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;
 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                 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) - size));
 451                 memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
 452                 event->mmap2.header.size += machine->id_hdr_size;
 453                 event->mmap2.pid = tgid;
 454                 event->mmap2.tid = pid;
 455
 456                 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
 457                         rc = -1;
 458                         break;
 459                 }
 460
 461                 if (truncation)
 462                         break;
 463         }
 464
 465         close(io.fd);
 466         return rc;
 467 }
 468
 469 #ifdef HAVE_FILE_HANDLE
 470 static int perf_event__synthesize_cgroup(struct perf_tool *tool,
 471                                          union perf_event *event,
 472                                          char *path, size_t mount_len,
 473                                          perf_event__handler_t process,
 474                                          struct machine *machine)
 475 {
 476         size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path);
 477         size_t path_len = strlen(path) - mount_len + 1;
 478         struct {
 479                 struct file_handle fh;
 480                 uint64_t cgroup_id;
 481         } handle;
 482         int mount_id;
 483
 484         while (path_len % sizeof(u64))
 485                 path[mount_len + path_len++] = '\0';
 486
 487         memset(&event->cgroup, 0, event_size);
 488
 489         event->cgroup.header.type = PERF_RECORD_CGROUP;
 490         event->cgroup.header.size = event_size + path_len + machine->id_hdr_size;
 491
 492         handle.fh.handle_bytes = sizeof(handle.cgroup_id);
 493         if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) {
 494                 pr_debug("stat failed: %s\n", path);
 495                 return -1;
 496         }
 497
 498         event->cgroup.id = handle.cgroup_id;
 499         strncpy(event->cgroup.path, path + mount_len, path_len);
 500         memset(event->cgroup.path + path_len, 0, machine->id_hdr_size);
 501
 502         if (perf_tool__process_synth_event(tool, event, machine, process) < 0) {
 503                 pr_debug("process synth event failed\n");
 504                 return -1;
 505         }
 506
 507         return 0;
 508 }
 509
 510 static int perf_event__walk_cgroup_tree(struct perf_tool *tool,
 511                                         union perf_event *event,
 512                                         char *path, size_t mount_len,
 513                                         perf_event__handler_t process,
 514                                         struct machine *machine)
 515 {
 516         size_t pos = strlen(path);
 517         DIR *d;
 518         struct dirent *dent;
 519         int ret = 0;
 520
 521         if (perf_event__synthesize_cgroup(tool, event, path, mount_len,
 522                                           process, machine) < 0)
 523                 return -1;
 524
 525         d = opendir(path);
 526         if (d == NULL) {
 527                 pr_debug("failed to open directory: %s\n", path);
 528                 return -1;
 529         }
 530
 531         while ((dent = readdir(d)) != NULL) {
 532                 if (dent->d_type != DT_DIR)
 533                         continue;
 534                 if (!strcmp(dent->d_name, ".") ||
 535                     !strcmp(dent->d_name, ".."))
 536                         continue;
 537
 538                 /* any sane path should be less than PATH_MAX */
 539                 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX)
 540                         continue;
 541
 542                 if (path[pos - 1] != '/')
 543                         strcat(path, "/");
 544                 strcat(path, dent->d_name);
 545
 546                 ret = perf_event__walk_cgroup_tree(tool, event, path,
 547                                                    mount_len, process, machine);
 548                 if (ret < 0)
 549                         break;
 550
 551                 path[pos] = '\0';
 552         }
 553
 554         closedir(d);
 555         return ret;
 556 }
 557
 558 int perf_event__synthesize_cgroups(struct perf_tool *tool,
 559                                    perf_event__handler_t process,
 560                                    struct machine *machine)
 561 {
 562         union perf_event event;
 563         char cgrp_root[PATH_MAX];
 564         size_t mount_len;  /* length of mount point in the path */
 565
 566         if (!tool || !tool->cgroup_events)
 567                 return 0;
 568
 569         if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) {
 570                 pr_debug("cannot find cgroup mount point\n");
 571                 return -1;
 572         }
 573
 574         mount_len = strlen(cgrp_root);
 575         /* make sure the path starts with a slash (after mount point) */
 576         strcat(cgrp_root, "/");
 577
 578         if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len,
 579                                          process, machine) < 0)
 580                 return -1;
 581
 582         return 0;
 583 }
 584 #else
 585 int perf_event__synthesize_cgroups(struct perf_tool *tool __maybe_unused,
 586                                    perf_event__handler_t process __maybe_unused,
 587                                    struct machine *machine __maybe_unused)
 588 {
 589         return -1;
 590 }
 591 #endif
 592
 593 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process,
 594                                    struct machine *machine)
 595 {
 596         int rc = 0;
 597         struct map *pos;
 598         struct maps *maps = machine__kernel_maps(machine);
 599         union perf_event *event = zalloc((sizeof(event->mmap) +
 600                                           machine->id_hdr_size));
 601         if (event == NULL) {
 602                 pr_debug("Not enough memory synthesizing mmap event "
 603                          "for kernel modules\n");
 604                 return -1;
 605         }
 606
 607         event->header.type = PERF_RECORD_MMAP;
 608
 609         /*
 610          * kernel uses 0 for user space maps, see kernel/perf_event.c
 611          * __perf_event_mmap
 612          */
 613         if (machine__is_host(machine))
 614                 event->header.misc = PERF_RECORD_MISC_KERNEL;
 615         else
 616                 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
 617
 618         maps__for_each_entry(maps, pos) {
 619                 size_t size;
 620
 621                 if (!__map__is_kmodule(pos))
 622                         continue;
 623
 624                 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
 625                 event->mmap.header.type = PERF_RECORD_MMAP;
 626                 event->mmap.header.size = (sizeof(event->mmap) -
 627                                         (sizeof(event->mmap.filename) - size));
 628                 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
 629                 event->mmap.header.size += machine->id_hdr_size;
 630                 event->mmap.start = pos->start;
 631                 event->mmap.len   = pos->end - pos->start;
 632                 event->mmap.pid   = machine->pid;
 633
 634                 memcpy(event->mmap.filename, pos->dso->long_name,
 635                        pos->dso->long_name_len + 1);
 636                 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) {
 637                         rc = -1;
 638                         break;
 639                 }
 640         }
 641
 642         free(event);
 643         return rc;
 644 }
 645
 646 static int __event__synthesize_thread(union perf_event *comm_event,
 647                                       union perf_event *mmap_event,
 648                                       union perf_event *fork_event,
 649                                       union perf_event *namespaces_event,
 650                                       pid_t pid, int full, perf_event__handler_t process,
 651                                       struct perf_tool *tool, struct machine *machine, bool mmap_data)
 652 {
 653         char filename[PATH_MAX];
 654         DIR *tasks;
 655         struct dirent *dirent;
 656         pid_t tgid, ppid;
 657         int rc = 0;
 658
