tools/perf/util/synthetic-events.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include "util/debug.h"
   4 #include "util/dso.h"
   5 #include "util/event.h"
   6 #include "util/evlist.h"
   7 #include "util/machine.h"
   8 #include "util/map.h"
   9 #include "util/map_symbol.h"
  10 #include "util/branch.h"
  11 #include "util/memswap.h"
  12 #include "util/namespaces.h"
  13 #include "util/session.h"
  14 #include "util/stat.h"
  15 #include "util/symbol.h"
  16 #include "util/synthetic-events.h"
  17 #include "util/target.h"
  18 #include "util/time-utils.h"
  19 #include "util/cgroup.h"
  20 #include <linux/bitops.h>
  21 #include <linux/kernel.h>
  22 #include <linux/string.h>
  23 #include <linux/zalloc.h>
  24 #include <linux/perf_event.h>
  25 #include <asm/bug.h>
  26 #include <perf/evsel.h>
  27 #include <internal/cpumap.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, char *comm, size_t len,
  73                                     pid_t *tgid, pid_t *ppid)
  74 {
  75         char bf[4096];
  76         int fd;
  77         size_t size = 0;
  78         ssize_t n;
  79         char *name, *tgids, *ppids;
  80
  81         *tgid = -1;
  82         *ppid = -1;
  83
  84         snprintf(bf, sizeof(bf), "/proc/%d/status", pid);
  85
  86         fd = open(bf, O_RDONLY);
  87         if (fd < 0) {
  88                 pr_debug("couldn't open %s\n", bf);
  89                 return -1;
  90         }
  91
  92         n = read(fd, bf, sizeof(bf) - 1);
  93         close(fd);
  94         if (n <= 0) {
  95                 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
  96                            pid);
  97                 return -1;
  98         }
  99         bf[n] = '\0';
 100
 101         name = strstr(bf, "Name:");
 102         tgids = strstr(bf, "Tgid:");
 103         ppids = strstr(bf, "PPid:");
 104
 105         if (name) {
 106                 char *nl;
 107
 108                 name = skip_spaces(name + 5);  /* strlen("Name:") */
 109                 nl = strchr(name, '\n');
 110                 if (nl)
 111                         *nl = '\0';
 112
 113                 size = strlen(name);
 114                 if (size >= len)
 115                         size = len - 1;
 116                 memcpy(comm, name, size);
 117                 comm[size] = '\0';
 118         } else {
 119                 pr_debug("Name: string not found for pid %d\n", pid);
 120         }
 121
 122         if (tgids) {
 123                 tgids += 5;  /* strlen("Tgid:") */
 124                 *tgid = atoi(tgids);
 125         } else {
 126                 pr_debug("Tgid: string not found for pid %d\n", pid);
 127         }
 128
 129         if (ppids) {
 130                 ppids += 5;  /* strlen("PPid:") */
 131                 *ppid = atoi(ppids);
 132         } else {
 133                 pr_debug("PPid: string not found for pid %d\n", pid);
 134         }
 135
 136         return 0;
 137 }
 138
 139 static int perf_event__prepare_comm(union perf_event *event, pid_t pid,
 140                                     struct machine *machine,
 141                                     pid_t *tgid, pid_t *ppid)
 142 {
 143         size_t size;
 144
 145         *ppid = -1;
 146
 147         memset(&event->comm, 0, sizeof(event->comm));
 148
 149         if (machine__is_host(machine)) {
 150                 if (perf_event__get_comm_ids(pid, event->comm.comm,
 151                                              sizeof(event->comm.comm),
 152                                              tgid, ppid) != 0) {
 153                         return -1;
 154                 }
 155         } else {
 156                 *tgid = machine->pid;
 157         }
 158
 159         if (*tgid < 0)
 160                 return -1;
 161
 162         event->comm.pid = *tgid;
 163         event->comm.header.type = PERF_RECORD_COMM;
 164
 165         size = strlen(event->comm.comm) + 1;
 166         size = PERF_ALIGN(size, sizeof(u64));
 167         memset(event->comm.comm + size, 0, machine->id_hdr_size);
 168         event->comm.header.size = (sizeof(event->comm) -
 169                                 (sizeof(event->comm.comm) - size) +
 170                                 machine->id_hdr_size);
 171         event->comm.tid = pid;
 172
 173         return 0;
 174 }
 175
 176 pid_t perf_event__synthesize_comm(struct perf_tool *tool,
 177                                          union perf_event *event, pid_t pid,
 178                                          perf_event__handler_t process,
 179                                          struct machine *machine)
 180 {
 181         pid_t tgid, ppid;
 182
 183         if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0)
 184                 return -1;
 185
 186         if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
 187                 return -1;
 188
 189         return tgid;
 190 }
 191
 192 static void perf_event__get_ns_link_info(pid_t pid, const char *ns,
 193                                          struct perf_ns_link_info *ns_link_info)
 194 {
 195         struct stat64 st;
 196         char proc_ns[128];
 197
 198         sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns);
 199         if (stat64(proc_ns, &st) == 0) {
 200                 ns_link_info->dev = st.st_dev;
 201                 ns_link_info->ino = st.st_ino;
 202         }
 203 }
 204
 205 int perf_event__synthesize_namespaces(struct perf_tool *tool,
 206                                       union perf_event *event,
 207                                       pid_t pid, pid_t tgid,
 208                                       perf_event__handler_t process,
 209                                       struct machine *machine)
 210 {
 211         u32 idx;
 212         struct perf_ns_link_info *ns_link_info;
 213
 214         if (!tool || !tool->namespace_events)
 215                 return 0;
 216
 217         memset(&event->namespaces, 0, (sizeof(event->namespaces) +
 218                (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 219                machine->id_hdr_size));
 220
 221         event->namespaces.pid = tgid;
 222         event->namespaces.tid = pid;
 223
 224         event->namespaces.nr_namespaces = NR_NAMESPACES;
 225
 226         ns_link_info = event->namespaces.link_info;
 227
 228         for (idx = 0; idx < event->namespaces.nr_namespaces; idx++)
 229                 perf_event__get_ns_link_info(pid, perf_ns__name(idx),
 230                                              &ns_link_info[idx]);
 231
 232         event->namespaces.header.type = PERF_RECORD_NAMESPACES;
 233
 234         event->namespaces.header.size = (sizeof(event->namespaces) +
 235                         (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 236                         machine->id_hdr_size);
 237
 238         if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
 239                 return -1;
 240
 241         return 0;
 242 }
 243
 244 static int perf_event__synthesize_fork(struct perf_tool *tool,
 245                                        union perf_event *event,
 246                                        pid_t pid, pid_t tgid, pid_t ppid,
 247                                        perf_event__handler_t process,
 248                                        struct machine *machine)
 249 {
 250         memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
 251
 252         /*
 253          * for main thread set parent to ppid from status file. For other
 254          * threads set parent pid to main thread. ie., assume main thread
 255          * spawns all threads in a process
 256         */
 257         if (tgid == pid) {
 258                 event->fork.ppid = ppid;
 259                 event->fork.ptid = ppid;
 260         } else {
 261                 event->fork.ppid = tgid;
 262                 event->fork.ptid = tgid;
 263         }
 264         event->fork.pid  = tgid;
 265         event->fork.tid  = pid;
 266         event->fork.header.type = PERF_RECORD_FORK;
 267         event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC;
 268
 269         event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
 270
 271         if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
 272                 return -1;
 273
 274         return 0;
 275 }
 276
 277 static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end,
 278                                 u32 *prot, u32 *flags, __u64 *offset,
 279                                 u32 *maj, u32 *min,
 280                                 __u64 *inode,
 281                                 ssize_t pathname_size, char *pathname)
 282 {
 283         __u64 temp;
 284         int ch;
 285         char *start_pathname = pathname;
 286
 287         if (io__get_hex(io, start) != '-')
 288                 return false;
 289         if (io__get_hex(io, end) != ' ')
 290                 return false;
 291
 292         /* map protection and flags bits */
 293         *prot = 0;
 294         ch = io__get_char(io);
 295         if (ch == 'r')
 296                 *prot |= PROT_READ;
 297         else if (ch != '-')
 298                 return false;
 299         ch = io__get_char(io);
 300         if (ch == 'w')
 301                 *prot |= PROT_WRITE;
 302         else if (ch != '-')
 303                 return false;
 304         ch = io__get_char(io);
 305         if (ch == 'x')
 306                 *prot |= PROT_EXEC;
 307         else if (ch != '-')
 308                 return false;
