1 // SPDX-License-Identifier: GPL-2.0-only
3 #include "util/cgroup.h"
5 #include "util/debug.h"
7 #include "util/event.h"
8 #include "util/evlist.h"
9 #include "util/machine.h"
11 #include "util/map_symbol.h"
12 #include "util/branch.h"
13 #include "util/memswap.h"
14 #include "util/namespaces.h"
15 #include "util/session.h"
16 #include "util/stat.h"
17 #include "util/symbol.h"
18 #include "util/synthetic-events.h"
19 #include "util/target.h"
20 #include "util/time-utils.h"
21 #include <linux/bitops.h>
22 #include <linux/kernel.h>
23 #include <linux/string.h>
24 #include <linux/zalloc.h>
25 #include <linux/perf_event.h>
27 #include <perf/evsel.h>
28 #include <perf/cpumap.h>
29 #include <internal/lib.h> // page_size
30 #include <internal/threadmap.h>
31 #include <perf/threadmap.h>
32 #include <symbol/kallsyms.h>
38 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */
39 #include <api/fs/fs.h>
41 #include <sys/types.h>
46 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500
48 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT;
50 int perf_tool__process_synth_event(struct perf_tool *tool,
51 union perf_event *event,
52 struct machine *machine,
53 perf_event__handler_t process)
55 struct perf_sample synth_sample = {
62 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK,
65 return process(tool, event, &synth_sample, machine);
69 * Assumes that the first 4095 bytes of /proc/pid/stat contains
70 * the comm, tgid and ppid.
72 static int perf_event__get_comm_ids(pid_t pid, pid_t tid, char *comm, size_t len,
73 pid_t *tgid, pid_t *ppid, bool *kernel)
79 char *name, *tgids, *ppids, *vmpeak, *threads;
85 snprintf(bf, sizeof(bf), "/proc/%d/task/%d/status", pid, tid);
87 snprintf(bf, sizeof(bf), "/proc/%d/status", tid);
89 fd = open(bf, O_RDONLY);
91 pr_debug("couldn't open %s\n", bf);
95 n = read(fd, bf, sizeof(bf) - 1);
98 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
104 name = strstr(bf, "Name:");
105 tgids = strstr(name ?: bf, "Tgid:");
106 ppids = strstr(tgids ?: bf, "PPid:");
107 vmpeak = strstr(ppids ?: bf, "VmPeak:");
112 threads = strstr(ppids ?: bf, "Threads:");
117 name = skip_spaces(name + 5); /* strlen("Name:") */
118 nl = strchr(name, '\n');
125 memcpy(comm, name, size);
128 pr_debug("Name: string not found for pid %d\n", tid);
132 tgids += 5; /* strlen("Tgid:") */
135 pr_debug("Tgid: string not found for pid %d\n", tid);
139 ppids += 5; /* strlen("PPid:") */
142 pr_debug("PPid: string not found for pid %d\n", tid);
145 if (!vmpeak && threads)
153 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, pid_t tid,
154 struct machine *machine,
155 pid_t *tgid, pid_t *ppid, bool *kernel)
161 memset(&event->comm, 0, sizeof(event->comm));
163 if (machine__is_host(machine)) {
164 if (perf_event__get_comm_ids(pid, tid, event->comm.comm,
165 sizeof(event->comm.comm),
166 tgid, ppid, kernel) != 0) {
170 *tgid = machine->pid;
176 event->comm.pid = *tgid;
177 event->comm.header.type = PERF_RECORD_COMM;
179 size = strlen(event->comm.comm) + 1;
180 size = PERF_ALIGN(size, sizeof(u64));
181 memset(event->comm.comm + size, 0, machine->id_hdr_size);
182 event->comm.header.size = (sizeof(event->comm) -
183 (sizeof(event->comm.comm) - size) +
184 machine->id_hdr_size);
185 event->comm.tid = tid;
190 pid_t perf_event__synthesize_comm(struct perf_tool *tool,
191 union perf_event *event, pid_t pid,
192 perf_event__handler_t process,
193 struct machine *machine)
198 if (perf_event__prepare_comm(event, 0, pid, machine, &tgid, &ppid,
199 &kernel_thread) != 0)
202 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
208 static void perf_event__get_ns_link_info(pid_t pid, const char *ns,
209 struct perf_ns_link_info *ns_link_info)
214 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns);
215 if (stat64(proc_ns, &st) == 0) {
216 ns_link_info->dev = st.st_dev;
217 ns_link_info->ino = st.st_ino;
221 int perf_event__synthesize_namespaces(struct perf_tool *tool,
222 union perf_event *event,
223 pid_t pid, pid_t tgid,
224 perf_event__handler_t process,
225 struct machine *machine)
228 struct perf_ns_link_info *ns_link_info;
230 if (!tool || !tool->namespace_events)
233 memset(&event->namespaces, 0, (sizeof(event->namespaces) +
234 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
235 machine->id_hdr_size));
237 event->namespaces.pid = tgid;
238 event->namespaces.tid = pid;
240 event->namespaces.nr_namespaces = NR_NAMESPACES;
242 ns_link_info = event->namespaces.link_info;
244 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++)
245 perf_event__get_ns_link_info(pid, perf_ns__name(idx),
248 event->namespaces.header.type = PERF_RECORD_NAMESPACES;
250 event->namespaces.header.size = (sizeof(event->namespaces) +
251 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
252 machine->id_hdr_size);
254 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
260 static int perf_event__synthesize_fork(struct perf_tool *tool,
261 union perf_event *event,
262 pid_t pid, pid_t tgid, pid_t ppid,
263 perf_event__handler_t process,
264 struct machine *machine)
266 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
269 * for main thread set parent to ppid from status file. For other
270 * threads set parent pid to main thread. ie., assume main thread
271 * spawns all threads in a process
274 event->fork.ppid = ppid;
275 event->fork.ptid = ppid;
277 event->fork.ppid = tgid;
278 event->fork.ptid = tgid;
280 event->fork.pid = tgid;
281 event->fork.tid = pid;
282 event->fork.header.type = PERF_RECORD_FORK;
283 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC;
285 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
287 if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
293 static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end,
294 u32 *prot, u32 *flags, __u64 *offset,
297 ssize_t pathname_size, char *pathname)
301 char *start_pathname = pathname;
303 if (io__get_hex(io, start) != '-')
305 if (io__get_hex(io, end) != ' ')
308 /* map protection and flags bits */
310 ch = io__get_char(io);
315 ch = io__get_char(io);
320 ch = io__get_char(io);
325 ch = io__get_char(io);
329 *flags = MAP_PRIVATE;
332 if (io__get_char(io) != ' ')
335 if (io__get_hex(io, offset) != ' ')
338 if (io__get_hex(io, &temp) != ':')
341 if (io__get_hex(io, &temp) != ' ')
345 ch = io__get_dec(io, inode);
351 ch = io__get_char(io);
356 if (ch == '\0' || ch == '\n' ||
357 (pathname + 1 - start_pathname) >= pathname_size) {
362 ch = io__get_char(io);
