5 #include <linux/kernel.h>
6 #include <linux/types.h>
10 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */
11 #include <api/fs/fs.h>
12 #include <linux/perf_event.h>
13 #include <linux/zalloc.h>
23 #include "thread_map.h"
24 #include "time-utils.h"
25 #include <linux/ctype.h>
28 #include "symbol/kallsyms.h"
32 #include "bpf-event.h"
36 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500
38 static const char *perf_event__names[] = {
40 [PERF_RECORD_MMAP] = "MMAP",
41 [PERF_RECORD_MMAP2] = "MMAP2",
42 [PERF_RECORD_LOST] = "LOST",
43 [PERF_RECORD_COMM] = "COMM",
44 [PERF_RECORD_EXIT] = "EXIT",
45 [PERF_RECORD_THROTTLE] = "THROTTLE",
46 [PERF_RECORD_UNTHROTTLE] = "UNTHROTTLE",
47 [PERF_RECORD_FORK] = "FORK",
48 [PERF_RECORD_READ] = "READ",
49 [PERF_RECORD_SAMPLE] = "SAMPLE",
50 [PERF_RECORD_AUX] = "AUX",
51 [PERF_RECORD_ITRACE_START] = "ITRACE_START",
52 [PERF_RECORD_LOST_SAMPLES] = "LOST_SAMPLES",
53 [PERF_RECORD_SWITCH] = "SWITCH",
54 [PERF_RECORD_SWITCH_CPU_WIDE] = "SWITCH_CPU_WIDE",
55 [PERF_RECORD_NAMESPACES] = "NAMESPACES",
56 [PERF_RECORD_KSYMBOL] = "KSYMBOL",
57 [PERF_RECORD_BPF_EVENT] = "BPF_EVENT",
58 [PERF_RECORD_HEADER_ATTR] = "ATTR",
59 [PERF_RECORD_HEADER_EVENT_TYPE] = "EVENT_TYPE",
60 [PERF_RECORD_HEADER_TRACING_DATA] = "TRACING_DATA",
61 [PERF_RECORD_HEADER_BUILD_ID] = "BUILD_ID",
62 [PERF_RECORD_FINISHED_ROUND] = "FINISHED_ROUND",
63 [PERF_RECORD_ID_INDEX] = "ID_INDEX",
64 [PERF_RECORD_AUXTRACE_INFO] = "AUXTRACE_INFO",
65 [PERF_RECORD_AUXTRACE] = "AUXTRACE",
66 [PERF_RECORD_AUXTRACE_ERROR] = "AUXTRACE_ERROR",
67 [PERF_RECORD_THREAD_MAP] = "THREAD_MAP",
68 [PERF_RECORD_CPU_MAP] = "CPU_MAP",
69 [PERF_RECORD_STAT_CONFIG] = "STAT_CONFIG",
70 [PERF_RECORD_STAT] = "STAT",
71 [PERF_RECORD_STAT_ROUND] = "STAT_ROUND",
72 [PERF_RECORD_EVENT_UPDATE] = "EVENT_UPDATE",
73 [PERF_RECORD_TIME_CONV] = "TIME_CONV",
74 [PERF_RECORD_HEADER_FEATURE] = "FEATURE",
75 [PERF_RECORD_COMPRESSED] = "COMPRESSED",
78 static const char *perf_ns__names[] = {
79 [NET_NS_INDEX] = "net",
80 [UTS_NS_INDEX] = "uts",
81 [IPC_NS_INDEX] = "ipc",
82 [PID_NS_INDEX] = "pid",
83 [USER_NS_INDEX] = "user",
84 [MNT_NS_INDEX] = "mnt",
85 [CGROUP_NS_INDEX] = "cgroup",
88 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT;
90 const char *perf_event__name(unsigned int id)
92 if (id >= ARRAY_SIZE(perf_event__names))
94 if (!perf_event__names[id])
96 return perf_event__names[id];
99 static const char *perf_ns__name(unsigned int id)
101 if (id >= ARRAY_SIZE(perf_ns__names))
103 return perf_ns__names[id];
106 int perf_tool__process_synth_event(struct perf_tool *tool,
107 union perf_event *event,
108 struct machine *machine,
109 perf_event__handler_t process)
111 struct perf_sample synth_sample = {
118 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK,
121 return process(tool, event, &synth_sample, machine);
125 * Assumes that the first 4095 bytes of /proc/pid/stat contains
126 * the comm, tgid and ppid.
