, sizeof(nri));
511 * write event string as passed on cmdline
513 ret = do_write_string(ff, perf_evsel__name(evsel));
517 * write unique ids for this event
519 ret = do_write(ff, evsel->id, evsel->ids * sizeof(u64));
526 static int write_cmdline(struct feat_fd *ff,
527 struct perf_evlist *evlist __maybe_unused)
529 char buf[MAXPATHLEN];
533 /* actual path to perf binary */
534 ret = readlink("/proc/self/exe", buf, sizeof(buf) - 1);
538 /* readlink() does not add null termination */
541 /* account for binary path */
542 n = perf_env.nr_cmdline + 1;
544 ret = do_write(ff, &n, sizeof(n));
548 ret = do_write_string(ff, buf);
552 for (i = 0 ; i < perf_env.nr_cmdline; i++) {
553 ret = do_write_string(ff, perf_env.cmdline_argv[i]);
560 #define CORE_SIB_FMT \
561 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
562 #define THRD_SIB_FMT \
563 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
568 char **core_siblings;
569 char **thread_siblings;
572 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
575 char filename[MAXPATHLEN];
576 char *buf = NULL, *p;
582 sprintf(filename, CORE_SIB_FMT, cpu);
583 fp = fopen(filename, "r");
587 sret = getline(&buf, &len, fp);
592 p = strchr(buf, '\n');
596 for (i = 0; i < tp->core_sib; i++) {
597 if (!strcmp(buf, tp->core_siblings[i]))
600 if (i == tp->core_sib) {
601 tp->core_siblings[i] = buf;
609 sprintf(filename, THRD_SIB_FMT, cpu);
610 fp = fopen(filename, "r");
614 if (getline(&buf, &len, fp) <= 0)
617 p = strchr(buf, '\n');
621 for (i = 0; i < tp->thread_sib; i++) {
622 if (!strcmp(buf, tp->thread_siblings[i]))
625 if (i == tp->thread_sib) {
626 tp->thread_siblings[i] = buf;
638 static void free_cpu_topo(struct cpu_topo *tp)
645 for (i = 0 ; i < tp->core_sib; i++)
646 zfree(&tp->core_siblings[i]);
648 for (i = 0 ; i < tp->thread_sib; i++)
649 zfree(&tp->thread_siblings[i]);
654 static struct cpu_topo *build_cpu_topology(void)
656 struct cpu_topo *tp = NULL;
664 ncpus = cpu__max_present_cpu();
666 /* build online CPU map */
667 map = cpu_map__new(NULL);
669 pr_debug("failed to get system cpumap\n");
673 nr = (u32)(ncpus & UINT_MAX);
675 sz = nr * sizeof(char *);
676 addr = calloc(1, sizeof(*tp) + 2 * sz);
682 tp->core_siblings = addr;
684 tp->thread_siblings = addr;
686 for (i = 0; i < nr; i++) {
687 if (!cpu_map__has(map, i))
690 ret = build_cpu_topo(tp, i);
704 static int write_cpu_topology(struct feat_fd *ff,
705 struct perf_evlist *evlist __maybe_unused)
711 tp = build_cpu_topology();
715 ret = do_write(ff, &tp->core_sib, sizeof(tp->core_sib));
719 for (i = 0; i < tp->core_sib; i++) {
720 ret = do_write_string(ff, tp->core_siblings[i]);
724 ret = do_write(ff, &tp->thread_sib, sizeof(tp->thread_sib));
728 for (i = 0; i < tp->thread_sib; i++) {
729 ret = do_write_string(ff, tp->thread_siblings[i]);
734 ret = perf_env__read_cpu_topology_map(&perf_env);
738 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
739 ret = do_write(ff, &perf_env.cpu[j].core_id,
740 sizeof(perf_env.cpu[j].core_id));
743 ret = do_write(ff, &perf_env.cpu[j].socket_id,
744 sizeof(perf_env.cpu[j].socket_id));
755 static int write_total_mem(struct feat_fd *ff,
756 struct perf_evlist *evlist __maybe_unused)
764 fp = fopen("/proc/meminfo", "r");
768 while (getline(&buf, &len, fp) > 0) {
769 ret = strncmp(buf, "MemTotal:", 9);
774 n = sscanf(buf, "%*s %"PRIu64, &mem);
776 ret = do_write(ff, &mem, sizeof(mem));
784 static int write_topo_node(struct feat_fd *ff, int node)
786 char str[MAXPATHLEN];
788 char *buf = NULL, *p;
791 u64 mem_total, mem_free, mem;
794 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
795 fp = fopen(str, "r");
799 while (getline(&buf, &len, fp) > 0) {
800 /* skip over invalid lines */
801 if (!strchr(buf, ':'))
803 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
805 if (!strcmp(field, "MemTotal:"))
807 if (!strcmp(field, "MemFree:"))
814 ret = do_write(ff, &mem_total, sizeof(u64));
818 ret = do_write(ff, &mem_free, sizeof(u64));
823 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
825 fp = fopen(str, "r");
829 if (getline(&buf, &len, fp) <= 0)
832 p = strchr(buf, '\n');
836 ret = do_write_string(ff, buf);
844 static int write_numa_topology(struct feat_fd *ff,
845 struct perf_evlist *evlist __maybe_unused)
850 struct cpu_map *node_map = NULL;
855 fp = fopen("/sys/devices/system/node/online", "r");
859 if (getline(&buf, &len, fp) <= 0)
862 c = strchr(buf, '\n');
866 node_map = cpu_map__new(buf);
870 nr = (u32)node_map->nr;
872 ret = do_write(ff, &nr, sizeof(nr));
876 for (i = 0; i < nr; i++) {
877 j = (u32)node_map->map[i];
878 ret = do_write(ff, &j, sizeof(j));
882 ret = write_topo_node(ff, j);
889 cpu_map__put(node_map);
896 * struct pmu_mappings {
905 static int write_pmu_mappings(struct feat_fd *ff,
906 struct perf_evlist *evlist __maybe_unused)
908 struct perf_pmu *pmu = NULL;
913 * Do a first pass to count number of pmu to avoid lseek so this
914 * works in pipe mode as well.
916 while ((pmu = perf_pmu__scan(pmu))) {
922 ret = do_write(ff, &pmu_num, sizeof(pmu_num));
926 while ((pmu = perf_pmu__scan(pmu))) {
930 ret = do_write(ff, &pmu->type, sizeof(pmu->type));
934 ret = do_write_string(ff, pmu->name);
945 * struct group_descs {
947 * struct group_desc {
954 static int write_group_desc(struct feat_fd *ff,
955 struct perf_evlist *evlist)
957 u32 nr_groups = evlist->nr_groups;
958 struct perf_evsel *evsel;
961 ret = do_write(ff, &nr_groups, sizeof(nr_groups));
965 evlist__for_each_entry(evlist, evsel) {
966 if (perf_evsel__is_group_leader(evsel) &&
967 evsel->nr_members > 1) {
968 const char *name = evsel->group_name ?: "{anon_group}";
969 u32 leader_idx = evsel->idx;
970 u32 nr_members = evsel->nr_members;
972 ret = do_write_string(ff, name);
976 ret = do_write(ff, &leader_idx, sizeof(leader_idx));
980 ret = do_write(ff, &nr_members, sizeof(nr_members));
989 * Return the CPU id as a raw string.
991 * Each architecture should provide a more precise id string that
992 * can be use to match the architecture's "mapfile".
994 char * __weak get_cpuid_str(struct perf_pmu *pmu __maybe_unused)
999 /* Return zero when the cpuid from the mapfile.csv matches the
1000 * cpuid string generated on this platform.
1001 * Otherwise return non-zero.