 309         ch = io__get_char(io);
 310         if (ch == 's')
 311                 *flags = MAP_SHARED;
 312         else if (ch == 'p')
 313                 *flags = MAP_PRIVATE;
 314         else
 315                 return false;
 316         if (io__get_char(io) != ' ')
 317                 return false;
 318
 319         if (io__get_hex(io, offset) != ' ')
 320                 return false;
 321
 322         if (io__get_hex(io, &temp) != ':')
 323                 return false;
 324         *maj = temp;
 325         if (io__get_hex(io, &temp) != ' ')
 326                 return false;
 327         *min = temp;
 328
 329         ch = io__get_dec(io, inode);
 330         if (ch != ' ') {
 331                 *pathname = '\0';
 332                 return ch == '\n';
 333         }
 334         do {
 335                 ch = io__get_char(io);
 336         } while (ch == ' ');
 337         while (true) {
 338                 if (ch < 0)
 339                         return false;
 340                 if (ch == '\0' || ch == '\n' ||
 341                     (pathname + 1 - start_pathname) >= pathname_size) {
 342                         *pathname = '\0';
 343                         return true;
 344                 }
 345                 *pathname++ = ch;
 346                 ch = io__get_char(io);
 347         }
 348 }
 349
 350 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
 351                                        union perf_event *event,
 352                                        pid_t pid, pid_t tgid,
 353                                        perf_event__handler_t process,
 354                                        struct machine *machine,
 355                                        bool mmap_data)
 356 {
 357         unsigned long long t;
 358         char bf[BUFSIZ];
 359         struct io io;
 360         bool truncation = false;
 361         unsigned long long timeout = proc_map_timeout * 1000000ULL;
 362         int rc = 0;
 363         const char *hugetlbfs_mnt = hugetlbfs__mountpoint();
 364         int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0;
 365
 366         if (machine__is_default_guest(machine))
 367                 return 0;
 368
 369         snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps",
 370                 machine->root_dir, pid, pid);
 371
 372         io.fd = open(bf, O_RDONLY, 0);
 373         if (io.fd < 0) {
 374                 /*
 375                  * We raced with a task exiting - just return:
 376                  */
 377                 pr_debug("couldn't open %s\n", bf);
 378                 return -1;
 379         }
 380         io__init(&io, io.fd, bf, sizeof(bf));
 381
 382         event->header.type = PERF_RECORD_MMAP2;
 383         t = rdclock();
 384
 385         while (!io.eof) {
 386                 static const char anonstr[] = "//anon";
 387                 size_t size, aligned_size;
 388
 389                 /* ensure null termination since stack will be reused. */
 390                 event->mmap2.filename[0] = '\0';
 391
 392                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 393                 if (!read_proc_maps_line(&io,
 394                                         &event->mmap2.start,
 395                                         &event->mmap2.len,
 396                                         &event->mmap2.prot,
 397                                         &event->mmap2.flags,
 398                                         &event->mmap2.pgoff,
 399                                         &event->mmap2.maj,
 400                                         &event->mmap2.min,
 401                                         &event->mmap2.ino,
 402                                         sizeof(event->mmap2.filename),
 403                                         event->mmap2.filename))
 404                         continue;
 405
 406                 if ((rdclock() - t) > timeout) {
 407                         pr_warning("Reading %s/proc/%d/task/%d/maps time out. "
 408                                    "You may want to increase "
 409                                    "the time limit by --proc-map-timeout\n",
 410                                    machine->root_dir, pid, pid);
 411                         truncation = true;
 412                         goto out;
 413                 }
 414
 415                 event->mmap2.ino_generation = 0;
 416
 417                 /*
 418                  * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
 419                  */
 420                 if (machine__is_host(machine))
 421                         event->header.misc = PERF_RECORD_MISC_USER;
 422                 else
 423                         event->header.misc = PERF_RECORD_MISC_GUEST_USER;
 424
 425                 if ((event->mmap2.prot & PROT_EXEC) == 0) {
 426                         if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0)
 427                                 continue;
 428
 429                         event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
 430                 }
 431
 432 out:
 433                 if (truncation)
 434                         event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
 435
 436                 if (!strcmp(event->mmap2.filename, ""))
 437                         strcpy(event->mmap2.filename, anonstr);
 438
 439                 if (hugetlbfs_mnt_len &&
 440                     !strncmp(event->mmap2.filename, hugetlbfs_mnt,
 441                              hugetlbfs_mnt_len)) {
 442                         strcpy(event->mmap2.filename, anonstr);
 443                         event->mmap2.flags |= MAP_HUGETLB;
 444                 }
 445
 446                 size = strlen(event->mmap2.filename) + 1;
 447                 aligned_size = PERF_ALIGN(size, sizeof(u64));
 448                 event->mmap2.len -= event->mmap.start;
 449                 event->mmap2.header.size = (sizeof(event->mmap2) -
 450                                         (sizeof(event->mmap2.filename) - aligned_size));
 451                 memset(event->mmap2.filename + size, 0, machine->id_hdr_size +
 452                         (aligned_size - size));
 453                 event->mmap2.header.size += machine->id_hdr_size;
 454                 event->mmap2.pid = tgid;
 455                 event->mmap2.tid = pid;
 456
 457                 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
 458                         rc = -1;
 459                         break;
 460                 }
 461
 462                 if (truncation)
 463                         break;
 464         }
 465
 466         close(io.fd);
 467         return rc;
 468 }
 469
 470 #ifdef HAVE_FILE_HANDLE
 471 static int perf_event__synthesize_cgroup(struct perf_tool *tool,
 472                                          union perf_event *event,
 473                                          char *path, size_t mount_len,
 474                                          perf_event__handler_t process,
 475                                          struct machine *machine)
 476 {
 477         size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path);
 478         size_t path_len = strlen(path) - mount_len + 1;
 479         struct {
 480                 struct file_handle fh;
 481                 uint64_t cgroup_id;
 482         } handle;
 483         int mount_id;
 484
 485         while (path_len % sizeof(u64))
 486                 path[mount_len + path_len++] = '\0';
 487
 488         memset(&event->cgroup, 0, event_size);
 489
 490         event->cgroup.header.type = PERF_RECORD_CGROUP;
 491         event->cgroup.header.size = event_size + path_len + machine->id_hdr_size;
 492
 493         handle.fh.handle_bytes = sizeof(handle.cgroup_id);
 494         if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) {
 495                 pr_debug("stat failed: %s\n", path);
 496                 return -1;
 497         }
 498
 499         event->cgroup.id = handle.cgroup_id;
 500         strncpy(event->cgroup.path, path + mount_len, path_len);
 501         memset(event->cgroup.path + path_len, 0, machine->id_hdr_size);
 502
 503         if (perf_tool__process_synth_event(tool, event, machine, process) < 0) {
 504                 pr_debug("process synth event failed\n");
 505                 return -1;
 506         }
 507
 508         return 0;
 509 }
 510
 511 static int perf_event__walk_cgroup_tree(struct perf_tool *tool,
 512                                         union perf_event *event,
 513                                         char *path, size_t mount_len,
 514                                         perf_event__handler_t process,
 515                                         struct machine *machine)
 516 {
 517         size_t pos = strlen(path);
 518         DIR *d;
 519         struct dirent *dent;
 520         int ret = 0;
 521
 522         if (perf_event__synthesize_cgroup(tool, event, path, mount_len,
 523                                           process, machine) < 0)
 524                 return -1;
 525
 526         d = opendir(path);
 527         if (d == NULL) {
 528                 pr_debug("failed to open directory: %s\n", path);
 529                 return -1;
 530         }
 531
 532         while ((dent = readdir(d)) != NULL) {
 533                 if (dent->d_type != DT_DIR)
 534                         continue;
 535                 if (!strcmp(dent->d_name, ".") ||
 536                     !strcmp(dent->d_name, ".."))