128 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len,
129 pid_t *tgid, pid_t *ppid)
131 char filename[PATH_MAX];
136 char *name, *tgids, *ppids;
141 snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
143 fd = open(filename, O_RDONLY);
145 pr_debug("couldn't open %s\n", filename);
149 n = read(fd, bf, sizeof(bf) - 1);
152 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
158 name = strstr(bf, "Name:");
159 tgids = strstr(bf, "Tgid:");
160 ppids = strstr(bf, "PPid:");
165 name = skip_spaces(name + 5); /* strlen("Name:") */
166 nl = strchr(name, '\n');
173 memcpy(comm, name, size);
176 pr_debug("Name: string not found for pid %d\n", pid);
180 tgids += 5; /* strlen("Tgid:") */
183 pr_debug("Tgid: string not found for pid %d\n", pid);
187 ppids += 5; /* strlen("PPid:") */
190 pr_debug("PPid: string not found for pid %d\n", pid);
196 static int perf_event__prepare_comm(union perf_event *event, pid_t pid,
197 struct machine *machine,
198 pid_t *tgid, pid_t *ppid)
204 memset(&event->comm, 0, sizeof(event->comm));
206 if (machine__is_host(machine)) {
207 if (perf_event__get_comm_ids(pid, event->comm.comm,
208 sizeof(event->comm.comm),
213 *tgid = machine->pid;
219 event->comm.pid = *tgid;
220 event->comm.header.type = PERF_RECORD_COMM;
222 size = strlen(event->comm.comm) + 1;
223 size = PERF_ALIGN(size, sizeof(u64));
224 memset(event->comm.comm + size, 0, machine->id_hdr_size);
225 event->comm.header.size = (sizeof(event->comm) -
226 (sizeof(event->comm.comm) - size) +
227 machine->id_hdr_size);
228 event->comm.tid = pid;
233 pid_t perf_event__synthesize_comm(struct perf_tool *tool,
234 union perf_event *event, pid_t pid,
235 perf_event__handler_t process,
236 struct machine *machine)
240 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0)
243 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
249 static void perf_event__get_ns_link_info(pid_t pid, const char *ns,
250 struct perf_ns_link_info *ns_link_info)
255 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns);
256 if (stat64(proc_ns, &st) == 0) {
257 ns_link_info->dev = st.st_dev;
258 ns_link_info->ino = st.st_ino;
262 int perf_event__synthesize_namespaces(struct perf_tool *tool,
263 union perf_event *event,
264 pid_t pid, pid_t tgid,
265 perf_event__handler_t process,
266 struct machine *machine)
269 struct perf_ns_link_info *ns_link_info;
271 if (!tool || !tool->namespace_events)
274 memset(&event->namespaces, 0, (sizeof(event->namespaces) +
275 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
276 machine->id_hdr_size));
278 event->namespaces.pid = tgid;
279 event->namespaces.tid = pid;
281 event->namespaces.nr_namespaces = NR_NAMESPACES;
283 ns_link_info = event->namespaces.link_info;
285 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++)
286 perf_event__get_ns_link_info(pid, perf_ns__name(idx),
289 event->namespaces.header.type = PERF_RECORD_NAMESPACES;
291 event->namespaces.header.size = (sizeof(event->namespaces) +
292 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
293 machine->id_hdr_size);
295 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
301 static int perf_event__synthesize_fork(struct perf_tool *tool,
302 union perf_event *event,
303 pid_t pid, pid_t tgid, pid_t ppid,
304 perf_event__handler_t process,
305 struct machine *machine)
307 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
310 * for main thread set parent to ppid from status file. For other
311 * threads set parent pid to main thread. ie., assume main thread
312 * spawns all threads in a process
315 event->fork.ppid = ppid;
316 event->fork.ptid = ppid;
318 event->fork.ppid = tgid;
319 event->fork.ptid = tgid;
321 event->fork.pid = tgid;
322 event->fork.tid = pid;
323 event->fork.header.type = PERF_RECORD_FORK;
324 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC;
326 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
328 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
334 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
335 union perf_event *event,
336 pid_t pid, pid_t tgid,
337 perf_event__handler_t process,
338 struct machine *machine,
341 char filename[PATH_MAX];
343 unsigned long long t;
344 bool truncation = false;
345 unsigned long long timeout = proc_map_timeout * 1000000ULL;
347 const char *hugetlbfs_mnt = hugetlbfs__mountpoint();
348 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0;
350 if (machine__is_default_guest(machine))
353 snprintf(filename, sizeof(filename), "%s/proc/%d/task/%d/maps",
354 machine->root_dir, pid, pid);
356 fp = fopen(filename, "r");
359 * We raced with a task exiting - just return:
361 pr_debug("couldn't open %s\n", filename);
365 event->header.type = PERF_RECORD_MMAP2;
371 char execname[PATH_MAX];
372 char anonstr[] = "//anon";
377 if (fgets(bf, sizeof(bf), fp) == NULL)
380 if ((rdclock() - t) > timeout) {
381 pr_warning("Reading %s time out. "
382 "You may want to increase "
383 "the time limit by --proc-map-timeout\n",
389 /* ensure null termination since stack will be reused. */
390 strcpy(execname, "");
392 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
393 n = sscanf(bf, "%"PRI_lx64"-%"PRI_lx64" %s %"PRI_lx64" %x:%x %u %[^\n]\n",
394 &event->mmap2.start, &event->mmap2.len, prot,
395 &event->mmap2.pgoff, &event->mmap2.maj,
400 * Anon maps don't have the execname.
