fs/coredump.c

   1 #include <linux/slab.h>
   2 #include <linux/file.h>
   3 #include <linux/fdtable.h>
   4 #include <linux/freezer.h>
   5 #include <linux/mm.h>
   6 #include <linux/stat.h>
   7 #include <linux/fcntl.h>
   8 #include <linux/swap.h>
   9 #include <linux/string.h>
  10 #include <linux/init.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/perf_event.h>
  13 #include <linux/highmem.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/key.h>
  16 #include <linux/personality.h>
  17 #include <linux/binfmts.h>
  18 #include <linux/coredump.h>
  19 #include <linux/utsname.h>
  20 #include <linux/pid_namespace.h>
  21 #include <linux/module.h>
  22 #include <linux/namei.h>
  23 #include <linux/mount.h>
  24 #include <linux/security.h>
  25 #include <linux/syscalls.h>
  26 #include <linux/tsacct_kern.h>
  27 #include <linux/cn_proc.h>
  28 #include <linux/audit.h>
  29 #include <linux/tracehook.h>
  30 #include <linux/kmod.h>
  31 #include <linux/fsnotify.h>
  32 #include <linux/fs_struct.h>
  33 #include <linux/pipe_fs_i.h>
  34 #include <linux/oom.h>
  35 #include <linux/compat.h>
  36 #include <linux/sched.h>
  37 #include <linux/fs.h>
  38 #include <linux/path.h>
  39 #include <linux/timekeeping.h>
  40
  41 #include <linux/uaccess.h>
  42 #include <asm/mmu_context.h>
  43 #include <asm/tlb.h>
  44 #include <asm/exec.h>
  45
  46 #include <trace/events/task.h>
  47 #include "internal.h"
  48
  49 #include <trace/events/sched.h>
  50
  51 int core_uses_pid;
  52 unsigned int core_pipe_limit;
  53 char core_pattern[CORENAME_MAX_SIZE] = "core";
  54 static int core_name_size = CORENAME_MAX_SIZE;
  55
  56 struct core_name {
  57         char *corename;
  58         int used, size;
  59 };
  60
  61 /* The maximal length of core_pattern is also specified in sysctl.c */
  62
  63 static int expand_corename(struct core_name *cn, int size)
  64 {
  65         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  66
  67         if (!corename)
  68                 return -ENOMEM;
  69
  70         if (size > core_name_size) /* racy but harmless */
  71                 core_name_size = size;
  72
  73         cn->size = ksize(corename);
  74         cn->corename = corename;
  75         return 0;
  76 }
  77
  78 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  79                                      va_list arg)
  80 {
  81         int free, need;
  82         va_list arg_copy;
  83
  84 again:
  85         free = cn->size - cn->used;
  86
  87         va_copy(arg_copy, arg);
  88         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  89         va_end(arg_copy);
  90
  91         if (need < free) {
  92                 cn->used += need;
  93                 return 0;
  94         }
  95
  96         if (!expand_corename(cn, cn->size + need - free + 1))
  97                 goto again;
  98
  99         return -ENOMEM;
 100 }
 101
 102 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 103 {
 104         va_list arg;
 105         int ret;
 106
 107         va_start(arg, fmt);
 108         ret = cn_vprintf(cn, fmt, arg);
 109         va_end(arg);
 110
 111         return ret;
 112 }
 113
 114 static __printf(2, 3)
 115 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 116 {
 117         int cur = cn->used;
 118         va_list arg;
 119         int ret;
 120
 121         va_start(arg, fmt);
 122         ret = cn_vprintf(cn, fmt, arg);
 123         va_end(arg);
 124
 125         if (ret == 0) {
 126                 /*
 127                  * Ensure that this coredump name component can't cause the
 128                  * resulting corefile path to consist of a ".." or ".".
 129                  */
 130                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 131                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 132                                 && cn->corename[cur+1] == '.'))
 133                         cn->corename[cur] = '!';
 134
 135                 /*
 136                  * Empty names are fishy and could be used to create a "//" in a
 137                  * corefile name, causing the coredump to happen one directory
 138                  * level too high. Enforce that all components of the core
 139                  * pattern are at least one character long.