1003 int __weak strcmp_cpuid_str(const char *mapcpuid, const char *cpuid)
1006 regmatch_t pmatch[1];
1009 if (regcomp(&re, mapcpuid, REG_EXTENDED) != 0) {
1010 /* Warn unable to generate match particular string. */
1011 pr_info("Invalid regular expression %s\n", mapcpuid);
1015 match = !regexec(&re, cpuid, 1, pmatch, 0);
1018 size_t match_len = (pmatch[0].rm_eo - pmatch[0].rm_so);
1020 /* Verify the entire string matched. */
1021 if (match_len == strlen(cpuid))
1028 * default get_cpuid(): nothing gets recorded
1029 * actual implementation must be in arch/$(SRCARCH)/util/header.c
1031 int __weak get_cpuid(char *buffer __maybe_unused, size_t sz __maybe_unused)
1036 static int write_cpuid(struct feat_fd *ff,
1037 struct perf_evlist *evlist __maybe_unused)
1042 ret = get_cpuid(buffer, sizeof(buffer));
1046 return do_write_string(ff, buffer);
1049 static int write_branch_stack(struct feat_fd *ff __maybe_unused,
1050 struct perf_evlist *evlist __maybe_unused)
1055 static int write_auxtrace(struct feat_fd *ff,
1056 struct perf_evlist *evlist __maybe_unused)
1058 struct perf_session *session;
1061 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
1064 session = container_of(ff->ph, struct perf_session, header);
1066 err = auxtrace_index__write(ff->fd, &session->auxtrace_index);
1068 pr_err("Failed to write auxtrace index\n");
1072 static int write_clockid(struct feat_fd *ff,
1073 struct perf_evlist *evlist __maybe_unused)
1075 return do_write(ff, &ff->ph->env.clockid_res_ns,
1076 sizeof(ff->ph->env.clockid_res_ns));
1079 static int cpu_cache_level__sort(const void *a, const void *b)
1081 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
1082 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
1084 return cache_a->level - cache_b->level;
1087 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
1089 if (a->level != b->level)
1092 if (a->line_size != b->line_size)
1095 if (a->sets != b->sets)
1098 if (a->ways != b->ways)
1101 if (strcmp(a->type, b->type))
1104 if (strcmp(a->size, b->size))
1107 if (strcmp(a->map, b->map))
1113 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
1115 char path[PATH_MAX], file[PATH_MAX];
1119 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
1120 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
1122 if (stat(file, &st))
1125 scnprintf(file, PATH_MAX, "%s/level", path);
1126 if (sysfs__read_int(file, (int *) &cache->level))
1129 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
1130 if (sysfs__read_int(file, (int *) &cache->line_size))
1133 scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
1134 if (sysfs__read_int(file, (int *) &cache->sets))
1137 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
1138 if (sysfs__read_int(file, (int *) &cache->ways))
1141 scnprintf(file, PATH_MAX, "%s/type", path);
1142 if (sysfs__read_str(file, &cache->type, &len))
1145 cache->type[len] = 0;
1146 cache->type = rtrim(cache->type);
1148 scnprintf(file, PATH_MAX, "%s/size", path);
1149 if (sysfs__read_str(file, &cache->size, &len)) {
1154 cache->size[len] = 0;
1155 cache->size = rtrim(cache->size);
1157 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
1158 if (sysfs__read_str(file, &cache->map, &len)) {
1164 cache->map[len] = 0;
1165 cache->map = rtrim(cache->map);
1169 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
1171 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
1174 static int build_caches(struct cpu_cache_level caches[], u32 size, u32 *cntp)
1181 ncpus = sysconf(_SC_NPROCESSORS_CONF);
1185 nr = (u32)(ncpus & UINT_MAX);
1187 for (cpu = 0; cpu < nr; cpu++) {
1188 for (level = 0; level < 10; level++) {
1189 struct cpu_cache_level c;
1192 err = cpu_cache_level__read(&c, cpu, level);
1199 for (i = 0; i < cnt; i++) {
1200 if (cpu_cache_level__cmp(&c, &caches[i]))
1207 cpu_cache_level__free(&c);
1209 if (WARN_ONCE(cnt == size, "way too many cpu caches.."))
1218 #define MAX_CACHES 2000
1220 static int write_cache(struct feat_fd *ff,
1221 struct perf_evlist *evlist __maybe_unused)
1223 struct cpu_cache_level caches[MAX_CACHES];
1224 u32 cnt = 0, i, version = 1;
1227 ret = build_caches(caches, MAX_CACHES, &cnt);
1231 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1233 ret = do_write(ff, &version, sizeof(u32));
1237 ret = do_write(ff, &cnt, sizeof(u32));
1241 for (i = 0; i < cnt; i++) {
1242 struct cpu_cache_level *c = &caches[i];
1245 ret = do_write(ff, &c->v, sizeof(u32)); \
1256 ret = do_write_string(ff, (const char *) c->v); \
1267 for (i = 0; i < cnt; i++)
1268 cpu_cache_level__free(&caches[i]);
1272 static int write_stat(struct feat_fd *ff __maybe_unused,
1273 struct perf_evlist *evlist __maybe_unused)
1278 static int write_sample_time(struct feat_fd *ff,
1279 struct perf_evlist *evlist)
1283 ret = do_write(ff, &evlist->first_sample_time,
1284 sizeof(evlist->first_sample_time));
1288 return do_write(ff, &evlist->last_sample_time,
1289 sizeof(evlist->last_sample_time));
1293 static int memory_node__read(struct memory_node *n, unsigned long idx)
1295 unsigned int phys, size = 0;
1296 char path[PATH_MAX];
1300 #define for_each_memory(mem, dir) \
1301 while ((ent = readdir(dir))) \
1302 if (strcmp(ent->d_name, ".") && \
1303 strcmp(ent->d_name, "..") && \
1304 sscanf(ent->d_name, "memory%u", &mem) == 1)
1306 scnprintf(path, PATH_MAX,
1307 "%s/devices/system/node/node%lu",
1308 sysfs__mountpoint(), idx);
1310 dir = opendir(path);
1312 pr_warning("failed: cant' open memory sysfs data\n");
1316 for_each_memory(phys, dir) {
1317 size = max(phys, size);
1322 n->set = bitmap_alloc(size);
1333 for_each_memory(phys, dir) {
1334 set_bit(phys, n->set);
1341 static int memory_node__sort(const void *a, const void *b)
1343 const struct memory_node *na = a;
1344 const struct memory_node *nb = b;
1346 return na->node - nb->node;
1349 static int build_mem_topology(struct memory_node *nodes, u64 size, u64 *cntp)
1351 char path[PATH_MAX];
1357 scnprintf(path, PATH_MAX, "%s/devices/system/node/",
1358 sysfs__mountpoint());
1360 dir = opendir(path);
1362 pr_debug2("%s: could't read %s, does this arch have topology information?\n",
1367 while (!ret && (ent = readdir(dir))) {
1371 if (!strcmp(ent->d_name, ".") ||
1372 !strcmp(ent->d_name, ".."))
1375 r = sscanf(ent->d_name, "node%u", &idx);
1379 if (WARN_ONCE(cnt >= size,
1380 "failed to write MEM_TOPOLOGY, way too many nodes\n"))
1383 ret = memory_node__read(&nodes[cnt++], idx);
1390 qsort(nodes, cnt, sizeof(nodes[0]), memory_node__sort);
1395 #define MAX_MEMORY_NODES 2000
1398 * The MEM_TOPOLOGY holds physical memory map for every
1399 * node in system. The format of data is as follows:
1401 * 0 - version | for future changes
1402 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes
1403 * 16 - count | number of nodes
1405 * For each node we store map of physical indexes for
1408 * 32 - node id | node index
1409 * 40 - size | size of bitmap
1410 * 48 - bitmap | bitmap of memory indexes that belongs to node
1412 static int write_mem_topology(struct feat_fd *ff __maybe_unused,
1413 struct perf_evlist *evlist __maybe_unused)
1415 static struct memory_node nodes[MAX_MEMORY_NODES];
1416 u64 bsize, version = 1, i, nr;
1419 ret = sysfs__read_xll("devices/system/memory/block_size_bytes",
1420 (unsigned long long *) &bsize);
1424 ret = build_mem_topology(&nodes[0], MAX_MEMORY_NODES, &nr);
1428 ret = do_write(ff, &version, sizeof(version));
1432 ret = do_write(ff, &bsize, sizeof(bsize));
1436 ret = do_write(ff, &nr, sizeof(nr));
1440 for (i = 0; i < nr; i++) {
1441 struct memory_node *n = &nodes[i];
1444 ret = do_write(ff, &n->v, sizeof(n->v)); \
1453 ret = do_write_bitmap(ff, n->set, n->size);
1462 static void print_hostname(struct feat_fd *ff, FILE *fp)
1464 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1467 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1469 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1472 static void print_arch(struct feat_fd *ff, FILE *fp)
1474 fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1477 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1479 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1482 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1484 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1485 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1488 static void print_version(struct feat_fd *ff, FILE *fp)
1490 fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1493 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1497 nr = ff->ph->env.nr_cmdline;
1499 fprintf(fp, "# cmdline : ");
1501 for (i = 0; i < nr; i++) {
1502 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]);
1504 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1508 char *quote = strchr(argv_i, '\'');
1512 fprintf(fp, "%s\\\'", argv_i);
1515 fprintf(fp, "%s ", argv_i);
1522 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1524 struct perf_header *ph = ff->ph;
1525 int cpu_nr = ph->env.nr_cpus_avail;
1529 nr = ph->env.nr_sibling_cores;
1530 str = ph->env.sibling_cores;
1532 for (i = 0; i < nr; i++) {
1533 fprintf(fp, "# sibling cores : %s\n", str);
1534 str += strlen(str) + 1;
1537 nr = ph->env.nr_sibling_threads;
1538 str = ph->env.sibling_threads;
1540 for (i = 0; i < nr; i++) {
1541 fprintf(fp, "# sibling threads : %s\n", str);
1542 str += strlen(str) + 1;
1545 if (ph->env.cpu != NULL) {
1546 for (i = 0; i < cpu_nr; i++)
1547 fprintf(fp, "# CPU %d: Core ID %d, Socket ID %d\n", i,
1548 ph->env.cpu[i].core_id, ph->env.cpu[i].socket_id);
1550 fprintf(fp, "# Core ID and Socket ID information is not available\n");
1553 static void print_clockid(struct feat_fd *ff, FILE *fp)
1555 fprintf(fp, "# clockid frequency: %"PRIu64" MHz\n",
1556 ff->ph->env.clockid_res_ns * 1000);
1559 static void free_event_desc(struct perf_evsel *events)
1561 struct perf_evsel *evsel;
1566 for (evsel = events; evsel->attr.size; evsel++) {
1567 zfree(&evsel->name);
1574 static struct perf_evsel *read_event_desc(struct feat_fd *ff)
1576 struct perf_evsel *evsel, *events = NULL;
1579 u32 nre, sz, nr, i, j;
1582 /* number of events */
1583 if (do_read_u32(ff, &nre))
1586 if (do_read_u32(ff, &sz))
1589 /* buffer to hold on file attr struct */
1594 /* the last event terminates with evsel->attr.size == 0: */
1595 events = calloc(nre + 1, sizeof(*events));
1599 msz = sizeof(evsel->attr);
1603 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1607 * must read entire on-file attr struct to
1608 * sync up with layout.
