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"
569 char **core_siblings
;
570 char **thread_siblings
;
573 static int build_cpu_topo(struct cpu_topo
*tp
, int cpu
)
576 char filename
[MAXPATHLEN
];
577 char *buf
= NULL
, *p
;
583 sprintf(filename
, CORE_SIB_FMT
, cpu
);
584 fp
= fopen(filename
, "r");
588 sret
= getline(&buf
, &len
, fp
);
593 p
= strchr(buf
, '\n');
597 for (i
= 0; i
< tp
->core_sib
; i
++) {
598 if (!strcmp(buf
, tp
->core_siblings
[i
]))
601 if (i
== tp
->core_sib
) {
602 tp
->core_siblings
[i
] = buf
;
610 sprintf(filename
, THRD_SIB_FMT
, cpu
);
611 fp
= fopen(filename
, "r");
615 if (getline(&buf
, &len
, fp
) <= 0)
618 p
= strchr(buf
, '\n');
622 for (i
= 0; i
< tp
->thread_sib
; i
++) {
623 if (!strcmp(buf
, tp
->thread_siblings
[i
]))
626 if (i
== tp
->thread_sib
) {
627 tp
->thread_siblings
[i
] = buf
;
639 static void free_cpu_topo(struct cpu_topo
*tp
)
646 for (i
= 0 ; i
< tp
->core_sib
; i
++)
647 zfree(&tp
->core_siblings
[i
]);
649 for (i
= 0 ; i
< tp
->thread_sib
; i
++)
650 zfree(&tp
->thread_siblings
[i
]);
655 static struct cpu_topo
*build_cpu_topology(void)
657 struct cpu_topo
*tp
= NULL
;
665 ncpus
= cpu__max_present_cpu();
667 /* build online CPU map */
668 map
= cpu_map__new(NULL
);
670 pr_debug("failed to get system cpumap\n");
674 nr
= (u32
)(ncpus
& UINT_MAX
);
676 sz
= nr
* sizeof(char *);
677 addr
= calloc(1, sizeof(*tp
) + 2 * sz
);
684 tp
->core_siblings
= addr
;
686 tp
->thread_siblings
= addr
;
688 for (i
= 0; i
< nr
; i
++) {
689 if (!cpu_map__has(map
, i
))
692 ret
= build_cpu_topo(tp
, i
);
706 static int write_cpu_topology(struct feat_fd
*ff
,
707 struct perf_evlist
*evlist __maybe_unused
)
713 tp
= build_cpu_topology();
717 ret
= do_write(ff
, &tp
->core_sib
, sizeof(tp
->core_sib
));
721 for (i
= 0; i
< tp
->core_sib
; i
++) {
722 ret
= do_write_string(ff
, tp
->core_siblings
[i
]);
726 ret
= do_write(ff
, &tp
->thread_sib
, sizeof(tp
->thread_sib
));
730 for (i
= 0; i
< tp
->thread_sib
; i
++) {
731 ret
= do_write_string(ff
, tp
->thread_siblings
[i
]);
736 ret
= perf_env__read_cpu_topology_map(&perf_env
);
740 for (j
= 0; j
< perf_env
.nr_cpus_avail
; j
++) {
741 ret
= do_write(ff
, &perf_env
.cpu
[j
].core_id
,
742 sizeof(perf_env
.cpu
[j
].core_id
));
745 ret
= do_write(ff
, &perf_env
.cpu
[j
].socket_id
,
746 sizeof(perf_env
.cpu
[j
].socket_id
));
757 static int write_total_mem(struct feat_fd
*ff
,
758 struct perf_evlist
*evlist __maybe_unused
)
766 fp
= fopen("/proc/meminfo", "r");
770 while (getline(&buf
, &len
, fp
) > 0) {
771 ret
= strncmp(buf
, "MemTotal:", 9);
776 n
= sscanf(buf
, "%*s %"PRIu64
, &mem
);
778 ret
= do_write(ff
, &mem
, sizeof(mem
));
786 static int write_topo_node(struct feat_fd
*ff
, int node
)
788 char str
[MAXPATHLEN
];
790 char *buf
= NULL
, *p
;
793 u64 mem_total
, mem_free
, mem
;
796 sprintf(str
, "/sys/devices/system/node/node%d/meminfo", node
);
797 fp
= fopen(str
, "r");
801 while (getline(&buf
, &len
, fp
) > 0) {
802 /* skip over invalid lines */
803 if (!strchr(buf
, ':'))
805 if (sscanf(buf
, "%*s %*d %31s %"PRIu64
, field
, &mem
) != 2)
807 if (!strcmp(field
, "MemTotal:"))
809 if (!strcmp(field
, "MemFree:"))
816 ret
= do_write(ff
, &mem_total
, sizeof(u64
));
820 ret
= do_write(ff
, &mem_free
, sizeof(u64
));
825 sprintf(str
, "/sys/devices/system/node/node%d/cpulist", node
);
827 fp
= fopen(str
, "r");
831 if (getline(&buf
, &len
, fp
) <= 0)
834 p
= strchr(buf
, '\n');
838 ret
= do_write_string(ff
, buf
);
846 static int write_numa_topology(struct feat_fd
*ff
,
847 struct perf_evlist
*evlist __maybe_unused
)
852 struct cpu_map
*node_map
= NULL
;
857 fp
= fopen("/sys/devices/system/node/online", "r");
861 if (getline(&buf
, &len
, fp
) <= 0)
864 c
= strchr(buf
, '\n');
868 node_map
= cpu_map__new(buf
);
872 nr
= (u32
)node_map
->nr
;
874 ret
= do_write(ff
, &nr
, sizeof(nr
));
878 for (i
= 0; i
< nr
; i
++) {
879 j
= (u32
)node_map
->map
[i
];
880 ret
= do_write(ff
, &j
, sizeof(j
));
884 ret
= write_topo_node(ff
, i
);
891 cpu_map__put(node_map
);
898 * struct pmu_mappings {
907 static int write_pmu_mappings(struct feat_fd
*ff
,
908 struct perf_evlist
*evlist __maybe_unused
)
910 struct perf_pmu
*pmu
= NULL
;
915 * Do a first pass to count number of pmu to avoid lseek so this
916 * works in pipe mode as well.
918 while ((pmu
= perf_pmu__scan(pmu
))) {
924 ret
= do_write(ff
, &pmu_num
, sizeof(pmu_num
));
928 while ((pmu
= perf_pmu__scan(pmu
))) {
932 ret
= do_write(ff
, &pmu
->type
, sizeof(pmu
->type
));
936 ret
= do_write_string(ff
, pmu
->name
);
947 * struct group_descs {
949 * struct group_desc {
956 static int write_group_desc(struct feat_fd
*ff
,
957 struct perf_evlist
*evlist
)
959 u32 nr_groups
= evlist
->nr_groups
;
960 struct perf_evsel
*evsel
;
963 ret
= do_write(ff
, &nr_groups
, sizeof(nr_groups
));
967 evlist__for_each_entry(evlist
, evsel
) {
968 if (perf_evsel__is_group_leader(evsel
) &&
969 evsel
->nr_members
> 1) {
970 const char *name
= evsel
->group_name
?: "{anon_group}";
971 u32 leader_idx
= evsel
->idx
;
972 u32 nr_members
= evsel
->nr_members
;
974 ret
= do_write_string(ff
, name
);
978 ret
= do_write(ff
, &leader_idx
, sizeof(leader_idx
));
982 ret
= do_write(ff
, &nr_members
, sizeof(nr_members
));
991 * default get_cpuid(): nothing gets recorded
992 * actual implementation must be in arch/$(SRCARCH)/util/header.c
994 int __weak
get_cpuid(char *buffer __maybe_unused
, size_t sz __maybe_unused
)
999 static int write_cpuid(struct feat_fd
*ff
,
1000 struct perf_evlist
*evlist __maybe_unused
)
1005 ret
= get_cpuid(buffer
, sizeof(buffer
));
1011 return do_write_string(ff
, buffer
);
1014 static int write_branch_stack(struct feat_fd
*ff __maybe_unused
,
1015 struct perf_evlist
*evlist __maybe_unused
)
1020 static int write_auxtrace(struct feat_fd
*ff
,
1021 struct perf_evlist
*evlist __maybe_unused
)
1023 struct perf_session
*session
;
1026 if (WARN(ff
->buf
, "Error: calling %s in pipe-mode.\n", __func__
))
1029 session
= container_of(ff
->ph
, struct perf_session
, header
);
1031 err
= auxtrace_index__write(ff
->fd
, &session
->auxtrace_index
);
1033 pr_err("Failed to write auxtrace index\n");
1037 static int write_clockid(struct feat_fd
*ff
,
1038 struct perf_evlist
*evlist __maybe_unused
)
1040 return do_write(ff
, &ff
->ph
->env
.clockid_res_ns
,
1041 sizeof(ff
->ph
->env
.clockid_res_ns
));
1044 static int cpu_cache_level__sort(const void *a
, const void *b
)
1046 struct cpu_cache_level
*cache_a
= (struct cpu_cache_level
*)a
;
1047 struct cpu_cache_level
*cache_b
= (struct cpu_cache_level
*)b
;
1049 return cache_a
->level
- cache_b
->level
;
1052 static bool cpu_cache_level__cmp(struct cpu_cache_level
*a
, struct cpu_cache_level
*b
)
1054 if (a
->level
!= b
->level
)
1057 if (a
->line_size
!= b
->line_size
)
1060 if (a
->sets
!= b
->sets
)
1063 if (a
->ways
!= b
->ways
)
1066 if (strcmp(a
->type
, b
->type
))
1069 if (strcmp(a
->size
, b
->size
))
1072 if (strcmp(a
->map
, b
->map
))
1078 static int cpu_cache_level__read(struct cpu_cache_level
*cache
, u32 cpu
, u16 level
)
1080 char path
[PATH_MAX
], file
[PATH_MAX
];
1084 scnprintf(path
, PATH_MAX
, "devices/system/cpu/cpu%d/cache/index%d/", cpu
, level
);
1085 scnprintf(file
, PATH_MAX
, "%s/%s", sysfs__mountpoint(), path
);
1087 if (stat(file
, &st
))
1090 scnprintf(file
, PATH_MAX
, "%s/level", path
);
1091 if (sysfs__read_int(file
, (int *) &cache
->level
))
1094 scnprintf(file
, PATH_MAX
, "%s/coherency_line_size", path
);
1095 if (sysfs__read_int(file
, (int *) &cache
->line_size
))
1098 scnprintf(file
, PATH_MAX
, "%s/number_of_sets", path
);
1099 if (sysfs__read_int(file
, (int *) &cache
->sets
))
1102 scnprintf(file
, PATH_MAX
, "%s/ways_of_associativity", path
);
1103 if (sysfs__read_int(file
, (int *) &cache
->ways
))
1106 scnprintf(file
, PATH_MAX
, "%s/type", path
);
1107 if (sysfs__read_str(file
, &cache
->type
, &len
))
1110 cache
->type
[len
] = 0;
1111 cache
->type
= rtrim(cache
->type
);
1113 scnprintf(file
, PATH_MAX
, "%s/size", path
);
1114 if (sysfs__read_str(file
, &cache
->size
, &len
)) {
1119 cache
->size
[len
] = 0;
1120 cache
->size
= rtrim(cache
->size
);
1122 scnprintf(file
, PATH_MAX
, "%s/shared_cpu_list", path
);
1123 if (sysfs__read_str(file
, &cache
->map
, &len
)) {
1129 cache
->map
[len
] = 0;
1130 cache
->map
= rtrim(cache
->map
);
1134 static void cpu_cache_level__fprintf(FILE *out
, struct cpu_cache_level
*c
)
1136 fprintf(out
, "L%d %-15s %8s [%s]\n", c
->level
, c
->type
, c
->size
, c
->map
);
1139 static int build_caches(struct cpu_cache_level caches
[], u32 size
, u32
*cntp
)
1146 ncpus
= sysconf(_SC_NPROCESSORS_CONF
);
1150 nr
= (u32
)(ncpus
& UINT_MAX
);
1152 for (cpu
= 0; cpu
< nr
; cpu
++) {
1153 for (level
= 0; level
< 10; level
++) {
1154 struct cpu_cache_level c
;
1157 err
= cpu_cache_level__read(&c
, cpu
, level
);
1164 for (i
= 0; i
< cnt
; i
++) {
1165 if (cpu_cache_level__cmp(&c
, &caches
[i
]))
1172 cpu_cache_level__free(&c
);
1174 if (WARN_ONCE(cnt
== size
, "way too many cpu caches.."))