 659         /* special case: only send one comm event using passed in pid */
 660         if (!full) {
 661                 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
 662                                                    process, machine);
 663
 664                 if (tgid == -1)
 665                         return -1;
 666
 667                 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid,
 668                                                       tgid, process, machine) < 0)
 669                         return -1;
 670
 671                 /*
 672                  * send mmap only for thread group leader
 673                  * see thread__init_maps()
 674                  */
 675                 if (pid == tgid &&
 676                     perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
 677                                                        process, machine, mmap_data))
 678                         return -1;
 679
 680                 return 0;
 681         }
 682
 683         if (machine__is_default_guest(machine))
 684                 return 0;
 685
 686         snprintf(filename, sizeof(filename), "%s/proc/%d/task",
 687                  machine->root_dir, pid);
 688
 689         tasks = opendir(filename);
 690         if (tasks == NULL) {
 691                 pr_debug("couldn't open %s\n", filename);
 692                 return 0;
 693         }
 694
 695         while ((dirent = readdir(tasks)) != NULL) {
 696                 char *end;
 697                 pid_t _pid;
 698
 699                 _pid = strtol(dirent->d_name, &end, 10);
 700                 if (*end)
 701                         continue;
 702
 703                 rc = -1;
 704                 if (perf_event__prepare_comm(comm_event, _pid, machine,
 705                                              &tgid, &ppid) != 0)
 706                         break;
 707
 708                 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
 709                                                 ppid, process, machine) < 0)
 710                         break;
 711
 712                 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid,
 713                                                       tgid, process, machine) < 0)
 714                         break;
 715
 716                 /*
 717                  * Send the prepared comm event
 718                  */
 719                 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0)
 720                         break;
 721
 722                 rc = 0;
 723                 if (_pid == pid) {
 724                         /* process the parent's maps too */
 725                         rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
 726                                                 process, machine, mmap_data);
 727                         if (rc)
 728                                 break;
 729                 }
 730         }
 731
 732         closedir(tasks);
 733         return rc;
 734 }
 735
 736 int perf_event__synthesize_thread_map(struct perf_tool *tool,
 737                                       struct perf_thread_map *threads,
 738                                       perf_event__handler_t process,
 739                                       struct machine *machine,
 740                                       bool mmap_data)
 741 {
 742         union perf_event *comm_event, *mmap_event, *fork_event;
 743         union perf_event *namespaces_event;
 744         int err = -1, thread, j;
 745
 746         comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
 747         if (comm_event == NULL)
 748                 goto out;
 749
 750         mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
 751         if (mmap_event == NULL)
 752                 goto out_free_comm;
 753
 754         fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
 755         if (fork_event == NULL)
 756                 goto out_free_mmap;
 757
 758         namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
 759                                   (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 760                                   machine->id_hdr_size);
 761         if (namespaces_event == NULL)
 762                 goto out_free_fork;
 763
 764         err = 0;
 765         for (thread = 0; thread < threads->nr; ++thread) {
 766                 if (__event__synthesize_thread(comm_event, mmap_event,
 767                                                fork_event, namespaces_event,
 768                                                perf_thread_map__pid(threads, thread), 0,
 769                                                process, tool, machine,
 770                                                mmap_data)) {
 771                         err = -1;
 772                         break;
 773                 }
 774
 775                 /*
 776                  * comm.pid is set to thread group id by
 777                  * perf_event__synthesize_comm
 778                  */
 779                 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) {
 780                         bool need_leader = true;
 781
 782                         /* is thread group leader in thread_map? */
 783                         for (j = 0; j < threads->nr; ++j) {
 784                                 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) {
 785                                         need_leader = false;
 786                                         break;
 787                                 }
 788                         }
 789
 790                         /* if not, generate events for it */
 791                         if (need_leader &&
 792                             __event__synthesize_thread(comm_event, mmap_event,
 793                                                        fork_event, namespaces_event,
 794                                                        comm_event->comm.pid, 0,
 795                                                        process, tool, machine,
 796                                                        mmap_data)) {
 797                                 err = -1;
 798                                 break;
 799                         }
 800                 }
 801         }
 802         free(namespaces_event);
 803 out_free_fork:
 804         free(fork_event);
 805 out_free_mmap:
 806         free(mmap_event);
 807 out_free_comm:
 808         free(comm_event);
 809 out:
 810         return err;
 811 }
 812
 813 static int __perf_event__synthesize_threads(struct perf_tool *tool,
 814                                             perf_event__handler_t process,
 815                                             struct machine *machine,
 816                                             bool mmap_data,
 817                                             struct dirent **dirent,
 818                                             int start,
 819                                             int num)
 820 {
 821         union perf_event *comm_event, *mmap_event, *fork_event;
 822         union perf_event *namespaces_event;
 823         int err = -1;
 824         char *end;
 825         pid_t pid;
 826         int i;
 827
 828         comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
 829         if (comm_event == NULL)
 830                 goto out;
 831
 832         mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size);
 833         if (mmap_event == NULL)
 834                 goto out_free_comm;
 835
 836         fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
 837         if (fork_event == NULL)
 838                 goto out_free_mmap;
 839
 840         namespaces_event = malloc(sizeof(namespaces_event->namespaces) +
 841                                   (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 842                                   machine->id_hdr_size);
 843         if (namespaces_event == NULL)
 844                 goto out_free_fork;
 845
 846         for (i = start; i < start + num; i++) {
 847                 if (!isdigit(dirent[i]->d_name[0]))
 848                         continue;
 849
 850                 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10);
 851                 /* only interested in proper numerical dirents */
 852                 if (*end)
 853                         continue;
 854                 /*
 855                  * We may race with exiting thread, so don't stop just because
 856                  * one thread couldn't be synthesized.
