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 <linux/ctype.h>
26 #include "symbol/kallsyms.h"
30 #include "bpf-event.h"
32 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500
34 static const char *perf_event__names[] = {
36 [PERF_RECORD_MMAP] = "MMAP",
37 [PERF_RECORD_MMAP2] = "MMAP2",
38 [PERF_RECORD_LOST] = "LOST",
39 [PERF_RECORD_COMM] = "COMM",
40 [PERF_RECORD_EXIT] = "EXIT",
41 [PERF_RECORD_THROTTLE] = "THROTTLE",
42 [PERF_RECORD_UNTHROTTLE] = "UNTHROTTLE",
43 [PERF_RECORD_FORK] = "FORK",
44 [PERF_RECORD_READ] = "READ",
45 [PERF_RECORD_SAMPLE] = "SAMPLE",
46 [PERF_RECORD_AUX] = "AUX",
47 [PERF_RECORD_ITRACE_START] = "ITRACE_START",
48 [PERF_RECORD_LOST_SAMPLES] = "LOST_SAMPLES",
49 [PERF_RECORD_SWITCH] = "SWITCH",
50 [PERF_RECORD_SWITCH_CPU_WIDE] = "SWITCH_CPU_WIDE",
51 [PERF_RECORD_NAMESPACES] = "NAMESPACES",
52 [PERF_RECORD_KSYMBOL] = "KSYMBOL",
53 [PERF_RECORD_BPF_EVENT] = "BPF_EVENT",
54 [PERF_RECORD_HEADER_ATTR] = "ATTR",
55 [PERF_RECORD_HEADER_EVENT_TYPE] = "EVENT_TYPE",
56 [PERF_RECORD_HEADER_TRACING_DATA] = "TRACING_DATA",
57 [PERF_RECORD_HEADER_BUILD_ID] = "BUILD_ID",
58 [PERF_RECORD_FINISHED_ROUND] = "FINISHED_ROUND",
59 [PERF_RECORD_ID_INDEX] = "ID_INDEX",
60 [PERF_RECORD_AUXTRACE_INFO] = "AUXTRACE_INFO",
61 [PERF_RECORD_AUXTRACE] = "AUXTRACE",
62 [PERF_RECORD_AUXTRACE_ERROR] = "AUXTRACE_ERROR",
63 [PERF_RECORD_THREAD_MAP] = "THREAD_MAP",
64 [PERF_RECORD_CPU_MAP] = "CPU_MAP",
65 [PERF_RECORD_STAT_CONFIG] = "STAT_CONFIG",
66 [PERF_RECORD_STAT] = "STAT",
67 [PERF_RECORD_STAT_ROUND] = "STAT_ROUND",
68 [PERF_RECORD_EVENT_UPDATE] = "EVENT_UPDATE",
69 [PERF_RECORD_TIME_CONV] = "TIME_CONV",
70 [PERF_RECORD_HEADER_FEATURE] = "FEATURE",
71 [PERF_RECORD_COMPRESSED] = "COMPRESSED",
74 static const char *perf_ns__names[] = {
75 [NET_NS_INDEX] = "net",
76 [UTS_NS_INDEX] = "uts",
77 [IPC_NS_INDEX] = "ipc",
78 [PID_NS_INDEX] = "pid",
79 [USER_NS_INDEX] = "user",
80 [MNT_NS_INDEX] = "mnt",
81 [CGROUP_NS_INDEX] = "cgroup",
84 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT;
86 const char *perf_event__name(unsigned int id)
88 if (id >= ARRAY_SIZE(perf_event__names))
90 if (!perf_event__names[id])
92 return perf_event__names[id];
95 static const char *perf_ns__name(unsigned int id)
97 if (id >= ARRAY_SIZE(perf_ns__names))
99 return perf_ns__names[id];
102 int perf_tool__process_synth_event(struct perf_tool *tool,
103 union perf_event *event,
104 struct machine *machine,
105 perf_event__handler_t process)
107 struct perf_sample synth_sample = {
114 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK,
117 return process(tool, event, &synth_sample, machine);
121 * Assumes that the first 4095 bytes of /proc/pid/stat contains
122 * the comm, tgid and ppid.