1183 #define MAX_CACHES 2000
1185 static int write_cache(struct feat_fd
*ff
,
1186 struct perf_evlist
*evlist __maybe_unused
)
1188 struct cpu_cache_level caches
[MAX_CACHES
];
1189 u32 cnt
= 0, i
, version
= 1;
1192 ret
= build_caches(caches
, MAX_CACHES
, &cnt
);
1196 qsort(&caches
, cnt
, sizeof(struct cpu_cache_level
), cpu_cache_level__sort
);
1198 ret
= do_write(ff
, &version
, sizeof(u32
));
1202 ret
= do_write(ff
, &cnt
, sizeof(u32
));
1206 for (i
= 0; i
< cnt
; i
++) {
1207 struct cpu_cache_level
*c
= &caches
[i
];
1210 ret = do_write(ff, &c->v, sizeof(u32)); \
1221 ret = do_write_string(ff, (const char *) c->v); \
1232 for (i
= 0; i
< cnt
; i
++)
1233 cpu_cache_level__free(&caches
[i
]);
1237 static int write_stat(struct feat_fd
*ff __maybe_unused
,
1238 struct perf_evlist
*evlist __maybe_unused
)
1243 static int write_sample_time(struct feat_fd
*ff
,
1244 struct perf_evlist
*evlist
)
1248 ret
= do_write(ff
, &evlist
->first_sample_time
,
1249 sizeof(evlist
->first_sample_time
));
1253 return do_write(ff
, &evlist
->last_sample_time
,
1254 sizeof(evlist
->last_sample_time
));
1258 static int memory_node__read(struct memory_node
*n
, unsigned long idx
)
1260 unsigned int phys
, size
= 0;
1261 char path
[PATH_MAX
];
1265 #define for_each_memory(mem, dir) \
1266 while ((ent = readdir(dir))) \
1267 if (strcmp(ent->d_name, ".") && \
1268 strcmp(ent->d_name, "..") && \
1269 sscanf(ent->d_name, "memory%u", &mem) == 1)
1271 scnprintf(path
, PATH_MAX
,
1272 "%s/devices/system/node/node%lu",
1273 sysfs__mountpoint(), idx
);
1275 dir
= opendir(path
);
1277 pr_warning("failed: cant' open memory sysfs data\n");
1281 for_each_memory(phys
, dir
) {
1282 size
= max(phys
, size
);
1287 n
->set
= bitmap_alloc(size
);
1298 for_each_memory(phys
, dir
) {
1299 set_bit(phys
, n
->set
);
1306 static int memory_node__sort(const void *a
, const void *b
)
1308 const struct memory_node
*na
= a
;
1309 const struct memory_node
*nb
= b
;
1311 return na
->node
- nb
->node
;
1314 static int build_mem_topology(struct memory_node
*nodes
, u64 size
, u64
*cntp
)
1316 char path
[PATH_MAX
];
1322 scnprintf(path
, PATH_MAX
, "%s/devices/system/node/",
1323 sysfs__mountpoint());
1325 dir
= opendir(path
);
1327 pr_debug2("%s: could't read %s, does this arch have topology information?\n",
1332 while (!ret
&& (ent
= readdir(dir
))) {
1336 if (!strcmp(ent
->d_name
, ".") ||
1337 !strcmp(ent
->d_name
, ".."))
1340 r
= sscanf(ent
->d_name
, "node%u", &idx
);
1344 if (WARN_ONCE(cnt
>= size
,
1345 "failed to write MEM_TOPOLOGY, way too many nodes\n"))
1348 ret
= memory_node__read(&nodes
[cnt
++], idx
);
1355 qsort(nodes
, cnt
, sizeof(nodes
[0]), memory_node__sort
);
1360 #define MAX_MEMORY_NODES 2000
1363 * The MEM_TOPOLOGY holds physical memory map for every
1364 * node in system. The format of data is as follows:
1366 * 0 - version | for future changes
1367 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes
1368 * 16 - count | number of nodes
1370 * For each node we store map of physical indexes for
1373 * 32 - node id | node index
1374 * 40 - size | size of bitmap
1375 * 48 - bitmap | bitmap of memory indexes that belongs to node
1377 static int write_mem_topology(struct feat_fd
*ff __maybe_unused
,
1378 struct perf_evlist
*evlist __maybe_unused
)
1380 static struct memory_node nodes
[MAX_MEMORY_NODES
];
1381 u64 bsize
, version
= 1, i
, nr
;
1384 ret
= sysfs__read_xll("devices/system/memory/block_size_bytes",
1385 (unsigned long long *) &bsize
);
1389 ret
= build_mem_topology(&nodes
[0], MAX_MEMORY_NODES
, &nr
);
1393 ret
= do_write(ff
, &version
, sizeof(version
));
1397 ret
= do_write(ff
, &bsize
, sizeof(bsize
));
1401 ret
= do_write(ff
, &nr
, sizeof(nr
));
1405 for (i
= 0; i
< nr
; i
++) {
1406 struct memory_node
*n
= &nodes
[i
];
1409 ret = do_write(ff, &n->v, sizeof(n->v)); \
1418 ret
= do_write_bitmap(ff
, n
->set
, n
->size
);
1427 static void print_hostname(struct feat_fd
*ff
, FILE *fp
)
1429 fprintf(fp
, "# hostname : %s\n", ff
->ph
->env
.hostname
);
1432 static void print_osrelease(struct feat_fd
*ff
, FILE *fp
)
1434 fprintf(fp
, "# os release : %s\n", ff
->ph
->env
.os_release
);
1437 static void print_arch(struct feat_fd
*ff
, FILE *fp
)
1439 fprintf(fp
, "# arch : %s\n", ff
->ph
->env
.arch
);
1442 static void print_cpudesc(struct feat_fd
*ff
, FILE *fp
)
1444 fprintf(fp
, "# cpudesc : %s\n", ff
->ph
->env
.cpu_desc
);
1447 static void print_nrcpus(struct feat_fd
*ff
, FILE *fp
)
1449 fprintf(fp
, "# nrcpus online : %u\n", ff
->ph
->env
.nr_cpus_online
);
1450 fprintf(fp
, "# nrcpus avail : %u\n", ff
->ph
->env
.nr_cpus_avail
);
1453 static void print_version(struct feat_fd
*ff
, FILE *fp
)
1455 fprintf(fp
, "# perf version : %s\n", ff
->ph
->env
.version
);
1458 static void print_cmdline(struct feat_fd
*ff
, FILE *fp
)
1462 nr
= ff
->ph
->env
.nr_cmdline
;
1464 fprintf(fp
, "# cmdline : ");
1466 for (i
= 0; i
< nr
; i
++) {
1467 char *argv_i
= strdup(ff
->ph
->env
.cmdline_argv
[i
]);
1469 fprintf(fp
, "%s ", ff
->ph
->env
.cmdline_argv
[i
]);
1473 char *quote
= strchr(argv_i
, '\'');
1477 fprintf(fp
, "%s\\\'", argv_i
);
1480 fprintf(fp
, "%s ", argv_i
);
1487 static void print_cpu_topology(struct feat_fd
*ff
, FILE *fp
)
1489 struct perf_header
*ph
= ff
->ph
;
1490 int cpu_nr
= ph
->env
.nr_cpus_avail
;
1494 nr
= ph
->env
.nr_sibling_cores
;
1495 str
= ph
->env
.sibling_cores
;
1497 for (i
= 0; i
< nr
; i
++) {
1498 fprintf(fp
, "# sibling cores : %s\n", str
);
1499 str
+= strlen(str
) + 1;
1502 nr
= ph
->env
.nr_sibling_threads
;
1503 str
= ph
->env
.sibling_threads
;
1505 for (i
= 0; i
< nr
; i
++) {
1506 fprintf(fp
, "# sibling threads : %s\n", str
);
1507 str
+= strlen(str
) + 1;
1510 if (ph
->env
.cpu
!= NULL
) {
1511 for (i
= 0; i
< cpu_nr
; i
++)
1512 fprintf(fp
, "# CPU %d: Core ID %d, Socket ID %d\n", i
,
1513 ph
->env
.cpu
[i
].core_id
, ph
->env
.cpu
[i
].socket_id
);
1515 fprintf(fp
, "# Core ID and Socket ID information is not available\n");
1518 static void print_clockid(struct feat_fd
*ff
, FILE *fp
)
1520 fprintf(fp
, "# clockid frequency: %"PRIu64
" MHz\n",
1521 ff
->ph
->env
.clockid_res_ns
* 1000);
1524 static void free_event_desc(struct perf_evsel
*events
)
1526 struct perf_evsel
*evsel
;
1531 for (evsel
= events
; evsel
->attr
.size
; evsel
++) {
1532 zfree(&evsel
->name
);
1539 static struct perf_evsel
*read_event_desc(struct feat_fd
*ff
)
1541 struct perf_evsel
*evsel
, *events
= NULL
;
1544 u32 nre
, sz
, nr
, i
, j
;
1547 /* number of events */
1548 if (do_read_u32(ff
, &nre
))
1551 if (do_read_u32(ff
, &sz
))
1554 /* buffer to hold on file attr struct */
1559 /* the last event terminates with evsel->attr.size == 0: */
1560 events
= calloc(nre
+ 1, sizeof(*events
));
1564 msz
= sizeof(evsel
->attr
);
1568 for (i
= 0, evsel
= events
; i
< nre
; evsel
++, i
++) {
1572 * must read entire on-file attr struct to
1573 * sync up with layout.
