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);
282 bitmap_zero(set, size);
286 for (i = 0; (u64) i < BITS_TO_U64(size); i++) {
287 ret = do_read_u64(ff, p + i);
299 static int write_tracing_data(struct feat_fd *ff,
300 struct perf_evlist *evlist)
302 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
305 return read_tracing_data(ff->fd, &evlist->entries);
308 static int write_build_id(struct feat_fd *ff,
309 struct perf_evlist *evlist __maybe_unused)
311 struct perf_session *session;
314 session = container_of(ff->ph, struct perf_session, header);
316 if (!perf_session__read_build_ids(session, true))
319 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
322 err = perf_session__write_buildid_table(session, ff);
324 pr_debug("failed to write buildid table\n");
327 perf_session__cache_build_ids(session);
332 static int write_hostname(struct feat_fd *ff,
333 struct perf_evlist *evlist __maybe_unused)
342 return do_write_string(ff, uts.nodename);
345 static int write_osrelease(struct feat_fd *ff,
346 struct perf_evlist *evlist __maybe_unused)
355 return do_write_string(ff, uts.release);
358 static int write_arch(struct feat_fd *ff,
359 struct perf_evlist *evlist __maybe_unused)
368 return do_write_string(ff, uts.machine);
371 static int write_version(struct feat_fd *ff,
372 struct perf_evlist *evlist __maybe_unused)
374 return do_write_string(ff, perf_version_string);
377 static int __write_cpudesc(struct feat_fd *ff, const char *cpuinfo_proc)
382 const char *search = cpuinfo_proc;
389 file = fopen("/proc/cpuinfo", "r");
393 while (getline(&buf, &len, file) > 0) {
394 ret = strncmp(buf, search, strlen(search));
406 p = strchr(buf, ':');
407 if (p && *(p+1) == ' ' && *(p+2))
413 /* squash extra space characters (branding string) */
420 while (*q && isspace(*q))
423 while ((*r++ = *q++));
427 ret = do_write_string(ff, s);
434 static int write_cpudesc(struct feat_fd *ff,
435 struct perf_evlist *evlist __maybe_unused)
437 const char *cpuinfo_procs[] = CPUINFO_PROC;
440 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
442 ret = __write_cpudesc(ff, cpuinfo_procs[i]);
450 static int write_nrcpus(struct feat_fd *ff,
451 struct perf_evlist *evlist __maybe_unused)
457 nrc = cpu__max_present_cpu();
459 nr = sysconf(_SC_NPROCESSORS_ONLN);
463 nra = (u32)(nr & UINT_MAX);
465 ret = do_write(ff, &nrc, sizeof(nrc));
469 return do_write(ff, &nra, sizeof(nra));
472 static int write_event_desc(struct feat_fd *ff,
473 struct perf_evlist *evlist)
475 struct perf_evsel *evsel;
479 nre = evlist->nr_entries;
482 * write number of events
484 ret = do_write(ff, &nre, sizeof(nre));
489 * size of perf_event_attr struct
491 sz = (u32)sizeof(evsel->attr);
492 ret = do_write(ff, &sz, sizeof(sz));
496 evlist__for_each_entry(evlist, evsel) {
497 ret = do_write(ff, &evsel->attr, sz);
501 * write number of unique id per event
502 * there is one id per instance of an event
504 * copy into an nri to be independent of the
508 ret = do_write(ff, &nri, sizeof(nri));
513 * write event string as passed on cmdline
515 ret = do_write_string(ff, perf_evsel__name(evsel));
519 * write unique ids for this event
521 ret = do_write(ff, evsel->id, evsel->ids * sizeof(u64));
528 static int write_cmdline(struct feat_fd *ff,
529 struct perf_evlist *evlist __maybe_unused)
531 char buf[MAXPATHLEN];
535 /* actual path to perf binary */
536 ret = readlink("/proc/self/exe", buf, sizeof(buf) - 1);
540 /* readlink() does not add null termination */
543 /* account for binary path */
544 n = perf_env.nr_cmdline + 1;
546 ret = do_write(ff, &n, sizeof(n));
550 ret = do_write_string(ff, buf);
554 for (i = 0 ; i < perf_env.nr_cmdline; i++) {
555 ret = do_write_string(ff, perf_env.cmdline_argv[i]);
562 #define CORE_SIB_FMT \
563 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
564 #define THRD_SIB_FMT \
565 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
571 char **core_siblings;
572 char **thread_siblings;
575 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
578 char filename[MAXPATHLEN];
579 char *buf = NULL, *p;
585 sprintf(filename, CORE_SIB_FMT, cpu);
586 fp = fopen(filename, "r");
590 sret = getline(&buf, &len, fp);
595 p = strchr(buf, '\n');
599 for (i = 0; i < tp->core_sib; i++) {
600 if (!strcmp(buf, tp->core_siblings[i]))
603 if (i == tp->core_sib) {
604 tp->core_siblings[i] = buf;
612 sprintf(filename, THRD_SIB_FMT, cpu);
613 fp = fopen(filename, "r");
617 if (getline(&buf, &len, fp) <= 0)
620 p = strchr(buf, '\n');
624 for (i = 0; i < tp->thread_sib; i++) {
625 if (!strcmp(buf, tp->thread_siblings[i]))
628 if (i == tp->thread_sib) {
629 tp->thread_siblings[i] = buf;
641 static void free_cpu_topo(struct cpu_topo *tp)
648 for (i = 0 ; i < tp->core_sib; i++)
649 zfree(&tp->core_siblings[i]);
651 for (i = 0 ; i < tp->thread_sib; i++)
652 zfree(&tp->thread_siblings[i]);
657 static struct cpu_topo *build_cpu_topology(void)
659 struct cpu_topo *tp = NULL;
667 ncpus = cpu__max_present_cpu();
669 /* build online CPU map */
670 map = cpu_map__new(NULL);
672 pr_debug("failed to get system cpumap\n");
676 nr = (u32)(ncpus & UINT_MAX);
678 sz = nr * sizeof(char *);
679 addr = calloc(1, sizeof(*tp) + 2 * sz);
686 tp->core_siblings = addr;
688 tp->thread_siblings = addr;
690 for (i = 0; i < nr; i++) {
691 if (!cpu_map__has(map, i))
694 ret = build_cpu_topo(tp, i);
708 static int write_cpu_topology(struct feat_fd *ff,
709 struct perf_evlist *evlist __maybe_unused)
715 tp = build_cpu_topology();
719 ret = do_write(ff, &tp->core_sib, sizeof(tp->core_sib));
723 for (i = 0; i < tp->core_sib; i++) {
724 ret = do_write_string(ff, tp->core_siblings[i]);
728 ret = do_write(ff, &tp->thread_sib, sizeof(tp->thread_sib));
732 for (i = 0; i < tp->thread_sib; i++) {
733 ret = do_write_string(ff, tp->thread_siblings[i]);
738 ret = perf_env__read_cpu_topology_map(&perf_env);
742 for (j = 0; j < perf_env.nr_cpus_avail; j++) {
743 ret = do_write(ff, &perf_env.cpu[j].core_id,
744 sizeof(perf_env.cpu[j].core_id));
747 ret = do_write(ff, &perf_env.cpu[j].socket_id,
748 sizeof(perf_env.cpu[j].socket_id));
759 static int write_total_mem(struct feat_fd *ff,
760 struct perf_evlist *evlist __maybe_unused)
768 fp = fopen("/proc/meminfo", "r");
772 while (getline(&buf, &len, fp) > 0) {
773 ret = strncmp(buf, "MemTotal:", 9);
778 n = sscanf(buf, "%*s %"PRIu64, &mem);
780 ret = do_write(ff, &mem, sizeof(mem));
788 static int write_topo_node(struct feat_fd *ff, int node)
790 char str[MAXPATHLEN];
792 char *buf = NULL, *p;
795 u64 mem_total, mem_free, mem;
798 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
799 fp = fopen(str, "r");
803 while (getline(&buf, &len, fp) > 0) {
804 /* skip over invalid lines */
805 if (!strchr(buf, ':'))
807 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
809 if (!strcmp(field, "MemTotal:"))
811 if (!strcmp(field, "MemFree:"))
818 ret = do_write(ff, &mem_total, sizeof(u64));
822 ret = do_write(ff, &mem_free, sizeof(u64));
827 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
829 fp = fopen(str, "r");
833 if (getline(&buf, &len, fp) <= 0)
836 p = strchr(buf, '\n');
840 ret = do_write_string(ff, buf);
848 static int write_numa_topology(struct feat_fd *ff,
849 struct perf_evlist *evlist __maybe_unused)
854 struct cpu_map *node_map = NULL;
859 fp = fopen("/sys/devices/system/node/online", "r");
863 if (getline(&buf, &len, fp) <= 0)
866 c = strchr(buf, '\n');
870 node_map = cpu_map__new(buf);
874 nr = (u32)node_map->nr;
876 ret = do_write(ff, &nr, sizeof(nr));
880 for (i = 0; i < nr; i++) {
881 j = (u32)node_map->map[i];
882 ret = do_write(ff, &j, sizeof(j));
886 ret = write_topo_node(ff, i);
893 cpu_map__put(node_map);
900 * struct pmu_mappings {
909 static int write_pmu_mappings(struct feat_fd *ff,
910 struct perf_evlist *evlist __maybe_unused)
912 struct perf_pmu *pmu = NULL;
917 * Do a first pass to count number of pmu to avoid lseek so this
918 * works in pipe mode as well.