1610 if (__do_read(ff, buf, sz))
1613 if (ff->ph->needs_swap)
1614 perf_event__attr_swap(buf);
1616 memcpy(&evsel->attr, buf, msz);
1618 if (do_read_u32(ff, &nr))
1621 if (ff->ph->needs_swap)
1622 evsel->needs_swap = true;
1624 evsel->name = do_read_string(ff);
1631 id = calloc(nr, sizeof(*id));
1637 for (j = 0 ; j < nr; j++) {
1638 if (do_read_u64(ff, id))
1647 free_event_desc(events);
1652 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1653 void *priv __maybe_unused)
1655 return fprintf(fp, ", %s = %s", name, val);
1658 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1660 struct perf_evsel *evsel, *events;
1665 events = ff->events;
1667 events = read_event_desc(ff);
1670 fprintf(fp, "# event desc: not available or unable to read\n");
1674 for (evsel = events; evsel->attr.size; evsel++) {
1675 fprintf(fp, "# event : name = %s, ", evsel->name);
1678 fprintf(fp, ", id = {");
1679 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1682 fprintf(fp, " %"PRIu64, *id);
1687 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1692 free_event_desc(events);
1696 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1698 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1701 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1704 struct numa_node *n;
1706 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1707 n = &ff->ph->env.numa_nodes[i];
1709 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1710 " free = %"PRIu64" kB\n",
1711 n->node, n->mem_total, n->mem_free);
1713 fprintf(fp, "# node%u cpu list : ", n->node);
1714 cpu_map__fprintf(n->map, fp);
1718 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1720 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1723 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1725 fprintf(fp, "# contains samples with branch stack\n");
1728 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1730 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1733 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1735 fprintf(fp, "# contains stat data\n");
1738 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1742 fprintf(fp, "# CPU cache info:\n");
1743 for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1745 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1749 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1751 const char *delimiter = "# pmu mappings: ";
1756 pmu_num = ff->ph->env.nr_pmu_mappings;
1758 fprintf(fp, "# pmu mappings: not available\n");
1762 str = ff->ph->env.pmu_mappings;
1765 type = strtoul(str, &tmp, 0);
1770 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1773 str += strlen(str) + 1;
1782 fprintf(fp, "# pmu mappings: unable to read\n");
1785 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1787 struct perf_session *session;
1788 struct perf_evsel *evsel;
1791 session = container_of(ff->ph, struct perf_session, header);
1793 evlist__for_each_entry(session->evlist, evsel) {
1794 if (perf_evsel__is_group_leader(evsel) &&
1795 evsel->nr_members > 1) {
1796 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1797 perf_evsel__name(evsel));
1799 nr = evsel->nr_members - 1;
1801 fprintf(fp, ",%s", perf_evsel__name(evsel));
1809 static void print_sample_time(struct feat_fd *ff, FILE *fp)
1811 struct perf_session *session;
1815 session = container_of(ff->ph, struct perf_session, header);
1817 timestamp__scnprintf_usec(session->evlist->first_sample_time,
1818 time_buf, sizeof(time_buf));
1819 fprintf(fp, "# time of first sample : %s\n", time_buf);
1821 timestamp__scnprintf_usec(session->evlist->last_sample_time,
1822 time_buf, sizeof(time_buf));
1823 fprintf(fp, "# time of last sample : %s\n", time_buf);
1825 d = (double)(session->evlist->last_sample_time -
1826 session->evlist->first_sample_time) / NSEC_PER_MSEC;
1828 fprintf(fp, "# sample duration : %10.3f ms\n", d);
1831 static void memory_node__fprintf(struct memory_node *n,
1832 unsigned long long bsize, FILE *fp)
1834 char buf_map[100], buf_size[50];
1835 unsigned long long size;
1837 size = bsize * bitmap_weight(n->set, n->size);
1838 unit_number__scnprintf(buf_size, 50, size);
1840 bitmap_scnprintf(n->set, n->size, buf_map, 100);
1841 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map);
1844 static void print_mem_topology(struct feat_fd *ff, FILE *fp)
1846 struct memory_node *nodes;
1849 nodes = ff->ph->env.memory_nodes;
1850 nr = ff->ph->env.nr_memory_nodes;
1852 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n",
1853 nr, ff->ph->env.memory_bsize);
1855 for (i = 0; i < nr; i++) {
1856 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp);
1860 static int __event_process_build_id(struct build_id_event *bev,
1862 struct perf_session *session)
1865 struct machine *machine;
1868 enum dso_kernel_type dso_type;
1870 machine = perf_session__findnew_machine(session, bev->pid);
1874 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1877 case PERF_RECORD_MISC_KERNEL:
1878 dso_type = DSO_TYPE_KERNEL;
1880 case PERF_RECORD_MISC_GUEST_KERNEL:
1881 dso_type = DSO_TYPE_GUEST_KERNEL;
1883 case PERF_RECORD_MISC_USER:
1884 case PERF_RECORD_MISC_GUEST_USER:
1885 dso_type = DSO_TYPE_USER;
1891 dso = machine__findnew_dso(machine, filename);
1893 char sbuild_id[SBUILD_ID_SIZE];
1895 dso__set_build_id(dso, &bev->build_id);
1897 if (dso_type != DSO_TYPE_USER) {
1898 struct kmod_path m = { .name = NULL, };
1900 if (!kmod_path__parse_name(&m, filename) && m.kmod)
1901 dso__set_module_info(dso, &m, machine);
1903 dso->kernel = dso_type;
1908 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1910 pr_debug("build id event received for %s: %s\n",
1911 dso->long_name, sbuild_id);
1920 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1921 int input, u64 offset, u64 size)
1923 struct perf_session *session = container_of(header, struct perf_session, header);
1925 struct perf_event_header header;
1926 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1929 struct build_id_event bev;
1930 char filename[PATH_MAX];
1931 u64 limit = offset + size;
1933 while (offset < limit) {
1936 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1939 if (header->needs_swap)
1940 perf_event_header__bswap(&old_bev.header);
1942 len = old_bev.header.size - sizeof(old_bev);
1943 if (readn(input, filename, len) != len)
1946 bev.header = old_bev.header;
1949 * As the pid is the missing value, we need to fill
1950 * it properly. The header.misc value give us nice hint.