1575 if (__do_read(ff
, buf
, sz
))
1578 if (ff
->ph
->needs_swap
)
1579 perf_event__attr_swap(buf
);
1581 memcpy(&evsel
->attr
, buf
, msz
);
1583 if (do_read_u32(ff
, &nr
))
1586 if (ff
->ph
->needs_swap
)
1587 evsel
->needs_swap
= true;
1589 evsel
->name
= do_read_string(ff
);
1596 id
= calloc(nr
, sizeof(*id
));
1602 for (j
= 0 ; j
< nr
; j
++) {
1603 if (do_read_u64(ff
, id
))
1612 free_event_desc(events
);
1617 static int __desc_attr__fprintf(FILE *fp
, const char *name
, const char *val
,
1618 void *priv __maybe_unused
)
1620 return fprintf(fp
, ", %s = %s", name
, val
);
1623 static void print_event_desc(struct feat_fd
*ff
, FILE *fp
)
1625 struct perf_evsel
*evsel
, *events
;
1630 events
= ff
->events
;
1632 events
= read_event_desc(ff
);
1635 fprintf(fp
, "# event desc: not available or unable to read\n");
1639 for (evsel
= events
; evsel
->attr
.size
; evsel
++) {
1640 fprintf(fp
, "# event : name = %s, ", evsel
->name
);
1643 fprintf(fp
, ", id = {");
1644 for (j
= 0, id
= evsel
->id
; j
< evsel
->ids
; j
++, id
++) {
1647 fprintf(fp
, " %"PRIu64
, *id
);
1652 perf_event_attr__fprintf(fp
, &evsel
->attr
, __desc_attr__fprintf
, NULL
);
1657 free_event_desc(events
);
1661 static void print_total_mem(struct feat_fd
*ff
, FILE *fp
)
1663 fprintf(fp
, "# total memory : %llu kB\n", ff
->ph
->env
.total_mem
);
1666 static void print_numa_topology(struct feat_fd
*ff
, FILE *fp
)
1669 struct numa_node
*n
;
1671 for (i
= 0; i
< ff
->ph
->env
.nr_numa_nodes
; i
++) {
1672 n
= &ff
->ph
->env
.numa_nodes
[i
];
1674 fprintf(fp
, "# node%u meminfo : total = %"PRIu64
" kB,"
1675 " free = %"PRIu64
" kB\n",
1676 n
->node
, n
->mem_total
, n
->mem_free
);
1678 fprintf(fp
, "# node%u cpu list : ", n
->node
);
1679 cpu_map__fprintf(n
->map
, fp
);
1683 static void print_cpuid(struct feat_fd
*ff
, FILE *fp
)
1685 fprintf(fp
, "# cpuid : %s\n", ff
->ph
->env
.cpuid
);
1688 static void print_branch_stack(struct feat_fd
*ff __maybe_unused
, FILE *fp
)
1690 fprintf(fp
, "# contains samples with branch stack\n");
1693 static void print_auxtrace(struct feat_fd
*ff __maybe_unused
, FILE *fp
)
1695 fprintf(fp
, "# contains AUX area data (e.g. instruction trace)\n");
1698 static void print_stat(struct feat_fd
*ff __maybe_unused
, FILE *fp
)
1700 fprintf(fp
, "# contains stat data\n");
1703 static void print_cache(struct feat_fd
*ff
, FILE *fp __maybe_unused
)
1707 fprintf(fp
, "# CPU cache info:\n");
1708 for (i
= 0; i
< ff
->ph
->env
.caches_cnt
; i
++) {
1710 cpu_cache_level__fprintf(fp
, &ff
->ph
->env
.caches
[i
]);
1714 static void print_pmu_mappings(struct feat_fd
*ff
, FILE *fp
)
1716 const char *delimiter
= "# pmu mappings: ";
1721 pmu_num
= ff
->ph
->env
.nr_pmu_mappings
;
1723 fprintf(fp
, "# pmu mappings: not available\n");
1727 str
= ff
->ph
->env
.pmu_mappings
;
1730 type
= strtoul(str
, &tmp
, 0);
1735 fprintf(fp
, "%s%s = %" PRIu32
, delimiter
, str
, type
);
1738 str
+= strlen(str
) + 1;
1747 fprintf(fp
, "# pmu mappings: unable to read\n");
1750 static void print_group_desc(struct feat_fd
*ff
, FILE *fp
)
1752 struct perf_session
*session
;
1753 struct perf_evsel
*evsel
;
1756 session
= container_of(ff
->ph
, struct perf_session
, header
);
1758 evlist__for_each_entry(session
->evlist
, evsel
) {
1759 if (perf_evsel__is_group_leader(evsel
) &&
1760 evsel
->nr_members
> 1) {
1761 fprintf(fp
, "# group: %s{%s", evsel
->group_name
?: "",
1762 perf_evsel__name(evsel
));
1764 nr
= evsel
->nr_members
- 1;
1766 fprintf(fp
, ",%s", perf_evsel__name(evsel
));
1774 static void print_sample_time(struct feat_fd
*ff
, FILE *fp
)
1776 struct perf_session
*session
;
1780 session
= container_of(ff
->ph
, struct perf_session
, header
);
1782 timestamp__scnprintf_usec(session
->evlist
->first_sample_time
,
1783 time_buf
, sizeof(time_buf
));
1784 fprintf(fp
, "# time of first sample : %s\n", time_buf
);
1786 timestamp__scnprintf_usec(session
->evlist
->last_sample_time
,
1787 time_buf
, sizeof(time_buf
));
1788 fprintf(fp
, "# time of last sample : %s\n", time_buf
);
1790 d
= (double)(session
->evlist
->last_sample_time
-
1791 session
->evlist
->first_sample_time
) / NSEC_PER_MSEC
;
1793 fprintf(fp
, "# sample duration : %10.3f ms\n", d
);
1796 static void memory_node__fprintf(struct memory_node
*n
,
1797 unsigned long long bsize
, FILE *fp
)
1799 char buf_map
[100], buf_size
[50];
1800 unsigned long long size
;
1802 size
= bsize
* bitmap_weight(n
->set
, n
->size
);
1803 unit_number__scnprintf(buf_size
, 50, size
);
1805 bitmap_scnprintf(n
->set
, n
->size
, buf_map
, 100);
1806 fprintf(fp
, "# %3" PRIu64
" [%s]: %s\n", n
->node
, buf_size
, buf_map
);
1809 static void print_mem_topology(struct feat_fd
*ff
, FILE *fp
)
1811 struct memory_node
*nodes
;
1814 nodes
= ff
->ph
->env
.memory_nodes
;
1815 nr
= ff
->ph
->env
.nr_memory_nodes
;
1817 fprintf(fp
, "# memory nodes (nr %d, block size 0x%llx):\n",
1818 nr
, ff
->ph
->env
.memory_bsize
);
1820 for (i
= 0; i
< nr
; i
++) {
1821 memory_node__fprintf(&nodes
[i
], ff
->ph
->env
.memory_bsize
, fp
);
1825 static int __event_process_build_id(struct build_id_event
*bev
,
1827 struct perf_session
*session
)
1830 struct machine
*machine
;
1833 enum dso_kernel_type dso_type
;
1835 machine
= perf_session__findnew_machine(session
, bev
->pid
);
1839 cpumode
= bev
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
1842 case PERF_RECORD_MISC_KERNEL
:
1843 dso_type
= DSO_TYPE_KERNEL
;
1845 case PERF_RECORD_MISC_GUEST_KERNEL
:
1846 dso_type
= DSO_TYPE_GUEST_KERNEL
;
1848 case PERF_RECORD_MISC_USER
:
1849 case PERF_RECORD_MISC_GUEST_USER
:
1850 dso_type
= DSO_TYPE_USER
;
1856 dso
= machine__findnew_dso(machine
, filename
);
1858 char sbuild_id
[SBUILD_ID_SIZE
];
1860 dso__set_build_id(dso
, &bev
->build_id
);
1862 if (dso_type
!= DSO_TYPE_USER
) {
1863 struct kmod_path m
= { .name
= NULL
, };
1865 if (!kmod_path__parse_name(&m
, filename
) && m
.kmod
)
1866 dso__set_module_info(dso
, &m
, machine
);
1868 dso
->kernel
= dso_type
;
1873 build_id__sprintf(dso
->build_id
, sizeof(dso
->build_id
),
1875 pr_debug("build id event received for %s: %s\n",
1876 dso
->long_name
, sbuild_id
);
1885 static int perf_header__read_build_ids_abi_quirk(struct perf_header
*header
,
1886 int input
, u64 offset
, u64 size
)
1888 struct perf_session
*session
= container_of(header
, struct perf_session
, header
);
1890 struct perf_event_header header
;
1891 u8 build_id
[PERF_ALIGN(BUILD_ID_SIZE
, sizeof(u64
))];
1894 struct build_id_event bev
;
1895 char filename
[PATH_MAX
];
1896 u64 limit
= offset
+ size
;
1898 while (offset
< limit
) {
1901 if (readn(input
, &old_bev
, sizeof(old_bev
)) != sizeof(old_bev
))
1904 if (header
->needs_swap
)
1905 perf_event_header__bswap(&old_bev
.header
);
1907 len
= old_bev
.header
.size
- sizeof(old_bev
);
1908 if (readn(input
, filename
, len
) != len
)
1911 bev
.header
= old_bev
.header
;
1914 * As the pid is the missing value, we need to fill
1915 * it properly. The header.misc value give us nice hint.
1917 bev
.pid
= HOST_KERNEL_ID
;
1918 if (bev
.header
.misc
== PERF_RECORD_MISC_GUEST_USER
||
1919 bev
.header
.misc
== PERF_RECORD_MISC_GUEST_KERNEL
)
1920 bev
.pid
= DEFAULT_GUEST_KERNEL_ID
;
1922 memcpy(bev
.build_id
, old_bev
.build_id
, sizeof(bev
.build_id
));
1923 __event_process_build_id(&bev
, filename
, session
);
1925 offset
+= bev
.header
.size
;
1931 static int perf_header__read_build_ids(struct perf_header
*header
,
1932 int input
, u64 offset
, u64 size
)
1934 struct perf_session
*session
= container_of(header
, struct perf_session
, header
);
1935 struct build_id_event bev
;
1936 char filename
[PATH_MAX
];
1937 u64 limit
= offset
+ size
, orig_offset
= offset
;
1940 while (offset
< limit
) {
1943 if (readn(input
, &bev
, sizeof(bev
)) != sizeof(bev
))
1946 if (header
->needs_swap
)
1947 perf_event_header__bswap(&bev
.header
);
1949 len
= bev
.header
.size
- sizeof(bev
);
1950 if (readn(input
, filename
, len
) != len
)
1953 * The a1645ce1 changeset:
1955 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1957 * Added a field to struct build_id_event that broke the file
1960 * Since the kernel build-id is the first entry, process the
1961 * table using the old format if the well known
1962 * '[kernel.kallsyms]' string for the kernel build-id has the
1963 * first 4 characters chopped off (where the pid_t sits).
