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 * send mmap only for thread group leader
546 * see thread__init_map_groups
549 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
550 process, machine, mmap_data,
557 if (machine__is_default_guest(machine))
560 snprintf(filename, sizeof(filename), "%s/proc/%d/task",
561 machine->root_dir, pid);
563 tasks = opendir(filename);
565 pr_debug("couldn't open %s\n", filename);
569 while ((dirent = readdir(tasks)) != NULL) {
573 _pid = strtol(dirent->d_name, &end, 10);
578 if (perf_event__prepare_comm(comm_event, _pid, machine,
582 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
583 ppid, process, machine) < 0)
586 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
587 tgid, process, machine) < 0)
591 * Send the prepared comm event
593 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
598 /* process the parent's maps too */
599 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
600 process, machine, mmap_data, proc_map_timeout);
610 int perf_event__synthesize_thread_map(struct perf_tool *tool,
611 struct thread_map *threads,
612 perf_event__handler_t process,
613 struct machine *machine,
615 unsigned int proc_map_timeout)
617 union perf_event *comm_event, *mmap_event, *fork_event;
618 union perf_event *namespaces_event;
619 int err = -1, thread, j;
621 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
622 if (comm_event == NULL)
625 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
626 if (mmap_event == NULL)
629 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
630 if (fork_event == NULL)
633 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
634 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
635 machine->id_hdr_size);
636 if (namespaces_event == NULL)
640 for (thread = 0; thread < threads->nr; ++thread) {
641 if (__event__synthesize_thread(comm_event, mmap_event,
642 fork_event, namespaces_event,
643 thread_map__pid(threads, thread), 0,
644 process, tool, machine,
645 mmap_data, proc_map_timeout)) {
651 * comm.pid is set to thread group id by
652 * perf_event__synthesize_comm
654 if ((int) comm_event->comm.pid != thread_map__pid(threads, thread)) {
655 bool need_leader = true;
657 /* is thread group leader in thread_map? */
658 for (j = 0; j < threads->nr; ++j) {
659 if ((int) comm_event->comm.pid == thread_map__pid(threads, j)) {
665 /* if not, generate events for it */
667 __event__synthesize_thread(comm_event, mmap_event,
668 fork_event, namespaces_event,
669 comm_event->comm.pid, 0,
670 process, tool, machine,
671 mmap_data, proc_map_timeout)) {
677 free(namespaces_event);
688 static int __perf_event__synthesize_threads(struct perf_tool *tool,
689 perf_event__handler_t process,
690 struct machine *machine,
692 unsigned int proc_map_timeout,
693 struct dirent **dirent,
697 union perf_event *comm_event, *mmap_event, *fork_event;
698 union perf_event *namespaces_event;
704 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
705 if (comm_event == NULL)
708 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
709 if (mmap_event == NULL)
712 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
713 if (fork_event == NULL)
716 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
717 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
718 machine->id_hdr_size);
719 if (namespaces_event == NULL)
722 for (i = start; i < start + num; i++) {
723 if (!isdigit(dirent[i]->d_name[0]))
726 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
727 /* only interested in proper numerical dirents */
731 * We may race with exiting thread, so don't stop just because
732 * one thread couldn't be synthesized.