366 static void perf_record_mmap2__read_build_id(struct perf_record_mmap2 *event,
367 struct machine *machine,
373 struct dso *dso = NULL;
378 rc = sysfs__read_build_id("/sys/kernel/notes", &bid);
385 id.ino_generation = event->ino_generation;
387 dso = dsos__findnew_id(&machine->dsos, event->filename, &id);
388 if (dso && dso->has_build_id) {
394 nsi = nsinfo__new(event->pid);
395 nsinfo__mountns_enter(nsi, &nc);
397 rc = filename__read_build_id(event->filename, &bid) > 0 ? 0 : -1;
399 nsinfo__mountns_exit(&nc);
404 memcpy(event->build_id, bid.data, sizeof(bid.data));
405 event->build_id_size = (u8) bid.size;
406 event->header.misc |= PERF_RECORD_MISC_MMAP_BUILD_ID;
407 event->__reserved_1 = 0;
408 event->__reserved_2 = 0;
410 if (dso && !dso->has_build_id)
411 dso__set_build_id(dso, &bid);
413 if (event->filename[0] == '/') {
414 pr_debug2("Failed to read build ID for %s\n",
421 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
422 union perf_event *event,
423 pid_t pid, pid_t tgid,
424 perf_event__handler_t process,
425 struct machine *machine,
428 unsigned long long t;
431 bool truncation = false;
432 unsigned long long timeout = proc_map_timeout * 1000000ULL;
434 const char *hugetlbfs_mnt = hugetlbfs__mountpoint();
435 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0;
437 if (machine__is_default_guest(machine))
440 snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps",
441 machine->root_dir, pid, pid);
443 io.fd = open(bf, O_RDONLY, 0);
446 * We raced with a task exiting - just return:
448 pr_debug("couldn't open %s\n", bf);
451 io__init(&io, io.fd, bf, sizeof(bf));
453 event->header.type = PERF_RECORD_MMAP2;
457 static const char anonstr[] = "//anon";
458 size_t size, aligned_size;
460 /* ensure null termination since stack will be reused. */
461 event->mmap2.filename[0] = '\0';
463 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
464 if (!read_proc_maps_line(&io,
473 sizeof(event->mmap2.filename),
474 event->mmap2.filename))
477 if ((rdclock() - t) > timeout) {
478 pr_warning("Reading %s/proc/%d/task/%d/maps time out. "
479 "You may want to increase "
480 "the time limit by --proc-map-timeout\n",
481 machine->root_dir, pid, pid);
486 event->mmap2.ino_generation = 0;
489 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
491 if (machine__is_host(machine))
492 event->header.misc = PERF_RECORD_MISC_USER;
494 event->header.misc = PERF_RECORD_MISC_GUEST_USER;
496 if ((event->mmap2.prot & PROT_EXEC) == 0) {
497 if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0)
500 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
505 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
507 if (!strcmp(event->mmap2.filename, ""))
508 strcpy(event->mmap2.filename, anonstr);
510 if (hugetlbfs_mnt_len &&
511 !strncmp(event->mmap2.filename, hugetlbfs_mnt,
512 hugetlbfs_mnt_len)) {
513 strcpy(event->mmap2.filename, anonstr);
514 event->mmap2.flags |= MAP_HUGETLB;
517 size = strlen(event->mmap2.filename) + 1;
518 aligned_size = PERF_ALIGN(size, sizeof(u64));
519 event->mmap2.len -= event->mmap.start;
520 event->mmap2.header.size = (sizeof(event->mmap2) -
521 (sizeof(event->mmap2.filename) - aligned_size));
522 memset(event->mmap2.filename + size, 0, machine->id_hdr_size +
523 (aligned_size - size));
524 event->mmap2.header.size += machine->id_hdr_size;
525 event->mmap2.pid = tgid;
526 event->mmap2.tid = pid;
528 if (symbol_conf.buildid_mmap2)
529 perf_record_mmap2__read_build_id(&event->mmap2, machine, false);
531 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
544 #ifdef HAVE_FILE_HANDLE
545 static int perf_event__synthesize_cgroup(struct perf_tool *tool,
546 union perf_event *event,
547 char *path, size_t mount_len,
548 perf_event__handler_t process,
549 struct machine *machine)
551 size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path);
552 size_t path_len = strlen(path) - mount_len + 1;
554 struct file_handle fh;
559 while (path_len % sizeof(u64))
560 path[mount_len + path_len++] = '\0';
562 memset(&event->cgroup, 0, event_size);
564 event->cgroup.header.type = PERF_RECORD_CGROUP;
565 event->cgroup.header.size = event_size + path_len + machine->id_hdr_size;
567 handle.fh.handle_bytes = sizeof(handle.cgroup_id);
568 if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) {
569 pr_debug("stat failed: %s\n", path);
573 event->cgroup.id = handle.cgroup_id;
574 strncpy(event->cgroup.path, path + mount_len, path_len);
575 memset(event->cgroup.path + path_len, 0, machine->id_hdr_size);
577 if (perf_tool__process_synth_event(tool, event, machine, process) < 0) {
578 pr_debug("process synth event failed\n");
585 static int perf_event__walk_cgroup_tree(struct perf_tool *tool,
586 union perf_event *event,
587 char *path, size_t mount_len,
588 perf_event__handler_t process,
589 struct machine *machine)
591 size_t pos = strlen(path);
596 if (perf_event__synthesize_cgroup(tool, event, path, mount_len,
597 process, machine) < 0)
602 pr_debug("failed to open directory: %s\n", path);
606 while ((dent = readdir(d)) != NULL) {
607 if (dent->d_type != DT_DIR)
609 if (!strcmp(dent->d_name, ".") ||
610 !strcmp(dent->d_name, ".."))
613 /* any sane path should be less than PATH_MAX */
614 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX)
617 if (path[pos - 1] != '/')
619 strcat(path, dent->d_name);
621 ret = perf_event__walk_cgroup_tree(tool, event, path,
622 mount_len, process, machine);
633 int perf_event__synthesize_cgroups(struct perf_tool *tool,
634 perf_event__handler_t process,
635 struct machine *machine)
637 union perf_event event;
638 char cgrp_root[PATH_MAX];
639 size_t mount_len; /* length of mount point in the path */
641 if (!tool || !tool->cgroup_events)
644 if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) {
645 pr_debug("cannot find cgroup mount point\n");
649 mount_len = strlen(cgrp_root);
650 /* make sure the path starts with a slash (after mount point) */
651 strcat(cgrp_root, "/");
653 if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len,
654 process, machine) < 0)
660 int perf_event__synthesize_cgroups(struct perf_tool *tool __maybe_unused,
661 perf_event__handler_t process __maybe_unused,
662 struct machine *machine __maybe_unused)
668 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process,
669 struct machine *machine)
673 struct maps *maps = machine__kernel_maps(machine);
674 union perf_event *event;
675 size_t size = symbol_conf.buildid_mmap2 ?