 537                         continue;
 538
 539                 /* any sane path should be less than PATH_MAX */
 540                 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX)
 541                         continue;
 542
 543                 if (path[pos - 1] != '/')
 544                         strcat(path, "/");
 545                 strcat(path, dent->d_name);
 546
 547                 ret = perf_event__walk_cgroup_tree(tool, event, path,
 548                                                    mount_len, process, machine);
 549                 if (ret < 0)
 550                         break;
 551
 552                 path[pos] = '\0';
 553         }
 554
 555         closedir(d);
 556         return ret;
 557 }
 558
 559 int perf_event__synthesize_cgroups(struct perf_tool *tool,
 560                                    perf_event__handler_t process,
 561                                    struct machine *machine)
 562 {
 563         union perf_event event;
 564         char cgrp_root[PATH_MAX];
 565         size_t mount_len;  /* length of mount point in the path */
 566
 567         if (!tool || !tool->cgroup_events)
 568                 return 0;
 569
 570         if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) {
 571                 pr_debug("cannot find cgroup mount point\n");
 572                 return -1;
 573         }
 574
 575         mount_len = strlen(cgrp_root);
 576         /* make sure the path starts with a slash (after mount point) */
 577         strcat(cgrp_root, "/");
 578
 579         if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len,
 580                                          process, machine) < 0)
 581                 return -1;
 582
 583         return 0;
 584 }
 585 #else
 586 int perf_event__synthesize_cgroups(struct perf_tool *tool __maybe_unused,
 587                                    perf_event__handler_t process __maybe_unused,
 588                                    struct machine *machine __maybe_unused)
 589 {
 590         return -1;
 591 }
 592 #endif
 593
 594 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process,
 595                                    struct machine *machine)
 596 {
 597         int rc = 0;
 598         struct map *pos;
 599         struct maps *maps = machine__kernel_maps(machine);
 600         union perf_event *event = zalloc((sizeof(event->mmap) +
 601                                           machine->id_hdr_size));
 602         if (event == NULL) {
 603                 pr_debug("Not enough memory synthesizing mmap event "
 604                          "for kernel modules\n");
 605                 return -1;
 606         }
 607
 608         event->header.type = PERF_RECORD_MMAP;
 609
 610         /*
 611          * kernel uses 0 for user space maps, see kernel/perf_event.c
 612          * __perf_event_mmap
 613          */
 614         if (machine__is_host(machine))
 615                 event->header.misc = PERF_RECORD_MISC_KERNEL;
 616         else
 617                 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
 618
 619         maps__for_each_entry(maps, pos) {
 620                 size_t size;
 621
 622                 if (!__map__is_kmodule(pos))
 623                         continue;
 624
 625                 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
 626                 event->mmap.header.type = PERF_RECORD_MMAP;
 627                 event->mmap.header.size = (sizeof(event->mmap) -
 628                                         (sizeof(event->mmap.filename) - size));
 629                 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
 630                 event->mmap.header.size += machine->id_hdr_size;
 631                 event->mmap.start = pos->start;
 632                 event->mmap.len   = pos->end - pos->start;
 633                 event->mmap.pid   = machine->pid;
 634
 635                 memcpy(event->mmap.filename, pos->dso->long_name,
 636                        pos->dso->long_name_len + 1);
 637                 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
 638                         rc = -1;
 639                         break;
 640                 }
 641         }
 642
 643         free(event);
 644         return rc;
 645 }
 646
 647 static int __event__synthesize_thread(union perf_event *comm_event,
 648                                       union perf_event *mmap_event,
 649                                       union perf_event *fork_event,
 650                                       union perf_event *namespaces_event,
 651                                       pid_t pid, int full, perf_event__handler_t process,
 652                                       struct perf_tool *tool, struct machine *machine, bool mmap_data)
 653 {
 654         char filename[PATH_MAX];
 655         DIR *tasks;
 656         struct dirent *dirent;
 657         pid_t tgid, ppid;
 658         int rc = 0;
 659
 660         /* special case: only send one comm event using passed in pid */
 661         if (!full) {
 662                 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
 663                                                    process, machine);
 664
 665                 if (tgid == -1)
 666                         return -1;
 667
 668                 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
 669                                                       tgid, process, machine) < 0)
 670                         return -1;
 671
 672                 /*
 673                  * send mmap only for thread group leader
 674                  * see thread__init_maps()
 675                  */
 676                 if (pid == tgid &&
 677                     perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
 678                                                        process, machine, mmap_data))
 679                         return -1;
 680
 681                 return 0;
 682         }
 683
 684         if (machine__is_default_guest(machine))
 685                 return 0;
 686
 687         snprintf(filename, sizeof(filename), "%s/proc/%d/task",
 688                  machine->root_dir, pid);
 689
 690         tasks = opendir(filename);
 691         if (tasks == NULL) {
 692                 pr_debug("couldn't open %s\n", filename);
 693                 return 0;
 694         }
 695
 696         while ((dirent = readdir(tasks)) != NULL) {
 697                 char *end;
 698                 pid_t _pid;
 699
 700                 _pid = strtol(dirent->d_name, &end, 10);
 701                 if (*end)
 702                         continue;
 703
 704                 rc = -1;
 705                 if (perf_event__prepare_comm(comm_event, _pid, machine,
 706                                              &tgid, &ppid) != 0)
 707                         break;
 708
 709                 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
 710                                                 ppid, process, machine) < 0)
 711                         break;
 712
 713                 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
 714                                                       tgid, process, machine) < 0)
 715                         break;
 716
 717                 /*
 718                  * Send the prepared comm event
 719                  */
 720                 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
 721                         break;
 722
 723                 rc = 0;
 724                 if (_pid == pid) {
 725                         /* process the parent's maps too */
 726                         rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
 727                                                 process, machine, mmap_data);
 728                         if (rc)
 729                                 break;
 730                 }
 731         }
 732
 733         closedir(tasks);
 734         return rc;
 735 }
 736
 737 int perf_event__synthesize_thread_map(struct perf_tool *tool,
 738                                       struct perf_thread_map *threads,
 739                                       perf_event__handler_t process,
 740                                       struct machine *machine,
 741                                       bool mmap_data)
 742 {
 743         union perf_event *comm_event, *mmap_event, *fork_event;
 744         union perf_event *namespaces_event;
 745         int err = -1, thread, j;
 746
 747         comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
 748         if (comm_event == NULL)
 749                 goto out;
 750
 751         mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
 752         if (mmap_event == NULL)
 753                 goto out_free_comm;
 754
 755         fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
 756         if (fork_event == NULL)
 757                 goto out_free_mmap;
 758
 759         namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
 760                                   (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 761                                   machine->id_hdr_size);
 762         if (namespaces_event == NULL)
 763                 goto out_free_fork;
 764
 765         err = 0;
 766         for (thread = 0; thread < threads->nr; ++thread) {
 767                 if (__event__synthesize_thread(comm_event, mmap_event,
 768                                                fork_event, namespaces_event,
 769                                                perf_thread_map__pid(threads, thread), 0,
 770                                                process, tool, machine,
 771                                                mmap_data)) {
 772                         err = -1;
 773                         break;
 774                 }
 775
 776                 /*
 777                  * comm.pid is set to thread group id by
 778                  * perf_event__synthesize_comm
 779                  */
 780                 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) {
 781                         bool need_leader = true;
 782
 783                         /* is thread group leader in thread_map? */
 784                         for (j = 0; j < threads->nr; ++j) {
 785                                 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) {
 786                                         need_leader = false;
 787                                         break;
 788                                 }
 789                         }
 790
 791                         /* if not, generate events for it */
 792                         if (need_leader &&
 793                             __event__synthesize_thread(comm_event, mmap_event,
 794                                                        fork_event, namespaces_event,
 795                                                        comm_event->comm.pid, 0,
 796                                                        process, tool, machine,
 797                                                        mmap_data)) {
 798                                 err = -1;
 799                                 break;
 800                         }
 801                 }
 802         }
 803         free(namespaces_event);
 804 out_free_fork:
 805         free(fork_event);
 806 out_free_mmap:
 807         free(mmap_event);
 808 out_free_comm:
 809         free(comm_event);
 810 out:
 811         return err;
 812 }
 813
 814 static int __perf_event__synthesize_threads(struct perf_tool *tool,
 815                                             perf_event__handler_t process,
 816                                             struct machine *machine,
 817                                             bool mmap_data,
 818                                             struct dirent **dirent,
 819                                             int start,
 820                                             int num)
 821 {
 822         union perf_event *comm_event, *mmap_event, *fork_event;
 823         union perf_event *namespaces_event;
 824         int err = -1;
 825         char *end;
 826         pid_t pid;
 827         int i;
 828
 829         comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
 830         if (comm_event == NULL)
 831                 goto out;
 832
 833         mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
 834         if (mmap_event == NULL)
 835                 goto out_free_comm;
 836
 837         fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
 838         if (fork_event == NULL)
 839                 goto out_free_mmap;
 840
 841         namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
 842                                   (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 843                                   machine->id_hdr_size);
 844         if (namespaces_event == NULL)
 845                 goto out_free_fork;
 846
 847         for (i = start; i < start + num; i++) {
 848                 if (!isdigit(dirent[i]->d_name[0]))
 849                         continue;
 850
 851                 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
 852                 /* only interested in proper numerical dirents */
 853                 if (*end)
 854                         continue;
 855                 /*
 856                  * We may race with exiting thread, so don't stop just because
 857                  * one thread couldn't be synthesized.