124 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len,
125 pid_t *tgid, pid_t *ppid)
127 char filename[PATH_MAX];
132 char *name, *tgids, *ppids;
137 snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
139 fd = open(filename, O_RDONLY);
141 pr_debug("couldn't open %s\n", filename);
145 n = read(fd, bf, sizeof(bf) - 1);
148 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
154 name = strstr(bf, "Name:");
155 tgids = strstr(bf, "Tgid:");
156 ppids = strstr(bf, "PPid:");
161 name = skip_spaces(name + 5); /* strlen("Name:") */
162 nl = strchr(name, '\n');
169 memcpy(comm, name, size);
172 pr_debug("Name: string not found for pid %d\n", pid);
176 tgids += 5; /* strlen("Tgid:") */
179 pr_debug("Tgid: string not found for pid %d\n", pid);
183 ppids += 5; /* strlen("PPid:") */
186 pr_debug("PPid: string not found for pid %d\n", pid);
192 static int perf_event__prepare_comm(union perf_event *event, pid_t pid,
193 struct machine *machine,
194 pid_t *tgid, pid_t *ppid)
200 memset(&event->comm, 0, sizeof(event->comm));
202 if (machine__is_host(machine)) {
203 if (perf_event__get_comm_ids(pid, event->comm.comm,
204 sizeof(event->comm.comm),
209 *tgid = machine->pid;
215 event->comm.pid = *tgid;
216 event->comm.header.type = PERF_RECORD_COMM;
218 size = strlen(event->comm.comm) + 1;
219 size = PERF_ALIGN(size, sizeof(u64));
220 memset(event->comm.comm + size, 0, machine->id_hdr_size);
221 event->comm.header.size = (sizeof(event->comm) -
222 (sizeof(event->comm.comm) - size) +
223 machine->id_hdr_size);
224 event->comm.tid = pid;
229 pid_t perf_event__synthesize_comm(struct perf_tool *tool,
230 union perf_event *event, pid_t pid,
231 perf_event__handler_t process,
232 struct machine *machine)
236 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0)
239 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
245 static void perf_event__get_ns_link_info(pid_t pid, const char *ns,
246 struct perf_ns_link_info *ns_link_info)
251 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns);
252 if (stat64(proc_ns, &st) == 0) {
253 ns_link_info->dev = st.st_dev;
254 ns_link_info->ino = st.st_ino;
258 int perf_event__synthesize_namespaces(struct perf_tool *tool,
259 union perf_event *event,
260 pid_t pid, pid_t tgid,
261 perf_event__handler_t process,
262 struct machine *machine)
265 struct perf_ns_link_info *ns_link_info;
267 if (!tool || !tool->namespace_events)
270 memset(&event->namespaces, 0, (sizeof(event->namespaces) +
271 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
272 machine->id_hdr_size));
274 event->namespaces.pid = tgid;
275 event->namespaces.tid = pid;
277 event->namespaces.nr_namespaces = NR_NAMESPACES;
279 ns_link_info = event->namespaces.link_info;
281 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++)
282 perf_event__get_ns_link_info(pid, perf_ns__name(idx),
285 event->namespaces.header.type = PERF_RECORD_NAMESPACES;
287 event->namespaces.header.size = (sizeof(event->namespaces) +
288 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
289 machine->id_hdr_size);
291 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
297 static int perf_event__synthesize_fork(struct perf_tool *tool,
298 union perf_event *event,
299 pid_t pid, pid_t tgid, pid_t ppid,
300 perf_event__handler_t process,
301 struct machine *machine)
303 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
306 * for main thread set parent to ppid from status file. For other
307 * threads set parent pid to main thread. ie., assume main thread
308 * spawns all threads in a process
311 event->fork.ppid = ppid;
312 event->fork.ptid = ppid;
314 event->fork.ppid = tgid;
315 event->fork.ptid = tgid;
317 event->fork.pid = tgid;
318 event->fork.tid = pid;
319 event->fork.header.type = PERF_RECORD_FORK;
320 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC;
322 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
324 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
330 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
331 union perf_event *event,
332 pid_t pid, pid_t tgid,
333 perf_event__handler_t process,
334 struct machine *machine,
337 char filename[PATH_MAX];
339 unsigned long long t;
340 bool truncation = false;
341 unsigned long long timeout = proc_map_timeout * 1000000ULL;
343 const char *hugetlbfs_mnt = hugetlbfs__mountpoint();
344 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0;
346 if (machine__is_default_guest(machine))
349 snprintf(filename, sizeof(filename), "%s/proc/%d/task/%d/maps",
350 machine->root_dir, pid, pid);
352 fp = fopen(filename, "r");
355 * We raced with a task exiting - just return:
357 pr_debug("couldn't open %s\n", filename);
361 event->header.type = PERF_RECORD_MMAP2;
367 char execname[PATH_MAX];
368 char anonstr[] = "//anon";
373 if (fgets(bf, sizeof(bf), fp) == NULL)
376 if ((rdclock() - t) > timeout) {
377 pr_warning("Reading %s time out. "
378 "You may want to increase "
379 "the time limit by --proc-map-timeout\n",
385 /* ensure null termination since stack will be reused. */
386 strcpy(execname, "");
388 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
389 n = sscanf(bf, "%"PRIx64"-%"PRIx64" %s %"PRIx64" %x:%x %u %[^\n]\n",
390 &event->mmap2.start, &event->mmap2.len, prot,
391 &event->mmap2.pgoff, &event->mmap2.maj,
396 * Anon maps don't have the execname.