1952 bev.pid = HOST_KERNEL_ID;
1953 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1954 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1955 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1957 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1958 __event_process_build_id(&bev, filename, session);
1960 offset += bev.header.size;
1966 static int perf_header__read_build_ids(struct perf_header *header,
1967 int input, u64 offset, u64 size)
1969 struct perf_session *session = container_of(header, struct perf_session, header);
1970 struct build_id_event bev;
1971 char filename[PATH_MAX];
1972 u64 limit = offset + size, orig_offset = offset;
1975 while (offset < limit) {
1978 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1981 if (header->needs_swap)
1982 perf_event_header__bswap(&bev.header);
1984 len = bev.header.size - sizeof(bev);
1985 if (readn(input, filename, len) != len)
1988 * The a1645ce1 changeset:
1990 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1992 * Added a field to struct build_id_event that broke the file
1995 * Since the kernel build-id is the first entry, process the
1996 * table using the old format if the well known
1997 * '[kernel.kallsyms]' string for the kernel build-id has the
1998 * first 4 characters chopped off (where the pid_t sits).
2000 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
2001 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
2003 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
2006 __event_process_build_id(&bev, filename, session);
2008 offset += bev.header.size;
2015 /* Macro for features that simply need to read and store a string. */
2016 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
2017 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
2019 ff->ph->env.__feat_env = do_read_string(ff); \
2020 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
2023 FEAT_PROCESS_STR_FUN(hostname, hostname);
2024 FEAT_PROCESS_STR_FUN(osrelease, os_release);
2025 FEAT_PROCESS_STR_FUN(version, version);
2026 FEAT_PROCESS_STR_FUN(arch, arch);
2027 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
2028 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
2030 static int process_tracing_data(struct feat_fd *ff, void *data)
2032 ssize_t ret = trace_report(ff->fd, data, false);
2034 return ret < 0 ? -1 : 0;
2037 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
2039 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
2040 pr_debug("Failed to read buildids, continuing...\n");
2044 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
2047 u32 nr_cpus_avail, nr_cpus_online;
2049 ret = do_read_u32(ff, &nr_cpus_avail);
2053 ret = do_read_u32(ff, &nr_cpus_online);
2056 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
2057 ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
2061 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
2066 ret = do_read_u64(ff, &total_mem);
2069 ff->ph->env.total_mem = (unsigned long long)total_mem;
2073 static struct perf_evsel *
2074 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
2076 struct perf_evsel *evsel;
2078 evlist__for_each_entry(evlist, evsel) {
2079 if (evsel->idx == idx)
2087 perf_evlist__set_event_name(struct perf_evlist *evlist,
2088 struct perf_evsel *event)
2090 struct perf_evsel *evsel;
2095 evsel = perf_evlist__find_by_index(evlist, event->idx);
2102 evsel->name = strdup(event->name);
2106 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
2108 struct perf_session *session;
2109 struct perf_evsel *evsel, *events = read_event_desc(ff);
2114 session = container_of(ff->ph, struct perf_session, header);
2116 if (session->data->is_pipe) {
2117 /* Save events for reading later by print_event_desc,
2118 * since they can't be read again in pipe mode. */
2119 ff->events = events;
2122 for (evsel = events; evsel->attr.size; evsel++)
2123 perf_evlist__set_event_name(session->evlist, evsel);
2125 if (!session->data->is_pipe)
2126 free_event_desc(events);
2131 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
2133 char *str, *cmdline = NULL, **argv = NULL;
2136 if (do_read_u32(ff, &nr))
2139 ff->ph->env.nr_cmdline = nr;
2141 cmdline = zalloc(ff->size + nr + 1);
2145 argv = zalloc(sizeof(char *) * (nr + 1));
2149 for (i = 0; i < nr; i++) {
2150 str = do_read_string(ff);
2154 argv[i] = cmdline + len;
2155 memcpy(argv[i], str, strlen(str) + 1);
2156 len += strlen(str) + 1;
2159 ff->ph->env.cmdline = cmdline;
2160 ff->ph->env.cmdline_argv = (const char **) argv;
2169 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
2174 int cpu_nr = ff->ph->env.nr_cpus_avail;
2176 struct perf_header *ph = ff->ph;
2177 bool do_core_id_test = true;
2179 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
2183 if (do_read_u32(ff, &nr))
2186 ph->env.nr_sibling_cores = nr;
2187 size += sizeof(u32);
2188 if (strbuf_init(&sb, 128) < 0)
2191 for (i = 0; i < nr; i++) {
2192 str = do_read_string(ff);
2196 /* include a NULL character at the end */
2197 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2199 size += string_size(str);
2202 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
2204 if (do_read_u32(ff, &nr))
2207 ph->env.nr_sibling_threads = nr;
2208 size += sizeof(u32);
2210 for (i = 0; i < nr; i++) {
2211 str = do_read_string(ff);
2215 /* include a NULL character at the end */
2216 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2218 size += string_size(str);
2221 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
2224 * The header may be from old perf,
2225 * which doesn't include core id and socket id information.
2227 if (ff->size <= size) {
2228 zfree(&ph->env.cpu);
2232 /* On s390 the socket_id number is not related to the numbers of cpus.
2233 * The socket_id number might be higher than the numbers of cpus.
2234 * This depends on the configuration.
2236 if (ph->env.arch && !strncmp(ph->env.arch, "s390", 4))
2237 do_core_id_test = false;
2239 for (i = 0; i < (u32)cpu_nr; i++) {
2240 if (do_read_u32(ff, &nr))
2243 ph->env.cpu[i].core_id = nr;
2245 if (do_read_u32(ff, &nr))
2248 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
2249 pr_debug("socket_id number is too big."
2250 "You may need to upgrade the perf tool.\n");
2254 ph->env.cpu[i].socket_id = nr;
2260 strbuf_release(&sb);
2262 zfree(&ph->env.cpu);
2266 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
2268 struct numa_node *nodes, *n;
2273 if (do_read_u32(ff, &nr))
2276 nodes = zalloc(sizeof(*nodes) * nr);
2280 for (i = 0; i < nr; i++) {
2284 if (do_read_u32(ff, &n->node))
2287 if (do_read_u64(ff, &n->mem_total))
2290 if (do_read_u64(ff, &n->mem_free))
2293 str = do_read_string(ff);
2297 n->map = cpu_map__new(str);
2303 ff->ph->env.nr_numa_nodes = nr;
2304 ff->ph->env.numa_nodes = nodes;
2312 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
2319 if (do_read_u32(ff, &pmu_num))
2323 pr_debug("pmu mappings not available\n");
2327 ff->ph->env.nr_pmu_mappings = pmu_num;
2328 if (strbuf_init(&sb, 128) < 0)
2332 if (do_read_u32(ff, &type))
2335 name = do_read_string(ff);
2339 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
2341 /* include a NULL character at the end */
2342 if (strbuf_add(&sb, "", 1) < 0)
2345 if (!strcmp(name, "msr"))
2346 ff->ph->env.msr_pmu_type = type;
2351 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2355 strbuf_release(&sb);
2359 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2362 u32 i, nr, nr_groups;
2363 struct perf_session *session;
2364 struct perf_evsel *evsel, *leader = NULL;
2371 if (do_read_u32(ff, &nr_groups))
2374 ff->ph->env.nr_groups = nr_groups;
2376 pr_debug("group desc not available\n");
2380 desc = calloc(nr_groups, sizeof(*desc));
2384 for (i = 0; i < nr_groups; i++) {
2385 desc[i].name = do_read_string(ff);
2389 if (do_read_u32(ff, &desc[i].leader_idx))
2392 if (do_read_u32(ff, &desc[i].nr_members))
2397 * Rebuild group relationship based on the group_desc
2399 session = container_of(ff->ph, struct perf_session, header);
2400 session->evlist->nr_groups = nr_groups;
2403 evlist__for_each_entry(session->evlist, evsel) {
2404 if (evsel->idx == (int) desc[i].leader_idx) {
2405 evsel->leader = evsel;
2406 /* {anon_group} is a dummy name */
2407 if (strcmp(desc[i].name, "{anon_group}")) {
2408 evsel->group_name = desc[i].name;
2409 desc[i].name = NULL;
2411 evsel->nr_members = desc[i].nr_members;
2413 if (i >= nr_groups || nr > 0) {
2414 pr_debug("invalid group desc\n");
2419 nr = evsel->nr_members - 1;
2422 /* This is a group member */
2423 evsel->leader = leader;
2429 if (i != nr_groups || nr != 0) {
2430 pr_debug("invalid group desc\n");
2436 for (i = 0; i < nr_groups; i++)
2437 zfree(&desc[i].name);
2443 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2445 struct perf_session *session;
2448 session = container_of(ff->ph, struct perf_session, header);
2450 err = auxtrace_index__process(ff->fd, ff->size, session,
2451 ff->ph->needs_swap);
2453 pr_err("Failed to process auxtrace index\n");
2457 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2459 struct cpu_cache_level *caches;
2460 u32 cnt, i, version;
2462 if (do_read_u32(ff, &version))
2468 if (do_read_u32(ff, &cnt))
2471 caches = zalloc(sizeof(*caches) * cnt);
2475 for (i = 0; i < cnt; i++) {
2476 struct cpu_cache_level c;
2479 if (do_read_u32(ff, &c.v))\
2480 goto out_free_caches; \
2489 c.v = do_read_string(ff); \
