perf lock contention: Update default map size to 16384
[platform/kernel/linux-starfive.git] / tools / perf / util / synthetic-events.c
1 // SPDX-License-Identifier: GPL-2.0-only 
2
3 #include "util/cgroup.h"
4 #include "util/data.h"
5 #include "util/debug.h"
6 #include "util/dso.h"
7 #include "util/event.h"
8 #include "util/evlist.h"
9 #include "util/machine.h"
10 #include "util/map.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>
26 #include <asm/bug.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>
33 #include <dirent.h>
34 #include <errno.h>
35 #include <inttypes.h>
36 #include <stdio.h>
37 #include <string.h>
38 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */
39 #include <api/fs/fs.h>
40 #include <api/io.h>
41 #include <sys/types.h>
42 #include <sys/stat.h>
43 #include <fcntl.h>
44 #include <unistd.h>
45
46 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500
47
48 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT;
49
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)
54 {
55         struct perf_sample synth_sample = {
56                 .pid       = -1,
57                 .tid       = -1,
58                 .time      = -1,
59                 .stream_id = -1,
60                 .cpu       = -1,
61                 .period    = 1,
62                 .cpumode   = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK,
63         };
64
65         return process(tool, event, &synth_sample, machine);
66 };
67
68 /*
69  * Assumes that the first 4095 bytes of /proc/pid/stat contains
70  * the comm, tgid and ppid.
71  */
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)
74 {
75         char bf[4096];
76         int fd;
77         size_t size = 0;
78         ssize_t n;
79         char *name, *tgids, *ppids, *vmpeak, *threads;
80
81         *tgid = -1;
82         *ppid = -1;
83
84         if (pid)
85                 snprintf(bf, sizeof(bf), "/proc/%d/task/%d/status", pid, tid);
86         else
87                 snprintf(bf, sizeof(bf), "/proc/%d/status", tid);
88
89         fd = open(bf, O_RDONLY);
90         if (fd < 0) {
91                 pr_debug("couldn't open %s\n", bf);
92                 return -1;
93         }
94
95         n = read(fd, bf, sizeof(bf) - 1);
96         close(fd);
97         if (n <= 0) {
98                 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
99                            tid);
100                 return -1;
101         }
102         bf[n] = '\0';
103
104         name = strstr(bf, "Name:");
105         tgids = strstr(name ?: bf, "Tgid:");
106         ppids = strstr(tgids ?: bf, "PPid:");
107         vmpeak = strstr(ppids ?: bf, "VmPeak:");
108
109         if (vmpeak)
110                 threads = NULL;
111         else
112                 threads = strstr(ppids ?: bf, "Threads:");
113
114         if (name) {
115                 char *nl;
116
117                 name = skip_spaces(name + 5);  /* strlen("Name:") */
118                 nl = strchr(name, '\n');
119                 if (nl)
120                         *nl = '\0';
121
122                 size = strlen(name);
123                 if (size >= len)
124                         size = len - 1;
125                 memcpy(comm, name, size);
126                 comm[size] = '\0';
127         } else {
128                 pr_debug("Name: string not found for pid %d\n", tid);
129         }
130
131         if (tgids) {
132                 tgids += 5;  /* strlen("Tgid:") */
133                 *tgid = atoi(tgids);
134         } else {
135                 pr_debug("Tgid: string not found for pid %d\n", tid);
136         }
137
138         if (ppids) {
139                 ppids += 5;  /* strlen("PPid:") */
140                 *ppid = atoi(ppids);
141         } else {
142                 pr_debug("PPid: string not found for pid %d\n", tid);
143         }
144
145         if (!vmpeak && threads)
146                 *kernel = true;
147         else
148                 *kernel = false;
149
150         return 0;
151 }
152
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)
156 {
157         size_t size;
158
159         *ppid = -1;
160
161         memset(&event->comm, 0, sizeof(event->comm));
162
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) {
167                         return -1;
168                 }
169         } else {
170                 *tgid = machine->pid;
171         }
172
173         if (*tgid < 0)
174                 return -1;
175
176         event->comm.pid = *tgid;
177         event->comm.header.type = PERF_RECORD_COMM;
178
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;
186
187         return 0;
188 }
189
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)
194 {
195         pid_t tgid, ppid;
196         bool kernel_thread;
197
198         if (perf_event__prepare_comm(event, 0, pid, machine, &tgid, &ppid,
199                                      &kernel_thread) != 0)
200                 return -1;
201
202         if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
203                 return -1;
204
205         return tgid;
206 }
207
208 static void perf_event__get_ns_link_info(pid_t pid, const char *ns,
209                                          struct perf_ns_link_info *ns_link_info)
210 {
211         struct stat64 st;
212         char proc_ns[128];
213
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;
218         }
219 }
220
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)
226 {
227         u32 idx;
228         struct perf_ns_link_info *ns_link_info;
229
230         if (!tool || !tool->namespace_events)
231                 return 0;
232
233         memset(&event->namespaces, 0, (sizeof(event->namespaces) +
234                (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
235                machine->id_hdr_size));
236
237         event->namespaces.pid = tgid;
238         event->namespaces.tid = pid;
239
240         event->namespaces.nr_namespaces = NR_NAMESPACES;
241
242         ns_link_info = event->namespaces.link_info;
243
244         for (idx = 0; idx < event->namespaces.nr_namespaces; idx++)
245                 perf_event__get_ns_link_info(pid, perf_ns__name(idx),
246                                              &ns_link_info[idx]);
247
248         event->namespaces.header.type = PERF_RECORD_NAMESPACES;
249
250         event->namespaces.header.size = (sizeof(event->namespaces) +
251                         (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
252                         machine->id_hdr_size);
253
254         if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
255                 return -1;
256
257         return 0;
258 }
259
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)
265 {
266         memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
267
268         /*
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
272         */
273         if (tgid == pid) {
274                 event->fork.ppid = ppid;
275                 event->fork.ptid = ppid;
276         } else {
277                 event->fork.ppid = tgid;
278                 event->fork.ptid = tgid;
279         }
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;
284
285         event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
286
287         if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
288                 return -1;
289
290         return 0;
291 }
292
293 static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end,
294                                 u32 *prot, u32 *flags, __u64 *offset,
295                                 u32 *maj, u32 *min,
296                                 __u64 *inode,
297                                 ssize_t pathname_size, char *pathname)
298 {
299         __u64 temp;
300         int ch;
301         char *start_pathname = pathname;
302
303         if (io__get_hex(io, start) != '-')
304                 return false;
305         if (io__get_hex(io, end) != ' ')
306                 return false;
307
308         /* map protection and flags bits */
309         *prot = 0;
310         ch = io__get_char(io);
311         if (ch == 'r')
312                 *prot |= PROT_READ;
313         else if (ch != '-')
314                 return false;
315         ch = io__get_char(io);
316         if (ch == 'w')
317                 *prot |= PROT_WRITE;
318         else if (ch != '-')
319                 return false;
320         ch = io__get_char(io);
321         if (ch == 'x')
322                 *prot |= PROT_EXEC;
323         else if (ch != '-')
324                 return false;
325         ch = io__get_char(io);
326         if (ch == 's')
327                 *flags = MAP_SHARED;
328         else if (ch == 'p')
329                 *flags = MAP_PRIVATE;
330         else
331                 return false;
332         if (io__get_char(io) != ' ')
333                 return false;
334
335         if (io__get_hex(io, offset) != ' ')
336                 return false;
337
338         if (io__get_hex(io, &temp) != ':')
339                 return false;
340         *maj = temp;
341         if (io__get_hex(io, &temp) != ' ')
342                 return false;
343         *min = temp;
344
345         ch = io__get_dec(io, inode);
346         if (ch != ' ') {
347                 *pathname = '\0';
348                 return ch == '\n';
349         }
350         do {
351                 ch = io__get_char(io);
352         } while (ch == ' ');
353         while (true) {
354                 if (ch < 0)
355                         return false;
356                 if (ch == '\0' || ch == '\n' ||
357                     (pathname + 1 - start_pathname) >= pathname_size) {
358                         *pathname = '\0';
359                         return true;
360                 }
361                 *pathname++ = ch;
362                 ch = io__get_char(io);
363         }
364 }
365
366 static void perf_record_mmap2__read_build_id(struct perf_record_mmap2 *event,
367                                              struct machine *machine,
368                                              bool is_kernel)
369 {
370         struct build_id bid;
371         struct nsinfo *nsi;
372         struct nscookie nc;
373         struct dso *dso = NULL;
374         struct dso_id id;
375         int rc;
376
377         if (is_kernel) {
378                 rc = sysfs__read_build_id("/sys/kernel/notes", &bid);
379                 goto out;
380         }
381
382         id.maj = event->maj;
383         id.min = event->min;
384         id.ino = event->ino;
385         id.ino_generation = event->ino_generation;
386
387         dso = dsos__findnew_id(&machine->dsos, event->filename, &id);
388         if (dso && dso->has_build_id) {
389                 bid = dso->bid;
390                 rc = 0;
391                 goto out;
392         }
393
394         nsi = nsinfo__new(event->pid);
395         nsinfo__mountns_enter(nsi, &nc);
396
397         rc = filename__read_build_id(event->filename, &bid) > 0 ? 0 : -1;
398
399         nsinfo__mountns_exit(&nc);
400         nsinfo__put(nsi);
401
402 out:
403         if (rc == 0) {
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;
409
410                 if (dso && !dso->has_build_id)
411                         dso__set_build_id(dso, &bid);
412         } else {
413                 if (event->filename[0] == '/') {
414                         pr_debug2("Failed to read build ID for %s\n",
415                                   event->filename);
416                 }
417         }
418         dso__put(dso);
419 }
420
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,
426                                        bool mmap_data)
427 {
428         unsigned long long t;
429         char bf[BUFSIZ];
430         struct io io;
431         bool truncation = false;
432         unsigned long long timeout = proc_map_timeout * 1000000ULL;
433         int rc = 0;
434         const char *hugetlbfs_mnt = hugetlbfs__mountpoint();
435         int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0;
436
437         if (machine__is_default_guest(machine))
438                 return 0;
439
440         snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps",
441                 machine->root_dir, pid, pid);
442
443         io.fd = open(bf, O_RDONLY, 0);
444         if (io.fd < 0) {
445                 /*
446                  * We raced with a task exiting - just return:
