7 #include <linux/list.h>
8 #include <linux/kernel.h>
9 #include <linux/bitops.h>
10 #include <sys/utsname.h>
16 #include "trace-event.h"
29 * must be a numerical value to let the endianness
30 * determine the memory layout. That way we are able
31 * to detect endianness when reading the perf.data file
34 * we check for legacy (PERFFILE) format.
36 static const char *__perf_magic1
= "PERFFILE";
37 static const u64 __perf_magic2
= 0x32454c4946524550ULL
;
38 static const u64 __perf_magic2_sw
= 0x50455246494c4532ULL
;
40 #define PERF_MAGIC __perf_magic2
42 struct perf_file_attr
{
43 struct perf_event_attr attr
;
44 struct perf_file_section ids
;
47 void perf_header__set_feat(struct perf_header
*header
, int feat
)
49 set_bit(feat
, header
->adds_features
);
52 void perf_header__clear_feat(struct perf_header
*header
, int feat
)
54 clear_bit(feat
, header
->adds_features
);
57 bool perf_header__has_feat(const struct perf_header
*header
, int feat
)
59 return test_bit(feat
, header
->adds_features
);
62 static int do_write(int fd
, const void *buf
, size_t size
)
65 int ret
= write(fd
, buf
, size
);
77 int write_padded(int fd
, const void *bf
, size_t count
, size_t count_aligned
)
79 static const char zero_buf
[NAME_ALIGN
];
80 int err
= do_write(fd
, bf
, count
);
83 err
= do_write(fd
, zero_buf
, count_aligned
- count
);
88 #define string_size(str) \
89 (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
91 static int do_write_string(int fd
, const char *str
)
96 olen
= strlen(str
) + 1;
97 len
= PERF_ALIGN(olen
, NAME_ALIGN
);
99 /* write len, incl. \0 */
100 ret
= do_write(fd
, &len
, sizeof(len
));
104 return write_padded(fd
, str
, olen
, len
);
107 static char *do_read_string(int fd
, struct perf_header
*ph
)
113 sz
= readn(fd
, &len
, sizeof(len
));
114 if (sz
< (ssize_t
)sizeof(len
))
124 ret
= readn(fd
, buf
, len
);
125 if (ret
== (ssize_t
)len
) {
127 * strings are padded by zeroes
128 * thus the actual strlen of buf
129 * may be less than len
138 static int write_tracing_data(int fd
, struct perf_header
*h __maybe_unused
,
139 struct perf_evlist
*evlist
)
141 return read_tracing_data(fd
, &evlist
->entries
);
145 static int write_build_id(int fd
, struct perf_header
*h
,
146 struct perf_evlist
*evlist __maybe_unused
)
148 struct perf_session
*session
;
151 session
= container_of(h
, struct perf_session
, header
);
153 if (!perf_session__read_build_ids(session
, true))
156 err
= perf_session__write_buildid_table(session
, fd
);
158 pr_debug("failed to write buildid table\n");
161 perf_session__cache_build_ids(session
);
166 static int write_hostname(int fd
, struct perf_header
*h __maybe_unused
,
167 struct perf_evlist
*evlist __maybe_unused
)
176 return do_write_string(fd
, uts
.nodename
);
179 static int write_osrelease(int fd
, struct perf_header
*h __maybe_unused
,
180 struct perf_evlist
*evlist __maybe_unused
)
189 return do_write_string(fd
, uts
.release
);
192 static int write_arch(int fd
, struct perf_header
*h __maybe_unused
,
193 struct perf_evlist
*evlist __maybe_unused
)
202 return do_write_string(fd
, uts
.machine
);
205 static int write_version(int fd
, struct perf_header
*h __maybe_unused
,
206 struct perf_evlist
*evlist __maybe_unused
)
208 return do_write_string(fd
, perf_version_string
);
211 static int __write_cpudesc(int fd
, const char *cpuinfo_proc
)
216 const char *search
= cpuinfo_proc
;
223 file
= fopen("/proc/cpuinfo", "r");
227 while (getline(&buf
, &len
, file
) > 0) {
228 ret
= strncmp(buf
, search
, strlen(search
));
240 p
= strchr(buf
, ':');
241 if (p
&& *(p
+1) == ' ' && *(p
+2))
247 /* squash extra space characters (branding string) */
254 while (*q
&& isspace(*q
))
257 while ((*r
++ = *q
++));
261 ret
= do_write_string(fd
, s
);
268 static int write_cpudesc(int fd
, struct perf_header
*h __maybe_unused
,
269 struct perf_evlist
*evlist __maybe_unused
)
272 #define CPUINFO_PROC {"model name", }
274 const char *cpuinfo_procs
[] = CPUINFO_PROC
;
277 for (i
= 0; i
< ARRAY_SIZE(cpuinfo_procs
); i
++) {
279 ret
= __write_cpudesc(fd
, cpuinfo_procs
[i
]);
287 static int write_nrcpus(int fd
, struct perf_header
*h __maybe_unused
,
288 struct perf_evlist
*evlist __maybe_unused
)
294 nr
= sysconf(_SC_NPROCESSORS_CONF
);
298 nrc
= (u32
)(nr
& UINT_MAX
);
300 nr
= sysconf(_SC_NPROCESSORS_ONLN
);
304 nra
= (u32
)(nr
& UINT_MAX
);
306 ret
= do_write(fd
, &nrc
, sizeof(nrc
));
310 return do_write(fd
, &nra
, sizeof(nra
));
313 static int write_event_desc(int fd
, struct perf_header
*h __maybe_unused
,
314 struct perf_evlist
*evlist
)
316 struct perf_evsel
*evsel
;
320 nre
= evlist
->nr_entries
;
323 * write number of events
325 ret
= do_write(fd
, &nre
, sizeof(nre
));
330 * size of perf_event_attr struct
332 sz
= (u32
)sizeof(evsel
->attr
);
333 ret
= do_write(fd
, &sz
, sizeof(sz
));
337 evlist__for_each(evlist
, evsel
) {
338 ret
= do_write(fd
, &evsel
->attr
, sz
);
342 * write number of unique id per event
343 * there is one id per instance of an event
345 * copy into an nri to be independent of the
349 ret
= do_write(fd
, &nri
, sizeof(nri
));
354 * write event string as passed on cmdline
356 ret
= do_write_string(fd
, perf_evsel__name(evsel
));
360 * write unique ids for this event
362 ret
= do_write(fd
, evsel
->id
, evsel
->ids
* sizeof(u64
));
369 static int write_cmdline(int fd
, struct perf_header
*h __maybe_unused
,
370 struct perf_evlist
*evlist __maybe_unused
)
372 char buf
[MAXPATHLEN
];
378 * actual atual path to perf binary
380 sprintf(proc
, "/proc/%d/exe", getpid());
381 ret
= readlink(proc
, buf
, sizeof(buf
));
385 /* readlink() does not add null termination */
388 /* account for binary path */
389 n
= perf_env
.nr_cmdline
+ 1;
391 ret
= do_write(fd
, &n
, sizeof(n
));
395 ret
= do_write_string(fd
, buf
);
399 for (i
= 0 ; i
< perf_env
.nr_cmdline
; i
++) {
400 ret
= do_write_string(fd
, perf_env
.cmdline_argv
[i
]);
407 #define CORE_SIB_FMT \
408 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
409 #define THRD_SIB_FMT \
410 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
416 char **core_siblings
;
417 char **thread_siblings
;
420 static int build_cpu_topo(struct cpu_topo
*tp
, int cpu
)
423 char filename
[MAXPATHLEN
];
424 char *buf
= NULL
, *p
;
430 sprintf(filename
, CORE_SIB_FMT
, cpu
);
431 fp
= fopen(filename
, "r");
435 sret
= getline(&buf
, &len
, fp
);
440 p
= strchr(buf
, '\n');
444 for (i
= 0; i
< tp
->core_sib
; i
++) {
445 if (!strcmp(buf
, tp
->core_siblings
[i
]))
448 if (i
== tp
->core_sib
) {
449 tp
->core_siblings
[i
] = buf
;
457 sprintf(filename
, THRD_SIB_FMT
, cpu
);
458 fp
= fopen(filename
, "r");
462 if (getline(&buf
, &len
, fp
) <= 0)
465 p
= strchr(buf
, '\n');
469 for (i
= 0; i
< tp
->thread_sib
; i
++) {
470 if (!strcmp(buf
, tp
->thread_siblings
[i
]))
473 if (i
== tp
->thread_sib
) {
474 tp
->thread_siblings
[i
] = buf
;
486 static void free_cpu_topo(struct cpu_topo
*tp
)
493 for (i
= 0 ; i
< tp
->core_sib
; i
++)
494 zfree(&tp
->core_siblings
[i
]);
496 for (i
= 0 ; i
< tp
->thread_sib
; i
++)
497 zfree(&tp
->thread_siblings
[i
]);
502 static struct cpu_topo
*build_cpu_topology(void)
511 ncpus
= sysconf(_SC_NPROCESSORS_CONF
);
515 nr
= (u32
)(ncpus
& UINT_MAX
);
517 sz
= nr
* sizeof(char *);
519 addr
= calloc(1, sizeof(*tp
) + 2 * sz
);
526 tp
->core_siblings
= addr
;
528 tp
->thread_siblings
= addr
;
530 for (i
= 0; i
< nr
; i
++) {
531 ret
= build_cpu_topo(tp
, i
);
542 static int write_cpu_topology(int fd