734 __event__synthesize_thread(comm_event, mmap_event, fork_event,
735 namespaces_event, pid, 1, process,
736 tool, machine, mmap_data,
741 free(namespaces_event);
752 struct synthesize_threads_arg {
753 struct perf_tool *tool;
754 perf_event__handler_t process;
755 struct machine *machine;
757 unsigned int proc_map_timeout;
758 struct dirent **dirent;
763 static void *synthesize_threads_worker(void *arg)
765 struct synthesize_threads_arg *args = arg;
767 __perf_event__synthesize_threads(args->tool, args->process,
768 args->machine, args->mmap_data,
769 args->proc_map_timeout, args->dirent,
770 args->start, args->num);
774 int perf_event__synthesize_threads(struct perf_tool *tool,
775 perf_event__handler_t process,
776 struct machine *machine,
778 unsigned int proc_map_timeout,
779 unsigned int nr_threads_synthesize)
781 struct synthesize_threads_arg *args = NULL;
782 pthread_t *synthesize_threads = NULL;
783 char proc_path[PATH_MAX];
784 struct dirent **dirent;
792 if (machine__is_default_guest(machine))
795 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
796 n = scandir(proc_path, &dirent, 0, alphasort);
800 if (nr_threads_synthesize == UINT_MAX)
801 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
803 thread_nr = nr_threads_synthesize;
805 if (thread_nr <= 1) {
806 err = __perf_event__synthesize_threads(tool, process,
815 synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
816 if (synthesize_threads == NULL)
819 args = calloc(sizeof(*args), thread_nr);
823 num_per_thread = n / thread_nr;
825 for (i = 0; i < thread_nr; i++) {
827 args[i].process = process;
828 args[i].machine = machine;
829 args[i].mmap_data = mmap_data;
830 args[i].proc_map_timeout = proc_map_timeout;
831 args[i].dirent = dirent;
833 for (i = 0; i < m; i++) {
834 args[i].num = num_per_thread + 1;
835 args[i].start = i * args[i].num;
838 base = args[i-1].start + args[i-1].num;
839 for (j = i; j < thread_nr; j++) {
840 args[j].num = num_per_thread;
841 args[j].start = base + (j - i) * args[i].num;
844 for (i = 0; i < thread_nr; i++) {
845 if (pthread_create(&synthesize_threads[i], NULL,
846 synthesize_threads_worker, &args[i]))
851 for (i = 0; i < thread_nr; i++)
852 pthread_join(synthesize_threads[i], NULL);
855 free(synthesize_threads);
857 for (i = 0; i < n; i++)
864 struct process_symbol_args {
869 static int find_symbol_cb(void *arg, const char *name, char type,
872 struct process_symbol_args *args = arg;
875 * Must be a function or at least an alias, as in PARISC64, where "_text" is
876 * an 'A' to the same address as "_stext".
878 if (!(kallsyms__is_function(type) ||
879 type == 'A') || strcmp(name, args->name))
886 int kallsyms__get_function_start(const char *kallsyms_filename,
887 const char *symbol_name, u64 *addr)
889 struct process_symbol_args args = { .name = symbol_name, };
891 if (kallsyms__parse(kallsyms_filename, &args, find_symbol_cb) <= 0)
898 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
899 perf_event__handler_t process __maybe_unused,
900 struct machine *machine __maybe_unused)
905 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
906 perf_event__handler_t process,
907 struct machine *machine)
910 struct map *map = machine__kernel_map(machine);
913 union perf_event *event;
915 if (symbol_conf.kptr_restrict)
921 * We should get this from /sys/kernel/sections/.text, but till that is
922 * available use this, and after it is use this as a fallback for older
925 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size));
927 pr_debug("Not enough memory synthesizing mmap event "
928 "for kernel modules\n");
932 if (machine__is_host(machine)) {
934 * kernel uses PERF_RECORD_MISC_USER for user space maps,
935 * see kernel/perf_event.c __perf_event_mmap
937 event->header.misc = PERF_RECORD_MISC_KERNEL;
939 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
942 kmap = map__kmap(map);
943 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
944 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
945 size = PERF_ALIGN(size, sizeof(u64));
946 event->mmap.header.type = PERF_RECORD_MMAP;
947 event->mmap.header.size = (sizeof(event->mmap) -
948 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
949 event->mmap.pgoff = kmap->ref_reloc_sym->addr;
950 event->mmap.start = map->start;
951 event->mmap.len = map->end - event->mmap.start;
952 event->mmap.pid = machine->pid;
954 err = perf_tool__process_synth_event(tool, event, machine, process);
960 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
961 perf_event__handler_t process,
962 struct machine *machine)
966 err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
970 return perf_event__synthesize_extra_kmaps(tool, process, machine);
973 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
974 struct thread_map *threads,
975 perf_event__handler_t process,
976 struct machine *machine)
978 union perf_event *event;
981 size = sizeof(event->thread_map);
982 size += threads->nr * sizeof(event->thread_map.entries[0]);
984 event = zalloc(size);
988 event->header.type = PERF_RECORD_THREAD_MAP;
989 event->header.size = size;
990 event->thread_map.nr = threads->nr;
992 for (i = 0; i < threads->nr; i++) {
993 struct thread_map_event_entry *entry = &event->thread_map.entries[i];
994 char *comm = thread_map__comm(threads, i);
999 entry->pid = thread_map__pid(threads, i);
1000 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
1003 err = process(tool, event, NULL, machine);
1009 static void synthesize_cpus(struct cpu_map_entries *cpus,
1010 struct cpu_map *map)
1016 for (i = 0; i < map->nr; i++)
1017 cpus->cpu[i] = map->map[i];
1020 static void synthesize_mask(struct cpu_map_mask *mask,
1021 struct cpu_map *map, int max)
1025 mask->nr = BITS_TO_LONGS(max);
1026 mask->long_size = sizeof(long);
1028 for (i = 0; i < map->nr; i++)
1029 set_bit(map->map[i], mask->mask);
1032 static size_t cpus_size(struct cpu_map *map)
1034 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16);
1037 static size_t mask_size(struct cpu_map *map, int *max)
1043 for (i = 0; i < map->nr; i++) {
1044 /* bit possition of the cpu is + 1 */
1045 int bit = map->map[i] + 1;
1051 return sizeof(struct cpu_map_mask) + BITS_TO_LONGS(*max) * sizeof(long);
1054 void *cpu_map_data__alloc(struct cpu_map *map, size_t *size, u16 *type, int *max)
1056 size_t size_cpus, size_mask;
1057 bool is_dummy = cpu_map__empty(map);
1060 * Both array and mask data have variable size based
1061 * on the number of cpus and their actual values.