2491 goto out_free_caches;
2501 ff->ph->env.caches = caches;
2502 ff->ph->env.caches_cnt = cnt;
2509 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused)
2511 struct perf_session *session;
2512 u64 first_sample_time, last_sample_time;
2515 session = container_of(ff->ph, struct perf_session, header);
2517 ret = do_read_u64(ff, &first_sample_time);
2521 ret = do_read_u64(ff, &last_sample_time);
2525 session->evlist->first_sample_time = first_sample_time;
2526 session->evlist->last_sample_time = last_sample_time;
2530 static int process_mem_topology(struct feat_fd *ff,
2531 void *data __maybe_unused)
2533 struct memory_node *nodes;
2534 u64 version, i, nr, bsize;
2537 if (do_read_u64(ff, &version))
2543 if (do_read_u64(ff, &bsize))
2546 if (do_read_u64(ff, &nr))
2549 nodes = zalloc(sizeof(*nodes) * nr);
2553 for (i = 0; i < nr; i++) {
2554 struct memory_node n;
2557 if (do_read_u64(ff, &n.v)) \
2565 if (do_read_bitmap(ff, &n.set, &n.size))
2571 ff->ph->env.memory_bsize = bsize;
2572 ff->ph->env.memory_nodes = nodes;
2573 ff->ph->env.nr_memory_nodes = nr;
2582 static int process_clockid(struct feat_fd *ff,
2583 void *data __maybe_unused)
2585 if (do_read_u64(ff, &ff->ph->env.clockid_res_ns))
2591 struct feature_ops {
2592 int (*write)(struct feat_fd *ff, struct perf_evlist *evlist);
2593 void (*print)(struct feat_fd *ff, FILE *fp);
2594 int (*process)(struct feat_fd *ff, void *data);
2600 #define FEAT_OPR(n, func, __full_only) \
2602 .name = __stringify(n), \
2603 .write = write_##func, \
2604 .print = print_##func, \
2605 .full_only = __full_only, \
2606 .process = process_##func, \
2607 .synthesize = true \
2610 #define FEAT_OPN(n, func, __full_only) \
2612 .name = __stringify(n), \
2613 .write = write_##func, \
2614 .print = print_##func, \
2615 .full_only = __full_only, \
2616 .process = process_##func \
2619 /* feature_ops not implemented: */
2620 #define print_tracing_data NULL
2621 #define print_build_id NULL
2623 #define process_branch_stack NULL
2624 #define process_stat NULL
2627 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2628 FEAT_OPN(TRACING_DATA, tracing_data, false),
2629 FEAT_OPN(BUILD_ID, build_id, false),
2630 FEAT_OPR(HOSTNAME, hostname, false),
2631 FEAT_OPR(OSRELEASE, osrelease, false),
2632 FEAT_OPR(VERSION, version, false),
2633 FEAT_OPR(ARCH, arch, false),
2634 FEAT_OPR(NRCPUS, nrcpus, false),
2635 FEAT_OPR(CPUDESC, cpudesc, false),
2636 FEAT_OPR(CPUID, cpuid, false),
2637 FEAT_OPR(TOTAL_MEM, total_mem, false),
2638 FEAT_OPR(EVENT_DESC, event_desc, false),
2639 FEAT_OPR(CMDLINE, cmdline, false),
2640 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true),
2641 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true),
2642 FEAT_OPN(BRANCH_STACK, branch_stack, false),
2643 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false),
2644 FEAT_OPR(GROUP_DESC, group_desc, false),
2645 FEAT_OPN(AUXTRACE, auxtrace, false),
2646 FEAT_OPN(STAT, stat, false),
2647 FEAT_OPN(CACHE, cache, true),
2648 FEAT_OPR(SAMPLE_TIME, sample_time, false),
2649 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true),
2650 FEAT_OPR(CLOCKID, clockid, false)
2653 struct header_print_data {
2655 bool full; /* extended list of headers */
2658 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2659 struct perf_header *ph,
2660 int feat, int fd, void *data)
2662 struct header_print_data *hd = data;
2665 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2666 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2667 "%d, continuing...\n", section->offset, feat);
2670 if (feat >= HEADER_LAST_FEATURE) {
2671 pr_warning("unknown feature %d\n", feat);
2674 if (!feat_ops[feat].print)
2677 ff = (struct feat_fd) {
2682 if (!feat_ops[feat].full_only || hd->full)
2683 feat_ops[feat].print(&ff, hd->fp);
2685 fprintf(hd->fp, "# %s info available, use -I to display\n",
2686 feat_ops[feat].name);
2691 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2693 struct header_print_data hd;
2694 struct perf_header *header = &session->header;
2695 int fd = perf_data__fd(session->data);
2703 ret = fstat(fd, &st);
2707 stctime = st.st_ctime;
2708 fprintf(fp, "# captured on : %s", ctime(&stctime));
2710 fprintf(fp, "# header version : %u\n", header->version);
2711 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset);
2712 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size);
2713 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset);
2715 perf_header__process_sections(header, fd, &hd,
2716 perf_file_section__fprintf_info);
2718 if (session->data->is_pipe)
2721 fprintf(fp, "# missing features: ");
2722 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
2724 fprintf(fp, "%s ", feat_ops[bit].name);
2731 static int do_write_feat(struct feat_fd *ff, int type,
2732 struct perf_file_section **p,
2733 struct perf_evlist *evlist)
2738 if (perf_header__has_feat(ff->ph, type)) {
2739 if (!feat_ops[type].write)
2742 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
2745 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
2747 err = feat_ops[type].write(ff, evlist);
2749 pr_debug("failed to write feature %s\n", feat_ops[type].name);
2751 /* undo anything written */
2752 lseek(ff->fd, (*p)->offset, SEEK_SET);
2756 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
2762 static int perf_header__adds_write(struct perf_header *header,
2763 struct perf_evlist *evlist, int fd)
2767 struct perf_file_section *feat_sec, *p;
2773 ff = (struct feat_fd){
2778 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2782 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2783 if (feat_sec == NULL)
2786 sec_size = sizeof(*feat_sec) * nr_sections;
2788 sec_start = header->feat_offset;
2789 lseek(fd, sec_start + sec_size, SEEK_SET);
2791 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2792 if (do_write_feat(&ff, feat, &p, evlist))
2793 perf_header__clear_feat(header, feat);
2796 lseek(fd, sec_start, SEEK_SET);
2798 * may write more than needed due to dropped feature, but
2799 * this is okay, reader will skip the missing entries
2801 err = do_write(&ff, feat_sec, sec_size);
2803 pr_debug("failed to write feature section\n");
2808 int perf_header__write_pipe(int fd)
2810 struct perf_pipe_file_header f_header;
2814 ff = (struct feat_fd){ .fd = fd };
2816 f_header = (struct perf_pipe_file_header){
2817 .magic = PERF_MAGIC,
2818 .size = sizeof(f_header),
2821 err = do_write(&ff, &f_header, sizeof(f_header));
2823 pr_debug("failed to write perf pipe header\n");
2830 int perf_session__write_header(struct perf_session *session,
2831 struct perf_evlist *evlist,
2832 int fd, bool at_exit)
2834 struct perf_file_header f_header;
2835 struct perf_file_attr f_attr;
2836 struct perf_header *header = &session->header;
2837 struct perf_evsel *evsel;
2842 ff = (struct feat_fd){ .fd = fd};
2843 lseek(fd, sizeof(f_header), SEEK_SET);
2845 evlist__for_each_entry(session->evlist, evsel) {
2846 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2847 err = do_write(&ff, evsel->id, evsel->ids * sizeof(u64));
2849 pr_debug("failed to write perf header\n");
2854 attr_offset = lseek(ff.fd, 0, SEEK_CUR);
2856 evlist__for_each_entry(evlist, evsel) {
2857 f_attr = (struct perf_file_attr){
2858 .attr = evsel->attr,
2860 .offset = evsel->id_offset,
2861 .size = evsel->ids * sizeof(u64),
2864 err = do_write(&ff, &f_attr, sizeof(f_attr));
2866 pr_debug("failed to write perf header attribute\n");
2871 if (!header->data_offset)
2872 header->data_offset = lseek(fd, 0, SEEK_CUR);
2873 header->feat_offset = header->data_offset + header->data_size;
2876 err = perf_header__adds_write(header, evlist, fd);
2881 f_header = (struct perf_file_header){
2882 .magic = PERF_MAGIC,
2883 .size = sizeof(f_header),
2884 .attr_size = sizeof(f_attr),
2886 .offset = attr_offset,
2887 .size = evlist->nr_entries * sizeof(f_attr),
2890 .offset = header->data_offset,
2891 .size = header->data_size,
2893 /* event_types is ignored, store zeros */
2896 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2898 lseek(fd, 0, SEEK_SET);
2899 err = do_write(&ff, &f_header, sizeof(f_header));
2901 pr_debug("failed to write perf header\n");
2904 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2909 static int perf_header__getbuffer64(struct perf_header *header,
2910 int fd, void *buf, size_t size)
2912 if (readn(fd, buf, size) <= 0)
2915 if (header->needs_swap)
2916 mem_bswap_64(buf, size);
2921 int perf_header__process_sections(struct perf_header *header, int fd,
2923 int (*process)(struct perf_file_section *section,
2924 struct perf_header *ph,
2925 int feat, int fd, void *data))
2927 struct perf_file_section *feat_sec, *sec;
2933 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2937 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2941 sec_size = sizeof(*feat_sec) * nr_sections;
2943 lseek(fd, header->feat_offset, SEEK_SET);
2945 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2949 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2950 err = process(sec++, header, feat, fd, data);
2960 static const int attr_file_abi_sizes[] = {
2961 [0] = PERF_ATTR_SIZE_VER0,
2962 [1] = PERF_ATTR_SIZE_VER1,
2963 [2] = PERF_ATTR_SIZE_VER2,
2964 [3] = PERF_ATTR_SIZE_VER3,
2965 [4] = PERF_ATTR_SIZE_VER4,
2970 * In the legacy file format, the magic number is not used to encode endianness.