 857                  */
 858                 __event__synthesize_thread(comm_event, mmap_event, fork_event,
 859                                            namespaces_event, pid, 1, process,
 860                                            tool, machine, mmap_data);
 861         }
 862         err = 0;
 863
 864         free(namespaces_event);
 865 out_free_fork:
 866         free(fork_event);
 867 out_free_mmap:
 868         free(mmap_event);
 869 out_free_comm:
 870         free(comm_event);
 871 out:
 872         return err;
 873 }
 874
 875 struct synthesize_threads_arg {
 876         struct perf_tool *tool;
 877         perf_event__handler_t process;
 878         struct machine *machine;
 879         bool mmap_data;
 880         struct dirent **dirent;
 881         int num;
 882         int start;
 883 };
 884
 885 static void *synthesize_threads_worker(void *arg)
 886 {
 887         struct synthesize_threads_arg *args = arg;
 888
 889         __perf_event__synthesize_threads(args->tool, args->process,
 890                                          args->machine, args->mmap_data,
 891                                          args->dirent,
 892                                          args->start, args->num);
 893         return NULL;
 894 }
 895
 896 int perf_event__synthesize_threads(struct perf_tool *tool,
 897                                    perf_event__handler_t process,
 898                                    struct machine *machine,
 899                                    bool mmap_data,
 900                                    unsigned int nr_threads_synthesize)
 901 {
 902         struct synthesize_threads_arg *args = NULL;
 903         pthread_t *synthesize_threads = NULL;
 904         char proc_path[PATH_MAX];
 905         struct dirent **dirent;
 906         int num_per_thread;
 907         int m, n, i, j;
 908         int thread_nr;
 909         int base = 0;
 910         int err = -1;
 911
 912
 913         if (machine__is_default_guest(machine))
 914                 return 0;
 915
 916         snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
 917         n = scandir(proc_path, &dirent, 0, alphasort);
 918         if (n < 0)
 919                 return err;
 920
 921         if (nr_threads_synthesize == UINT_MAX)
 922                 thread_nr = sysconf(_SC_NPROCESSORS_ONLN);
 923         else
 924                 thread_nr = nr_threads_synthesize;
 925
 926         if (thread_nr <= 1) {
 927                 err = __perf_event__synthesize_threads(tool, process,
 928                                                        machine, mmap_data,
 929                                                        dirent, base, n);
 930                 goto free_dirent;
 931         }
 932         if (thread_nr > n)
 933                 thread_nr = n;
 934
 935         synthesize_threads = calloc(sizeof(pthread_t), thread_nr);
 936         if (synthesize_threads == NULL)
 937                 goto free_dirent;
 938
 939         args = calloc(sizeof(*args), thread_nr);
 940         if (args == NULL)
 941                 goto free_threads;
 942
 943         num_per_thread = n / thread_nr;
 944         m = n % thread_nr;
 945         for (i = 0; i < thread_nr; i++) {
 946                 args[i].tool = tool;
 947                 args[i].process = process;
 948                 args[i].machine = machine;
 949                 args[i].mmap_data = mmap_data;
 950                 args[i].dirent = dirent;
 951         }
 952         for (i = 0; i < m; i++) {
 953                 args[i].num = num_per_thread + 1;
 954                 args[i].start = i * args[i].num;
 955         }
 956         if (i != 0)
 957                 base = args[i-1].start + args[i-1].num;
 958         for (j = i; j < thread_nr; j++) {
 959                 args[j].num = num_per_thread;
 960                 args[j].start = base + (j - i) * args[i].num;
 961         }
 962
 963         for (i = 0; i < thread_nr; i++) {
 964                 if (pthread_create(&synthesize_threads[i], NULL,
 965                                    synthesize_threads_worker, &args[i]))
 966                         goto out_join;
 967         }
 968         err = 0;
 969 out_join:
 970         for (i = 0; i < thread_nr; i++)
 971                 pthread_join(synthesize_threads[i], NULL);
 972         free(args);
 973 free_threads:
 974         free(synthesize_threads);
 975 free_dirent:
 976         for (i = 0; i < n; i++)
 977                 zfree(&dirent[i]);
 978         free(dirent);
 979
 980         return err;
 981 }
 982
 983 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused,
 984                                               perf_event__handler_t process __maybe_unused,
 985                                               struct machine *machine __maybe_unused)
 986 {
 987         return 0;
 988 }
 989
 990 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
 991                                                 perf_event__handler_t process,
 992                                                 struct machine *machine)
 993 {
 994         size_t size;
 995         struct map *map = machine__kernel_map(machine);
 996         struct kmap *kmap;
 997         int err;
 998         union perf_event *event;
 999
1000         if (map == NULL)
1001                 return -1;
1002
1003         kmap = map__kmap(map);
1004         if (!kmap->ref_reloc_sym)
1005                 return -1;
1006
1007         /*
1008          * We should get this from /sys/kernel/sections/.text, but till that is
1009          * available use this, and after it is use this as a fallback for older
1010          * kernels.