405 event->mmap2.ino = (u64)ino;
408 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
410 if (machine__is_host(machine))
411 event->header.misc = PERF_RECORD_MISC_USER;
413 event->header.misc = PERF_RECORD_MISC_GUEST_USER;
415 /* map protection and flags bits */
416 event->mmap2.prot = 0;
417 event->mmap2.flags = 0;
419 event->mmap2.prot |= PROT_READ;
421 event->mmap2.prot |= PROT_WRITE;
423 event->mmap2.prot |= PROT_EXEC;
426 event->mmap2.flags |= MAP_SHARED;
428 event->mmap2.flags |= MAP_PRIVATE;
430 if (prot[2] != 'x') {
431 if (!mmap_data || prot[0] != 'r')
434 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
439 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
441 if (!strcmp(execname, ""))
442 strcpy(execname, anonstr);
444 if (hugetlbfs_mnt_len &&
445 !strncmp(execname, hugetlbfs_mnt, hugetlbfs_mnt_len)) {
446 strcpy(execname, anonstr);
447 event->mmap2.flags |= MAP_HUGETLB;
450 size = strlen(execname) + 1;
451 memcpy(event->mmap2.filename, execname, size);
452 size = PERF_ALIGN(size, sizeof(u64));
453 event->mmap2.len -= event->mmap.start;
454 event->mmap2.header.size = (sizeof(event->mmap2) -
455 (sizeof(event->mmap2.filename) - size));
456 memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
457 event->mmap2.header.size += machine->id_hdr_size;
458 event->mmap2.pid = tgid;
459 event->mmap2.tid = pid;
461 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
474 int perf_event__synthesize_modules(struct perf_tool *tool,
475 perf_event__handler_t process,
476 struct machine *machine)
480 struct maps *maps = machine__kernel_maps(machine);
481 union perf_event *event = zalloc((sizeof(event->mmap) +
482 machine->id_hdr_size));
484 pr_debug("Not enough memory synthesizing mmap event "
485 "for kernel modules\n");
489 event->header.type = PERF_RECORD_MMAP;
492 * kernel uses 0 for user space maps, see kernel/perf_event.c
495 if (machine__is_host(machine))
496 event->header.misc = PERF_RECORD_MISC_KERNEL;
498 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
500 for (pos = maps__first(maps); pos; pos = map__next(pos)) {
503 if (!__map__is_kmodule(pos))
506 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
507 event->mmap.header.type = PERF_RECORD_MMAP;
508 event->mmap.header.size = (sizeof(event->mmap) -
509 (sizeof(event->mmap.filename) - size));
510 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
511 event->mmap.header.size += machine->id_hdr_size;
512 event->mmap.start = pos->start;
513 event->mmap.len = pos->end - pos->start;
514 event->mmap.pid = machine->pid;
516 memcpy(event->mmap.filename, pos->dso->long_name,
517 pos->dso->long_name_len + 1);
518 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
528 static int __event__synthesize_thread(union perf_event *comm_event,
529 union perf_event *mmap_event,
530 union perf_event *fork_event,
531 union perf_event *namespaces_event,
533 perf_event__handler_t process,
534 struct perf_tool *tool,
535 struct machine *machine,
538 char filename[PATH_MAX];
540 struct dirent *dirent;
544 /* special case: only send one comm event using passed in pid */
546 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
552 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
553 tgid, process, machine) < 0)
557 * send mmap only for thread group leader
558 * see thread__init_map_groups
561 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
562 process, machine, mmap_data))
568 if (machine__is_default_guest(machine))
571 snprintf(filename, sizeof(filename), "%s/proc/%d/task",
572 machine->root_dir, pid);
574 tasks = opendir(filename);
576 pr_debug("couldn't open %s\n", filename);
580 while ((dirent = readdir(tasks)) != NULL) {
584 _pid = strtol(dirent->d_name, &end, 10);
589 if (perf_event__prepare_comm(comm_event, _pid, machine,
593 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
594 ppid, process, machine) < 0)
597 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
598 tgid, process, machine) < 0)
602 * Send the prepared comm event
604 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
609 /* process the parent's maps too */
610 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
611 process, machine, mmap_data);
621 int perf_event__synthesize_thread_map(struct perf_tool *tool,
622 struct perf_thread_map *threads,
623 perf_event__handler_t process,
624 struct machine *machine,
627 union perf_event *comm_event, *mmap_event, *fork_event;
628 union perf_event *namespaces_event;
629 int err = -1, thread, j;
631 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
632 if (comm_event == NULL)
635 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
636 if (mmap_event == NULL)
639 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
640 if (fork_event == NULL)
643 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
644 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
645 machine->id_hdr_size);
646 if (namespaces_event == NULL)
650 for (thread = 0; thread < threads->nr; ++thread) {
651 if (__event__synthesize_thread(comm_event, mmap_event,
652 fork_event, namespaces_event,
653 perf_thread_map__pid(threads, thread), 0,
654 process, tool, machine,
661 * comm.pid is set to thread group id by
662 * perf_event__synthesize_comm
664 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) {
665 bool need_leader = true;
667 /* is thread group leader in thread_map? */
668 for (j = 0; j < threads->nr; ++j) {
669 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) {
675 /* if not, generate events for it */
677 __event__synthesize_thread(comm_event, mmap_event,
678 fork_event, namespaces_event,
679 comm_event->comm.pid, 0,
680 process, tool, machine,
687 free(namespaces_event);
698 static int __perf_event__synthesize_threads(struct perf_tool *tool,
699 perf_event__handler_t process,
700 struct machine *machine,
702 struct dirent **dirent,
706 union perf_event *comm_event, *mmap_event, *fork_event;
707 union perf_event *namespaces_event;
713 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
714 if (comm_event == NULL)
717 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
718 if (mmap_event == NULL)
721 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
722 if (fork_event == NULL)
725 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
726 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
727 machine->id_hdr_size);
728 if (namespaces_event == NULL)
731 for (i = start; i < start + num; i++) {
732 if (!isdigit(dirent[i]->d_name[0]))
735 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
736 /* only interested in proper numerical dirents */
740 * We may race with exiting thread, so don't stop just because
741 * one thread couldn't be synthesized.
