fs/coredump.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/slab.h>
   3 #include <linux/file.h>
   4 #include <linux/fdtable.h>
   5 #include <linux/freezer.h>
   6 #include <linux/mm.h>
   7 #include <linux/stat.h>
   8 #include <linux/fcntl.h>
   9 #include <linux/swap.h>
  10 #include <linux/string.h>
  11 #include <linux/init.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/perf_event.h>
  14 #include <linux/highmem.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/key.h>
  17 #include <linux/personality.h>
  18 #include <linux/binfmts.h>
  19 #include <linux/coredump.h>
  20 #include <linux/sched/coredump.h>
  21 #include <linux/sched/signal.h>
  22 #include <linux/sched/task_stack.h>
  23 #include <linux/utsname.h>
  24 #include <linux/pid_namespace.h>
  25 #include <linux/module.h>
  26 #include <linux/namei.h>
  27 #include <linux/mount.h>
  28 #include <linux/security.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/tsacct_kern.h>
  31 #include <linux/cn_proc.h>
  32 #include <linux/audit.h>
  33 #include <linux/tracehook.h>
  34 #include <linux/kmod.h>
  35 #include <linux/fsnotify.h>
  36 #include <linux/fs_struct.h>
  37 #include <linux/pipe_fs_i.h>
  38 #include <linux/oom.h>
  39 #include <linux/compat.h>
  40 #include <linux/fs.h>
  41 #include <linux/path.h>
  42 #include <linux/timekeeping.h>
  43
  44 #include <linux/uaccess.h>
  45 #include <asm/mmu_context.h>
  46 #include <asm/tlb.h>
  47 #include <asm/exec.h>
  48
  49 #include <trace/events/task.h>
  50 #include "internal.h"
  51
  52 #include <trace/events/sched.h>
  53
  54 int core_uses_pid;
  55 unsigned int core_pipe_limit;
  56 char core_pattern[CORENAME_MAX_SIZE] = "core";
  57 static int core_name_size = CORENAME_MAX_SIZE;
  58
  59 struct core_name {
  60         char *corename;
  61         int used, size;
  62 };
  63
  64 /* The maximal length of core_pattern is also specified in sysctl.c */
  65
  66 static int expand_corename(struct core_name *cn, int size)
  67 {
  68         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  69
  70         if (!corename)
  71                 return -ENOMEM;
  72
  73         if (size > core_name_size) /* racy but harmless */
  74                 core_name_size = size;
  75
  76         cn->size = ksize(corename);
  77         cn->corename = corename;
  78         return 0;
  79 }
  80
  81 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  82                                      va_list arg)
  83 {
  84         int free, need;
  85         va_list arg_copy;
  86
  87 again:
  88         free = cn->size - cn->used;
  89
  90         va_copy(arg_copy, arg);
  91         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  92         va_end(arg_copy);
  93
  94         if (need < free) {
  95                 cn->used += need;
  96                 return 0;
  97         }
  98
  99         if (!expand_corename(cn, cn->size + need - free + 1))
 100                 goto again;
 101
 102         return -ENOMEM;
 103 }
 104
 105 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 106 {
 107         va_list arg;
 108         int ret;
 109
 110         va_start(arg, fmt);
 111         ret = cn_vprintf(cn, fmt, arg);
 112         va_end(arg);
 113
 114         return ret;
 115 }
 116
 117 static __printf(2, 3)
 118 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 119 {
 120         int cur = cn->used;
 121         va_list arg;
 122         int ret;
 123
 124         va_start(arg, fmt);
 125         ret = cn_vprintf(cn, fmt, arg);
 126         va_end(arg);
 127
 128         if (ret == 0) {
 129                 /*
 130                  * Ensure that this coredump name component can't cause the
 131                  * resulting corefile path to consist of a ".." or ".".
 132                  */
 133                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 134                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 135                                 && cn->corename[cur+1] == '.'))
 136                         cn->corename[cur] = '!';
 137
 138                 /*
 139                  * Empty names are fishy and could be used to create a "//" in a
 140                  * corefile name, causing the coredump to happen one directory
 141                  * level too high. Enforce that all components of the core
 142                  * pattern are at least one character long.