 140                  */
 141                 if (cn->used == cur)
 142                         ret = cn_printf(cn, "!");
 143         }
 144
 145         for (; cur < cn->used; ++cur) {
 146                 if (cn->corename[cur] == '/')
 147                         cn->corename[cur] = '!';
 148         }
 149         return ret;
 150 }
 151
 152 static int cn_print_exe_file(struct core_name *cn)
 153 {
 154         struct file *exe_file;
 155         char *pathbuf, *path;
 156         int ret;
 157
 158         exe_file = get_mm_exe_file(current->mm);
 159         if (!exe_file)
 160                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 161
 162         pathbuf = kmalloc(PATH_MAX, GFP_TEMPORARY);
 163         if (!pathbuf) {
 164                 ret = -ENOMEM;
 165                 goto put_exe_file;
 166         }
 167
 168         path = file_path(exe_file, pathbuf, PATH_MAX);
 169         if (IS_ERR(path)) {
 170                 ret = PTR_ERR(path);
 171                 goto free_buf;
 172         }
 173
 174         ret = cn_esc_printf(cn, "%s", path);
 175
 176 free_buf:
 177         kfree(pathbuf);
 178 put_exe_file:
 179         fput(exe_file);
 180         return ret;
 181 }
 182
 183 /* format_corename will inspect the pattern parameter, and output a
 184  * name into corename, which must have space for at least
 185  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 186  */
 187 static int format_corename(struct core_name *cn, struct coredump_params *cprm)
 188 {
 189         const struct cred *cred = current_cred();
 190         const char *pat_ptr = core_pattern;
 191         int ispipe = (*pat_ptr == '|');
 192         int pid_in_pattern = 0;
 193         int err = 0;
 194
 195         cn->used = 0;
 196         cn->corename = NULL;
 197         if (expand_corename(cn, core_name_size))
 198                 return -ENOMEM;
 199         cn->corename[0] = '\0';
 200
 201         if (ispipe)
 202                 ++pat_ptr;
 203
 204         /* Repeat as long as we have more pattern to process and more output
 205            space */
 206         while (*pat_ptr) {
 207                 if (*pat_ptr != '%') {
 208                         err = cn_printf(cn, "%c", *pat_ptr++);
 209                 } else {
 210                         switch (*++pat_ptr) {
 211                         /* single % at the end, drop that */
 212                         case 0:
 213                                 goto out;
 214                         /* Double percent, output one percent */
 215                         case '%':
 216                                 err = cn_printf(cn, "%c", '%');
 217                                 break;
 218                         /* pid */
 219                         case 'p':
 220                                 pid_in_pattern = 1;
 221                                 err = cn_printf(cn, "%d",
 222                                               task_tgid_vnr(current));
 223                                 break;
 224                         /* global pid */
 225                         case 'P':
 226                                 err = cn_printf(cn, "%d",
 227                                               task_tgid_nr(current));
 228                                 break;
 229                         case 'i':
 230                                 err = cn_printf(cn, "%d",
 231                                               task_pid_vnr(current));
 232                                 break;
 233                         case 'I':
 234                                 err = cn_printf(cn, "%d",
 235                                               task_pid_nr(current));
 236                                 break;
 237                         /* uid */
 238                         case 'u':
 239                                 err = cn_printf(cn, "%u",
 240                                                 from_kuid(&init_user_ns,
 241                                                           cred->uid));
 242                                 break;
 243                         /* gid */
 244                         case 'g':
 245                                 err = cn_printf(cn, "%u",
 246                                                 from_kgid(&init_user_ns,
 247                                                           cred->gid));
 248                                 break;
 249                         case 'd':
 250                                 err = cn_printf(cn, "%d",
 251                                         __get_dumpable(cprm->mm_flags));
 252                                 break;
 253                         /* signal that caused the coredump */
 254                         case 's':
 255                                 err = cn_printf(cn, "%d",
 256                                                 cprm->siginfo->si_signo);
 257                                 break;
 258                         /* UNIX time of coredump */
 259                         case 't': {
 260                                 time64_t time;
 261
 262                                 time = ktime_get_real_seconds();
 263                                 err = cn_printf(cn, "%lld", time);