1965 if (memcmp(filename
, "nel.kallsyms]", 13) == 0) {
1966 if (lseek(input
, orig_offset
, SEEK_SET
) == (off_t
)-1)
1968 return perf_header__read_build_ids_abi_quirk(header
, input
, offset
, size
);
1971 __event_process_build_id(&bev
, filename
, session
);
1973 offset
+= bev
.header
.size
;
1980 /* Macro for features that simply need to read and store a string. */
1981 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
1982 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
1984 ff->ph->env.__feat_env = do_read_string(ff); \
1985 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
1988 FEAT_PROCESS_STR_FUN(hostname
, hostname
);
1989 FEAT_PROCESS_STR_FUN(osrelease
, os_release
);
1990 FEAT_PROCESS_STR_FUN(version
, version
);
1991 FEAT_PROCESS_STR_FUN(arch
, arch
);
1992 FEAT_PROCESS_STR_FUN(cpudesc
, cpu_desc
);
1993 FEAT_PROCESS_STR_FUN(cpuid
, cpuid
);
1995 static int process_tracing_data(struct feat_fd
*ff
, void *data
)
1997 ssize_t ret
= trace_report(ff
->fd
, data
, false);
1999 return ret
< 0 ? -1 : 0;
2002 static int process_build_id(struct feat_fd
*ff
, void *data __maybe_unused
)
2004 if (perf_header__read_build_ids(ff
->ph
, ff
->fd
, ff
->offset
, ff
->size
))
2005 pr_debug("Failed to read buildids, continuing...\n");
2009 static int process_nrcpus(struct feat_fd
*ff
, void *data __maybe_unused
)
2012 u32 nr_cpus_avail
, nr_cpus_online
;
2014 ret
= do_read_u32(ff
, &nr_cpus_avail
);
2018 ret
= do_read_u32(ff
, &nr_cpus_online
);
2021 ff
->ph
->env
.nr_cpus_avail
= (int)nr_cpus_avail
;
2022 ff
->ph
->env
.nr_cpus_online
= (int)nr_cpus_online
;
2026 static int process_total_mem(struct feat_fd
*ff
, void *data __maybe_unused
)
2031 ret
= do_read_u64(ff
, &total_mem
);
2034 ff
->ph
->env
.total_mem
= (unsigned long long)total_mem
;
2038 static struct perf_evsel
*
2039 perf_evlist__find_by_index(struct perf_evlist
*evlist
, int idx
)
2041 struct perf_evsel
*evsel
;
2043 evlist__for_each_entry(evlist
, evsel
) {
2044 if (evsel
->idx
== idx
)
2052 perf_evlist__set_event_name(struct perf_evlist
*evlist
,
2053 struct perf_evsel
*event
)
2055 struct perf_evsel
*evsel
;
2060 evsel
= perf_evlist__find_by_index(evlist
, event
->idx
);
2067 evsel
->name
= strdup(event
->name
);
2071 process_event_desc(struct feat_fd
*ff
, void *data __maybe_unused
)
2073 struct perf_session
*session
;
2074 struct perf_evsel
*evsel
, *events
= read_event_desc(ff
);
2079 session
= container_of(ff
->ph
, struct perf_session
, header
);
2081 if (session
->data
->is_pipe
) {
2082 /* Save events for reading later by print_event_desc,
2083 * since they can't be read again in pipe mode. */
2084 ff
->events
= events
;
2087 for (evsel
= events
; evsel
->attr
.size
; evsel
++)
2088 perf_evlist__set_event_name(session
->evlist
, evsel
);
2090 if (!session
->data
->is_pipe
)
2091 free_event_desc(events
);
2096 static int process_cmdline(struct feat_fd
*ff
, void *data __maybe_unused
)
2098 char *str
, *cmdline
= NULL
, **argv
= NULL
;
2101 if (do_read_u32(ff
, &nr
))
2104 ff
->ph
->env
.nr_cmdline
= nr
;
2106 cmdline
= zalloc(ff
->size
+ nr
+ 1);
2110 argv
= zalloc(sizeof(char *) * (nr
+ 1));
2114 for (i
= 0; i
< nr
; i
++) {
2115 str
= do_read_string(ff
);
2119 argv
[i
] = cmdline
+ len
;
2120 memcpy(argv
[i
], str
, strlen(str
) + 1);
2121 len
+= strlen(str
) + 1;
2124 ff
->ph
->env
.cmdline
= cmdline
;
2125 ff
->ph
->env
.cmdline_argv
= (const char **) argv
;
2134 static int process_cpu_topology(struct feat_fd
*ff
, void *data __maybe_unused
)
2139 int cpu_nr
= ff
->ph
->env
.nr_cpus_avail
;
2141 struct perf_header
*ph
= ff
->ph
;
2142 bool do_core_id_test
= true;
2144 ph
->env
.cpu
= calloc(cpu_nr
, sizeof(*ph
->env
.cpu
));
2148 if (do_read_u32(ff
, &nr
))
2151 ph
->env
.nr_sibling_cores
= nr
;
2152 size
+= sizeof(u32
);
2153 if (strbuf_init(&sb
, 128) < 0)
2156 for (i
= 0; i
< nr
; i
++) {
2157 str
= do_read_string(ff
);
2161 /* include a NULL character at the end */
2162 if (strbuf_add(&sb
, str
, strlen(str
) + 1) < 0)
2164 size
+= string_size(str
);
2167 ph
->env
.sibling_cores
= strbuf_detach(&sb
, NULL
);
2169 if (do_read_u32(ff
, &nr
))
2172 ph
->env
.nr_sibling_threads
= nr
;
2173 size
+= sizeof(u32
);
2175 for (i
= 0; i
< nr
; i
++) {
2176 str
= do_read_string(ff
);
2180 /* include a NULL character at the end */
2181 if (strbuf_add(&sb
, str
, strlen(str
) + 1) < 0)
2183 size
+= string_size(str
);
2186 ph
->env
.sibling_threads
= strbuf_detach(&sb
, NULL
);
2189 * The header may be from old perf,
2190 * which doesn't include core id and socket id information.
2192 if (ff
->size
<= size
) {
2193 zfree(&ph
->env
.cpu
);
2197 /* On s390 the socket_id number is not related to the numbers of cpus.
2198 * The socket_id number might be higher than the numbers of cpus.
2199 * This depends on the configuration.
2201 if (ph
->env
.arch
&& !strncmp(ph
->env
.arch
, "s390", 4))
2202 do_core_id_test
= false;
2204 for (i
= 0; i
< (u32
)cpu_nr
; i
++) {
2205 if (do_read_u32(ff
, &nr
))
2208 ph
->env
.cpu
[i
].core_id
= nr
;
2210 if (do_read_u32(ff
, &nr
))
2213 if (do_core_id_test
&& nr
!= (u32
)-1 && nr
> (u32
)cpu_nr
) {
2214 pr_debug("socket_id number is too big."
2215 "You may need to upgrade the perf tool.\n");
2219 ph
->env
.cpu
[i
].socket_id
= nr
;
2225 strbuf_release(&sb
);
2227 zfree(&ph
->env
.cpu
);
2231 static int process_numa_topology(struct feat_fd
*ff
, void *data __maybe_unused
)
2233 struct numa_node
*nodes
, *n
;
2238 if (do_read_u32(ff
, &nr
))
2241 nodes
= zalloc(sizeof(*nodes
) * nr
);
2245 for (i
= 0; i
< nr
; i
++) {
2249 if (do_read_u32(ff
, &n
->node
))
2252 if (do_read_u64(ff
, &n
->mem_total
))
2255 if (do_read_u64(ff
, &n
->mem_free
))
2258 str
= do_read_string(ff
);
2262 n
->map
= cpu_map__new(str
);
2268 ff
->ph
->env
.nr_numa_nodes
= nr
;
2269 ff
->ph
->env
.numa_nodes
= nodes
;
2277 static int process_pmu_mappings(struct feat_fd
*ff
, void *data __maybe_unused
)
2284 if (do_read_u32(ff
, &pmu_num
))
2288 pr_debug("pmu mappings not available\n");
2292 ff
->ph
->env
.nr_pmu_mappings
= pmu_num
;
2293 if (strbuf_init(&sb
, 128) < 0)
2297 if (do_read_u32(ff
, &type
))
2300 name
= do_read_string(ff
);
2304 if (strbuf_addf(&sb
, "%u:%s", type
, name
) < 0)
2306 /* include a NULL character at the end */
2307 if (strbuf_add(&sb
, "", 1) < 0)
2310 if (!strcmp(name
, "msr"))
2311 ff
->ph
->env
.msr_pmu_type
= type
;
2316 ff
->ph
->env
.pmu_mappings
= strbuf_detach(&sb
, NULL
);
2320 strbuf_release(&sb
);
2324 static int process_group_desc(struct feat_fd
*ff
, void *data __maybe_unused
)
2327 u32 i
, nr
, nr_groups
;
2328 struct perf_session
*session
;
2329 struct perf_evsel
*evsel
, *leader
= NULL
;
2336 if (do_read_u32(ff
, &nr_groups
))
2339 ff
->ph
->env
.nr_groups
= nr_groups
;
2341 pr_debug("group desc not available\n");
2345 desc
= calloc(nr_groups
, sizeof(*desc
));
2349 for (i
= 0; i
< nr_groups
; i
++) {
2350 desc
[i
].name
= do_read_string(ff
);
2354 if (do_read_u32(ff
, &desc
[i
].leader_idx
))
2357 if (do_read_u32(ff
, &desc
[i
].nr_members
))
2362 * Rebuild group relationship based on the group_desc
2364 session
= container_of(ff
->ph
, struct perf_session
, header
);
2365 session
->evlist
->nr_groups
= nr_groups
;
2368 evlist__for_each_entry(session
->evlist
, evsel
) {
2369 if (evsel
->idx
== (int) desc
[i
].leader_idx
) {
2370 evsel
->leader
= evsel
;
2371 /* {anon_group} is a dummy name */
2372 if (strcmp(desc
[i
].name
, "{anon_group}")) {
2373 evsel
->group_name
= desc
[i
].name
;
2374 desc
[i
].name
= NULL
;
2376 evsel
->nr_members
= desc
[i
].nr_members
;
2378 if (i
>= nr_groups
|| nr
> 0) {
2379 pr_debug("invalid group desc\n");
2384 nr
= evsel
->nr_members
- 1;
2387 /* This is a group member */
2388 evsel
->leader
= leader
;
2394 if (i
!= nr_groups
|| nr
!= 0) {
2395 pr_debug("invalid group desc\n");
2401 for (i
= 0; i
< nr_groups
; i
++)
2402 zfree(&desc
[i
].name
);
2408 static int process_auxtrace(struct feat_fd
*ff
, void *data __maybe_unused
)
2410 struct perf_session
*session
;
2413 session
= container_of(ff
->ph
, struct perf_session
, header
);
2415 err
= auxtrace_index__process(ff
->fd
, ff
->size
, session
,
2416 ff
->ph
->needs_swap
);
2418 pr_err("Failed to process auxtrace index\n");
2422 static int process_cache(struct feat_fd
*ff
, void *data __maybe_unused
)
2424 struct cpu_cache_level
*caches
;
2425 u32 cnt
, i
, version
;
2427 if (do_read_u32(ff
, &version
))
2433 if (do_read_u32(ff
, &cnt
))
2436 caches
= zalloc(sizeof(*caches
) * cnt
);
2440 for (i
= 0; i
< cnt
; i
++) {
2441 struct cpu_cache_level c
;
2444 if (do_read_u32(ff, &c.v))\
2445 goto out_free_caches; \
2454 c.v = do_read_string(ff); \
2456 goto out_free_caches;
2466 ff
->ph
->env
.caches
= caches
;
2467 ff
->ph
->env
.caches_cnt
= cnt
;
2474 static int process_sample_time(struct feat_fd
*ff
, void *data __maybe_unused
)
2476 struct perf_session
*session
;
2477 u64 first_sample_time
, last_sample_time
;
2480 session
= container_of(ff
->ph
, struct perf_session
, header
);
2482 ret
= do_read_u64(ff
, &first_sample_time
);
2486 ret
= do_read_u64(ff
, &last_sample_time
);
2490 session
->evlist
->first_sample_time
= first_sample_time
;
2491 session
->evlist
->last_sample_time
= last_sample_time
;
2495 static int process_mem_topology(struct feat_fd
*ff
,
