fs/coredump.c

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