1011          */
1012         event = zalloc((sizeof(event->mmap) + machine->id_hdr_size));
1013         if (event == NULL) {
1014                 pr_debug("Not enough memory synthesizing mmap event "
1015                          "for kernel modules\n");
1016                 return -1;
1017         }
1018
1019         if (machine__is_host(machine)) {
1020                 /*
1021                  * kernel uses PERF_RECORD_MISC_USER for user space maps,
1022                  * see kernel/perf_event.c __perf_event_mmap
1023                  */
1024                 event->header.misc = PERF_RECORD_MISC_KERNEL;
1025         } else {
1026                 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
1027         }
1028
1029         size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
1030                         "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1;
1031         size = PERF_ALIGN(size, sizeof(u64));
1032         event->mmap.header.type = PERF_RECORD_MMAP;
1033         event->mmap.header.size = (sizeof(event->mmap) -
1034                         (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
1035         event->mmap.pgoff = kmap->ref_reloc_sym->addr;
1036         event->mmap.start = map->start;
1037         event->mmap.len   = map->end - event->mmap.start;
1038         event->mmap.pid   = machine->pid;
1039
1040         err = perf_tool__process_synth_event(tool, event, machine, process);
1041         free(event);
1042
1043         return err;
1044 }
1045
1046 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
1047                                        perf_event__handler_t process,
1048                                        struct machine *machine)
1049 {
1050         int err;
1051
1052         err = __perf_event__synthesize_kernel_mmap(tool, process, machine);
1053         if (err < 0)
1054                 return err;
1055
1056         return perf_event__synthesize_extra_kmaps(tool, process, machine);
1057 }
1058
1059 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
1060                                       struct perf_thread_map *threads,
1061                                       perf_event__handler_t process,
1062                                       struct machine *machine)
1063 {
1064         union perf_event *event;
1065         int i, err, size;
1066
1067         size  = sizeof(event->thread_map);
1068         size += threads->nr * sizeof(event->thread_map.entries[0]);
1069
1070         event = zalloc(size);
1071         if (!event)
1072                 return -ENOMEM;
1073
1074         event->header.type = PERF_RECORD_THREAD_MAP;
1075         event->header.size = size;
1076         event->thread_map.nr = threads->nr;
1077
1078         for (i = 0; i < threads->nr; i++) {
1079                 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i];
1080                 char *comm = perf_thread_map__comm(threads, i);
1081
1082                 if (!comm)
1083                         comm = (char *) "";
1084
1085                 entry->pid = perf_thread_map__pid(threads, i);
1086                 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
1087         }
1088
1089         err = process(tool, event, NULL, machine);
1090
1091         free(event);
1092         return err;
1093 }
1094
1095 static void synthesize_cpus(struct cpu_map_entries *cpus,
1096                             struct perf_cpu_map *map)
1097 {
1098         int i;
1099
1100         cpus->nr = map->nr;
1101
1102         for (i = 0; i < map->nr; i++)
1103                 cpus->cpu[i] = map->map[i];
1104 }
1105
1106 static void synthesize_mask(struct perf_record_record_cpu_map *mask,
1107                             struct perf_cpu_map *map, int max)
1108 {
1109         int i;
1110
1111         mask->nr = BITS_TO_LONGS(max);
1112         mask->long_size = sizeof(long);
1113
1114         for (i = 0; i < map->nr; i++)
1115                 set_bit(map->map[i], mask->mask);
1116 }
1117
1118 static size_t cpus_size(struct perf_cpu_map *map)
1119 {
1120         return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16);
1121 }
1122
1123 static size_t mask_size(struct perf_cpu_map *map, int *max)
1124 {
1125         int i;
1126
1127         *max = 0;
1128
1129         for (i = 0; i < map->nr; i++) {
1130                 /* bit possition of the cpu is + 1 */
1131                 int bit = map->map[i] + 1;
1132
1133                 if (bit > *max)
1134                         *max = bit;
1135         }
1136
1137         return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long);
1138 }
1139
1140 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max)
1141 {
1142         size_t size_cpus, size_mask;
1143         bool is_dummy = perf_cpu_map__empty(map);
1144
1145         /*
1146          * Both array and mask data have variable size based
1147          * on the number of cpus and their actual values.
1148          * The size of the 'struct perf_record_cpu_map_data' is:
1149          *
1150          *   array = size of 'struct cpu_map_entries' +
1151          *           number of cpus * sizeof(u64)
1152          *
1153          *   mask  = size of 'struct perf_record_record_cpu_map' +
1154          *           maximum cpu bit converted to size of longs
1155          *
1156          * and finaly + the size of 'struct perf_record_cpu_map_data'.
1157          */
1158         size_cpus = cpus_size(map);
1159         size_mask = mask_size(map, max);
1160
1161         if (is_dummy || (size_cpus < size_mask)) {
1162                 *size += size_cpus;
1163                 *type  = PERF_CPU_MAP__CPUS;
1164         } else {
1165                 *size += size_mask;
1166                 *type  = PERF_CPU_MAP__MASK;
1167         }
1168
1169         *size += sizeof(struct perf_record_cpu_map_data);
1170         *size = PERF_ALIGN(*size, sizeof(u64));
1171         return zalloc(*size);
1172 }
1173
1174 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
1175                               u16 type, int max)
1176 {
1177         data->type = type;
1178
1179         switch (type) {
1180         case PERF_CPU_MAP__CPUS:
1181                 synthesize_cpus((struct cpu_map_entries *) data->data, map);
1182                 break;
1183         case PERF_CPU_MAP__MASK:
1184                 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max);
1185         default:
1186                 break;
1187         }
1188 }
1189
1190 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map)
1191 {
1192         size_t size = sizeof(struct perf_record_cpu_map);
1193         struct perf_record_cpu_map *event;
1194         int max;
1195         u16 type;
1196
1197         event = cpu_map_data__alloc(map, &size, &type, &max);
1198         if (!event)
1199                 return NULL;
1200
1201         event->header.type = PERF_RECORD_CPU_MAP;
1202         event->header.size = size;
1203         event->data.type   = type;
1204
1205         cpu_map_data__synthesize(&event->data, map, type, max);
1206         return event;
1207 }
1208
1209 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
1210                                    struct perf_cpu_map *map,
1211                                    perf_event__handler_t process,
1212                                    struct machine *machine)
1213 {
1214         struct perf_record_cpu_map *event;
1215         int err;
1216
1217         event = cpu_map_event__new(map);
1218         if (!event)
1219                 return -ENOMEM;
1220