401 event->mmap2.ino = (u64)ino;
404 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
406 if (machine__is_host(machine))
407 event->header.misc = PERF_RECORD_MISC_USER;
409 event->header.misc = PERF_RECORD_MISC_GUEST_USER;
411 /* map protection and flags bits */
412 event->mmap2.prot = 0;
413 event->mmap2.flags = 0;
415 event->mmap2.prot |= PROT_READ;
417 event->mmap2.prot |= PROT_WRITE;
419 event->mmap2.prot |= PROT_EXEC;
422 event->mmap2.flags |= MAP_SHARED;
424 event->mmap2.flags |= MAP_PRIVATE;
426 if (prot[2] != 'x') {
427 if (!mmap_data || prot[0] != 'r')
430 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
435 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
437 if (!strcmp(execname, ""))
438 strcpy(execname, anonstr);
440 if (hugetlbfs_mnt_len &&
441 !strncmp(execname, hugetlbfs_mnt, hugetlbfs_mnt_len)) {
442 strcpy(execname, anonstr);
443 event->mmap2.flags |= MAP_HUGETLB;
446 size = strlen(execname) + 1;
447 memcpy(event->mmap2.filename, execname, size);
448 size = PERF_ALIGN(size, sizeof(u64));
449 event->mmap2.len -= event->mmap.start;
450 event->mmap2.header.size = (sizeof(event->mmap2) -
451 (sizeof(event->mmap2.filename) - size));
452 memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
453 event->mmap2.header.size += machine->id_hdr_size;
454 event->mmap2.pid = tgid;
455 event->mmap2.tid = pid;
457 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
470 int perf_event__synthesize_modules(struct perf_tool *tool,
471 perf_event__handler_t process,
472 struct machine *machine)
476 struct maps *maps = machine__kernel_maps(machine);
477 union perf_event *event = zalloc((sizeof(event->mmap) +
478 machine->id_hdr_size));
480 pr_debug("Not enough memory synthesizing mmap event "
481 "for kernel modules\n");
485 event->header.type = PERF_RECORD_MMAP;
488 * kernel uses 0 for user space maps, see kernel/perf_event.c
491 if (machine__is_host(machine))
492 event->header.misc = PERF_RECORD_MISC_KERNEL;
494 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
496 for (pos = maps__first(maps); pos; pos = map__next(pos)) {
499 if (!__map__is_kmodule(pos))
502 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
503 event->mmap.header.type = PERF_RECORD_MMAP;
504 event->mmap.header.size = (sizeof(event->mmap) -
505 (sizeof(event->mmap.filename) - size));
506 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
507 event->mmap.header.size += machine->id_hdr_size;
508 event->mmap.start = pos->start;
509 event->mmap.len = pos->end - pos->start;
510 event->mmap.pid = machine->pid;
512 memcpy(event->mmap.filename, pos->dso->long_name,
513 pos->dso->long_name_len + 1);
514 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
524 static int __event__synthesize_thread(union perf_event *comm_event,
525 union perf_event *mmap_event,
526 union perf_event *fork_event,
527 union perf_event *namespaces_event,
529 perf_event__handler_t process,
530 struct perf_tool *tool,
531 struct machine *machine,
534 char filename[PATH_MAX];
536 struct dirent *dirent;
540 /* special case: only send one comm event using passed in pid */
542 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
548 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
549 tgid, process, machine) < 0)
553 * send mmap only for thread group leader
554 * see thread__init_map_groups
557 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
558 process, machine, mmap_data))
564 if (machine__is_default_guest(machine))
567 snprintf(filename, sizeof(filename), "%s/proc/%d/task",
568 machine->root_dir, pid);
570 tasks = opendir(filename);
572 pr_debug("couldn't open %s\n", filename);
576 while ((dirent = readdir(tasks)) != NULL) {
580 _pid = strtol(dirent->d_name, &end, 10);
585 if (perf_event__prepare_comm(comm_event, _pid, machine,
589 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
590 ppid, process, machine) < 0)
593 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
594 tgid, process, machine) < 0)
598 * Send the prepared comm event
600 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
605 /* process the parent's maps too */
606 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
607 process, machine, mmap_data);
617 int perf_event__synthesize_thread_map(struct perf_tool *tool,
618 struct thread_map *threads,
619 perf_event__handler_t process,
620 struct machine *machine,
623 union perf_event *comm_event, *mmap_event, *fork_event;
624 union perf_event *namespaces_event;
625 int err = -1, thread, j;