, struct perf_header
*h __maybe_unused
,
543 struct perf_evlist
*evlist __maybe_unused
)
549 tp
= build_cpu_topology();
553 ret
= do_write(fd
, &tp
->core_sib
, sizeof(tp
->core_sib
));
557 for (i
= 0; i
< tp
->core_sib
; i
++) {
558 ret
= do_write_string(fd
, tp
->core_siblings
[i
]);
562 ret
= do_write(fd
, &tp
->thread_sib
, sizeof(tp
->thread_sib
));
566 for (i
= 0; i
< tp
->thread_sib
; i
++) {
567 ret
= do_write_string(fd
, tp
->thread_siblings
[i
]);
572 ret
= perf_env__read_cpu_topology_map(&perf_env
);
576 for (j
= 0; j
< perf_env
.nr_cpus_avail
; j
++) {
577 ret
= do_write(fd
, &perf_env
.cpu
[j
].core_id
,
578 sizeof(perf_env
.cpu
[j
].core_id
));
581 ret
= do_write(fd
, &perf_env
.cpu
[j
].socket_id
,
582 sizeof(perf_env
.cpu
[j
].socket_id
));
593 static int write_total_mem(int fd
, struct perf_header
*h __maybe_unused
,
594 struct perf_evlist
*evlist __maybe_unused
)
602 fp
= fopen("/proc/meminfo", "r");
606 while (getline(&buf
, &len
, fp
) > 0) {
607 ret
= strncmp(buf
, "MemTotal:", 9);
612 n
= sscanf(buf
, "%*s %"PRIu64
, &mem
);
614 ret
= do_write(fd
, &mem
, sizeof(mem
));
622 static int write_topo_node(int fd
, int node
)
624 char str
[MAXPATHLEN
];
626 char *buf
= NULL
, *p
;
629 u64 mem_total
, mem_free
, mem
;
632 sprintf(str
, "/sys/devices/system/node/node%d/meminfo", node
);
633 fp
= fopen(str
, "r");
637 while (getline(&buf
, &len
, fp
) > 0) {
638 /* skip over invalid lines */
639 if (!strchr(buf
, ':'))
641 if (sscanf(buf
, "%*s %*d %31s %"PRIu64
, field
, &mem
) != 2)
643 if (!strcmp(field
, "MemTotal:"))
645 if (!strcmp(field
, "MemFree:"))
652 ret
= do_write(fd
, &mem_total
, sizeof(u64
));
656 ret
= do_write(fd
, &mem_free
, sizeof(u64
));
661 sprintf(str
, "/sys/devices/system/node/node%d/cpulist", node
);
663 fp
= fopen(str
, "r");
667 if (getline(&buf
, &len
, fp
) <= 0)
670 p
= strchr(buf
, '\n');
674 ret
= do_write_string(fd
, buf
);
682 static int write_numa_topology(int fd
, struct perf_header
*h __maybe_unused
,
683 struct perf_evlist
*evlist __maybe_unused
)
688 struct cpu_map
*node_map
= NULL
;
693 fp
= fopen("/sys/devices/system/node/online", "r");
697 if (getline(&buf
, &len
, fp
) <= 0)
700 c
= strchr(buf
, '\n');
704 node_map
= cpu_map__new(buf
);
708 nr
= (u32
)node_map
->nr
;
710 ret
= do_write(fd
, &nr
, sizeof(nr
));
714 for (i
= 0; i
< nr
; i
++) {
715 j
= (u32
)node_map
->map
[i
];
716 ret
= do_write(fd
, &j
, sizeof(j
));
720 ret
= write_topo_node(fd
, i
);
727 cpu_map__put(node_map
);
734 * struct pmu_mappings {
743 static int write_pmu_mappings(int fd
, struct perf_header
*h __maybe_unused
,
744 struct perf_evlist
*evlist __maybe_unused
)
746 struct perf_pmu
*pmu
= NULL
;
747 off_t offset
= lseek(fd
, 0, SEEK_CUR
);
751 /* write real pmu_num later */
752 ret
= do_write(fd
, &pmu_num
, sizeof(pmu_num
));
756 while ((pmu
= perf_pmu__scan(pmu
))) {
761 ret
= do_write(fd
, &pmu
->type
, sizeof(pmu
->type
));
765 ret
= do_write_string(fd
, pmu
->name
);
770 if (pwrite(fd
, &pmu_num
, sizeof(pmu_num
), offset
) != sizeof(pmu_num
)) {
772 lseek(fd
, offset
, SEEK_SET
);
782 * struct group_descs {
784 * struct group_desc {
791 static int write_group_desc(int fd
, struct perf_header
*h __maybe_unused
,
792 struct perf_evlist
*evlist
)
794 u32 nr_groups
= evlist
->nr_groups
;
795 struct perf_evsel
*evsel
;
798 ret
= do_write(fd
, &nr_groups
, sizeof(nr_groups
));
802 evlist__for_each(evlist
, evsel
) {
803 if (perf_evsel__is_group_leader(evsel
) &&
804 evsel
->nr_members
> 1) {
805 const char *name
= evsel
->group_name
?: "{anon_group}";
806 u32 leader_idx
= evsel
->idx
;
807 u32 nr_members
= evsel
->nr_members
;
809 ret
= do_write_string(fd
, name
);
813 ret
= do_write(fd
, &leader_idx
, sizeof(leader_idx
));
817 ret
= do_write(fd
, &nr_members
, sizeof(nr_members
));
826 * default get_cpuid(): nothing gets recorded
827 * actual implementation must be in arch/$(ARCH)/util/header.c
829 int __attribute__ ((weak
)) get_cpuid(char *buffer __maybe_unused
,
830 size_t sz __maybe_unused
)
835 static int write_cpuid(int fd
, struct perf_header
*h __maybe_unused
,
836 struct perf_evlist
*evlist __maybe_unused
)
841 ret
= get_cpuid(buffer
, sizeof(buffer
));
847 return do_write_string(fd
, buffer
);
850 static int write_branch_stack(int fd __maybe_unused
,
851 struct perf_header
*h __maybe_unused
,
852 struct perf_evlist
*evlist __maybe_unused
)
857 static int write_auxtrace(int fd
, struct perf_header
*h
,
858 struct perf_evlist
*evlist __maybe_unused
)
860 struct perf_session
*session
;
863 session
= container_of(h
, struct perf_session
, header
);
865 err
= auxtrace_index__write(fd
, &session
->auxtrace_index
);
867 pr_err("Failed to write auxtrace index\n");
871 static void print_hostname(struct perf_header
*ph
, int fd __maybe_unused
,
874 fprintf(fp
, "# hostname : %s\n", ph
->env
.hostname
);
877 static void print_osrelease(struct perf_header
*ph
, int fd __maybe_unused
,
880 fprintf(fp
, "# os release : %s\n", ph
->env
.os_release
);
883 static void print_arch(struct perf_header
*ph
, int fd __maybe_unused
, FILE *fp
)
885 fprintf(fp
, "# arch : %s\n", ph
->env
.arch
);
888 static void print_cpudesc(struct perf_header
*ph
, int fd __maybe_unused
,
891 fprintf(fp
, "# cpudesc : %s\n", ph
->env
.cpu_desc
);
894 static void print_nrcpus(struct perf_header
*ph
, int fd __maybe_unused
,
897 fprintf(fp
, "# nrcpus online : %u\n", ph
->env
.nr_cpus_online
);
898 fprintf(fp
, "# nrcpus avail : %u\n", ph
->env
.nr_cpus_avail
);
901 static void print_version(struct perf_header
*ph
, int fd __maybe_unused
,
904 fprintf(fp
, "# perf version : %s\n", ph
->env
.version
);
907 static void print_cmdline(struct perf_header
*ph
, int fd __maybe_unused
,
912 nr
= ph
->env
.nr_cmdline
;
914 fprintf(fp
, "# cmdline : ");
916 for (i
= 0; i
< nr
; i
++)
917 fprintf(fp
, "%s ", ph
->env
.cmdline_argv
[i
]);
921 static void print_cpu_topology(struct perf_header
*ph
, int fd __maybe_unused
,
926 int cpu_nr
= ph
->env
.nr_cpus_online
;
928 nr
= ph
->env
.nr_sibling_cores
;
929 str
= ph
->env
.sibling_cores
;
931 for (i
= 0; i
< nr
; i
++) {
932 fprintf(fp
, "# sibling cores : %s\n", str
);
933 str
+= strlen(str
) + 1;
936 nr
= ph
->env
.nr_sibling_threads
;
937 str
= ph
->env
.sibling_threads
;
939 for (i
= 0; i
< nr
; i
++) {
940 fprintf(fp
, "# sibling threads : %s\n", str
);
941 str
+= strlen(str
) + 1;
944 if (ph
->env
.cpu
!= NULL
) {
945 for (i
= 0; i
< cpu_nr
; i
++)
946 fprintf(fp
, "# CPU %d: Core ID %d, Socket ID %d\n", i
,
947 ph
->env
.cpu
[i
].core_id
, ph
->env
.cpu
[i
].socket_id
);
949 fprintf(fp
, "# Core ID and Socket ID information is not available\n");
952 static void free_event_desc(struct perf_evsel
*events
)
954 struct perf_evsel
*evsel
;
959 for (evsel
= events
; evsel
->attr
.size
; evsel
++) {
967 static struct perf_evsel
*
968 read_event_desc(struct perf_header
*ph
, int fd
)
970 struct perf_evsel
*evsel
, *events
= NULL
;
973 u32 nre
, sz
, nr
, i
, j
;
977 /* number of events */
978 ret
= readn(fd
, &nre
, sizeof(nre
));
979 if (ret
!= (ssize_t
)sizeof(nre
))
985 ret
= readn(fd
, &sz
, sizeof(sz
));
986 if (ret
!= (ssize_t
)sizeof(sz
))
992 /* buffer to hold on file attr struct */
997 /* the last event terminates with evsel->attr.size == 0: */
998 events
= calloc(nre
+ 1, sizeof(*events
));
1002 msz
= sizeof(evsel
->attr
);
1006 for (i
= 0, evsel
= events
; i
< nre
; evsel
++, i
++) {
1010 * must read entire on-file attr struct to
1011 * sync up with layout.
