1 // SPDX-License-Identifier: GPL-2.0
6 #include <linux/kernel.h>
7 #include <linux/types.h>
11 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */
12 #include <api/fs/fs.h>
13 #include <linux/perf_event.h>
22 #include "thread_map.h"
23 #include "sane_ctype.h"
24 #include "symbol/kallsyms.h"
28 static const char *perf_event__names[] = {
30 [PERF_RECORD_MMAP] = "MMAP",
31 [PERF_RECORD_MMAP2] = "MMAP2",
32 [PERF_RECORD_LOST] = "LOST",
33 [PERF_RECORD_COMM] = "COMM",
34 [PERF_RECORD_EXIT] = "EXIT",
35 [PERF_RECORD_THROTTLE] = "THROTTLE",
36 [PERF_RECORD_UNTHROTTLE] = "UNTHROTTLE",
37 [PERF_RECORD_FORK] = "FORK",
38 [PERF_RECORD_READ] = "READ",
39 [PERF_RECORD_SAMPLE] = "SAMPLE",
40 [PERF_RECORD_AUX] = "AUX",
41 [PERF_RECORD_ITRACE_START] = "ITRACE_START",
42 [PERF_RECORD_LOST_SAMPLES] = "LOST_SAMPLES",
43 [PERF_RECORD_SWITCH] = "SWITCH",
44 [PERF_RECORD_SWITCH_CPU_WIDE] = "SWITCH_CPU_WIDE",
45 [PERF_RECORD_NAMESPACES] = "NAMESPACES",
46 [PERF_RECORD_HEADER_ATTR] = "ATTR",
47 [PERF_RECORD_HEADER_EVENT_TYPE] = "EVENT_TYPE",
48 [PERF_RECORD_HEADER_TRACING_DATA] = "TRACING_DATA",
49 [PERF_RECORD_HEADER_BUILD_ID] = "BUILD_ID",
50 [PERF_RECORD_FINISHED_ROUND] = "FINISHED_ROUND",
51 [PERF_RECORD_ID_INDEX] = "ID_INDEX",
52 [PERF_RECORD_AUXTRACE_INFO] = "AUXTRACE_INFO",
53 [PERF_RECORD_AUXTRACE] = "AUXTRACE",
54 [PERF_RECORD_AUXTRACE_ERROR] = "AUXTRACE_ERROR",
55 [PERF_RECORD_THREAD_MAP] = "THREAD_MAP",
56 [PERF_RECORD_CPU_MAP] = "CPU_MAP",
57 [PERF_RECORD_STAT_CONFIG] = "STAT_CONFIG",
58 [PERF_RECORD_STAT] = "STAT",
59 [PERF_RECORD_STAT_ROUND] = "STAT_ROUND",
60 [PERF_RECORD_EVENT_UPDATE] = "EVENT_UPDATE",
61 [PERF_RECORD_TIME_CONV] = "TIME_CONV",
62 [PERF_RECORD_HEADER_FEATURE] = "FEATURE",
65 static const char *perf_ns__names[] = {
66 [NET_NS_INDEX] = "net",
67 [UTS_NS_INDEX] = "uts",
68 [IPC_NS_INDEX] = "ipc",
69 [PID_NS_INDEX] = "pid",
70 [USER_NS_INDEX] = "user",
71 [MNT_NS_INDEX] = "mnt",
72 [CGROUP_NS_INDEX] = "cgroup",
75 const char *perf_event__name(unsigned int id)
77 if (id >= ARRAY_SIZE(perf_event__names))
79 if (!perf_event__names[id])
81 return perf_event__names[id];
84 static const char *perf_ns__name(unsigned int id)
86 if (id >= ARRAY_SIZE(perf_ns__names))
88 return perf_ns__names[id];
91 int perf_tool__process_synth_event(struct perf_tool *tool,
92 union perf_event *event,
93 struct machine *machine,
94 perf_event__handler_t process)
96 struct perf_sample synth_sample = {
103 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK,
106 return process(tool, event, &synth_sample, machine);
110 * Assumes that the first 4095 bytes of /proc/pid/stat contains
111 * the comm, tgid and ppid.