676 sizeof(event->mmap2) : sizeof(event->mmap);
678 event = zalloc(size + machine->id_hdr_size);
680 pr_debug("Not enough memory synthesizing mmap event "
681 "for kernel modules\n");
686 * kernel uses 0 for user space maps, see kernel/perf_event.c
689 if (machine__is_host(machine))
690 event->header.misc = PERF_RECORD_MISC_KERNEL;
692 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
694 maps__for_each_entry(maps, pos) {
695 if (!__map__is_kmodule(pos))
698 if (symbol_conf.buildid_mmap2) {
699 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
700 event->mmap2.header.type = PERF_RECORD_MMAP2;
701 event->mmap2.header.size = (sizeof(event->mmap2) -
702 (sizeof(event->mmap2.filename) - size));
703 memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
704 event->mmap2.header.size += machine->id_hdr_size;
705 event->mmap2.start = pos->start;
706 event->mmap2.len = pos->end - pos->start;
707 event->mmap2.pid = machine->pid;
709 memcpy(event->mmap2.filename, pos->dso->long_name,
710 pos->dso->long_name_len + 1);
712 perf_record_mmap2__read_build_id(&event->mmap2, machine, false);
714 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
715 event->mmap.header.type = PERF_RECORD_MMAP;
716 event->mmap.header.size = (sizeof(event->mmap) -
717 (sizeof(event->mmap.filename) - size));
718 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
719 event->mmap.header.size += machine->id_hdr_size;
720 event->mmap.start = pos->start;
721 event->mmap.len = pos->end - pos->start;
722 event->mmap.pid = machine->pid;
724 memcpy(event->mmap.filename, pos->dso->long_name,
725 pos->dso->long_name_len + 1);
728 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
738 static int filter_task(const struct dirent *dirent)
740 return isdigit(dirent->d_name[0]);
743 static int __event__synthesize_thread(union perf_event *comm_event,
744 union perf_event *mmap_event,
745 union perf_event *fork_event,
746 union perf_event *namespaces_event,
747 pid_t pid, int full, perf_event__handler_t process,
748 struct perf_tool *tool, struct machine *machine,
749 bool needs_mmap, bool mmap_data)
751 char filename[PATH_MAX];
752 struct dirent **dirent;
757 /* special case: only send one comm event using passed in pid */
759 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
765 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
766 tgid, process, machine) < 0)
770 * send mmap only for thread group leader
771 * see thread__init_maps()
773 if (pid == tgid && needs_mmap &&
774 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
775 process, machine, mmap_data))
781 if (machine__is_default_guest(machine))
784 snprintf(filename, sizeof(filename), "%s/proc/%d/task",
785 machine->root_dir, pid);
787 n = scandir(filename, &dirent, filter_task, NULL);
791 for (i = 0; i < n; i++) {
794 bool kernel_thread = false;
796 _pid = strtol(dirent[i]->d_name, &end, 10);
800 /* some threads may exit just after scan, ignore it */
801 if (perf_event__prepare_comm(comm_event, pid, _pid, machine,
802 &tgid, &ppid, &kernel_thread) != 0)
806 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
807 ppid, process, machine) < 0)
810 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
811 tgid, process, machine) < 0)
815 * Send the prepared comm event
817 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
821 if (_pid == pid && !kernel_thread && needs_mmap) {
822 /* process the parent's maps too */
823 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
824 process, machine, mmap_data);
830 for (i = 0; i < n; i++)
837 int perf_event__synthesize_thread_map(struct perf_tool *tool,
838 struct perf_thread_map *threads,
839 perf_event__handler_t process,
840 struct machine *machine,
841 bool needs_mmap, bool mmap_data)
843 union perf_event *comm_event, *mmap_event, *fork_event;
844 union perf_event *namespaces_event;
845 int err = -1, thread, j;
847 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
848 if (comm_event == NULL)
851 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
852 if (mmap_event == NULL)
855 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
856 if (fork_event == NULL)
859 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
860 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
861 machine->id_hdr_size);
862 if (namespaces_event == NULL)
866 for (thread = 0; thread < threads->nr; ++thread) {
867 if (__event__synthesize_thread(comm_event, mmap_event,
868 fork_event, namespaces_event,
869 perf_thread_map__pid(threads, thread), 0,
870 process, tool, machine,
871 needs_mmap, mmap_data)) {
877 * comm.pid is set to thread group id by
878 * perf_event__synthesize_comm
880 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) {
881 bool need_leader = true;
883 /* is thread group leader in thread_map? */
884 for (j = 0; j < threads->nr; ++j) {
885 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) {
891 /* if not, generate events for it */
893 __event__synthesize_thread(comm_event, mmap_event,
894 fork_event, namespaces_event,
895 comm_event->comm.pid, 0,
896 process, tool, machine,
897 needs_mmap, mmap_data)) {
903 free(namespaces_event);
914 static int __perf_event__synthesize_threads(struct perf_tool *tool,
915 perf_event__handler_t process,
916 struct machine *machine,
919 struct dirent **dirent,
923 union perf_event *comm_event, *mmap_event, *fork_event;
924 union perf_event *namespaces_event;
930 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
931 if (comm_event == NULL)
934 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
935 if (mmap_event == NULL)
938 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
939 if (fork_event == NULL)
942 namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
943 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
944 machine->id_hdr_size);
945 if (namespaces_event == NULL)
948 for (i = start; i < start + num; i++) {
949 if (!isdigit(dirent[i]->d_name[0]))
952 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
953 /* only interested in proper numerical dirents */
957 * We may race with exiting thread, so don't stop just because
958 * one thread couldn't be synthesized.