447                  */
448                 pr_debug("couldn't open %s\n", bf);
449                 return -1;
450         }
451         io__init(&io, io.fd, bf, sizeof(bf));
452
453         event->header.type = PERF_RECORD_MMAP2;
454         t = rdclock();
455
456         while (!io.eof) {
457                 static const char anonstr[] = "//anon";
458                 size_t size, aligned_size;
459
460                 /* ensure null termination since stack will be reused. */
461                 event->mmap2.filename[0] = '\0';
462
463                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
464                 if (!read_proc_maps_line(&io,
465                                         &event->mmap2.start,
466                                         &event->mmap2.len,
467                                         &event->mmap2.prot,
468                                         &event->mmap2.flags,
469                                         &event->mmap2.pgoff,
470                                         &event->mmap2.maj,
471                                         &event->mmap2.min,
472                                         &event->mmap2.ino,
473                                         sizeof(event->mmap2.filename),
474                                         event->mmap2.filename))
475                         continue;
476
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);
482                         truncation = true;
483                         goto out;
484                 }
485
486                 event->mmap2.ino_generation = 0;
487
488                 /*
489                  * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
490                  */
491                 if (machine__is_host(machine))
492                         event->header.misc = PERF_RECORD_MISC_USER;
493                 else
494                         event->header.misc = PERF_RECORD_MISC_GUEST_USER;
495
496                 if ((event->mmap2.prot & PROT_EXEC) == 0) {
497                         if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0)
498                                 continue;
499
500                         event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
501                 }
502
503 out:
504                 if (truncation)
505                         event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
506
507                 if (!strcmp(event->mmap2.filename, ""))
508                         strcpy(event->mmap2.filename, anonstr);
509
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;
515                 }
516
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;
527
528                 if (symbol_conf.buildid_mmap2)
529                         perf_record_mmap2__read_build_id(&event->mmap2, machine, false);
530
531                 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
532                         rc = -1;
533                         break;
534                 }
535
536                 if (truncation)
537                         break;
538         }
539
540         close(io.fd);
541         return rc;
542 }
543
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)
550 {
551         size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path);
552         size_t path_len = strlen(path) - mount_len + 1;
553         struct {
554                 struct file_handle fh;
555                 uint64_t cgroup_id;
556         } handle;
557         int mount_id;
558
559         while (path_len % sizeof(u64))
560                 path[mount_len + path_len++] = '\0';
561
562         memset(&event->cgroup, 0, event_size);
563
564         event->cgroup.header.type = PERF_RECORD_CGROUP;
565         event->cgroup.header.size = event_size + path_len + machine->id_hdr_size;
566
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);
570                 return -1;
571         }
572
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);
576
577         if (perf_tool__process_synth_event(tool, event, machine, process) < 0) {
578                 pr_debug("process synth event failed\n");
579                 return -1;
580         }
581
582         return 0;
583 }
584
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)
590 {
591         size_t pos = strlen(path);
592         DIR *d;
593         struct dirent *dent;
594         int ret = 0;
595
596         if (perf_event__synthesize_cgroup(tool, event, path, mount_len,
597                                           process, machine) < 0)
598                 return -1;
599
600         d = opendir(path);
601         if (d == NULL) {
602                 pr_debug("failed to open directory: %s\n", path);
603                 return -1;
604         }
605
606         while ((dent = readdir(d)) != NULL) {
607                 if (dent->d_type != DT_DIR)
608                         continue;
609                 if (!strcmp(dent->d_name, ".") ||
610                     !strcmp(dent->d_name, ".."))
611                         continue;
612
613                 /* any sane path should be less than PATH_MAX */
614                 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX)
615                         continue;
616
617                 if (path[pos - 1] != '/')
618                         strcat(path, "/");
619                 strcat(path, dent->d_name);
620
621                 ret = perf_event__walk_cgroup_tree(tool, event, path,
622                                                    mount_len, process, machine);
623                 if (ret < 0)
624                         break;
625
626                 path[pos] = '\0';
627         }
628
629         closedir(d);
630         return ret;
631 }
632
633 int perf_event__synthesize_cgroups(struct perf_tool *tool,
634                                    perf_event__handler_t process,
635                                    struct machine *machine)
636 {
637         union perf_event event;
638         char cgrp_root[PATH_MAX];
639         size_t mount_len;  /* length of mount point in the path */
640
641         if (!tool || !tool->cgroup_events)
642                 return 0;
643
644         if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) {
645                 pr_debug("cannot find cgroup mount point\n");
646                 return -1;
647         }
648
649         mount_len = strlen(cgrp_root);
650         /* make sure the path starts with a slash (after mount point) */
651         strcat(cgrp_root, "/");
652
653         if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len,
654                                          process, machine) < 0)
655                 return -1;
656
657         return 0;
658 }
659 #else
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)
663 {
664         return -1;
665 }
666 #endif
667
668 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process,
669                                    struct machine *machine)
670 {
671         int rc = 0;
672         struct map_rb_node *pos;
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);
677
678         event = zalloc(size + machine->id_hdr_size);
679         if (event == NULL) {
680                 pr_debug("Not enough memory synthesizing mmap event "
681                          "for kernel modules\n");
682                 return -1;
683         }
684
685         /*
686          * kernel uses 0 for user space maps, see kernel/perf_event.c
687          * __perf_event_mmap
688          */
689         if (machine__is_host(machine))
690                 event->header.misc = PERF_RECORD_MISC_KERNEL;
691         else
692                 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
693
694         maps__for_each_entry(maps, pos) {
695                 struct map *map = pos->map;
696                 struct dso *dso;
697
698                 if (!__map__is_kmodule(map))
699                         continue;
700
701                 dso = map__dso(map);
702                 if (symbol_conf.buildid_mmap2) {
703                         size = PERF_ALIGN(dso->long_name_len + 1, sizeof(u64));
704                         event->mmap2.header.type = PERF_RECORD_MMAP2;
705                         event->mmap2.header.size = (sizeof(event->mmap2) -
706                                                 (sizeof(event->mmap2.filename) - size));
707                         memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
708                         event->mmap2.header.size += machine->id_hdr_size;
709                         event->mmap2.start = map__start(map);
710                         event->mmap2.len   = map__size(map);
711                         event->mmap2.pid   = machine->pid;
712
713                         memcpy(event->mmap2.filename, dso->long_name, dso->long_name_len + 1);
714
715                         perf_record_mmap2__read_build_id(&event->mmap2, machine, false);
716                 } else {
717                         size = PERF_ALIGN(dso->long_name_len + 1, sizeof(u64));
718                         event->mmap.header.type = PERF_RECORD_MMAP;
719                         event->mmap.header.size = (sizeof(event->mmap) -
720                                                 (sizeof(event->mmap.filename) - size));
721                         memset(event->mmap.filename + size, 0, machine->id_hdr_size);
722                         event->mmap.header.size += machine->id_hdr_size;
723                         event->mmap.start = map__start(map);
724                         event->mmap.len   = map__size(map);
725                         event->mmap.pid   = machine->pid;
726
727                         memcpy(event->mmap.filename, dso->long_name, dso->long_name_len + 1);
728                 }
729
730                 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
731                         rc = -1;
732                         break;
733                 }
734         }
735
736         free(event);
737         return rc;
738 }
739
740 static int filter_task(const struct dirent *dirent)
741 {
742         return isdigit(dirent->d_name[0]);
743 }
744
745 static int __event__synthesize_thread(union perf_event *comm_event,
746                                       union perf_event *mmap_event,
747                                       union perf_event *fork_event,
748                                       union perf_event *namespaces_event,
749                                       pid_t pid, int full, perf_event__handler_t process,
750                                       struct perf_tool *tool, struct machine *machine,
751                                       bool needs_mmap, bool mmap_data)
752 {
753         char filename[PATH_MAX];
754         struct dirent **dirent;
755         pid_t tgid, ppid;
756         int rc = 0;
757         int i, n;
758
759         /* special case: only send one comm event using passed in pid */
760         if (!full) {
761                 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
762                                                    process, machine);
763
764                 if (tgid == -1)
765                         return -1;
766
767                 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
768                                                       tgid, process, machine) < 0)
769                         return -1;
770
771                 /*
772                  * send mmap only for thread group leader
773                  * see thread__init_maps()
774                  */
775                 if (pid == tgid && needs_mmap &&
776                     perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
777                                                        process, machine, mmap_data))
778                         return -1;
779
780                 return 0;
781         }
782
783         if (machine__is_default_guest(machine))
784                 return 0;
785
786         snprintf(filename, sizeof(filename), "%s/proc/%d/task",
787                  machine->root_dir, pid);
788
789         n = scandir(filename, &dirent, filter_task, NULL);
790         if (n < 0)
791                 return n;
792
793         for (i = 0; i < n; i++) {
794                 char *end;
795                 pid_t _pid;
796                 bool kernel_thread = false;
797
798                 _pid = strtol(dirent[i]->d_name, &end, 10);
799                 if (*end)
800                         continue;
801