627 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
628 if (comm_event == NULL)
631 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
632 if (mmap_event == NULL)
635 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
636 if (fork_event == NULL)
639 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
640 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
641 machine->id_hdr_size);
642 if (namespaces_event == NULL)
646 for (thread = 0; thread < threads->nr; ++thread) {
647 if (__event__synthesize_thread(comm_event, mmap_event,
648 fork_event, namespaces_event,
649 thread_map__pid(threads, thread), 0,
650 process, tool, machine,
657 * comm.pid is set to thread group id by
658 * perf_event__synthesize_comm
660 if ((int) comm_event->comm.pid != thread_map__pid(threads, thread)) {
661 bool need_leader = true;
663 /* is thread group leader in thread_map? */
664 for (j = 0; j < threads->nr; ++j) {
665 if ((int) comm_event->comm.pid == thread_map__pid(threads, j)) {
671 /* if not, generate events for it */
673 __event__synthesize_thread(comm_event, mmap_event,
674 fork_event, namespaces_event,
675 comm_event->comm.pid, 0,
676 process, tool, machine,
683 free(namespaces_event);
694 static int __perf_event__synthesize_threads(struct perf_tool *tool,
695 perf_event__handler_t process,
696 struct machine *machine,
698 struct dirent **dirent,
702 union perf_event *comm_event, *mmap_event, *fork_event;
703 union perf_event *namespaces_event;
709 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
710 if (comm_event == NULL)
713 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
714 if (mmap_event == NULL)
717 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
718 if (fork_event == NULL)
721 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
722 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
723 machine->id_hdr_size);
724 if (namespaces_event == NULL)
727 for (i = start; i < start + num; i++) {
728 if (!isdigit(dirent[i]->d_name[0]))
731 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
732 /* only interested in proper numerical dirents */
736 * We may race with exiting thread, so don't stop just because
737 * one thread couldn't be synthesized.
739 __event__synthesize_thread(comm_event, mmap_event, fork_event,
740 namespaces_event, pid, 1, process,
741 tool, machine, mmap_data);
745 free(namespaces_event);
756 struct synthesize_threads_arg {
757 struct perf_tool *tool;
758 perf_event__handler_t process;
759 struct machine *machine;
761 struct dirent **dirent;
766 static void *synthesize_threads_worker(void *arg)
768 struct synthesize_threads_arg *args = arg;
770 __perf_event__synthesize_threads(args->tool, args->process,
771 args->machine, args->mmap_data,
773 args->start, args->num);
777 int perf_event__synthesize_threads(struct perf_tool *tool,
778 perf_event__handler_t process,
779 struct machine *machine,
781 unsigned int nr_threads_synthesize)
783 struct synthesize_threads_arg *args = NULL;
784 pthread_t *synthesize_threads = NULL;
785 char proc_path[PATH_MAX];
786 struct dirent **dirent;
794 if (machine__is_default_guest(machine))
797 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
798 n = scandir(proc_path, &dirent, 0, alphasort);
802 if (nr_threads_synthesize == UINT_MAX)
803 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
805 thread_nr = nr_threads_synthesize;
807 if (thread_nr <= 1) {
808 err = __perf_event__synthesize_threads(tool, process,
816 synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
817 if (synthesize_threads == NULL)
820 args = calloc(sizeof(*args), thread_nr);
824 num_per_thread = n / thread_nr;
826 for (i = 0; i < thread_nr; i++) {
828 args[i].process = process;
829 args[i].machine = machine;
830 args[i].mmap_data = mmap_data;
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 *size = PERF_ALIGN(*size, sizeof(u64));
1085 return zalloc(*size);
1088 void cpu_map_data__synthesize(struct cpu_map_data *data, struct cpu_map *map,
1094 case PERF_CPU_MAP__CPUS:
1095 synthesize_cpus((struct cpu_map_entries *) data->data, map);
1097 case PERF_CPU_MAP__MASK:
1098 synthesize_mask((struct cpu_map_mask *) data->data, map, max);
1104 static struct cpu_map_event* cpu_map_event__new(struct cpu_map *map)
1106 size_t size = sizeof(struct cpu_map_event);
1107 struct cpu_map_event *event;
1111 event = cpu_map_data__alloc(map, &size, &type, &max);
1115 event->header.type = PERF_RECORD_CPU_MAP;
1116 event->header.size = size;