 143                  */
 144                 if (cn->used == cur)
 145                         ret = cn_printf(cn, "!");
 146         }
 147
 148         for (; cur < cn->used; ++cur) {
 149                 if (cn->corename[cur] == '/')
 150                         cn->corename[cur] = '!';
 151         }
 152         return ret;
 153 }
 154
 155 static int cn_print_exe_file(struct core_name *cn)
 156 {
 157         struct file *exe_file;
 158         char *pathbuf, *path;
 159         int ret;
 160
 161         exe_file = get_mm_exe_file(current->mm);
 162         if (!exe_file)
 163                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 164
 165         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 166         if (!pathbuf) {
 167                 ret = -ENOMEM;
 168                 goto put_exe_file;
 169         }
 170
 171         path = file_path(exe_file, pathbuf, PATH_MAX);
 172         if (IS_ERR(path)) {
 173                 ret = PTR_ERR(path);
 174                 goto free_buf;
 175         }
 176
 177         ret = cn_esc_printf(cn, "%s", path);
 178
 179 free_buf:
 180         kfree(pathbuf);
 181 put_exe_file:
 182         fput(exe_file);
 183         return ret;
 184 }
 185
 186 /* format_corename will inspect the pattern parameter, and output a
 187  * name into corename, which must have space for at least
 188  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 189  */
 190 static int format_corename(struct core_name *cn, struct coredump_params *cprm)
 191 {
 192         const struct cred *cred = current_cred();
 193         const char *pat_ptr = core_pattern;
 194         int ispipe = (*pat_ptr == '|');
 195         int pid_in_pattern = 0;
 196         int err = 0;
 197
 198         cn->used = 0;
 199         cn->corename = NULL;
 200         if (expand_corename(cn, core_name_size))
 201                 return -ENOMEM;
 202         cn->corename[0] = '\0';
 203
 204         if (ispipe)
 205                 ++pat_ptr;
 206
 207         /* Repeat as long as we have more pattern to process and more output
 208            space */
 209         while (*pat_ptr) {
 210                 if (*pat_ptr != '%') {
 211                         err = cn_printf(cn, "%c", *pat_ptr++);
 212                 } else {
 213                         switch (*++pat_ptr) {
 214                         /* single % at the end, drop that */
 215                         case 0:
 216                                 goto out;
 217                         /* Double percent, output one percent */
 218                         case '%':
 219                                 err = cn_printf(cn, "%c", '%');
 220                                 break;
 221                         /* pid */
 222                         case 'p':
 223                                 pid_in_pattern = 1;
 224                                 err = cn_printf(cn, "%d",
 225                                               task_tgid_vnr(current));
 226                                 break;
 227                         /* global pid */
 228                         case 'P':
 229                                 err = cn_printf(cn, "%d",
 230                                               task_tgid_nr(current));
 231                                 break;
 232                         case 'i':
 233                                 err = cn_printf(cn, "%d",
 234                                               task_pid_vnr(current));
 235                                 break;
 236                         case 'I':
 237                                 err = cn_printf(cn, "%d",
 238                                               task_pid_nr(current));
 239                                 break;
 240                         /* uid */
 241                         case 'u':
 242                                 err = cn_printf(cn, "%u",
 243                                                 from_kuid(&init_user_ns,
 244                                                           cred->uid));
 245                                 break;
 246                         /* gid */
 247                         case 'g':
 248                                 err = cn_printf(cn, "%u",
 249                                                 from_kgid(&init_user_ns,
 250                                                           cred->gid));
 251                                 break;
 252                         case 'd':
 253                                 err = cn_printf(cn, "%d",
 254                                         __get_dumpable(cprm->mm_flags));
 255                                 break;
 256                         /* signal that caused the coredump */
 257                         case 's':
 258                                 err = cn_printf(cn, "%d",
 259                                                 cprm->siginfo->si_signo);
 260                                 break;
 261                         /* UNIX time of coredump */
 262                         case 't': {
 263                                 time64_t time;
 264
 265                                 time = ktime_get_real_seconds();
 266                                 err = cn_printf(cn, "%lld", time);