802                 /* some threads may exit just after scan, ignore it */
803                 if (perf_event__prepare_comm(comm_event, pid, _pid, machine,
804                                              &tgid, &ppid, &kernel_thread) != 0)
805                         continue;
806
807                 rc = -1;
808                 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
809                                                 ppid, process, machine) < 0)
810                         break;
811
812                 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
813                                                       tgid, process, machine) < 0)
814                         break;
815
816                 /*
817                  * Send the prepared comm event
818                  */
819                 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
820                         break;
821
822                 rc = 0;
823                 if (_pid == pid && !kernel_thread && needs_mmap) {
824                         /* process the parent's maps too */
825                         rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
826                                                 process, machine, mmap_data);
827                         if (rc)
828                                 break;
829                 }
830         }
831
832         for (i = 0; i < n; i++)
833                 zfree(&dirent[i]);
834         free(dirent);
835
836         return rc;
837 }
838
839 int perf_event__synthesize_thread_map(struct perf_tool *tool,
840                                       struct perf_thread_map *threads,
841                                       perf_event__handler_t process,
842                                       struct machine *machine,
843                                       bool needs_mmap, bool mmap_data)
844 {
845         union perf_event *comm_event, *mmap_event, *fork_event;
846         union perf_event *namespaces_event;
847         int err = -1, thread, j;
848
849         comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
850         if (comm_event == NULL)
851                 goto out;
852
853         mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
854         if (mmap_event == NULL)
855                 goto out_free_comm;
856
857         fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
858         if (fork_event == NULL)
859                 goto out_free_mmap;
860
861         namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
862                                   (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
863                                   machine->id_hdr_size);
864         if (namespaces_event == NULL)
865                 goto out_free_fork;
866
867         err = 0;
868         for (thread = 0; thread < threads->nr; ++thread) {
869                 if (__event__synthesize_thread(comm_event, mmap_event,
870                                                fork_event, namespaces_event,
871                                                perf_thread_map__pid(threads, thread), 0,
872                                                process, tool, machine,
873                                                needs_mmap, mmap_data)) {
874                         err = -1;
875                         break;
876                 }
877
878                 /*
879                  * comm.pid is set to thread group id by
880                  * perf_event__synthesize_comm
881                  */
882                 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) {
883                         bool need_leader = true;
884
885                         /* is thread group leader in thread_map? */
886                         for (j = 0; j < threads->nr; ++j) {
887                                 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) {
888                                         need_leader = false;
889                                         break;
890                                 }
891                         }
892
893                         /* if not, generate events for it */
894                         if (need_leader &&
895                             __event__synthesize_thread(comm_event, mmap_event,
896                                                        fork_event, namespaces_event,
897                                                        comm_event->comm.pid, 0,
898                                                        process, tool, machine,
899                                                        needs_mmap, mmap_data)) {
900                                 err = -1;
901                                 break;
902                         }
903                 }
904         }
905         free(namespaces_event);
906 out_free_fork:
907         free(fork_event);
908 out_free_mmap:
909         free(mmap_event);
910 out_free_comm:
911         free(comm_event);
912 out:
913         return err;
914 }
915
916 static int __perf_event__synthesize_threads(struct perf_tool *tool,
917                                             perf_event__handler_t process,
918                                             struct machine *machine,
919                                             bool needs_mmap,
920                                             bool mmap_data,
921                                             struct dirent **dirent,
922                                             int start,
923                                             int num)
924 {
925         union perf_event *comm_event, *mmap_event, *fork_event;
926         union perf_event *namespaces_event;
927         int err = -1;
928         char *end;
929         pid_t pid;
930         int i;
931
932         comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
933         if (comm_event == NULL)
934                 goto out;
935
936         mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
937         if (mmap_event == NULL)
938                 goto out_free_comm;
939
940         fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
941         if (fork_event == NULL)
942                 goto out_free_mmap;
943
944         namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
945                                   (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
946                                   machine->id_hdr_size);
947         if (namespaces_event == NULL)
948                 goto out_free_fork;
949
950         for (i = start; i < start + num; i++) {
951                 if (!isdigit(dirent[i]->d_name[0]))
952                         continue;
953
954                 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
955                 /* only interested in proper numerical dirents */
956                 if (*end)
957                         continue;
958                 /*
959                  * We may race with exiting thread, so don't stop just because
960                  * one thread couldn't be synthesized.
961                  */
962                 __event__synthesize_thread(comm_event, mmap_event, fork_event,
963                                            namespaces_event, pid, 1, process,
964                                            tool, machine, needs_mmap, mmap_data);
965         }
966         err = 0;
967
968         free(namespaces_event);
969 out_free_fork:
970         free(fork_event);
971 out_free_mmap:
972         free(mmap_event);
973 out_free_comm:
974         free(comm_event);
975 out:
976         return err;
977 }
978
979 struct synthesize_threads_arg {
980         struct perf_tool *tool;
981         perf_event__handler_t process;
982         struct machine *machine;
983         bool needs_mmap;
984         bool mmap_data;
985         struct dirent **dirent;
986         int num;
987         int start;
988 };
989
990 static void *synthesize_threads_worker(void *arg)
991 {
992         struct synthesize_threads_arg *args = arg;
993
994         __perf_event__synthesize_threads(args->tool, args->process,
995                                          args->machine,
996                                          args->needs_mmap, args->mmap_data,
997                                          args->dirent,
998                                          args->start, args->num);
999         return NULL;
1000 }
1001
1002 int perf_event__synthesize_threads(struct perf_tool *tool,
1003                                    perf_event__handler_t process,
1004                                    struct machine *machine,
1005                                    bool needs_mmap, bool mmap_data,
1006                                    unsigned int nr_threads_synthesize)
1007 {
1008         struct synthesize_threads_arg *args = NULL;
1009         pthread_t *synthesize_threads = NULL;
1010         char proc_path[PATH_MAX];
1011         struct dirent **dirent;
1012         int num_per_thread;
1013         int m, n, i, j;
1014         int thread_nr;
1015         int base = 0;
1016         int err = -1;
1017
1018
1019         if (machine__is_default_guest(machine))
1020                 return 0;
1021
1022         snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
1023         n = scandir(proc_path, &dirent, filter_task, NULL);
1024         if (n < 0)
1025                 return err;
1026
1027         if (nr_threads_synthesize == UINT_MAX)
1028                 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
1029         else
1030                 thread_nr = nr_threads_synthesize;
1031
1032         if (thread_nr <= 1) {
1033                 err = __perf_event__synthesize_threads(tool, process,
1034                                                        machine,
1035                                                        needs_mmap, mmap_data,
1036                                                        dirent, base, n);
1037                 goto free_dirent;
1038         }
1039         if (thread_nr > n)
1040                 thread_nr = n;
1041
1042         synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
1043         if (synthesize_threads == NULL)
1044                 goto free_dirent;
1045
1046         args = calloc(sizeof(*args), thread_nr);
1047         if (args == NULL)
1048                 goto free_threads;
1049
1050         num_per_thread = n / thread_nr;
1051         m = n % thread_nr;
1052         for (i = 0; i < thread_nr; i++) {
1053                 args[i].tool = tool;
1054                 args[i].process = process;
1055                 args[i].machine = machine;
1056                 args[i].needs_mmap = needs_mmap;
1057                 args[i].mmap_data = mmap_data;
1058                 args[i].dirent = dirent;
1059         }
1060         for (i = 0; i < m; i++) {
1061                 args[i].num = num_per_thread + 1;
1062                 args[i].start = i * args[i].num;
1063         }
1064         if (i != 0)
1065                 base = args[i-1].start + args[i-1].num;
1066         for (j = i; j < thread_nr; j++) {
1067                 args[j].num = num_per_thread;
1068                 args[j].start = base + (j - i) * args[i].num;
1069         }
1070
1071         for (i = 0; i < thread_nr; i++) {
1072                 if (pthread_create(&synthesize_threads[i], NULL,
1073                                    synthesize_threads_worker, &args[i]))
1074                         goto out_join;
1075         }
1076         err = 0;
1077 out_join:
1078         for (i = 0; i < thread_nr; i++)
1079                 pthread_join(synthesize_threads[i], NULL);
1080         free(args);
1081 free_threads:
1082         free(synthesize_threads);
1083 free_dirent:
1084         for (i = 0; i < n; i++)
1085                 zfree(&dirent[i]);
1086         free(dirent);
1087
1088         return err;
1089 }
1090
1091 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
1092                                               perf_event__handler_t process __maybe_unused,
1093                                               struct machine *machine __maybe_unused)
1094 {
1095         return 0;
1096 }
1097
1098 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
1099                                                 perf_event__handler_t process,
1100                                                 struct machine *machine)
1101 {
1102         union perf_event *event;
1103         size_t size = symbol_conf.buildid_mmap2 ?