 264                                 break;
 265                         }
 266                         /* hostname */
 267                         case 'h':
 268                                 down_read(&uts_sem);
 269                                 err = cn_esc_printf(cn, "%s",
 270                                               utsname()->nodename);
 271                                 up_read(&uts_sem);
 272                                 break;
 273                         /* executable */
 274                         case 'e':
 275                                 err = cn_esc_printf(cn, "%s", current->comm);
 276                                 break;
 277                         case 'E':
 278                                 err = cn_print_exe_file(cn);
 279                                 break;
 280                         /* core limit size */
 281                         case 'c':
 282                                 err = cn_printf(cn, "%lu",
 283                                               rlimit(RLIMIT_CORE));
 284                                 break;
 285                         default:
 286                                 break;
 287                         }
 288                         ++pat_ptr;
 289                 }
 290
 291                 if (err)
 292                         return err;
 293         }
 294
 295 out:
 296         /* Backward compatibility with core_uses_pid:
 297          *
 298          * If core_pattern does not include a %p (as is the default)
 299          * and core_uses_pid is set, then .%pid will be appended to
 300          * the filename. Do not do this for piped commands. */
 301         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 302                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 303                 if (err)
 304                         return err;
 305         }
 306         return ispipe;
 307 }
 308
 309 static int zap_process(struct task_struct *start, int exit_code, int flags)
 310 {
 311         struct task_struct *t;
 312         int nr = 0;
 313
 314         /* ignore all signals except SIGKILL, see prepare_signal() */
 315         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
 316         start->signal->group_exit_code = exit_code;
 317         start->signal->group_stop_count = 0;
 318
 319         for_each_thread(start, t) {
 320                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 321                 if (t != current && t->mm) {
 322                         sigaddset(&t->pending.signal, SIGKILL);
 323                         signal_wake_up(t, 1);
 324                         nr++;
 325                 }
 326         }
 327
 328         return nr;
 329 }
 330
 331 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 332                         struct core_state *core_state, int exit_code)
 333 {
 334         struct task_struct *g, *p;
 335         unsigned long flags;
 336         int nr = -EAGAIN;
 337
 338         spin_lock_irq(&tsk->sighand->siglock);
 339         if (!signal_group_exit(tsk->signal)) {
 340                 mm->core_state = core_state;
 341                 tsk->signal->group_exit_task = tsk;
 342                 nr = zap_process(tsk, exit_code, 0);
 343                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 344         }
 345         spin_unlock_irq(&tsk->sighand->siglock);
 346         if (unlikely(nr < 0))
 347                 return nr;
 348
 349         tsk->flags |= PF_DUMPCORE;
 350         if (atomic_read(&mm->mm_users) == nr + 1)
 351                 goto done;
 352         /*
 353          * We should find and kill all tasks which use this mm, and we should
 354          * count them correctly into ->nr_threads. We don't take tasklist
 355          * lock, but this is safe wrt:
 356          *
 357          * fork:
 358          *      None of sub-threads can fork after zap_process(leader). All
 359          *      processes which were created before this point should be
 360          *      visible to zap_threads() because copy_process() adds the new
 361          *      process to the tail of init_task.tasks list, and lock/unlock
 362          *      of ->siglock provides a memory barrier.
 363          *
 364          * do_exit:
 365          *      The caller holds mm->mmap_sem. This means that the task which
 366          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 367          *      its ->mm.
 368          *
 369          * de_thread:
 370          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 371          *      we must see either old or new leader, this does not matter.
 372          *      However, it can change p->sighand, so lock_task_sighand(p)
 373          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
 374          *      it can't fail.
 375          *
 376          *      Note also that "g" can be the old leader with ->mm == NULL
 377          *      and already unhashed and thus removed from ->thread_group.
 378          *      This is OK, __unhash_process()->list_del_rcu() does not
 379          *      clear the ->next pointer, we will find the new leader via
 380          *      next_thread().