960 __event__synthesize_thread(comm_event, mmap_event, fork_event,
961 namespaces_event, pid, 1, process,
962 tool, machine, needs_mmap, mmap_data);
966 free(namespaces_event);
977 struct synthesize_threads_arg {
978 struct perf_tool *tool;
979 perf_event__handler_t process;
980 struct machine *machine;
983 struct dirent **dirent;
988 static void *synthesize_threads_worker(void *arg)
990 struct synthesize_threads_arg *args = arg;
992 __perf_event__synthesize_threads(args->tool, args->process,
994 args->needs_mmap, args->mmap_data,
996 args->start, args->num);
1000 int perf_event__synthesize_threads(struct perf_tool *tool,
1001 perf_event__handler_t process,
1002 struct machine *machine,
1003 bool needs_mmap, bool mmap_data,
1004 unsigned int nr_threads_synthesize)
1006 struct synthesize_threads_arg *args = NULL;
1007 pthread_t *synthesize_threads = NULL;
1008 char proc_path[PATH_MAX];
1009 struct dirent **dirent;
1017 if (machine__is_default_guest(machine))
1020 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
1021 n = scandir(proc_path, &dirent, filter_task, NULL);
1025 if (nr_threads_synthesize == UINT_MAX)
1026 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
1028 thread_nr = nr_threads_synthesize;
1030 if (thread_nr <= 1) {
1031 err = __perf_event__synthesize_threads(tool, process,
1033 needs_mmap, mmap_data,
1040 synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
1041 if (synthesize_threads == NULL)
1044 args = calloc(sizeof(*args), thread_nr);
1048 num_per_thread = n / thread_nr;
1050 for (i = 0; i < thread_nr; i++) {
1051 args[i].tool = tool;
1052 args[i].process = process;
1053 args[i].machine = machine;
1054 args[i].needs_mmap = needs_mmap;
1055 args[i].mmap_data = mmap_data;
1056 args[i].dirent = dirent;
1058 for (i = 0; i < m; i++) {
1059 args[i].num = num_per_thread + 1;
1060 args[i].start = i * args[i].num;
1063 base = args[i-1].start + args[i-1].num;
1064 for (j = i; j < thread_nr; j++) {
1065 args[j].num = num_per_thread;
1066 args[j].start = base + (j - i) * args[i].num;
1069 for (i = 0; i < thread_nr; i++) {
1070 if (pthread_create(&synthesize_threads[i], NULL,
1071 synthesize_threads_worker, &args[i]))
1076 for (i = 0; i < thread_nr; i++)
1077 pthread_join(synthesize_threads[i], NULL);
1080 free(synthesize_threads);
1082 for (i = 0; i < n; i++)
1089 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
1090 perf_event__handler_t process __maybe_unused,
1091 struct machine *machine __maybe_unused)
1096 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
1097 perf_event__handler_t process,
1098 struct machine *machine)
1100 union perf_event *event;
1101 size_t size = symbol_conf.buildid_mmap2 ?
1102 sizeof(event->mmap2) : sizeof(event->mmap);
1103 struct map *map = machine__kernel_map(machine);
1110 kmap = map__kmap(map);
1111 if (!kmap->ref_reloc_sym)
1115 * We should get this from /sys/kernel/sections/.text, but till that is
1116 * available use this, and after it is use this as a fallback for older
1119 event = zalloc(size + machine->id_hdr_size);
1120 if (event == NULL) {
1121 pr_debug("Not enough memory synthesizing mmap event "
1122 "for kernel modules\n");
1126 if (machine__is_host(machine)) {
1128 * kernel uses PERF_RECORD_MISC_USER for user space maps,
1129 * see kernel/perf_event.c __perf_event_mmap
1131 event->header.misc = PERF_RECORD_MISC_KERNEL;
1133 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
1136 if (symbol_conf.buildid_mmap2) {
1137 size = snprintf(event->mmap2.filename, sizeof(event->mmap2.filename),
1138 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
1139 size = PERF_ALIGN(size, sizeof(u64));
1140 event->mmap2.header.type = PERF_RECORD_MMAP2;
1141 event->mmap2.header.size = (sizeof(event->mmap2) -
1142 (sizeof(event->mmap2.filename) - size) + machine->id_hdr_size);
1143 event->mmap2.pgoff = kmap->ref_reloc_sym->addr;
1144 event->mmap2.start = map->start;
1145 event->mmap2.len = map->end - event->mmap.start;
1146 event->mmap2.pid = machine->pid;
1148 perf_record_mmap2__read_build_id(&event->mmap2, machine, true);
1150 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
1151 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
1152 size = PERF_ALIGN(size, sizeof(u64));
1153 event->mmap.header.type = PERF_RECORD_MMAP;
1154 event->mmap.header.size = (sizeof(event->mmap) -
1155 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
1156 event->mmap.pgoff = kmap->ref_reloc_sym->addr;
1157 event->mmap.start = map->start;
1158 event->mmap.len = map->end - event->mmap.start;
1159 event->mmap.pid = machine->pid;
1162 err = perf_tool__process_synth_event(tool, event, machine, process);
1168 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
1169 perf_event__handler_t process,
1170 struct machine *machine)
1174 err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
1178 return perf_event__synthesize_extra_kmaps(tool, process, machine);
1181 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
1182 struct perf_thread_map *threads,
1183 perf_event__handler_t process,
1184 struct machine *machine)
1186 union perf_event *event;
1189 size = sizeof(event->thread_map);
1190 size += threads->nr * sizeof(event->thread_map.entries[0]);
1192 event = zalloc(size);
1196 event->header.type = PERF_RECORD_THREAD_MAP;
1197 event->header.size = size;
1198 event->thread_map.nr = threads->nr;
1200 for (i = 0; i < threads->nr; i++) {
1201 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i];
1202 char *comm = perf_thread_map__comm(threads, i);
1207 entry->pid = perf_thread_map__pid(threads, i);
1208 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
1211 err = process(tool, event, NULL, machine);
1217 struct synthesize_cpu_map_data {
1218 const struct perf_cpu_map *map;
1225 struct perf_record_cpu_map_data *data;
1228 static void synthesize_cpus(struct synthesize_cpu_map_data *data)
1230 data->data->type = PERF_CPU_MAP__CPUS;
1231 data->data->cpus_data.nr = data->nr;
1232 for (int i = 0; i < data->nr; i++)
1233 data->data->cpus_data.cpu[i] = perf_cpu_map__cpu(data->map, i).cpu;
1236 static void synthesize_mask(struct synthesize_cpu_map_data *data)
1239 struct perf_cpu cpu;
1241 /* Due to padding, the 4bytes per entry mask variant is always smaller. */
1242 data->data->type = PERF_CPU_MAP__MASK;
1243 data->data->mask32_data.nr = BITS_TO_U32(data->max_cpu);
1244 data->data->mask32_data.long_size = 4;
1246 perf_cpu_map__for_each_cpu(cpu, idx, data->map) {
1247 int bit_word = cpu.cpu / 32;
1248 u32 bit_mask = 1U << (cpu.cpu & 31);
1250 data->data->mask32_data.mask[bit_word] |= bit_mask;
1254 static void synthesize_range_cpus(struct synthesize_cpu_map_data *data)
1256 data->data->type = PERF_CPU_MAP__RANGE_CPUS;
1257 data->data->range_cpu_data.any_cpu = data->has_any_cpu;
1258 data->data->range_cpu_data.start_cpu = data->min_cpu;
1259 data->data->range_cpu_data.end_cpu = data->max_cpu;
1262 static void *cpu_map_data__alloc(struct synthesize_cpu_map_data *syn_data,
1265 size_t size_cpus, size_mask;
1267 syn_data->nr = perf_cpu_map__nr(syn_data->map);
1268 syn_data->has_any_cpu = (perf_cpu_map__cpu(syn_data->map, 0).cpu == -1) ? 1 : 0;
1270 syn_data->min_cpu = perf_cpu_map__cpu(syn_data->map, syn_data->has_any_cpu).cpu;
1271 syn_data->max_cpu = perf_cpu_map__max(syn_data->map).cpu;
1272 if (syn_data->max_cpu - syn_data->min_cpu + 1 == syn_data->nr - syn_data->has_any_cpu) {