920 while ((pmu = perf_pmu__scan(pmu))) {
926 ret = do_write(ff, &pmu_num, sizeof(pmu_num));
930 while ((pmu = perf_pmu__scan(pmu))) {
934 ret = do_write(ff, &pmu->type, sizeof(pmu->type));
938 ret = do_write_string(ff, pmu->name);
949 * struct group_descs {
951 * struct group_desc {
958 static int write_group_desc(struct feat_fd *ff,
959 struct perf_evlist *evlist)
961 u32 nr_groups = evlist->nr_groups;
962 struct perf_evsel *evsel;
965 ret = do_write(ff, &nr_groups, sizeof(nr_groups));
969 evlist__for_each_entry(evlist, evsel) {
970 if (perf_evsel__is_group_leader(evsel) &&
971 evsel->nr_members > 1) {
972 const char *name = evsel->group_name ?: "{anon_group}";
973 u32 leader_idx = evsel->idx;
974 u32 nr_members = evsel->nr_members;
976 ret = do_write_string(ff, name);
980 ret = do_write(ff, &leader_idx, sizeof(leader_idx));
984 ret = do_write(ff, &nr_members, sizeof(nr_members));
993 * default get_cpuid(): nothing gets recorded
994 * actual implementation must be in arch/$(SRCARCH)/util/header.c
996 int __weak get_cpuid(char *buffer __maybe_unused, size_t sz __maybe_unused)
1001 static int write_cpuid(struct feat_fd *ff,
1002 struct perf_evlist *evlist __maybe_unused)
1007 ret = get_cpuid(buffer, sizeof(buffer));
1013 return do_write_string(ff, buffer);
1016 static int write_branch_stack(struct feat_fd *ff __maybe_unused,
1017 struct perf_evlist *evlist __maybe_unused)
1022 static int write_auxtrace(struct feat_fd *ff,
1023 struct perf_evlist *evlist __maybe_unused)
1025 struct perf_session *session;
1028 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
1031 session = container_of(ff->ph, struct perf_session, header);
1033 err = auxtrace_index__write(ff->fd, &session->auxtrace_index);
1035 pr_err("Failed to write auxtrace index\n");
1039 static int cpu_cache_level__sort(const void *a, const void *b)
1041 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
1042 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
1044 return cache_a->level - cache_b->level;
1047 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
1049 if (a->level != b->level)
1052 if (a->line_size != b->line_size)
1055 if (a->sets != b->sets)
1058 if (a->ways != b->ways)
1061 if (strcmp(a->type, b->type))
1064 if (strcmp(a->size, b->size))
1067 if (strcmp(a->map, b->map))
1073 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
1075 char path[PATH_MAX], file[PATH_MAX];
1079 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
1080 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
1082 if (stat(file, &st))
1085 scnprintf(file, PATH_MAX, "%s/level", path);
1086 if (sysfs__read_int(file, (int *) &cache->level))
1089 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
1090 if (sysfs__read_int(file, (int *) &cache->line_size))
1093 scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
1094 if (sysfs__read_int(file, (int *) &cache->sets))
1097 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
1098 if (sysfs__read_int(file, (int *) &cache->ways))
1101 scnprintf(file, PATH_MAX, "%s/type", path);
1102 if (sysfs__read_str(file, &cache->type, &len))
1105 cache->type[len] = 0;
1106 cache->type = rtrim(cache->type);
1108 scnprintf(file, PATH_MAX, "%s/size", path);
1109 if (sysfs__read_str(file, &cache->size, &len)) {
1114 cache->size[len] = 0;
1115 cache->size = rtrim(cache->size);
1117 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
1118 if (sysfs__read_str(file, &cache->map, &len)) {
1124 cache->map[len] = 0;
1125 cache->map = rtrim(cache->map);
1129 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
1131 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
1134 static int build_caches(struct cpu_cache_level caches[], u32 size, u32 *cntp)
1141 ncpus = sysconf(_SC_NPROCESSORS_CONF);
1145 nr = (u32)(ncpus & UINT_MAX);
1147 for (cpu = 0; cpu < nr; cpu++) {
1148 for (level = 0; level < 10; level++) {
1149 struct cpu_cache_level c;
1152 err = cpu_cache_level__read(&c, cpu, level);
1159 for (i = 0; i < cnt; i++) {
1160 if (cpu_cache_level__cmp(&c, &caches[i]))
1167 cpu_cache_level__free(&c);
1169 if (WARN_ONCE(cnt == size, "way too many cpu caches.."))
1178 #define MAX_CACHES 2000
1180 static int write_cache(struct feat_fd *ff,
1181 struct perf_evlist *evlist __maybe_unused)
1183 struct cpu_cache_level caches[MAX_CACHES];
1184 u32 cnt = 0, i, version = 1;
1187 ret = build_caches(caches, MAX_CACHES, &cnt);
1191 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1193 ret = do_write(ff, &version, sizeof(u32));
1197 ret = do_write(ff, &cnt, sizeof(u32));
1201 for (i = 0; i < cnt; i++) {
1202 struct cpu_cache_level *c = &caches[i];
1205 ret = do_write(ff, &c->v, sizeof(u32)); \
1216 ret = do_write_string(ff, (const char *) c->v); \
1227 for (i = 0; i < cnt; i++)
1228 cpu_cache_level__free(&caches[i]);
1232 static int write_stat(struct feat_fd *ff __maybe_unused,
1233 struct perf_evlist *evlist __maybe_unused)
1238 static int write_sample_time(struct feat_fd *ff,
1239 struct perf_evlist *evlist)
1243 ret = do_write(ff, &evlist->first_sample_time,
1244 sizeof(evlist->first_sample_time));
1248 return do_write(ff, &evlist->last_sample_time,
1249 sizeof(evlist->last_sample_time));
1253 static int memory_node__read(struct memory_node *n, unsigned long idx)
1255 unsigned int phys, size = 0;
1256 char path[PATH_MAX];
1260 #define for_each_memory(mem, dir) \
1261 while ((ent = readdir(dir))) \
1262 if (strcmp(ent->d_name, ".") && \
1263 strcmp(ent->d_name, "..") && \
1264 sscanf(ent->d_name, "memory%u", &mem) == 1)
1266 scnprintf(path, PATH_MAX,
1267 "%s/devices/system/node/node%lu",
1268 sysfs__mountpoint(), idx);
1270 dir = opendir(path);
1272 pr_warning("failed: cant' open memory sysfs data\n");
1276 for_each_memory(phys, dir) {
1277 size = max(phys, size);
1282 n->set = bitmap_alloc(size);
1288 bitmap_zero(n->set, size);
1294 for_each_memory(phys, dir) {
1295 set_bit(phys, n->set);
1302 static int memory_node__sort(const void *a, const void *b)
1304 const struct memory_node *na = a;
1305 const struct memory_node *nb = b;
1307 return na->node - nb->node;
1310 static int build_mem_topology(struct memory_node *nodes, u64 size, u64 *cntp)
1312 char path[PATH_MAX];
1318 scnprintf(path, PATH_MAX, "%s/devices/system/node/",
1319 sysfs__mountpoint());
1321 dir = opendir(path);
1323 pr_debug2("%s: could't read %s, does this arch have topology information?\n",
1328 while (!ret && (ent = readdir(dir))) {
1332 if (!strcmp(ent->d_name, ".") ||
1333 !strcmp(ent->d_name, ".."))