 381          */
 382         rcu_read_lock();
 383         for_each_process(g) {
 384                 if (g == tsk->group_leader)
 385                         continue;
 386                 if (g->flags & PF_KTHREAD)
 387                         continue;
 388
 389                 for_each_thread(g, p) {
 390                         if (unlikely(!p->mm))
 391                                 continue;
 392                         if (unlikely(p->mm == mm)) {
 393                                 lock_task_sighand(p, &flags);
 394                                 nr += zap_process(p, exit_code,
 395                                                         SIGNAL_GROUP_EXIT);
 396                                 unlock_task_sighand(p, &flags);
 397                         }
 398                         break;
 399                 }
 400         }
 401         rcu_read_unlock();
 402 done:
 403         atomic_set(&core_state->nr_threads, nr);
 404         return nr;
 405 }
 406
 407 static int coredump_wait(int exit_code, struct core_state *core_state)
 408 {
 409         struct task_struct *tsk = current;
 410         struct mm_struct *mm = tsk->mm;
 411         int core_waiters = -EBUSY;
 412
 413         init_completion(&core_state->startup);
 414         core_state->dumper.task = tsk;
 415         core_state->dumper.next = NULL;
 416
 417         if (down_write_killable(&mm->mmap_sem))
 418                 return -EINTR;
 419
 420         if (!mm->core_state)
 421                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 422         up_write(&mm->mmap_sem);
 423
 424         if (core_waiters > 0) {
 425                 struct core_thread *ptr;
 426
 427                 freezer_do_not_count();
 428                 wait_for_completion(&core_state->startup);
 429                 freezer_count();
 430                 /*
 431                  * Wait for all the threads to become inactive, so that
 432                  * all the thread context (extended register state, like
 433                  * fpu etc) gets copied to the memory.
 434                  */
 435                 ptr = core_state->dumper.next;
 436                 while (ptr != NULL) {
 437                         wait_task_inactive(ptr->task, 0);
 438                         ptr = ptr->next;
 439                 }
 440         }
 441
 442         return core_waiters;
 443 }
 444
 445 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
 446 {
 447         struct core_thread *curr, *next;
 448         struct task_struct *task;
 449
 450         spin_lock_irq(&current->sighand->siglock);
 451         if (core_dumped && !__fatal_signal_pending(current))
 452                 current->signal->group_exit_code |= 0x80;
 453         current->signal->group_exit_task = NULL;
 454         current->signal->flags = SIGNAL_GROUP_EXIT;
 455         spin_unlock_irq(&current->sighand->siglock);
 456
 457         next = mm->core_state->dumper.next;
 458         while ((curr = next) != NULL) {
 459                 next = curr->next;
 460                 task = curr->task;
 461                 /*
 462                  * see exit_mm(), curr->task must not see
 463                  * ->task == NULL before we read ->next.
 464                  */
 465                 smp_mb();
 466                 curr->task = NULL;
 467                 wake_up_process(task);
 468         }
 469
 470         mm->core_state = NULL;
 471 }
 472
 473 static bool dump_interrupted(void)
 474 {
 475         /*
 476          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 477          * can do try_to_freeze() and check __fatal_signal_pending(),
 478          * but then we need to teach dump_write() to restart and clear
 479          * TIF_SIGPENDING.
 480          */
 481         return signal_pending(current);
 482 }
 483
 484 static void wait_for_dump_helpers(struct file *file)
 485 {
 486         struct pipe_inode_info *pipe = file->private_data;
 487
 488         pipe_lock(pipe);
 489         pipe->readers++;
 490         pipe->writers--;
 491         wake_up_interruptible_sync(&pipe->wait);
 492         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 493         pipe_unlock(pipe);
 494
 495         /*
 496          * We actually want wait_event_freezable() but then we need
 497          * to clear TIF_SIGPENDING and improve dump_interrupted().
 498          */
 499         wait_event_interruptible(pipe->wait, pipe->readers == 1);
 500
 501         pipe_lock(pipe);
 502         pipe->readers--;
 503         pipe->writers++;
 504         pipe_unlock(pipe);
 505 }
 506
 507 /*
 508  * umh_pipe_setup
 509  * helper function to customize the process used
 510  * to collect the core in userspace.  Specifically
 511  * it sets up a pipe and installs it as fd 0 (stdin)
 512  * for the process.  Returns 0 on success, or
 513  * PTR_ERR on failure.