743 __event__synthesize_thread(comm_event, mmap_event, fork_event,
744 namespaces_event, pid, 1, process,
745 tool, machine, mmap_data);
749 free(namespaces_event);
760 struct synthesize_threads_arg {
761 struct perf_tool *tool;
762 perf_event__handler_t process;
763 struct machine *machine;
765 struct dirent **dirent;
770 static void *synthesize_threads_worker(void *arg)
772 struct synthesize_threads_arg *args = arg;
774 __perf_event__synthesize_threads(args->tool, args->process,
775 args->machine, args->mmap_data,
777 args->start, args->num);
781 int perf_event__synthesize_threads(struct perf_tool *tool,
782 perf_event__handler_t process,
783 struct machine *machine,
785 unsigned int nr_threads_synthesize)
787 struct synthesize_threads_arg *args = NULL;
788 pthread_t *synthesize_threads = NULL;
789 char proc_path[PATH_MAX];
790 struct dirent **dirent;
798 if (machine__is_default_guest(machine))
801 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
802 n = scandir(proc_path, &dirent, 0, alphasort);
806 if (nr_threads_synthesize == UINT_MAX)
807 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
809 thread_nr = nr_threads_synthesize;
811 if (thread_nr <= 1) {
812 err = __perf_event__synthesize_threads(tool, process,
820 synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
821 if (synthesize_threads == NULL)
824 args = calloc(sizeof(*args), thread_nr);
828 num_per_thread = n / thread_nr;
830 for (i = 0; i < thread_nr; i++) {
832 args[i].process = process;
833 args[i].machine = machine;
834 args[i].mmap_data = mmap_data;
835 args[i].dirent = dirent;
837 for (i = 0; i < m; i++) {
838 args[i].num = num_per_thread + 1;
839 args[i].start = i * args[i].num;
842 base = args[i-1].start + args[i-1].num;
843 for (j = i; j < thread_nr; j++) {
844 args[j].num = num_per_thread;
845 args[j].start = base + (j - i) * args[i].num;
848 for (i = 0; i < thread_nr; i++) {
849 if (pthread_create(&synthesize_threads[i], NULL,
850 synthesize_threads_worker, &args[i]))
855 for (i = 0; i < thread_nr; i++)
856 pthread_join(synthesize_threads[i], NULL);
859 free(synthesize_threads);
861 for (i = 0; i < n; i++)
868 struct process_symbol_args {
873 static int find_symbol_cb(void *arg, const char *name, char type,
876 struct process_symbol_args *args = arg;
879 * Must be a function or at least an alias, as in PARISC64, where "_text" is
880 * an 'A' to the same address as "_stext".
882 if (!(kallsyms__is_function(type) ||
883 type == 'A') || strcmp(name, args->name))
890 int kallsyms__get_function_start(const char *kallsyms_filename,
891 const char *symbol_name, u64 *addr)
893 struct process_symbol_args args = { .name = symbol_name, };
895 if (kallsyms__parse(kallsyms_filename, &args, find_symbol_cb) <= 0)
902 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
903 perf_event__handler_t process __maybe_unused,
904 struct machine *machine __maybe_unused)
909 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
910 perf_event__handler_t process,
911 struct machine *machine)
914 struct map *map = machine__kernel_map(machine);
917 union perf_event *event;
922 kmap = map__kmap(map);
923 if (!kmap->ref_reloc_sym)
927 * We should get this from /sys/kernel/sections/.text, but till that is
928 * available use this, and after it is use this as a fallback for older
931 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size));
933 pr_debug("Not enough memory synthesizing mmap event "
934 "for kernel modules\n");
938 if (machine__is_host(machine)) {
940 * kernel uses PERF_RECORD_MISC_USER for user space maps,
941 * see kernel/perf_event.c __perf_event_mmap
943 event->header.misc = PERF_RECORD_MISC_KERNEL;
945 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
948 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
949 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
950 size = PERF_ALIGN(size, sizeof(u64));
951 event->mmap.header.type = PERF_RECORD_MMAP;
952 event->mmap.header.size = (sizeof(event->mmap) -
953 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
954 event->mmap.pgoff = kmap->ref_reloc_sym->addr;
955 event->mmap.start = map->start;
956 event->mmap.len = map->end - event->mmap.start;
957 event->mmap.pid = machine->pid;
959 err = perf_tool__process_synth_event(tool, event, machine, process);
965 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
966 perf_event__handler_t process,
967 struct machine *machine)
971 err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
975 return perf_event__synthesize_extra_kmaps(tool, process, machine);
978 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
979 struct perf_thread_map *threads,
980 perf_event__handler_t process,
981 struct machine *machine)
983 union perf_event *event;
986 size = sizeof(event->thread_map);
987 size += threads->nr * sizeof(event->thread_map.entries[0]);
989 event = zalloc(size);
993 event->header.type = PERF_RECORD_THREAD_MAP;
994 event->header.size = size;
995 event->thread_map.nr = threads->nr;
997 for (i = 0; i < threads->nr; i++) {