2971 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2972 * on ABI revisions, we need to try all combinations for all endianness to
2973 * detect the endianness.
2975 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2977 uint64_t ref_size, attr_size;
2980 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2981 ref_size = attr_file_abi_sizes[i]
2982 + sizeof(struct perf_file_section);
2983 if (hdr_sz != ref_size) {
2984 attr_size = bswap_64(hdr_sz);
2985 if (attr_size != ref_size)
2988 ph->needs_swap = true;
2990 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2995 /* could not determine endianness */
2999 #define PERF_PIPE_HDR_VER0 16
3001 static const size_t attr_pipe_abi_sizes[] = {
3002 [0] = PERF_PIPE_HDR_VER0,
3007 * In the legacy pipe format, there is an implicit assumption that endiannesss
3008 * between host recording the samples, and host parsing the samples is the
3009 * same. This is not always the case given that the pipe output may always be
3010 * redirected into a file and analyzed on a different machine with possibly a
3011 * different endianness and perf_event ABI revsions in the perf tool itself.
3013 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
3018 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
3019 if (hdr_sz != attr_pipe_abi_sizes[i]) {
3020 attr_size = bswap_64(hdr_sz);
3021 if (attr_size != hdr_sz)
3024 ph->needs_swap = true;
3026 pr_debug("Pipe ABI%d perf.data file detected\n", i);
3032 bool is_perf_magic(u64 magic)
3034 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
3035 || magic == __perf_magic2
3036 || magic == __perf_magic2_sw)
3042 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
3043 bool is_pipe, struct perf_header *ph)
3047 /* check for legacy format */
3048 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
3050 ph->version = PERF_HEADER_VERSION_1;
3051 pr_debug("legacy perf.data format\n");
3053 return try_all_pipe_abis(hdr_sz, ph);
3055 return try_all_file_abis(hdr_sz, ph);
3058 * the new magic number serves two purposes:
3059 * - unique number to identify actual perf.data files
3060 * - encode endianness of file
3062 ph->version = PERF_HEADER_VERSION_2;
3064 /* check magic number with one endianness */
3065 if (magic == __perf_magic2)
3068 /* check magic number with opposite endianness */
3069 if (magic != __perf_magic2_sw)
3072 ph->needs_swap = true;
3077 int perf_file_header__read(struct perf_file_header *header,
3078 struct perf_header *ph, int fd)
3082 lseek(fd, 0, SEEK_SET);
3084 ret = readn(fd, header, sizeof(*header));
3088 if (check_magic_endian(header->magic,
3089 header->attr_size, false, ph) < 0) {
3090 pr_debug("magic/endian check failed\n");
3094 if (ph->needs_swap) {
3095 mem_bswap_64(header, offsetof(struct perf_file_header,
3099 if (header->size != sizeof(*header)) {
3100 /* Support the previous format */
3101 if (header->size == offsetof(typeof(*header), adds_features))
3102 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3105 } else if (ph->needs_swap) {
3107 * feature bitmap is declared as an array of unsigned longs --
3108 * not good since its size can differ between the host that
3109 * generated the data file and the host analyzing the file.
3111 * We need to handle endianness, but we don't know the size of
3112 * the unsigned long where the file was generated. Take a best
3113 * guess at determining it: try 64-bit swap first (ie., file
3114 * created on a 64-bit host), and check if the hostname feature
3115 * bit is set (this feature bit is forced on as of fbe96f2).
3116 * If the bit is not, undo the 64-bit swap and try a 32-bit
3117 * swap. If the hostname bit is still not set (e.g., older data
3118 * file), punt and fallback to the original behavior --
3119 * clearing all feature bits and setting buildid.
3121 mem_bswap_64(&header->adds_features,
3122 BITS_TO_U64(HEADER_FEAT_BITS));
3124 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3126 mem_bswap_64(&header->adds_features,
3127 BITS_TO_U64(HEADER_FEAT_BITS));
3130 mem_bswap_32(&header->adds_features,
3131 BITS_TO_U32(HEADER_FEAT_BITS));
3134 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3135 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3136 set_bit(HEADER_BUILD_ID, header->adds_features);
3140 memcpy(&ph->adds_features, &header->adds_features,
3141 sizeof(ph->adds_features));
3143 ph->data_offset = header->data.offset;
3144 ph->data_size = header->data.size;
3145 ph->feat_offset = header->data.offset + header->data.size;
3149 static int perf_file_section__process(struct perf_file_section *section,
3150 struct perf_header *ph,
3151 int feat, int fd, void *data)
3153 struct feat_fd fdd = {
3156 .size = section->size,
3157 .offset = section->offset,
3160 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3161 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3162 "%d, continuing...\n", section->offset, feat);
3166 if (feat >= HEADER_LAST_FEATURE) {
3167 pr_debug("unknown feature %d, continuing...\n", feat);
3171 if (!feat_ops[feat].process)
3174 return feat_ops[feat].process(&fdd, data);
3177 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
3178 struct perf_header *ph, int fd,
3181 struct feat_fd ff = {
3182 .fd = STDOUT_FILENO,
3187 ret = readn(fd, header, sizeof(*header));
3191 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
3192 pr_debug("endian/magic failed\n");
3197 header->size = bswap_64(header->size);
3199 if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
3205 static int perf_header__read_pipe(struct perf_session *session)
3207 struct perf_header *header = &session->header;
3208 struct perf_pipe_file_header f_header;
3210 if (perf_file_header__read_pipe(&f_header, header,
3211 perf_data__fd(session->data),
3212 session->repipe) < 0) {
3213 pr_debug("incompatible file format\n");
3220 static int read_attr(int fd, struct perf_header *ph,
3221 struct perf_file_attr *f_attr)
3223 struct perf_event_attr *attr = &f_attr->attr;
3225 size_t our_sz = sizeof(f_attr->attr);
3228 memset(f_attr, 0, sizeof(*f_attr));
3230 /* read minimal guaranteed structure */
3231 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
3233 pr_debug("cannot read %d bytes of header attr\n",
3234 PERF_ATTR_SIZE_VER0);
3238 /* on file perf_event_attr size */
3246 sz = PERF_ATTR_SIZE_VER0;
3247 } else if (sz > our_sz) {
3248 pr_debug("file uses a more recent and unsupported ABI"
3249 " (%zu bytes extra)\n", sz - our_sz);
3252 /* what we have not yet read and that we know about */
3253 left = sz - PERF_ATTR_SIZE_VER0;
3256 ptr += PERF_ATTR_SIZE_VER0;
3258 ret = readn(fd, ptr, left);
3260 /* read perf_file_section, ids are read in caller */
3261 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
3263 return ret <= 0 ? -1 : 0;
3266 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
3267 struct tep_handle *pevent)
3269 struct tep_event *event;
3272 /* already prepared */
3273 if (evsel->tp_format)
3276 if (pevent == NULL) {
3277 pr_debug("broken or missing trace data\n");
3281 event = tep_find_event(pevent, evsel->attr.config);
3282 if (event == NULL) {
3283 pr_debug("cannot find event format for %d\n", (int)evsel->attr.config);
3288 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
3289 evsel->name = strdup(bf);
3290 if (evsel->name == NULL)
3294 evsel->tp_format = event;
3298 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
3299 struct tep_handle *pevent)
3301 struct perf_evsel *pos;
3303 evlist__for_each_entry(evlist, pos) {
3304 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
3305 perf_evsel__prepare_tracepoint_event(pos, pevent))
3312 int perf_session__read_header(struct perf_session *session)
3314 struct perf_data *data = session->data;
3315 struct perf_header *header = &session->header;
3316 struct perf_file_header f_header;
3317 struct perf_file_attr f_attr;
3319 int nr_attrs, nr_ids, i, j;
3320 int fd = perf_data__fd(data);
3322 session->evlist = perf_evlist__new();
3323 if (session->evlist == NULL)
3326 session->evlist->env = &header->env;
3327 session->machines.host.env = &header->env;
3328 if (perf_data__is_pipe(data))
3329 return perf_header__read_pipe(session);
3331 if (perf_file_header__read(&f_header, header, fd) < 0)
3335 * Sanity check that perf.data was written cleanly; data size is
3336 * initialized to 0 and updated only if the on_exit function is run.