1104                         sizeof(event->mmap2) : sizeof(event->mmap);
1105         struct map *map = machine__kernel_map(machine);
1106         struct kmap *kmap;
1107         int err;
1108
1109         if (map == NULL)
1110                 return -1;
1111
1112         kmap = map__kmap(map);
1113         if (!kmap->ref_reloc_sym)
1114                 return -1;
1115
1116         /*
1117          * We should get this from /sys/kernel/sections/.text, but till that is
1118          * available use this, and after it is use this as a fallback for older
1119          * kernels.
1120          */
1121         event = zalloc(size + machine->id_hdr_size);
1122         if (event == NULL) {
1123                 pr_debug("Not enough memory synthesizing mmap event "
1124                          "for kernel modules\n");
1125                 return -1;
1126         }
1127
1128         if (machine__is_host(machine)) {
1129                 /*
1130                  * kernel uses PERF_RECORD_MISC_USER for user space maps,
1131                  * see kernel/perf_event.c __perf_event_mmap
1132                  */
1133                 event->header.misc = PERF_RECORD_MISC_KERNEL;
1134         } else {
1135                 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
1136         }
1137
1138         if (symbol_conf.buildid_mmap2) {
1139                 size = snprintf(event->mmap2.filename, sizeof(event->mmap2.filename),
1140                                 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
1141                 size = PERF_ALIGN(size, sizeof(u64));
1142                 event->mmap2.header.type = PERF_RECORD_MMAP2;
1143                 event->mmap2.header.size = (sizeof(event->mmap2) -
1144                                 (sizeof(event->mmap2.filename) - size) + machine->id_hdr_size);
1145                 event->mmap2.pgoff = kmap->ref_reloc_sym->addr;
1146                 event->mmap2.start = map__start(map);
1147                 event->mmap2.len   = map__end(map) - event->mmap.start;
1148                 event->mmap2.pid   = machine->pid;
1149
1150                 perf_record_mmap2__read_build_id(&event->mmap2, machine, true);
1151         } else {
1152                 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
1153                                 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
1154                 size = PERF_ALIGN(size, sizeof(u64));
1155                 event->mmap.header.type = PERF_RECORD_MMAP;
1156                 event->mmap.header.size = (sizeof(event->mmap) -
1157                                 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
1158                 event->mmap.pgoff = kmap->ref_reloc_sym->addr;
1159                 event->mmap.start = map__start(map);
1160                 event->mmap.len   = map__end(map) - event->mmap.start;
1161                 event->mmap.pid   = machine->pid;
1162         }
1163
1164         err = perf_tool__process_synth_event(tool, event, machine, process);
1165         free(event);
1166
1167         return err;
1168 }
1169
1170 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
1171                                        perf_event__handler_t process,
1172                                        struct machine *machine)
1173 {
1174         int err;
1175
1176         err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
1177         if (err < 0)
1178                 return err;
1179
1180         return perf_event__synthesize_extra_kmaps(tool, process, machine);
1181 }
1182
1183 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
1184                                       struct perf_thread_map *threads,
1185                                       perf_event__handler_t process,
1186                                       struct machine *machine)
1187 {
1188         union perf_event *event;
1189         int i, err, size;
1190
1191         size  = sizeof(event->thread_map);
1192         size += threads->nr * sizeof(event->thread_map.entries[0]);
1193
1194         event = zalloc(size);
1195         if (!event)
1196                 return -ENOMEM;
1197
1198         event->header.type = PERF_RECORD_THREAD_MAP;
1199         event->header.size = size;
1200         event->thread_map.nr = threads->nr;
1201
1202         for (i = 0; i < threads->nr; i++) {
1203                 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i];
1204                 char *comm = perf_thread_map__comm(threads, i);
1205
1206                 if (!comm)
1207                         comm = (char *) "";
1208
1209                 entry->pid = perf_thread_map__pid(threads, i);
1210                 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
1211         }
1212
1213         err = process(tool, event, NULL, machine);
1214
1215         free(event);
1216         return err;
1217 }
1218
1219 struct synthesize_cpu_map_data {
1220         const struct perf_cpu_map *map;
1221         int nr;
1222         int min_cpu;
1223         int max_cpu;
1224         int has_any_cpu;
1225         int type;
1226         size_t size;
1227         struct perf_record_cpu_map_data *data;
1228 };
1229
1230 static void synthesize_cpus(struct synthesize_cpu_map_data *data)
1231 {
1232         data->data->type = PERF_CPU_MAP__CPUS;
1233         data->data->cpus_data.nr = data->nr;
1234         for (int i = 0; i < data->nr; i++)
1235                 data->data->cpus_data.cpu[i] = perf_cpu_map__cpu(data->map, i).cpu;
1236 }
1237
1238 static void synthesize_mask(struct synthesize_cpu_map_data *data)
1239 {
1240         int idx;
1241         struct perf_cpu cpu;
1242
1243         /* Due to padding, the 4bytes per entry mask variant is always smaller. */
1244         data->data->type = PERF_CPU_MAP__MASK;
1245         data->data->mask32_data.nr = BITS_TO_U32(data->max_cpu);
1246         data->data->mask32_data.long_size = 4;
1247
1248         perf_cpu_map__for_each_cpu(cpu, idx, data->map) {
1249                 int bit_word = cpu.cpu / 32;
1250                 u32 bit_mask = 1U << (cpu.cpu & 31);
1251
1252                 data->data->mask32_data.mask[bit_word] |= bit_mask;
1253         }
1254 }
1255
1256 static void synthesize_range_cpus(struct synthesize_cpu_map_data *data)
1257 {
1258         data->data->type = PERF_CPU_MAP__RANGE_CPUS;
1259         data->data->range_cpu_data.any_cpu = data->has_any_cpu;
1260         data->data->range_cpu_data.start_cpu = data->min_cpu;
1261         data->data->range_cpu_data.end_cpu = data->max_cpu;
1262 }
1263
1264 static void *cpu_map_data__alloc(struct synthesize_cpu_map_data *syn_data,
1265                                  size_t header_size)
1266 {
1267         size_t size_cpus, size_mask;
1268
1269         syn_data->nr = perf_cpu_map__nr(syn_data->map);
1270         syn_data->has_any_cpu = (perf_cpu_map__cpu(syn_data->map, 0).cpu == -1) ? 1 : 0;
1271
1272         syn_data->min_cpu = perf_cpu_map__cpu(syn_data->map, syn_data->has_any_cpu).cpu;
1273         syn_data->max_cpu = perf_cpu_map__max(syn_data->map).cpu;
1274         if (syn_data->max_cpu - syn_data->min_cpu + 1 == syn_data->nr - syn_data->has_any_cpu) {
1275                 /* A consecutive range of CPUs can be encoded using a range. */
1276                 assert(sizeof(u16) + sizeof(struct perf_record_range_cpu_map) == sizeof(u64));
1277                 syn_data->type = PERF_CPU_MAP__RANGE_CPUS;
1278                 syn_data->size = header_size + sizeof(u64);
1279                 return zalloc(syn_data->size);
1280         }
1281
1282         size_cpus = sizeof(u16) + sizeof(struct cpu_map_entries) + syn_data->nr * sizeof(u16);
1283         /* Due to padding, the 4bytes per entry mask variant is always smaller. */
1284         size_mask = sizeof(u16) + sizeof(struct perf_record_mask_cpu_map32) +
1285                 BITS_TO_U32(syn_data->max_cpu) * sizeof(__u32);
1286         if (syn_data->has_any_cpu || size_cpus < size_mask) {
1287                 /* Follow the CPU map encoding. */
1288                 syn_data->type = PERF_CPU_MAP__CPUS;
1289                 syn_data->size = header_size + PERF_ALIGN(size_cpus, sizeof(u64));
1290                 return zalloc(syn_data->size);
1291         }
1292         /* Encode using a bitmask. */
1293         syn_data->type = PERF_CPU_MAP__MASK;
1294         syn_data->size = header_size + PERF_ALIGN(size_mask, sizeof(u64));
1295         return zalloc(syn_data->size);
1296 }
1297
1298 static void cpu_map_data__synthesize(struct synthesize_cpu_map_data *data)
1299 {
1300         switch (data->type) {
1301         case PERF_CPU_MAP__CPUS:
1302                 synthesize_cpus(data);
1303                 break;
1304         case PERF_CPU_MAP__MASK:
1305                 synthesize_mask(data);
1306                 break;
1307         case PERF_CPU_MAP__RANGE_CPUS:
1308                 synthesize_range_cpus(data);
1309                 break;
1310         default:
1311                 break;
1312         }
1313 }
1314
1315 static struct perf_record_cpu_map *cpu_map_event__new(const struct perf_cpu_map *map)
1316 {
1317         struct synthesize_cpu_map_data syn_data = { .map = map };
1318         struct perf_record_cpu_map *event;
1319
1320
1321         event = cpu_map_data__alloc(&syn_data, sizeof(struct perf_event_header));
1322         if (!event)
1323                 return NULL;
1324