2496 void *data __maybe_unused
)
2498 struct memory_node
*nodes
;
2499 u64 version
, i
, nr
, bsize
;
2502 if (do_read_u64(ff
, &version
))
2508 if (do_read_u64(ff
, &bsize
))
2511 if (do_read_u64(ff
, &nr
))
2514 nodes
= zalloc(sizeof(*nodes
) * nr
);
2518 for (i
= 0; i
< nr
; i
++) {
2519 struct memory_node n
;
2522 if (do_read_u64(ff, &n.v)) \
2530 if (do_read_bitmap(ff
, &n
.set
, &n
.size
))
2536 ff
->ph
->env
.memory_bsize
= bsize
;
2537 ff
->ph
->env
.memory_nodes
= nodes
;
2538 ff
->ph
->env
.nr_memory_nodes
= nr
;
2547 static int process_clockid(struct feat_fd
*ff
,
2548 void *data __maybe_unused
)
2550 if (do_read_u64(ff
, &ff
->ph
->env
.clockid_res_ns
))
2556 struct feature_ops
{
2557 int (*write
)(struct feat_fd
*ff
, struct perf_evlist
*evlist
);
2558 void (*print
)(struct feat_fd
*ff
, FILE *fp
);
2559 int (*process
)(struct feat_fd
*ff
, void *data
);
2565 #define FEAT_OPR(n, func, __full_only) \
2567 .name = __stringify(n), \
2568 .write = write_##func, \
2569 .print = print_##func, \
2570 .full_only = __full_only, \
2571 .process = process_##func, \
2572 .synthesize = true \
2575 #define FEAT_OPN(n, func, __full_only) \
2577 .name = __stringify(n), \
2578 .write = write_##func, \
2579 .print = print_##func, \
2580 .full_only = __full_only, \
2581 .process = process_##func \
2584 /* feature_ops not implemented: */
2585 #define print_tracing_data NULL
2586 #define print_build_id NULL
2588 #define process_branch_stack NULL
2589 #define process_stat NULL
2592 static const struct feature_ops feat_ops
[HEADER_LAST_FEATURE
] = {
2593 FEAT_OPN(TRACING_DATA
, tracing_data
, false),
2594 FEAT_OPN(BUILD_ID
, build_id
, false),
2595 FEAT_OPR(HOSTNAME
, hostname
, false),
2596 FEAT_OPR(OSRELEASE
, osrelease
, false),
2597 FEAT_OPR(VERSION
, version
, false),
2598 FEAT_OPR(ARCH
, arch
, false),
2599 FEAT_OPR(NRCPUS
, nrcpus
, false),
2600 FEAT_OPR(CPUDESC
, cpudesc
, false),
2601 FEAT_OPR(CPUID
, cpuid
, false),
2602 FEAT_OPR(TOTAL_MEM
, total_mem
, false),
2603 FEAT_OPR(EVENT_DESC
, event_desc
, false),
2604 FEAT_OPR(CMDLINE
, cmdline
, false),
2605 FEAT_OPR(CPU_TOPOLOGY
, cpu_topology
, true),
2606 FEAT_OPR(NUMA_TOPOLOGY
, numa_topology
, true),
2607 FEAT_OPN(BRANCH_STACK
, branch_stack
, false),
2608 FEAT_OPR(PMU_MAPPINGS
, pmu_mappings
, false),
2609 FEAT_OPR(GROUP_DESC
, group_desc
, false),
2610 FEAT_OPN(AUXTRACE
, auxtrace
, false),
2611 FEAT_OPN(STAT
, stat
, false),
2612 FEAT_OPN(CACHE
, cache
, true),
2613 FEAT_OPR(SAMPLE_TIME
, sample_time
, false),
2614 FEAT_OPR(MEM_TOPOLOGY
, mem_topology
, true),
2615 FEAT_OPR(CLOCKID
, clockid
, false)
2618 struct header_print_data
{
2620 bool full
; /* extended list of headers */
2623 static int perf_file_section__fprintf_info(struct perf_file_section
*section
,
2624 struct perf_header
*ph
,
2625 int feat
, int fd
, void *data
)
2627 struct header_print_data
*hd
= data
;
2630 if (lseek(fd
, section
->offset
, SEEK_SET
) == (off_t
)-1) {
2631 pr_debug("Failed to lseek to %" PRIu64
" offset for feature "
2632 "%d, continuing...\n", section
->offset
, feat
);
2635 if (feat
>= HEADER_LAST_FEATURE
) {
2636 pr_warning("unknown feature %d\n", feat
);
2639 if (!feat_ops
[feat
].print
)
2642 ff
= (struct feat_fd
) {
2647 if (!feat_ops
[feat
].full_only
|| hd
->full
)
2648 feat_ops
[feat
].print(&ff
, hd
->fp
);
2650 fprintf(hd
->fp
, "# %s info available, use -I to display\n",
2651 feat_ops
[feat
].name
);
2656 int perf_header__fprintf_info(struct perf_session
*session
, FILE *fp
, bool full
)
2658 struct header_print_data hd
;
2659 struct perf_header
*header
= &session
->header
;
2660 int fd
= perf_data__fd(session
->data
);
2667 ret
= fstat(fd
, &st
);
2671 fprintf(fp
, "# captured on : %s", ctime(&st
.st_ctime
));
2673 fprintf(fp
, "# header version : %u\n", header
->version
);
2674 fprintf(fp
, "# data offset : %" PRIu64
"\n", header
->data_offset
);
2675 fprintf(fp
, "# data size : %" PRIu64
"\n", header
->data_size
);
2676 fprintf(fp
, "# feat offset : %" PRIu64
"\n", header
->feat_offset
);
2678 perf_header__process_sections(header
, fd
, &hd
,
2679 perf_file_section__fprintf_info
);
2681 if (session
->data
->is_pipe
)
2684 fprintf(fp
, "# missing features: ");
2685 for_each_clear_bit(bit
, header
->adds_features
, HEADER_LAST_FEATURE
) {
2687 fprintf(fp
, "%s ", feat_ops
[bit
].name
);
2694 static int do_write_feat(struct feat_fd
*ff
, int type
,
2695 struct perf_file_section
**p
,
2696 struct perf_evlist
*evlist
)
2701 if (perf_header__has_feat(ff
->ph
, type
)) {
2702 if (!feat_ops
[type
].write
)
2705 if (WARN(ff
->buf
, "Error: calling %s in pipe-mode.\n", __func__
))
2708 (*p
)->offset
= lseek(ff
->fd
, 0, SEEK_CUR
);
2710 err
= feat_ops
[type
].write(ff
, evlist
);
2712 pr_debug("failed to write feature %s\n", feat_ops
[type
].name
);
2714 /* undo anything written */
2715 lseek(ff
->fd
, (*p
)->offset
, SEEK_SET
);
2719 (*p
)->size
= lseek(ff
->fd
, 0, SEEK_CUR
) - (*p
)->offset
;
2725 static int perf_header__adds_write(struct perf_header
*header
,
2726 struct perf_evlist
*evlist
, int fd
)
2730 struct perf_file_section
*feat_sec
, *p
;
2736 ff
= (struct feat_fd
){
2741 nr_sections
= bitmap_weight(header
->adds_features
, HEADER_FEAT_BITS
);
2745 feat_sec
= p
= calloc(nr_sections
, sizeof(*feat_sec
));
2746 if (feat_sec
== NULL
)
2749 sec_size
= sizeof(*feat_sec
) * nr_sections
;
2751 sec_start
= header
->feat_offset
;
2752 lseek(fd
, sec_start
+ sec_size
, SEEK_SET
);
2754 for_each_set_bit(feat
, header
->adds_features
, HEADER_FEAT_BITS
) {
2755 if (do_write_feat(&ff
, feat
, &p
, evlist
))
2756 perf_header__clear_feat(header
, feat
);
2759 lseek(fd
, sec_start
, SEEK_SET
);
2761 * may write more than needed due to dropped feature, but
2762 * this is okay, reader will skip the mising entries
2764 err
= do_write(&ff
, feat_sec
, sec_size
);
2766 pr_debug("failed to write feature section\n");
2771 int perf_header__write_pipe(int fd
)
2773 struct perf_pipe_file_header f_header
;
2777 ff
= (struct feat_fd
){ .fd
= fd
};
2779 f_header
= (struct perf_pipe_file_header
){
2780 .magic
= PERF_MAGIC
,
2781 .size
= sizeof(f_header
),
2784 err
= do_write(&ff
, &f_header
, sizeof(f_header
));
2786 pr_debug("failed to write perf pipe header\n");
2793 int perf_session__write_header(struct perf_session
*session
,
2794 struct perf_evlist
*evlist
,
2795 int fd
, bool at_exit
)
2797 struct perf_file_header f_header
;
2798 struct perf_file_attr f_attr
;
2799 struct perf_header
*header
= &session
->header
;
2800 struct perf_evsel
*evsel
;
2805 ff
= (struct feat_fd
){ .fd
= fd
};
2806 lseek(fd
, sizeof(f_header
), SEEK_SET
);
2808 evlist__for_each_entry(session
->evlist
, evsel
) {
2809 evsel
->id_offset
= lseek(fd
, 0, SEEK_CUR
);
2810 err
= do_write(&ff
, evsel
->id
, evsel
->ids
* sizeof(u64
));
2812 pr_debug("failed to write perf header\n");
2817 attr_offset
= lseek(ff
.fd
, 0, SEEK_CUR
);
2819 evlist__for_each_entry(evlist
, evsel
) {
2820 f_attr
= (struct perf_file_attr
){
2821 .attr
= evsel
->attr
,
2823 .offset
= evsel
->id_offset
,
2824 .size
= evsel
->ids
* sizeof(u64
),
2827 err
= do_write(&ff
, &f_attr
, sizeof(f_attr
));
2829 pr_debug("failed to write perf header attribute\n");
2834 if (!header
->data_offset
)
2835 header
->data_offset
= lseek(fd
, 0, SEEK_CUR
);
2836 header
->feat_offset
= header
->data_offset
+ header
->data_size
;
2839 err
= perf_header__adds_write(header
, evlist
, fd
);
2844 f_header
= (struct perf_file_header
){
2845 .magic
= PERF_MAGIC
,
2846 .size
= sizeof(f_header
),
2847 .attr_size
= sizeof(f_attr
),
2849 .offset
= attr_offset
,
2850 .size
= evlist
->nr_entries
* sizeof(f_attr
),
2853 .offset
= header
->data_offset
,
2854 .size
= header
->data_size
,
2856 /* event_types is ignored, store zeros */
2859 memcpy(&f_header
.adds_features
, &header
->adds_features
, sizeof(header
->adds_features
));
2861 lseek(fd
, 0, SEEK_SET
);
2862 err
= do_write(&ff
, &f_header
, sizeof(f_header
));
2864 pr_debug("failed to write perf header\n");
2867 lseek(fd
, header
->data_offset
+ header
->data_size
, SEEK_SET
);
2872 static int perf_header__getbuffer64(struct perf_header
*header
,
2873 int fd
, void *buf
, size_t size
)
2875 if (readn(fd
, buf
, size
) <= 0)
2878 if (header
->needs_swap
)
2879 mem_bswap_64(buf
, size
);
2884 int perf_header__process_sections(struct perf_header
*header
, int fd
,
2886 int (*process
)(struct perf_file_section
*section
,
2887 struct perf_header
*ph
,
2888 int feat
, int fd
, void *data
))
2890 struct perf_file_section
*feat_sec
, *sec
;
2896 nr_sections
= bitmap_weight(header
->adds_features
, HEADER_FEAT_BITS
);
2900 feat_sec
= sec
= calloc(nr_sections
, sizeof(*feat_sec
));
2904 sec_size
= sizeof(*feat_sec
) * nr_sections
;
2906 lseek(fd
, header
->feat_offset
, SEEK_SET
);
2908 err
= perf_header__getbuffer64(header
, fd
, feat_sec
, sec_size
);
2912 for_each_set_bit(feat
, header
->adds_features
, HEADER_LAST_FEATURE
) {
2913 err
= process(sec
++, header
, feat
, fd
, data
);
2923 static const int attr_file_abi_sizes
[] = {
2924 [0] = PERF_ATTR_SIZE_VER0
,
2925 [1] = PERF_ATTR_SIZE_VER1
,
2926 [2] = PERF_ATTR_SIZE_VER2
,
2927 [3] = PERF_ATTR_SIZE_VER3
,
2928 [4] = PERF_ATTR_SIZE_VER4
,
2933 * In the legacy file format, the magic number is not used to encode endianness.