1117 event->data.type = type;
1119 cpu_map_data__synthesize(&event->data, map, type, max);
1123 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1124 struct cpu_map *map,
1125 perf_event__handler_t process,
1126 struct machine *machine)
1128 struct cpu_map_event *event;
1131 event = cpu_map_event__new(map);
1135 err = process(tool, (union perf_event *) event, NULL, machine);
1141 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1142 struct perf_stat_config *config,
1143 perf_event__handler_t process,
1144 struct machine *machine)
1146 struct stat_config_event *event;
1147 int size, i = 0, err;
1149 size = sizeof(*event);
1150 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1152 event = zalloc(size);
1156 event->header.type = PERF_RECORD_STAT_CONFIG;
1157 event->header.size = size;
1158 event->nr = PERF_STAT_CONFIG_TERM__MAX;
1160 #define ADD(__term, __val) \
1161 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \
1162 event->data[i].val = __val; \
1165 ADD(AGGR_MODE, config->aggr_mode)
1166 ADD(INTERVAL, config->interval)
1167 ADD(SCALE, config->scale)
1169 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1170 "stat config terms unbalanced\n");
1173 err = process(tool, (union perf_event *) event, NULL, machine);
1179 int perf_event__synthesize_stat(struct perf_tool *tool,
1180 u32 cpu, u32 thread, u64 id,
1181 struct perf_counts_values *count,
1182 perf_event__handler_t process,
1183 struct machine *machine)
1185 struct stat_event event;
1187 event.header.type = PERF_RECORD_STAT;
1188 event.header.size = sizeof(event);
1189 event.header.misc = 0;
1193 event.thread = thread;
1194 event.val = count->val;
1195 event.ena = count->ena;
1196 event.run = count->run;
1198 return process(tool, (union perf_event *) &event, NULL, machine);
1201 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1202 u64 evtime, u64 type,
1203 perf_event__handler_t process,
1204 struct machine *machine)
1206 struct stat_round_event event;
1208 event.header.type = PERF_RECORD_STAT_ROUND;
1209 event.header.size = sizeof(event);
1210 event.header.misc = 0;
1212 event.time = evtime;
1215 return process(tool, (union perf_event *) &event, NULL, machine);
1218 void perf_event__read_stat_config(struct perf_stat_config *config,
1219 struct stat_config_event *event)
1223 for (i = 0; i < event->nr; i++) {
1225 switch (event->data[i].tag) {
1226 #define CASE(__term, __val) \
1227 case PERF_STAT_CONFIG_TERM__##__term: \
1228 config->__val = event->data[i].val; \
1231 CASE(AGGR_MODE, aggr_mode)
1233 CASE(INTERVAL, interval)
1236 pr_warning("unknown stat config term %" PRIu64 "\n",
1237 event->data[i].tag);
1242 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp)
1246 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC)
1251 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid);
1254 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp)
1257 struct perf_ns_link_info *ns_link_info;
1258 u32 nr_namespaces, idx;
1260 ns_link_info = event->namespaces.link_info;
1261 nr_namespaces = event->namespaces.nr_namespaces;
1263 ret += fprintf(fp, " %d/%d - nr_namespaces: %u\n\t\t[",
1264 event->namespaces.pid,
1265 event->namespaces.tid,
1268 for (idx = 0; idx < nr_namespaces; idx++) {
1269 if (idx && (idx % 4 == 0))
1270 ret += fprintf(fp, "\n\t\t ");
1272 ret += fprintf(fp, "%u/%s: %" PRIu64 "/%#" PRIx64 "%s", idx,
1273 perf_ns__name(idx), (u64)ns_link_info[idx].dev,
1274 (u64)ns_link_info[idx].ino,
1275 ((idx + 1) != nr_namespaces) ? ", " : "]\n");
1281 int perf_event__process_comm(struct perf_tool *tool __maybe_unused,
1282 union perf_event *event,
1283 struct perf_sample *sample,
1284 struct machine *machine)
1286 return machine__process_comm_event(machine, event, sample);
1289 int perf_event__process_namespaces(struct perf_tool *tool __maybe_unused,
1290 union perf_event *event,
1291 struct perf_sample *sample,
1292 struct machine *machine)
1294 return machine__process_namespaces_event(machine, event, sample);
1297 int perf_event__process_lost(struct perf_tool *tool __maybe_unused,
1298 union perf_event *event,
1299 struct perf_sample *sample,
1300 struct machine *machine)
1302 return machine__process_lost_event(machine, event, sample);
1305 int perf_event__process_aux(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_aux_event(machine, event);
1313 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused,
1314 union perf_event *event,
1315 struct perf_sample *sample __maybe_unused,