1336 r = sscanf(ent->d_name, "node%u", &idx);
1340 if (WARN_ONCE(cnt >= size,
1341 "failed to write MEM_TOPOLOGY, way too many nodes\n"))
1344 ret = memory_node__read(&nodes[cnt++], idx);
1351 qsort(nodes, cnt, sizeof(nodes[0]), memory_node__sort);
1356 #define MAX_MEMORY_NODES 2000
1359 * The MEM_TOPOLOGY holds physical memory map for every
1360 * node in system. The format of data is as follows:
1362 * 0 - version | for future changes
1363 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes
1364 * 16 - count | number of nodes
1366 * For each node we store map of physical indexes for
1369 * 32 - node id | node index
1370 * 40 - size | size of bitmap
1371 * 48 - bitmap | bitmap of memory indexes that belongs to node
1373 static int write_mem_topology(struct feat_fd *ff __maybe_unused,
1374 struct perf_evlist *evlist __maybe_unused)
1376 static struct memory_node nodes[MAX_MEMORY_NODES];
1377 u64 bsize, version = 1, i, nr;
1380 ret = sysfs__read_xll("devices/system/memory/block_size_bytes",
1381 (unsigned long long *) &bsize);
1385 ret = build_mem_topology(&nodes[0], MAX_MEMORY_NODES, &nr);
1389 ret = do_write(ff, &version, sizeof(version));
1393 ret = do_write(ff, &bsize, sizeof(bsize));
1397 ret = do_write(ff, &nr, sizeof(nr));
1401 for (i = 0; i < nr; i++) {
1402 struct memory_node *n = &nodes[i];
1405 ret = do_write(ff, &n->v, sizeof(n->v)); \
1414 ret = do_write_bitmap(ff, n->set, n->size);
1423 static void print_hostname(struct feat_fd *ff, FILE *fp)
1425 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1428 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1430 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1433 static void print_arch(struct feat_fd *ff, FILE *fp)
1435 fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1438 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1440 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1443 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1445 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1446 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1449 static void print_version(struct feat_fd *ff, FILE *fp)
1451 fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1454 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1458 nr = ff->ph->env.nr_cmdline;
1460 fprintf(fp, "# cmdline : ");
1462 for (i = 0; i < nr; i++) {
1463 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]);
1465 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1469 char *quote = strchr(argv_i, '\'');
1473 fprintf(fp, "%s\\\'", argv_i);
1476 fprintf(fp, "%s ", argv_i);
1483 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1485 struct perf_header *ph = ff->ph;
1486 int cpu_nr = ph->env.nr_cpus_avail;
1490 nr = ph->env.nr_sibling_cores;
1491 str = ph->env.sibling_cores;
1493 for (i = 0; i < nr; i++) {
1494 fprintf(fp, "# sibling cores : %s\n", str);
1495 str += strlen(str) + 1;
1498 nr = ph->env.nr_sibling_threads;
1499 str = ph->env.sibling_threads;
1501 for (i = 0; i < nr; i++) {
1502 fprintf(fp, "# sibling threads : %s\n", str);
1503 str += strlen(str) + 1;
1506 if (ph->env.cpu != NULL) {
1507 for (i = 0; i < cpu_nr; i++)
1508 fprintf(fp, "# CPU %d: Core ID %d, Socket ID %d\n", i,
1509 ph->env.cpu[i].core_id, ph->env.cpu[i].socket_id);
1511 fprintf(fp, "# Core ID and Socket ID information is not available\n");
1514 static void free_event_desc(struct perf_evsel *events)
1516 struct perf_evsel *evsel;
1521 for (evsel = events; evsel->attr.size; evsel++) {
1522 zfree(&evsel->name);
1529 static struct perf_evsel *read_event_desc(struct feat_fd *ff)
1531 struct perf_evsel *evsel, *events = NULL;
1534 u32 nre, sz, nr, i, j;
1537 /* number of events */
1538 if (do_read_u32(ff, &nre))
1541 if (do_read_u32(ff, &sz))
1544 /* buffer to hold on file attr struct */
1549 /* the last event terminates with evsel->attr.size == 0: */
1550 events = calloc(nre + 1, sizeof(*events));
1554 msz = sizeof(evsel->attr);
1558 for (i = 0, evsel = events; i < nre; evsel++, i++) {
1562 * must read entire on-file attr struct to
1563 * sync up with layout.
1565 if (__do_read(ff, buf, sz))
1568 if (ff->ph->needs_swap)
1569 perf_event__attr_swap(buf);
1571 memcpy(&evsel->attr, buf, msz);
1573 if (do_read_u32(ff, &nr))
1576 if (ff->ph->needs_swap)
1577 evsel->needs_swap = true;
1579 evsel->name = do_read_string(ff);
1586 id = calloc(nr, sizeof(*id));
1592 for (j = 0 ; j < nr; j++) {
1593 if (do_read_u64(ff, id))
1602 free_event_desc(events);
1607 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1608 void *priv __maybe_unused)
1610 return fprintf(fp, ", %s = %s", name, val);
1613 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1615 struct perf_evsel *evsel, *events;
1620 events = ff->events;
1622 events = read_event_desc(ff);
1625 fprintf(fp, "# event desc: not available or unable to read\n");
1629 for (evsel = events; evsel->attr.size; evsel++) {
1630 fprintf(fp, "# event : name = %s, ", evsel->name);
1633 fprintf(fp, ", id = {");
1634 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1637 fprintf(fp, " %"PRIu64, *id);
1642 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1647 free_event_desc(events);
1651 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1653 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1656 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1659 struct numa_node *n;
1661 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1662 n = &ff->ph->env.numa_nodes[i];
1664 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB,"
1665 " free = %"PRIu64" kB\n",
1666 n->node, n->mem_total, n->mem_free);
1668 fprintf(fp, "# node%u cpu list : ", n->node);
1669 cpu_map__fprintf(n->map, fp);
1673 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1675 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1678 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1680 fprintf(fp, "# contains samples with branch stack\n");
1683 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1685 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1688 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1690 fprintf(fp, "# contains stat data\n");
1693 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1697 fprintf(fp, "# CPU cache info:\n");
1698 for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1700 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1704 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1706 const char *delimiter = "# pmu mappings: ";
1711 pmu_num = ff->ph->env.nr_pmu_mappings;
1713 fprintf(fp, "# pmu mappings: not available\n");
1717 str = ff->ph->env.pmu_mappings;
1720 type = strtoul(str, &tmp, 0);
1725 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1728 str += strlen(str) + 1;
1737 fprintf(fp, "# pmu mappings: unable to read\n");
1740 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1742 struct perf_session *session;
1743 struct perf_evsel *evsel;
1746 session = container_of(ff->ph, struct perf_session, header);
1748 evlist__for_each_entry(session->evlist, evsel) {
1749 if (perf_evsel__is_group_leader(evsel) &&
1750 evsel->nr_members > 1) {
1751 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1752 perf_evsel__name(evsel));
1754 nr = evsel->nr_members - 1;
1756 fprintf(fp, ",%s", perf_evsel__name(evsel));
1764 static void print_sample_time(struct feat_fd *ff, FILE *fp)
1766 struct perf_session *session;
1770 session = container_of(ff->ph, struct perf_session, header);
1772 timestamp__scnprintf_usec(session->evlist->first_sample_time,
1773 time_buf, sizeof(time_buf));
1774 fprintf(fp, "# time of first sample : %s\n", time_buf);
1776 timestamp__scnprintf_usec(session->evlist->last_sample_time,
1777 time_buf, sizeof(time_buf));
1778 fprintf(fp, "# time of last sample : %s\n", time_buf);
1780 d = (double)(session->evlist->last_sample_time -
1781 session->evlist->first_sample_time) / NSEC_PER_MSEC;
1783 fprintf(fp, "# sample duration : %10.3f ms\n", d);
1786 static void memory_node__fprintf(struct memory_node *n,
1787 unsigned long long bsize, FILE *fp)
1789 char buf_map[100], buf_size[50];
1790 unsigned long long size;
1792 size = bsize * bitmap_weight(n->set, n->size);
1793 unit_number__scnprintf(buf_size, 50, size);
1795 bitmap_scnprintf(n->set, n->size, buf_map, 100);
1796 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map);
1799 static void print_mem_topology(struct feat_fd *ff, FILE *fp)
1801 struct memory_node *nodes;
1804 nodes = ff->ph->env.memory_nodes;
1805 nr = ff->ph->env.nr_memory_nodes;
1807 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n",
1808 nr, ff->ph->env.memory_bsize);
1810 for (i = 0; i < nr; i++) {
1811 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp);
1815 static int __event_process_build_id(struct build_id_event *bev,
1817 struct perf_session *session)
1820 struct machine *machine;
1823 enum dso_kernel_type dso_type;
1825 machine = perf_session__findnew_machine(session, bev->pid);
1829 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1832 case PERF_RECORD_MISC_KERNEL:
1833 dso_type = DSO_TYPE_KERNEL;
1835 case PERF_RECORD_MISC_GUEST_KERNEL:
1836 dso_type = DSO_TYPE_GUEST_KERNEL;
1838 case PERF_RECORD_MISC_USER:
1839 case PERF_RECORD_MISC_GUEST_USER:
1840 dso_type = DSO_TYPE_USER;
1846 dso = machine__findnew_dso(machine, filename);
1848 char sbuild_id[SBUILD_ID_SIZE];
1850 dso__set_build_id(dso, &bev->build_id);
1852 if (dso_type != DSO_TYPE_USER) {
1853 struct kmod_path m = { .name = NULL, };
1855 if (!kmod_path__parse_name(&m, filename) && m.kmod)
1856 dso__set_module_info(dso, &m, machine);
1858 dso->kernel = dso_type;
1863 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1865 pr_debug("build id event received for %s: %s\n",
1866 dso->long_name, sbuild_id);
1875 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1876 int input, u64 offset, u64 size)
1878 struct perf_session *session = container_of(header, struct perf_session, header);
1880 struct perf_event_header header;
1881 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1884 struct build_id_event bev;
1885 char filename[PATH_MAX];
1886 u64 limit = offset + size;
1888 while (offset < limit) {
1891 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1894 if (header->needs_swap)
1895 perf_event_header__bswap(&old_bev.header);
1897 len = old_bev.header.size - sizeof(old_bev);
1898 if (readn(input, filename, len) != len)
1901 bev.header = old_bev.header;
1904 * As the pid is the missing value, we need to fill
1905 * it properly. The header.misc value give us nice hint.