2934 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2935 * on ABI revisions, we need to try all combinations for all endianness to
2936 * detect the endianness.
2938 static int try_all_file_abis(uint64_t hdr_sz
, struct perf_header
*ph
)
2940 uint64_t ref_size
, attr_size
;
2943 for (i
= 0 ; attr_file_abi_sizes
[i
]; i
++) {
2944 ref_size
= attr_file_abi_sizes
[i
]
2945 + sizeof(struct perf_file_section
);
2946 if (hdr_sz
!= ref_size
) {
2947 attr_size
= bswap_64(hdr_sz
);
2948 if (attr_size
!= ref_size
)
2951 ph
->needs_swap
= true;
2953 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2958 /* could not determine endianness */
2962 #define PERF_PIPE_HDR_VER0 16
2964 static const size_t attr_pipe_abi_sizes
[] = {
2965 [0] = PERF_PIPE_HDR_VER0
,
2970 * In the legacy pipe format, there is an implicit assumption that endiannesss
2971 * between host recording the samples, and host parsing the samples is the
2972 * same. This is not always the case given that the pipe output may always be
2973 * redirected into a file and analyzed on a different machine with possibly a
2974 * different endianness and perf_event ABI revsions in the perf tool itself.
2976 static int try_all_pipe_abis(uint64_t hdr_sz
, struct perf_header
*ph
)
2981 for (i
= 0 ; attr_pipe_abi_sizes
[i
]; i
++) {
2982 if (hdr_sz
!= attr_pipe_abi_sizes
[i
]) {
2983 attr_size
= bswap_64(hdr_sz
);
2984 if (attr_size
!= hdr_sz
)
2987 ph
->needs_swap
= true;
2989 pr_debug("Pipe ABI%d perf.data file detected\n", i
);
2995 bool is_perf_magic(u64 magic
)
2997 if (!memcmp(&magic
, __perf_magic1
, sizeof(magic
))
2998 || magic
== __perf_magic2
2999 || magic
== __perf_magic2_sw
)
3005 static int check_magic_endian(u64 magic
, uint64_t hdr_sz
,
3006 bool is_pipe
, struct perf_header
*ph
)
3010 /* check for legacy format */
3011 ret
= memcmp(&magic
, __perf_magic1
, sizeof(magic
));
3013 ph
->version
= PERF_HEADER_VERSION_1
;
3014 pr_debug("legacy perf.data format\n");
3016 return try_all_pipe_abis(hdr_sz
, ph
);
3018 return try_all_file_abis(hdr_sz
, ph
);
3021 * the new magic number serves two purposes:
3022 * - unique number to identify actual perf.data files
3023 * - encode endianness of file
3025 ph
->version
= PERF_HEADER_VERSION_2
;
3027 /* check magic number with one endianness */
3028 if (magic
== __perf_magic2
)
3031 /* check magic number with opposite endianness */
3032 if (magic
!= __perf_magic2_sw
)
3035 ph
->needs_swap
= true;
3040 int perf_file_header__read(struct perf_file_header
*header
,
3041 struct perf_header
*ph
, int fd
)
3045 lseek(fd
, 0, SEEK_SET
);
3047 ret
= readn(fd
, header
, sizeof(*header
));
3051 if (check_magic_endian(header
->magic
,
3052 header
->attr_size
, false, ph
) < 0) {
3053 pr_debug("magic/endian check failed\n");
3057 if (ph
->needs_swap
) {
3058 mem_bswap_64(header
, offsetof(struct perf_file_header
,
3062 if (header
->size
!= sizeof(*header
)) {
3063 /* Support the previous format */
3064 if (header
->size
== offsetof(typeof(*header
), adds_features
))
3065 bitmap_zero(header
->adds_features
, HEADER_FEAT_BITS
);
3068 } else if (ph
->needs_swap
) {
3070 * feature bitmap is declared as an array of unsigned longs --
3071 * not good since its size can differ between the host that
3072 * generated the data file and the host analyzing the file.
3074 * We need to handle endianness, but we don't know the size of
3075 * the unsigned long where the file was generated. Take a best
3076 * guess at determining it: try 64-bit swap first (ie., file
3077 * created on a 64-bit host), and check if the hostname feature
3078 * bit is set (this feature bit is forced on as of fbe96f2).
3079 * If the bit is not, undo the 64-bit swap and try a 32-bit
3080 * swap. If the hostname bit is still not set (e.g., older data
3081 * file), punt and fallback to the original behavior --
3082 * clearing all feature bits and setting buildid.
3084 mem_bswap_64(&header
->adds_features
,
3085 BITS_TO_U64(HEADER_FEAT_BITS
));
3087 if (!test_bit(HEADER_HOSTNAME
, header
->adds_features
)) {
3089 mem_bswap_64(&header
->adds_features
,
3090 BITS_TO_U64(HEADER_FEAT_BITS
));
3093 mem_bswap_32(&header
->adds_features
,
3094 BITS_TO_U32(HEADER_FEAT_BITS
));
3097 if (!test_bit(HEADER_HOSTNAME
, header
->adds_features
)) {
3098 bitmap_zero(header
->adds_features
, HEADER_FEAT_BITS
);
3099 set_bit(HEADER_BUILD_ID
, header
->adds_features
);
3103 memcpy(&ph
->adds_features
, &header
->adds_features
,
3104 sizeof(ph
->adds_features
));
3106 ph
->data_offset
= header
->data
.offset
;
3107 ph
->data_size
= header
->data
.size
;
3108 ph
->feat_offset
= header
->data
.offset
+ header
->data
.size
;
3112 static int perf_file_section__process(struct perf_file_section
*section
,
3113 struct perf_header
*ph
,
3114 int feat
, int fd
, void *data
)
3116 struct feat_fd fdd
= {
3119 .size
= section
->size
,
3120 .offset
= section
->offset
,
3123 if (lseek(fd
, section
->offset
, SEEK_SET
) == (off_t
)-1) {
3124 pr_debug("Failed to lseek to %" PRIu64
" offset for feature "
3125 "%d, continuing...\n", section
->offset
, feat
);
3129 if (feat
>= HEADER_LAST_FEATURE
) {
3130 pr_debug("unknown feature %d, continuing...\n", feat
);
3134 if (!feat_ops
[feat
].process
)
3137 return feat_ops
[feat
].process(&fdd
, data
);
3140 static int perf_file_header__read_pipe(struct perf_pipe_file_header
*header
,
3141 struct perf_header
*ph
, int fd
,
3144 struct feat_fd ff
= {
3145 .fd
= STDOUT_FILENO
,
3150 ret
= readn(fd
, header
, sizeof(*header
));
3154 if (check_magic_endian(header
->magic
, header
->size
, true, ph
) < 0) {
3155 pr_debug("endian/magic failed\n");
3160 header
->size
= bswap_64(header
->size
);
3162 if (repipe
&& do_write(&ff
, header
, sizeof(*header
)) < 0)
3168 static int perf_header__read_pipe(struct perf_session
*session
)
3170 struct perf_header
*header
= &session
->header
;
3171 struct perf_pipe_file_header f_header
;
3173 if (perf_file_header__read_pipe(&f_header
, header
,
3174 perf_data__fd(session
->data
),
3175 session
->repipe
) < 0) {
3176 pr_debug("incompatible file format\n");
3183 static int read_attr(int fd
, struct perf_header
*ph
,
3184 struct perf_file_attr
*f_attr
)
3186 struct perf_event_attr
*attr
= &f_attr
->attr
;
3188 size_t our_sz
= sizeof(f_attr
->attr
);
3191 memset(f_attr
, 0, sizeof(*f_attr
));
3193 /* read minimal guaranteed structure */
3194 ret
= readn(fd
, attr
, PERF_ATTR_SIZE_VER0
);
3196 pr_debug("cannot read %d bytes of header attr\n",
3197 PERF_ATTR_SIZE_VER0
);
3201 /* on file perf_event_attr size */
3209 sz
= PERF_ATTR_SIZE_VER0
;
3210 } else if (sz
> our_sz
) {
3211 pr_debug("file uses a more recent and unsupported ABI"
3212 " (%zu bytes extra)\n", sz
- our_sz
);
3215 /* what we have not yet read and that we know about */
3216 left
= sz
- PERF_ATTR_SIZE_VER0
;
3219 ptr
+= PERF_ATTR_SIZE_VER0
;
3221 ret
= readn(fd
, ptr
, left
);
3223 /* read perf_file_section, ids are read in caller */
3224 ret
= readn(fd
, &f_attr
->ids
, sizeof(f_attr
->ids
));
3226 return ret
<= 0 ? -1 : 0;
3229 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel
*evsel
,
3230 struct tep_handle
*pevent
)
3232 struct tep_event_format
*event
;
3235 /* already prepared */
3236 if (evsel
->tp_format
)
3239 if (pevent
== NULL
) {
3240 pr_debug("broken or missing trace data\n");
3244 event
= tep_find_event(pevent
, evsel
->attr
.config
);
3245 if (event
== NULL
) {
3246 pr_debug("cannot find event format for %d\n", (int)evsel
->attr
.config
);
3251 snprintf(bf
, sizeof(bf
), "%s:%s", event
->system
, event
->name
);
3252 evsel
->name
= strdup(bf
);
3253 if (evsel
->name
== NULL
)
3257 evsel
->tp_format
= event
;
3261 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist
*evlist
,
3262 struct tep_handle
*pevent
)
3264 struct perf_evsel
*pos
;
3266 evlist__for_each_entry(evlist
, pos
) {
3267 if (pos
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
3268 perf_evsel__prepare_tracepoint_event(pos
, pevent
))
3275 int perf_session__read_header(struct perf_session
*session
)
3277 struct perf_data
*data
= session
->data
;
3278 struct perf_header
*header
= &session
->header
;
3279 struct perf_file_header f_header
;
3280 struct perf_file_attr f_attr
;
3282 int nr_attrs
, nr_ids
, i
, j
;
3283 int fd
= perf_data__fd(data
);
3285 session
->evlist
= perf_evlist__new();
3286 if (session
->evlist
== NULL
)
3289 session
->evlist
->env
= &header
->env
;
3290 session
->machines
.host
.env
= &header
->env
;
3291 if (perf_data__is_pipe(data
))
3292 return perf_header__read_pipe(session
);
3294 if (perf_file_header__read(&f_header
, header
, fd
) < 0)
3298 * Sanity check that perf.data was written cleanly; data size is
3299 * initialized to 0 and updated only if the on_exit function is run.