1325         syn_data.data = &event->data;
1326         event->header.type = PERF_RECORD_CPU_MAP;
1327         event->header.size = syn_data.size;
1328         cpu_map_data__synthesize(&syn_data);
1329         return event;
1330 }
1331
1332
1333 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1334                                    const struct perf_cpu_map *map,
1335                                    perf_event__handler_t process,
1336                                    struct machine *machine)
1337 {
1338         struct perf_record_cpu_map *event;
1339         int err;
1340
1341         event = cpu_map_event__new(map);
1342         if (!event)
1343                 return -ENOMEM;
1344
1345         err = process(tool, (union perf_event *) event, NULL, machine);
1346
1347         free(event);
1348         return err;
1349 }
1350
1351 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1352                                        struct perf_stat_config *config,
1353                                        perf_event__handler_t process,
1354                                        struct machine *machine)
1355 {
1356         struct perf_record_stat_config *event;
1357         int size, i = 0, err;
1358
1359         size  = sizeof(*event);
1360         size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1361
1362         event = zalloc(size);
1363         if (!event)
1364                 return -ENOMEM;
1365
1366         event->header.type = PERF_RECORD_STAT_CONFIG;
1367         event->header.size = size;
1368         event->nr          = PERF_STAT_CONFIG_TERM__MAX;
1369
1370 #define ADD(__term, __val)                                      \
1371         event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term;   \
1372         event->data[i].val = __val;                             \
1373         i++;
1374
1375         ADD(AGGR_MODE,  config->aggr_mode)
1376         ADD(INTERVAL,   config->interval)
1377         ADD(SCALE,      config->scale)
1378
1379         WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1380                   "stat config terms unbalanced\n");
1381 #undef ADD
1382
1383         err = process(tool, (union perf_event *) event, NULL, machine);
1384
1385         free(event);
1386         return err;
1387 }
1388
1389 int perf_event__synthesize_stat(struct perf_tool *tool,
1390                                 struct perf_cpu cpu, u32 thread, u64 id,
1391                                 struct perf_counts_values *count,
1392                                 perf_event__handler_t process,
1393                                 struct machine *machine)
1394 {
1395         struct perf_record_stat event;
1396
1397         event.header.type = PERF_RECORD_STAT;
1398         event.header.size = sizeof(event);
1399         event.header.misc = 0;
1400
1401         event.id        = id;
1402         event.cpu       = cpu.cpu;
1403         event.thread    = thread;
1404         event.val       = count->val;
1405         event.ena       = count->ena;
1406         event.run       = count->run;
1407
1408         return process(tool, (union perf_event *) &event, NULL, machine);
1409 }
1410
1411 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1412                                       u64 evtime, u64 type,
1413                                       perf_event__handler_t process,
1414                                       struct machine *machine)
1415 {
1416         struct perf_record_stat_round event;
1417
1418         event.header.type = PERF_RECORD_STAT_ROUND;
1419         event.header.size = sizeof(event);
1420         event.header.misc = 0;
1421
1422         event.time = evtime;
1423         event.type = type;
1424
1425         return process(tool, (union perf_event *) &event, NULL, machine);
1426 }
1427
1428 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format)
1429 {
1430         size_t sz, result = sizeof(struct perf_record_sample);
1431
1432         if (type & PERF_SAMPLE_IDENTIFIER)
1433                 result += sizeof(u64);
1434
1435         if (type & PERF_SAMPLE_IP)
1436                 result += sizeof(u64);
1437
1438         if (type & PERF_SAMPLE_TID)
1439                 result += sizeof(u64);
1440
1441         if (type & PERF_SAMPLE_TIME)
1442                 result += sizeof(u64);
1443
1444         if (type & PERF_SAMPLE_ADDR)
1445                 result += sizeof(u64);
1446
1447         if (type & PERF_SAMPLE_ID)
1448                 result += sizeof(u64);
1449
1450         if (type & PERF_SAMPLE_STREAM_ID)
1451                 result += sizeof(u64);
1452
1453         if (type & PERF_SAMPLE_CPU)
1454                 result += sizeof(u64);
1455
1456         if (type & PERF_SAMPLE_PERIOD)
1457                 result += sizeof(u64);
1458
1459         if (type & PERF_SAMPLE_READ) {
1460                 result += sizeof(u64);
1461                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1462                         result += sizeof(u64);
1463                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1464                         result += sizeof(u64);
1465                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1466                 if (read_format & PERF_FORMAT_GROUP) {
1467                         sz = sample_read_value_size(read_format);
1468                         result += sz * sample->read.group.nr;
1469                 } else {
1470                         result += sizeof(u64);
1471                         if (read_format & PERF_FORMAT_LOST)
1472                                 result += sizeof(u64);
1473                 }
1474         }
1475
1476         if (type & PERF_SAMPLE_CALLCHAIN) {
1477                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1478                 result += sz;
1479         }
1480
1481         if (type & PERF_SAMPLE_RAW) {
1482                 result += sizeof(u32);
1483                 result += sample->raw_size;
1484         }
1485
1486         if (type & PERF_SAMPLE_BRANCH_STACK) {
1487                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1488                 /* nr, hw_idx */
1489                 sz += 2 * sizeof(u64);
1490                 result += sz;
1491         }
1492
1493         if (type & PERF_SAMPLE_REGS_USER) {
1494                 if (sample->user_regs.abi) {
1495                         result += sizeof(u64);
1496                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1497                         result += sz;
1498                 } else {
1499                         result += sizeof(u64);
1500                 }
1501         }
1502
1503         if (type & PERF_SAMPLE_STACK_USER) {
1504                 sz = sample->user_stack.size;
1505                 result += sizeof(u64);
1506                 if (sz) {
1507                         result += sz;
1508                         result += sizeof(u64);
1509                 }
1510         }
1511
1512         if (type & PERF_SAMPLE_WEIGHT_TYPE)
1513                 result += sizeof(u64);
1514
1515         if (type & PERF_SAMPLE_DATA_SRC)
1516                 result += sizeof(u64);
1517
1518         if (type & PERF_SAMPLE_TRANSACTION)
1519                 result += sizeof(u64);
1520
1521         if (type & PERF_SAMPLE_REGS_INTR) {
1522                 if (sample->intr_regs.abi) {
1523                         result += sizeof(u64);
1524                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1525                         result += sz;
1526                 } else {
1527                         result += sizeof(u64);
1528                 }
1529         }
1530
1531         if (type & PERF_SAMPLE_PHYS_ADDR)
1532                 result += sizeof(u64);
1533
1534         if (type & PERF_SAMPLE_CGROUP)
1535                 result += sizeof(u64);
1536
1537         if (type & PERF_SAMPLE_DATA_PAGE_SIZE)
1538                 result += sizeof(u64);
1539
1540         if (type & PERF_SAMPLE_CODE_PAGE_SIZE)
1541                 result += sizeof(u64);
1542
1543         if (type & PERF_SAMPLE_AUX) {
1544                 result += sizeof(u64);
1545                 result += sample->aux_sample.size;
1546         }
1547
1548         return result;
1549 }
1550
1551 void __weak arch_perf_synthesize_sample_weight(const struct perf_sample *data,
1552                                                __u64 *array, u64 type __maybe_unused)
1553 {
1554         *array = data->weight;
1555 }
1556
1557 static __u64 *copy_read_group_values(__u64 *array, __u64 read_format,
1558                                      const struct perf_sample *sample)
1559 {
1560         size_t sz = sample_read_value_size(read_format);
1561         struct sample_read_value *v = sample->read.group.values;
1562
1563         sample_read_group__for_each(v, sample->read.group.nr, read_format) {
1564                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1565                 memcpy(array, v, sz);
1566                 array = (void *)array + sz;
1567         }
1568         return array;
1569 }
1570
1571 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format,
1572                                   const struct perf_sample *sample)
1573 {
1574         __u64 *array;
1575         size_t sz;
1576         /*
1577          * used for cross-endian analysis. See git commit 65014ab3
1578          * for why this goofiness is needed.