1273 /* A consecutive range of CPUs can be encoded using a range. */
1274 assert(sizeof(u16) + sizeof(struct perf_record_range_cpu_map) == sizeof(u64));
1275 syn_data->type = PERF_CPU_MAP__RANGE_CPUS;
1276 syn_data->size = header_size + sizeof(u64);
1277 return zalloc(syn_data->size);
1280 size_cpus = sizeof(u16) + sizeof(struct cpu_map_entries) + syn_data->nr * sizeof(u16);
1281 /* Due to padding, the 4bytes per entry mask variant is always smaller. */
1282 size_mask = sizeof(u16) + sizeof(struct perf_record_mask_cpu_map32) +
1283 BITS_TO_U32(syn_data->max_cpu) * sizeof(__u32);
1284 if (syn_data->has_any_cpu || size_cpus < size_mask) {
1285 /* Follow the CPU map encoding. */
1286 syn_data->type = PERF_CPU_MAP__CPUS;
1287 syn_data->size = header_size + PERF_ALIGN(size_cpus, sizeof(u64));
1288 return zalloc(syn_data->size);
1290 /* Encode using a bitmask. */
1291 syn_data->type = PERF_CPU_MAP__MASK;
1292 syn_data->size = header_size + PERF_ALIGN(size_mask, sizeof(u64));
1293 return zalloc(syn_data->size);
1296 static void cpu_map_data__synthesize(struct synthesize_cpu_map_data *data)
1298 switch (data->type) {
1299 case PERF_CPU_MAP__CPUS:
1300 synthesize_cpus(data);
1302 case PERF_CPU_MAP__MASK:
1303 synthesize_mask(data);
1305 case PERF_CPU_MAP__RANGE_CPUS:
1306 synthesize_range_cpus(data);
1313 static struct perf_record_cpu_map *cpu_map_event__new(const struct perf_cpu_map *map)
1315 struct synthesize_cpu_map_data syn_data = { .map = map };
1316 struct perf_record_cpu_map *event;
1319 event = cpu_map_data__alloc(&syn_data, sizeof(struct perf_event_header));
1323 syn_data.data = &event->data;
1324 event->header.type = PERF_RECORD_CPU_MAP;
1325 event->header.size = syn_data.size;
1326 cpu_map_data__synthesize(&syn_data);
1331 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1332 const struct perf_cpu_map *map,
1333 perf_event__handler_t process,
1334 struct machine *machine)
1336 struct perf_record_cpu_map *event;
1339 event = cpu_map_event__new(map);
1343 err = process(tool, (union perf_event *) event, NULL, machine);
1349 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1350 struct perf_stat_config *config,
1351 perf_event__handler_t process,
1352 struct machine *machine)
1354 struct perf_record_stat_config *event;
1355 int size, i = 0, err;
1357 size = sizeof(*event);
1358 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1360 event = zalloc(size);
1364 event->header.type = PERF_RECORD_STAT_CONFIG;
1365 event->header.size = size;
1366 event->nr = PERF_STAT_CONFIG_TERM__MAX;
1368 #define ADD(__term, __val) \
1369 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \
1370 event->data[i].val = __val; \
1373 ADD(AGGR_MODE, config->aggr_mode)
1374 ADD(INTERVAL, config->interval)
1375 ADD(SCALE, config->scale)
1377 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1378 "stat config terms unbalanced\n");
1381 err = process(tool, (union perf_event *) event, NULL, machine);
1387 int perf_event__synthesize_stat(struct perf_tool *tool,
1388 struct perf_cpu cpu, u32 thread, u64 id,
1389 struct perf_counts_values *count,
1390 perf_event__handler_t process,
1391 struct machine *machine)
1393 struct perf_record_stat event;
1395 event.header.type = PERF_RECORD_STAT;
1396 event.header.size = sizeof(event);
1397 event.header.misc = 0;
1400 event.cpu = cpu.cpu;
1401 event.thread = thread;
1402 event.val = count->val;
1403 event.ena = count->ena;
1404 event.run = count->run;
1406 return process(tool, (union perf_event *) &event, NULL, machine);
1409 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1410 u64 evtime, u64 type,
1411 perf_event__handler_t process,
1412 struct machine *machine)
1414 struct perf_record_stat_round event;
1416 event.header.type = PERF_RECORD_STAT_ROUND;
1417 event.header.size = sizeof(event);
1418 event.header.misc = 0;
1420 event.time = evtime;
1423 return process(tool, (union perf_event *) &event, NULL, machine);
1426 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format)
1428 size_t sz, result = sizeof(struct perf_record_sample);
1430 if (type & PERF_SAMPLE_IDENTIFIER)
1431 result += sizeof(u64);
1433 if (type & PERF_SAMPLE_IP)
1434 result += sizeof(u64);
1436 if (type & PERF_SAMPLE_TID)
1437 result += sizeof(u64);
1439 if (type & PERF_SAMPLE_TIME)
1440 result += sizeof(u64);
1442 if (type & PERF_SAMPLE_ADDR)
1443 result += sizeof(u64);
1445 if (type & PERF_SAMPLE_ID)
1446 result += sizeof(u64);
1448 if (type & PERF_SAMPLE_STREAM_ID)
1449 result += sizeof(u64);
1451 if (type & PERF_SAMPLE_CPU)
1452 result += sizeof(u64);
1454 if (type & PERF_SAMPLE_PERIOD)
1455 result += sizeof(u64);
1457 if (type & PERF_SAMPLE_READ) {
1458 result += sizeof(u64);
1459 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1460 result += sizeof(u64);
1461 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1462 result += sizeof(u64);
1463 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1464 if (read_format & PERF_FORMAT_GROUP) {
1465 sz = sample_read_value_size(read_format);
1466 result += sz * sample->read.group.nr;
1468 result += sizeof(u64);
1469 if (read_format & PERF_FORMAT_LOST)
1470 result += sizeof(u64);
1474 if (type & PERF_SAMPLE_CALLCHAIN) {
1475 sz = (sample->callchain->nr + 1) * sizeof(u64);
1479 if (type & PERF_SAMPLE_RAW) {
1480 result += sizeof(u32);
1481 result += sample->raw_size;
1484 if (type & PERF_SAMPLE_BRANCH_STACK) {
1485 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1487 sz += 2 * sizeof(u64);
1491 if (type & PERF_SAMPLE_REGS_USER) {
1492 if (sample->user_regs.abi) {
1493 result += sizeof(u64);
1494 sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1497 result += sizeof(u64);
1501 if (type & PERF_SAMPLE_STACK_USER) {
1502 sz = sample->user_stack.size;
1503 result += sizeof(u64);
1506 result += sizeof(u64);
1510 if (type & PERF_SAMPLE_WEIGHT_TYPE)
1511 result += sizeof(u64);
1513 if (type & PERF_SAMPLE_DATA_SRC)
1514 result += sizeof(u64);
1516 if (type & PERF_SAMPLE_TRANSACTION)
1517 result += sizeof(u64);
1519 if (type & PERF_SAMPLE_REGS_INTR) {
1520 if (sample->intr_regs.abi) {
1521 result += sizeof(u64);
1522 sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1525 result += sizeof(u64);
1529 if (type & PERF_SAMPLE_PHYS_ADDR)
1530 result += sizeof(u64);
1532 if (type & PERF_SAMPLE_CGROUP)
1533 result += sizeof(u64);
1535 if (type & PERF_SAMPLE_DATA_PAGE_SIZE)
1536 result += sizeof(u64);
1538 if (type & PERF_SAMPLE_CODE_PAGE_SIZE)
1539 result += sizeof(u64);
1541 if (type & PERF_SAMPLE_AUX) {
1542 result += sizeof(u64);
1543 result += sample->aux_sample.size;
1549 void __weak arch_perf_synthesize_sample_weight(const struct perf_sample *data,
1550 __u64 *array, u64 type __maybe_unused)
1552 *array = data->weight;
1555 static __u64 *copy_read_group_values(__u64 *array, __u64 read_format,
1556 const struct perf_sample *sample)
1558 size_t sz = sample_read_value_size(read_format);
1559 struct sample_read_value *v = sample->read.group.values;
1561 sample_read_group__for_each(v, sample->read.group.nr, read_format) {
1562 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1563 memcpy(array, v, sz);
1564 array = (void *)array + sz;
1569 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format,
1570 const struct perf_sample *sample)
1575 * used for cross-endian analysis. See git commit 65014ab3
1576 * for why this goofiness is needed.