1316 struct machine *machine)
1318 return machine__process_itrace_start_event(machine, event);
1321 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused,
1322 union perf_event *event,
1323 struct perf_sample *sample,
1324 struct machine *machine)
1326 return machine__process_lost_samples_event(machine, event, sample);
1329 int perf_event__process_switch(struct perf_tool *tool __maybe_unused,
1330 union perf_event *event,
1331 struct perf_sample *sample __maybe_unused,
1332 struct machine *machine)
1334 return machine__process_switch_event(machine, event);
1337 int perf_event__process_ksymbol(struct perf_tool *tool __maybe_unused,
1338 union perf_event *event,
1339 struct perf_sample *sample __maybe_unused,
1340 struct machine *machine)
1342 return machine__process_ksymbol(machine, event, sample);
1345 int perf_event__process_bpf_event(struct perf_tool *tool __maybe_unused,
1346 union perf_event *event,
1347 struct perf_sample *sample __maybe_unused,
1348 struct machine *machine)
1350 return machine__process_bpf_event(machine, event, sample);
1353 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp)
1355 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64 "]: %c %s\n",
1356 event->mmap.pid, event->mmap.tid, event->mmap.start,
1357 event->mmap.len, event->mmap.pgoff,
1358 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x',
1359 event->mmap.filename);
1362 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp)
1364 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64
1365 " %02x:%02x %"PRIu64" %"PRIu64"]: %c%c%c%c %s\n",
1366 event->mmap2.pid, event->mmap2.tid, event->mmap2.start,
1367 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj,
1368 event->mmap2.min, event->mmap2.ino,
1369 event->mmap2.ino_generation,
1370 (event->mmap2.prot & PROT_READ) ? 'r' : '-',
1371 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-',
1372 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-',
1373 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p',
1374 event->mmap2.filename);
1377 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp)
1379 struct thread_map *threads = thread_map__new_event(&event->thread_map);
1382 ret = fprintf(fp, " nr: ");
1385 ret += thread_map__fprintf(threads, fp);
1387 ret += fprintf(fp, "failed to get threads from event\n");
1389 thread_map__put(threads);
1393 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp)
1395 struct cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data);
1398 ret = fprintf(fp, ": ");
1401 ret += cpu_map__fprintf(cpus, fp);
1403 ret += fprintf(fp, "failed to get cpumap from event\n");
1409 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused,
1410 union perf_event *event,
1411 struct perf_sample *sample,
1412 struct machine *machine)
1414 return machine__process_mmap_event(machine, event, sample);
1417 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused,
1418 union perf_event *event,
1419 struct perf_sample *sample,
1420 struct machine *machine)
1422 return machine__process_mmap2_event(machine, event, sample);
1425 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp)
1427 return fprintf(fp, "(%d:%d):(%d:%d)\n",
1428 event->fork.pid, event->fork.tid,
1429 event->fork.ppid, event->fork.ptid);
1432 int perf_event__process_fork(struct perf_tool *tool __maybe_unused,
1433 union perf_event *event,
1434 struct perf_sample *sample,
1435 struct machine *machine)
1437 return machine__process_fork_event(machine, event, sample);
1440 int perf_event__process_exit(struct perf_tool *tool __maybe_unused,
1441 union perf_event *event,
1442 struct perf_sample *sample,
1443 struct machine *machine)
1445 return machine__process_exit_event(machine, event, sample);
1448 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp)
1450 return fprintf(fp, " offset: %#"PRIx64" size: %#"PRIx64" flags: %#"PRIx64" [%s%s%s]\n",
1451 event->aux.aux_offset, event->aux.aux_size,
1453 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "",
1454 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : "",
1455 event->aux.flags & PERF_AUX_FLAG_PARTIAL ? "P" : "");
1458 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp)
1460 return fprintf(fp, " pid: %u tid: %u\n",
1461 event->itrace_start.pid, event->itrace_start.tid);
1464 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp)
1466 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1467 const char *in_out = !out ? "IN " :
1468 !(event->header.misc & PERF_RECORD_MISC_SWITCH_OUT_PREEMPT) ?