 858                  */
 859                 __event__synthesize_thread(comm_event, mmap_event, fork_event,
 860                                            namespaces_event, pid, 1, process,
 861                                            tool, machine, mmap_data);
 862         }
 863         err = 0;
 864
 865         free(namespaces_event);
 866 out_free_fork:
 867         free(fork_event);
 868 out_free_mmap:
 869         free(mmap_event);
 870 out_free_comm:
 871         free(comm_event);
 872 out:
 873         return err;
 874 }
 875
 876 struct synthesize_threads_arg {
 877         struct perf_tool *tool;
 878         perf_event__handler_t process;
 879         struct machine *machine;
 880         bool mmap_data;
 881         struct dirent **dirent;
 882         int num;
 883         int start;
 884 };
 885
 886 static void *synthesize_threads_worker(void *arg)
 887 {
 888         struct synthesize_threads_arg *args = arg;
 889
 890         __perf_event__synthesize_threads(args->tool, args->process,
 891                                          args->machine, args->mmap_data,
 892                                          args->dirent,
 893                                          args->start, args->num);
 894         return NULL;
 895 }
 896
 897 int perf_event__synthesize_threads(struct perf_tool *tool,
 898                                    perf_event__handler_t process,
 899                                    struct machine *machine,
 900                                    bool mmap_data,
 901                                    unsigned int nr_threads_synthesize)
 902 {
 903         struct synthesize_threads_arg *args = NULL;
 904         pthread_t *synthesize_threads = NULL;
 905         char proc_path[PATH_MAX];
 906         struct dirent **dirent;
 907         int num_per_thread;
 908         int m, n, i, j;
 909         int thread_nr;
 910         int base = 0;
 911         int err = -1;
 912
 913
 914         if (machine__is_default_guest(machine))
 915                 return 0;
 916
 917         snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
 918         n = scandir(proc_path, &dirent, 0, alphasort);
 919         if (n < 0)
 920                 return err;
 921
 922         if (nr_threads_synthesize == UINT_MAX)
 923                 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
 924         else
 925                 thread_nr = nr_threads_synthesize;
 926
 927         if (thread_nr <= 1) {
 928                 err = __perf_event__synthesize_threads(tool, process,
 929                                                        machine, mmap_data,
 930                                                        dirent, base, n);
 931                 goto free_dirent;
 932         }
 933         if (thread_nr > n)
 934                 thread_nr = n;
 935
 936         synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
 937         if (synthesize_threads == NULL)
 938                 goto free_dirent;
 939
 940         args = calloc(sizeof(*args), thread_nr);
 941         if (args == NULL)
 942                 goto free_threads;
 943
 944         num_per_thread = n / thread_nr;
 945         m = n % thread_nr;
 946         for (i = 0; i < thread_nr; i++) {
 947                 args[i].tool = tool;
 948                 args[i].process = process;
 949                 args[i].machine = machine;
 950                 args[i].mmap_data = mmap_data;
 951                 args[i].dirent = dirent;
 952         }
 953         for (i = 0; i < m; i++) {
 954                 args[i].num = num_per_thread + 1;
 955                 args[i].start = i * args[i].num;
 956         }
 957         if (i != 0)
 958                 base = args[i-1].start + args[i-1].num;
 959         for (j = i; j < thread_nr; j++) {
 960                 args[j].num = num_per_thread;
 961                 args[j].start = base + (j - i) * args[i].num;
 962         }
 963
 964         for (i = 0; i < thread_nr; i++) {
 965                 if (pthread_create(&synthesize_threads[i], NULL,
 966                                    synthesize_threads_worker, &args[i]))
 967                         goto out_join;
 968         }
 969         err = 0;
 970 out_join:
 971         for (i = 0; i < thread_nr; i++)
 972                 pthread_join(synthesize_threads[i], NULL);
 973         free(args);
 974 free_threads:
 975         free(synthesize_threads);
 976 free_dirent:
 977         for (i = 0; i < n; i++)
 978                 zfree(&dirent[i]);
 979         free(dirent);
 980
 981         return err;
 982 }
 983
 984 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
 985                                               perf_event__handler_t process __maybe_unused,
 986                                               struct machine *machine __maybe_unused)
 987 {
 988         return 0;
 989 }
 990
 991 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
 992                                                 perf_event__handler_t process,
 993                                                 struct machine *machine)
 994 {
 995         size_t size;
 996         struct map *map = machine__kernel_map(machine);
 997         struct kmap *kmap;
 998         int err;
 999         union perf_event *event;
1000
1001         if (map == NULL)
1002                 return -1;
1003
1004         kmap = map__kmap(map);
1005         if (!kmap->ref_reloc_sym)
1006                 return -1;
1007
1008         /*
1009          * We should get this from /sys/kernel/sections/.text, but till that is
1010          * available use this, and after it is use this as a fallback for older
1011          * kernels.