1580 array = event->sample.array;
1582 if (type & PERF_SAMPLE_IDENTIFIER) {
1583 *array = sample->id;
1587 if (type & PERF_SAMPLE_IP) {
1588 *array = sample->ip;
1592 if (type & PERF_SAMPLE_TID) {
1593 u.val32[0] = sample->pid;
1594 u.val32[1] = sample->tid;
1599 if (type & PERF_SAMPLE_TIME) {
1600 *array = sample->time;
1604 if (type & PERF_SAMPLE_ADDR) {
1605 *array = sample->addr;
1609 if (type & PERF_SAMPLE_ID) {
1610 *array = sample->id;
1614 if (type & PERF_SAMPLE_STREAM_ID) {
1615 *array = sample->stream_id;
1619 if (type & PERF_SAMPLE_CPU) {
1620 u.val32[0] = sample->cpu;
1626 if (type & PERF_SAMPLE_PERIOD) {
1627 *array = sample->period;
1631 if (type & PERF_SAMPLE_READ) {
1632 if (read_format & PERF_FORMAT_GROUP)
1633 *array = sample->read.group.nr;
1635 *array = sample->read.one.value;
1638 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1639 *array = sample->read.time_enabled;
1643 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1644 *array = sample->read.time_running;
1648 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1649 if (read_format & PERF_FORMAT_GROUP) {
1650 array = copy_read_group_values(array, read_format,
1653 *array = sample->read.one.id;
1656 if (read_format & PERF_FORMAT_LOST) {
1657 *array = sample->read.one.lost;
1663 if (type & PERF_SAMPLE_CALLCHAIN) {
1664 sz = (sample->callchain->nr + 1) * sizeof(u64);
1665 memcpy(array, sample->callchain, sz);
1666 array = (void *)array + sz;
1669 if (type & PERF_SAMPLE_RAW) {
1670 u.val32[0] = sample->raw_size;
1672 array = (void *)array + sizeof(u32);
1674 memcpy(array, sample->raw_data, sample->raw_size);
1675 array = (void *)array + sample->raw_size;
1678 if (type & PERF_SAMPLE_BRANCH_STACK) {
1679 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1681 sz += 2 * sizeof(u64);
1682 memcpy(array, sample->branch_stack, sz);
1683 array = (void *)array + sz;
1686 if (type & PERF_SAMPLE_REGS_USER) {
1687 if (sample->user_regs.abi) {
1688 *array++ = sample->user_regs.abi;
1689 sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1690 memcpy(array, sample->user_regs.regs, sz);
1691 array = (void *)array + sz;
1697 if (type & PERF_SAMPLE_STACK_USER) {
1698 sz = sample->user_stack.size;
1701 memcpy(array, sample->user_stack.data, sz);
1702 array = (void *)array + sz;
1707 if (type & PERF_SAMPLE_WEIGHT_TYPE) {
1708 arch_perf_synthesize_sample_weight(sample, array, type);
1712 if (type & PERF_SAMPLE_DATA_SRC) {
1713 *array = sample->data_src;
1717 if (type & PERF_SAMPLE_TRANSACTION) {
1718 *array = sample->transaction;
1722 if (type & PERF_SAMPLE_REGS_INTR) {
1723 if (sample->intr_regs.abi) {
1724 *array++ = sample->intr_regs.abi;
1725 sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1726 memcpy(array, sample->intr_regs.regs, sz);
1727 array = (void *)array + sz;
1733 if (type & PERF_SAMPLE_PHYS_ADDR) {
1734 *array = sample->phys_addr;
1738 if (type & PERF_SAMPLE_CGROUP) {
1739 *array = sample->cgroup;
1743 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
1744 *array = sample->data_page_size;
1748 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
1749 *array = sample->code_page_size;
1753 if (type & PERF_SAMPLE_AUX) {
1754 sz = sample->aux_sample.size;
1756 memcpy(array, sample->aux_sample.data, sz);
1757 array = (void *)array + sz;
1763 int perf_event__synthesize_id_sample(__u64 *array, u64 type, const struct perf_sample *sample)
1765 __u64 *start = array;
1768 * used for cross-endian analysis. See git commit 65014ab3
1769 * for why this goofiness is needed.
