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/ctype.h>
  11 #include <linux/string.h>
  12 #include <linux/init.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/perf_event.h>
  15 #include <linux/highmem.h>
  16 #include <linux/spinlock.h>
  17 #include <linux/key.h>
  18 #include <linux/personality.h>
  19 #include <linux/binfmts.h>
  20 #include <linux/coredump.h>
  21 #include <linux/sched/coredump.h>
  22 #include <linux/sched/signal.h>
  23 #include <linux/sched/task_stack.h>
  24 #include <linux/utsname.h>
  25 #include <linux/pid_namespace.h>
  26 #include <linux/module.h>
  27 #include <linux/namei.h>
  28 #include <linux/mount.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/tsacct_kern.h>
  32 #include <linux/cn_proc.h>
  33 #include <linux/audit.h>
  34 #include <linux/tracehook.h>
  35 #include <linux/kmod.h>
  36 #include <linux/fsnotify.h>
  37 #include <linux/fs_struct.h>
  38 #include <linux/pipe_fs_i.h>
  39 #include <linux/oom.h>
  40 #include <linux/compat.h>
  41 #include <linux/fs.h>
  42 #include <linux/path.h>
  43 #include <linux/timekeeping.h>
  44
  45 #include <linux/uaccess.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/tlb.h>
  48 #include <asm/exec.h>
  49
  50 #include <trace/events/task.h>
  51 #include "internal.h"
  52
  53 #include <trace/events/sched.h>
  54
  55 int core_uses_pid;
  56 unsigned int core_pipe_limit;
  57 char core_pattern[CORENAME_MAX_SIZE] = "core";
  58 static int core_name_size = CORENAME_MAX_SIZE;
  59
  60 struct core_name {
  61         char *corename;
  62         int used, size;
  63 };
  64
  65 /* The maximal length of core_pattern is also specified in sysctl.c */
  66
  67 static int expand_corename(struct core_name *cn, int size)
  68 {
  69         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  70
  71         if (!corename)
  72                 return -ENOMEM;
  73
  74         if (size > core_name_size) /* racy but harmless */
  75                 core_name_size = size;
  76
  77         cn->size = ksize(corename);
  78         cn->corename = corename;
  79         return 0;
  80 }
  81
  82 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  83                                      va_list arg)
  84 {
  85         int free, need;
  86         va_list arg_copy;
  87
  88 again:
  89         free = cn->size - cn->used;
  90
  91         va_copy(arg_copy, arg);
  92         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  93         va_end(arg_copy);
  94
  95         if (need < free) {
  96                 cn->used += need;
  97                 return 0;
  98         }
  99
 100         if (!expand_corename(cn, cn->size + need - free + 1))
 101                 goto again;
 102
 103         return -ENOMEM;
 104 }
 105
 106 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 107 {
 108         va_list arg;
 109         int ret;
 110
 111         va_start(arg, fmt);
 112         ret = cn_vprintf(cn, fmt, arg);
 113         va_end(arg);
 114
 115         return ret;
 116 }
 117
 118 static __printf(2, 3)
 119 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 120 {
 121         int cur = cn->used;
 122         va_list arg;
 123         int ret;
 124
 125         va_start(arg, fmt);
 126         ret = cn_vprintf(cn, fmt, arg);
 127         va_end(arg);
 128
 129         if (ret == 0) {
 130                 /*
 131                  * Ensure that this coredump name component can't cause the
 132                  * resulting corefile path to consist of a ".." or ".".
 133                  */
 134                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 135                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 136                                 && cn->corename[cur+1] == '.'))
 137                         cn->corename[cur] = '!';
 138
 139                 /*
 140                  * Empty names are fishy and could be used to create a "//" in a
 141                  * corefile name, causing the coredump to happen one directory
 142                  * level too high. Enforce that all components of the core
 143                  * pattern are at least one character long.