1907 bev.pid = HOST_KERNEL_ID;
1908 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1909 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1910 bev.pid = DEFAULT_GUEST_KERNEL_ID;
1912 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1913 __event_process_build_id(&bev, filename, session);
1915 offset += bev.header.size;
1921 static int perf_header__read_build_ids(struct perf_header *header,
1922 int input, u64 offset, u64 size)
1924 struct perf_session *session = container_of(header, struct perf_session, header);
1925 struct build_id_event bev;
1926 char filename[PATH_MAX];
1927 u64 limit = offset + size, orig_offset = offset;
1930 while (offset < limit) {
1933 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1936 if (header->needs_swap)
1937 perf_event_header__bswap(&bev.header);
1939 len = bev.header.size - sizeof(bev);
1940 if (readn(input, filename, len) != len)
1943 * The a1645ce1 changeset:
1945 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1947 * Added a field to struct build_id_event that broke the file
1950 * Since the kernel build-id is the first entry, process the
1951 * table using the old format if the well known
1952 * '[kernel.kallsyms]' string for the kernel build-id has the
1953 * first 4 characters chopped off (where the pid_t sits).
1955 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1956 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1958 return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1961 __event_process_build_id(&bev, filename, session);
1963 offset += bev.header.size;
1970 /* Macro for features that simply need to read and store a string. */
1971 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
1972 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
1974 ff->ph->env.__feat_env = do_read_string(ff); \
1975 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
1978 FEAT_PROCESS_STR_FUN(hostname, hostname);
1979 FEAT_PROCESS_STR_FUN(osrelease, os_release);
1980 FEAT_PROCESS_STR_FUN(version, version);
1981 FEAT_PROCESS_STR_FUN(arch, arch);
1982 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
1983 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
1985 static int process_tracing_data(struct feat_fd *ff, void *data)
1987 ssize_t ret = trace_report(ff->fd, data, false);
1989 return ret < 0 ? -1 : 0;
1992 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
1994 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
1995 pr_debug("Failed to read buildids, continuing...\n");
1999 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
2002 u32 nr_cpus_avail, nr_cpus_online;
2004 ret = do_read_u32(ff, &nr_cpus_avail);
2008 ret = do_read_u32(ff, &nr_cpus_online);
2011 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
2012 ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
2016 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
2021 ret = do_read_u64(ff, &total_mem);
2024 ff->ph->env.total_mem = (unsigned long long)total_mem;
2028 static struct perf_evsel *
2029 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
2031 struct perf_evsel *evsel;
2033 evlist__for_each_entry(evlist, evsel) {
2034 if (evsel->idx == idx)
2042 perf_evlist__set_event_name(struct perf_evlist *evlist,
2043 struct perf_evsel *event)
2045 struct perf_evsel *evsel;
2050 evsel = perf_evlist__find_by_index(evlist, event->idx);
2057 evsel->name = strdup(event->name);
2061 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
2063 struct perf_session *session;
2064 struct perf_evsel *evsel, *events = read_event_desc(ff);
2069 session = container_of(ff->ph, struct perf_session, header);
2071 if (session->data->is_pipe) {
2072 /* Save events for reading later by print_event_desc,
2073 * since they can't be read again in pipe mode. */
2074 ff->events = events;
2077 for (evsel = events; evsel->attr.size; evsel++)
2078 perf_evlist__set_event_name(session->evlist, evsel);
2080 if (!session->data->is_pipe)
2081 free_event_desc(events);
2086 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
2088 char *str, *cmdline = NULL, **argv = NULL;
2091 if (do_read_u32(ff, &nr))
2094 ff->ph->env.nr_cmdline = nr;
2096 cmdline = zalloc(ff->size + nr + 1);
2100 argv = zalloc(sizeof(char *) * (nr + 1));
2104 for (i = 0; i < nr; i++) {
2105 str = do_read_string(ff);
2109 argv[i] = cmdline + len;
2110 memcpy(argv[i], str, strlen(str) + 1);
2111 len += strlen(str) + 1;
2114 ff->ph->env.cmdline = cmdline;
2115 ff->ph->env.cmdline_argv = (const char **) argv;
2124 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
2129 int cpu_nr = ff->ph->env.nr_cpus_avail;
2131 struct perf_header *ph = ff->ph;
2132 bool do_core_id_test = true;
2134 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
2138 if (do_read_u32(ff, &nr))
2141 ph->env.nr_sibling_cores = nr;
2142 size += sizeof(u32);
2143 if (strbuf_init(&sb, 128) < 0)
2146 for (i = 0; i < nr; i++) {
2147 str = do_read_string(ff);
2151 /* include a NULL character at the end */
2152 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2154 size += string_size(str);
2157 ph->env.sibling_cores = strbuf_detach(&sb, NULL);
2159 if (do_read_u32(ff, &nr))
2162 ph->env.nr_sibling_threads = nr;
2163 size += sizeof(u32);
2165 for (i = 0; i < nr; i++) {
2166 str = do_read_string(ff);
2170 /* include a NULL character at the end */
2171 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
2173 size += string_size(str);
2176 ph->env.sibling_threads = strbuf_detach(&sb, NULL);
2179 * The header may be from old perf,
2180 * which doesn't include core id and socket id information.
2182 if (ff->size <= size) {
2183 zfree(&ph->env.cpu);
2187 /* On s390 the socket_id number is not related to the numbers of cpus.
2188 * The socket_id number might be higher than the numbers of cpus.
2189 * This depends on the configuration.
2191 if (ph->env.arch && !strncmp(ph->env.arch, "s390", 4))
2192 do_core_id_test = false;
2194 for (i = 0; i < (u32)cpu_nr; i++) {
2195 if (do_read_u32(ff, &nr))
2198 ph->env.cpu[i].core_id = nr;
2200 if (do_read_u32(ff, &nr))
2203 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
2204 pr_debug("socket_id number is too big."