1469 "OUT " : "OUT preempt";
1471 if (event->header.type == PERF_RECORD_SWITCH)
1472 return fprintf(fp, " %s\n", in_out);
1474 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n",
1475 in_out, out ? "next" : "prev",
1476 event->context_switch.next_prev_pid,
1477 event->context_switch.next_prev_tid);
1480 static size_t perf_event__fprintf_lost(union perf_event *event, FILE *fp)
1482 return fprintf(fp, " lost %" PRIu64 "\n", event->lost.lost);
1485 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp)
1487 return fprintf(fp, " addr %" PRIx64 " len %u type %u flags 0x%x name %s\n",
1488 event->ksymbol_event.addr, event->ksymbol_event.len,
1489 event->ksymbol_event.ksym_type,
1490 event->ksymbol_event.flags, event->ksymbol_event.name);
1493 size_t perf_event__fprintf_bpf_event(union perf_event *event, FILE *fp)
1495 return fprintf(fp, " type %u, flags %u, id %u\n",
1496 event->bpf_event.type, event->bpf_event.flags,
1497 event->bpf_event.id);
1500 size_t perf_event__fprintf(union perf_event *event, FILE *fp)
1502 size_t ret = fprintf(fp, "PERF_RECORD_%s",
1503 perf_event__name(event->header.type));
1505 switch (event->header.type) {
1506 case PERF_RECORD_COMM:
1507 ret += perf_event__fprintf_comm(event, fp);
1509 case PERF_RECORD_FORK:
1510 case PERF_RECORD_EXIT:
1511 ret += perf_event__fprintf_task(event, fp);
1513 case PERF_RECORD_MMAP:
1514 ret += perf_event__fprintf_mmap(event, fp);
1516 case PERF_RECORD_NAMESPACES:
1517 ret += perf_event__fprintf_namespaces(event, fp);
1519 case PERF_RECORD_MMAP2:
1520 ret += perf_event__fprintf_mmap2(event, fp);
1522 case PERF_RECORD_AUX:
1523 ret += perf_event__fprintf_aux(event, fp);
1525 case PERF_RECORD_ITRACE_START:
1526 ret += perf_event__fprintf_itrace_start(event, fp);
1528 case PERF_RECORD_SWITCH:
1529 case PERF_RECORD_SWITCH_CPU_WIDE:
1530 ret += perf_event__fprintf_switch(event, fp);
1532 case PERF_RECORD_LOST:
1533 ret += perf_event__fprintf_lost(event, fp);
1535 case PERF_RECORD_KSYMBOL:
1536 ret += perf_event__fprintf_ksymbol(event, fp);
1538 case PERF_RECORD_BPF_EVENT:
1539 ret += perf_event__fprintf_bpf_event(event, fp);
1542 ret += fprintf(fp, "\n");
1548 int perf_event__process(struct perf_tool *tool __maybe_unused,
1549 union perf_event *event,
1550 struct perf_sample *sample,
1551 struct machine *machine)
1553 return machine__process_event(machine, event, sample);
1556 struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
1557 struct addr_location *al)
1559 struct map_groups *mg = thread->mg;
1560 struct machine *machine = mg->machine;
1561 bool load_map = false;
1563 al->machine = machine;
1564 al->thread = thread;
1566 al->cpumode = cpumode;
1569 if (machine == NULL) {
1574 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
1576 mg = &machine->kmaps;
1578 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) {
1580 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
1582 mg = &machine->kmaps;
1584 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) {
1590 if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
1591 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
1593 al->filtered |= (1 << HIST_FILTER__GUEST);
1594 if ((cpumode == PERF_RECORD_MISC_USER ||
1595 cpumode == PERF_RECORD_MISC_KERNEL) &&
1597 al->filtered |= (1 << HIST_FILTER__HOST);
1602 al->map = map_groups__find(mg, al->addr);
1603 if (al->map != NULL) {
1605 * Kernel maps might be changed when loading symbols so loading
1606 * must be done prior to using kernel maps.