 267                                 break;
 268                         }
 269                         /* hostname */
 270                         case 'h':
 271                                 down_read(&uts_sem);
 272                                 err = cn_esc_printf(cn, "%s",
 273                                               utsname()->nodename);
 274                                 up_read(&uts_sem);
 275                                 break;
 276                         /* executable */
 277                         case 'e':
 278                                 err = cn_esc_printf(cn, "%s", current->comm);
 279                                 break;
 280                         case 'E':
 281                                 err = cn_print_exe_file(cn);
 282                                 break;
 283                         /* core limit size */
 284                         case 'c':
 285                                 err = cn_printf(cn, "%lu",
 286                                               rlimit(RLIMIT_CORE));
 287                                 break;
 288                         default:
 289                                 break;
 290                         }
 291                         ++pat_ptr;
 292                 }
 293
 294                 if (err)
 295                         return err;
 296         }
 297
 298 out:
 299         /* Backward compatibility with core_uses_pid:
 300          *
 301          * If core_pattern does not include a %p (as is the default)
 302          * and core_uses_pid is set, then .%pid will be appended to
 303          * the filename. Do not do this for piped commands. */
 304         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 305                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 306                 if (err)
 307                         return err;
 308         }
 309         return ispipe;
 310 }
 311
 312 static int zap_process(struct task_struct *start, int exit_code, int flags)
 313 {
 314         struct task_struct *t;
 315         int nr = 0;
 316
 317         /* ignore all signals except SIGKILL, see prepare_signal() */
 318         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
 319         start->signal->group_exit_code = exit_code;
 320         start->signal->group_stop_count = 0;
 321
 322         for_each_thread(start, t) {
 323                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 324                 if (t != current && t->mm) {
 325                         sigaddset(&t->pending.signal, SIGKILL);
 326                         signal_wake_up(t, 1);
 327                         nr++;
 328                 }
 329         }
 330
 331         return nr;
 332 }
 333
 334 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 335                         struct core_state *core_state, int exit_code)
 336 {
 337         struct task_struct *g, *p;
 338         unsigned long flags;
 339         int nr = -EAGAIN;
 340
 341         spin_lock_irq(&tsk->sighand->siglock);
 342         if (!signal_group_exit(tsk->signal)) {
 343                 mm->core_state = core_state;
 344                 tsk->signal->group_exit_task = tsk;
 345                 nr = zap_process(tsk, exit_code, 0);
 346                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 347         }
 348         spin_unlock_irq(&tsk->sighand->siglock);
 349         if (unlikely(nr < 0))
 350                 return nr;
 351
 352         tsk->flags |= PF_DUMPCORE;
 353         if (atomic_read(&mm->mm_users) == nr + 1)
 354                 goto done;
 355         /*
 356          * We should find and kill all tasks which use this mm, and we should
 357          * count them correctly into ->nr_threads. We don't take tasklist
 358          * lock, but this is safe wrt:
 359          *
 360          * fork:
 361          *      None of sub-threads can fork after zap_process(leader). All
 362          *      processes which were created before this point should be
 363          *      visible to zap_threads() because copy_process() adds the new
 364          *      process to the tail of init_task.tasks list, and lock/unlock
 365          *      of ->siglock provides a memory barrier.
 366          *
 367          * do_exit:
 368          *      The caller holds mm->mmap_sem. This means that the task which
 369          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 370          *      its ->mm.
 371          *
 372          * de_thread:
 373          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 374          *      we must see either old or new leader, this does not matter.
 375          *      However, it can change p->sighand, so lock_task_sighand(p)
 376          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
 377          *      it can't fail.
 378          *
 379          *      Note also that "g" can be the old leader with ->mm == NULL
 380          *      and already unhashed and thus removed from ->thread_group.
 381          *      This is OK, __unhash_process()->list_del_rcu() does not
 382          *      clear the ->next pointer, we will find the new leader via
 383          *      next_thread().