113 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len,
114 pid_t *tgid, pid_t *ppid)
116 char filename[PATH_MAX];
121 char *name, *tgids, *ppids;
126 snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
128 fd = open(filename, O_RDONLY);
130 pr_debug("couldn't open %s\n", filename);
134 n = read(fd, bf, sizeof(bf) - 1);
137 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
143 name = strstr(bf, "Name:");
144 tgids = strstr(bf, "Tgid:");
145 ppids = strstr(bf, "PPid:");
150 name += 5; /* strlen("Name:") */
153 nl = strchr(name, '\n');
160 memcpy(comm, name, size);
163 pr_debug("Name: string not found for pid %d\n", pid);
167 tgids += 5; /* strlen("Tgid:") */
170 pr_debug("Tgid: string not found for pid %d\n", pid);
174 ppids += 5; /* strlen("PPid:") */
177 pr_debug("PPid: string not found for pid %d\n", pid);
183 static int perf_event__prepare_comm(union perf_event *event, pid_t pid,
184 struct machine *machine,
185 pid_t *tgid, pid_t *ppid)
191 memset(&event->comm, 0, sizeof(event->comm));
193 if (machine__is_host(machine)) {
194 if (perf_event__get_comm_ids(pid, event->comm.comm,
195 sizeof(event->comm.comm),
200 *tgid = machine->pid;
206 event->comm.pid = *tgid;
207 event->comm.header.type = PERF_RECORD_COMM;
209 size = strlen(event->comm.comm) + 1;
210 size = PERF_ALIGN(size, sizeof(u64));
211 memset(event->comm.comm + size, 0, machine->id_hdr_size);
212 event->comm.header.size = (sizeof(event->comm) -
213 (sizeof(event->comm.comm) - size) +
214 machine->id_hdr_size);
215 event->comm.tid = pid;
220 pid_t perf_event__synthesize_comm(struct perf_tool *tool,
221 union perf_event *event, pid_t pid,
222 perf_event__handler_t process,
223 struct machine *machine)
227 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0)
230 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
236 static void perf_event__get_ns_link_info(pid_t pid, const char *ns,
237 struct perf_ns_link_info *ns_link_info)
242 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns);
243 if (stat64(proc_ns, &st) == 0) {
244 ns_link_info->dev = st.st_dev;
245 ns_link_info->ino = st.st_ino;
249 int perf_event__synthesize_namespaces(struct perf_tool *tool,
250 union perf_event *event,
251 pid_t pid, pid_t tgid,
252 perf_event__handler_t process,
253 struct machine *machine)
256 struct perf_ns_link_info *ns_link_info;
258 if (!tool || !tool->namespace_events)
261 memset(&event->namespaces, 0, (sizeof(event->namespaces) +
262 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
263 machine->id_hdr_size));
265 event->namespaces.pid = tgid;
266 event->namespaces.tid = pid;
268 event->namespaces.nr_namespaces = NR_NAMESPACES;
270 ns_link_info = event->namespaces.link_info;
272 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++)
273 perf_event__get_ns_link_info(pid, perf_ns__name(idx),
276 event->namespaces.header.type = PERF_RECORD_NAMESPACES;
278 event->namespaces.header.size = (sizeof(event->namespaces) +
279 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
280 machine->id_hdr_size);
282 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
288 static int perf_event__synthesize_fork(struct perf_tool *tool,
289 union perf_event *event,
290 pid_t pid, pid_t tgid, pid_t ppid,
291 perf_event__handler_t process,
292 struct machine *machine)
294 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
297 * for main thread set parent to ppid from status file. For other
298 * threads set parent pid to main thread. ie., assume main thread
299 * spawns all threads in a process
302 event->fork.ppid = ppid;
303 event->fork.ptid = ppid;
305 event->fork.ppid = tgid;
306 event->fork.ptid = tgid;
308 event->fork.pid = tgid;
309 event->fork.tid = pid;
310 event->fork.header.type = PERF_RECORD_FORK;
312 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
314 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
320 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
321 union perf_event *event,
322 pid_t pid, pid_t tgid,
323 perf_event__handler_t process,
324 struct machine *machine,
326 unsigned int proc_map_timeout)
328 char filename[PATH_MAX];
330 unsigned long long t;
331 bool truncation = false;
332 unsigned long long timeout = proc_map_timeout * 1000000ULL;
334 const char *hugetlbfs_mnt = hugetlbfs__mountpoint();
335 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0;
337 if (machine__is_default_guest(machine))
340 snprintf(filename, sizeof(filename), "%s/proc/%d/task/%d/maps",
341 machine->root_dir, pid, pid);
343 fp = fopen(filename, "r");
346 * We raced with a task exiting - just return:
348 pr_debug("couldn't open %s\n", filename);
352 event->header.type = PERF_RECORD_MMAP2;
358 char execname[PATH_MAX];
359 char anonstr[] = "//anon";
364 if (fgets(bf, sizeof(bf), fp) == NULL)
367 if ((rdclock() - t) > timeout) {
368 pr_warning("Reading %s time out. "
369 "You may want to increase "
370 "the time limit by --proc-map-timeout\n",
376 /* ensure null termination since stack will be reused. */
377 strcpy(execname, "");
379 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
380 n = sscanf(bf, "%"PRIx64"-%"PRIx64" %s %"PRIx64" %x:%x %u %[^\n]\n",
381 &event->mmap2.start, &event->mmap2.len, prot,
382 &event->mmap2.pgoff, &event->mmap2.maj,
387 * Anon maps don't have the execname.