1012          */
1013         event = zalloc((sizeof(event->mmap) + machine->id_hdr_size));
1014         if (event == NULL) {
1015                 pr_debug("Not enough memory synthesizing mmap event "
1016                          "for kernel modules\n");
1017                 return -1;
1018         }
1019
1020         if (machine__is_host(machine)) {
1021                 /*
1022                  * kernel uses PERF_RECORD_MISC_USER for user space maps,
1023                  * see kernel/perf_event.c __perf_event_mmap
1024                  */
1025                 event->header.misc = PERF_RECORD_MISC_KERNEL;
1026         } else {
1027                 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
1028         }
1029
1030         size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
1031                         "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
1032         size = PERF_ALIGN(size, sizeof(u64));
1033         event->mmap.header.type = PERF_RECORD_MMAP;
1034         event->mmap.header.size = (sizeof(event->mmap) -
1035                         (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
1036         event->mmap.pgoff = kmap->ref_reloc_sym->addr;
1037         event->mmap.start = map->start;
1038         event->mmap.len   = map->end - event->mmap.start;
1039         event->mmap.pid   = machine->pid;
1040
1041         err = perf_tool__process_synth_event(tool, event, machine, process);
1042         free(event);
1043
1044         return err;
1045 }
1046
1047 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
1048                                        perf_event__handler_t process,
1049                                        struct machine *machine)
1050 {
1051         int err;
1052
1053         err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
1054         if (err < 0)
1055                 return err;
1056
1057         return perf_event__synthesize_extra_kmaps(tool, process, machine);
1058 }
1059
1060 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
1061                                       struct perf_thread_map *threads,
1062                                       perf_event__handler_t process,
1063                                       struct machine *machine)
1064 {
1065         union perf_event *event;
1066         int i, err, size;
1067
1068         size  = sizeof(event->thread_map);
1069         size += threads->nr * sizeof(event->thread_map.entries[0]);
1070
1071         event = zalloc(size);
1072         if (!event)
1073                 return -ENOMEM;
1074
1075         event->header.type = PERF_RECORD_THREAD_MAP;
1076         event->header.size = size;
1077         event->thread_map.nr = threads->nr;
1078
1079         for (i = 0; i < threads->nr; i++) {
1080                 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i];
1081                 char *comm = perf_thread_map__comm(threads, i);
1082
1083                 if (!comm)
1084                         comm = (char *) "";
1085
1086                 entry->pid = perf_thread_map__pid(threads, i);
1087                 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
1088         }
1089
1090         err = process(tool, event, NULL, machine);
1091
1092         free(event);
1093         return err;
1094 }
1095
1096 static void synthesize_cpus(struct cpu_map_entries *cpus,
1097                             struct perf_cpu_map *map)
1098 {
1099         int i;
1100
1101         cpus->nr = map->nr;
1102
1103         for (i = 0; i < map->nr; i++)
1104                 cpus->cpu[i] = map->map[i];
1105 }
1106
1107 static void synthesize_mask(struct perf_record_record_cpu_map *mask,
1108                             struct perf_cpu_map *map, int max)
1109 {
1110         int i;
1111
1112         mask->nr = BITS_TO_LONGS(max);
1113         mask->long_size = sizeof(long);
1114
1115         for (i = 0; i < map->nr; i++)
1116                 set_bit(map->map[i], mask->mask);
1117 }
1118
1119 static size_t cpus_size(struct perf_cpu_map *map)
1120 {
1121         return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16);
1122 }
1123
1124 static size_t mask_size(struct perf_cpu_map *map, int *max)
1125 {
1126         int i;
1127
1128         *max = 0;
1129
1130         for (i = 0; i < map->nr; i++) {
1131                 /* bit possition of the cpu is + 1 */
1132                 int bit = map->map[i] + 1;
1133
1134                 if (bit > *max)
1135                         *max = bit;
1136         }
1137
1138         return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long);
1139 }
1140
1141 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max)
1142 {
1143         size_t size_cpus, size_mask;
1144         bool is_dummy = perf_cpu_map__empty(map);
1145
1146         /*
1147          * Both array and mask data have variable size based
1148          * on the number of cpus and their actual values.
1149          * The size of the 'struct perf_record_cpu_map_data' is:
1150          *
1151          *   array = size of 'struct cpu_map_entries' +
1152          *           number of cpus * sizeof(u64)
1153          *
1154          *   mask  = size of 'struct perf_record_record_cpu_map' +
1155          *           maximum cpu bit converted to size of longs
1156          *
1157          * and finaly + the size of 'struct perf_record_cpu_map_data'.
1158          */
1159         size_cpus = cpus_size(map);
1160         size_mask = mask_size(map, max);
1161
1162         if (is_dummy || (size_cpus < size_mask)) {
1163                 *size += size_cpus;
1164                 *type  = PERF_CPU_MAP__CPUS;
1165         } else {
1166                 *size += size_mask;
1167                 *type  = PERF_CPU_MAP__MASK;
1168         }
1169
1170         *size += sizeof(struct perf_record_cpu_map_data);
1171         *size = PERF_ALIGN(*size, sizeof(u64));
1172         return zalloc(*size);
1173 }
1174
1175 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
1176                               u16 type, int max)
1177 {
1178         data->type = type;
1179
1180         switch (type) {
1181         case PERF_CPU_MAP__CPUS:
1182                 synthesize_cpus((struct cpu_map_entries *) data->data, map);
1183                 break;
1184         case PERF_CPU_MAP__MASK:
1185                 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max);
1186         default:
1187                 break;
1188         }
1189 }
1190
1191 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map)
1192 {
1193         size_t size = sizeof(struct perf_record_cpu_map);
1194         struct perf_record_cpu_map *event;
1195         int max;
1196         u16 type;
1197
1198         event = cpu_map_data__alloc(map, &size, &type, &max);
1199         if (!event)
1200                 return NULL;
1201
1202         event->header.type = PERF_RECORD_CPU_MAP;
1203         event->header.size = size;
1204         event->data.type   = type;
1205
1206         cpu_map_data__synthesize(&event->data, map, type, max);
1207         return event;
1208 }
1209
1210 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1211                                    struct perf_cpu_map *map,
1212                                    perf_event__handler_t process,
1213                                    struct machine *machine)
1214 {
1215         struct perf_record_cpu_map *event;
1216         int err;
1217
1218         event = cpu_map_event__new(map);
1219         if (!event)
1220                 return -ENOMEM;
1221
1222         err = process(tool, (union perf_event *) event, NULL, machine);
1223
1224         free(event);
1225         return err;
1226 }
1227
1228 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1229                                        struct perf_stat_config *config,
1230                                        perf_event__handler_t process,
1231                                        struct machine *machine)
1232 {
1233         struct perf_record_stat_config *event;
1234         int size, i = 0, err;
1235
1236         size  = sizeof(*event);
1237         size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1238
1239         event = zalloc(size);
1240         if (!event)
1241                 return -ENOMEM;
1242
1243         event->header.type = PERF_RECORD_STAT_CONFIG;
1244         event->header.size = size;
1245         event->nr          = PERF_STAT_CONFIG_TERM__MAX;
1246
1247 #define ADD(__term, __val)                                      \
1248         event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term;   \
1249         event->data[i].val = __val;                             \
1250         i++;
1251
1252         ADD(AGGR_MODE,  config->aggr_mode)
1253         ADD(INTERVAL,   config->interval)
1254         ADD(SCALE,      config->scale)
1255
1256         WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1257                   "stat config terms unbalanced\n");
1258 #undef ADD
1259
1260         err = process(tool, (union perf_event *) event, NULL, machine);
1261
1262         free(event);
1263         return err;
1264 }
1265
1266 int perf_event__synthesize_stat(struct perf_tool *tool,
1267                                 u32 cpu, u32 thread, u64 id,
1268                                 struct perf_counts_values *count,
1269                                 perf_event__handler_t process,
1270                                 struct machine *machine)
1271 {
1272         struct perf_record_stat event;
1273
1274         event.header.type = PERF_RECORD_STAT;
1275         event.header.size = sizeof(event);
1276         event.header.misc = 0;
1277
1278         event.id        = id;
1279         event.cpu       = cpu;
1280         event.thread    = thread;
1281         event.val       = count->val;
1282         event.ena       = count->ena;
1283         event.run       = count->run;
1284
1285         return process(tool, (union perf_event *) &event, NULL, machine);