1221         err = process(tool, (union perf_event *) event, NULL, machine);
1222
1223         free(event);
1224         return err;
1225 }
1226
1227 int perf_event__synthesize_stat_config(struct perf_tool *tool,
1228                                        struct perf_stat_config *config,
1229                                        perf_event__handler_t process,
1230                                        struct machine *machine)
1231 {
1232         struct perf_record_stat_config *event;
1233         int size, i = 0, err;
1234
1235         size  = sizeof(*event);
1236         size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
1237
1238         event = zalloc(size);
1239         if (!event)
1240                 return -ENOMEM;
1241
1242         event->header.type = PERF_RECORD_STAT_CONFIG;
1243         event->header.size = size;
1244         event->nr          = PERF_STAT_CONFIG_TERM__MAX;
1245
1246 #define ADD(__term, __val)                                      \
1247         event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term;   \
1248         event->data[i].val = __val;                             \
1249         i++;
1250
1251         ADD(AGGR_MODE,  config->aggr_mode)
1252         ADD(INTERVAL,   config->interval)
1253         ADD(SCALE,      config->scale)
1254
1255         WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
1256                   "stat config terms unbalanced\n");
1257 #undef ADD
1258
1259         err = process(tool, (union perf_event *) event, NULL, machine);
1260
1261         free(event);
1262         return err;
1263 }
1264
1265 int perf_event__synthesize_stat(struct perf_tool *tool,
1266                                 u32 cpu, u32 thread, u64 id,
1267                                 struct perf_counts_values *count,
1268                                 perf_event__handler_t process,
1269                                 struct machine *machine)
1270 {
1271         struct perf_record_stat event;
1272
1273         event.header.type = PERF_RECORD_STAT;
1274         event.header.size = sizeof(event);
1275         event.header.misc = 0;
1276
1277         event.id        = id;
1278         event.cpu       = cpu;
1279         event.thread    = thread;
1280         event.val       = count->val;
1281         event.ena       = count->ena;
1282         event.run       = count->run;
1283
1284         return process(tool, (union perf_event *) &event, NULL, machine);
1285 }
1286
1287 int perf_event__synthesize_stat_round(struct perf_tool *tool,
1288                                       u64 evtime, u64 type,
1289                                       perf_event__handler_t process,
1290                                       struct machine *machine)
1291 {
1292         struct perf_record_stat_round event;
1293
1294         event.header.type = PERF_RECORD_STAT_ROUND;
1295         event.header.size = sizeof(event);
1296         event.header.misc = 0;
1297
1298         event.time = evtime;
1299         event.type = type;
1300
1301         return process(tool, (union perf_event *) &event, NULL, machine);
1302 }
1303
1304 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format)
1305 {
1306         size_t sz, result = sizeof(struct perf_record_sample);
1307
1308         if (type & PERF_SAMPLE_IDENTIFIER)
1309                 result += sizeof(u64);
1310
1311         if (type & PERF_SAMPLE_IP)
1312                 result += sizeof(u64);
1313
1314         if (type & PERF_SAMPLE_TID)
1315                 result += sizeof(u64);
1316
1317         if (type & PERF_SAMPLE_TIME)
1318                 result += sizeof(u64);
1319
1320         if (type & PERF_SAMPLE_ADDR)
1321                 result += sizeof(u64);
1322
1323         if (type & PERF_SAMPLE_ID)
1324                 result += sizeof(u64);
1325
1326         if (type & PERF_SAMPLE_STREAM_ID)
1327                 result += sizeof(u64);
1328
1329         if (type & PERF_SAMPLE_CPU)
1330                 result += sizeof(u64);
1331
1332         if (type & PERF_SAMPLE_PERIOD)
1333                 result += sizeof(u64);
1334
1335         if (type & PERF_SAMPLE_READ) {
1336                 result += sizeof(u64);
1337                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1338                         result += sizeof(u64);
1339                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1340                         result += sizeof(u64);
1341                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1342                 if (read_format & PERF_FORMAT_GROUP) {
1343                         sz = sample->read.group.nr *
1344                              sizeof(struct sample_read_value);
1345                         result += sz;
1346                 } else {
1347                         result += sizeof(u64);
1348                 }
1349         }
1350
1351         if (type & PERF_SAMPLE_CALLCHAIN) {
1352                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1353                 result += sz;
1354         }
1355
1356         if (type & PERF_SAMPLE_RAW) {
1357                 result += sizeof(u32);
1358                 result += sample->raw_size;
1359         }
1360
1361         if (type & PERF_SAMPLE_BRANCH_STACK) {
1362                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1363                 /* nr, hw_idx */
1364                 sz += 2 * sizeof(u64);
1365                 result += sz;
1366         }
1367
1368         if (type & PERF_SAMPLE_REGS_USER) {
1369                 if (sample->user_regs.abi) {
1370                         result += sizeof(u64);
1371                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1372                         result += sz;
1373                 } else {
1374                         result += sizeof(u64);
1375                 }
1376         }
1377
1378         if (type & PERF_SAMPLE_STACK_USER) {
1379                 sz = sample->user_stack.size;
1380                 result += sizeof(u64);
1381                 if (sz) {
1382                         result += sz;
1383                         result += sizeof(u64);
1384                 }
1385         }
1386
1387         if (type & PERF_SAMPLE_WEIGHT)
1388                 result += sizeof(u64);
1389
1390         if (type & PERF_SAMPLE_DATA_SRC)
1391                 result += sizeof(u64);
1392
1393         if (type & PERF_SAMPLE_TRANSACTION)
1394                 result += sizeof(u64);
1395
1396         if (type & PERF_SAMPLE_REGS_INTR) {
1397                 if (sample->intr_regs.abi) {
1398                         result += sizeof(u64);
1399                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1400                         result += sz;
1401                 } else {
1402                         result += sizeof(u64);
1403                 }
1404         }
1405
1406         if (type & PERF_SAMPLE_PHYS_ADDR)
1407                 result += sizeof(u64);
1408
1409         if (type & PERF_SAMPLE_CGROUP)
1410                 result += sizeof(u64);
1411
1412         if (type & PERF_SAMPLE_AUX) {
1413                 result += sizeof(u64);
1414                 result += sample->aux_sample.size;
1415         }
1416
1417         return result;
1418 }
1419
1420 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format,
1421                                   const struct perf_sample *sample)
1422 {
1423         __u64 *array;
1424         size_t sz;
1425         /*
1426          * used for cross-endian analysis. See git commit 65014ab3
1427          * for why this goofiness is needed.