1013 ret
= readn(fd
, buf
, sz
);
1014 if (ret
!= (ssize_t
)sz
)
1018 perf_event__attr_swap(buf
);
1020 memcpy(&evsel
->attr
, buf
, msz
);
1022 ret
= readn(fd
, &nr
, sizeof(nr
));
1023 if (ret
!= (ssize_t
)sizeof(nr
))
1026 if (ph
->needs_swap
) {
1028 evsel
->needs_swap
= true;
1031 evsel
->name
= do_read_string(fd
, ph
);
1036 id
= calloc(nr
, sizeof(*id
));
1042 for (j
= 0 ; j
< nr
; j
++) {
1043 ret
= readn(fd
, id
, sizeof(*id
));
1044 if (ret
!= (ssize_t
)sizeof(*id
))
1047 *id
= bswap_64(*id
);
1055 free_event_desc(events
);
1060 static int __desc_attr__fprintf(FILE *fp
, const char *name
, const char *val
,
1061 void *priv
__attribute__((unused
)))
1063 return fprintf(fp
, ", %s = %s", name
, val
);
1066 static void print_event_desc(struct perf_header
*ph
, int fd
, FILE *fp
)
1068 struct perf_evsel
*evsel
, *events
= read_event_desc(ph
, fd
);
1073 fprintf(fp
, "# event desc: not available or unable to read\n");
1077 for (evsel
= events
; evsel
->attr
.size
; evsel
++) {
1078 fprintf(fp
, "# event : name = %s, ", evsel
->name
);
1081 fprintf(fp
, ", id = {");
1082 for (j
= 0, id
= evsel
->id
; j
< evsel
->ids
; j
++, id
++) {
1085 fprintf(fp
, " %"PRIu64
, *id
);
1090 perf_event_attr__fprintf(fp
, &evsel
->attr
, __desc_attr__fprintf
, NULL
);
1095 free_event_desc(events
);
1098 static void print_total_mem(struct perf_header
*ph
, int fd __maybe_unused
,
1101 fprintf(fp
, "# total memory : %Lu kB\n", ph
->env
.total_mem
);
1104 static void print_numa_topology(struct perf_header
*ph
, int fd __maybe_unused
,
1109 uint64_t mem_total
, mem_free
;
1112 nr
= ph
->env
.nr_numa_nodes
;
1113 str
= ph
->env
.numa_nodes
;
1115 for (i
= 0; i
< nr
; i
++) {
1117 c
= strtoul(str
, &tmp
, 0);
1122 mem_total
= strtoull(str
, &tmp
, 0);
1127 mem_free
= strtoull(str
, &tmp
, 0);
1131 fprintf(fp
, "# node%u meminfo : total = %"PRIu64
" kB,"
1132 " free = %"PRIu64
" kB\n",
1133 c
, mem_total
, mem_free
);
1136 fprintf(fp
, "# node%u cpu list : %s\n", c
, str
);
1138 str
+= strlen(str
) + 1;
1142 fprintf(fp
, "# numa topology : not available\n");
1145 static void print_cpuid(struct perf_header
*ph
, int fd __maybe_unused
, FILE *fp
)
1147 fprintf(fp
, "# cpuid : %s\n", ph
->env
.cpuid
);
1150 static void print_branch_stack(struct perf_header
*ph __maybe_unused
,
1151 int fd __maybe_unused
, FILE *fp
)
1153 fprintf(fp
, "# contains samples with branch stack\n");
1156 static void print_auxtrace(struct perf_header
*ph __maybe_unused
,
1157 int fd __maybe_unused
, FILE *fp
)
1159 fprintf(fp
, "# contains AUX area data (e.g. instruction trace)\n");
1162 static void print_pmu_mappings(struct perf_header
*ph
, int fd __maybe_unused
,
1165 const char *delimiter
= "# pmu mappings: ";
1170 pmu_num
= ph
->env
.nr_pmu_mappings
;
1172 fprintf(fp
, "# pmu mappings: not available\n");
1176 str
= ph
->env
.pmu_mappings
;
1179 type
= strtoul(str
, &tmp
, 0);
1184 fprintf(fp
, "%s%s = %" PRIu32
, delimiter
, str
, type
);
1187 str
+= strlen(str
) + 1;
1196 fprintf(fp
, "# pmu mappings: unable to read\n");
1199 static void print_group_desc(struct perf_header
*ph
, int fd __maybe_unused
,
1202 struct perf_session
*session
;
1203 struct perf_evsel
*evsel
;
1206 session
= container_of(ph
, struct perf_session
, header
);
1208 evlist__for_each(session
->evlist
, evsel
) {
1209 if (perf_evsel__is_group_leader(evsel
) &&
1210 evsel
->nr_members
> 1) {
1211 fprintf(fp
, "# group: %s{%s", evsel
->group_name
?: "",
1212 perf_evsel__name(evsel
));
1214 nr
= evsel
->nr_members
- 1;
1216 fprintf(fp
, ",%s", perf_evsel__name(evsel
));
1224 static int __event_process_build_id(struct build_id_event
*bev
,
1226 struct perf_session
*session
)
1229 struct machine
*machine
;
1232 enum dso_kernel_type dso_type
;
1234 machine
= perf_session__findnew_machine(session
, bev
->pid
);
1238 cpumode
= bev
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
1241 case PERF_RECORD_MISC_KERNEL
:
1242 dso_type
= DSO_TYPE_KERNEL
;
1244 case PERF_RECORD_MISC_GUEST_KERNEL
:
1245 dso_type
= DSO_TYPE_GUEST_KERNEL
;
1247 case PERF_RECORD_MISC_USER
:
1248 case PERF_RECORD_MISC_GUEST_USER
:
1249 dso_type
= DSO_TYPE_USER
;
1255 dso
= machine__findnew_dso(machine
, filename
);
1257 char sbuild_id
[BUILD_ID_SIZE
* 2 + 1];
1259 dso__set_build_id(dso
, &bev
->build_id
);
1261 if (!is_kernel_module(filename
, cpumode
))
1262 dso
->kernel
= dso_type
;
1264 build_id__sprintf(dso
->build_id
, sizeof(dso
->build_id
),
1266 pr_debug("build id event received for %s: %s\n",
1267 dso
->long_name
, sbuild_id
);
1276 static int perf_header__read_build_ids_abi_quirk(struct perf_header
*header
,
1277 int input
, u64 offset
, u64 size
)
1279 struct perf_session
*session
= container_of(header
, struct perf_session
, header
);
1281 struct perf_event_header header
;
1282 u8 build_id
[PERF_ALIGN(BUILD_ID_SIZE
, sizeof(u64
))];
1285 struct build_id_event bev
;
1286 char filename
[PATH_MAX
];
1287 u64 limit
= offset
+ size
;
1289 while (offset
< limit
) {
1292 if (readn(input
, &old_bev
, sizeof(old_bev
)) != sizeof(old_bev
))
1295 if (header
->needs_swap
)
1296 perf_event_header__bswap(&old_bev
.header
);
1298 len
= old_bev
.header
.size
- sizeof(old_bev
);
1299 if (readn(input
, filename
, len
) != len
)
1302 bev
.header
= old_bev
.header
;
1305 * As the pid is the missing value, we need to fill
1306 * it properly. The header.misc value give us nice hint.