 384          */
 385         rcu_read_lock();
 386         for_each_process(g) {
 387                 if (g == tsk->group_leader)
 388                         continue;
 389                 if (g->flags & PF_KTHREAD)
 390                         continue;
 391
 392                 for_each_thread(g, p) {
 393                         if (unlikely(!p->mm))
 394                                 continue;
 395                         if (unlikely(p->mm == mm)) {
 396                                 lock_task_sighand(p, &flags);
 397                                 nr += zap_process(p, exit_code,
 398                                                         SIGNAL_GROUP_EXIT);
 399                                 unlock_task_sighand(p, &flags);
 400                         }
 401                         break;
 402                 }
 403         }
 404         rcu_read_unlock();
 405 done:
 406         atomic_set(&core_state->nr_threads, nr);
 407         return nr;
 408 }
 409
 410 static int coredump_wait(int exit_code, struct core_state *core_state)
 411 {
 412         struct task_struct *tsk = current;
 413         struct mm_struct *mm = tsk->mm;
 414         int core_waiters = -EBUSY;
 415
 416         init_completion(&core_state->startup);
 417         core_state->dumper.task = tsk;
 418         core_state->dumper.next = NULL;
 419
 420         if (down_write_killable(&mm->mmap_sem))
 421                 return -EINTR;
 422
 423         if (!mm->core_state)
 424                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 425         up_write(&mm->mmap_sem);
 426
 427         if (core_waiters > 0) {
 428                 struct core_thread *ptr;
 429
 430                 freezer_do_not_count();
 431                 wait_for_completion(&core_state->startup);
 432                 freezer_count();
 433                 /*
 434                  * Wait for all the threads to become inactive, so that
 435                  * all the thread context (extended register state, like
 436                  * fpu etc) gets copied to the memory.
 437                  */
 438                 ptr = core_state->dumper.next;
 439                 while (ptr != NULL) {
 440                         wait_task_inactive(ptr->task, 0);
 441                         ptr = ptr->next;
 442                 }
 443         }
 444
 445         return core_waiters;
 446 }
 447
 448 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
 449 {
 450         struct core_thread *curr, *next;
 451         struct task_struct *task;
 452
 453         spin_lock_irq(&current->sighand->siglock);
 454         if (core_dumped && !__fatal_signal_pending(current))
 455                 current->signal->group_exit_code |= 0x80;
 456         current->signal->group_exit_task = NULL;
 457         current->signal->flags = SIGNAL_GROUP_EXIT;
 458         spin_unlock_irq(&current->sighand->siglock);
 459
 460         next = mm->core_state->dumper.next;
 461         while ((curr = next) != NULL) {
 462                 next = curr->next;
 463                 task = curr->task;
 464                 /*
 465                  * see exit_mm(), curr->task must not see
 466                  * ->task == NULL before we read ->next.
 467                  */
 468                 smp_mb();
 469                 curr->task = NULL;
 470                 wake_up_process(task);
 471         }
 472
 473         mm->core_state = NULL;
 474 }
 475
 476 static bool dump_interrupted(void)
 477 {
 478         /*
 479          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 480          * can do try_to_freeze() and check __fatal_signal_pending(),
 481          * but then we need to teach dump_write() to restart and clear
 482          * TIF_SIGPENDING.
 483          */
 484         return signal_pending(current);
 485 }
 486
 487 static void wait_for_dump_helpers(struct file *file)
 488 {
 489         struct pipe_inode_info *pipe = file->private_data;
 490
 491         pipe_lock(pipe);
 492         pipe->readers++;
 493         pipe->writers--;
 494         wake_up_interruptible_sync(&pipe->wait);
 495         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 496         pipe_unlock(pipe);
 497
 498         /*
 499          * We actually want wait_event_freezable() but then we need
 500          * to clear TIF_SIGPENDING and improve dump_interrupted().