2205 "You may need to upgrade the perf tool.\n");
2209 ph->env.cpu[i].socket_id = nr;
2215 strbuf_release(&sb);
2217 zfree(&ph->env.cpu);
2221 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
2223 struct numa_node *nodes, *n;
2228 if (do_read_u32(ff, &nr))
2231 nodes = zalloc(sizeof(*nodes) * nr);
2235 for (i = 0; i < nr; i++) {
2239 if (do_read_u32(ff, &n->node))
2242 if (do_read_u64(ff, &n->mem_total))
2245 if (do_read_u64(ff, &n->mem_free))
2248 str = do_read_string(ff);
2252 n->map = cpu_map__new(str);
2258 ff->ph->env.nr_numa_nodes = nr;
2259 ff->ph->env.numa_nodes = nodes;
2267 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
2274 if (do_read_u32(ff, &pmu_num))
2278 pr_debug("pmu mappings not available\n");
2282 ff->ph->env.nr_pmu_mappings = pmu_num;
2283 if (strbuf_init(&sb, 128) < 0)
2287 if (do_read_u32(ff, &type))
2290 name = do_read_string(ff);
2294 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
2296 /* include a NULL character at the end */
2297 if (strbuf_add(&sb, "", 1) < 0)
2300 if (!strcmp(name, "msr"))
2301 ff->ph->env.msr_pmu_type = type;
2306 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2310 strbuf_release(&sb);
2314 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2317 u32 i, nr, nr_groups;
2318 struct perf_session *session;
2319 struct perf_evsel *evsel, *leader = NULL;
2326 if (do_read_u32(ff, &nr_groups))
2329 ff->ph->env.nr_groups = nr_groups;
2331 pr_debug("group desc not available\n");
2335 desc = calloc(nr_groups, sizeof(*desc));
2339 for (i = 0; i < nr_groups; i++) {
2340 desc[i].name = do_read_string(ff);
2344 if (do_read_u32(ff, &desc[i].leader_idx))
2347 if (do_read_u32(ff, &desc[i].nr_members))
2352 * Rebuild group relationship based on the group_desc
2354 session = container_of(ff->ph, struct perf_session, header);
2355 session->evlist->nr_groups = nr_groups;
2358 evlist__for_each_entry(session->evlist, evsel) {
2359 if (evsel->idx == (int) desc[i].leader_idx) {
2360 evsel->leader = evsel;
2361 /* {anon_group} is a dummy name */
2362 if (strcmp(desc[i].name, "{anon_group}")) {
2363 evsel->group_name = desc[i].name;
2364 desc[i].name = NULL;
2366 evsel->nr_members = desc[i].nr_members;
2368 if (i >= nr_groups || nr > 0) {
2369 pr_debug("invalid group desc\n");
2374 nr = evsel->nr_members - 1;
2377 /* This is a group member */
2378 evsel->leader = leader;
2384 if (i != nr_groups || nr != 0) {
2385 pr_debug("invalid group desc\n");
2391 for (i = 0; i < nr_groups; i++)
2392 zfree(&desc[i].name);
2398 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2400 struct perf_session *session;
2403 session = container_of(ff->ph, struct perf_session, header);
2405 err = auxtrace_index__process(ff->fd, ff->size, session,
2406 ff->ph->needs_swap);
2408 pr_err("Failed to process auxtrace index\n");
2412 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2414 struct cpu_cache_level *caches;
2415 u32 cnt, i, version;
2417 if (do_read_u32(ff, &version))
2423 if (do_read_u32(ff, &cnt))
2426 caches = zalloc(sizeof(*caches) * cnt);
2430 for (i = 0; i < cnt; i++) {
2431 struct cpu_cache_level c;
2434 if (do_read_u32(ff, &c.v))\
2435 goto out_free_caches; \
2444 c.v = do_read_string(ff); \
2446 goto out_free_caches;
2456 ff->ph->env.caches = caches;
2457 ff->ph->env.caches_cnt = cnt;
2464 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused)
2466 struct perf_session *session;
2467 u64 first_sample_time, last_sample_time;
2470 session = container_of(ff->ph, struct perf_session, header);
2472 ret = do_read_u64(ff, &first_sample_time);
2476 ret = do_read_u64(ff, &last_sample_time);
2480 session->evlist->first_sample_time = first_sample_time;
2481 session->evlist->last_sample_time = last_sample_time;
2485 static int process_mem_topology(struct feat_fd *ff,
2486 void *data __maybe_unused)
2488 struct memory_node *nodes;
2489 u64 version, i, nr, bsize;
2492 if (do_read_u64(ff, &version))
2498 if (do_read_u64(ff, &bsize))
2501 if (do_read_u64(ff, &nr))
2504 nodes = zalloc(sizeof(*nodes) * nr);
2508 for (i = 0; i < nr; i++) {
2509 struct memory_node n;
2512 if (do_read_u64(ff, &n.v)) \
2520 if (do_read_bitmap(ff, &n.set, &n.size))
2526 ff->ph->env.memory_bsize = bsize;
2527 ff->ph->env.memory_nodes = nodes;
2528 ff->ph->env.nr_memory_nodes = nr;
2537 struct feature_ops {
2538 int (*write)(struct feat_fd *ff, struct perf_evlist *evlist);
2539 void (*print)(struct feat_fd *ff, FILE *fp);
2540 int (*process)(struct feat_fd *ff, void *data);
2546 #define FEAT_OPR(n, func, __full_only) \
2548 .name = __stringify(n), \
2549 .write = write_##func, \
2550 .print = print_##func, \
2551 .full_only = __full_only, \
2552 .process = process_##func, \
2553 .synthesize = true \
2556 #define FEAT_OPN(n, func, __full_only) \
2558 .name = __stringify(n), \
2559 .write = write_##func, \
2560 .print = print_##func, \
2561 .full_only = __full_only, \
2562 .process = process_##func \
2565 /* feature_ops not implemented: */
2566 #define print_tracing_data NULL
2567 #define print_build_id NULL
2569 #define process_branch_stack NULL
2570 #define process_stat NULL
2573 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2574 FEAT_OPN(TRACING_DATA, tracing_data, false),
2575 FEAT_OPN(BUILD_ID, build_id, false),
2576 FEAT_OPR(HOSTNAME, hostname, false),
2577 FEAT_OPR(OSRELEASE, osrelease, false),
2578 FEAT_OPR(VERSION, version, false),
2579 FEAT_OPR(ARCH, arch, false),
2580 FEAT_OPR(NRCPUS, nrcpus, false),
2581 FEAT_OPR(CPUDESC, cpudesc, false),
2582 FEAT_OPR(CPUID, cpuid, false),
2583 FEAT_OPR(TOTAL_MEM, total_mem, false),
2584 FEAT_OPR(EVENT_DESC, event_desc, false),
2585 FEAT_OPR(CMDLINE, cmdline, false),
2586 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true),
2587 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true),
2588 FEAT_OPN(BRANCH_STACK, branch_stack, false),
2589 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false),
2590 FEAT_OPN(GROUP_DESC, group_desc, false),
2591 FEAT_OPN(AUXTRACE, auxtrace, false),
2592 FEAT_OPN(STAT, stat, false),
2593 FEAT_OPN(CACHE, cache, true),
2594 FEAT_OPR(SAMPLE_TIME, sample_time, false),
2595 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true),
2598 struct header_print_data {
2600 bool full; /* extended list of headers */
2603 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2604 struct perf_header *ph,
2605 int feat, int fd, void *data)
2607 struct header_print_data *hd = data;
2610 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2611 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2612 "%d, continuing...\n", section->offset, feat);
2615 if (feat >= HEADER_LAST_FEATURE) {
2616 pr_warning("unknown feature %d\n", feat);
2619 if (!feat_ops[feat].print)
2622 ff = (struct feat_fd) {
2627 if (!feat_ops[feat].full_only || hd->full)
2628 feat_ops[feat].print(&ff, hd->fp);
2630 fprintf(hd->fp, "# %s info available, use -I to display\n",
2631 feat_ops[feat].name);
2636 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2638 struct header_print_data hd;
2639 struct perf_header *header = &session->header;
2640 int fd = perf_data__fd(session->data);
2647 ret = fstat(fd, &st);
2651 fprintf(fp, "# captured on : %s", ctime(&st.st_ctime));
2653 fprintf(fp, "# header version : %u\n", header->version);
2654 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset);
2655 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size);
2656 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset);
2658 perf_header__process_sections(header, fd, &hd,
2659 perf_file_section__fprintf_info);
2661 if (session->data->is_pipe)
2664 fprintf(fp, "# missing features: ");
2665 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
2667 fprintf(fp, "%s ", feat_ops[bit].name);
2674 static int do_write_feat(struct feat_fd *ff, int type,
2675 struct perf_file_section **p,
2676 struct perf_evlist *evlist)
2681 if (perf_header__has_feat(ff->ph, type)) {
2682 if (!feat_ops[type].write)
2685 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
2688 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
2690 err = feat_ops[type].write(ff, evlist);
2692 pr_debug("failed to write feature %s\n", feat_ops[type].name);
2694 /* undo anything written */
2695 lseek(ff->fd, (*p)->offset, SEEK_SET);
2699 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
2705 static int perf_header__adds_write(struct perf_header *header,
2706 struct perf_evlist *evlist, int fd)
2710 struct perf_file_section *feat_sec, *p;
2716 ff = (struct feat_fd){
2721 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2725 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2726 if (feat_sec == NULL)
2729 sec_size = sizeof(*feat_sec) * nr_sections;
2731 sec_start = header->feat_offset;
2732 lseek(fd, sec_start + sec_size, SEEK_SET);
2734 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2735 if (do_write_feat(&ff, feat, &p, evlist))
2736 perf_header__clear_feat(header, feat);
2739 lseek(fd, sec_start, SEEK_SET);
2741 * may write more than needed due to dropped feature, but
2742 * this is okay, reader will skip the mising entries
2744 err = do_write(&ff, feat_sec, sec_size);
2746 pr_debug("failed to write feature section\n");
2751 int perf_header__write_pipe(int fd)
2753 struct perf_pipe_file_header f_header;
2757 ff = (struct feat_fd){ .fd = fd };
2759 f_header = (struct perf_pipe_file_header){
2760 .magic = PERF_MAGIC,
2761 .size = sizeof(f_header),
2764 err = do_write(&ff, &f_header, sizeof(f_header));
2766 pr_debug("failed to write perf pipe header\n");
2773 int perf_session__write_header(struct perf_session *session,
2774 struct perf_evlist *evlist,
2775 int fd, bool at_exit)
2777 struct perf_file_header f_header;
2778 struct perf_file_attr f_attr;
2779 struct perf_header *header = &session->header;
2780 struct perf_evsel *evsel;
2785 ff = (struct feat_fd){ .fd = fd};
2786 lseek(fd, sizeof(f_header), SEEK_SET);
2788 evlist__for_each_entry(session->evlist, evsel) {
2789 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2790 err = do_write(&ff, evsel->id, evsel->ids * sizeof(u64));
2792 pr_debug("failed to write perf header\n");
2797 attr_offset = lseek(ff.fd, 0, SEEK_CUR);
2799 evlist__for_each_entry(evlist, evsel) {
2800 f_attr = (struct perf_file_attr){
2801 .attr = evsel->attr,
2803 .offset = evsel->id_offset,
2804 .size = evsel->ids * sizeof(u64),
2807 err = do_write(&ff, &f_attr, sizeof(f_attr));