 514  * Note that it also sets the core limit to 1.  This
 515  * is a special value that we use to trap recursive
 516  * core dumps
 517  */
 518 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 519 {
 520         struct file *files[2];
 521         struct coredump_params *cp = (struct coredump_params *)info->data;
 522         int err = create_pipe_files(files, 0);
 523         if (err)
 524                 return err;
 525
 526         cp->file = files[1];
 527
 528         err = replace_fd(0, files[0], 0);
 529         fput(files[0]);
 530         /* and disallow core files too */
 531         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 532
 533         return err;
 534 }
 535
 536 void do_coredump(const siginfo_t *siginfo)
 537 {
 538         struct core_state core_state;
 539         struct core_name cn;
 540         struct mm_struct *mm = current->mm;
 541         struct linux_binfmt * binfmt;
 542         const struct cred *old_cred;
 543         struct cred *cred;
 544         int retval = 0;
 545         int ispipe;
 546         struct files_struct *displaced;
 547         /* require nonrelative corefile path and be extra careful */
 548         bool need_suid_safe = false;
 549         bool core_dumped = false;
 550         static atomic_t core_dump_count = ATOMIC_INIT(0);
 551         struct coredump_params cprm = {
 552                 .siginfo = siginfo,
 553                 .regs = signal_pt_regs(),
 554                 .limit = rlimit(RLIMIT_CORE),
 555                 /*
 556                  * We must use the same mm->flags while dumping core to avoid
 557                  * inconsistency of bit flags, since this flag is not protected
 558                  * by any locks.
 559                  */
 560                 .mm_flags = mm->flags,
 561         };
 562
 563         audit_core_dumps(siginfo->si_signo);
 564
 565         binfmt = mm->binfmt;
 566         if (!binfmt || !binfmt->core_dump)
 567                 goto fail;
 568         if (!__get_dumpable(cprm.mm_flags))
 569                 goto fail;
 570
 571         cred = prepare_creds();
 572         if (!cred)
 573                 goto fail;
 574         /*
 575          * We cannot trust fsuid as being the "true" uid of the process
 576          * nor do we know its entire history. We only know it was tainted
 577          * so we dump it as root in mode 2, and only into a controlled
 578          * environment (pipe handler or fully qualified path).
 579          */
 580         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 581                 /* Setuid core dump mode */
 582                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 583                 need_suid_safe = true;
 584         }
 585
 586         retval = coredump_wait(siginfo->si_signo, &core_state);
 587         if (retval < 0)
 588                 goto fail_creds;
 589
 590         old_cred = override_creds(cred);
 591
 592         ispipe = format_corename(&cn, &cprm);
 593
 594         if (ispipe) {
 595                 int dump_count;
 596                 char **helper_argv;
 597                 struct subprocess_info *sub_info;
 598
 599                 if (ispipe < 0) {
 600                         printk(KERN_WARNING "format_corename failed\n");
 601                         printk(KERN_WARNING "Aborting core\n");
 602                         goto fail_unlock;
 603                 }
 604
 605                 if (cprm.limit == 1) {
 606                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 607                          *
 608                          * Normally core limits are irrelevant to pipes, since
 609                          * we're not writing to the file system, but we use
 610                          * cprm.limit of 1 here as a special value, this is a
 611                          * consistent way to catch recursive crashes.
 612                          * We can still crash if the core_pattern binary sets
 613                          * RLIM_CORE = !1, but it runs as root, and can do
 614                          * lots of stupid things.
 615                          *
 616                          * Note that we use task_tgid_vnr here to grab the pid
 617                          * of the process group leader.  That way we get the
 618                          * right pid if a thread in a multi-threaded
 619                          * core_pattern process dies.