1428          */
1429         union u64_swap u;
1430
1431         array = event->sample.array;
1432
1433         if (type & PERF_SAMPLE_IDENTIFIER) {
1434                 *array = sample->id;
1435                 array++;
1436         }
1437
1438         if (type & PERF_SAMPLE_IP) {
1439                 *array = sample->ip;
1440                 array++;
1441         }
1442
1443         if (type & PERF_SAMPLE_TID) {
1444                 u.val32[0] = sample->pid;
1445                 u.val32[1] = sample->tid;
1446                 *array = u.val64;
1447                 array++;
1448         }
1449
1450         if (type & PERF_SAMPLE_TIME) {
1451                 *array = sample->time;
1452                 array++;
1453         }
1454
1455         if (type & PERF_SAMPLE_ADDR) {
1456                 *array = sample->addr;
1457                 array++;
1458         }
1459
1460         if (type & PERF_SAMPLE_ID) {
1461                 *array = sample->id;
1462                 array++;
1463         }
1464
1465         if (type & PERF_SAMPLE_STREAM_ID) {
1466                 *array = sample->stream_id;
1467                 array++;
1468         }
1469
1470         if (type & PERF_SAMPLE_CPU) {
1471                 u.val32[0] = sample->cpu;
1472                 u.val32[1] = 0;
1473                 *array = u.val64;
1474                 array++;
1475         }
1476
1477         if (type & PERF_SAMPLE_PERIOD) {
1478                 *array = sample->period;
1479                 array++;
1480         }
1481
1482         if (type & PERF_SAMPLE_READ) {
1483                 if (read_format & PERF_FORMAT_GROUP)
1484                         *array = sample->read.group.nr;
1485                 else
1486                         *array = sample->read.one.value;
1487                 array++;
1488
1489                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1490                         *array = sample->read.time_enabled;
1491                         array++;
1492                 }
1493
1494                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1495                         *array = sample->read.time_running;
1496                         array++;
1497                 }
1498
1499                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1500                 if (read_format & PERF_FORMAT_GROUP) {
1501                         sz = sample->read.group.nr *
1502                              sizeof(struct sample_read_value);
1503                         memcpy(array, sample->read.group.values, sz);
1504                         array = (void *)array + sz;
1505                 } else {
1506                         *array = sample->read.one.id;
1507                         array++;
1508                 }
1509         }
1510
1511         if (type & PERF_SAMPLE_CALLCHAIN) {
1512                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1513                 memcpy(array, sample->callchain, sz);
1514                 array = (void *)array + sz;
1515         }
1516
1517         if (type & PERF_SAMPLE_RAW) {
1518                 u.val32[0] = sample->raw_size;
1519                 *array = u.val64;
1520                 array = (void *)array + sizeof(u32);
1521
1522                 memcpy(array, sample->raw_data, sample->raw_size);
1523                 array = (void *)array + sample->raw_size;
1524         }
1525
1526         if (type & PERF_SAMPLE_BRANCH_STACK) {
1527                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1528                 /* nr, hw_idx */
1529                 sz += 2 * sizeof(u64);
1530                 memcpy(array, sample->branch_stack, sz);
1531                 array = (void *)array + sz;
1532         }
1533
1534         if (type & PERF_SAMPLE_REGS_USER) {
1535                 if (sample->user_regs.abi) {
1536                         *array++ = sample->user_regs.abi;
1537                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
1538                         memcpy(array, sample->user_regs.regs, sz);
1539                         array = (void *)array + sz;
1540                 } else {
1541                         *array++ = 0;
1542                 }
1543         }
1544
1545         if (type & PERF_SAMPLE_STACK_USER) {
1546                 sz = sample->user_stack.size;
1547                 *array++ = sz;
1548                 if (sz) {
1549                         memcpy(array, sample->user_stack.data, sz);
1550                         array = (void *)array + sz;
1551                         *array++ = sz;
1552                 }
1553         }
1554
1555         if (type & PERF_SAMPLE_WEIGHT) {
1556                 *array = sample->weight;
1557                 array++;
1558         }
1559
1560         if (type & PERF_SAMPLE_DATA_SRC) {
1561                 *array = sample->data_src;
1562                 array++;
1563         }
1564
1565         if (type & PERF_SAMPLE_TRANSACTION) {
1566                 *array = sample->transaction;
1567                 array++;
1568         }
1569
1570         if (type & PERF_SAMPLE_REGS_INTR) {
1571                 if (sample->intr_regs.abi) {
1572                         *array++ = sample->intr_regs.abi;
1573                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
1574                         memcpy(array, sample->intr_regs.regs, sz);
1575                         array = (void *)array + sz;
1576                 } else {
1577                         *array++ = 0;
1578                 }
1579         }
1580
1581         if (type & PERF_SAMPLE_PHYS_ADDR) {
1582                 *array = sample->phys_addr;
1583                 array++;
1584         }
1585
1586         if (type & PERF_SAMPLE_CGROUP) {
1587                 *array = sample->cgroup;
1588                 array++;
1589         }
1590
1591         if (type & PERF_SAMPLE_AUX) {
1592                 sz = sample->aux_sample.size;
1593                 *array++ = sz;
1594                 memcpy(array, sample->aux_sample.data, sz);
1595                 array = (void *)array + sz;
1596         }
1597
1598         return 0;
1599 }
1600
1601 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process,
1602                                     struct evlist *evlist, struct machine *machine)
1603 {
1604         union perf_event *ev;
1605         struct evsel *evsel;
1606         size_t nr = 0, i = 0, sz, max_nr, n;
1607         int err;
1608
1609         pr_debug2("Synthesizing id index\n");
1610
1611         max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) /
1612                  sizeof(struct id_index_entry);
1613
1614         evlist__for_each_entry(evlist, evsel)
1615                 nr += evsel->core.ids;
1616
1617         n = nr > max_nr ? max_nr : nr;
1618         sz = sizeof(struct perf_record_id_index) + n * sizeof(struct id_index_entry);
1619         ev = zalloc(sz);
1620         if (!ev)
1621                 return -ENOMEM;
1622
1623         ev->id_index.header.type = PERF_RECORD_ID_INDEX;
1624         ev->id_index.header.size = sz;
1625         ev->id_index.nr = n;
1626
1627         evlist__for_each_entry(evlist, evsel) {
1628                 u32 j;
1629
1630                 for (j = 0; j < evsel->core.ids; j++) {
1631                         struct id_index_entry *e;
1632                         struct perf_sample_id *sid;
1633
1634                         if (i >= n) {
1635                                 err = process(tool, ev, NULL, machine);
1636                                 if (err)
1637                                         goto out_err;
1638                                 nr -= n;
1639                                 i = 0;
1640                         }
1641
1642                         e = &ev->id_index.entries[i++];
1643
1644                         e->id = evsel->core.id[j];
1645
1646                         sid = evlist__id2sid(evlist, e->id);
1647                         if (!sid) {
1648                                 free(ev);
1649                                 return -ENOENT;
1650                         }