1579          */
1580         union u64_swap u;
1581
1582         array = event->sample.array;
1583
1584         if (type & PERF_SAMPLE_IDENTIFIER) {
1585                 *array = sample->id;
1586                 array++;
1587         }
1588
1589         if (type & PERF_SAMPLE_IP) {
1590                 *array = sample->ip;
1591                 array++;
1592         }
1593
1594         if (type & PERF_SAMPLE_TID) {
1595                 u.val32[0] = sample->pid;
1596                 u.val32[1] = sample->tid;
1597                 *array = u.val64;
1598                 array++;
1599         }
1600
1601         if (type & PERF_SAMPLE_TIME) {
1602                 *array = sample->time;
1603                 array++;
1604         }
1605
1606         if (type & PERF_SAMPLE_ADDR) {
1607                 *array = sample->addr;
1608                 array++;
1609         }
1610
1611         if (type & PERF_SAMPLE_ID) {
1612                 *array = sample->id;
1613                 array++;
1614         }
1615
1616         if (type & PERF_SAMPLE_STREAM_ID) {
1617                 *array = sample->stream_id;
1618                 array++;
1619         }
1620
1621         if (type & PERF_SAMPLE_CPU) {
1622                 u.val32[0] = sample->cpu;
1623                 u.val32[1] = 0;
1624                 *array = u.val64;
1625                 array++;
1626         }
1627
1628         if (type & PERF_SAMPLE_PERIOD) {
1629                 *array = sample->period;
1630                 array++;
1631         }
1632
1633         if (type & PERF_SAMPLE_READ) {
1634                 if (read_format & PERF_FORMAT_GROUP)
1635                         *array = sample->read.group.nr;
1636                 else
1637                         *array = sample->read.one.value;
1638                 array++;
1639
1640                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1641                         *array = sample->read.time_enabled;
1642                         array++;
1643                 }
1644
1645                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1646                         *array = sample->read.time_running;
1647                         array++;
1648                 }
1649
1650                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1651                 if (read_format & PERF_FORMAT_GROUP) {
1652                         array = copy_read_group_values(array, read_format,
1653                                                        sample);
1654                 } else {
1655                         *array = sample->read.one.id;
1656                         array++;
1657
1658                         if (read_format & PERF_FORMAT_LOST) {
1659                                 *array = sample->read.one.lost;
1660                                 array++;
1661                         }
1662                 }
1663         }
1664
1665         if (type & PERF_SAMPLE_CALLCHAIN) {
1666                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1667                 memcpy(array, sample->callchain, sz);
1668                 array = (void *)array + sz;
1669         }
1670
1671         if (type & PERF_SAMPLE_RAW) {
1672                 u.val32[0] = sample->raw_size;
1673                 *array = u.val64;
1674                 array = (void *)array + sizeof(u32);
1675
1676                 memcpy(array, sample->raw_data, sample->raw_size);
1677                 array = (void *)array + sample->raw_size;
1678         }
1679
1680         if (type & PERF_SAMPLE_BRANCH_STACK) {
1681                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1682                 /* nr, hw_idx */
1683                 sz += 2 * sizeof(u64);
1684                 memcpy(array, sample->branch_stack, sz);
1685                 array = (void *)array + sz;
1686         }
1687
1688         if (type & PERF_SAMPLE_REGS_USER) {
1689                 if (sample->user_regs.abi) {
1690                         *array++ = sample->user_regs.abi;
1691                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1692                         memcpy(array, sample->user_regs.regs, sz);
1693                         array = (void *)array + sz;
1694                 } else {
1695                         *array++ = 0;
1696                 }
1697         }
1698
1699         if (type & PERF_SAMPLE_STACK_USER) {
1700                 sz = sample->user_stack.size;
1701                 *array++ = sz;
1702                 if (sz) {
1703                         memcpy(array, sample->user_stack.data, sz);
1704                         array = (void *)array + sz;
1705                         *array++ = sz;
1706                 }
1707         }
1708
1709         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
1710                 arch_perf_synthesize_sample_weight(sample, array, type);
1711                 array++;
1712         }
1713
1714         if (type & PERF_SAMPLE_DATA_SRC) {
1715                 *array = sample->data_src;
1716                 array++;
1717         }
1718
1719         if (type & PERF_SAMPLE_TRANSACTION) {
1720                 *array = sample->transaction;
1721                 array++;
1722         }
1723
1724         if (type & PERF_SAMPLE_REGS_INTR) {
1725                 if (sample->intr_regs.abi) {
1726                         *array++ = sample->intr_regs.abi;
1727                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1728                         memcpy(array, sample->intr_regs.regs, sz);
1729                         array = (void *)array + sz;
1730                 } else {
1731                         *array++ = 0;
1732                 }
1733         }
1734
1735         if (type & PERF_SAMPLE_PHYS_ADDR) {
1736                 *array = sample->phys_addr;
1737                 array++;
1738         }
1739
1740         if (type & PERF_SAMPLE_CGROUP) {
1741                 *array = sample->cgroup;
1742                 array++;
1743         }
1744
1745         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
1746                 *array = sample->data_page_size;
1747                 array++;
1748         }
1749
1750         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
1751                 *array = sample->code_page_size;
1752                 array++;
1753         }
1754
1755         if (type & PERF_SAMPLE_AUX) {
1756                 sz = sample->aux_sample.size;
1757                 *array++ = sz;
1758                 memcpy(array, sample->aux_sample.data, sz);
1759                 array = (void *)array + sz;
1760         }
1761
1762         return 0;
1763 }
1764
1765 int perf_event__synthesize_id_sample(__u64 *array, u64 type, const struct perf_sample *sample)
1766 {
1767         __u64 *start = array;
1768
1769         /*
1770          * used for cross-endian analysis. See git commit 65014ab3
1771          * for why this goofiness is needed.
1772          */
1773         union u64_swap u;
1774
1775         if (type & PERF_SAMPLE_TID) {
1776                 u.val32[0] = sample->pid;
1777                 u.val32[1] = sample->tid;
1778                 *array = u.val64;
1779                 array++;
1780         }
1781
1782         if (type & PERF_SAMPLE_TIME) {
1783                 *array = sample->time;
1784                 array++;
1785         }
1786
1787         if (type & PERF_SAMPLE_ID) {
1788                 *array = sample->id;
1789                 array++;
1790         }
1791
1792         if (type & PERF_SAMPLE_STREAM_ID) {
1793                 *array = sample->stream_id;
1794                 array++;
1795         }
1796
1797         if (type & PERF_SAMPLE_CPU) {
1798                 u.val32[0] = sample->cpu;
1799                 u.val32[1] = 0;
1800                 *array = u.val64;
1801                 array++;
1802         }
1803
1804         if (type & PERF_SAMPLE_IDENTIFIER) {
1805                 *array = sample->id;
1806                 array++;
1807         }
1808
1809         return (void *)array - (void *)start;
1810 }
1811
1812 int __perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process,
1813                                       struct evlist *evlist, struct machine *machine, size_t from)
1814 {
1815         union perf_event *ev;
1816         struct evsel *evsel;
1817         size_t nr = 0, i = 0, sz, max_nr, n, pos;
1818         size_t e1_sz = sizeof(struct id_index_entry);
1819         size_t e2_sz = sizeof(struct id_index_entry_2);
1820         size_t etot_sz = e1_sz + e2_sz;
1821         bool e2_needed = false;
1822         int err;
1823
1824         max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / etot_sz;
1825
1826         pos = 0;
1827         evlist__for_each_entry(evlist, evsel) {
1828                 if (pos++ < from)
1829                         continue;
1830                 nr += evsel->core.ids;
1831         }
1832
1833         if (!nr)
1834                 return 0;
1835
1836         pr_debug2("Synthesizing id index\n");
1837
1838         n = nr > max_nr ? max_nr : nr;
1839         sz = sizeof(struct perf_record_id_index) + n * etot_sz;
1840         ev = zalloc(sz);
1841         if (!ev)
1842                 return -ENOMEM;
1843
1844         sz = sizeof(struct perf_record_id_index) + n * e1_sz;
1845
1846         ev->id_index.header.type = PERF_RECORD_ID_INDEX;
1847         ev->id_index.nr = n;
1848
1849         pos = 0;
1850         evlist__for_each_entry(evlist, evsel) {
1851                 u32 j;
1852
1853                 if (pos++ < from)
1854                         continue;
1855                 for (j = 0; j < evsel->core.ids; j++, i++) {
1856                         struct id_index_entry *e;
1857                         struct id_index_entry_2 *e2;
1858                         struct perf_sample_id *sid;
1859
1860                         if (i >= n) {
1861                                 ev->id_index.header.size = sz + (e2_needed ? n * e2_sz : 0);
1862                                 err = process(tool, ev, NULL, machine);
1863                                 if (err)
1864                                         goto out_err;
1865                                 nr -= n;
1866                                 i = 0;
1867                                 e2_needed = false;
1868                         }
1869
1870                         e = &ev->id_index.entries[i];
1871
1872                         e->id = evsel->core.id[j];
1873
1874                         sid = evlist__id2sid(evlist, e->id);
1875                         if (!sid) {
1876                                 free(ev);
1877                                 return -ENOENT;
1878                         }
1879
1880                         e->idx = sid->idx;
1881                         e->cpu = sid->cpu.cpu;
1882                         e->tid = sid->tid;
1883
1884                         if (sid->machine_pid)
1885                                 e2_needed = true;
1886
1887                         e2 = (void *)ev + sz;
1888                         e2[i].machine_pid = sid->machine_pid;
1889                         e2[i].vcpu        = sid->vcpu.cpu;
1890                 }
1891         }
1892
1893         sz = sizeof(struct perf_record_id_index) + nr * e1_sz;
1894         ev->id_index.header.size = sz + (e2_needed ? nr * e2_sz : 0);
1895         ev->id_index.nr = nr;
1896
1897         err = process(tool, ev, NULL, machine);
1898 out_err:
1899         free(ev);
1900
1901         return err;
1902 }
1903
1904 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process,
1905                                     struct evlist *evlist, struct machine *machine)
1906 {
1907         return __perf_event__synthesize_id_index(tool, process, evlist, machine, 0);
1908 }
1909
1910 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1911                                   struct target *target, struct perf_thread_map *threads,
1912                                   perf_event__handler_t process, bool needs_mmap,
1913                                   bool data_mmap, unsigned int nr_threads_synthesize)
1914 {
1915         /*
1916          * When perf runs in non-root PID namespace, and the namespace's proc FS
1917          * is not mounted, nsinfo__is_in_root_namespace() returns false.