 620                          */
 621                         printk(KERN_WARNING
 622                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 623                                 task_tgid_vnr(current), current->comm);
 624                         printk(KERN_WARNING "Aborting core\n");
 625                         goto fail_unlock;
 626                 }
 627                 cprm.limit = RLIM_INFINITY;
 628
 629                 dump_count = atomic_inc_return(&core_dump_count);
 630                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 631                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 632                                task_tgid_vnr(current), current->comm);
 633                         printk(KERN_WARNING "Skipping core dump\n");
 634                         goto fail_dropcount;
 635                 }
 636
 637                 helper_argv = argv_split(GFP_KERNEL, cn.corename, NULL);
 638                 if (!helper_argv) {
 639                         printk(KERN_WARNING "%s failed to allocate memory\n",
 640                                __func__);
 641                         goto fail_dropcount;
 642                 }
 643
 644                 retval = -ENOMEM;
 645                 sub_info = call_usermodehelper_setup(helper_argv[0],
 646                                                 helper_argv, NULL, GFP_KERNEL,
 647                                                 umh_pipe_setup, NULL, &cprm);
 648                 if (sub_info)
 649                         retval = call_usermodehelper_exec(sub_info,
 650                                                           UMH_WAIT_EXEC);
 651
 652                 argv_free(helper_argv);
 653                 if (retval) {
 654                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 655                                cn.corename);
 656                         goto close_fail;
 657                 }
 658         } else {
 659                 struct inode *inode;
 660                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 661                                  O_LARGEFILE | O_EXCL;
 662
 663                 if (cprm.limit < binfmt->min_coredump)
 664                         goto fail_unlock;
 665
 666                 if (need_suid_safe && cn.corename[0] != '/') {
 667                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 668                                 "to fully qualified path!\n",
 669                                 task_tgid_vnr(current), current->comm);
 670                         printk(KERN_WARNING "Skipping core dump\n");
 671                         goto fail_unlock;
 672                 }
 673
 674                 /*
 675                  * Unlink the file if it exists unless this is a SUID
 676                  * binary - in that case, we're running around with root
 677                  * privs and don't want to unlink another user's coredump.
 678                  */
 679                 if (!need_suid_safe) {
 680                         mm_segment_t old_fs;
 681
 682                         old_fs = get_fs();
 683                         set_fs(KERNEL_DS);
 684                         /*
 685                          * If it doesn't exist, that's fine. If there's some
 686                          * other problem, we'll catch it at the filp_open().
 687                          */
 688                         (void) sys_unlink((const char __user *)cn.corename);
 689                         set_fs(old_fs);
 690                 }
 691
 692                 /*
 693                  * There is a race between unlinking and creating the
 694                  * file, but if that causes an EEXIST here, that's
 695                  * fine - another process raced with us while creating
 696                  * the corefile, and the other process won. To userspace,
 697                  * what matters is that at least one of the two processes
 698                  * writes its coredump successfully, not which one.
 699                  */
 700                 if (need_suid_safe) {
 701                         /*
 702                          * Using user namespaces, normal user tasks can change
 703                          * their current->fs->root to point to arbitrary
 704                          * directories. Since the intention of the "only dump
 705                          * with a fully qualified path" rule is to control where
 706                          * coredumps may be placed using root privileges,
 707                          * current->fs->root must not be used. Instead, use the
 708                          * root directory of init_task.
 709                          */
 710                         struct path root;
 711
 712                         task_lock(&init_task);
 713                         get_fs_root(init_task.fs, &root);
 714                         task_unlock(&init_task);
 715                         cprm.file = file_open_root(root.dentry, root.mnt,
 716                                 cn.corename, open_flags, 0600);
 717                         path_put(&root);
 718                 } else {
 719                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 720                 }
 721                 if (IS_ERR(cprm.file))
 722                         goto fail_unlock;
 723
 724                 inode = file_inode(cprm.file);
 725                 if (inode->i_nlink > 1)
 726                         goto close_fail;
 727                 if (d_unhashed(cprm.file->f_path.dentry))
 728                         goto close_fail;
 729                 /*
 730                  * AK: actually i see no reason to not allow this for named
 731                  * pipes etc, but keep the previous behaviour for now.