1308 bev
.pid
= HOST_KERNEL_ID
;
1309 if (bev
.header
.misc
== PERF_RECORD_MISC_GUEST_USER
||
1310 bev
.header
.misc
== PERF_RECORD_MISC_GUEST_KERNEL
)
1311 bev
.pid
= DEFAULT_GUEST_KERNEL_ID
;
1313 memcpy(bev
.build_id
, old_bev
.build_id
, sizeof(bev
.build_id
));
1314 __event_process_build_id(&bev
, filename
, session
);
1316 offset
+= bev
.header
.size
;
1322 static int perf_header__read_build_ids(struct perf_header
*header
,
1323 int input
, u64 offset
, u64 size
)
1325 struct perf_session
*session
= container_of(header
, struct perf_session
, header
);
1326 struct build_id_event bev
;
1327 char filename
[PATH_MAX
];
1328 u64 limit
= offset
+ size
, orig_offset
= offset
;
1331 while (offset
< limit
) {
1334 if (readn(input
, &bev
, sizeof(bev
)) != sizeof(bev
))
1337 if (header
->needs_swap
)
1338 perf_event_header__bswap(&bev
.header
);
1340 len
= bev
.header
.size
- sizeof(bev
);
1341 if (readn(input
, filename
, len
) != len
)
1344 * The a1645ce1 changeset:
1346 * "perf: 'perf kvm' tool for monitoring guest performance from host"
1348 * Added a field to struct build_id_event that broke the file
1351 * Since the kernel build-id is the first entry, process the
1352 * table using the old format if the well known
1353 * '[kernel.kallsyms]' string for the kernel build-id has the
1354 * first 4 characters chopped off (where the pid_t sits).
1356 if (memcmp(filename
, "nel.kallsyms]", 13) == 0) {
1357 if (lseek(input
, orig_offset
, SEEK_SET
) == (off_t
)-1)
1359 return perf_header__read_build_ids_abi_quirk(header
, input
, offset
, size
);
1362 __event_process_build_id(&bev
, filename
, session
);
1364 offset
+= bev
.header
.size
;
1371 static int process_tracing_data(struct perf_file_section
*section __maybe_unused
,
1372 struct perf_header
*ph __maybe_unused
,
1375 ssize_t ret
= trace_report(fd
, data
, false);
1376 return ret
< 0 ? -1 : 0;
1379 static int process_build_id(struct perf_file_section
*section
,
1380 struct perf_header
*ph
, int fd
,
1381 void *data __maybe_unused
)
1383 if (perf_header__read_build_ids(ph
, fd
, section
->offset
, section
->size
))
1384 pr_debug("Failed to read buildids, continuing...\n");
1388 static int process_hostname(struct perf_file_section
*section __maybe_unused
,
1389 struct perf_header
*ph
, int fd
,
1390 void *data __maybe_unused
)
1392 ph
->env
.hostname
= do_read_string(fd
, ph
);
1393 return ph
->env
.hostname
? 0 : -ENOMEM
;
1396 static int process_osrelease(struct perf_file_section
*section __maybe_unused
,
1397 struct perf_header
*ph
, int fd
,
1398 void *data __maybe_unused
)
1400 ph
->env
.os_release
= do_read_string(fd
, ph
);
1401 return ph
->env
.os_release
? 0 : -ENOMEM
;
1404 static int process_version(struct perf_file_section
*section __maybe_unused
,
1405 struct perf_header
*ph
, int fd
,
1406 void *data __maybe_unused
)
1408 ph
->env
.version
= do_read_string(fd
, ph
);
1409 return ph
->env
.version
? 0 : -ENOMEM
;
1412 static int process_arch(struct perf_file_section
*section __maybe_unused
,
1413 struct perf_header
*ph
, int fd
,
1414 void *data __maybe_unused
)
1416 ph
->env
.arch
= do_read_string(fd
, ph
);
1417 return ph
->env
.arch
? 0 : -ENOMEM
;
1420 static int process_nrcpus(struct perf_file_section
*section __maybe_unused
,
1421 struct perf_header
*ph
, int fd
,
1422 void *data __maybe_unused
)
1427 ret
= readn(fd
, &nr
, sizeof(nr
));
1428 if (ret
!= sizeof(nr
))
1434 ph
->env
.nr_cpus_avail
= nr
;
1436 ret
= readn(fd
, &nr
, sizeof(nr
));
1437 if (ret
!= sizeof(nr
))
1443 ph
->env
.nr_cpus_online
= nr
;
1447 static int process_cpudesc(struct perf_file_section
*section __maybe_unused
,
1448 struct perf_header
*ph
, int fd
,
1449 void *data __maybe_unused
)
1451 ph
->env
.cpu_desc
= do_read_string(fd
, ph
);
1452 return ph
->env
.cpu_desc
? 0 : -ENOMEM
;
1455 static int process_cpuid(struct perf_file_section
*section __maybe_unused
,
1456 struct perf_header
*ph
, int fd
,
1457 void *data __maybe_unused
)
1459 ph
->env
.cpuid
= do_read_string(fd
, ph
);
1460 return ph
->env
.cpuid
? 0 : -ENOMEM
;
1463 static int process_total_mem(struct perf_file_section
*section __maybe_unused
,
1464 struct perf_header
*ph
, int fd
,
1465 void *data __maybe_unused
)
1470 ret
= readn(fd
, &mem
, sizeof(mem
));
1471 if (ret
!= sizeof(mem
))
1475 mem
= bswap_64(mem
);
1477 ph
->env
.total_mem
= mem
;
1481 static struct perf_evsel
*
1482 perf_evlist__find_by_index(struct perf_evlist
*evlist
, int idx
)
1484 struct perf_evsel
*evsel
;
1486 evlist__for_each(evlist
, evsel
) {
1487 if (evsel
->idx
== idx
)
1495 perf_evlist__set_event_name(struct perf_evlist
*evlist
,
1496 struct perf_evsel
*event
)
1498 struct perf_evsel
*evsel
;
1503 evsel
= perf_evlist__find_by_index(evlist
, event
->idx
);
1510 evsel
->name
= strdup(event
->name
);
1514 process_event_desc(struct perf_file_section
*section __maybe_unused
,
1515 struct perf_header
*header
, int fd
,
1516 void *data __maybe_unused
)
1518 struct perf_session
*session
;
1519 struct perf_evsel
*evsel
, *events
= read_event_desc(header
, fd
);
1524 session
= container_of(header
, struct perf_session
, header
);
1525 for (evsel
= events
; evsel
->attr
.size
; evsel
++)
1526 perf_evlist__set_event_name(session
->evlist
, evsel
);
1528 free_event_desc(events
);
1533 static int process_cmdline(struct perf_file_section
*section
,
1534 struct perf_header
*ph
, int fd
,
1535 void *data __maybe_unused
)
1538 char *str
, *cmdline
= NULL
, **argv
= NULL
;
1541 ret
= readn(fd
, &nr
, sizeof(nr
));
1542 if (ret
!= sizeof(nr
))
1548 ph
->env
.nr_cmdline
= nr
;
1550 cmdline
= zalloc(section
->size
+ nr
+ 1);
1554 argv
= zalloc(sizeof(char *) * (nr
+ 1));
1558 for (i
= 0; i
< nr
; i
++) {
1559 str
= do_read_string(fd
, ph
);
1563 argv
[i
] = cmdline
+ len
;
1564 memcpy(argv
[i
], str
, strlen(str
) + 1);
1565 len
+= strlen(str
) + 1;
1568 ph
->env
.cmdline
= cmdline
;
1569 ph
->env
.cmdline_argv
= (const char **) argv
;
1578 static int process_cpu_topology(struct perf_file_section
*section
,
1579 struct perf_header
*ph
, int fd
,
1580 void *data __maybe_unused
)
1586 int cpu_nr
= ph
->env
.nr_cpus_online
;
1589 ph
->env
.cpu
= calloc(cpu_nr
, sizeof(*ph
->env
.cpu
));
1593 ret
= readn(fd
, &nr
, sizeof(nr
));
1594 if (ret
!= sizeof(nr
))
1600 ph
->env
.nr_sibling_cores
= nr
;
1601 size
+= sizeof(u32
);
1602 strbuf_init(&sb
, 128);
1604 for (i
= 0; i
< nr
; i
++) {
1605 str
= do_read_string(fd
, ph
);
1609 /* include a NULL character at the end */
1610 strbuf_add(&sb
, str
, strlen(str
) + 1);
1611 size
+= string_size(str
);
1614 ph
->env
.sibling_cores
= strbuf_detach(&sb
, NULL
);
1616 ret
= readn(fd
, &nr
, sizeof(nr
));
1617 if (ret
!= sizeof(nr
))
1623 ph
->env
.nr_sibling_threads
= nr
;
1624 size
+= sizeof(u32
);
1626 for (i
= 0; i
< nr
; i
++) {
1627 str
= do_read_string(fd
, ph
);
1631 /* include a NULL character at the end */
1632 strbuf_add(&sb
, str
, strlen(str
) + 1);
1633 size
+= string_size(str
);
1636 ph
->env
.sibling_threads
= strbuf_detach(&sb
, NULL
);
1639 * The header may be from old perf,
1640 * which doesn't include core id and socket id information.
1642 if (section
->size
<= size
) {
1643 zfree(&ph
->env
.cpu
);
1647 for (i
= 0; i
< (u32
)cpu_nr
; i
++) {
1648 ret
= readn(fd
, &nr
, sizeof(nr
));
1649 if (ret
!= sizeof(nr
))
1655 if (nr
> (u32
)cpu_nr
) {
1656 pr_debug("core_id number is too big."
1657 "You may need to upgrade the perf tool.\n");
1660 ph
->env
.cpu
[i
].core_id
= nr
;
1662 ret
= readn(fd
, &nr
, sizeof(nr
));
1663 if (ret
!= sizeof(nr
))
1669 if (nr
> (u32
)cpu_nr
) {
1670 pr_debug("socket_id number is too big."