1610 al->addr = al->map->map_ip(al->map, al->addr);
1617 * For branch stacks or branch samples, the sample cpumode might not be correct
1618 * because it applies only to the sample 'ip' and not necessary to 'addr' or
1619 * branch stack addresses. If possible, use a fallback to deal with those cases.
1621 struct map *thread__find_map_fb(struct thread *thread, u8 cpumode, u64 addr,
1622 struct addr_location *al)
1624 struct map *map = thread__find_map(thread, cpumode, addr, al);
1625 struct machine *machine = thread->mg->machine;
1626 u8 addr_cpumode = machine__addr_cpumode(machine, cpumode, addr);
1628 if (map || addr_cpumode == cpumode)
1631 return thread__find_map(thread, addr_cpumode, addr, al);
1634 struct symbol *thread__find_symbol(struct thread *thread, u8 cpumode,
1635 u64 addr, struct addr_location *al)
1638 if (thread__find_map(thread, cpumode, addr, al))
1639 al->sym = map__find_symbol(al->map, al->addr);
1643 struct symbol *thread__find_symbol_fb(struct thread *thread, u8 cpumode,
1644 u64 addr, struct addr_location *al)
1647 if (thread__find_map_fb(thread, cpumode, addr, al))
1648 al->sym = map__find_symbol(al->map, al->addr);
1653 * Callers need to drop the reference to al->thread, obtained in
1654 * machine__findnew_thread()
1656 int machine__resolve(struct machine *machine, struct addr_location *al,
1657 struct perf_sample *sample)
1659 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1665 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
1666 thread__find_map(thread, sample->cpumode, sample->ip, al);
1667 dump_printf(" ...... dso: %s\n",
1668 al->map ? al->map->dso->long_name :
1669 al->level == 'H' ? "[hypervisor]" : "<not found>");
1671 if (thread__is_filtered(thread))
1672 al->filtered |= (1 << HIST_FILTER__THREAD);
1675 al->cpu = sample->cpu;
1680 struct perf_env *env = machine->env;
1682 if (env && env->cpu)
1683 al->socket = env->cpu[al->cpu].socket_id;
1687 struct dso *dso = al->map->dso;
1689 if (symbol_conf.dso_list &&
1690 (!dso || !(strlist__has_entry(symbol_conf.dso_list,
1692 (dso->short_name != dso->long_name &&
1693 strlist__has_entry(symbol_conf.dso_list,
1694 dso->long_name))))) {
1695 al->filtered |= (1 << HIST_FILTER__DSO);
1698 al->sym = map__find_symbol(al->map, al->addr);
1701 if (symbol_conf.sym_list &&
1702 (!al->sym || !strlist__has_entry(symbol_conf.sym_list,
1704 al->filtered |= (1 << HIST_FILTER__SYMBOL);
1711 * The preprocess_sample method will return with reference counts for the
1712 * in it, when done using (and perhaps getting ref counts if needing to
1713 * keep a pointer to one of those entries) it must be paired with
1714 * addr_location__put(), so that the refcounts can be decremented.
1716 void addr_location__put(struct addr_location *al)
1718 thread__zput(al->thread);
1721 bool is_bts_event(struct perf_event_attr *attr)
1723 return attr->type == PERF_TYPE_HARDWARE &&
1724 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1725 attr->sample_period == 1;
1728 bool sample_addr_correlates_sym(struct perf_event_attr *attr)
1730 if (attr->type == PERF_TYPE_SOFTWARE &&
1731 (attr->config == PERF_COUNT_SW_PAGE_FAULTS ||
1732 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
1733 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ))
1736 if (is_bts_event(attr))
1742 void thread__resolve(struct thread *thread, struct addr_location *al,
1743 struct perf_sample *sample)
1745 thread__find_map_fb(thread, sample->cpumode, sample->addr, al);
1747 al->cpu = sample->cpu;
1751 al->sym = map__find_symbol(al->map, al->addr);