 501          */
 502         wait_event_interruptible(pipe->wait, pipe->readers == 1);
 503
 504         pipe_lock(pipe);
 505         pipe->readers--;
 506         pipe->writers++;
 507         pipe_unlock(pipe);
 508 }
 509
 510 /*
 511  * umh_pipe_setup
 512  * helper function to customize the process used
 513  * to collect the core in userspace.  Specifically
 514  * it sets up a pipe and installs it as fd 0 (stdin)
 515  * for the process.  Returns 0 on success, or
 516  * PTR_ERR on failure.
 517  * Note that it also sets the core limit to 1.  This
 518  * is a special value that we use to trap recursive
 519  * core dumps
 520  */
 521 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 522 {
 523         struct file *files[2];
 524         struct coredump_params *cp = (struct coredump_params *)info->data;
 525         int err = create_pipe_files(files, 0);
 526         if (err)
 527                 return err;
 528
 529         cp->file = files[1];
 530
 531         err = replace_fd(0, files[0], 0);
 532         fput(files[0]);
 533         /* and disallow core files too */
 534         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 535
 536         return err;
 537 }
 538
 539 void do_coredump(const kernel_siginfo_t *siginfo)
 540 {
 541         struct core_state core_state;
 542         struct core_name cn;
 543         struct mm_struct *mm = current->mm;
 544         struct linux_binfmt * binfmt;
 545         const struct cred *old_cred;
 546         struct cred *cred;
 547         int retval = 0;
 548         int ispipe;
 549         struct files_struct *displaced;
 550         /* require nonrelative corefile path and be extra careful */
 551         bool need_suid_safe = false;
 552         bool core_dumped = false;
 553         static atomic_t core_dump_count = ATOMIC_INIT(0);
 554         struct coredump_params cprm = {
 555                 .siginfo = siginfo,
 556                 .regs = signal_pt_regs(),
 557                 .limit = rlimit(RLIMIT_CORE),
 558                 /*
 559                  * We must use the same mm->flags while dumping core to avoid
 560                  * inconsistency of bit flags, since this flag is not protected
 561                  * by any locks.
 562                  */
 563                 .mm_flags = mm->flags,
 564         };
 565
 566         audit_core_dumps(siginfo->si_signo);
 567
 568         binfmt = mm->binfmt;
 569         if (!binfmt || !binfmt->core_dump)
 570                 goto fail;
 571         if (!__get_dumpable(cprm.mm_flags))
 572                 goto fail;
 573
 574         cred = prepare_creds();
 575         if (!cred)
 576                 goto fail;
 577         /*
 578          * We cannot trust fsuid as being the "true" uid of the process
 579          * nor do we know its entire history. We only know it was tainted
 580          * so we dump it as root in mode 2, and only into a controlled
 581          * environment (pipe handler or fully qualified path).
 582          */
 583         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 584                 /* Setuid core dump mode */
 585                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 586                 need_suid_safe = true;
 587         }
 588
 589         retval = coredump_wait(siginfo->si_signo, &core_state);
 590         if (retval < 0)
 591                 goto fail_creds;
 592
 593         old_cred = override_creds(cred);
 594
 595         ispipe = format_corename(&cn, &cprm);
 596
 597         if (ispipe) {
 598                 int dump_count;
 599                 char **helper_argv;
 600                 struct subprocess_info *sub_info;
 601
 602                 if (ispipe < 0) {
 603                         printk(KERN_WARNING "format_corename failed\n");
 604                         printk(KERN_WARNING "Aborting core\n");
 605                         goto fail_unlock;
 606                 }
 607
 608                 if (cprm.limit == 1) {
 609                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 610                          *
 611                          * Normally core limits are irrelevant to pipes, since
 612                          * we're not writing to the file system, but we use
 613                          * cprm.limit of 1 here as a special value, this is a
 614                          * consistent way to catch recursive crashes.
 615                          * We can still crash if the core_pattern binary sets
 616                          * RLIM_CORE = !1, but it runs as root, and can do
 617                          * lots of stupid things.
 618                          *
 619                          * Note that we use task_tgid_vnr here to grab the pid
 620                          * of the process group leader.  That way we get the
 621                          * right pid if a thread in a multi-threaded
 622                          * core_pattern process dies.