1671 "You may need to upgrade the perf tool.\n");
1675 ph
->env
.cpu
[i
].socket_id
= nr
;
1681 strbuf_release(&sb
);
1683 zfree(&ph
->env
.cpu
);
1687 static int process_numa_topology(struct perf_file_section
*section __maybe_unused
,
1688 struct perf_header
*ph
, int fd
,
1689 void *data __maybe_unused
)
1694 uint64_t mem_total
, mem_free
;
1698 ret
= readn(fd
, &nr
, sizeof(nr
));
1699 if (ret
!= sizeof(nr
))
1705 ph
->env
.nr_numa_nodes
= nr
;
1706 strbuf_init(&sb
, 256);
1708 for (i
= 0; i
< nr
; i
++) {
1710 ret
= readn(fd
, &node
, sizeof(node
));
1711 if (ret
!= sizeof(node
))
1714 ret
= readn(fd
, &mem_total
, sizeof(u64
));
1715 if (ret
!= sizeof(u64
))
1718 ret
= readn(fd
, &mem_free
, sizeof(u64
));
1719 if (ret
!= sizeof(u64
))
1722 if (ph
->needs_swap
) {
1723 node
= bswap_32(node
);
1724 mem_total
= bswap_64(mem_total
);
1725 mem_free
= bswap_64(mem_free
);
1728 strbuf_addf(&sb
, "%u:%"PRIu64
":%"PRIu64
":",
1729 node
, mem_total
, mem_free
);
1731 str
= do_read_string(fd
, ph
);
1735 /* include a NULL character at the end */
1736 strbuf_add(&sb
, str
, strlen(str
) + 1);
1739 ph
->env
.numa_nodes
= strbuf_detach(&sb
, NULL
);
1743 strbuf_release(&sb
);
1747 static int process_pmu_mappings(struct perf_file_section
*section __maybe_unused
,
1748 struct perf_header
*ph
, int fd
,
1749 void *data __maybe_unused
)
1757 ret
= readn(fd
, &pmu_num
, sizeof(pmu_num
));
1758 if (ret
!= sizeof(pmu_num
))
1762 pmu_num
= bswap_32(pmu_num
);
1765 pr_debug("pmu mappings not available\n");
1769 ph
->env
.nr_pmu_mappings
= pmu_num
;
1770 strbuf_init(&sb
, 128);
1773 if (readn(fd
, &type
, sizeof(type
)) != sizeof(type
))
1776 type
= bswap_32(type
);
1778 name
= do_read_string(fd
, ph
);
1782 strbuf_addf(&sb
, "%u:%s", type
, name
);
1783 /* include a NULL character at the end */
1784 strbuf_add(&sb
, "", 1);
1786 if (!strcmp(name
, "msr"))
1787 ph
->env
.msr_pmu_type
= type
;
1792 ph
->env
.pmu_mappings
= strbuf_detach(&sb
, NULL
);
1796 strbuf_release(&sb
);
1800 static int process_group_desc(struct perf_file_section
*section __maybe_unused
,
1801 struct perf_header
*ph
, int fd
,
1802 void *data __maybe_unused
)
1805 u32 i
, nr
, nr_groups
;
1806 struct perf_session
*session
;
1807 struct perf_evsel
*evsel
, *leader
= NULL
;
1814 if (readn(fd
, &nr_groups
, sizeof(nr_groups
)) != sizeof(nr_groups
))
1818 nr_groups
= bswap_32(nr_groups
);
1820 ph
->env
.nr_groups
= nr_groups
;
1822 pr_debug("group desc not available\n");
1826 desc
= calloc(nr_groups
, sizeof(*desc
));
1830 for (i
= 0; i
< nr_groups
; i
++) {
1831 desc
[i
].name
= do_read_string(fd
, ph
);
1835 if (readn(fd
, &desc
[i
].leader_idx
, sizeof(u32
)) != sizeof(u32
))
1838 if (readn(fd
, &desc
[i
].nr_members
, sizeof(u32
)) != sizeof(u32
))
1841 if (ph
->needs_swap
) {
1842 desc
[i
].leader_idx
= bswap_32(desc
[i
].leader_idx
);
1843 desc
[i
].nr_members
= bswap_32(desc
[i
].nr_members
);
1848 * Rebuild group relationship based on the group_desc
1850 session
= container_of(ph
, struct perf_session
, header
);
1851 session
->evlist
->nr_groups
= nr_groups
;
1854 evlist__for_each(session
->evlist
, evsel
) {
1855 if (evsel
->idx
== (int) desc
[i
].leader_idx
) {
1856 evsel
->leader
= evsel
;
1857 /* {anon_group} is a dummy name */
1858 if (strcmp(desc
[i
].name
, "{anon_group}")) {
1859 evsel
->group_name
= desc
[i
].name
;
1860 desc
[i
].name
= NULL
;
1862 evsel
->nr_members
= desc
[i
].nr_members
;
1864 if (i
>= nr_groups
|| nr
> 0) {
1865 pr_debug("invalid group desc\n");
1870 nr
= evsel
->nr_members
- 1;
1873 /* This is a group member */
1874 evsel
->leader
= leader
;
1880 if (i
!= nr_groups
|| nr
!= 0) {
1881 pr_debug("invalid group desc\n");
1887 for (i
= 0; i
< nr_groups
; i
++)
1888 zfree(&desc
[i
].name
);
1894 static int process_auxtrace(struct perf_file_section
*section
,
1895 struct perf_header
*ph
, int fd
,
1896 void *data __maybe_unused
)
1898 struct perf_session
*session
;
1901 session
= container_of(ph
, struct perf_session
, header
);
1903 err
= auxtrace_index__process(fd
, section
->size
, session
,
1906 pr_err("Failed to process auxtrace index\n");
1910 struct feature_ops
{
1911 int (*write
)(int fd
, struct perf_header
*h
, struct perf_evlist
*evlist
);
1912 void (*print
)(struct perf_header
*h
, int fd
, FILE *fp
);
1913 int (*process
)(struct perf_file_section
*section
,
1914 struct perf_header
*h
, int fd
, void *data
);
1919 #define FEAT_OPA(n, func) \
1920 [n] = { .name = #n, .write = write_##func, .print = print_##func }
1921 #define FEAT_OPP(n, func) \
1922 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1923 .process = process_##func }
1924 #define FEAT_OPF(n, func) \
1925 [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1926 .process = process_##func, .full_only = true }
1928 /* feature_ops not implemented: */
1929 #define print_tracing_data NULL
1930 #define print_build_id NULL
1932 static const struct feature_ops feat_ops
[HEADER_LAST_FEATURE
] = {
1933 FEAT_OPP(HEADER_TRACING_DATA
, tracing_data
),
1934 FEAT_OPP(HEADER_BUILD_ID
, build_id
),
1935 FEAT_OPP(HEADER_HOSTNAME
, hostname
),
1936 FEAT_OPP(HEADER_OSRELEASE
, osrelease
),
1937 FEAT_OPP(HEADER_VERSION
, version
),
1938 FEAT_OPP(HEADER_ARCH
, arch
),
1939 FEAT_OPP(HEADER_NRCPUS
, nrcpus
),
1940 FEAT_OPP(HEADER_CPUDESC
, cpudesc
),
1941 FEAT_OPP(HEADER_CPUID
, cpuid
),
1942 FEAT_OPP(HEADER_TOTAL_MEM
, total_mem
),
1943 FEAT_OPP(HEADER_EVENT_DESC
, event_desc
),
1944 FEAT_OPP(HEADER_CMDLINE
, cmdline
),
1945 FEAT_OPF(HEADER_CPU_TOPOLOGY
, cpu_topology
),
1946 FEAT_OPF(HEADER_NUMA_TOPOLOGY
, numa_topology
),
1947 FEAT_OPA(HEADER_BRANCH_STACK
, branch_stack
),
1948 FEAT_OPP(HEADER_PMU_MAPPINGS
, pmu_mappings
),
1949 FEAT_OPP(HEADER_GROUP_DESC
, group_desc
),
1950 FEAT_OPP(HEADER_AUXTRACE
, auxtrace
),
1953 struct header_print_data
{
1955 bool full
; /* extended list of headers */
1958 static int perf_file_section__fprintf_info(struct perf_file_section
*section
,
1959 struct perf_header
*ph
,
1960 int feat
, int fd
, void *data
)
1962 struct header_print_data
*hd
= data
;
1964 if (lseek(fd
, section
->offset
, SEEK_SET
) == (off_t
)-1) {
1965 pr_debug("Failed to lseek to %" PRIu64
" offset for feature "
1966 "%d, continuing...\n", section
->offset
, feat
);
1969 if (feat
>= HEADER_LAST_FEATURE
) {
1970 pr_warning("unknown feature %d\n", feat
);
1973 if (!feat_ops
[feat
].print
)
1976 if (!feat_ops
[feat
].full_only
|| hd
->full
)
1977 feat_ops
[feat
].print(ph
, fd
, hd
->fp
);
1979 fprintf(hd
->fp
, "# %s info available, use -I to display\n",
1980 feat_ops
[feat
].name
);
1985 int perf_header__fprintf_info(struct perf_session
*session
, FILE *fp
, bool full
)
1987 struct header_print_data hd
;
1988 struct perf_header
*header
= &session
->header
;
1989 int fd
= perf_data_file__fd(session
->file
);
1993 perf_header__process_sections(header
, fd
, &hd
,
1994 perf_file_section__fprintf_info
);
1998 static int do_write_feat(int fd
, struct perf_header
*h
, int type
,
1999 struct perf_file_section
**p
,
2000 struct perf_evlist
*evlist
)
2005 if (perf_header__has_feat(h
, type
)) {
2006 if (!feat_ops
[type
].write
)
2009 (*p
)->offset
= lseek(fd
, 0, SEEK_CUR
);
2011 err
= feat_ops
[type
].write(fd
, h
, evlist
);
2013 pr_debug("failed to write feature %d\n", type
);
2015 /* undo anything written */
2016 lseek(fd
, (*p
)->offset
, SEEK_SET
);
2020 (*p
)->size
= lseek(fd
, 0, SEEK_CUR
) - (*p
)->offset
;
2026 static int perf_header__adds_write(struct perf_header