3300 * If data size is still 0 then the file contains only partial
3301 * information. Just warn user and process it as much as it can.
3303 if (f_header
.data
.size
== 0) {
3304 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
3305 "Was the 'perf record' command properly terminated?\n",
3309 nr_attrs
= f_header
.attrs
.size
/ f_header
.attr_size
;
3310 lseek(fd
, f_header
.attrs
.offset
, SEEK_SET
);
3312 for (i
= 0; i
< nr_attrs
; i
++) {
3313 struct perf_evsel
*evsel
;
3316 if (read_attr(fd
, header
, &f_attr
) < 0)
3319 if (header
->needs_swap
) {
3320 f_attr
.ids
.size
= bswap_64(f_attr
.ids
.size
);
3321 f_attr
.ids
.offset
= bswap_64(f_attr
.ids
.offset
);
3322 perf_event__attr_swap(&f_attr
.attr
);
3325 tmp
= lseek(fd
, 0, SEEK_CUR
);
3326 evsel
= perf_evsel__new(&f_attr
.attr
);
3329 goto out_delete_evlist
;
3331 evsel
->needs_swap
= header
->needs_swap
;
3333 * Do it before so that if perf_evsel__alloc_id fails, this
3334 * entry gets purged too at perf_evlist__delete().
3336 perf_evlist__add(session
->evlist
, evsel
);
3338 nr_ids
= f_attr
.ids
.size
/ sizeof(u64
);
3340 * We don't have the cpu and thread maps on the header, so
3341 * for allocating the perf_sample_id table we fake 1 cpu and
3342 * hattr->ids threads.
3344 if (perf_evsel__alloc_id(evsel
, 1, nr_ids
))
3345 goto out_delete_evlist
;
3347 lseek(fd
, f_attr
.ids
.offset
, SEEK_SET
);
3349 for (j
= 0; j
< nr_ids
; j
++) {
3350 if (perf_header__getbuffer64(header
, fd
, &f_id
, sizeof(f_id
)))
3353 perf_evlist__id_add(session
->evlist
, evsel
, 0, j
, f_id
);
3356 lseek(fd
, tmp
, SEEK_SET
);
3359 perf_header__process_sections(header
, fd
, &session
->tevent
,
3360 perf_file_section__process
);
3362 if (perf_evlist__prepare_tracepoint_events(session
->evlist
,
3363 session
->tevent
.pevent
))
3364 goto out_delete_evlist
;
3371 perf_evlist__delete(session
->evlist
);
3372 session
->evlist
= NULL
;
3376 int perf_event__synthesize_attr(struct perf_tool
*tool
,
3377 struct perf_event_attr
*attr
, u32 ids
, u64
*id
,
3378 perf_event__handler_t process
)
3380 union perf_event
*ev
;
3384 size
= sizeof(struct perf_event_attr
);
3385 size
= PERF_ALIGN(size
, sizeof(u64
));
3386 size
+= sizeof(struct perf_event_header
);
3387 size
+= ids
* sizeof(u64
);
3394 ev
->attr
.attr
= *attr
;
3395 memcpy(ev
->attr
.id
, id
, ids
* sizeof(u64
));
3397 ev
->attr
.header
.type
= PERF_RECORD_HEADER_ATTR
;
3398 ev
->attr
.header
.size
= (u16
)size
;
3400 if (ev
->attr
.header
.size
== size
)
3401 err
= process(tool
, ev
, NULL
, NULL
);
3410 int perf_event__synthesize_features(struct perf_tool
*tool
,
3411 struct perf_session
*session
,
3412 struct perf_evlist
*evlist
,
3413 perf_event__handler_t process
)
3415 struct perf_header
*header
= &session
->header
;
3417 struct feature_event
*fe
;
3421 sz_hdr
= sizeof(fe
->header
);
3422 sz
= sizeof(union perf_event
);
3423 /* get a nice alignment */
3424 sz
= PERF_ALIGN(sz
, page_size
);
3426 memset(&ff
, 0, sizeof(ff
));
3428 ff
.buf
= malloc(sz
);
3432 ff
.size
= sz
- sz_hdr
;
3434 for_each_set_bit(feat
, header
->adds_features
, HEADER_FEAT_BITS
) {
3435 if (!feat_ops
[feat
].synthesize
) {
3436 pr_debug("No record header feature for header :%d\n", feat
);
3440 ff
.offset
= sizeof(*fe
);
3442 ret
= feat_ops
[feat
].write(&ff
, evlist
);
3443 if (ret
|| ff
.offset
<= (ssize_t
)sizeof(*fe
)) {
3444 pr_debug("Error writing feature\n");
3447 /* ff.buf may have changed due to realloc in do_write() */
3449 memset(fe
, 0, sizeof(*fe
));
3452 fe
->header
.type
= PERF_RECORD_HEADER_FEATURE
;
3453 fe
->header
.size
= ff
.offset
;
3455 ret
= process(tool
, ff
.buf
, NULL
, NULL
);
3462 /* Send HEADER_LAST_FEATURE mark. */
3464 fe
->feat_id
= HEADER_LAST_FEATURE
;
3465 fe
->header
.type
= PERF_RECORD_HEADER_FEATURE
;
3466 fe
->header
.size
= sizeof(*fe
);
3468 ret
= process(tool
, ff
.buf
, NULL
, NULL
);
3474 int perf_event__process_feature(struct perf_session
*session
,
3475 union perf_event
*event
)
3477 struct perf_tool
*tool
= session
->tool
;
3478 struct feat_fd ff
= { .fd
= 0 };
3479 struct feature_event
*fe
= (struct feature_event
*)event
;
3480 int type
= fe
->header
.type
;
3481 u64 feat
= fe
->feat_id
;
3483 if (type
< 0 || type
>= PERF_RECORD_HEADER_MAX
) {
3484 pr_warning("invalid record type %d in pipe-mode\n", type
);
3487 if (feat
== HEADER_RESERVED
|| feat
>= HEADER_LAST_FEATURE
) {
3488 pr_warning("invalid record type %d in pipe-mode\n", type
);
3492 if (!feat_ops
[feat
].process
)
3495 ff
.buf
= (void *)fe
->data
;
3496 ff
.size
= event
->header
.size
- sizeof(event
->header
);
3497 ff
.ph
= &session
->header
;
3499 if (feat_ops
[feat
].process(&ff
, NULL
))
3502 if (!feat_ops
[feat
].print
|| !tool
->show_feat_hdr
)
3505 if (!feat_ops
[feat
].full_only
||
3506 tool
->show_feat_hdr
>= SHOW_FEAT_HEADER_FULL_INFO
) {
3507 feat_ops
[feat
].print(&ff
, stdout
);
3509 fprintf(stdout
, "# %s info available, use -I to display\n",
3510 feat_ops
[feat
].name
);
3516 static struct event_update_event
*
3517 event_update_event__new(size_t size
, u64 type
, u64 id
)
3519 struct event_update_event
*ev
;
3521 size
+= sizeof(*ev
);
3522 size
= PERF_ALIGN(size
, sizeof(u64
));
3526 ev
->header
.type
= PERF_RECORD_EVENT_UPDATE
;
3527 ev
->header
.size
= (u16
)size
;
3535 perf_event__synthesize_event_update_unit(struct perf_tool
*tool
,
3536 struct perf_evsel
*evsel
,
3537 perf_event__handler_t process
)
3539 struct event_update_event
*ev
;
3540 size_t size
= strlen(evsel
->unit
);
3543 ev
= event_update_event__new(size
+ 1, PERF_EVENT_UPDATE__UNIT
, evsel
->id
[0]);
3547 strncpy(ev
->data
, evsel
->unit
, size
);
3548 err
= process(tool
, (union perf_event
*)ev
, NULL
, NULL
);
3554 perf_event__synthesize_event_update_scale(struct perf_tool
*tool
,
3555 struct perf_evsel
*evsel
,
3556 perf_event__handler_t process
)
3558 struct event_update_event
*ev
;
3559 struct event_update_event_scale
*ev_data
;
3562 ev
= event_update_event__new(sizeof(*ev_data
), PERF_EVENT_UPDATE__SCALE
, evsel
->id
[0]);
3566 ev_data
= (struct event_update_event_scale
*) ev
->data
;
3567 ev_data
->scale
= evsel
->scale
;
3568 err
= process(tool
, (union perf_event
*) ev
, NULL
, NULL
);
3574 perf_event__synthesize_event_update_name(struct perf_tool
*tool
,
3575 struct perf_evsel
*evsel
,
3576 perf_event__handler_t process
)
3578 struct event_update_event
*ev
;
3579 size_t len
= strlen(evsel
->name
);
3582 ev
= event_update_event__new(len
+ 1, PERF_EVENT_UPDATE__NAME
, evsel
->id
[0]);
3586 strncpy(ev
->data
, evsel
->name
, len
);
3587 err
= process(tool
, (union perf_event
*) ev
, NULL
, NULL
);
3593 perf_event__synthesize_event_update_cpus(struct perf_tool
*tool
,
3594 struct perf_evsel
*evsel
,
3595 perf_event__handler_t process
)
3597 size_t size
= sizeof(struct event_update_event
);
3598 struct event_update_event
*ev
;
3602 if (!evsel
->own_cpus
)
3605 ev
= cpu_map_data__alloc(evsel
->own_cpus
, &size
, &type
, &max
);
3609 ev
->header
.type
= PERF_RECORD_EVENT_UPDATE
;
3610 ev
->header
.size
= (u16
)size
;
3611 ev
->type
= PERF_EVENT_UPDATE__CPUS
;
3612 ev
->id
= evsel
->id
[0];
3614 cpu_map_data__synthesize((struct cpu_map_data
*) ev
->data
,
3618 err
= process(tool
, (union perf_event
*) ev
, NULL
, NULL
);
3623 size_t perf_event__fprintf_event_update(union perf_event
*event
, FILE *fp
)
3625 struct event_update_event
*ev
= &event
->event_update
;
3626 struct event_update_event_scale
*ev_scale
;
3627 struct event_update_event_cpus
*ev_cpus
;
3628 struct cpu_map