392 event->mmap2.ino = (u64)ino;
395 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
397 if (machine__is_host(machine))
398 event->header.misc = PERF_RECORD_MISC_USER;
400 event->header.misc = PERF_RECORD_MISC_GUEST_USER;
402 /* map protection and flags bits */
403 event->mmap2.prot = 0;
404 event->mmap2.flags = 0;
406 event->mmap2.prot |= PROT_READ;
408 event->mmap2.prot |= PROT_WRITE;
410 event->mmap2.prot |= PROT_EXEC;
413 event->mmap2.flags |= MAP_SHARED;
415 event->mmap2.flags |= MAP_PRIVATE;
417 if (prot[2] != 'x') {
418 if (!mmap_data || prot[0] != 'r')
421 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
426 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
428 if (!strcmp(execname, ""))
429 strcpy(execname, anonstr);
431 if (hugetlbfs_mnt_len &&
432 !strncmp(execname, hugetlbfs_mnt, hugetlbfs_mnt_len)) {
433 strcpy(execname, anonstr);
434 event->mmap2.flags |= MAP_HUGETLB;
437 size = strlen(execname) + 1;
438 memcpy(event->mmap2.filename, execname, size);
439 size = PERF_ALIGN(size, sizeof(u64));
440 event->mmap2.len -= event->mmap.start;
441 event->mmap2.header.size = (sizeof(event->mmap2) -
442 (sizeof(event->mmap2.filename) - size));
443 memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
444 event->mmap2.header.size += machine->id_hdr_size;
445 event->mmap2.pid = tgid;
446 event->mmap2.tid = pid;
448 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
461 int perf_event__synthesize_modules(struct perf_tool *tool,
462 perf_event__handler_t process,
463 struct machine *machine)
467 struct maps *maps = machine__kernel_maps(machine);
468 union perf_event *event = zalloc((sizeof(event->mmap) +
469 machine->id_hdr_size));
471 pr_debug("Not enough memory synthesizing mmap event "
472 "for kernel modules\n");
476 event->header.type = PERF_RECORD_MMAP;
479 * kernel uses 0 for user space maps, see kernel/perf_event.c
482 if (machine__is_host(machine))
483 event->header.misc = PERF_RECORD_MISC_KERNEL;
485 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
487 for (pos = maps__first(maps); pos; pos = map__next(pos)) {
490 if (!__map__is_kmodule(pos))
493 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
494 event->mmap.header.type = PERF_RECORD_MMAP;
495 event->mmap.header.size = (sizeof(event->mmap) -
496 (sizeof(event->mmap.filename) - size));
497 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
498 event->mmap.header.size += machine->id_hdr_size;
499 event->mmap.start = pos->start;
500 event->mmap.len = pos->end - pos->start;
501 event->mmap.pid = machine->pid;
503 memcpy(event->mmap.filename, pos->dso->long_name,
504 pos->dso->long_name_len + 1);
505 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
515 static int __event__synthesize_thread(union perf_event *comm_event,
516 union perf_event *mmap_event,
517 union perf_event *fork_event,
518 union perf_event *namespaces_event,
520 perf_event__handler_t process,
521 struct perf_tool *tool,
522 struct machine *machine,
524 unsigned int proc_map_timeout)
526 char filename[PATH_MAX];
528 struct dirent *dirent;
532 /* special case: only send one comm event using passed in pid */
534 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
540 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
541 tgid, process, machine) < 0)
545 return perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
546 process, machine, mmap_data,
550 if (machine__is_default_guest(machine))
553 snprintf(filename, sizeof(filename), "%s/proc/%d/task",
554 machine->root_dir, pid);
556 tasks = opendir(filename);
558 pr_debug("couldn't open %s\n", filename);
562 while ((dirent = readdir(tasks)) != NULL) {
566 _pid = strtol(dirent->d_name, &end, 10);
571 if (perf_event__prepare_comm(comm_event, _pid, machine,
575 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
576 ppid, process, machine) < 0)
579 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
580 tgid, process, machine) < 0)
584 * Send the prepared comm event
586 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
591 /* process the parent's maps too */
592 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
593 process, machine, mmap_data, proc_map_timeout);
603 int perf_event__synthesize_thread_map(struct perf_tool *tool,
604 struct thread_map *threads,
605 perf_event__handler_t process,
606 struct machine *machine,
608 unsigned int proc_map_timeout)
610 union perf_event *comm_event, *mmap_event, *fork_event;
611 union perf_event *namespaces_event;
612 int err = -1, thread, j;
614 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
615 if (comm_event == NULL)
618 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
619 if (mmap_event == NULL)
622 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
623 if (fork_event == NULL)
626 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
627 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
628 machine->id_hdr_size);
629 if (namespaces_event == NULL)
633 for (thread = 0; thread < threads->nr; ++thread) {
634 if (__event__synthesize_thread(comm_event, mmap_event,
635 fork_event, namespaces_event,
636 thread_map__pid(threads, thread), 0,
637 process, tool, machine,
638 mmap_data, proc_map_timeout)) {
644 * comm.pid is set to thread group id by
645 * perf_event__synthesize_comm
647 if ((int) comm_event->comm.pid != thread_map__pid(threads, thread)) {
648 bool need_leader = true;
650 /* is thread group leader in thread_map? */
651 for (j = 0; j < threads->nr; ++j) {
652 if ((int) comm_event->comm.pid == thread_map__pid(threads, j)) {
658 /* if not, generate events for it */
660 __event__synthesize_thread(comm_event, mmap_event,
661 fork_event, namespaces_event,
662 comm_event->comm.pid, 0,
663 process, tool, machine,
664 mmap_data, proc_map_timeout)) {
670 free(namespaces_event);
681 static int __perf_event__synthesize_threads(struct perf_tool *tool,
682 perf_event__handler_t process,
683 struct machine *machine,
685 unsigned int proc_map_timeout,
686 struct dirent **dirent,
690 union perf_event *comm_event, *mmap_event, *fork_event;
691 union perf_event *namespaces_event;
697 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
698 if (comm_event == NULL)
701 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
702 if (mmap_event == NULL)
705 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
706 if (fork_event == NULL)
709 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
710 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
711 machine->id_hdr_size);
712 if (namespaces_event == NULL)
715 for (i = start; i < start + num; i++) {
716 if (!isdigit(dirent[i]->d_name[0]))
719 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
720 /* only interested in proper numerical dirents */
724 * We may race with exiting thread, so don't stop just because
725 * one thread couldn't be synthesized.