*header
,
2027 struct perf_evlist
*evlist
, int fd
)
2030 struct perf_file_section
*feat_sec
, *p
;
2036 nr_sections
= bitmap_weight(header
->adds_features
, HEADER_FEAT_BITS
);
2040 feat_sec
= p
= calloc(nr_sections
, sizeof(*feat_sec
));
2041 if (feat_sec
== NULL
)
2044 sec_size
= sizeof(*feat_sec
) * nr_sections
;
2046 sec_start
= header
->feat_offset
;
2047 lseek(fd
, sec_start
+ sec_size
, SEEK_SET
);
2049 for_each_set_bit(feat
, header
->adds_features
, HEADER_FEAT_BITS
) {
2050 if (do_write_feat(fd
, header
, feat
, &p
, evlist
))
2051 perf_header__clear_feat(header
, feat
);
2054 lseek(fd
, sec_start
, SEEK_SET
);
2056 * may write more than needed due to dropped feature, but
2057 * this is okay, reader will skip the mising entries
2059 err
= do_write(fd
, feat_sec
, sec_size
);
2061 pr_debug("failed to write feature section\n");
2066 int perf_header__write_pipe(int fd
)
2068 struct perf_pipe_file_header f_header
;
2071 f_header
= (struct perf_pipe_file_header
){
2072 .magic
= PERF_MAGIC
,
2073 .size
= sizeof(f_header
),
2076 err
= do_write(fd
, &f_header
, sizeof(f_header
));
2078 pr_debug("failed to write perf pipe header\n");
2085 int perf_session__write_header(struct perf_session
*session
,
2086 struct perf_evlist
*evlist
,
2087 int fd
, bool at_exit
)
2089 struct perf_file_header f_header
;
2090 struct perf_file_attr f_attr
;
2091 struct perf_header
*header
= &session
->header
;
2092 struct perf_evsel
*evsel
;
2096 lseek(fd
, sizeof(f_header
), SEEK_SET
);
2098 evlist__for_each(session
->evlist
, evsel
) {
2099 evsel
->id_offset
= lseek(fd
, 0, SEEK_CUR
);
2100 err
= do_write(fd
, evsel
->id
, evsel
->ids
* sizeof(u64
));
2102 pr_debug("failed to write perf header\n");
2107 attr_offset
= lseek(fd
, 0, SEEK_CUR
);
2109 evlist__for_each(evlist
, evsel
) {
2110 f_attr
= (struct perf_file_attr
){
2111 .attr
= evsel
->attr
,
2113 .offset
= evsel
->id_offset
,
2114 .size
= evsel
->ids
* sizeof(u64
),
2117 err
= do_write(fd
, &f_attr
, sizeof(f_attr
));
2119 pr_debug("failed to write perf header attribute\n");
2124 if (!header
->data_offset
)
2125 header
->data_offset
= lseek(fd
, 0, SEEK_CUR
);
2126 header
->feat_offset
= header
->data_offset
+ header
->data_size
;
2129 err
= perf_header__adds_write(header
, evlist
, fd
);
2134 f_header
= (struct perf_file_header
){
2135 .magic
= PERF_MAGIC
,
2136 .size
= sizeof(f_header
),
2137 .attr_size
= sizeof(f_attr
),
2139 .offset
= attr_offset
,
2140 .size
= evlist
->nr_entries
* sizeof(f_attr
),
2143 .offset
= header
->data_offset
,
2144 .size
= header
->data_size
,
2146 /* event_types is ignored, store zeros */
2149 memcpy(&f_header
.adds_features
, &header
->adds_features
, sizeof(header
->adds_features
));
2151 lseek(fd
, 0, SEEK_SET
);
2152 err
= do_write(fd
, &f_header
, sizeof(f_header
));
2154 pr_debug("failed to write perf header\n");
2157 lseek(fd
, header
->data_offset
+ header
->data_size
, SEEK_SET
);
2162 static int perf_header__getbuffer64(struct perf_header
*header
,
2163 int fd
, void *buf
, size_t size
)
2165 if (readn(fd
, buf
, size
) <= 0)
2168 if (header
->needs_swap
)
2169 mem_bswap_64(buf
, size
);
2174 int perf_header__process_sections(struct perf_header
*header
, int fd
,
2176 int (*process
)(struct perf_file_section
*section
,
2177 struct perf_header
*ph
,
2178 int feat
, int fd
, void *data
))
2180 struct perf_file_section
*feat_sec
, *sec
;
2186 nr_sections
= bitmap_weight(header
->adds_features
, HEADER_FEAT_BITS
);
2190 feat_sec
= sec
= calloc(nr_sections
, sizeof(*feat_sec
));
2194 sec_size
= sizeof(*feat_sec
) * nr_sections
;
2196 lseek(fd
, header
->feat_offset
, SEEK_SET
);
2198 err
= perf_header__getbuffer64(header
, fd
, feat_sec
, sec_size
);
2202 for_each_set_bit(feat
, header
->adds_features
, HEADER_LAST_FEATURE
) {
2203 err
= process(sec
++, header
, feat
, fd
, data
);
2213 static const int attr_file_abi_sizes
[] = {
2214 [0] = PERF_ATTR_SIZE_VER0
,
2215 [1] = PERF_ATTR_SIZE_VER1
,
2216 [2] = PERF_ATTR_SIZE_VER2
,
2217 [3] = PERF_ATTR_SIZE_VER3
,
2218 [4] = PERF_ATTR_SIZE_VER4
,
2223 * In the legacy file format, the magic number is not used to encode endianness.
2224 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2225 * on ABI revisions, we need to try all combinations for all endianness to
2226 * detect the endianness.
2228 static int try_all_file_abis(uint64_t hdr_sz
, struct perf_header
*ph
)
2230 uint64_t ref_size
, attr_size
;
2233 for (i
= 0 ; attr_file_abi_sizes
[i
]; i
++) {
2234 ref_size
= attr_file_abi_sizes
[i
]
2235 + sizeof(struct perf_file_section
);
2236 if (hdr_sz
!= ref_size
) {
2237 attr_size
= bswap_64(hdr_sz
);
2238 if (attr_size
!= ref_size
)
2241 ph
->needs_swap
= true;
2243 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2248 /* could not determine endianness */
2252 #define PERF_PIPE_HDR_VER0 16
2254 static const size_t attr_pipe_abi_sizes
[] = {
2255 [0] = PERF_PIPE_HDR_VER0
,
2260 * In the legacy pipe format, there is an implicit assumption that endiannesss
2261 * between host recording the samples, and host parsing the samples is the
2262 * same. This is not always the case given that the pipe output may always be
2263 * redirected into a file and analyzed on a different machine with possibly a
2264 * different endianness and perf_event ABI revsions in the perf tool itself.
2266 static int try_all_pipe_abis(uint64_t hdr_sz
, struct perf_header
*ph
)
2271 for (i
= 0 ; attr_pipe_abi_sizes
[i
]; i
++) {
2272 if (hdr_sz
!= attr_pipe_abi_sizes
[i
]) {
2273 attr_size
= bswap_64(hdr_sz
);
2274 if (attr_size
!= hdr_sz
)
2277 ph
->needs_swap
= true;
2279 pr_debug("Pipe ABI%d perf.data file detected\n", i
);
2285 bool is_perf_magic(u64 magic
)
2287 if (!memcmp(&magic
, __perf_magic1
, sizeof(magic
))
2288 || magic
== __perf_magic2
2289 || magic
== __perf_magic2_sw
)
2295 static int check_magic_endian(u64 magic
, uint64_t hdr_sz
,
2296 bool is_pipe
, struct perf_header
*ph
)
2300 /* check for legacy format */
2301 ret
= memcmp(&magic
, __perf_magic1
, sizeof(magic
));
2303 ph
->version
= PERF_HEADER_VERSION_1
;
2304 pr_debug("legacy perf.data format\n");
2306 return try_all_pipe_abis(hdr_sz
, ph
);
2308 return try_all_file_abis(hdr_sz
, ph
);
2311 * the new magic number serves two purposes:
2312 * - unique number to identify actual perf.data files
2313 * - encode endianness of file
2315 ph
->version
= PERF_HEADER_VERSION_2
;
2317 /* check magic number with one endianness */
2318 if (magic
== __perf_magic2
)
2321 /* check magic number with opposite endianness */
2322 if (magic
!= __perf_magic2_sw
)
2325 ph
->needs_swap
= true;
2330 int perf_file_header__read(struct perf_file_header
*header
,
2331 struct perf_header
*ph
, int fd
)
2335 lseek(fd
, 0, SEEK_SET
);
2337 ret
= readn(fd
, header
, sizeof(*header
));
2341 if (check_magic_endian(header
->magic
,
2342 header
->attr_size
, false, ph
) < 0) {
2343 pr_debug("magic/endian check failed\n");
2347 if (ph
->needs_swap
) {
2348 mem_bswap_64(header
, offsetof(struct perf_file_header
,
2352 if (header
->size
!= sizeof(*header
)) {
2353 /* Support the previous format */
2354 if (header
->size
== offsetof(typeof(*header
), adds_features
))
2355 bitmap_zero(header
->adds_features
, HEADER_FEAT_BITS
);
2358 } else if (ph
->needs_swap
) {
2360 * feature bitmap is declared as an array of unsigned longs --
2361 * not good since its size can differ between the host that
2362 * generated the data file and the host analyzing the file.