1651
1652                         e->idx = sid->idx;
1653                         e->cpu = sid->cpu;
1654                         e->tid = sid->tid;
1655                 }
1656         }
1657
1658         sz = sizeof(struct perf_record_id_index) + nr * sizeof(struct id_index_entry);
1659         ev->id_index.header.size = sz;
1660         ev->id_index.nr = nr;
1661
1662         err = process(tool, ev, NULL, machine);
1663 out_err:
1664         free(ev);
1665
1666         return err;
1667 }
1668
1669 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1670                                   struct target *target, struct perf_thread_map *threads,
1671                                   perf_event__handler_t process, bool data_mmap,
1672                                   unsigned int nr_threads_synthesize)
1673 {
1674         if (target__has_task(target))
1675                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
1676         else if (target__has_cpu(target))
1677                 return perf_event__synthesize_threads(tool, process,
1678                                                       machine, data_mmap,
1679                                                       nr_threads_synthesize);
1680         /* command specified */
1681         return 0;
1682 }
1683
1684 int machine__synthesize_threads(struct machine *machine, struct target *target,
1685                                 struct perf_thread_map *threads, bool data_mmap,
1686                                 unsigned int nr_threads_synthesize)
1687 {
1688         return __machine__synthesize_threads(machine, NULL, target, threads,
1689                                              perf_event__process, data_mmap,
1690                                              nr_threads_synthesize);
1691 }
1692
1693 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id)
1694 {
1695         struct perf_record_event_update *ev;
1696
1697         size += sizeof(*ev);
1698         size  = PERF_ALIGN(size, sizeof(u64));
1699
1700         ev = zalloc(size);
1701         if (ev) {
1702                 ev->header.type = PERF_RECORD_EVENT_UPDATE;
1703                 ev->header.size = (u16)size;
1704                 ev->type        = type;
1705                 ev->id          = id;
1706         }
1707         return ev;
1708 }
1709
1710 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel,
1711                                              perf_event__handler_t process)
1712 {
1713         size_t size = strlen(evsel->unit);
1714         struct perf_record_event_update *ev;
1715         int err;
1716
1717         ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]);
1718         if (ev == NULL)
1719                 return -ENOMEM;
1720
1721         strlcpy(ev->data, evsel->unit, size + 1);
1722         err = process(tool, (union perf_event *)ev, NULL, NULL);
1723         free(ev);
1724         return err;
1725 }
1726
1727 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel,
1728                                               perf_event__handler_t process)
1729 {
1730         struct perf_record_event_update *ev;
1731         struct perf_record_event_update_scale *ev_data;
1732         int err;
1733
1734         ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]);
1735         if (ev == NULL)
1736                 return -ENOMEM;
1737
1738         ev_data = (struct perf_record_event_update_scale *)ev->data;
1739         ev_data->scale = evsel->scale;
1740         err = process(tool, (union perf_event *)ev, NULL, NULL);
1741         free(ev);
1742         return err;
1743 }
1744
1745 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel,
1746                                              perf_event__handler_t process)
1747 {
1748         struct perf_record_event_update *ev;
1749         size_t len = strlen(evsel->name);
1750         int err;
1751
1752         ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]);
1753         if (ev == NULL)
1754                 return -ENOMEM;
1755
1756         strlcpy(ev->data, evsel->name, len + 1);
1757         err = process(tool, (union perf_event *)ev, NULL, NULL);
1758         free(ev);
1759         return err;
1760 }
1761
1762 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel,
1763                                              perf_event__handler_t process)
1764 {
1765         size_t size = sizeof(struct perf_record_event_update);
1766         struct perf_record_event_update *ev;
1767         int max, err;
1768         u16 type;
1769
1770         if (!evsel->core.own_cpus)
1771                 return 0;
1772
1773         ev = cpu_map_data__alloc(evsel->core.own_cpus, &size, &type, &max);
1774         if (!ev)
1775                 return -ENOMEM;
1776
1777         ev->header.type = PERF_RECORD_EVENT_UPDATE;
1778         ev->header.size = (u16)size;
1779         ev->type        = PERF_EVENT_UPDATE__CPUS;
1780         ev->id          = evsel->core.id[0];
1781
1782         cpu_map_data__synthesize((struct perf_record_cpu_map_data *)ev->data,
1783                                  evsel->core.own_cpus, type, max);
1784
1785         err = process(tool, (union perf_event *)ev, NULL, NULL);
1786         free(ev);
1787         return err;
1788 }
1789
1790 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist,
1791                                  perf_event__handler_t process)
1792 {
1793         struct evsel *evsel;
1794         int err = 0;
1795
1796         evlist__for_each_entry(evlist, evsel) {
1797                 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids,
1798                                                   evsel->core.id, process);
1799                 if (err) {
1800                         pr_debug("failed to create perf header attribute\n");
1801                         return err;
1802                 }
1803         }
1804
1805         return err;
1806 }
1807
1808 static bool has_unit(struct evsel *evsel)
1809 {
1810         return evsel->unit && *evsel->unit;
1811 }
1812
1813 static bool has_scale(struct evsel *evsel)
1814 {
1815         return evsel->scale != 1;
1816 }
1817
1818 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list,
1819                                       perf_event__handler_t process, bool is_pipe)
1820 {
1821         struct evsel *evsel;
1822         int err;
1823
1824         /*
1825          * Synthesize other events stuff not carried within
1826          * attr event - unit, scale, name
1827          */
1828         evlist__for_each_entry(evsel_list, evsel) {
1829                 if (!evsel->supported)
1830                         continue;
1831
1832                 /*
1833                  * Synthesize unit and scale only if it's defined.
1834                  */
1835                 if (has_unit(evsel)) {
1836                         err = perf_event__synthesize_event_update_unit(tool, evsel, process);
1837                         if (err < 0) {
1838                                 pr_err("Couldn't synthesize evsel unit.\n");
1839                                 return err;
1840                         }
1841                 }
1842
1843                 if (has_scale(evsel)) {
1844                         err = perf_event__synthesize_event_update_scale(tool, evsel, process);
1845                         if (err < 0) {
1846                                 pr_err("Couldn't synthesize evsel evsel.\n");
1847                                 return err;
1848                         }
1849                 }
1850
1851                 if (evsel->core.own_cpus) {
1852                         err = perf_event__synthesize_event_update_cpus(tool, evsel, process);
1853                         if (err < 0) {
1854                                 pr_err("Couldn't synthesize evsel cpus.\n");
1855                                 return err;
1856                         }
1857                 }
1858
1859                 /*
1860                  * Name is needed only for pipe output,
1861                  * perf.data carries event names.