1286 }
1287
1288 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1289                                       u64 evtime, u64 type,
1290                                       perf_event__handler_t process,
1291                                       struct machine *machine)
1292 {
1293         struct perf_record_stat_round event;
1294
1295         event.header.type = PERF_RECORD_STAT_ROUND;
1296         event.header.size = sizeof(event);
1297         event.header.misc = 0;
1298
1299         event.time = evtime;
1300         event.type = type;
1301
1302         return process(tool, (union perf_event *) &event, NULL, machine);
1303 }
1304
1305 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format)
1306 {
1307         size_t sz, result = sizeof(struct perf_record_sample);
1308
1309         if (type & PERF_SAMPLE_IDENTIFIER)
1310                 result += sizeof(u64);
1311
1312         if (type & PERF_SAMPLE_IP)
1313                 result += sizeof(u64);
1314
1315         if (type & PERF_SAMPLE_TID)
1316                 result += sizeof(u64);
1317
1318         if (type & PERF_SAMPLE_TIME)
1319                 result += sizeof(u64);
1320
1321         if (type & PERF_SAMPLE_ADDR)
1322                 result += sizeof(u64);
1323
1324         if (type & PERF_SAMPLE_ID)
1325                 result += sizeof(u64);
1326
1327         if (type & PERF_SAMPLE_STREAM_ID)
1328                 result += sizeof(u64);
1329
1330         if (type & PERF_SAMPLE_CPU)
1331                 result += sizeof(u64);
1332
1333         if (type & PERF_SAMPLE_PERIOD)
1334                 result += sizeof(u64);
1335
1336         if (type & PERF_SAMPLE_READ) {
1337                 result += sizeof(u64);
1338                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1339                         result += sizeof(u64);
1340                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1341                         result += sizeof(u64);
1342                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1343                 if (read_format & PERF_FORMAT_GROUP) {
1344                         sz = sample->read.group.nr *
1345                              sizeof(struct sample_read_value);
1346                         result += sz;
1347                 } else {
1348                         result += sizeof(u64);
1349                 }
1350         }
1351
1352         if (type & PERF_SAMPLE_CALLCHAIN) {
1353                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1354                 result += sz;
1355         }
1356
1357         if (type & PERF_SAMPLE_RAW) {
1358                 result += sizeof(u32);
1359                 result += sample->raw_size;
1360         }
1361
1362         if (type & PERF_SAMPLE_BRANCH_STACK) {
1363                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1364                 /* nr, hw_idx */
1365                 sz += 2 * sizeof(u64);
1366                 result += sz;
1367         }
1368
1369         if (type & PERF_SAMPLE_REGS_USER) {
1370                 if (sample->user_regs.abi) {
1371                         result += sizeof(u64);
1372                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1373                         result += sz;
1374                 } else {
1375                         result += sizeof(u64);
1376                 }
1377         }
1378
1379         if (type & PERF_SAMPLE_STACK_USER) {
1380                 sz = sample->user_stack.size;
1381                 result += sizeof(u64);
1382                 if (sz) {
1383                         result += sz;
1384                         result += sizeof(u64);
1385                 }
1386         }
1387
1388         if (type & PERF_SAMPLE_WEIGHT)
1389                 result += sizeof(u64);
1390
1391         if (type & PERF_SAMPLE_DATA_SRC)
1392                 result += sizeof(u64);
1393
1394         if (type & PERF_SAMPLE_TRANSACTION)
1395                 result += sizeof(u64);
1396
1397         if (type & PERF_SAMPLE_REGS_INTR) {
1398                 if (sample->intr_regs.abi) {
1399                         result += sizeof(u64);
1400                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1401                         result += sz;
1402                 } else {
1403                         result += sizeof(u64);
1404                 }
1405         }
1406
1407         if (type & PERF_SAMPLE_PHYS_ADDR)
1408                 result += sizeof(u64);
1409
1410         if (type & PERF_SAMPLE_CGROUP)
1411                 result += sizeof(u64);
1412
1413         if (type & PERF_SAMPLE_AUX) {
1414                 result += sizeof(u64);
1415                 result += sample->aux_sample.size;
1416         }
1417
1418         return result;
1419 }
1420
1421 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format,
1422                                   const struct perf_sample *sample)
1423 {
1424         __u64 *array;
1425         size_t sz;
1426         /*
1427          * used for cross-endian analysis. See git commit 65014ab3
1428          * for why this goofiness is needed.
1429          */
1430         union u64_swap u;
1431
1432         array = event->sample.array;
1433
1434         if (type & PERF_SAMPLE_IDENTIFIER) {
1435                 *array = sample->id;
1436                 array++;
1437         }
1438
1439         if (type & PERF_SAMPLE_IP) {
1440                 *array = sample->ip;
1441                 array++;
1442         }
1443
1444         if (type & PERF_SAMPLE_TID) {
1445                 u.val32[0] = sample->pid;
1446                 u.val32[1] = sample->tid;
1447                 *array = u.val64;
1448                 array++;
1449         }
1450
1451         if (type & PERF_SAMPLE_TIME) {
1452                 *array = sample->time;
1453                 array++;
1454         }
1455
1456         if (type & PERF_SAMPLE_ADDR) {
1457                 *array = sample->addr;
1458                 array++;
1459         }
1460
1461         if (type & PERF_SAMPLE_ID) {
1462                 *array = sample->id;
1463                 array++;
1464         }
1465
1466         if (type & PERF_SAMPLE_STREAM_ID) {
1467                 *array = sample->stream_id;
1468                 array++;
1469         }
1470
1471         if (type & PERF_SAMPLE_CPU) {
1472                 u.val32[0] = sample->cpu;
1473                 u.val32[1] = 0;
1474                 *array = u.val64;
1475                 array++;
1476         }
1477
1478         if (type & PERF_SAMPLE_PERIOD) {
1479                 *array = sample->period;
1480                 array++;
1481         }
1482
1483         if (type & PERF_SAMPLE_READ) {
1484                 if (read_format & PERF_FORMAT_GROUP)
1485                         *array = sample->read.group.nr;
1486                 else
1487                         *array = sample->read.one.value;
1488                 array++;
1489
1490                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1491                         *array = sample->read.time_enabled;
1492                         array++;
1493                 }
1494
1495                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1496                         *array = sample->read.time_running;
1497                         array++;
1498                 }
1499
1500                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1501                 if (read_format & PERF_FORMAT_GROUP) {
1502                         sz = sample->read.group.nr *
1503                              sizeof(struct sample_read_value);
1504                         memcpy(array, sample->read.group.values, sz);
1505                         array = (void *)array + sz;
1506                 } else {
1507                         *array = sample->read.one.id;
1508                         array++;
1509                 }
1510         }
1511
1512         if (type & PERF_SAMPLE_CALLCHAIN) {
1513                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1514                 memcpy(array, sample->callchain, sz);
1515                 array = (void *)array + sz;
1516         }
1517
1518         if (type & PERF_SAMPLE_RAW) {
1519                 u.val32[0] = sample->raw_size;
1520                 *array = u.val64;
1521                 array = (void *)array + sizeof(u32);
1522
1523                 memcpy(array, sample->raw_data, sample->raw_size);
1524                 array = (void *)array + sample->raw_size;
1525         }
1526
1527         if (type & PERF_SAMPLE_BRANCH_STACK) {
1528                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1529                 /* nr, hw_idx */
1530                 sz += 2 * sizeof(u64);
1531                 memcpy(array, sample->branch_stack, sz);
1532                 array = (void *)array + sz;
1533         }
1534
1535         if (type & PERF_SAMPLE_REGS_USER) {
1536                 if (sample->user_regs.abi) {
1537                         *array++ = sample->user_regs.abi;
1538                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1539                         memcpy(array, sample->user_regs.regs, sz);
1540                         array = (void *)array + sz;
1541                 } else {
1542                         *array++ = 0;
1543                 }
1544         }
1545
1546         if (type & PERF_SAMPLE_STACK_USER) {
1547                 sz = sample->user_stack.size;
1548                 *array++ = sz;
1549                 if (sz) {
1550                         memcpy(array, sample->user_stack.data, sz);
1551                         array = (void *)array + sz;
1552                         *array++ = sz;
1553                 }
1554         }
1555
1556         if (type & PERF_SAMPLE_WEIGHT) {
1557                 *array = sample->weight;
1558                 array++;
1559         }
1560
1561         if (type & PERF_SAMPLE_DATA_SRC) {
1562                 *array = sample->data_src;
1563                 array++;