1062 * The size of the 'struct cpu_map_data' is:
1064 * array = size of 'struct cpu_map_entries' +
1065 * number of cpus * sizeof(u64)
1067 * mask = size of 'struct cpu_map_mask' +
1068 * maximum cpu bit converted to size of longs
1070 * and finaly + the size of 'struct cpu_map_data'.
1072 size_cpus = cpus_size(map);
1073 size_mask = mask_size(map, max);
1075 if (is_dummy || (size_cpus < size_mask)) {
1077 *type = PERF_CPU_MAP__CPUS;
1080 *type = PERF_CPU_MAP__MASK;
1083 *size += sizeof(struct cpu_map_data);
1084 return zalloc(*size);
1087 void cpu_map_data__synthesize(struct cpu_map_data *data, struct cpu_map *map,
1093 case PERF_CPU_MAP__CPUS:
1094 synthesize_cpus((struct cpu_map_entries *) data->data, map);
1096 case PERF_CPU_MAP__MASK:
1097 synthesize_mask((struct cpu_map_mask *) data->data, map, max);
1103 static struct cpu_map_event* cpu_map_event__new(struct cpu_map *map)
1105 size_t size = sizeof(struct cpu_map_event);
1106 struct cpu_map_event *event;
1110 event = cpu_map_data__alloc(map, &size, &type, &max);
1114 event->header.type = PERF_RECORD_CPU_MAP;
1115 event->header.size = size;
1116 event->data.type = type;
1118 cpu_map_data__synthesize(&event->data, map, type, max);
1122 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1123 struct cpu_map *map,
1124 perf_event__handler_t process,
1125 struct machine *machine)
1127 struct cpu_map_event *event;
1130 event = cpu_map_event__new(map);
1134 err = process(tool, (union perf_event *) event, NULL, machine);
1140 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1141 struct perf_stat_config *config,
1142 perf_event__handler_t process,
1143 struct machine *machine)
1145 struct stat_config_event *event;
1146 int size, i = 0, err;
1148 size = sizeof(*event);
1149 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1151 event = zalloc(size);
1155 event->header.type = PERF_RECORD_STAT_CONFIG;
1156 event->header.size = size;
1157 event->nr = PERF_STAT_CONFIG_TERM__MAX;
1159 #define ADD(__term, __val) \
1160 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \
1161 event->data[i].val = __val; \
1164 ADD(AGGR_MODE, config->aggr_mode)
1165 ADD(INTERVAL, config->interval)
1166 ADD(SCALE, config->scale)
1168 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1169 "stat config terms unbalanced\n");
1172 err = process(tool, (union perf_event *) event, NULL, machine);
1178 int perf_event__synthesize_stat(struct perf_tool *tool,
1179 u32 cpu, u32 thread, u64 id,
1180 struct perf_counts_values *count,
1181 perf_event__handler_t process,
1182 struct machine *machine)
1184 struct stat_event event;
1186 event.header.type = PERF_RECORD_STAT;
1187 event.header.size = sizeof(event);
1188 event.header.misc = 0;
1192 event.thread = thread;
1193 event.val = count->val;
1194 event.ena = count->ena;
1195 event.run = count->run;
1197 return process(tool, (union perf_event *) &event, NULL, machine);
1200 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1201 u64 evtime, u64 type,
1202 perf_event__handler_t process,
1203 struct machine *machine)
1205 struct stat_round_event event;
1207 event.header.type = PERF_RECORD_STAT_ROUND;
1208 event.header.size = sizeof(event);
1209 event.header.misc = 0;
1211 event.time = evtime;
1214 return process(tool, (union perf_event *) &event, NULL, machine);
1217 void perf_event__read_stat_config(struct perf_stat_config *config,
1218 struct stat_config_event *event)
1222 for (i = 0; i < event->nr; i++) {
1224 switch (event->data[i].tag) {
1225 #define CASE(__term, __val) \
1226 case PERF_STAT_CONFIG_TERM__##__term: \
1227 config->__val = event->data[i].val; \
1230 CASE(AGGR_MODE, aggr_mode)
1232 CASE(INTERVAL, interval)
1235 pr_warning("unknown stat config term %" PRIu64 "\n",
1236 event->data[i].tag);
1241 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp)
1245 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC)
1250 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid);
1253 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp)
1256 struct perf_ns_link_info *ns_link_info;
1257 u32 nr_namespaces, idx;
1259 ns_link_info = event->namespaces.link_info;
1260 nr_namespaces = event->namespaces.nr_namespaces;