3337 * If data size is still 0 then the file contains only partial
3338 * information. Just warn user and process it as much as it can.
3340 if (f_header.data.size == 0) {
3341 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3342 "Was the 'perf record' command properly terminated?\n",
3346 nr_attrs = f_header.attrs.size / f_header.attr_size;
3347 lseek(fd, f_header.attrs.offset, SEEK_SET);
3349 for (i = 0; i < nr_attrs; i++) {
3350 struct perf_evsel *evsel;
3353 if (read_attr(fd, header, &f_attr) < 0)
3356 if (header->needs_swap) {
3357 f_attr.ids.size = bswap_64(f_attr.ids.size);
3358 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
3359 perf_event__attr_swap(&f_attr.attr);
3362 tmp = lseek(fd, 0, SEEK_CUR);
3363 evsel = perf_evsel__new(&f_attr.attr);
3366 goto out_delete_evlist;
3368 evsel->needs_swap = header->needs_swap;
3370 * Do it before so that if perf_evsel__alloc_id fails, this
3371 * entry gets purged too at perf_evlist__delete().
3373 perf_evlist__add(session->evlist, evsel);
3375 nr_ids = f_attr.ids.size / sizeof(u64);
3377 * We don't have the cpu and thread maps on the header, so
3378 * for allocating the perf_sample_id table we fake 1 cpu and
3379 * hattr->ids threads.
3381 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
3382 goto out_delete_evlist;
3384 lseek(fd, f_attr.ids.offset, SEEK_SET);
3386 for (j = 0; j < nr_ids; j++) {
3387 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
3390 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
3393 lseek(fd, tmp, SEEK_SET);
3396 perf_header__process_sections(header, fd, &session->tevent,
3397 perf_file_section__process);
3399 if (perf_evlist__prepare_tracepoint_events(session->evlist,
3400 session->tevent.pevent))
3401 goto out_delete_evlist;
3408 perf_evlist__delete(session->evlist);
3409 session->evlist = NULL;
3413 int perf_event__synthesize_attr(struct perf_tool *tool,
3414 struct perf_event_attr *attr, u32 ids, u64 *id,
3415 perf_event__handler_t process)
3417 union perf_event *ev;
3421 size = sizeof(struct perf_event_attr);
3422 size = PERF_ALIGN(size, sizeof(u64));
3423 size += sizeof(struct perf_event_header);
3424 size += ids * sizeof(u64);
3431 ev->attr.attr = *attr;
3432 memcpy(ev->attr.id, id, ids * sizeof(u64));
3434 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
3435 ev->attr.header.size = (u16)size;
3437 if (ev->attr.header.size == size)
3438 err = process(tool, ev, NULL, NULL);
3447 int perf_event__synthesize_features(struct perf_tool *tool,
3448 struct perf_session *session,
3449 struct perf_evlist *evlist,
3450 perf_event__handler_t process)
3452 struct perf_header *header = &session->header;
3454 struct feature_event *fe;
3458 sz_hdr = sizeof(fe->header);
3459 sz = sizeof(union perf_event);
3460 /* get a nice alignment */
3461 sz = PERF_ALIGN(sz, page_size);
3463 memset(&ff, 0, sizeof(ff));
3465 ff.buf = malloc(sz);
3469 ff.size = sz - sz_hdr;
3471 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3472 if (!feat_ops[feat].synthesize) {
3473 pr_debug("No record header feature for header :%d\n", feat);
3477 ff.offset = sizeof(*fe);
3479 ret = feat_ops[feat].write(&ff, evlist);
3480 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
3481 pr_debug("Error writing feature\n");
3484 /* ff.buf may have changed due to realloc in do_write() */
3486 memset(fe, 0, sizeof(*fe));
3489 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3490 fe->header.size = ff.offset;
3492 ret = process(tool, ff.buf, NULL, NULL);
3499 /* Send HEADER_LAST_FEATURE mark. */
3501 fe->feat_id = HEADER_LAST_FEATURE;
3502 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3503 fe->header.size = sizeof(*fe);
3505 ret = process(tool, ff.buf, NULL, NULL);
3511 int perf_event__process_feature(struct perf_session *session,
3512 union perf_event *event)
3514 struct perf_tool *tool = session->tool;
3515 struct feat_fd ff = { .fd = 0 };
3516 struct feature_event *fe = (struct feature_event *)event;
3517 int type = fe->header.type;
3518 u64 feat = fe->feat_id;
3520 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3521 pr_warning("invalid record type %d in pipe-mode\n", type);
3524 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) {
3525 pr_warning("invalid record type %d in pipe-mode\n", type);
3529 if (!feat_ops[feat].process)
3532 ff.buf = (void *)fe->data;
3533 ff.size = event->header.size - sizeof(event->header);
3534 ff.ph = &session->header;
3536 if (feat_ops[feat].process(&ff, NULL))
3539 if (!feat_ops[feat].print || !tool->show_feat_hdr)
3542 if (!feat_ops[feat].full_only ||
3543 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3544 feat_ops[feat].print(&ff, stdout);
3546 fprintf(stdout, "# %s info available, use -I to display\n",
3547 feat_ops[feat].name);
3553 static struct event_update_event *
3554 event_update_event__new(size_t size, u64 type, u64 id)
3556 struct event_update_event *ev;
3558 size += sizeof(*ev);
3559 size = PERF_ALIGN(size, sizeof(u64));
3563 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3564 ev->header.size = (u16)size;
3572 perf_event__synthesize_event_update_unit(struct perf_tool *tool,
3573 struct perf_evsel *evsel,
3574 perf_event__handler_t process)
3576 struct event_update_event *ev;
3577 size_t size = strlen(evsel->unit);
3580 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->id[0]);
3584 strlcpy(ev->data, evsel->unit, size + 1);
3585 err = process(tool, (union perf_event *)ev, NULL, NULL);
3591 perf_event__synthesize_event_update_scale(struct perf_tool *tool,
3592 struct perf_evsel *evsel,
3593 perf_event__handler_t process)
3595 struct event_update_event *ev;
3596 struct event_update_event_scale *ev_data;
3599 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->id[0]);
3603 ev_data = (struct event_update_event_scale *) ev->data;
3604 ev_data->scale = evsel->scale;
3605 err = process(tool, (union perf_event*) ev, NULL, NULL);
3611 perf_event__synthesize_event_update_name(struct perf_tool *tool,
3612 struct perf_evsel *evsel,
3613 perf_event__handler_t process)
3615 struct event_update_event *ev;
3616 size_t len = strlen(evsel->name);
3619 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->id[0]);
3623 strlcpy(ev->data, evsel->name, len + 1);
3624 err = process(tool, (union perf_event*) ev, NULL, NULL);
3630 perf_event__synthesize_event_update_cpus(struct perf_tool *tool,
3631 struct perf_evsel *evsel,
3632 perf_event__handler_t process)
3634 size_t size = sizeof(struct event_update_event);
3635 struct event_update_event *ev;
3639 if (!evsel->own_cpus)
3642 ev = cpu_map_data__alloc(evsel->own_cpus, &size, &type, &max);
3646 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3647 ev->header.size = (u16)size;
3648 ev->type = PERF_EVENT_UPDATE__CPUS;
3649 ev->id = evsel->id[0];
3651 cpu_map_data__synthesize((struct cpu_map_data *) ev->data,
3655 err = process(tool, (union perf_event*) ev, NULL, NULL);
3660 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3662 struct event_update_event *ev = &event->event_update;
3663 struct event_update_event_scale *ev_scale;
3664 struct event_update_event_cpus *ev_cpus;