2809 pr_debug("failed to write perf header attribute\n");
2814 if (!header->data_offset)
2815 header->data_offset = lseek(fd, 0, SEEK_CUR);
2816 header->feat_offset = header->data_offset + header->data_size;
2819 err = perf_header__adds_write(header, evlist, fd);
2824 f_header = (struct perf_file_header){
2825 .magic = PERF_MAGIC,
2826 .size = sizeof(f_header),
2827 .attr_size = sizeof(f_attr),
2829 .offset = attr_offset,
2830 .size = evlist->nr_entries * sizeof(f_attr),
2833 .offset = header->data_offset,
2834 .size = header->data_size,
2836 /* event_types is ignored, store zeros */
2839 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2841 lseek(fd, 0, SEEK_SET);
2842 err = do_write(&ff, &f_header, sizeof(f_header));
2844 pr_debug("failed to write perf header\n");
2847 lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2852 static int perf_header__getbuffer64(struct perf_header *header,
2853 int fd, void *buf, size_t size)
2855 if (readn(fd, buf, size) <= 0)
2858 if (header->needs_swap)
2859 mem_bswap_64(buf, size);
2864 int perf_header__process_sections(struct perf_header *header, int fd,
2866 int (*process)(struct perf_file_section *section,
2867 struct perf_header *ph,
2868 int feat, int fd, void *data))
2870 struct perf_file_section *feat_sec, *sec;
2876 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2880 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2884 sec_size = sizeof(*feat_sec) * nr_sections;
2886 lseek(fd, header->feat_offset, SEEK_SET);
2888 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2892 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2893 err = process(sec++, header, feat, fd, data);
2903 static const int attr_file_abi_sizes[] = {
2904 [0] = PERF_ATTR_SIZE_VER0,
2905 [1] = PERF_ATTR_SIZE_VER1,
2906 [2] = PERF_ATTR_SIZE_VER2,
2907 [3] = PERF_ATTR_SIZE_VER3,
2908 [4] = PERF_ATTR_SIZE_VER4,
2913 * In the legacy file format, the magic number is not used to encode endianness.
2914 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2915 * on ABI revisions, we need to try all combinations for all endianness to
2916 * detect the endianness.
2918 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2920 uint64_t ref_size, attr_size;
2923 for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2924 ref_size = attr_file_abi_sizes[i]
2925 + sizeof(struct perf_file_section);
2926 if (hdr_sz != ref_size) {
2927 attr_size = bswap_64(hdr_sz);
2928 if (attr_size != ref_size)
2931 ph->needs_swap = true;
2933 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2938 /* could not determine endianness */
2942 #define PERF_PIPE_HDR_VER0 16
2944 static const size_t attr_pipe_abi_sizes[] = {
2945 [0] = PERF_PIPE_HDR_VER0,
2950 * In the legacy pipe format, there is an implicit assumption that endiannesss
2951 * between host recording the samples, and host parsing the samples is the
2952 * same. This is not always the case given that the pipe output may always be
2953 * redirected into a file and analyzed on a different machine with possibly a
2954 * different endianness and perf_event ABI revsions in the perf tool itself.
2956 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2961 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2962 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2963 attr_size = bswap_64(hdr_sz);
2964 if (attr_size != hdr_sz)
2967 ph->needs_swap = true;
2969 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2975 bool is_perf_magic(u64 magic)
2977 if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2978 || magic == __perf_magic2
2979 || magic == __perf_magic2_sw)
2985 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2986 bool is_pipe, struct perf_header *ph)
2990 /* check for legacy format */
2991 ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2993 ph->version = PERF_HEADER_VERSION_1;
2994 pr_debug("legacy perf.data format\n");
2996 return try_all_pipe_abis(hdr_sz, ph);
2998 return try_all_file_abis(hdr_sz, ph);
3001 * the new magic number serves two purposes:
3002 * - unique number to identify actual perf.data files
3003 * - encode endianness of file
3005 ph->version = PERF_HEADER_VERSION_2;
3007 /* check magic number with one endianness */
3008 if (magic == __perf_magic2)
3011 /* check magic number with opposite endianness */
3012 if (magic != __perf_magic2_sw)
3015 ph->needs_swap = true;
3020 int perf_file_header__read(struct perf_file_header *header,
3021 struct perf_header *ph, int fd)
3025 lseek(fd, 0, SEEK_SET);
3027 ret = readn(fd, header, sizeof(*header));
3031 if (check_magic_endian(header->magic,
3032 header->attr_size, false, ph) < 0) {
3033 pr_debug("magic/endian check failed\n");
3037 if (ph->needs_swap) {
3038 mem_bswap_64(header, offsetof(struct perf_file_header,
3042 if (header->size != sizeof(*header)) {
3043 /* Support the previous format */
3044 if (header->size == offsetof(typeof(*header), adds_features))
3045 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3048 } else if (ph->needs_swap) {
3050 * feature bitmap is declared as an array of unsigned longs --
3051 * not good since its size can differ between the host that
3052 * generated the data file and the host analyzing the file.
3054 * We need to handle endianness, but we don't know the size of
3055 * the unsigned long where the file was generated. Take a best
3056 * guess at determining it: try 64-bit swap first (ie., file
3057 * created on a 64-bit host), and check if the hostname feature
3058 * bit is set (this feature bit is forced on as of fbe96f2).
3059 * If the bit is not, undo the 64-bit swap and try a 32-bit
3060 * swap. If the hostname bit is still not set (e.g., older data
3061 * file), punt and fallback to the original behavior --
3062 * clearing all feature bits and setting buildid.
3064 mem_bswap_64(&header->adds_features,
3065 BITS_TO_U64(HEADER_FEAT_BITS));
3067 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3069 mem_bswap_64(&header->adds_features,
3070 BITS_TO_U64(HEADER_FEAT_BITS));
3073 mem_bswap_32(&header->adds_features,
3074 BITS_TO_U32(HEADER_FEAT_BITS));
3077 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
3078 bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
3079 set_bit(HEADER_BUILD_ID, header->adds_features);
3083 memcpy(&ph->adds_features, &header->adds_features,
3084 sizeof(ph->adds_features));
3086 ph->data_offset = header->data.offset;
3087 ph->data_size = header->data.size;
3088 ph->feat_offset = header->data.offset + header->data.size;
3092 static int perf_file_section__process(struct perf_file_section *section,
3093 struct perf_header *ph,
3094 int feat, int fd, void *data)
3096 struct feat_fd fdd = {
3099 .size = section->size,
3100 .offset = section->offset,
3103 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
3104 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
3105 "%d, continuing...\n", section->offset, feat);
3109 if (feat >= HEADER_LAST_FEATURE) {
3110 pr_debug("unknown feature %d, continuing...\n", feat);
3114 if (!feat_ops[feat].process)
3117 return feat_ops[feat].process(&fdd, data);
3120 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
3121 struct perf_header *ph, int fd,
3124 struct feat_fd ff = {
3125 .fd = STDOUT_FILENO,
3130 ret = readn(fd, header, sizeof(*header));
3134 if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
3135 pr_debug("endian/magic failed\n");
3140 header->size = bswap_64(header->size);
3142 if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
3148 static int perf_header__read_pipe(struct perf_session *session)
3150 struct perf_header *header = &session->header;
3151 struct perf_pipe_file_header f_header;
3153 if (perf_file_header__read_pipe(&f_header, header,
3154 perf_data__fd(session->data),
3155 session->repipe) < 0) {
3156 pr_debug("incompatible file format\n");
3163 static int read_attr(int fd, struct perf_header *ph,
3164 struct perf_file_attr *f_attr)
3166 struct perf_event_attr *attr = &f_attr->attr;
3168 size_t our_sz = sizeof(f_attr->attr);
3171 memset(f_attr, 0, sizeof(*f_attr));
3173 /* read minimal guaranteed structure */
3174 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
3176 pr_debug("cannot read %d bytes of header attr\n",
3177 PERF_ATTR_SIZE_VER0);
3181 /* on file perf_event_attr size */
3189 sz = PERF_ATTR_SIZE_VER0;
3190 } else if (sz > our_sz) {
3191 pr_debug("file uses a more recent and unsupported ABI"
3192 " (%zu bytes extra)\n", sz - our_sz);
3195 /* what we have not yet read and that we know about */
3196 left = sz - PERF_ATTR_SIZE_VER0;
3199 ptr += PERF_ATTR_SIZE_VER0;
3201 ret = readn(fd, ptr, left);
3203 /* read perf_file_section, ids are read in caller */
3204 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
3206 return ret <= 0 ? -1 : 0;
3209 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
3210 struct pevent *pevent)
3212 struct event_format *event;
3215 /* already prepared */
3216 if (evsel->tp_format)
3219 if (pevent == NULL) {
3220 pr_debug("broken or missing trace data\n");
3224 event = pevent_find_event(pevent, evsel->attr.config);
3225 if (event == NULL) {
3226 pr_debug("cannot find event format for %d\n", (int)evsel->attr.config);
3231 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
3232 evsel->name = strdup(bf);
3233 if (evsel->name == NULL)
3237 evsel->tp_format = event;
3241 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
3242 struct pevent *pevent)
3244 struct perf_evsel *pos;
3246 evlist__for_each_entry(evlist, pos) {
3247 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
3248 perf_evsel__prepare_tracepoint_event(pos, pevent))
3255 int perf_session__read_header(struct perf_session *session)
3257 struct perf_data *data = session->data;
3258 struct perf_header *header = &session->header;
3259 struct perf_file_header f_header;
3260 struct perf_file_attr f_attr;
3262 int nr_attrs, nr_ids, i, j;
3263 int fd = perf_data__fd(data);
3265 session->evlist = perf_evlist__new();
3266 if (session->evlist == NULL)
3269 session->evlist->env = &header->env;
3270 session->machines.host.env = &header->env;
3271 if (perf_data__is_pipe(data))
3272 return perf_header__read_pipe(session);
3274 if (perf_file_header__read(&f_header, header, fd) < 0)
3278 * Sanity check that perf.data was written cleanly; data size is
3279 * initialized to 0 and updated only if the on_exit function is run.