2364 * We need to handle endianness, but we don't know the size of
2365 * the unsigned long where the file was generated. Take a best
2366 * guess at determining it: try 64-bit swap first (ie., file
2367 * created on a 64-bit host), and check if the hostname feature
2368 * bit is set (this feature bit is forced on as of fbe96f2).
2369 * If the bit is not, undo the 64-bit swap and try a 32-bit
2370 * swap. If the hostname bit is still not set (e.g., older data
2371 * file), punt and fallback to the original behavior --
2372 * clearing all feature bits and setting buildid.
2374 mem_bswap_64(&header
->adds_features
,
2375 BITS_TO_U64(HEADER_FEAT_BITS
));
2377 if (!test_bit(HEADER_HOSTNAME
, header
->adds_features
)) {
2379 mem_bswap_64(&header
->adds_features
,
2380 BITS_TO_U64(HEADER_FEAT_BITS
));
2383 mem_bswap_32(&header
->adds_features
,
2384 BITS_TO_U32(HEADER_FEAT_BITS
));
2387 if (!test_bit(HEADER_HOSTNAME
, header
->adds_features
)) {
2388 bitmap_zero(header
->adds_features
, HEADER_FEAT_BITS
);
2389 set_bit(HEADER_BUILD_ID
, header
->adds_features
);
2393 memcpy(&ph
->adds_features
, &header
->adds_features
,
2394 sizeof(ph
->adds_features
));
2396 ph
->data_offset
= header
->data
.offset
;
2397 ph
->data_size
= header
->data
.size
;
2398 ph
->feat_offset
= header
->data
.offset
+ header
->data
.size
;
2402 static int perf_file_section__process(struct perf_file_section
*section
,
2403 struct perf_header
*ph
,
2404 int feat
, int fd
, void *data
)
2406 if (lseek(fd
, section
->offset
, SEEK_SET
) == (off_t
)-1) {
2407 pr_debug("Failed to lseek to %" PRIu64
" offset for feature "
2408 "%d, continuing...\n", section
->offset
, feat
);
2412 if (feat
>= HEADER_LAST_FEATURE
) {
2413 pr_debug("unknown feature %d, continuing...\n", feat
);
2417 if (!feat_ops
[feat
].process
)
2420 return feat_ops
[feat
].process(section
, ph
, fd
, data
);
2423 static int perf_file_header__read_pipe(struct perf_pipe_file_header
*header
,
2424 struct perf_header
*ph
, int fd
,
2429 ret
= readn(fd
, header
, sizeof(*header
));
2433 if (check_magic_endian(header
->magic
, header
->size
, true, ph
) < 0) {
2434 pr_debug("endian/magic failed\n");
2439 header
->size
= bswap_64(header
->size
);
2441 if (repipe
&& do_write(STDOUT_FILENO
, header
, sizeof(*header
)) < 0)
2447 static int perf_header__read_pipe(struct perf_session
*session
)
2449 struct perf_header
*header
= &session
->header
;
2450 struct perf_pipe_file_header f_header
;
2452 if (perf_file_header__read_pipe(&f_header
, header
,
2453 perf_data_file__fd(session
->file
),
2454 session
->repipe
) < 0) {
2455 pr_debug("incompatible file format\n");
2462 static int read_attr(int fd
, struct perf_header
*ph
,
2463 struct perf_file_attr
*f_attr
)
2465 struct perf_event_attr
*attr
= &f_attr
->attr
;
2467 size_t our_sz
= sizeof(f_attr
->attr
);
2470 memset(f_attr
, 0, sizeof(*f_attr
));
2472 /* read minimal guaranteed structure */
2473 ret
= readn(fd
, attr
, PERF_ATTR_SIZE_VER0
);
2475 pr_debug("cannot read %d bytes of header attr\n",
2476 PERF_ATTR_SIZE_VER0
);
2480 /* on file perf_event_attr size */
2488 sz
= PERF_ATTR_SIZE_VER0
;
2489 } else if (sz
> our_sz
) {
2490 pr_debug("file uses a more recent and unsupported ABI"
2491 " (%zu bytes extra)\n", sz
- our_sz
);
2494 /* what we have not yet read and that we know about */
2495 left
= sz
- PERF_ATTR_SIZE_VER0
;
2498 ptr
+= PERF_ATTR_SIZE_VER0
;
2500 ret
= readn(fd
, ptr
, left
);
2502 /* read perf_file_section, ids are read in caller */
2503 ret
= readn(fd
, &f_attr
->ids
, sizeof(f_attr
->ids
));
2505 return ret
<= 0 ? -1 : 0;
2508 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel
*evsel
,
2509 struct pevent
*pevent
)
2511 struct event_format
*event
;
2514 /* already prepared */
2515 if (evsel
->tp_format
)
2518 if (pevent
== NULL
) {
2519 pr_debug("broken or missing trace data\n");
2523 event
= pevent_find_event(pevent
, evsel
->attr
.config
);
2528 snprintf(bf
, sizeof(bf
), "%s:%s", event
->system
, event
->name
);
2529 evsel
->name
= strdup(bf
);
2530 if (evsel
->name
== NULL
)
2534 evsel
->tp_format
= event
;
2538 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist
*evlist
,
2539 struct pevent
*pevent
)
2541 struct perf_evsel
*pos
;
2543 evlist__for_each(evlist
, pos
) {
2544 if (pos
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
2545 perf_evsel__prepare_tracepoint_event(pos
, pevent
))
2552 int perf_session__read_header(struct perf_session
*session
)
2554 struct perf_data_file
*file
= session
->file
;
2555 struct perf_header
*header
= &session
->header
;
2556 struct perf_file_header f_header
;
2557 struct perf_file_attr f_attr
;
2559 int nr_attrs
, nr_ids
, i
, j
;
2560 int fd
= perf_data_file__fd(file
);
2562 session
->evlist
= perf_evlist__new();
2563 if (session
->evlist
== NULL
)
2566 session
->evlist
->env
= &header
->env
;
2567 session
->machines
.host
.env
= &header
->env
;
2568 if (perf_data_file__is_pipe(file
))
2569 return perf_header__read_pipe(session
);
2571 if (perf_file_header__read(&f_header
, header
, fd
) < 0)
2575 * Sanity check that perf.data was written cleanly; data size is
2576 * initialized to 0 and updated only if the on_exit function is run.
2577 * If data size is still 0 then the file contains only partial
2578 * information. Just warn user and process it as much as it can.
2580 if (f_header
.data
.size
== 0) {
2581 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2582 "Was the 'perf record' command properly terminated?\n",
2586 nr_attrs
= f_header
.attrs
.size
/ f_header
.attr_size
;
2587 lseek(fd
, f_header
.attrs
.offset
, SEEK_SET
);
2589 for (i
= 0; i
< nr_attrs
; i
++) {
2590 struct perf_evsel
*evsel
;
2593 if (read_attr(fd
, header
, &f_attr
) < 0)
2596 if (header
->needs_swap
) {
2597 f_attr
.ids
.size
= bswap_64(f_attr
.ids
.size
);
2598 f_attr
.ids
.offset
= bswap_64(f_attr
.ids
.offset
);
2599 perf_event__attr_swap(&f_attr
.attr
);
2602 tmp
= lseek(fd
, 0, SEEK_CUR
);
2603 evsel
= perf_evsel__new(&f_attr
.attr
);
2606 goto out_delete_evlist
;
2608 evsel
->needs_swap
= header
->needs_swap
;
2610 * Do it before so that if perf_evsel__alloc_id fails, this
2611 * entry gets purged too at perf_evlist__delete().
2613 perf_evlist__add(session
->evlist
, evsel
);
2615 nr_ids
= f_attr
.ids
.size
/ sizeof(u64
);
2617 * We don't have the cpu and thread maps on the header, so
2618 * for allocating the perf_sample_id table we fake 1 cpu and
2619 * hattr->ids threads.