1773 if (type & PERF_SAMPLE_TID) {
1774 u.val32[0] = sample->pid;
1775 u.val32[1] = sample->tid;
1780 if (type & PERF_SAMPLE_TIME) {
1781 *array = sample->time;
1785 if (type & PERF_SAMPLE_ID) {
1786 *array = sample->id;
1790 if (type & PERF_SAMPLE_STREAM_ID) {
1791 *array = sample->stream_id;
1795 if (type & PERF_SAMPLE_CPU) {
1796 u.val32[0] = sample->cpu;
1802 if (type & PERF_SAMPLE_IDENTIFIER) {
1803 *array = sample->id;
1807 return (void *)array - (void *)start;
1810 int __perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process,
1811 struct evlist *evlist, struct machine *machine, size_t from)
1813 union perf_event *ev;
1814 struct evsel *evsel;
1815 size_t nr = 0, i = 0, sz, max_nr, n, pos;
1816 size_t e1_sz = sizeof(struct id_index_entry);
1817 size_t e2_sz = sizeof(struct id_index_entry_2);
1818 size_t etot_sz = e1_sz + e2_sz;
1819 bool e2_needed = false;
1822 max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / etot_sz;
1825 evlist__for_each_entry(evlist, evsel) {
1828 nr += evsel->core.ids;
1834 pr_debug2("Synthesizing id index\n");
1836 n = nr > max_nr ? max_nr : nr;
1837 sz = sizeof(struct perf_record_id_index) + n * etot_sz;
1842 sz = sizeof(struct perf_record_id_index) + n * e1_sz;
1844 ev->id_index.header.type = PERF_RECORD_ID_INDEX;
1845 ev->id_index.nr = n;
1848 evlist__for_each_entry(evlist, evsel) {
1853 for (j = 0; j < evsel->core.ids; j++, i++) {
1854 struct id_index_entry *e;
1855 struct id_index_entry_2 *e2;
1856 struct perf_sample_id *sid;
1859 ev->id_index.header.size = sz + (e2_needed ? n * e2_sz : 0);
1860 err = process(tool, ev, NULL, machine);
1868 e = &ev->id_index.entries[i];
1870 e->id = evsel->core.id[j];
1872 sid = evlist__id2sid(evlist, e->id);
1879 e->cpu = sid->cpu.cpu;
1882 if (sid->machine_pid)
1885 e2 = (void *)ev + sz;
1886 e2[i].machine_pid = sid->machine_pid;
1887 e2[i].vcpu = sid->vcpu.cpu;
1891 sz = sizeof(struct perf_record_id_index) + nr * e1_sz;
1892 ev->id_index.header.size = sz + (e2_needed ? nr * e2_sz : 0);
1893 ev->id_index.nr = nr;
1895 err = process(tool, ev, NULL, machine);
1902 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process,
1903 struct evlist *evlist, struct machine *machine)
1905 return __perf_event__synthesize_id_index(tool, process, evlist, machine, 0);
1908 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1909 struct target *target, struct perf_thread_map *threads,
1910 perf_event__handler_t process, bool needs_mmap,
1911 bool data_mmap, unsigned int nr_threads_synthesize)
1914 * When perf runs in non-root PID namespace, and the namespace's proc FS
1915 * is not mounted, nsinfo__is_in_root_namespace() returns false.
1916 * In this case, the proc FS is coming for the parent namespace, thus
1917 * perf tool will wrongly gather process info from its parent PID
1920 * To avoid the confusion that the perf tool runs in a child PID
1921 * namespace but it synthesizes thread info from its parent PID
1922 * namespace, returns failure with warning.
1924 if (!nsinfo__is_in_root_namespace()) {
1925 pr_err("Perf runs in non-root PID namespace but it tries to ");
1926 pr_err("gather process info from its parent PID namespace.\n");
1927 pr_err("Please mount the proc file system properly, e.g. ");
1928 pr_err("add the option '--mount-proc' for unshare command.\n");
1932 if (target__has_task(target))
1933 return perf_event__synthesize_thread_map(tool, threads, process, machine,
1934 needs_mmap, data_mmap);
1935 else if (target__has_cpu(target))
1936 return perf_event__synthesize_threads(tool, process, machine,
1937 needs_mmap, data_mmap,
1938 nr_threads_synthesize);
1939 /* command specified */
1943 int machine__synthesize_threads(struct machine *machine, struct target *target,
1944 struct perf_thread_map *threads, bool needs_mmap,
1945 bool data_mmap, unsigned int nr_threads_synthesize)
1947 return __machine__synthesize_threads(machine, NULL, target, threads,
1948 perf_event__process, needs_mmap,
1949 data_mmap, nr_threads_synthesize);
1952 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id)
1954 struct perf_record_event_update *ev;
1956 size += sizeof(*ev);
1957 size = PERF_ALIGN(size, sizeof(u64));
1961 ev->header.type = PERF_RECORD_EVENT_UPDATE;
1962 ev->header.size = (u16)size;
1969 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel,
1970 perf_event__handler_t process)
1972 size_t size = strlen(evsel->unit);
1973 struct perf_record_event_update *ev;
1976 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]);
1980 strlcpy(ev->unit, evsel->unit, size + 1);
1981 err = process(tool, (union perf_event *)ev, NULL, NULL);
1986 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel,
1987 perf_event__handler_t process)
1989 struct perf_record_event_update *ev;
1990 struct perf_record_event_update_scale *ev_data;
1993 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]);
1997 ev->scale.scale = evsel->scale;
1998 err = process(tool, (union perf_event *)ev, NULL, NULL);
2003 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel,
2004 perf_event__handler_t process)
2006 struct perf_record_event_update *ev;
2007 size_t len = strlen(evsel->name);
2010 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]);
2014 strlcpy(ev->name, evsel->name, len + 1);
2015 err = process(tool, (union perf_event *)ev, NULL, NULL);
2020 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel,
2021 perf_event__handler_t process)
2023 struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };
2024 struct perf_record_event_update *ev;
2027 ev = cpu_map_data__alloc(&syn_data, sizeof(struct perf_event_header) + 2 * sizeof(u64));
2031 syn_data.data = &ev->cpus.cpus;
2032 ev->header.type = PERF_RECORD_EVENT_UPDATE;
2033 ev->header.size = (u16)syn_data.size;
2034 ev->type = PERF_EVENT_UPDATE__CPUS;
2035 ev->id = evsel->core.id[0];
2036 cpu_map_data__synthesize(&syn_data);
2038 err = process(tool, (union perf_event *)ev, NULL, NULL);
2043 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist,
2044 perf_event__handler_t process)
2046 struct evsel *evsel;
2049 evlist__for_each_entry(evlist, evsel) {
2050 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids,
2051 evsel->core.id, process);
2053 pr_debug("failed to create perf header attribute\n");
2061 static bool has_unit(struct evsel *evsel)
2063 return evsel->unit && *evsel->unit;
2066 static bool has_scale(struct evsel *evsel)
2068 return evsel->scale != 1;
2071 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list,
2072 perf_event__handler_t process, bool is_pipe)
2074 struct evsel *evsel;
2078 * Synthesize other events stuff not carried within
2079 * attr event - unit, scale, name
2081 evlist__for_each_entry(evsel_list, evsel) {
2082 if (!evsel->supported)
2086 * Synthesize unit and scale only if it's defined.
2088 if (has_unit(evsel)) {
2089 err = perf_event__synthesize_event_update_unit(tool, evsel, process);
2091 pr_err("Couldn't synthesize evsel unit.\n");
2096 if (has_scale(evsel)) {
2097 err = perf_event__synthesize_event_update_scale(tool, evsel, process);
2099 pr_err("Couldn't synthesize evsel evsel.\n");
2104 if (evsel->core.own_cpus) {
2105 err = perf_event__synthesize_event_update_cpus(tool, evsel, process);
2107 pr_err("Couldn't synthesize evsel cpus.\n");
2113 * Name is needed only for pipe output,
2114 * perf.data carries event names.
