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