727 __event__synthesize_thread(comm_event, mmap_event, fork_event,
728 namespaces_event, pid, 1, process,
729 tool, machine, mmap_data,
734 free(namespaces_event);
745 struct synthesize_threads_arg {
746 struct perf_tool *tool;
747 perf_event__handler_t process;
748 struct machine *machine;
750 unsigned int proc_map_timeout;
751 struct dirent **dirent;
756 static void *synthesize_threads_worker(void *arg)
758 struct synthesize_threads_arg *args = arg;
760 __perf_event__synthesize_threads(args->tool, args->process,
761 args->machine, args->mmap_data,
762 args->proc_map_timeout, args->dirent,
763 args->start, args->num);
767 int perf_event__synthesize_threads(struct perf_tool *tool,
768 perf_event__handler_t process,
769 struct machine *machine,
771 unsigned int proc_map_timeout,
772 unsigned int nr_threads_synthesize)
774 struct synthesize_threads_arg *args = NULL;
775 pthread_t *synthesize_threads = NULL;
776 char proc_path[PATH_MAX];
777 struct dirent **dirent;
785 if (machine__is_default_guest(machine))
788 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
789 n = scandir(proc_path, &dirent, 0, alphasort);
793 if (nr_threads_synthesize == UINT_MAX)
794 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
796 thread_nr = nr_threads_synthesize;
798 if (thread_nr <= 1) {
799 err = __perf_event__synthesize_threads(tool, process,
808 synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
809 if (synthesize_threads == NULL)
812 args = calloc(sizeof(*args), thread_nr);
816 num_per_thread = n / thread_nr;
818 for (i = 0; i < thread_nr; i++) {
820 args[i].process = process;
821 args[i].machine = machine;
822 args[i].mmap_data = mmap_data;
823 args[i].proc_map_timeout = proc_map_timeout;
824 args[i].dirent = dirent;
826 for (i = 0; i < m; i++) {
827 args[i].num = num_per_thread + 1;
828 args[i].start = i * args[i].num;
831 base = args[i-1].start + args[i-1].num;
832 for (j = i; j < thread_nr; j++) {
833 args[j].num = num_per_thread;
834 args[j].start = base + (j - i) * args[i].num;
837 for (i = 0; i < thread_nr; i++) {
838 if (pthread_create(&synthesize_threads[i], NULL,
839 synthesize_threads_worker, &args[i]))
844 for (i = 0; i < thread_nr; i++)
845 pthread_join(synthesize_threads[i], NULL);
848 free(synthesize_threads);
850 for (i = 0; i < n; i++)
857 struct process_symbol_args {
862 static int find_symbol_cb(void *arg, const char *name, char type,
865 struct process_symbol_args *args = arg;
868 * Must be a function or at least an alias, as in PARISC64, where "_text" is
869 * an 'A' to the same address as "_stext".