998 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i];
999 char *comm = perf_thread_map__comm(threads, i);
1004 entry->pid = perf_thread_map__pid(threads, i);
1005 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
1008 err = process(tool, event, NULL, machine);
1014 static void synthesize_cpus(struct cpu_map_entries *cpus,
1015 struct perf_cpu_map *map)
1021 for (i = 0; i < map->nr; i++)
1022 cpus->cpu[i] = map->map[i];
1025 static void synthesize_mask(struct perf_record_record_cpu_map *mask,
1026 struct perf_cpu_map *map, int max)
1030 mask->nr = BITS_TO_LONGS(max);
1031 mask->long_size = sizeof(long);
1033 for (i = 0; i < map->nr; i++)
1034 set_bit(map->map[i], mask->mask);
1037 static size_t cpus_size(struct perf_cpu_map *map)
1039 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16);
1042 static size_t mask_size(struct perf_cpu_map *map, int *max)
1048 for (i = 0; i < map->nr; i++) {
1049 /* bit possition of the cpu is + 1 */
1050 int bit = map->map[i] + 1;
1056 return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long);
1059 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max)
1061 size_t size_cpus, size_mask;
1062 bool is_dummy = perf_cpu_map__empty(map);
1065 * Both array and mask data have variable size based
1066 * on the number of cpus and their actual values.
1067 * The size of the 'struct perf_record_cpu_map_data' is:
1069 * array = size of 'struct cpu_map_entries' +
1070 * number of cpus * sizeof(u64)
1072 * mask = size of 'struct perf_record_record_cpu_map' +
1073 * maximum cpu bit converted to size of longs
1075 * and finaly + the size of 'struct perf_record_cpu_map_data'.
1077 size_cpus = cpus_size(map);
1078 size_mask = mask_size(map, max);
1080 if (is_dummy || (size_cpus < size_mask)) {
1082 *type = PERF_CPU_MAP__CPUS;
1085 *type = PERF_CPU_MAP__MASK;
1088 *size += sizeof(struct perf_record_cpu_map_data);
1089 *size = PERF_ALIGN(*size, sizeof(u64));
1090 return zalloc(*size);
1093 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
1099 case PERF_CPU_MAP__CPUS:
1100 synthesize_cpus((struct cpu_map_entries *) data->data, map);
1102 case PERF_CPU_MAP__MASK:
1103 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max);
1109 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map)
1111 size_t size = sizeof(struct perf_record_cpu_map);
1112 struct perf_record_cpu_map *event;
1116 event = cpu_map_data__alloc(map, &size, &type, &max);
1120 event->header.type = PERF_RECORD_CPU_MAP;
1121 event->header.size = size;
1122 event->data.type = type;
1124 cpu_map_data__synthesize(&event->data, map, type, max);
1128 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1129 struct perf_cpu_map *map,
1130 perf_event__handler_t process,
1131 struct machine *machine)
1133 struct perf_record_cpu_map *event;
1136 event = cpu_map_event__new(map);
1140 err = process(tool, (union perf_event *) event, NULL, machine);
1146 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1147 struct perf_stat_config *config,
1148 perf_event__handler_t process,
1149 struct machine *machine)
1151 struct perf_record_stat_config *event;
1152 int size, i = 0, err;
1154 size = sizeof(*event);
1155 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1157 event = zalloc(size);
1161 event->header.type = PERF_RECORD_STAT_CONFIG;
1162 event->header.size = size;
1163 event->nr = PERF_STAT_CONFIG_TERM__MAX;
1165 #define ADD(__term, __val) \
1166 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \
1167 event->data[i].val = __val; \
1170 ADD(AGGR_MODE, config->aggr_mode)
1171 ADD(INTERVAL, config->interval)
1172 ADD(SCALE, config->scale)
1174 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1175 "stat config terms unbalanced\n");
1178 err = process(tool, (union perf_event *) event, NULL, machine);
1184 int perf_event__synthesize_stat(struct perf_tool *tool,
1185 u32 cpu, u32 thread, u64 id,
1186 struct perf_counts_values *count,
1187 perf_event__handler_t process,
1188 struct machine *machine)
1190 struct perf_record_stat event;
1192 event.header.type = PERF_RECORD_STAT;
1193 event.header.size = sizeof(event);
1194 event.header.misc = 0;
1198 event.thread = thread;
1199 event.val = count->val;
1200 event.ena = count->ena;
1201 event.run = count->run;
1203 return process(tool, (union perf_event *) &event, NULL, machine);
1206 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1207 u64 evtime, u64 type,
1208 perf_event__handler_t process,
1209 struct machine *machine)
1211 struct perf_record_stat_round event;
1213 event.header.type = PERF_RECORD_STAT_ROUND;
1214 event.header.size = sizeof(event);
1215 event.header.misc = 0;
1217 event.time = evtime;
1220 return process(tool, (union perf_event *) &event, NULL, machine);
1223 void perf_event__read_stat_config(struct perf_stat_config *config,
1224 struct perf_record_stat_config *event)
1228 for (i = 0; i < event->nr; i++) {
1230 switch (event->data[i].tag) {
1231 #define CASE(__term, __val) \
1232 case PERF_STAT_CONFIG_TERM__##__term: \
1233 config->__val = event->data[i].val; \