 623                          */
 624                         printk(KERN_WARNING
 625                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 626                                 task_tgid_vnr(current), current->comm);
 627                         printk(KERN_WARNING "Aborting core\n");
 628                         goto fail_unlock;
 629                 }
 630                 cprm.limit = RLIM_INFINITY;
 631
 632                 dump_count = atomic_inc_return(&core_dump_count);
 633                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 634                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 635                                task_tgid_vnr(current), current->comm);
 636                         printk(KERN_WARNING "Skipping core dump\n");
 637                         goto fail_dropcount;
 638                 }
 639
 640                 helper_argv = argv_split(GFP_KERNEL, cn.corename, NULL);
 641                 if (!helper_argv) {
 642                         printk(KERN_WARNING "%s failed to allocate memory\n",
 643                                __func__);
 644                         goto fail_dropcount;
 645                 }
 646
 647                 retval = -ENOMEM;
 648                 sub_info = call_usermodehelper_setup(helper_argv[0],
 649                                                 helper_argv, NULL, GFP_KERNEL,
 650                                                 umh_pipe_setup, NULL, &cprm);
 651                 if (sub_info)
 652                         retval = call_usermodehelper_exec(sub_info,
 653                                                           UMH_WAIT_EXEC);
 654
 655                 argv_free(helper_argv);
 656                 if (retval) {
 657                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 658                                cn.corename);
 659                         goto close_fail;
 660                 }
 661         } else {
 662                 struct inode *inode;
 663                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 664                                  O_LARGEFILE | O_EXCL;
 665
 666                 if (cprm.limit < binfmt->min_coredump)
 667                         goto fail_unlock;
 668
 669                 if (need_suid_safe && cn.corename[0] != '/') {
 670                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 671                                 "to fully qualified path!\n",
 672                                 task_tgid_vnr(current), current->comm);
 673                         printk(KERN_WARNING "Skipping core dump\n");
 674                         goto fail_unlock;
 675                 }
 676
 677                 /*
 678                  * Unlink the file if it exists unless this is a SUID
 679                  * binary - in that case, we're running around with root
 680                  * privs and don't want to unlink another user's coredump.
 681                  */
 682                 if (!need_suid_safe) {
 683                         /*
 684                          * If it doesn't exist, that's fine. If there's some
 685                          * other problem, we'll catch it at the filp_open().
 686                          */
 687                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 688                 }
 689
 690                 /*
 691                  * There is a race between unlinking and creating the
 692                  * file, but if that causes an EEXIST here, that's
 693                  * fine - another process raced with us while creating
 694                  * the corefile, and the other process won. To userspace,
 695                  * what matters is that at least one of the two processes
 696                  * writes its coredump successfully, not which one.
 697                  */
 698                 if (need_suid_safe) {
 699                         /*
 700                          * Using user namespaces, normal user tasks can change
 701                          * their current->fs->root to point to arbitrary
 702                          * directories. Since the intention of the "only dump
 703                          * with a fully qualified path" rule is to control where
 704                          * coredumps may be placed using root privileges,
 705                          * current->fs->root must not be used. Instead, use the
 706                          * root directory of init_task.
 707                          */
 708                         struct path root;
 709
 710                         task_lock(&init_task);
 711                         get_fs_root(init_task.fs, &root);
 712                         task_unlock(&init_task);
 713                         cprm.file = file_open_root(root.dentry, root.mnt,
 714                                 cn.corename, open_flags, 0600);
 715                         path_put(&root);
 716                 } else {
 717                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 718                 }
 719                 if (IS_ERR(cprm.file))
 720                         goto fail_unlock;
 721
 722                 inode = file_inode(cprm.file);
 723                 if (inode->i_nlink > 1)
 724                         goto close_fail;
 725                 if (d_unhashed(cprm.file->f_path.dentry))
 726                         goto close_fail;
 727                 /*
 728                  * AK: actually i see no reason to not allow this for named
 729                  * pipes etc, but keep the previous behaviour for now.