3665 struct cpu_map *map;
3668 ret = fprintf(fp, "\n... id: %" PRIu64 "\n", ev->id);
3671 case PERF_EVENT_UPDATE__SCALE:
3672 ev_scale = (struct event_update_event_scale *) ev->data;
3673 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3675 case PERF_EVENT_UPDATE__UNIT:
3676 ret += fprintf(fp, "... unit: %s\n", ev->data);
3678 case PERF_EVENT_UPDATE__NAME:
3679 ret += fprintf(fp, "... name: %s\n", ev->data);
3681 case PERF_EVENT_UPDATE__CPUS:
3682 ev_cpus = (struct event_update_event_cpus *) ev->data;
3683 ret += fprintf(fp, "... ");
3685 map = cpu_map__new_data(&ev_cpus->cpus);
3687 ret += cpu_map__fprintf(map, fp);
3689 ret += fprintf(fp, "failed to get cpus\n");
3692 ret += fprintf(fp, "... unknown type\n");
3699 int perf_event__synthesize_attrs(struct perf_tool *tool,
3700 struct perf_evlist *evlist,
3701 perf_event__handler_t process)
3703 struct perf_evsel *evsel;
3706 evlist__for_each_entry(evlist, evsel) {
3707 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
3708 evsel->id, process);
3710 pr_debug("failed to create perf header attribute\n");
3718 static bool has_unit(struct perf_evsel *counter)
3720 return counter->unit && *counter->unit;
3723 static bool has_scale(struct perf_evsel *counter)
3725 return counter->scale != 1;
3728 int perf_event__synthesize_extra_attr(struct perf_tool *tool,
3729 struct perf_evlist *evsel_list,
3730 perf_event__handler_t process,
3733 struct perf_evsel *counter;
3737 * Synthesize other events stuff not carried within
3738 * attr event - unit, scale, name
3740 evlist__for_each_entry(evsel_list, counter) {
3741 if (!counter->supported)
3745 * Synthesize unit and scale only if it's defined.
3747 if (has_unit(counter)) {
3748 err = perf_event__synthesize_event_update_unit(tool, counter, process);
3750 pr_err("Couldn't synthesize evsel unit.\n");
3755 if (has_scale(counter)) {
3756 err = perf_event__synthesize_event_update_scale(tool, counter, process);
3758 pr_err("Couldn't synthesize evsel counter.\n");
3763 if (counter->own_cpus) {
3764 err = perf_event__synthesize_event_update_cpus(tool, counter, process);
3766 pr_err("Couldn't synthesize evsel cpus.\n");
3772 * Name is needed only for pipe output,
3773 * perf.data carries event names.
3776 err = perf_event__synthesize_event_update_name(tool, counter, process);
3778 pr_err("Couldn't synthesize evsel name.\n");
3786 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3787 union perf_event *event,
3788 struct perf_evlist **pevlist)
3791 struct perf_evsel *evsel;
3792 struct perf_evlist *evlist = *pevlist;
3794 if (evlist == NULL) {
3795 *pevlist = evlist = perf_evlist__new();
3800 evsel = perf_evsel__new(&event->attr.attr);
3804 perf_evlist__add(evlist, evsel);
3806 ids = event->header.size;
3807 ids -= (void *)&event->attr.id - (void *)event;
3808 n_ids = ids / sizeof(u64);
3810 * We don't have the cpu and thread maps on the header, so
3811 * for allocating the perf_sample_id table we fake 1 cpu and
3812 * hattr->ids threads.
3814 if (perf_evsel__alloc_id(evsel, 1, n_ids))
3817 for (i = 0; i < n_ids; i++) {
3818 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
3824 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3825 union perf_event *event,
3826 struct perf_evlist **pevlist)
3828 struct event_update_event *ev = &event->event_update;
3829 struct event_update_event_scale *ev_scale;
3830 struct event_update_event_cpus *ev_cpus;
3831 struct perf_evlist *evlist;
3832 struct perf_evsel *evsel;
3833 struct cpu_map *map;
3835 if (!pevlist || *pevlist == NULL)
3840 evsel = perf_evlist__id2evsel(evlist, ev->id);
3845 case PERF_EVENT_UPDATE__UNIT:
3846 evsel->unit = strdup(ev->data);
3848 case PERF_EVENT_UPDATE__NAME:
3849 evsel->name = strdup(ev->data);
3851 case PERF_EVENT_UPDATE__SCALE:
3852 ev_scale = (struct event_update_event_scale *) ev->data;
3853 evsel->scale = ev_scale->scale;
3855 case PERF_EVENT_UPDATE__CPUS:
3856 ev_cpus = (struct event_update_event_cpus *) ev->data;
3858 map = cpu_map__new_data(&ev_cpus->cpus);
3860 evsel->own_cpus = map;
3862 pr_err("failed to get event_update cpus\n");
3870 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
3871 struct perf_evlist *evlist,
3872 perf_event__handler_t process)
3874 union perf_event ev;
3875 struct tracing_data *tdata;
3876 ssize_t size = 0, aligned_size = 0, padding;
3878 int err __maybe_unused = 0;
3881 * We are going to store the size of the data followed
3882 * by the data contents. Since the fd descriptor is a pipe,
3883 * we cannot seek back to store the size of the data once
3884 * we know it. Instead we:
3886 * - write the tracing data to the temp file
3887 * - get/write the data size to pipe
3888 * - write the tracing data from the temp file
3891 tdata = tracing_data_get(&evlist->entries, fd, true);
3895 memset(&ev, 0, sizeof(ev));
3897 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
3899 aligned_size = PERF_ALIGN(size, sizeof(u64));
3900 padding = aligned_size - size;
3901 ev.tracing_data.header.size = sizeof(ev.tracing_data);
3902 ev.tracing_data.size = aligned_size;
3904 process(tool, &ev, NULL, NULL);
3907 * The put function will copy all the tracing data
3908 * stored in temp file to the pipe.
3910 tracing_data_put(tdata);
3912 ff = (struct feat_fd){ .fd = fd };
3913 if (write_padded(&ff, NULL, 0, padding))
3916 return aligned_size;
3919 int perf_event__process_tracing_data(struct perf_session *session,
3920 union perf_event *event)
3922 ssize_t size_read, padding, size = event->tracing_data.size;
3923 int fd = perf_data__fd(session->data);
3924 off_t offset = lseek(fd, 0, SEEK_CUR);
3927 /* setup for reading amidst mmap */
3928 lseek(fd, offset + sizeof(struct tracing_data_event),
3931 size_read = trace_report(fd, &session->tevent,
3933 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3935 if (readn(fd, buf, padding) < 0) {
3936 pr_err("%s: reading input file", __func__);
3939 if (session->repipe) {
3940 int retw = write(STDOUT_FILENO, buf, padding);
3941 if (retw <= 0 || retw != padding) {
3942 pr_err("%s: repiping tracing data padding", __func__);
3947 if (size_read + padding != size) {
3948 pr_err("%s: tracing data size mismatch", __func__);
3952 perf_evlist__prepare_tracepoint_events(session->evlist,
3953 session->tevent.pevent);
3955 return size_read + padding;
3958 int perf_event__synthesize_build_id(struct perf_tool *tool,
3959 struct dso *pos, u16 misc,
3960 perf_event__handler_t process,
3961 struct machine *machine)
3963 union perf_event ev;
3970 memset(&ev, 0, sizeof(ev));
3972 len = pos->long_name_len + 1;
3973 len = PERF_ALIGN(len, NAME_ALIGN);
3974 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3975 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3976 ev.build_id.header.misc = misc;
3977 ev.build_id.pid = machine->pid;
3978 ev.build_id.header.size = sizeof(ev.build_id) + len;
3979 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3981 err = process(tool, &ev, NULL, machine);
3986 int perf_event__process_build_id(struct perf_session *session,
3987 union perf_event *event)
3989 __event_process_build_id(&event->build_id,
3990 event->build_id.filename,