871 if (!(kallsyms__is_function(type) ||
872 type == 'A') || strcmp(name, args->name))
879 int kallsyms__get_function_start(const char *kallsyms_filename,
880 const char *symbol_name, u64 *addr)
882 struct process_symbol_args args = { .name = symbol_name, };
884 if (kallsyms__parse(kallsyms_filename, &args, find_symbol_cb) <= 0)
891 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
892 perf_event__handler_t process __maybe_unused,
893 struct machine *machine __maybe_unused)
898 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
899 perf_event__handler_t process,
900 struct machine *machine)
903 struct map *map = machine__kernel_map(machine);
906 union perf_event *event;
908 if (symbol_conf.kptr_restrict)
914 * We should get this from /sys/kernel/sections/.text, but till that is
915 * available use this, and after it is use this as a fallback for older
918 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size));
920 pr_debug("Not enough memory synthesizing mmap event "
921 "for kernel modules\n");
925 if (machine__is_host(machine)) {
927 * kernel uses PERF_RECORD_MISC_USER for user space maps,
928 * see kernel/perf_event.c __perf_event_mmap
930 event->header.misc = PERF_RECORD_MISC_KERNEL;
932 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
935 kmap = map__kmap(map);
936 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
937 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
938 size = PERF_ALIGN(size, sizeof(u64));
939 event->mmap.header.type = PERF_RECORD_MMAP;
940 event->mmap.header.size = (sizeof(event->mmap) -
941 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
942 event->mmap.pgoff = kmap->ref_reloc_sym->addr;
943 event->mmap.start = map->start;
944 event->mmap.len = map->end - event->mmap.start;
945 event->mmap.pid = machine->pid;
947 err = perf_tool__process_synth_event(tool, event, machine, process);
953 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
954 perf_event__handler_t process,
955 struct machine *machine)
959 err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
963 return perf_event__synthesize_extra_kmaps(tool, process, machine);
966 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
967 struct thread_map *threads,
968 perf_event__handler_t process,
969 struct machine *machine)
971 union perf_event *event;
974 size = sizeof(event->thread_map);
975 size += threads->nr * sizeof(event->thread_map.entries[0]);
977 event = zalloc(size);
981 event->header.type = PERF_RECORD_THREAD_MAP;
982 event->header.size = size;
983 event->thread_map.nr = threads->nr;
985 for (i = 0; i < threads->nr; i++) {
986 struct thread_map_event_entry *entry = &event->thread_map.entries[i];
987 char *comm = thread_map__comm(threads, i);
992 entry->pid = thread_map__pid(threads, i);
993 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
996 err = process(tool, event, NULL, machine);
1002 static void synthesize_cpus(struct cpu_map_entries *cpus,
1003 struct cpu_map *map)
1009 for (i = 0; i < map->nr; i++)
1010 cpus->cpu[i] = map->map[i];
1013 static void synthesize_mask(struct cpu_map_mask *mask,
1014 struct cpu_map *map, int max)
1018 mask->nr = BITS_TO_LONGS(max);
1019 mask->long_size = sizeof(long);
1021 for (i = 0; i < map->nr; i++)
1022 set_bit(map->map[i], mask->mask);
1025 static size_t cpus_size(struct cpu_map *map)
1027 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16);
1030 static size_t mask_size(struct cpu_map *map, int *max)
1036 for (i = 0; i < map->nr; i++) {
1037 /* bit possition of the cpu is + 1 */
1038 int bit = map->map[i] + 1;
1044 return sizeof(struct cpu_map_mask) + BITS_TO_LONGS(*max) * sizeof(long);
1047 void *cpu_map_data__alloc(struct cpu_map *map, size_t *size, u16 *type, int *max)
1049 size_t size_cpus, size_mask;
1050 bool is_dummy = cpu_map__empty(map);
1053 * Both array and mask data have variable size based
1054 * on the number of cpus and their actual values.
1055 * The size of the 'struct cpu_map_data' is:
1057 * array = size of 'struct cpu_map_entries' +
1058 * number of cpus * sizeof(u64)
1060 * mask = size of 'struct cpu_map_mask' +
1061 * maximum cpu bit converted to size of longs
1063 * and finaly + the size of 'struct cpu_map_data'.