1262 ret += fprintf(fp, " %d/%d - nr_namespaces: %u\n\t\t[",
1263 event->namespaces.pid,
1264 event->namespaces.tid,
1267 for (idx = 0; idx < nr_namespaces; idx++) {
1268 if (idx && (idx % 4 == 0))
1269 ret += fprintf(fp, "\n\t\t ");
1271 ret += fprintf(fp, "%u/%s: %" PRIu64 "/%#" PRIx64 "%s", idx,
1272 perf_ns__name(idx), (u64)ns_link_info[idx].dev,
1273 (u64)ns_link_info[idx].ino,
1274 ((idx + 1) != nr_namespaces) ? ", " : "]\n");
1280 int perf_event__process_comm(struct perf_tool *tool __maybe_unused,
1281 union perf_event *event,
1282 struct perf_sample *sample,
1283 struct machine *machine)
1285 return machine__process_comm_event(machine, event, sample);
1288 int perf_event__process_namespaces(struct perf_tool *tool __maybe_unused,
1289 union perf_event *event,
1290 struct perf_sample *sample,
1291 struct machine *machine)
1293 return machine__process_namespaces_event(machine, event, sample);
1296 int perf_event__process_lost(struct perf_tool *tool __maybe_unused,
1297 union perf_event *event,
1298 struct perf_sample *sample,
1299 struct machine *machine)
1301 return machine__process_lost_event(machine, event, sample);
1304 int perf_event__process_aux(struct perf_tool *tool __maybe_unused,
1305 union perf_event *event,
1306 struct perf_sample *sample __maybe_unused,
1307 struct machine *machine)
1309 return machine__process_aux_event(machine, event);
1312 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused,
1313 union perf_event *event,
1314 struct perf_sample *sample __maybe_unused,
1315 struct machine *machine)
1317 return machine__process_itrace_start_event(machine, event);
1320 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused,
1321 union perf_event *event,
1322 struct perf_sample *sample,
1323 struct machine *machine)
1325 return machine__process_lost_samples_event(machine, event, sample);
1328 int perf_event__process_switch(struct perf_tool *tool __maybe_unused,
1329 union perf_event *event,
1330 struct perf_sample *sample __maybe_unused,
1331 struct machine *machine)
1333 return machine__process_switch_event(machine, event);
1336 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp)
1338 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64 "]: %c %s\n",
1339 event->mmap.pid, event->mmap.tid, event->mmap.start,
1340 event->mmap.len, event->mmap.pgoff,
1341 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x',
1342 event->mmap.filename);
1345 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp)
1347 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64
1348 " %02x:%02x %"PRIu64" %"PRIu64"]: %c%c%c%c %s\n",
1349 event->mmap2.pid, event->mmap2.tid, event->mmap2.start,
1350 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj,
1351 event->mmap2.min, event->mmap2.ino,
1352 event->mmap2.ino_generation,
1353 (event->mmap2.prot & PROT_READ) ? 'r' : '-',
1354 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-',
1355 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-',
1356 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p',
1357 event->mmap2.filename);
1360 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp)
1362 struct thread_map *threads = thread_map__new_event(&event->thread_map);
1365 ret = fprintf(fp, " nr: ");
1368 ret += thread_map__fprintf(threads, fp);
1370 ret += fprintf(fp, "failed to get threads from event\n");
1372 thread_map__put(threads);
1376 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp)
1378 struct cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data);
1381 ret = fprintf(fp, ": ");
1384 ret += cpu_map__fprintf(cpus, fp);
1386 ret += fprintf(fp, "failed to get cpumap from event\n");
1392 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused,
1393 union perf_event *event,
1394 struct perf_sample *sample,
1395 struct machine *machine)
1397 return machine__process_mmap_event(machine, event, sample);
1400 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused,
1401 union perf_event *event,
1402 struct perf_sample *sample,
1403 struct machine *machine)
1405 return machine__process_mmap2_event(machine, event, sample);
1408 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp)
1410 return fprintf(fp, "(%d:%d):(%d:%d)\n",