1236 CASE(AGGR_MODE, aggr_mode)
1238 CASE(INTERVAL, interval)
1241 pr_warning("unknown stat config term %" PRI_lu64 "\n",
1242 event->data[i].tag);
1247 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp)
1251 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC)
1256 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid);
1259 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp)
1262 struct perf_ns_link_info *ns_link_info;
1263 u32 nr_namespaces, idx;
1265 ns_link_info = event->namespaces.link_info;
1266 nr_namespaces = event->namespaces.nr_namespaces;
1268 ret += fprintf(fp, " %d/%d - nr_namespaces: %u\n\t\t[",
1269 event->namespaces.pid,
1270 event->namespaces.tid,
1273 for (idx = 0; idx < nr_namespaces; idx++) {
1274 if (idx && (idx % 4 == 0))
1275 ret += fprintf(fp, "\n\t\t ");
1277 ret += fprintf(fp, "%u/%s: %" PRIu64 "/%#" PRIx64 "%s", idx,
1278 perf_ns__name(idx), (u64)ns_link_info[idx].dev,
1279 (u64)ns_link_info[idx].ino,
1280 ((idx + 1) != nr_namespaces) ? ", " : "]\n");
1286 int perf_event__process_comm(struct perf_tool *tool __maybe_unused,
1287 union perf_event *event,
1288 struct perf_sample *sample,
1289 struct machine *machine)
1291 return machine__process_comm_event(machine, event, sample);
1294 int perf_event__process_namespaces(struct perf_tool *tool __maybe_unused,
1295 union perf_event *event,
1296 struct perf_sample *sample,
1297 struct machine *machine)
1299 return machine__process_namespaces_event(machine, event, sample);
1302 int perf_event__process_lost(struct perf_tool *tool __maybe_unused,
1303 union perf_event *event,
1304 struct perf_sample *sample,
1305 struct machine *machine)
1307 return machine__process_lost_event(machine, event, sample);
1310 int perf_event__process_aux(struct perf_tool *tool __maybe_unused,
1311 union perf_event *event,
1312 struct perf_sample *sample __maybe_unused,
1313 struct machine *machine)
1315 return machine__process_aux_event(machine, event);
1318 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused,
1319 union perf_event *event,
1320 struct perf_sample *sample __maybe_unused,
1321 struct machine *machine)
1323 return machine__process_itrace_start_event(machine, event);
1326 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused,
1327 union perf_event *event,
1328 struct perf_sample *sample,
1329 struct machine *machine)
1331 return machine__process_lost_samples_event(machine, event, sample);
1334 int perf_event__process_switch(struct perf_tool *tool __maybe_unused,
1335 union perf_event *event,
1336 struct perf_sample *sample __maybe_unused,
1337 struct machine *machine)
1339 return machine__process_switch_event(machine, event);
1342 int perf_event__process_ksymbol(struct perf_tool *tool __maybe_unused,
1343 union perf_event *event,
1344 struct perf_sample *sample __maybe_unused,
1345 struct machine *machine)
1347 return machine__process_ksymbol(machine, event, sample);
1350 int perf_event__process_bpf(struct perf_tool *tool __maybe_unused,
1351 union perf_event *event,
1352 struct perf_sample *sample,
1353 struct machine *machine)
1355 return machine__process_bpf(machine, event, sample);
1358 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp)
1360 return fprintf(fp, " %d/%d: [%#" PRI_lx64 "(%#" PRI_lx64 ") @ %#" PRI_lx64 "]: %c %s\n",
1361 event->mmap.pid, event->mmap.tid, event->mmap.start,
1362 event->mmap.len, event->mmap.pgoff,
1363 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x',
1364 event->mmap.filename);
1367 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp)
1369 return fprintf(fp, " %d/%d: [%#" PRI_lx64 "(%#" PRI_lx64 ") @ %#" PRI_lx64
1370 " %02x:%02x %"PRI_lu64" %"PRI_lu64"]: %c%c%c%c %s\n",
1371 event->mmap2.pid, event->mmap2.tid, event->mmap2.start,
1372 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj,
1373 event->mmap2.min, event->mmap2.ino,
1374 event->mmap2.ino_generation,
1375 (event->mmap2.prot & PROT_READ) ? 'r' : '-',
1376 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-',
1377 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-',
1378 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p',
1379 event->mmap2.filename);
1382 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp)
1384 struct perf_thread_map *threads = thread_map__new_event(&event->thread_map);
1387 ret = fprintf(fp, " nr: ");
1390 ret += thread_map__fprintf(threads, fp);
1392 ret += fprintf(fp, "failed to get threads from event\n");
1394 perf_thread_map__put(threads);
1398 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp)
1400 struct perf_cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data);
1403 ret = fprintf(fp, ": ");
1406 ret += cpu_map__fprintf(cpus, fp);
1408 ret += fprintf(fp, "failed to get cpumap from event\n");
1410 perf_cpu_map__put(cpus);
1414 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused,
1415 union perf_event *event,
1416 struct perf_sample *sample,
1417 struct machine *machine)
1419 return machine__process_mmap_event(machine, event, sample);
1422 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused,