1065 size_cpus = cpus_size(map);
1066 size_mask = mask_size(map, max);
1068 if (is_dummy || (size_cpus < size_mask)) {
1070 *type = PERF_CPU_MAP__CPUS;
1073 *type = PERF_CPU_MAP__MASK;
1076 *size += sizeof(struct cpu_map_data);
1077 return zalloc(*size);
1080 void cpu_map_data__synthesize(struct cpu_map_data *data, struct cpu_map *map,
1086 case PERF_CPU_MAP__CPUS:
1087 synthesize_cpus((struct cpu_map_entries *) data->data, map);
1089 case PERF_CPU_MAP__MASK:
1090 synthesize_mask((struct cpu_map_mask *) data->data, map, max);
1096 static struct cpu_map_event* cpu_map_event__new(struct cpu_map *map)
1098 size_t size = sizeof(struct cpu_map_event);
1099 struct cpu_map_event *event;
1103 event = cpu_map_data__alloc(map, &size, &type, &max);
1107 event->header.type = PERF_RECORD_CPU_MAP;
1108 event->header.size = size;
1109 event->data.type = type;
1111 cpu_map_data__synthesize(&event->data, map, type, max);
1115 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1116 struct cpu_map *map,
1117 perf_event__handler_t process,
1118 struct machine *machine)
1120 struct cpu_map_event *event;
1123 event = cpu_map_event__new(map);
1127 err = process(tool, (union perf_event *) event, NULL, machine);
1133 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1134 struct perf_stat_config *config,
1135 perf_event__handler_t process,
1136 struct machine *machine)
1138 struct stat_config_event *event;
1139 int size, i = 0, err;
1141 size = sizeof(*event);
1142 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1144 event = zalloc(size);
1148 event->header.type = PERF_RECORD_STAT_CONFIG;
1149 event->header.size = size;
1150 event->nr = PERF_STAT_CONFIG_TERM__MAX;
1152 #define ADD(__term, __val) \
1153 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \
1154 event->data[i].val = __val; \
1157 ADD(AGGR_MODE, config->aggr_mode)
1158 ADD(INTERVAL, config->interval)
1159 ADD(SCALE, config->scale)
1161 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1162 "stat config terms unbalanced\n");
1165 err = process(tool, (union perf_event *) event, NULL, machine);
1171 int perf_event__synthesize_stat(struct perf_tool *tool,
1172 u32 cpu, u32 thread, u64 id,
1173 struct perf_counts_values *count,
1174 perf_event__handler_t process,
1175 struct machine *machine)
1177 struct stat_event event;
1179 event.header.type = PERF_RECORD_STAT;
1180 event.header.size = sizeof(event);
1181 event.header.misc = 0;
1185 event.thread = thread;
1186 event.val = count->val;
1187 event.ena = count->ena;
1188 event.run = count->run;
1190 return process(tool, (union perf_event *) &event, NULL, machine);
1193 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1194 u64 evtime, u64 type,
1195 perf_event__handler_t process,
1196 struct machine *machine)
1198 struct stat_round_event event;
1200 event.header.type = PERF_RECORD_STAT_ROUND;
1201 event.header.size = sizeof(event);
1202 event.header.misc = 0;
1204 event.time = evtime;
1207 return process(tool, (union perf_event *) &event, NULL, machine);
1210 void perf_event__read_stat_config(struct perf_stat_config *config,
1211 struct stat_config_event *event)
1215 for (i = 0; i < event->nr; i++) {
1217 switch (event->data[i].tag) {
1218 #define CASE(__term, __val) \
1219 case PERF_STAT_CONFIG_TERM__##__term: \
1220 config->__val = event->data[i].val; \
1223 CASE(AGGR_MODE, aggr_mode)
1225 CASE(INTERVAL, interval)
1228 pr_warning("unknown stat config term %" PRIu64 "\n",
1229 event->data[i].tag);
1234 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp)
1238 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC)
1243 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid);
1246 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp)
1249 struct perf_ns_link_info *ns_link_info;
1250 u32 nr_namespaces, idx;
1252 ns_link_info = event->namespaces.link_info;
1253 nr_namespaces = event->namespaces.nr_namespaces;
1255 ret += fprintf(fp, " %d/%d - nr_namespaces: %u\n\t\t[",
1256 event->namespaces.pid,
1257 event->namespaces.tid,
1260 for (idx = 0; idx < nr_namespaces; idx++) {
1261 if (idx && (idx % 4 == 0))
1262 ret += fprintf(fp, "\n\t\t ");
1264 ret += fprintf(fp, "%u/%s: %" PRIu64 "/%#" PRIx64 "%s", idx,
1265 perf_ns__name(idx), (u64)ns_link_info[idx].dev,
1266 (u64)ns_link_info[idx].ino,
1267 ((idx + 1) != nr_namespaces) ? ", " : "]\n");
1273 int perf_event__process_comm(struct perf_tool *tool __maybe_unused,
1274 union perf_event *event,
1275 struct perf_sample *sample,
1276 struct machine *machine)
1278 return machine__process_comm_event(machine, event, sample);
1281 int perf_event__process_namespaces(struct perf_tool *tool __maybe_unused,
1282 union perf_event *event,
1283 struct perf_sample *sample,
1284 struct machine *machine)
1286 return machine__process_namespaces_event(machine, event, sample);
1289 int perf_event__process_lost(struct perf_tool *tool __maybe_unused,
1290 union perf_event *event,
1291 struct perf_sample *sample,
1292 struct machine *machine)
1294 return machine__process_lost_event(machine, event, sample);
1297 int perf_event__process_aux(struct perf_tool *tool __maybe_unused,
1298 union perf_event *event,