1564         }
1565
1566         if (type & PERF_SAMPLE_TRANSACTION) {
1567                 *array = sample->transaction;
1568                 array++;
1569         }
1570
1571         if (type & PERF_SAMPLE_REGS_INTR) {
1572                 if (sample->intr_regs.abi) {
1573                         *array++ = sample->intr_regs.abi;
1574                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1575                         memcpy(array, sample->intr_regs.regs, sz);
1576                         array = (void *)array + sz;
1577                 } else {
1578                         *array++ = 0;
1579                 }
1580         }
1581
1582         if (type & PERF_SAMPLE_PHYS_ADDR) {
1583                 *array = sample->phys_addr;
1584                 array++;
1585         }
1586
1587         if (type & PERF_SAMPLE_CGROUP) {
1588                 *array = sample->cgroup;
1589                 array++;
1590         }
1591
1592         if (type & PERF_SAMPLE_AUX) {
1593                 sz = sample->aux_sample.size;
1594                 *array++ = sz;
1595                 memcpy(array, sample->aux_sample.data, sz);
1596                 array = (void *)array + sz;
1597         }
1598
1599         return 0;
1600 }
1601
1602 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process,
1603                                     struct evlist *evlist, struct machine *machine)
1604 {
1605         union perf_event *ev;
1606         struct evsel *evsel;
1607         size_t nr = 0, i = 0, sz, max_nr, n;
1608         int err;
1609
1610         pr_debug2("Synthesizing id index\n");
1611
1612         max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) /
1613                  sizeof(struct id_index_entry);
1614
1615         evlist__for_each_entry(evlist, evsel)
1616                 nr += evsel->core.ids;
1617
1618         n = nr > max_nr ? max_nr : nr;
1619         sz = sizeof(struct perf_record_id_index) + n * sizeof(struct id_index_entry);
1620         ev = zalloc(sz);
1621         if (!ev)
1622                 return -ENOMEM;
1623
1624         ev->id_index.header.type = PERF_RECORD_ID_INDEX;
1625         ev->id_index.header.size = sz;
1626         ev->id_index.nr = n;
1627
1628         evlist__for_each_entry(evlist, evsel) {
1629                 u32 j;
1630
1631                 for (j = 0; j < evsel->core.ids; j++) {
1632                         struct id_index_entry *e;
1633                         struct perf_sample_id *sid;
1634
1635                         if (i >= n) {
1636                                 err = process(tool, ev, NULL, machine);
1637                                 if (err)
1638                                         goto out_err;
1639                                 nr -= n;
1640                                 i = 0;
1641                         }
1642
1643                         e = &ev->id_index.entries[i++];
1644
1645                         e->id = evsel->core.id[j];
1646
1647                         sid = perf_evlist__id2sid(evlist, e->id);
1648                         if (!sid) {
1649                                 free(ev);
1650                                 return -ENOENT;
1651                         }
1652
1653                         e->idx = sid->idx;
1654                         e->cpu = sid->cpu;
1655                         e->tid = sid->tid;
1656                 }
1657         }
1658
1659         sz = sizeof(struct perf_record_id_index) + nr * sizeof(struct id_index_entry);
1660         ev->id_index.header.size = sz;
1661         ev->id_index.nr = nr;
1662
1663         err = process(tool, ev, NULL, machine);
1664 out_err:
1665         free(ev);
1666
1667         return err;
1668 }
1669
1670 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1671                                   struct target *target, struct perf_thread_map *threads,
1672                                   perf_event__handler_t process, bool data_mmap,
1673                                   unsigned int nr_threads_synthesize)
1674 {
1675         if (target__has_task(target))
1676                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
1677         else if (target__has_cpu(target))
1678                 return perf_event__synthesize_threads(tool, process,
1679                                                       machine, data_mmap,
1680                                                       nr_threads_synthesize);
1681         /* command specified */
1682         return 0;
1683 }
1684
1685 int machine__synthesize_threads(struct machine *machine, struct target *target,
1686                                 struct perf_thread_map *threads, bool data_mmap,
1687                                 unsigned int nr_threads_synthesize)
1688 {
1689         return __machine__synthesize_threads(machine, NULL, target, threads,
1690                                              perf_event__process, data_mmap,
1691                                              nr_threads_synthesize);
1692 }
1693
1694 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id)
1695 {
1696         struct perf_record_event_update *ev;
1697
1698         size += sizeof(*ev);
1699         size  = PERF_ALIGN(size, sizeof(u64));
1700
1701         ev = zalloc(size);
1702         if (ev) {
1703                 ev->header.type = PERF_RECORD_EVENT_UPDATE;
1704                 ev->header.size = (u16)size;
1705                 ev->type        = type;
1706                 ev->id          = id;
1707         }
1708         return ev;
1709 }
1710
1711 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel,
1712                                              perf_event__handler_t process)
1713 {
1714         size_t size = strlen(evsel->unit);
1715         struct perf_record_event_update *ev;
1716         int err;
1717
1718         ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]);
1719         if (ev == NULL)
1720                 return -ENOMEM;
1721
1722         strlcpy(ev->data, evsel->unit, size + 1);
1723         err = process(tool, (union perf_event *)ev, NULL, NULL);
1724         free(ev);
1725         return err;
1726 }
1727
1728 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel,
1729                                               perf_event__handler_t process)
1730 {
1731         struct perf_record_event_update *ev;
1732         struct perf_record_event_update_scale *ev_data;
1733         int err;
1734
1735         ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]);
1736         if (ev == NULL)
1737                 return -ENOMEM;
1738
1739         ev_data = (struct perf_record_event_update_scale *)ev->data;
1740         ev_data->scale = evsel->scale;
1741         err = process(tool, (union perf_event *)ev, NULL, NULL);
1742         free(ev);
1743         return err;
1744 }
1745
1746 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel,
1747                                              perf_event__handler_t process)
1748 {
1749         struct perf_record_event_update *ev;
1750         size_t len = strlen(evsel->name);
1751         int err;
1752
1753         ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]);
1754         if (ev == NULL)
1755                 return -ENOMEM;
1756
1757         strlcpy(ev->data, evsel->name, len + 1);
1758         err = process(tool, (union perf_event *)ev, NULL, NULL);
1759         free(ev);
1760         return err;
1761 }
1762
1763 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel,
1764                                              perf_event__handler_t process)
1765 {
1766         size_t size = sizeof(struct perf_record_event_update);
1767         struct perf_record_event_update *ev;
1768         int max, err;
1769         u16 type;
1770
1771         if (!evsel->core.own_cpus)
1772                 return 0;
1773
1774         ev = cpu_map_data__alloc(evsel->core.own_cpus, &size, &type, &max);
1775         if (!ev)
1776                 return -ENOMEM;
1777
1778         ev->header.type = PERF_RECORD_EVENT_UPDATE;
1779         ev->header.size = (u16)size;
1780         ev->type        = PERF_EVENT_UPDATE__CPUS;
1781         ev->id          = evsel->core.id[0];
1782
1783         cpu_map_data__synthesize((struct perf_record_cpu_map_data *)ev->data,
1784                                  evsel->core.own_cpus, type, max);
1785
1786         err = process(tool, (union perf_event *)ev, NULL, NULL);
1787         free(ev);
1788         return err;
1789 }
1790
1791 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist,
1792                                  perf_event__handler_t process)
1793 {
1794         struct evsel *evsel;
1795         int err = 0;
1796
1797         evlist__for_each_entry(evlist, evsel) {
1798                 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids,
1799                                                   evsel->core.id, process);
1800                 if (err) {
1801                         pr_debug("failed to create perf header attribute\n");
1802                         return err;
1803                 }
1804         }
1805
1806         return err;
1807 }
1808
1809 static bool has_unit(struct evsel *evsel)
1810 {
1811         return evsel->unit && *evsel->unit;
1812 }
1813
1814 static bool has_scale(struct evsel *evsel)
1815 {
1816         return evsel->scale != 1;
1817 }
1818
1819 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list,
1820                                       perf_event__handler_t process, bool is_pipe)
1821 {
1822         struct evsel *evsel;
1823         int err;
1824
1825         /*
1826          * Synthesize other events stuff not carried within
1827          * attr event - unit, scale, name
1828          */
1829         evlist__for_each_entry(evsel_list, evsel) {
1830                 if (!evsel->supported)
1831                         continue;
1832
1833                 /*
1834                  * Synthesize unit and scale only if it's defined.