1423 union perf_event *event,
1424 struct perf_sample *sample,
1425 struct machine *machine)
1427 return machine__process_mmap2_event(machine, event, sample);
1430 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp)
1432 return fprintf(fp, "(%d:%d):(%d:%d)\n",
1433 event->fork.pid, event->fork.tid,
1434 event->fork.ppid, event->fork.ptid);
1437 int perf_event__process_fork(struct perf_tool *tool __maybe_unused,
1438 union perf_event *event,
1439 struct perf_sample *sample,
1440 struct machine *machine)
1442 return machine__process_fork_event(machine, event, sample);
1445 int perf_event__process_exit(struct perf_tool *tool __maybe_unused,
1446 union perf_event *event,
1447 struct perf_sample *sample,
1448 struct machine *machine)
1450 return machine__process_exit_event(machine, event, sample);
1453 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp)
1455 return fprintf(fp, " offset: %#"PRI_lx64" size: %#"PRI_lx64" flags: %#"PRI_lx64" [%s%s%s]\n",
1456 event->aux.aux_offset, event->aux.aux_size,
1458 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "",
1459 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : "",
1460 event->aux.flags & PERF_AUX_FLAG_PARTIAL ? "P" : "");
1463 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp)
1465 return fprintf(fp, " pid: %u tid: %u\n",
1466 event->itrace_start.pid, event->itrace_start.tid);
1469 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp)
1471 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1472 const char *in_out = !out ? "IN " :
1473 !(event->header.misc & PERF_RECORD_MISC_SWITCH_OUT_PREEMPT) ?
1474 "OUT " : "OUT preempt";
1476 if (event->header.type == PERF_RECORD_SWITCH)
1477 return fprintf(fp, " %s\n", in_out);
1479 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n",
1480 in_out, out ? "next" : "prev",
1481 event->context_switch.next_prev_pid,
1482 event->context_switch.next_prev_tid);
1485 static size_t perf_event__fprintf_lost(union perf_event *event, FILE *fp)
1487 return fprintf(fp, " lost %" PRI_lu64 "\n", event->lost.lost);
1490 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp)
1492 return fprintf(fp, " addr %" PRI_lx64 " len %u type %u flags 0x%x name %s\n",
1493 event->ksymbol.addr, event->ksymbol.len,
1494 event->ksymbol.ksym_type,
1495 event->ksymbol.flags, event->ksymbol.name);
1498 size_t perf_event__fprintf_bpf(union perf_event *event, FILE *fp)
1500 return fprintf(fp, " type %u, flags %u, id %u\n",
1501 event->bpf.type, event->bpf.flags, event->bpf.id);
1504 size_t perf_event__fprintf(union perf_event *event, FILE *fp)
1506 size_t ret = fprintf(fp, "PERF_RECORD_%s",
1507 perf_event__name(event->header.type));
1509 switch (event->header.type) {
1510 case PERF_RECORD_COMM:
1511 ret += perf_event__fprintf_comm(event, fp);
1513 case PERF_RECORD_FORK:
1514 case PERF_RECORD_EXIT:
1515 ret += perf_event__fprintf_task(event, fp);
1517 case PERF_RECORD_MMAP:
1518 ret += perf_event__fprintf_mmap(event, fp);
1520 case PERF_RECORD_NAMESPACES:
1521 ret += perf_event__fprintf_namespaces(event, fp);
1523 case PERF_RECORD_MMAP2:
1524 ret += perf_event__fprintf_mmap2(event, fp);
1526 case PERF_RECORD_AUX:
1527 ret += perf_event__fprintf_aux(event, fp);
1529 case PERF_RECORD_ITRACE_START:
1530 ret += perf_event__fprintf_itrace_start(event, fp);
1532 case PERF_RECORD_SWITCH:
1533 case PERF_RECORD_SWITCH_CPU_WIDE:
1534 ret += perf_event__fprintf_switch(event, fp);
1536 case PERF_RECORD_LOST:
1537 ret += perf_event__fprintf_lost(event, fp);
1539 case PERF_RECORD_KSYMBOL:
1540 ret += perf_event__fprintf_ksymbol(event, fp);
1542 case PERF_RECORD_BPF_EVENT:
1543 ret += perf_event__fprintf_bpf(event, fp);
1546 ret += fprintf(fp, "\n");
1552 int perf_event__process(struct perf_tool *tool __maybe_unused,
1553 union perf_event *event,
1554 struct perf_sample *sample,
1555 struct machine *machine)
1557 return machine__process_event(machine, event, sample);
1560 struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
1561 struct addr_location *al)
1563 struct map_groups *mg = thread->mg;
1564 struct machine *machine = mg->machine;
1565 bool load_map = false;
1567 al->machine = machine;
1568 al->thread = thread;
1570 al->cpumode = cpumode;
1573 if (machine == NULL) {
1578 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
1580 mg = &machine->kmaps;
1582 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) {
1584 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
1586 mg = &machine->kmaps;
1588 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) {
1594 if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
1595 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
1597 al->filtered |= (1 << HIST_FILTER__GUEST);
1598 if ((cpumode == PERF_RECORD_MISC_USER ||
1599 cpumode == PERF_RECORD_MISC_KERNEL) &&
1601 al->filtered |= (1 << HIST_FILTER__HOST);
1606 al->map = map_groups__find(mg, al->addr);
1607 if (al->map != NULL) {
1609 * Kernel maps might be changed when loading symbols so loading
1610 * must be done prior to using kernel maps.