1862                  */
1863                 if (is_pipe) {
1864                         err = perf_event__synthesize_event_update_name(tool, evsel, process);
1865                         if (err < 0) {
1866                                 pr_err("Couldn't synthesize evsel name.\n");
1867                                 return err;
1868                         }
1869                 }
1870         }
1871         return 0;
1872 }
1873
1874 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr,
1875                                 u32 ids, u64 *id, perf_event__handler_t process)
1876 {
1877         union perf_event *ev;
1878         size_t size;
1879         int err;
1880
1881         size = sizeof(struct perf_event_attr);
1882         size = PERF_ALIGN(size, sizeof(u64));
1883         size += sizeof(struct perf_event_header);
1884         size += ids * sizeof(u64);
1885
1886         ev = zalloc(size);
1887
1888         if (ev == NULL)
1889                 return -ENOMEM;
1890
1891         ev->attr.attr = *attr;
1892         memcpy(ev->attr.id, id, ids * sizeof(u64));
1893
1894         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
1895         ev->attr.header.size = (u16)size;
1896
1897         if (ev->attr.header.size == size)
1898                 err = process(tool, ev, NULL, NULL);
1899         else
1900                 err = -E2BIG;
1901
1902         free(ev);
1903
1904         return err;
1905 }
1906
1907 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist,
1908                                         perf_event__handler_t process)
1909 {
1910         union perf_event ev;
1911         struct tracing_data *tdata;
1912         ssize_t size = 0, aligned_size = 0, padding;
1913         struct feat_fd ff;
1914
1915         /*
1916          * We are going to store the size of the data followed
1917          * by the data contents. Since the fd descriptor is a pipe,
1918          * we cannot seek back to store the size of the data once
1919          * we know it. Instead we:
1920          *
1921          * - write the tracing data to the temp file
1922          * - get/write the data size to pipe
1923          * - write the tracing data from the temp file
1924          *   to the pipe
1925          */
1926         tdata = tracing_data_get(&evlist->core.entries, fd, true);
1927         if (!tdata)
1928                 return -1;
1929
1930         memset(&ev, 0, sizeof(ev));
1931
1932         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
1933         size = tdata->size;
1934         aligned_size = PERF_ALIGN(size, sizeof(u64));
1935         padding = aligned_size - size;
1936         ev.tracing_data.header.size = sizeof(ev.tracing_data);
1937         ev.tracing_data.size = aligned_size;
1938
1939         process(tool, &ev, NULL, NULL);
1940
1941         /*
1942          * The put function will copy all the tracing data
1943          * stored in temp file to the pipe.
1944          */
1945         tracing_data_put(tdata);
1946
1947         ff = (struct feat_fd){ .fd = fd };
1948         if (write_padded(&ff, NULL, 0, padding))
1949                 return -1;
1950
1951         return aligned_size;
1952 }
1953
1954 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc,
1955                                     perf_event__handler_t process, struct machine *machine)
1956 {
1957         union perf_event ev;
1958         size_t len;
1959
1960         if (!pos->hit)
1961                 return 0;
1962
1963         memset(&ev, 0, sizeof(ev));
1964
1965         len = pos->long_name_len + 1;
1966         len = PERF_ALIGN(len, NAME_ALIGN);
1967         memcpy(&ev.build_id.build_id, pos->bid.data, sizeof(pos->bid.data));
1968         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
1969         ev.build_id.header.misc = misc;
1970         ev.build_id.pid = machine->pid;
1971         ev.build_id.header.size = sizeof(ev.build_id) + len;
1972         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
1973
1974         return process(tool, &ev, NULL, machine);
1975 }
1976
1977 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool,
1978                                        struct evlist *evlist, perf_event__handler_t process, bool attrs)
1979 {
1980         int err;
1981
1982         if (attrs) {
1983                 err = perf_event__synthesize_attrs(tool, evlist, process);
1984                 if (err < 0) {
1985                         pr_err("Couldn't synthesize attrs.\n");
1986                         return err;
1987                 }
1988         }
1989
1990         err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs);
1991         err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL);
1992         if (err < 0) {
1993                 pr_err("Couldn't synthesize thread map.\n");
1994                 return err;
1995         }
1996
1997         err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus, process, NULL);
1998         if (err < 0) {
1999                 pr_err("Couldn't synthesize thread map.\n");
2000                 return err;
2001         }
2002
2003         err = perf_event__synthesize_stat_config(tool, config, process, NULL);
2004         if (err < 0) {
2005                 pr_err("Couldn't synthesize config.\n");
2006                 return err;
2007         }
2008
2009         return 0;
2010 }
2011
2012 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE];
2013
2014 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session,
2015                                     struct evlist *evlist, perf_event__handler_t process)
2016 {
2017         struct perf_header *header = &session->header;
2018         struct perf_record_header_feature *fe;
2019         struct feat_fd ff;
2020         size_t sz, sz_hdr;
2021         int feat, ret;
2022
2023         sz_hdr = sizeof(fe->header);
2024         sz = sizeof(union perf_event);
2025         /* get a nice alignment */
2026         sz = PERF_ALIGN(sz, page_size);
2027
2028         memset(&ff, 0, sizeof(ff));
2029
2030         ff.buf = malloc(sz);
2031         if (!ff.buf)
2032                 return -ENOMEM;
2033
2034         ff.size = sz - sz_hdr;
2035         ff.ph = &session->header;
2036
2037         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2038                 if (!feat_ops[feat].synthesize) {
2039                         pr_debug("No record header feature for header :%d\n", feat);
2040                         continue;
2041                 }
2042
2043                 ff.offset = sizeof(*fe);
2044
2045                 ret = feat_ops[feat].write(&ff, evlist);
2046                 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
2047                         pr_debug("Error writing feature\n");
2048                         continue;
2049                 }
2050                 /* ff.buf may have changed due to realloc in do_write() */
2051                 fe = ff.buf;
2052                 memset(fe, 0, sizeof(*fe));
2053
2054                 fe->feat_id = feat;
2055                 fe->header.type = PERF_RECORD_HEADER_FEATURE;
2056                 fe->header.size = ff.offset;
2057
2058                 ret = process(tool, ff.buf, NULL, NULL);
2059                 if (ret) {
2060                         free(ff.buf);
2061                         return ret;
2062                 }
2063         }
2064
2065         /* Send HEADER_LAST_FEATURE mark. */
2066         fe = ff.buf;
2067         fe->feat_id     = HEADER_LAST_FEATURE;
2068         fe->header.type = PERF_RECORD_HEADER_FEATURE;
2069         fe->header.size = sizeof(*fe);
2070
2071         ret = process(tool, ff.buf, NULL, NULL);
2072
2073         free(ff.buf);
2074         return ret;
2075 }