1918          * In this case, the proc FS is coming for the parent namespace, thus
1919          * perf tool will wrongly gather process info from its parent PID
1920          * namespace.
1921          *
1922          * To avoid the confusion that the perf tool runs in a child PID
1923          * namespace but it synthesizes thread info from its parent PID
1924          * namespace, returns failure with warning.
1925          */
1926         if (!nsinfo__is_in_root_namespace()) {
1927                 pr_err("Perf runs in non-root PID namespace but it tries to ");
1928                 pr_err("gather process info from its parent PID namespace.\n");
1929                 pr_err("Please mount the proc file system properly, e.g. ");
1930                 pr_err("add the option '--mount-proc' for unshare command.\n");
1931                 return -EPERM;
1932         }
1933
1934         if (target__has_task(target))
1935                 return perf_event__synthesize_thread_map(tool, threads, process, machine,
1936                                                          needs_mmap, data_mmap);
1937         else if (target__has_cpu(target))
1938                 return perf_event__synthesize_threads(tool, process, machine,
1939                                                       needs_mmap, data_mmap,
1940                                                       nr_threads_synthesize);
1941         /* command specified */
1942         return 0;
1943 }
1944
1945 int machine__synthesize_threads(struct machine *machine, struct target *target,
1946                                 struct perf_thread_map *threads, bool needs_mmap,
1947                                 bool data_mmap, unsigned int nr_threads_synthesize)
1948 {
1949         return __machine__synthesize_threads(machine, NULL, target, threads,
1950                                              perf_event__process, needs_mmap,
1951                                              data_mmap, nr_threads_synthesize);
1952 }
1953
1954 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id)
1955 {
1956         struct perf_record_event_update *ev;
1957
1958         size += sizeof(*ev);
1959         size  = PERF_ALIGN(size, sizeof(u64));
1960
1961         ev = zalloc(size);
1962         if (ev) {
1963                 ev->header.type = PERF_RECORD_EVENT_UPDATE;
1964                 ev->header.size = (u16)size;
1965                 ev->type        = type;
1966                 ev->id          = id;
1967         }
1968         return ev;
1969 }
1970
1971 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel,
1972                                              perf_event__handler_t process)
1973 {
1974         size_t size = strlen(evsel->unit);
1975         struct perf_record_event_update *ev;
1976         int err;
1977
1978         ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]);
1979         if (ev == NULL)
1980                 return -ENOMEM;
1981
1982         strlcpy(ev->unit, evsel->unit, size + 1);
1983         err = process(tool, (union perf_event *)ev, NULL, NULL);
1984         free(ev);
1985         return err;
1986 }
1987
1988 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel,
1989                                               perf_event__handler_t process)
1990 {
1991         struct perf_record_event_update *ev;
1992         struct perf_record_event_update_scale *ev_data;
1993         int err;
1994
1995         ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]);
1996         if (ev == NULL)
1997                 return -ENOMEM;
1998
1999         ev->scale.scale = evsel->scale;
2000         err = process(tool, (union perf_event *)ev, NULL, NULL);
2001         free(ev);
2002         return err;
2003 }
2004
2005 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel,
2006                                              perf_event__handler_t process)
2007 {
2008         struct perf_record_event_update *ev;
2009         size_t len = strlen(evsel__name(evsel));
2010         int err;
2011
2012         ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]);
2013         if (ev == NULL)
2014                 return -ENOMEM;
2015
2016         strlcpy(ev->name, evsel->name, len + 1);
2017         err = process(tool, (union perf_event *)ev, NULL, NULL);
2018         free(ev);
2019         return err;
2020 }
2021
2022 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel,
2023                                              perf_event__handler_t process)
2024 {
2025         struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };
2026         struct perf_record_event_update *ev;
2027         int err;
2028
2029         ev = cpu_map_data__alloc(&syn_data, sizeof(struct perf_event_header) + 2 * sizeof(u64));
2030         if (!ev)
2031                 return -ENOMEM;
2032
2033         syn_data.data = &ev->cpus.cpus;
2034         ev->header.type = PERF_RECORD_EVENT_UPDATE;
2035         ev->header.size = (u16)syn_data.size;
2036         ev->type        = PERF_EVENT_UPDATE__CPUS;
2037         ev->id          = evsel->core.id[0];
2038         cpu_map_data__synthesize(&syn_data);
2039
2040         err = process(tool, (union perf_event *)ev, NULL, NULL);
2041         free(ev);
2042         return err;
2043 }
2044
2045 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist,
2046                                  perf_event__handler_t process)
2047 {
2048         struct evsel *evsel;
2049         int err = 0;
2050
2051         evlist__for_each_entry(evlist, evsel) {
2052                 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids,
2053                                                   evsel->core.id, process);
2054                 if (err) {
2055                         pr_debug("failed to create perf header attribute\n");
2056                         return err;
2057                 }
2058         }
2059
2060         return err;
2061 }
2062
2063 static bool has_unit(struct evsel *evsel)
2064 {
2065         return evsel->unit && *evsel->unit;
2066 }
2067
2068 static bool has_scale(struct evsel *evsel)
2069 {
2070         return evsel->scale != 1;
2071 }
2072
2073 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list,
2074                                       perf_event__handler_t process, bool is_pipe)
2075 {
2076         struct evsel *evsel;
2077         int err;
2078
2079         /*
2080          * Synthesize other events stuff not carried within
2081          * attr event - unit, scale, name
2082          */
2083         evlist__for_each_entry(evsel_list, evsel) {
2084                 if (!evsel->supported)
2085                         continue;
2086
2087                 /*
2088                  * Synthesize unit and scale only if it's defined.
2089                  */
2090                 if (has_unit(evsel)) {
2091                         err = perf_event__synthesize_event_update_unit(tool, evsel, process);
2092                         if (err < 0) {
2093                                 pr_err("Couldn't synthesize evsel unit.\n");
2094                                 return err;
2095                         }
2096                 }
2097
2098                 if (has_scale(evsel)) {
2099                         err = perf_event__synthesize_event_update_scale(tool, evsel, process);
2100                         if (err < 0) {
2101                                 pr_err("Couldn't synthesize evsel evsel.\n");
2102                                 return err;
2103                         }
2104                 }
2105
2106                 if (evsel->core.own_cpus) {
2107                         err = perf_event__synthesize_event_update_cpus(tool, evsel, process);
2108                         if (err < 0) {
2109                                 pr_err("Couldn't synthesize evsel cpus.\n");
2110                                 return err;
2111                         }
2112                 }
2113
2114                 /*
2115                  * Name is needed only for pipe output,
2116                  * perf.data carries event names.