2117 err = perf_event__synthesize_event_update_name(tool, evsel, process);
2119 pr_err("Couldn't synthesize evsel name.\n");
2127 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr,
2128 u32 ids, u64 *id, perf_event__handler_t process)
2130 union perf_event *ev;
2134 size = sizeof(struct perf_event_attr);
2135 size = PERF_ALIGN(size, sizeof(u64));
2136 size += sizeof(struct perf_event_header);
2137 size += ids * sizeof(u64);
2144 ev->attr.attr = *attr;
2145 memcpy(ev->attr.id, id, ids * sizeof(u64));
2147 ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2148 ev->attr.header.size = (u16)size;
2150 if (ev->attr.header.size == size)
2151 err = process(tool, ev, NULL, NULL);
2160 #ifdef HAVE_LIBTRACEEVENT
2161 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist,
2162 perf_event__handler_t process)
2164 union perf_event ev;
2165 struct tracing_data *tdata;
2166 ssize_t size = 0, aligned_size = 0, padding;
2170 * We are going to store the size of the data followed
2171 * by the data contents. Since the fd descriptor is a pipe,
2172 * we cannot seek back to store the size of the data once
2173 * we know it. Instead we:
2175 * - write the tracing data to the temp file
2176 * - get/write the data size to pipe
2177 * - write the tracing data from the temp file
2180 tdata = tracing_data_get(&evlist->core.entries, fd, true);
2184 memset(&ev, 0, sizeof(ev));
2186 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2188 aligned_size = PERF_ALIGN(size, sizeof(u64));
2189 padding = aligned_size - size;
2190 ev.tracing_data.header.size = sizeof(ev.tracing_data);
2191 ev.tracing_data.size = aligned_size;
2193 process(tool, &ev, NULL, NULL);
2196 * The put function will copy all the tracing data
2197 * stored in temp file to the pipe.
2199 tracing_data_put(tdata);
2201 ff = (struct feat_fd){ .fd = fd };
2202 if (write_padded(&ff, NULL, 0, padding))
2205 return aligned_size;
2209 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc,
2210 perf_event__handler_t process, struct machine *machine)
2212 union perf_event ev;
2218 memset(&ev, 0, sizeof(ev));
2220 len = pos->long_name_len + 1;
2221 len = PERF_ALIGN(len, NAME_ALIGN);
2222 memcpy(&ev.build_id.build_id, pos->bid.data, sizeof(pos->bid.data));
2223 ev.build_id.size = pos->bid.size;
2224 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2225 ev.build_id.header.misc = misc | PERF_RECORD_MISC_BUILD_ID_SIZE;
2226 ev.build_id.pid = machine->pid;
2227 ev.build_id.header.size = sizeof(ev.build_id) + len;
2228 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2230 return process(tool, &ev, NULL, machine);
2233 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool,
2234 struct evlist *evlist, perf_event__handler_t process, bool attrs)
2239 err = perf_event__synthesize_attrs(tool, evlist, process);
2241 pr_err("Couldn't synthesize attrs.\n");
2246 err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs);
2247 err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL);
2249 pr_err("Couldn't synthesize thread map.\n");
2253 err = perf_event__synthesize_cpu_map(tool, evlist->core.user_requested_cpus, process, NULL);
2255 pr_err("Couldn't synthesize thread map.\n");
2259 err = perf_event__synthesize_stat_config(tool, config, process, NULL);
2261 pr_err("Couldn't synthesize config.\n");
2268 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
2270 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session,
2271 struct evlist *evlist, perf_event__handler_t process)
2273 struct perf_header *header = &session->header;
2274 struct perf_record_header_feature *fe;
2279 sz_hdr = sizeof(fe->header);
2280 sz = sizeof(union perf_event);
2281 /* get a nice alignment */
2282 sz = PERF_ALIGN(sz, page_size);
2284 memset(&ff, 0, sizeof(ff));
2286 ff.buf = malloc(sz);
2290 ff.size = sz - sz_hdr;
2291 ff.ph = &session->header;
2293 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2294 if (!feat_ops[feat].synthesize) {
2295 pr_debug("No record header feature for header :%d\n", feat);
2299 ff.offset = sizeof(*fe);
2301 ret = feat_ops[feat].write(&ff, evlist);
2302 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
2303 pr_debug("Error writing feature\n");
2306 /* ff.buf may have changed due to realloc in do_write() */
2308 memset(fe, 0, sizeof(*fe));
2311 fe->header.type = PERF_RECORD_HEADER_FEATURE;
2312 fe->header.size = ff.offset;
2314 ret = process(tool, ff.buf, NULL, NULL);
2321 /* Send HEADER_LAST_FEATURE mark. */
2323 fe->feat_id = HEADER_LAST_FEATURE;
2324 fe->header.type = PERF_RECORD_HEADER_FEATURE;
2325 fe->header.size = sizeof(*fe);
2327 ret = process(tool, ff.buf, NULL, NULL);
2333 int perf_event__synthesize_for_pipe(struct perf_tool *tool,
2334 struct perf_session *session,
2335 struct perf_data *data,
2336 perf_event__handler_t process)
2340 struct evlist *evlist = session->evlist;
2343 * We need to synthesize events first, because some
2344 * features works on top of them (on report side).
2346 err = perf_event__synthesize_attrs(tool, evlist, process);
2348 pr_err("Couldn't synthesize attrs.\n");
2353 err = perf_event__synthesize_features(tool, session, evlist, process);
2355 pr_err("Couldn't synthesize features.\n");
2360 #ifdef HAVE_LIBTRACEEVENT
2361 if (have_tracepoints(&evlist->core.entries)) {
2362 int fd = perf_data__fd(data);
2365 * FIXME err <= 0 here actually means that
2366 * there were no tracepoints so its not really
2367 * an error, just that we don't need to
2368 * synthesize anything. We really have to
2369 * return this more properly and also
2370 * propagate errors that now are calling die()
2372 err = perf_event__synthesize_tracing_data(tool, fd, evlist,
2375 pr_err("Couldn't record tracing data.\n");
2387 int parse_synth_opt(char *synth)
2395 for (q = synth; (p = strsep(&q, ",")); p = q) {
2396 if (!strcasecmp(p, "no") || !strcasecmp(p, "none"))
2399 if (!strcasecmp(p, "all"))
2400 return PERF_SYNTH_ALL;
2402 if (!strcasecmp(p, "task"))
2403 ret |= PERF_SYNTH_TASK;
2404 else if (!strcasecmp(p, "mmap"))
2405 ret |= PERF_SYNTH_TASK | PERF_SYNTH_MMAP;
2406 else if (!strcasecmp(p, "cgroup"))
2407 ret |= PERF_SYNTH_CGROUP;