1614 al->addr = al->map->map_ip(al->map, al->addr);
1621 * For branch stacks or branch samples, the sample cpumode might not be correct
1622 * because it applies only to the sample 'ip' and not necessary to 'addr' or
1623 * branch stack addresses. If possible, use a fallback to deal with those cases.
1625 struct map *thread__find_map_fb(struct thread *thread, u8 cpumode, u64 addr,
1626 struct addr_location *al)
1628 struct map *map = thread__find_map(thread, cpumode, addr, al);
1629 struct machine *machine = thread->mg->machine;
1630 u8 addr_cpumode = machine__addr_cpumode(machine, cpumode, addr);
1632 if (map || addr_cpumode == cpumode)
1635 return thread__find_map(thread, addr_cpumode, addr, al);
1638 struct symbol *thread__find_symbol(struct thread *thread, u8 cpumode,
1639 u64 addr, struct addr_location *al)
1642 if (thread__find_map(thread, cpumode, addr, al))
1643 al->sym = map__find_symbol(al->map, al->addr);
1647 struct symbol *thread__find_symbol_fb(struct thread *thread, u8 cpumode,
1648 u64 addr, struct addr_location *al)
1651 if (thread__find_map_fb(thread, cpumode, addr, al))
1652 al->sym = map__find_symbol(al->map, al->addr);
1657 * Callers need to drop the reference to al->thread, obtained in
1658 * machine__findnew_thread()
1660 int machine__resolve(struct machine *machine, struct addr_location *al,
1661 struct perf_sample *sample)
1663 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1669 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
1670 thread__find_map(thread, sample->cpumode, sample->ip, al);
1671 dump_printf(" ...... dso: %s\n",
1672 al->map ? al->map->dso->long_name :
1673 al->level == 'H' ? "[hypervisor]" : "<not found>");
1675 if (thread__is_filtered(thread))
1676 al->filtered |= (1 << HIST_FILTER__THREAD);
1679 al->cpu = sample->cpu;
1684 struct perf_env *env = machine->env;
1686 if (env && env->cpu)
1687 al->socket = env->cpu[al->cpu].socket_id;
1691 struct dso *dso = al->map->dso;
1693 if (symbol_conf.dso_list &&
1694 (!dso || !(strlist__has_entry(symbol_conf.dso_list,
1696 (dso->short_name != dso->long_name &&
1697 strlist__has_entry(symbol_conf.dso_list,
1698 dso->long_name))))) {
1699 al->filtered |= (1 << HIST_FILTER__DSO);
1702 al->sym = map__find_symbol(al->map, al->addr);
1705 if (symbol_conf.sym_list &&
1706 (!al->sym || !strlist__has_entry(symbol_conf.sym_list,
1708 al->filtered |= (1 << HIST_FILTER__SYMBOL);
1715 * The preprocess_sample method will return with reference counts for the
1716 * in it, when done using (and perhaps getting ref counts if needing to
1717 * keep a pointer to one of those entries) it must be paired with
1718 * addr_location__put(), so that the refcounts can be decremented.
1720 void addr_location__put(struct addr_location *al)
1722 thread__zput(al->thread);
1725 bool is_bts_event(struct perf_event_attr *attr)
1727 return attr->type == PERF_TYPE_HARDWARE &&
1728 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1729 attr->sample_period == 1;
1732 bool sample_addr_correlates_sym(struct perf_event_attr *attr)
1734 if (attr->type == PERF_TYPE_SOFTWARE &&
1735 (attr->config == PERF_COUNT_SW_PAGE_FAULTS ||
1736 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
1737 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ))
1740 if (is_bts_event(attr))
1746 void thread__resolve(struct thread *thread, struct addr_location *al,
1747 struct perf_sample *sample)
1749 thread__find_map_fb(thread, sample->cpumode, sample->addr, al);
1751 al->cpu = sample->cpu;
1755 al->sym = map__find_symbol(al->map, al->addr);