2117                  */
2118                 if (is_pipe) {
2119                         err = perf_event__synthesize_event_update_name(tool, evsel, process);
2120                         if (err < 0) {
2121                                 pr_err("Couldn't synthesize evsel name.\n");
2122                                 return err;
2123                         }
2124                 }
2125         }
2126         return 0;
2127 }
2128
2129 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr,
2130                                 u32 ids, u64 *id, perf_event__handler_t process)
2131 {
2132         union perf_event *ev;
2133         size_t size;
2134         int err;
2135
2136         size = sizeof(struct perf_event_attr);
2137         size = PERF_ALIGN(size, sizeof(u64));
2138         size += sizeof(struct perf_event_header);
2139         size += ids * sizeof(u64);
2140
2141         ev = zalloc(size);
2142
2143         if (ev == NULL)
2144                 return -ENOMEM;
2145
2146         ev->attr.attr = *attr;
2147         memcpy(ev->attr.id, id, ids * sizeof(u64));
2148
2149         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2150         ev->attr.header.size = (u16)size;
2151
2152         if (ev->attr.header.size == size)
2153                 err = process(tool, ev, NULL, NULL);
2154         else
2155                 err = -E2BIG;
2156
2157         free(ev);
2158
2159         return err;
2160 }
2161
2162 #ifdef HAVE_LIBTRACEEVENT
2163 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist,
2164                                         perf_event__handler_t process)
2165 {
2166         union perf_event ev;
2167         struct tracing_data *tdata;
2168         ssize_t size = 0, aligned_size = 0, padding;
2169         struct feat_fd ff;
2170
2171         /*
2172          * We are going to store the size of the data followed
2173          * by the data contents. Since the fd descriptor is a pipe,
2174          * we cannot seek back to store the size of the data once
2175          * we know it. Instead we:
2176          *
2177          * - write the tracing data to the temp file
2178          * - get/write the data size to pipe
2179          * - write the tracing data from the temp file
2180          *   to the pipe
2181          */
2182         tdata = tracing_data_get(&evlist->core.entries, fd, true);
2183         if (!tdata)
2184                 return -1;
2185
2186         memset(&ev, 0, sizeof(ev));
2187
2188         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2189         size = tdata->size;
2190         aligned_size = PERF_ALIGN(size, sizeof(u64));
2191         padding = aligned_size - size;
2192         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2193         ev.tracing_data.size = aligned_size;
2194
2195         process(tool, &ev, NULL, NULL);
2196
2197         /*
2198          * The put function will copy all the tracing data
2199          * stored in temp file to the pipe.
2200          */
2201         tracing_data_put(tdata);
2202
2203         ff = (struct feat_fd){ .fd = fd };
2204         if (write_padded(&ff, NULL, 0, padding))
2205                 return -1;
2206
2207         return aligned_size;
2208 }
2209 #endif
2210
2211 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc,
2212                                     perf_event__handler_t process, struct machine *machine)
2213 {
2214         union perf_event ev;
2215         size_t len;
2216
2217         if (!pos->hit)
2218                 return 0;
2219
2220         memset(&ev, 0, sizeof(ev));
2221
2222         len = pos->long_name_len + 1;
2223         len = PERF_ALIGN(len, NAME_ALIGN);
2224         ev.build_id.size = min(pos->bid.size, sizeof(pos->bid.data));
2225         memcpy(&ev.build_id.build_id, pos->bid.data, ev.build_id.size);
2226         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2227         ev.build_id.header.misc = misc | PERF_RECORD_MISC_BUILD_ID_SIZE;
2228         ev.build_id.pid = machine->pid;
2229         ev.build_id.header.size = sizeof(ev.build_id) + len;
2230         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2231
2232         return process(tool, &ev, NULL, machine);
2233 }
2234
2235 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool,
2236                                        struct evlist *evlist, perf_event__handler_t process, bool attrs)
2237 {
2238         int err;
2239
2240         if (attrs) {
2241                 err = perf_event__synthesize_attrs(tool, evlist, process);
2242                 if (err < 0) {
2243                         pr_err("Couldn't synthesize attrs.\n");
2244                         return err;
2245                 }
2246         }
2247
2248         err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs);
2249         err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL);
2250         if (err < 0) {
2251                 pr_err("Couldn't synthesize thread map.\n");
2252                 return err;
2253         }
2254
2255         err = perf_event__synthesize_cpu_map(tool, evlist->core.user_requested_cpus, process, NULL);
2256         if (err < 0) {
2257                 pr_err("Couldn't synthesize thread map.\n");
2258                 return err;
2259         }
2260
2261         err = perf_event__synthesize_stat_config(tool, config, process, NULL);
2262         if (err < 0) {
2263                 pr_err("Couldn't synthesize config.\n");
2264                 return err;
2265         }
2266
2267         return 0;
2268 }
2269
2270 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
2271
2272 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session,
2273                                     struct evlist *evlist, perf_event__handler_t process)
2274 {
2275         struct perf_header *header = &session->header;
2276         struct perf_record_header_feature *fe;
2277         struct feat_fd ff;
2278         size_t sz, sz_hdr;
2279         int feat, ret;
2280
2281         sz_hdr = sizeof(fe->header);
2282         sz = sizeof(union perf_event);
2283         /* get a nice alignment */
2284         sz = PERF_ALIGN(sz, page_size);
2285
2286         memset(&ff, 0, sizeof(ff));
2287
2288         ff.buf = malloc(sz);
2289         if (!ff.buf)
2290                 return -ENOMEM;
2291
2292         ff.size = sz - sz_hdr;
2293         ff.ph = &session->header;
2294
2295         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2296                 if (!feat_ops[feat].synthesize) {
2297                         pr_debug("No record header feature for header :%d\n", feat);
2298                         continue;
2299                 }
2300
2301                 ff.offset = sizeof(*fe);
2302
2303                 ret = feat_ops[feat].write(&ff, evlist);
2304                 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
2305                         pr_debug("Error writing feature\n");
2306                         continue;
2307                 }
2308                 /* ff.buf may have changed due to realloc in do_write() */
2309                 fe = ff.buf;
2310                 memset(fe, 0, sizeof(*fe));
2311
2312                 fe->feat_id = feat;
2313                 fe->header.type = PERF_RECORD_HEADER_FEATURE;
2314                 fe->header.size = ff.offset;
2315
2316                 ret = process(tool, ff.buf, NULL, NULL);
2317                 if (ret) {
2318                         free(ff.buf);
2319                         return ret;
2320                 }
2321         }
2322
2323         /* Send HEADER_LAST_FEATURE mark. */
2324         fe = ff.buf;
2325         fe->feat_id     = HEADER_LAST_FEATURE;
2326         fe->header.type = PERF_RECORD_HEADER_FEATURE;
2327         fe->header.size = sizeof(*fe);
2328
2329         ret = process(tool, ff.buf, NULL, NULL);
2330
2331         free(ff.buf);
2332         return ret;
2333 }
2334
2335 int perf_event__synthesize_for_pipe(struct perf_tool *tool,
2336                                     struct perf_session *session,
2337                                     struct perf_data *data,
2338                                     perf_event__handler_t process)
2339 {
2340         int err;
2341         int ret = 0;
2342         struct evlist *evlist = session->evlist;
2343
2344         /*
2345          * We need to synthesize events first, because some
2346          * features works on top of them (on report side).
2347          */
2348         err = perf_event__synthesize_attrs(tool, evlist, process);
2349         if (err < 0) {
2350                 pr_err("Couldn't synthesize attrs.\n");
2351                 return err;
2352         }
2353         ret += err;
2354
2355         err = perf_event__synthesize_features(tool, session, evlist, process);
2356         if (err < 0) {
2357                 pr_err("Couldn't synthesize features.\n");
2358                 return err;
2359         }
2360         ret += err;
2361
2362 #ifdef HAVE_LIBTRACEEVENT
2363         if (have_tracepoints(&evlist->core.entries)) {
2364                 int fd = perf_data__fd(data);
2365
2366                 /*
2367                  * FIXME err <= 0 here actually means that
2368                  * there were no tracepoints so its not really
2369                  * an error, just that we don't need to
2370                  * synthesize anything.  We really have to
2371                  * return this more properly and also
2372                  * propagate errors that now are calling die()
2373                  */
2374                 err = perf_event__synthesize_tracing_data(tool, fd, evlist,
2375                                                           process);
2376                 if (err <= 0) {
2377                         pr_err("Couldn't record tracing data.\n");
2378                         return err;
2379                 }
2380                 ret += err;
2381         }
2382 #else
2383         (void)data;
2384 #endif
2385
2386         return ret;
2387 }
2388
2389 int parse_synth_opt(char *synth)
2390 {
2391         char *p, *q;
2392         int ret = 0;
2393
2394         if (synth == NULL)
2395                 return -1;
2396
2397         for (q = synth; (p = strsep(&q, ",")); p = q) {
2398                 if (!strcasecmp(p, "no") || !strcasecmp(p, "none"))
2399                         return 0;
2400
2401                 if (!strcasecmp(p, "all"))
2402                         return PERF_SYNTH_ALL;
2403
2404                 if (!strcasecmp(p, "task"))
2405                         ret |= PERF_SYNTH_TASK;
2406                 else if (!strcasecmp(p, "mmap"))
2407                         ret |= PERF_SYNTH_TASK | PERF_SYNTH_MMAP;
2408                 else if (!strcasecmp(p, "cgroup"))
2409                         ret |= PERF_SYNTH_CGROUP;
2410                 else
2411                         return -1;
2412         }
2413
2414         return ret;
2415 }