1299 struct perf_sample *sample __maybe_unused,
1300 struct machine *machine)
1302 return machine__process_aux_event(machine, event);
1305 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused,
1306 union perf_event *event,
1307 struct perf_sample *sample __maybe_unused,
1308 struct machine *machine)
1310 return machine__process_itrace_start_event(machine, event);
1313 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused,
1314 union perf_event *event,
1315 struct perf_sample *sample,
1316 struct machine *machine)
1318 return machine__process_lost_samples_event(machine, event, sample);
1321 int perf_event__process_switch(struct perf_tool *tool __maybe_unused,
1322 union perf_event *event,
1323 struct perf_sample *sample __maybe_unused,
1324 struct machine *machine)
1326 return machine__process_switch_event(machine, event);
1329 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp)
1331 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64 "]: %c %s\n",
1332 event->mmap.pid, event->mmap.tid, event->mmap.start,
1333 event->mmap.len, event->mmap.pgoff,
1334 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x',
1335 event->mmap.filename);
1338 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp)
1340 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64
1341 " %02x:%02x %"PRIu64" %"PRIu64"]: %c%c%c%c %s\n",
1342 event->mmap2.pid, event->mmap2.tid, event->mmap2.start,
1343 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj,
1344 event->mmap2.min, event->mmap2.ino,
1345 event->mmap2.ino_generation,
1346 (event->mmap2.prot & PROT_READ) ? 'r' : '-',
1347 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-',
1348 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-',
1349 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p',
1350 event->mmap2.filename);
1353 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp)
1355 struct thread_map *threads = thread_map__new_event(&event->thread_map);
1358 ret = fprintf(fp, " nr: ");
1361 ret += thread_map__fprintf(threads, fp);
1363 ret += fprintf(fp, "failed to get threads from event\n");
1365 thread_map__put(threads);
1369 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp)
1371 struct cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data);
1374 ret = fprintf(fp, ": ");
1377 ret += cpu_map__fprintf(cpus, fp);
1379 ret += fprintf(fp, "failed to get cpumap from event\n");
1385 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused,
1386 union perf_event *event,
1387 struct perf_sample *sample,
1388 struct machine *machine)
1390 return machine__process_mmap_event(machine, event, sample);
1393 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused,
1394 union perf_event *event,
1395 struct perf_sample *sample,
1396 struct machine *machine)
1398 return machine__process_mmap2_event(machine, event, sample);
1401 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp)
1403 return fprintf(fp, "(%d:%d):(%d:%d)\n",
1404 event->fork.pid, event->fork.tid,
1405 event->fork.ppid, event->fork.ptid);
1408 int perf_event__process_fork(struct perf_tool *tool __maybe_unused,
1409 union perf_event *event,
1410 struct perf_sample *sample,
1411 struct machine *machine)
1413 return machine__process_fork_event(machine, event, sample);
1416 int perf_event__process_exit(struct perf_tool *tool __maybe_unused,
1417 union perf_event *event,
1418 struct perf_sample *sample,
1419 struct machine *machine)
1421 return machine__process_exit_event(machine, event, sample);
1424 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp)
1426 return fprintf(fp, " offset: %#"PRIx64" size: %#"PRIx64" flags: %#"PRIx64" [%s%s%s]\n",
1427 event->aux.aux_offset, event->aux.aux_size,
1429 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "",
1430 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : "",
1431 event->aux.flags & PERF_AUX_FLAG_PARTIAL ? "P" : "");
1434 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp)
1436 return fprintf(fp, " pid: %u tid: %u\n",
1437 event->itrace_start.pid, event->itrace_start.tid);
1440 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp)
1442 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1443 const char *in_out = !out ? "IN " :
1444 !(event->header.misc & PERF_RECORD_MISC_SWITCH_OUT_PREEMPT) ?
1445 "OUT " : "OUT preempt";
1447 if (event->header.type == PERF_RECORD_SWITCH)
1448 return fprintf(fp, " %s\n", in_out);
1450 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n",
1451 in_out, out ? "next" : "prev",
1452 event->context_switch.next_prev_pid,
1453 event->context_switch.next_prev_tid);
1456 static size_t perf_event__fprintf_lost(union perf_event *event, FILE *fp)
1458 return fprintf(fp, " lost %" PRIu64 "\n", event->lost.lost);
1461 size_t perf_event__fprintf(union perf_event *event, FILE *fp)
1463 size_t ret = fprintf(fp, "PERF_RECORD_%s",
1464 perf_event__name(event->header.type));
1466 switch (event->header.type) {
1467 case PERF_RECORD_COMM:
1468 ret += perf_event__fprintf_comm(event, fp);
1470 case PERF_RECORD_FORK:
1471 case PERF_RECORD_EXIT:
1472 ret += perf_event__fprintf_task(event, fp);
1474 case PERF_RECORD_MMAP:
1475 ret += perf_event__fprintf_mmap(event, fp);
1477 case PERF_RECORD_NAMESPACES:
1478 ret += perf_event__fprintf_namespaces(event, fp);
1480 case PERF_RECORD_MMAP2:
1481 ret += perf_event__fprintf_mmap2(event, fp);
1483 case PERF_RECORD_AUX:
1484 ret += perf_event__fprintf_aux(event, fp);
1486 case PERF_RECORD_ITRACE_START:
1487 ret += perf_event__fprintf_itrace_start(event, fp);
1489 case PERF_RECORD_SWITCH:
1490 case PERF_RECORD_SWITCH_CPU_WIDE:
1491 ret += perf_event__fprintf_switch(event, fp);
1493 case PERF_RECORD_LOST:
1494 ret += perf_event__fprintf_lost(event, fp);
1497 ret += fprintf(fp, "\n");
1503 int perf_event__process(struct perf_tool *tool __maybe_unused,
1504 union perf_event *event,
1505 struct perf_sample *sample,
1506 struct machine *machine)
1508 return machine__process_event(machine, event, sample);
1511 struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
1512 struct addr_location *al)
1514 struct map_groups *mg = thread->mg;
1515 struct machine *machine = mg->machine;
1516 bool load_map = false;
1518 al->machine = machine;
1519 al->thread = thread;
1521 al->cpumode = cpumode;
1524 if (machine == NULL) {
1529 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
1531 mg = &machine->kmaps;
1533 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) {
1535 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
1537 mg = &machine->kmaps;
1539 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) {
1545 if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
1546 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
1548 al->filtered |= (1 << HIST_FILTER__GUEST);
1549 if ((cpumode == PERF_RECORD_MISC_USER ||
1550 cpumode == PERF_RECORD_MISC_KERNEL) &&
1552 al->filtered |= (1 << HIST_FILTER__HOST);
1557 al->map = map_groups__find(mg, al->addr);
1558 if (al->map == NULL) {
1560 * If this is outside of all known maps, and is a negative
1561 * address, try to look it up in the kernel dso, as it might be
1562 * a vsyscall or vdso (which executes in user-mode).
1564 * XXX This is nasty, we should have a symbol list in the
1565 * "[vdso]" dso, but for now lets use the old trick of looking
1566 * in the whole kernel symbol list.
1568 if (cpumode == PERF_RECORD_MISC_USER && machine &&
1569 mg != &machine->kmaps &&
1570 machine__kernel_ip(machine, al->addr)) {
1571 mg = &machine->kmaps;
1577 * Kernel maps might be changed when loading symbols so loading
1578 * must be done prior to using kernel maps.
1582 al->addr = al->map->map_ip(al->map, al->addr);
1588 struct symbol *thread__find_symbol(struct thread *thread, u8 cpumode,
1589 u64 addr, struct addr_location *al)
1592 if (thread__find_map(thread, cpumode, addr, al))
1593 al->sym = map__find_symbol(al->map, al->addr);
1598 * Callers need to drop the reference to al->thread, obtained in
1599 * machine__findnew_thread()
1601 int machine__resolve(struct machine *machine, struct addr_location *al,
1602 struct perf_sample *sample)
1604 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1610 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
1611 thread__find_map(thread, sample->cpumode, sample->ip, al);
1612 dump_printf(" ...... dso: %s\n",
1613 al->map ? al->map->dso->long_name :
1614 al->level == 'H' ? "[hypervisor]" : "<not found>");
1616 if (thread__is_filtered(thread))
1617 al->filtered |= (1 << HIST_FILTER__THREAD);
1620 al->cpu = sample->cpu;
1625 struct perf_env *env = machine->env;
1627 if (env && env->cpu)
1628 al->socket = env->cpu[al->cpu].socket_id;
1632 struct dso *dso = al->map->dso;
1634 if (symbol_conf.dso_list &&
1635 (!dso || !(strlist__has_entry(symbol_conf.dso_list,
1637 (dso->short_name != dso->long_name &&
1638 strlist__has_entry(symbol_conf.dso_list,
1639 dso->long_name))))) {
1640 al->filtered |= (1 << HIST_FILTER__DSO);
1643 al->sym = map__find_symbol(al->map, al->addr);
1646 if (symbol_conf.sym_list &&
1647 (!al->sym || !strlist__has_entry(symbol_conf.sym_list,
1649 al->filtered |= (1 << HIST_FILTER__SYMBOL);
1656 * The preprocess_sample method will return with reference counts for the
1657 * in it, when done using (and perhaps getting ref counts if needing to
1658 * keep a pointer to one of those entries) it must be paired with
1659 * addr_location__put(), so that the refcounts can be decremented.
1661 void addr_location__put(struct addr_location *al)
1663 thread__zput(al->thread);
1666 bool is_bts_event(struct perf_event_attr *attr)
1668 return attr->type == PERF_TYPE_HARDWARE &&
1669 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1670 attr->sample_period == 1;
1673 bool sample_addr_correlates_sym(struct perf_event_attr *attr)
1675 if (attr->type == PERF_TYPE_SOFTWARE &&
1676 (attr->config == PERF_COUNT_SW_PAGE_FAULTS ||
1677 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
1678 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ))
1681 if (is_bts_event(attr))
1687 void thread__resolve(struct thread *thread, struct addr_location *al,
1688 struct perf_sample *sample)
1690 thread__find_map(thread, sample->cpumode, sample->addr, al);
1692 al->cpu = sample->cpu;
1696 al->sym = map__find_symbol(al->map, al->addr);