 732                  */
 733                 if (!S_ISREG(inode->i_mode))
 734                         goto close_fail;
 735                 /*
 736                  * Don't dump core if the filesystem changed owner or mode
 737                  * of the file during file creation. This is an issue when
 738                  * a process dumps core while its cwd is e.g. on a vfat
 739                  * filesystem.
 740                  */
 741                 if (!uid_eq(inode->i_uid, current_fsuid()))
 742                         goto close_fail;
 743                 if ((inode->i_mode & 0677) != 0600)
 744                         goto close_fail;
 745                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 746                         goto close_fail;
 747                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
 748                         goto close_fail;
 749         }
 750
 751         /* get us an unshared descriptor table; almost always a no-op */
 752         retval = unshare_files(&displaced);
 753         if (retval)
 754                 goto close_fail;
 755         if (displaced)
 756                 put_files_struct(displaced);
 757         if (!dump_interrupted()) {
 758                 file_start_write(cprm.file);
 759                 core_dumped = binfmt->core_dump(&cprm);
 760                 file_end_write(cprm.file);
 761         }
 762         if (ispipe && core_pipe_limit)
 763                 wait_for_dump_helpers(cprm.file);
 764 close_fail:
 765         if (cprm.file)
 766                 filp_close(cprm.file, NULL);
 767 fail_dropcount:
 768         if (ispipe)
 769                 atomic_dec(&core_dump_count);
 770 fail_unlock:
 771         kfree(cn.corename);
 772         coredump_finish(mm, core_dumped);
 773         revert_creds(old_cred);
 774 fail_creds:
 775         put_cred(cred);
 776 fail:
 777         return;
 778 }
 779
 780 /*
 781  * Core dumping helper functions.  These are the only things you should
 782  * do on a core-file: use only these functions to write out all the
 783  * necessary info.
 784  */
 785 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 786 {
 787         struct file *file = cprm->file;
 788         loff_t pos = file->f_pos;
 789         ssize_t n;
 790         if (cprm->written + nr > cprm->limit)
 791                 return 0;
 792         while (nr) {
 793                 if (dump_interrupted())
 794                         return 0;
 795                 n = __kernel_write(file, addr, nr, &pos);
 796                 if (n <= 0)
 797                         return 0;
 798                 file->f_pos = pos;
 799                 cprm->written += n;
 800                 cprm->pos += n;
 801                 nr -= n;
 802         }
 803         return 1;
 804 }
 805 EXPORT_SYMBOL(dump_emit);
 806
 807 int dump_skip(struct coredump_params *cprm, size_t nr)
 808 {
 809         static char zeroes[PAGE_SIZE];
 810         struct file *file = cprm->file;
 811         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 812                 if (dump_interrupted() ||
 813                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
 814                         return 0;
 815                 cprm->pos += nr;
 816                 return 1;
 817         } else {
 818                 while (nr > PAGE_SIZE) {
 819                         if (!dump_emit(cprm, zeroes, PAGE_SIZE))
 820                                 return 0;
 821                         nr -= PAGE_SIZE;
 822                 }
 823                 return dump_emit(cprm, zeroes, nr);
 824         }
 825 }
 826 EXPORT_SYMBOL(dump_skip);
 827
 828 int dump_align(struct coredump_params *cprm, int align)
 829 {
 830         unsigned mod = cprm->pos & (align - 1);
 831         if (align & (align - 1))
 832                 return 0;
 833         return mod ? dump_skip(cprm, align - mod) : 1;
 834 }
 835 EXPORT_SYMBOL(dump_align);
 836
 837 /*
 838  * Ensures that file size is big enough to contain the current file
 839  * postion. This prevents gdb from complaining about a truncated file
 840  * if the last "write" to the file was dump_skip.
 841  */
 842 void dump_truncate(struct coredump_params *cprm)
 843 {
 844         struct file *file = cprm->file;
 845         loff_t offset;
 846
 847         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 848                 offset = file->f_op->llseek(file, 0, SEEK_CUR);
 849                 if (i_size_read(file->f_mapping->host) < offset)
 850                         do_truncate(file->f_path.dentry, offset, 0, file);
 851         }
 852 }
 853 EXPORT_SYMBOL(dump_truncate);