1411 event->fork.pid, event->fork.tid,
1412 event->fork.ppid, event->fork.ptid);
1415 int perf_event__process_fork(struct perf_tool *tool __maybe_unused,
1416 union perf_event *event,
1417 struct perf_sample *sample,
1418 struct machine *machine)
1420 return machine__process_fork_event(machine, event, sample);
1423 int perf_event__process_exit(struct perf_tool *tool __maybe_unused,
1424 union perf_event *event,
1425 struct perf_sample *sample,
1426 struct machine *machine)
1428 return machine__process_exit_event(machine, event, sample);
1431 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp)
1433 return fprintf(fp, " offset: %#"PRIx64" size: %#"PRIx64" flags: %#"PRIx64" [%s%s%s]\n",
1434 event->aux.aux_offset, event->aux.aux_size,
1436 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "",
1437 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : "",
1438 event->aux.flags & PERF_AUX_FLAG_PARTIAL ? "P" : "");
1441 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp)
1443 return fprintf(fp, " pid: %u tid: %u\n",
1444 event->itrace_start.pid, event->itrace_start.tid);
1447 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp)
1449 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1450 const char *in_out = !out ? "IN " :
1451 !(event->header.misc & PERF_RECORD_MISC_SWITCH_OUT_PREEMPT) ?
1452 "OUT " : "OUT preempt";
1454 if (event->header.type == PERF_RECORD_SWITCH)
1455 return fprintf(fp, " %s\n", in_out);
1457 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n",
1458 in_out, out ? "next" : "prev",
1459 event->context_switch.next_prev_pid,
1460 event->context_switch.next_prev_tid);
1463 static size_t perf_event__fprintf_lost(union perf_event *event, FILE *fp)
1465 return fprintf(fp, " lost %" PRIu64 "\n", event->lost.lost);
1468 size_t perf_event__fprintf(union perf_event *event, FILE *fp)
1470 size_t ret = fprintf(fp, "PERF_RECORD_%s",
1471 perf_event__name(event->header.type));
1473 switch (event->header.type) {
1474 case PERF_RECORD_COMM:
1475 ret += perf_event__fprintf_comm(event, fp);
1477 case PERF_RECORD_FORK:
1478 case PERF_RECORD_EXIT:
1479 ret += perf_event__fprintf_task(event, fp);
1481 case PERF_RECORD_MMAP:
1482 ret += perf_event__fprintf_mmap(event, fp);
1484 case PERF_RECORD_NAMESPACES:
1485 ret += perf_event__fprintf_namespaces(event, fp);
1487 case PERF_RECORD_MMAP2:
1488 ret += perf_event__fprintf_mmap2(event, fp);
1490 case PERF_RECORD_AUX:
1491 ret += perf_event__fprintf_aux(event, fp);
1493 case PERF_RECORD_ITRACE_START:
1494 ret += perf_event__fprintf_itrace_start(event, fp);
1496 case PERF_RECORD_SWITCH:
1497 case PERF_RECORD_SWITCH_CPU_WIDE:
1498 ret += perf_event__fprintf_switch(event, fp);
1500 case PERF_RECORD_LOST:
1501 ret += perf_event__fprintf_lost(event, fp);
1504 ret += fprintf(fp, "\n");
1510 int perf_event__process(struct perf_tool *tool __maybe_unused,
1511 union perf_event *event,
1512 struct perf_sample *sample,
1513 struct machine *machine)
1515 return machine__process_event(machine, event, sample);
1518 struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
1519 struct addr_location *al)
1521 struct map_groups *mg = thread->mg;
1522 struct machine *machine = mg->machine;
1523 bool load_map = false;
1525 al->machine = machine;
1526 al->thread = thread;
1528 al->cpumode = cpumode;
1531 if (machine == NULL) {
1536 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
1538 mg = &machine->kmaps;
1540 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) {
1542 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
1544 mg = &machine->kmaps;
1546 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) {
1552 if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
1553 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
1555 al->filtered |= (1 << HIST_FILTER__GUEST);
1556 if ((cpumode == PERF_RECORD_MISC_USER ||
1557 cpumode == PERF_RECORD_MISC_KERNEL) &&
1559 al->filtered |= (1 << HIST_FILTER__HOST);
1564 al->map = map_groups__find(mg, al->addr);
1565 if (al->map == NULL) {
1567 * If this is outside of all known maps, and is a negative
1568 * address, try to look it up in the kernel dso, as it might be
1569 * a vsyscall or vdso (which executes in user-mode).