1835                  */
1836                 if (has_unit(evsel)) {
1837                         err = perf_event__synthesize_event_update_unit(tool, evsel, process);
1838                         if (err < 0) {
1839                                 pr_err("Couldn't synthesize evsel unit.\n");
1840                                 return err;
1841                         }
1842                 }
1843
1844                 if (has_scale(evsel)) {
1845                         err = perf_event__synthesize_event_update_scale(tool, evsel, process);
1846                         if (err < 0) {
1847                                 pr_err("Couldn't synthesize evsel evsel.\n");
1848                                 return err;
1849                         }
1850                 }
1851
1852                 if (evsel->core.own_cpus) {
1853                         err = perf_event__synthesize_event_update_cpus(tool, evsel, process);
1854                         if (err < 0) {
1855                                 pr_err("Couldn't synthesize evsel cpus.\n");
1856                                 return err;
1857                         }
1858                 }
1859
1860                 /*
1861                  * Name is needed only for pipe output,
1862                  * perf.data carries event names.
1863                  */
1864                 if (is_pipe) {
1865                         err = perf_event__synthesize_event_update_name(tool, evsel, process);
1866                         if (err < 0) {
1867                                 pr_err("Couldn't synthesize evsel name.\n");
1868                                 return err;
1869                         }
1870                 }
1871         }
1872         return 0;
1873 }
1874
1875 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr,
1876                                 u32 ids, u64 *id, perf_event__handler_t process)
1877 {
1878         union perf_event *ev;
1879         size_t size;
1880         int err;
1881
1882         size = sizeof(struct perf_event_attr);
1883         size = PERF_ALIGN(size, sizeof(u64));
1884         size += sizeof(struct perf_event_header);
1885         size += ids * sizeof(u64);
1886
1887         ev = zalloc(size);
1888
1889         if (ev == NULL)
1890                 return -ENOMEM;
1891
1892         ev->attr.attr = *attr;
1893         memcpy(ev->attr.id, id, ids * sizeof(u64));
1894
1895         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
1896         ev->attr.header.size = (u16)size;
1897
1898         if (ev->attr.header.size == size)
1899                 err = process(tool, ev, NULL, NULL);
1900         else
1901                 err = -E2BIG;
1902
1903         free(ev);
1904
1905         return err;
1906 }
1907
1908 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist,
1909                                         perf_event__handler_t process)
1910 {
1911         union perf_event ev;
1912         struct tracing_data *tdata;
1913         ssize_t size = 0, aligned_size = 0, padding;
1914         struct feat_fd ff;
1915
1916         /*
1917          * We are going to store the size of the data followed
1918          * by the data contents. Since the fd descriptor is a pipe,
1919          * we cannot seek back to store the size of the data once
1920          * we know it. Instead we:
1921          *
1922          * - write the tracing data to the temp file
1923          * - get/write the data size to pipe
1924          * - write the tracing data from the temp file
1925          *   to the pipe
1926          */
1927         tdata = tracing_data_get(&evlist->core.entries, fd, true);
1928         if (!tdata)
1929                 return -1;
1930
1931         memset(&ev, 0, sizeof(ev));
1932
1933         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
1934         size = tdata->size;
1935         aligned_size = PERF_ALIGN(size, sizeof(u64));
1936         padding = aligned_size - size;
1937         ev.tracing_data.header.size = sizeof(ev.tracing_data);
1938         ev.tracing_data.size = aligned_size;
1939
1940         process(tool, &ev, NULL, NULL);
1941
1942         /*
1943          * The put function will copy all the tracing data
1944          * stored in temp file to the pipe.
1945          */
1946         tracing_data_put(tdata);
1947
1948         ff = (struct feat_fd){ .fd = fd };
1949         if (write_padded(&ff, NULL, 0, padding))
1950                 return -1;
1951
1952         return aligned_size;
1953 }
1954
1955 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc,
1956                                     perf_event__handler_t process, struct machine *machine)
1957 {
1958         union perf_event ev;
1959         size_t len;
1960
1961         if (!pos->hit)
1962                 return 0;
1963
1964         memset(&ev, 0, sizeof(ev));
1965
1966         len = pos->long_name_len + 1;
1967         len = PERF_ALIGN(len, NAME_ALIGN);
1968         memcpy(&ev.build_id.build_id, pos->bid.data, sizeof(pos->bid.data));
1969         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
1970         ev.build_id.header.misc = misc;
1971         ev.build_id.pid = machine->pid;
1972         ev.build_id.header.size = sizeof(ev.build_id) + len;
1973         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
1974
1975         return process(tool, &ev, NULL, machine);
1976 }
1977
1978 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool,
1979                                        struct evlist *evlist, perf_event__handler_t process, bool attrs)
1980 {
1981         int err;
1982
1983         if (attrs) {
1984                 err = perf_event__synthesize_attrs(tool, evlist, process);
1985                 if (err < 0) {
1986                         pr_err("Couldn't synthesize attrs.\n");
1987                         return err;
1988                 }
1989         }
1990
1991         err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs);
1992         err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL);
1993         if (err < 0) {
1994                 pr_err("Couldn't synthesize thread map.\n");
1995                 return err;
1996         }
1997
1998         err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus, process, NULL);
1999         if (err < 0) {
2000                 pr_err("Couldn't synthesize thread map.\n");
2001                 return err;
2002         }
2003
2004         err = perf_event__synthesize_stat_config(tool, config, process, NULL);
2005         if (err < 0) {
2006                 pr_err("Couldn't synthesize config.\n");
2007                 return err;
2008         }
2009
2010         return 0;
2011 }
2012
2013 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
2014
2015 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session,
2016                                     struct evlist *evlist, perf_event__handler_t process)
2017 {
2018         struct perf_header *header = &session->header;
2019         struct perf_record_header_feature *fe;
2020         struct feat_fd ff;
2021         size_t sz, sz_hdr;
2022         int feat, ret;
2023
2024         sz_hdr = sizeof(fe->header);
2025         sz = sizeof(union perf_event);
2026         /* get a nice alignment */
2027         sz = PERF_ALIGN(sz, page_size);
2028
2029         memset(&ff, 0, sizeof(ff));
2030
2031         ff.buf = malloc(sz);
2032         if (!ff.buf)
2033                 return -ENOMEM;
2034
2035         ff.size = sz - sz_hdr;
2036         ff.ph = &session->header;
2037
2038         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2039                 if (!feat_ops[feat].synthesize) {
2040                         pr_debug("No record header feature for header :%d\n", feat);
2041                         continue;
2042                 }
2043
2044                 ff.offset = sizeof(*fe);
2045
2046                 ret = feat_ops[feat].write(&ff, evlist);
2047                 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
2048                         pr_debug("Error writing feature\n");
2049                         continue;
2050                 }
2051                 /* ff.buf may have changed due to realloc in do_write() */
2052                 fe = ff.buf;
2053                 memset(fe, 0, sizeof(*fe));
2054
2055                 fe->feat_id = feat;
2056                 fe->header.type = PERF_RECORD_HEADER_FEATURE;
2057                 fe->header.size = ff.offset;
2058
2059                 ret = process(tool, ff.buf, NULL, NULL);
2060                 if (ret) {
2061                         free(ff.buf);
2062                         return ret;
2063                 }
2064         }
2065
2066         /* Send HEADER_LAST_FEATURE mark. */
2067         fe = ff.buf;
2068         fe->feat_id     = HEADER_LAST_FEATURE;
2069         fe->header.type = PERF_RECORD_HEADER_FEATURE;
2070         fe->header.size = sizeof(*fe);
2071
2072         ret = process(tool, ff.buf, NULL, NULL);
2073
2074         free(ff.buf);
2075         return ret;
2076 }