 730                  */
 731                 if (!S_ISREG(inode->i_mode))
 732                         goto close_fail;
 733                 /*
 734                  * Don't dump core if the filesystem changed owner or mode
 735                  * of the file during file creation. This is an issue when
 736                  * a process dumps core while its cwd is e.g. on a vfat
 737                  * filesystem.
 738                  */
 739                 if (!uid_eq(inode->i_uid, current_fsuid()))
 740                         goto close_fail;
 741                 if ((inode->i_mode & 0677) != 0600)
 742                         goto close_fail;
 743                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 744                         goto close_fail;
 745                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
 746                         goto close_fail;
 747         }
 748
 749         /* get us an unshared descriptor table; almost always a no-op */
 750         retval = unshare_files(&displaced);
 751         if (retval)
 752                 goto close_fail;
 753         if (displaced)
 754                 put_files_struct(displaced);
 755         if (!dump_interrupted()) {
 756                 file_start_write(cprm.file);
 757                 core_dumped = binfmt->core_dump(&cprm);
 758                 file_end_write(cprm.file);
 759         }
 760         if (ispipe && core_pipe_limit)
 761                 wait_for_dump_helpers(cprm.file);
 762 close_fail:
 763         if (cprm.file)
 764                 filp_close(cprm.file, NULL);
 765 fail_dropcount:
 766         if (ispipe)
 767                 atomic_dec(&core_dump_count);
 768 fail_unlock:
 769         kfree(cn.corename);
 770         coredump_finish(mm, core_dumped);
 771         revert_creds(old_cred);
 772 fail_creds:
 773         put_cred(cred);
 774 fail:
 775         return;
 776 }
 777
 778 /*
 779  * Core dumping helper functions.  These are the only things you should
 780  * do on a core-file: use only these functions to write out all the
 781  * necessary info.
 782  */
 783 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 784 {
 785         struct file *file = cprm->file;
 786         loff_t pos = file->f_pos;
 787         ssize_t n;
 788         if (cprm->written + nr > cprm->limit)
 789                 return 0;
 790         while (nr) {
 791                 if (dump_interrupted())
 792                         return 0;
 793                 n = __kernel_write(file, addr, nr, &pos);
 794                 if (n <= 0)
 795                         return 0;
 796                 file->f_pos = pos;
 797                 cprm->written += n;
 798                 cprm->pos += n;
 799                 nr -= n;
 800         }
 801         return 1;
 802 }
 803 EXPORT_SYMBOL(dump_emit);
 804
 805 int dump_skip(struct coredump_params *cprm, size_t nr)
 806 {
 807         static char zeroes[PAGE_SIZE];
 808         struct file *file = cprm->file;
 809         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 810                 if (dump_interrupted() ||
 811                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
 812                         return 0;
 813                 cprm->pos += nr;
 814                 return 1;
 815         } else {
 816                 while (nr > PAGE_SIZE) {
 817                         if (!dump_emit(cprm, zeroes, PAGE_SIZE))
 818                                 return 0;
 819                         nr -= PAGE_SIZE;
 820                 }
 821                 return dump_emit(cprm, zeroes, nr);
 822         }
 823 }
 824 EXPORT_SYMBOL(dump_skip);
 825
 826 int dump_align(struct coredump_params *cprm, int align)
 827 {
 828         unsigned mod = cprm->pos & (align - 1);
 829         if (align & (align - 1))
 830                 return 0;
 831         return mod ? dump_skip(cprm, align - mod) : 1;
 832 }
 833 EXPORT_SYMBOL(dump_align);
 834
 835 /*
 836  * Ensures that file size is big enough to contain the current file
 837  * postion. This prevents gdb from complaining about a truncated file
 838  * if the last "write" to the file was dump_skip.
 839  */
 840 void dump_truncate(struct coredump_params *cprm)
 841 {
 842         struct file *file = cprm->file;
 843         loff_t offset;
 844
 845         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 846                 offset = file->f_op->llseek(file, 0, SEEK_CUR);
 847                 if (i_size_read(file->f_mapping->host) < offset)
 848                         do_truncate(file->f_path.dentry, offset, 0, file);
 849         }
 850 }
 851 EXPORT_SYMBOL(dump_truncate);