1571 * XXX This is nasty, we should have a symbol list in the
1572 * "[vdso]" dso, but for now lets use the old trick of looking
1573 * in the whole kernel symbol list.
1575 if (cpumode == PERF_RECORD_MISC_USER && machine &&
1576 mg != &machine->kmaps &&
1577 machine__kernel_ip(machine, al->addr)) {
1578 mg = &machine->kmaps;
1584 * Kernel maps might be changed when loading symbols so loading
1585 * must be done prior to using kernel maps.
1589 al->addr = al->map->map_ip(al->map, al->addr);
1595 struct symbol *thread__find_symbol(struct thread *thread, u8 cpumode,
1596 u64 addr, struct addr_location *al)
1599 if (thread__find_map(thread, cpumode, addr, al))
1600 al->sym = map__find_symbol(al->map, al->addr);
1605 * Callers need to drop the reference to al->thread, obtained in
1606 * machine__findnew_thread()
1608 int machine__resolve(struct machine *machine, struct addr_location *al,
1609 struct perf_sample *sample)
1611 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1617 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
1618 thread__find_map(thread, sample->cpumode, sample->ip, al);
1619 dump_printf(" ...... dso: %s\n",
1620 al->map ? al->map->dso->long_name :
1621 al->level == 'H' ? "[hypervisor]" : "<not found>");
1623 if (thread__is_filtered(thread))
1624 al->filtered |= (1 << HIST_FILTER__THREAD);
1627 al->cpu = sample->cpu;
1632 struct perf_env *env = machine->env;
1634 if (env && env->cpu)
1635 al->socket = env->cpu[al->cpu].socket_id;
1639 struct dso *dso = al->map->dso;
1641 if (symbol_conf.dso_list &&
1642 (!dso || !(strlist__has_entry(symbol_conf.dso_list,
1644 (dso->short_name != dso->long_name &&
1645 strlist__has_entry(symbol_conf.dso_list,
1646 dso->long_name))))) {
1647 al->filtered |= (1 << HIST_FILTER__DSO);
1650 al->sym = map__find_symbol(al->map, al->addr);
1653 if (symbol_conf.sym_list &&
1654 (!al->sym || !strlist__has_entry(symbol_conf.sym_list,
1656 al->filtered |= (1 << HIST_FILTER__SYMBOL);
1663 * The preprocess_sample method will return with reference counts for the
1664 * in it, when done using (and perhaps getting ref counts if needing to
1665 * keep a pointer to one of those entries) it must be paired with
1666 * addr_location__put(), so that the refcounts can be decremented.
1668 void addr_location__put(struct addr_location *al)
1670 thread__zput(al->thread);
1673 bool is_bts_event(struct perf_event_attr *attr)
1675 return attr->type == PERF_TYPE_HARDWARE &&
1676 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1677 attr->sample_period == 1;
1680 bool sample_addr_correlates_sym(struct perf_event_attr *attr)
1682 if (attr->type == PERF_TYPE_SOFTWARE &&
1683 (attr->config == PERF_COUNT_SW_PAGE_FAULTS ||
1684 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
1685 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ))
1688 if (is_bts_event(attr))
1694 void thread__resolve(struct thread *thread, struct addr_location *al,
1695 struct perf_sample *sample)
1697 thread__find_map(thread, sample->cpumode, sample->addr, al);
1699 al->cpu = sample->cpu;
1703 al->sym = map__find_symbol(al->map, al->addr);