2621 if (perf_evsel__alloc_id(evsel
, 1, nr_ids
))
2622 goto out_delete_evlist
;
2624 lseek(fd
, f_attr
.ids
.offset
, SEEK_SET
);
2626 for (j
= 0; j
< nr_ids
; j
++) {
2627 if (perf_header__getbuffer64(header
, fd
, &f_id
, sizeof(f_id
)))
2630 perf_evlist__id_add(session
->evlist
, evsel
, 0, j
, f_id
);
2633 lseek(fd
, tmp
, SEEK_SET
);
2636 symbol_conf
.nr_events
= nr_attrs
;
2638 perf_header__process_sections(header
, fd
, &session
->tevent
,
2639 perf_file_section__process
);
2641 if (perf_evlist__prepare_tracepoint_events(session
->evlist
,
2642 session
->tevent
.pevent
))
2643 goto out_delete_evlist
;
2650 perf_evlist__delete(session
->evlist
);
2651 session
->evlist
= NULL
;
2655 int perf_event__synthesize_attr(struct perf_tool
*tool
,
2656 struct perf_event_attr
*attr
, u32 ids
, u64
*id
,
2657 perf_event__handler_t process
)
2659 union perf_event
*ev
;
2663 size
= sizeof(struct perf_event_attr
);
2664 size
= PERF_ALIGN(size
, sizeof(u64
));
2665 size
+= sizeof(struct perf_event_header
);
2666 size
+= ids
* sizeof(u64
);
2673 ev
->attr
.attr
= *attr
;
2674 memcpy(ev
->attr
.id
, id
, ids
* sizeof(u64
));
2676 ev
->attr
.header
.type
= PERF_RECORD_HEADER_ATTR
;
2677 ev
->attr
.header
.size
= (u16
)size
;
2679 if (ev
->attr
.header
.size
== size
)
2680 err
= process(tool
, ev
, NULL
, NULL
);
2689 static struct event_update_event
*
2690 event_update_event__new(size_t size
, u64 type
, u64 id
)
2692 struct event_update_event
*ev
;
2694 size
+= sizeof(*ev
);
2695 size
= PERF_ALIGN(size
, sizeof(u64
));
2699 ev
->header
.type
= PERF_RECORD_EVENT_UPDATE
;
2700 ev
->header
.size
= (u16
)size
;
2708 perf_event__synthesize_event_update_unit(struct perf_tool
*tool
,
2709 struct perf_evsel
*evsel
,
2710 perf_event__handler_t process
)
2712 struct event_update_event
*ev
;
2713 size_t size
= strlen(evsel
->unit
);
2716 ev
= event_update_event__new(size
+ 1, PERF_EVENT_UPDATE__UNIT
, evsel
->id
[0]);
2720 strncpy(ev
->data
, evsel
->unit
, size
);
2721 err
= process(tool
, (union perf_event
*)ev
, NULL
, NULL
);
2726 int perf_event__synthesize_attrs(struct perf_tool
*tool
,
2727 struct perf_session
*session
,
2728 perf_event__handler_t process
)
2730 struct perf_evsel
*evsel
;
2733 evlist__for_each(session
->evlist
, evsel
) {
2734 err
= perf_event__synthesize_attr(tool
, &evsel
->attr
, evsel
->ids
,
2735 evsel
->id
, process
);
2737 pr_debug("failed to create perf header attribute\n");
2745 int perf_event__process_attr(struct perf_tool
*tool __maybe_unused
,
2746 union perf_event
*event
,
2747 struct perf_evlist
**pevlist
)
2750 struct perf_evsel
*evsel
;
2751 struct perf_evlist
*evlist
= *pevlist
;
2753 if (evlist
== NULL
) {
2754 *pevlist
= evlist
= perf_evlist__new();
2759 evsel
= perf_evsel__new(&event
->attr
.attr
);
2763 perf_evlist__add(evlist
, evsel
);
2765 ids
= event
->header
.size
;
2766 ids
-= (void *)&event
->attr
.id
- (void *)event
;
2767 n_ids
= ids
/ sizeof(u64
);
2769 * We don't have the cpu and thread maps on the header, so
2770 * for allocating the perf_sample_id table we fake 1 cpu and
2771 * hattr->ids threads.
2773 if (perf_evsel__alloc_id(evsel
, 1, n_ids
))
2776 for (i
= 0; i
< n_ids
; i
++) {
2777 perf_evlist__id_add(evlist
, evsel
, 0, i
, event
->attr
.id
[i
]);
2780 symbol_conf
.nr_events
= evlist
->nr_entries
;
2785 int perf_event__process_event_update(struct perf_tool
*tool __maybe_unused
,
2786 union perf_event
*event
,
2787 struct perf_evlist
**pevlist
)
2789 struct event_update_event
*ev
= &event
->event_update
;
2790 struct perf_evlist
*evlist
;
2791 struct perf_evsel
*evsel
;
2793 if (!pevlist
|| *pevlist
== NULL
)
2798 evsel
= perf_evlist__id2evsel(evlist
, ev
->id
);
2803 case PERF_EVENT_UPDATE__UNIT
:
2804 evsel
->unit
= strdup(ev
->data
);
2812 int perf_event__synthesize_tracing_data(struct perf_tool
*tool
, int fd
,
2813 struct perf_evlist
*evlist
,
2814 perf_event__handler_t process
)
2816 union perf_event ev
;
2817 struct tracing_data
*tdata
;
2818 ssize_t size
= 0, aligned_size
= 0, padding
;
2819 int err __maybe_unused
= 0;
2822 * We are going to store the size of the data followed
2823 * by the data contents. Since the fd descriptor is a pipe,
2824 * we cannot seek back to store the size of the data once
2825 * we know it. Instead we:
2827 * - write the tracing data to the temp file
2828 * - get/write the data size to pipe
2829 * - write the tracing data from the temp file
2832 tdata
= tracing_data_get(&evlist
->entries
, fd
, true);
2836 memset(&ev
, 0, sizeof(ev
));
2838 ev
.tracing_data
.header
.type
= PERF_RECORD_HEADER_TRACING_DATA
;
2840 aligned_size
= PERF_ALIGN(size
, sizeof(u64
));
2841 padding
= aligned_size
- size
;
2842 ev
.tracing_data
.header
.size
= sizeof(ev
.tracing_data
);
2843 ev
.tracing_data
.size
= aligned_size
;
2845 process(tool
, &ev
, NULL
, NULL
);
2848 * The put function will copy all the tracing data
2849 * stored in temp file to the pipe.
2851 tracing_data_put(tdata
);
2853 write_padded(fd
, NULL
, 0, padding
);
2855 return aligned_size
;
2858 int perf_event__process_tracing_data(struct perf_tool
*tool __maybe_unused
,
2859 union perf_event
*event
,
2860 struct perf_session
*session
)
2862 ssize_t size_read
, padding
, size
= event
->tracing_data
.size
;
2863 int fd
= perf_data_file__fd(session
->file
);
2864 off_t offset
= lseek(fd
, 0, SEEK_CUR
);
2867 /* setup for reading amidst mmap */
2868 lseek(fd
, offset
+ sizeof(struct tracing_data_event
),
2871 size_read
= trace_report(fd
, &session
->tevent
,
2873 padding
= PERF_ALIGN(size_read
, sizeof(u64
)) - size_read
;
2875 if (readn(fd
, buf
, padding
) < 0) {
2876 pr_err("%s: reading input file", __func__
);
2879 if (session
->repipe
) {
2880 int retw
= write(STDOUT_FILENO
, buf
, padding
);
2881 if (retw
<= 0 || retw
!= padding
) {
2882 pr_err("%s: repiping tracing data padding", __func__
);
2887 if (size_read
+ padding
!= size
) {
2888 pr_err("%s: tracing data size mismatch", __func__
);
2892 perf_evlist__prepare_tracepoint_events(session
->evlist
,
2893 session
->tevent
.pevent
);
2895 return size_read
+ padding
;
2898 int perf_event__synthesize_build_id(struct perf_tool
*tool
,
2899 struct dso
*pos
, u16 misc
,
2900 perf_event__handler_t process
,
2901 struct machine
*machine
)
2903 union perf_event ev
;
2910 memset(&ev
, 0, sizeof(ev
));
2912 len
= pos
->long_name_len
+ 1;
2913 len
= PERF_ALIGN(len
, NAME_ALIGN
);
2914 memcpy(&ev
.build_id
.build_id
, pos
->build_id
, sizeof(pos
->build_id
));
2915 ev
.build_id
.header
.type
= PERF_RECORD_HEADER_BUILD_ID
;
2916 ev
.build_id
.header
.misc
= misc
;
2917 ev
.build_id
.pid
= machine
->pid
;
2918 ev
.build_id
.header
.size
= sizeof(ev
.build_id
) + len
;
2919 memcpy(&ev
.build_id
.filename
, pos
->long_name
, pos
->long_name_len
);
2921 err
= process(tool
, &ev
, NULL
, machine
);
2926 int perf_event__process_build_id(struct perf_tool
*tool __maybe_unused
,
2927 union perf_event
*event
,
2928 struct perf_session
*session
)
2930 __event_process_build_id(&event
->build_id
,
2931 event
->build_id
.filename
,