3280 * If data size is still 0 then the file contains only partial
3281 * information. Just warn user and process it as much as it can.
3283 if (f_header.data.size == 0) {
3284 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3285 "Was the 'perf record' command properly terminated?\n",
3289 nr_attrs = f_header.attrs.size / f_header.attr_size;
3290 lseek(fd, f_header.attrs.offset, SEEK_SET);
3292 for (i = 0; i < nr_attrs; i++) {
3293 struct perf_evsel *evsel;
3296 if (read_attr(fd, header, &f_attr) < 0)
3299 if (header->needs_swap) {
3300 f_attr.ids.size = bswap_64(f_attr.ids.size);
3301 f_attr.ids.offset = bswap_64(f_attr.ids.offset);
3302 perf_event__attr_swap(&f_attr.attr);
3305 tmp = lseek(fd, 0, SEEK_CUR);
3306 evsel = perf_evsel__new(&f_attr.attr);
3309 goto out_delete_evlist;
3311 evsel->needs_swap = header->needs_swap;
3313 * Do it before so that if perf_evsel__alloc_id fails, this
3314 * entry gets purged too at perf_evlist__delete().
3316 perf_evlist__add(session->evlist, evsel);
3318 nr_ids = f_attr.ids.size / sizeof(u64);
3320 * We don't have the cpu and thread maps on the header, so
3321 * for allocating the perf_sample_id table we fake 1 cpu and
3322 * hattr->ids threads.
3324 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
3325 goto out_delete_evlist;
3327 lseek(fd, f_attr.ids.offset, SEEK_SET);
3329 for (j = 0; j < nr_ids; j++) {
3330 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
3333 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
3336 lseek(fd, tmp, SEEK_SET);
3339 perf_header__process_sections(header, fd, &session->tevent,
3340 perf_file_section__process);
3342 if (perf_evlist__prepare_tracepoint_events(session->evlist,
3343 session->tevent.pevent))
3344 goto out_delete_evlist;
3351 perf_evlist__delete(session->evlist);
3352 session->evlist = NULL;
3356 int perf_event__synthesize_attr(struct perf_tool *tool,
3357 struct perf_event_attr *attr, u32 ids, u64 *id,
3358 perf_event__handler_t process)
3360 union perf_event *ev;
3364 size = sizeof(struct perf_event_attr);
3365 size = PERF_ALIGN(size, sizeof(u64));
3366 size += sizeof(struct perf_event_header);
3367 size += ids * sizeof(u64);
3374 ev->attr.attr = *attr;
3375 memcpy(ev->attr.id, id, ids * sizeof(u64));
3377 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
3378 ev->attr.header.size = (u16)size;
3380 if (ev->attr.header.size == size)
3381 err = process(tool, ev, NULL, NULL);
3390 int perf_event__synthesize_features(struct perf_tool *tool,
3391 struct perf_session *session,
3392 struct perf_evlist *evlist,
3393 perf_event__handler_t process)
3395 struct perf_header *header = &session->header;
3397 struct feature_event *fe;
3401 sz_hdr = sizeof(fe->header);
3402 sz = sizeof(union perf_event);
3403 /* get a nice alignment */
3404 sz = PERF_ALIGN(sz, page_size);
3406 memset(&ff, 0, sizeof(ff));
3408 ff.buf = malloc(sz);
3412 ff.size = sz - sz_hdr;
3414 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3415 if (!feat_ops[feat].synthesize) {
3416 pr_debug("No record header feature for header :%d\n", feat);
3420 ff.offset = sizeof(*fe);
3422 ret = feat_ops[feat].write(&ff, evlist);
3423 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
3424 pr_debug("Error writing feature\n");
3427 /* ff.buf may have changed due to realloc in do_write() */
3429 memset(fe, 0, sizeof(*fe));
3432 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3433 fe->header.size = ff.offset;
3435 ret = process(tool, ff.buf, NULL, NULL);
3442 /* Send HEADER_LAST_FEATURE mark. */
3444 fe->feat_id = HEADER_LAST_FEATURE;
3445 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3446 fe->header.size = sizeof(*fe);
3448 ret = process(tool, ff.buf, NULL, NULL);
3454 int perf_event__process_feature(struct perf_tool *tool,
3455 union perf_event *event,
3456 struct perf_session *session __maybe_unused)
3458 struct feat_fd ff = { .fd = 0 };
3459 struct feature_event *fe = (struct feature_event *)event;
3460 int type = fe->header.type;
3461 u64 feat = fe->feat_id;
3463 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3464 pr_warning("invalid record type %d in pipe-mode\n", type);
3467 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) {
3468 pr_warning("invalid record type %d in pipe-mode\n", type);
3472 if (!feat_ops[feat].process)
3475 ff.buf = (void *)fe->data;
3476 ff.size = event->header.size - sizeof(event->header);
3477 ff.ph = &session->header;
3479 if (feat_ops[feat].process(&ff, NULL))
3482 if (!feat_ops[feat].print || !tool->show_feat_hdr)
3485 if (!feat_ops[feat].full_only ||
3486 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3487 feat_ops[feat].print(&ff, stdout);
3489 fprintf(stdout, "# %s info available, use -I to display\n",
3490 feat_ops[feat].name);
3496 static struct event_update_event *
3497 event_update_event__new(size_t size, u64 type, u64 id)
3499 struct event_update_event *ev;
3501 size += sizeof(*ev);
3502 size = PERF_ALIGN(size, sizeof(u64));
3506 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3507 ev->header.size = (u16)size;
3515 perf_event__synthesize_event_update_unit(struct perf_tool *tool,
3516 struct perf_evsel *evsel,
3517 perf_event__handler_t process)
3519 struct event_update_event *ev;
3520 size_t size = strlen(evsel->unit);
3523 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->id[0]);
3527 strncpy(ev->data, evsel->unit, size);
3528 err = process(tool, (union perf_event *)ev, NULL, NULL);
3534 perf_event__synthesize_event_update_scale(struct perf_tool *tool,
3535 struct perf_evsel *evsel,
3536 perf_event__handler_t process)
3538 struct event_update_event *ev;
3539 struct event_update_event_scale *ev_data;
3542 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->id[0]);
3546 ev_data = (struct event_update_event_scale *) ev->data;
3547 ev_data->scale = evsel->scale;
3548 err = process(tool, (union perf_event*) ev, NULL, NULL);
3554 perf_event__synthesize_event_update_name(struct perf_tool *tool,
3555 struct perf_evsel *evsel,
3556 perf_event__handler_t process)
3558 struct event_update_event *ev;
3559 size_t len = strlen(evsel->name);
3562 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->id[0]);
3566 strncpy(ev->data, evsel->name, len);
3567 err = process(tool, (union perf_event*) ev, NULL, NULL);
3573 perf_event__synthesize_event_update_cpus(struct perf_tool *tool,
3574 struct perf_evsel *evsel,
3575 perf_event__handler_t process)
3577 size_t size = sizeof(struct event_update_event);
3578 struct event_update_event *ev;
3582 if (!evsel->own_cpus)
3585 ev = cpu_map_data__alloc(evsel->own_cpus, &size, &type, &max);
3589 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3590 ev->header.size = (u16)size;
3591 ev->type = PERF_EVENT_UPDATE__CPUS;
3592 ev->id = evsel->id[0];
3594 cpu_map_data__synthesize((struct cpu_map_data *) ev->data,
3598 err = process(tool, (union perf_event*) ev, NULL, NULL);
3603 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3605 struct event_update_event *ev = &event->event_update;
3606 struct event_update_event_scale *ev_scale;
3607 struct event_update_event_cpus *ev_cpus;
3608 struct cpu_map *map;