*map
;
3631 ret
= fprintf(fp
, "\n... id: %" PRIu64
"\n", ev
->id
);
3634 case PERF_EVENT_UPDATE__SCALE
:
3635 ev_scale
= (struct event_update_event_scale
*) ev
->data
;
3636 ret
+= fprintf(fp
, "... scale: %f\n", ev_scale
->scale
);
3638 case PERF_EVENT_UPDATE__UNIT
:
3639 ret
+= fprintf(fp
, "... unit: %s\n", ev
->data
);
3641 case PERF_EVENT_UPDATE__NAME
:
3642 ret
+= fprintf(fp
, "... name: %s\n", ev
->data
);
3644 case PERF_EVENT_UPDATE__CPUS
:
3645 ev_cpus
= (struct event_update_event_cpus
*) ev
->data
;
3646 ret
+= fprintf(fp
, "... ");
3648 map
= cpu_map__new_data(&ev_cpus
->cpus
);
3650 ret
+= cpu_map__fprintf(map
, fp
);
3652 ret
+= fprintf(fp
, "failed to get cpus\n");
3655 ret
+= fprintf(fp
, "... unknown type\n");
3662 int perf_event__synthesize_attrs(struct perf_tool
*tool
,
3663 struct perf_evlist
*evlist
,
3664 perf_event__handler_t process
)
3666 struct perf_evsel
*evsel
;
3669 evlist__for_each_entry(evlist
, evsel
) {
3670 err
= perf_event__synthesize_attr(tool
, &evsel
->attr
, evsel
->ids
,
3671 evsel
->id
, process
);
3673 pr_debug("failed to create perf header attribute\n");
3681 static bool has_unit(struct perf_evsel
*counter
)
3683 return counter
->unit
&& *counter
->unit
;
3686 static bool has_scale(struct perf_evsel
*counter
)
3688 return counter
->scale
!= 1;
3691 int perf_event__synthesize_extra_attr(struct perf_tool
*tool
,
3692 struct perf_evlist
*evsel_list
,
3693 perf_event__handler_t process
,
3696 struct perf_evsel
*counter
;
3700 * Synthesize other events stuff not carried within
3701 * attr event - unit, scale, name
3703 evlist__for_each_entry(evsel_list
, counter
) {
3704 if (!counter
->supported
)
3708 * Synthesize unit and scale only if it's defined.
3710 if (has_unit(counter
)) {
3711 err
= perf_event__synthesize_event_update_unit(tool
, counter
, process
);
3713 pr_err("Couldn't synthesize evsel unit.\n");
3718 if (has_scale(counter
)) {
3719 err
= perf_event__synthesize_event_update_scale(tool
, counter
, process
);
3721 pr_err("Couldn't synthesize evsel counter.\n");
3726 if (counter
->own_cpus
) {
3727 err
= perf_event__synthesize_event_update_cpus(tool
, counter
, process
);
3729 pr_err("Couldn't synthesize evsel cpus.\n");
3735 * Name is needed only for pipe output,
3736 * perf.data carries event names.
3739 err
= perf_event__synthesize_event_update_name(tool
, counter
, process
);
3741 pr_err("Couldn't synthesize evsel name.\n");
3749 int perf_event__process_attr(struct perf_tool
*tool __maybe_unused
,
3750 union perf_event
*event
,
3751 struct perf_evlist
**pevlist
)
3754 struct perf_evsel
*evsel
;
3755 struct perf_evlist
*evlist
= *pevlist
;
3757 if (evlist
== NULL
) {
3758 *pevlist
= evlist
= perf_evlist__new();
3763 evsel
= perf_evsel__new(&event
->attr
.attr
);
3767 perf_evlist__add(evlist
, evsel
);
3769 ids
= event
->header
.size
;
3770 ids
-= (void *)&event
->attr
.id
- (void *)event
;
3771 n_ids
= ids
/ sizeof(u64
);
3773 * We don't have the cpu and thread maps on the header, so
3774 * for allocating the perf_sample_id table we fake 1 cpu and
3775 * hattr->ids threads.
3777 if (perf_evsel__alloc_id(evsel
, 1, n_ids
))
3780 for (i
= 0; i
< n_ids
; i
++) {
3781 perf_evlist__id_add(evlist
, evsel
, 0, i
, event
->attr
.id
[i
]);
3787 int perf_event__process_event_update(struct perf_tool
*tool __maybe_unused
,
3788 union perf_event
*event
,
3789 struct perf_evlist
**pevlist
)
3791 struct event_update_event
*ev
= &event
->event_update
;
3792 struct event_update_event_scale
*ev_scale
;
3793 struct event_update_event_cpus
*ev_cpus
;
3794 struct perf_evlist
*evlist
;
3795 struct perf_evsel
*evsel
;
3796 struct cpu_map
*map
;
3798 if (!pevlist
|| *pevlist
== NULL
)
3803 evsel
= perf_evlist__id2evsel(evlist
, ev
->id
);
3808 case PERF_EVENT_UPDATE__UNIT
:
3809 evsel
->unit
= strdup(ev
->data
);
3811 case PERF_EVENT_UPDATE__NAME
:
3812 evsel
->name
= strdup(ev
->data
);
3814 case PERF_EVENT_UPDATE__SCALE
:
3815 ev_scale
= (struct event_update_event_scale
*) ev
->data
;
3816 evsel
->scale
= ev_scale
->scale
;
3818 case PERF_EVENT_UPDATE__CPUS
:
3819 ev_cpus
= (struct event_update_event_cpus
*) ev
->data
;
3821 map
= cpu_map__new_data(&ev_cpus
->cpus
);
3823 evsel
->own_cpus
= map
;
3825 pr_err("failed to get event_update cpus\n");
3833 int perf_event__synthesize_tracing_data(struct perf_tool
*tool
, int fd
,
3834 struct perf_evlist
*evlist
,
3835 perf_event__handler_t process
)
3837 union perf_event ev
;
3838 struct tracing_data
*tdata
;
3839 ssize_t size
= 0, aligned_size
= 0, padding
;
3841 int err __maybe_unused
= 0;
3844 * We are going to store the size of the data followed
3845 * by the data contents. Since the fd descriptor is a pipe,
3846 * we cannot seek back to store the size of the data once
3847 * we know it. Instead we:
3849 * - write the tracing data to the temp file
3850 * - get/write the data size to pipe
3851 * - write the tracing data from the temp file
3854 tdata
= tracing_data_get(&evlist
->entries
, fd
, true);
3858 memset(&ev
, 0, sizeof(ev
));
3860 ev
.tracing_data
.header
.type
= PERF_RECORD_HEADER_TRACING_DATA
;
3862 aligned_size
= PERF_ALIGN(size
, sizeof(u64
));
3863 padding
= aligned_size
- size
;
3864 ev
.tracing_data
.header
.size
= sizeof(ev
.tracing_data
);
3865 ev
.tracing_data
.size
= aligned_size
;
3867 process(tool
, &ev
, NULL
, NULL
);
3870 * The put function will copy all the tracing data
3871 * stored in temp file to the pipe.
3873 tracing_data_put(tdata
);
3875 ff
= (struct feat_fd
){ .fd
= fd
};
3876 if (write_padded(&ff
, NULL
, 0, padding
))
3879 return aligned_size
;
3882 int perf_event__process_tracing_data(struct perf_session
*session
,
3883 union perf_event
*event
)
3885 ssize_t size_read
, padding
, size
= event
->tracing_data
.size
;
3886 int fd
= perf_data__fd(session
->data
);
3887 off_t offset
= lseek(fd
, 0, SEEK_CUR
);
3890 /* setup for reading amidst mmap */
3891 lseek(fd
, offset
+ sizeof(struct tracing_data_event
),
3894 size_read
= trace_report(fd
, &session
->tevent
,
3896 padding
= PERF_ALIGN(size_read
, sizeof(u64
)) - size_read
;
3898 if (readn(fd
, buf
, padding
) < 0) {
3899 pr_err("%s: reading input file", __func__
);
3902 if (session
->repipe
) {
3903 int retw
= write(STDOUT_FILENO
, buf
, padding
);
3904 if (retw
<= 0 || retw
!= padding
) {
3905 pr_err("%s: repiping tracing data padding", __func__
);
3910 if (size_read
+ padding
!= size
) {
3911 pr_err("%s: tracing data size mismatch", __func__
);
3915 perf_evlist__prepare_tracepoint_events(session
->evlist
,
3916 session
->tevent
.pevent
);
3918 return size_read
+ padding
;
3921 int perf_event__synthesize_build_id(struct perf_tool
*tool
,
3922 struct dso
*pos
, u16 misc
,
3923 perf_event__handler_t process
,
3924 struct machine
*machine
)
3926 union perf_event ev
;
3933 memset(&ev
, 0, sizeof(ev
));
3935 len
= pos
->long_name_len
+ 1;
3936 len
= PERF_ALIGN(len
, NAME_ALIGN
);
3937 memcpy(&ev
.build_id
.build_id
, pos
->build_id
, sizeof(pos
->build_id
));
3938 ev
.build_id
.header
.type
= PERF_RECORD_HEADER_BUILD_ID
;
3939 ev
.build_id
.header
.misc
= misc
;
3940 ev
.build_id
.pid
= machine
->pid
;
3941 ev
.build_id
.header
.size
= sizeof(ev
.build_id
) + len
;
3942 memcpy(&ev
.build_id
.filename
, pos
->long_name
, pos
->long_name_len
);
3944 err
= process(tool
, &ev
, NULL
, machine
);
3949 int perf_event__process_build_id(struct perf_session
*session
,
3950 union perf_event
*event
)
3952 __event_process_build_id(&event
->build_id
,
3953 event
->build_id
.filename
,