3611 ret = fprintf(fp, "\n... id: %" PRIu64 "\n", ev->id);
3614 case PERF_EVENT_UPDATE__SCALE:
3615 ev_scale = (struct event_update_event_scale *) ev->data;
3616 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3618 case PERF_EVENT_UPDATE__UNIT:
3619 ret += fprintf(fp, "... unit: %s\n", ev->data);
3621 case PERF_EVENT_UPDATE__NAME:
3622 ret += fprintf(fp, "... name: %s\n", ev->data);
3624 case PERF_EVENT_UPDATE__CPUS:
3625 ev_cpus = (struct event_update_event_cpus *) ev->data;
3626 ret += fprintf(fp, "... ");
3628 map = cpu_map__new_data(&ev_cpus->cpus);
3630 ret += cpu_map__fprintf(map, fp);
3632 ret += fprintf(fp, "failed to get cpus\n");
3635 ret += fprintf(fp, "... unknown type\n");
3642 int perf_event__synthesize_attrs(struct perf_tool *tool,
3643 struct perf_session *session,
3644 perf_event__handler_t process)
3646 struct perf_evsel *evsel;
3649 evlist__for_each_entry(session->evlist, evsel) {
3650 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
3651 evsel->id, process);
3653 pr_debug("failed to create perf header attribute\n");
3661 static bool has_unit(struct perf_evsel *counter)
3663 return counter->unit && *counter->unit;
3666 static bool has_scale(struct perf_evsel *counter)
3668 return counter->scale != 1;
3671 int perf_event__synthesize_extra_attr(struct perf_tool *tool,
3672 struct perf_evlist *evsel_list,
3673 perf_event__handler_t process,
3676 struct perf_evsel *counter;
3680 * Synthesize other events stuff not carried within
3681 * attr event - unit, scale, name
3683 evlist__for_each_entry(evsel_list, counter) {
3684 if (!counter->supported)
3688 * Synthesize unit and scale only if it's defined.
3690 if (has_unit(counter)) {
3691 err = perf_event__synthesize_event_update_unit(tool, counter, process);
3693 pr_err("Couldn't synthesize evsel unit.\n");
3698 if (has_scale(counter)) {
3699 err = perf_event__synthesize_event_update_scale(tool, counter, process);
3701 pr_err("Couldn't synthesize evsel counter.\n");
3706 if (counter->own_cpus) {
3707 err = perf_event__synthesize_event_update_cpus(tool, counter, process);
3709 pr_err("Couldn't synthesize evsel cpus.\n");
3715 * Name is needed only for pipe output,
3716 * perf.data carries event names.
3719 err = perf_event__synthesize_event_update_name(tool, counter, process);
3721 pr_err("Couldn't synthesize evsel name.\n");
3729 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3730 union perf_event *event,
3731 struct perf_evlist **pevlist)
3734 struct perf_evsel *evsel;
3735 struct perf_evlist *evlist = *pevlist;
3737 if (evlist == NULL) {
3738 *pevlist = evlist = perf_evlist__new();
3743 evsel = perf_evsel__new(&event->attr.attr);
3747 perf_evlist__add(evlist, evsel);
3749 ids = event->header.size;
3750 ids -= (void *)&event->attr.id - (void *)event;
3751 n_ids = ids / sizeof(u64);
3753 * We don't have the cpu and thread maps on the header, so
3754 * for allocating the perf_sample_id table we fake 1 cpu and
3755 * hattr->ids threads.
3757 if (perf_evsel__alloc_id(evsel, 1, n_ids))
3760 for (i = 0; i < n_ids; i++) {
3761 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
3767 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3768 union perf_event *event,
3769 struct perf_evlist **pevlist)
3771 struct event_update_event *ev = &event->event_update;
3772 struct event_update_event_scale *ev_scale;
3773 struct event_update_event_cpus *ev_cpus;
3774 struct perf_evlist *evlist;
3775 struct perf_evsel *evsel;
3776 struct cpu_map *map;
3778 if (!pevlist || *pevlist == NULL)
3783 evsel = perf_evlist__id2evsel(evlist, ev->id);
3788 case PERF_EVENT_UPDATE__UNIT:
3789 evsel->unit = strdup(ev->data);
3791 case PERF_EVENT_UPDATE__NAME:
3792 evsel->name = strdup(ev->data);
3794 case PERF_EVENT_UPDATE__SCALE:
3795 ev_scale = (struct event_update_event_scale *) ev->data;
3796 evsel->scale = ev_scale->scale;
3798 case PERF_EVENT_UPDATE__CPUS:
3799 ev_cpus = (struct event_update_event_cpus *) ev->data;
3801 map = cpu_map__new_data(&ev_cpus->cpus);
3803 evsel->own_cpus = map;
3805 pr_err("failed to get event_update cpus\n");
3813 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
3814 struct perf_evlist *evlist,
3815 perf_event__handler_t process)
3817 union perf_event ev;
3818 struct tracing_data *tdata;
3819 ssize_t size = 0, aligned_size = 0, padding;
3821 int err __maybe_unused = 0;
3824 * We are going to store the size of the data followed
3825 * by the data contents. Since the fd descriptor is a pipe,
3826 * we cannot seek back to store the size of the data once
3827 * we know it. Instead we:
3829 * - write the tracing data to the temp file
3830 * - get/write the data size to pipe
3831 * - write the tracing data from the temp file
3834 tdata = tracing_data_get(&evlist->entries, fd, true);
3838 memset(&ev, 0, sizeof(ev));
3840 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
3842 aligned_size = PERF_ALIGN(size, sizeof(u64));
3843 padding = aligned_size - size;
3844 ev.tracing_data.header.size = sizeof(ev.tracing_data);
3845 ev.tracing_data.size = aligned_size;
3847 process(tool, &ev, NULL, NULL);
3850 * The put function will copy all the tracing data
3851 * stored in temp file to the pipe.
3853 tracing_data_put(tdata);
3855 ff = (struct feat_fd){ .fd = fd };
3856 if (write_padded(&ff, NULL, 0, padding))
3859 return aligned_size;
3862 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
3863 union perf_event *event,
3864 struct perf_session *session)
3866 ssize_t size_read, padding, size = event->tracing_data.size;
3867 int fd = perf_data__fd(session->data);
3868 off_t offset = lseek(fd, 0, SEEK_CUR);
3871 /* setup for reading amidst mmap */
3872 lseek(fd, offset + sizeof(struct tracing_data_event),
3875 size_read = trace_report(fd, &session->tevent,
3877 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3879 if (readn(fd, buf, padding) < 0) {
3880 pr_err("%s: reading input file", __func__);
3883 if (session->repipe) {
3884 int retw = write(STDOUT_FILENO, buf, padding);
3885 if (retw <= 0 || retw != padding) {
3886 pr_err("%s: repiping tracing data padding", __func__);
3891 if (size_read + padding != size) {
3892 pr_err("%s: tracing data size mismatch", __func__);
3896 perf_evlist__prepare_tracepoint_events(session->evlist,
3897 session->tevent.pevent);
3899 return size_read + padding;
3902 int perf_event__synthesize_build_id(struct perf_tool *tool,
3903 struct dso *pos, u16 misc,
3904 perf_event__handler_t process,
3905 struct machine *machine)
3907 union perf_event ev;
3914 memset(&ev, 0, sizeof(ev));
3916 len = pos->long_name_len + 1;
3917 len = PERF_ALIGN(len, NAME_ALIGN);
3918 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3919 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3920 ev.build_id.header.misc = misc;
3921 ev.build_id.pid = machine->pid;
3922 ev.build_id.header.size = sizeof(ev.build_id) + len;
3923 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3925 err = process(tool, &ev, NULL, machine);
3930 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3931 union perf_event *event,
3932 struct perf_session *session)
3934 __event_process_build_id(&event->build_id,
3935 event->build_id.filename,