 144                  */
 145                 if (cn->used == cur)
 146                         ret = cn_printf(cn, "!");
 147         }
 148
 149         for (; cur < cn->used; ++cur) {
 150                 if (cn->corename[cur] == '/')
 151                         cn->corename[cur] = '!';
 152         }
 153         return ret;
 154 }
 155
 156 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 157 {
 158         struct file *exe_file;
 159         char *pathbuf, *path, *ptr;
 160         int ret;
 161
 162         exe_file = get_mm_exe_file(current->mm);
 163         if (!exe_file)
 164                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 165
 166         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 167         if (!pathbuf) {
 168                 ret = -ENOMEM;
 169                 goto put_exe_file;
 170         }
 171
 172         path = file_path(exe_file, pathbuf, PATH_MAX);
 173         if (IS_ERR(path)) {
 174                 ret = PTR_ERR(path);
 175                 goto free_buf;
 176         }
 177
 178         if (name_only) {
 179                 ptr = strrchr(path, '/');
 180                 if (ptr)
 181                         path = ptr + 1;
 182         }
 183         ret = cn_esc_printf(cn, "%s", path);
 184
 185 free_buf:
 186         kfree(pathbuf);
 187 put_exe_file:
 188         fput(exe_file);
 189         return ret;
 190 }
 191
 192 /* format_corename will inspect the pattern parameter, and output a
 193  * name into corename, which must have space for at least
 194  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 195  */
 196 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 197                            size_t **argv, int *argc)
 198 {
 199         const struct cred *cred = current_cred();
 200         const char *pat_ptr = core_pattern;
 201         int ispipe = (*pat_ptr == '|');
 202         bool was_space = false;
 203         int pid_in_pattern = 0;
 204         int err = 0;
 205
 206         cn->used = 0;
 207         cn->corename = NULL;
 208         if (expand_corename(cn, core_name_size))
 209                 return -ENOMEM;
 210         cn->corename[0] = '\0';
 211
 212         if (ispipe) {
 213                 int argvs = sizeof(core_pattern) / 2;
 214                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 215                 if (!(*argv))
 216                         return -ENOMEM;
 217                 (*argv)[(*argc)++] = 0;
 218                 ++pat_ptr;
 219                 if (!(*pat_ptr))
 220                         return -ENOMEM;
 221         }
 222
 223         /* Repeat as long as we have more pattern to process and more output
 224            space */
 225         while (*pat_ptr) {
 226                 /*
 227                  * Split on spaces before doing template expansion so that
 228                  * %e and %E don't get split if they have spaces in them
 229                  */
 230                 if (ispipe) {
 231                         if (isspace(*pat_ptr)) {
 232                                 was_space = true;
 233                                 pat_ptr++;
 234                                 continue;
 235                         } else if (was_space) {
 236                                 was_space = false;
 237                                 err = cn_printf(cn, "%c", '\0');
 238                                 if (err)
 239                                         return err;
 240                                 (*argv)[(*argc)++] = cn->used;
 241                         }
 242                 }
 243                 if (*pat_ptr != '%') {
 244                         err = cn_printf(cn, "%c", *pat_ptr++);
 245                 } else {
 246                         switch (*++pat_ptr) {
 247                         /* single % at the end, drop that */
 248                         case 0:
 249                                 goto out;
 250                         /* Double percent, output one percent */
 251                         case '%':
 252                                 err = cn_printf(cn, "%c", '%');
 253                                 break;
 254                         /* pid */
 255                         case 'p':
 256                                 pid_in_pattern = 1;
 257                                 err = cn_printf(cn, "%d",
 258                                               task_tgid_vnr(current));
 259                                 break;
 260                         /* global pid */
 261                         case 'P':
 262                                 err = cn_printf(cn, "%d",
 263                                               task_tgid_nr(current));
 264                                 break;
 265                         case 'i':
 266                                 err = cn_printf(cn, "%d",
 267                                               task_pid_vnr(current));
 268                                 break;
 269                         case 'I':
 270                                 err = cn_printf(cn, "%d",
 271                                               task_pid_nr(current));
 272                                 break;
 273                         /* uid */
 274                         case 'u':
 275                                 err = cn_printf(cn, "%u",
 276                                                 from_kuid(&init_user_ns,
 277                                                           cred->uid));
 278                                 break;
 279                         /* gid */
 280                         case 'g':
 281                                 err = cn_printf(cn, "%u",
 282                                                 from_kgid(&init_user_ns,
 283                                                           cred->gid));
 284                                 break;
 285                         case 'd':
 286                                 err = cn_printf(cn, "%d",
 287                                         __get_dumpable(cprm->mm_flags));
 288                                 break;
 289                         /* signal that caused the coredump */
 290                         case 's':
 291                                 err = cn_printf(cn, "%d",
 292                                                 cprm->siginfo->si_signo);
 293                                 break;
 294                         /* UNIX time of coredump */
 295                         case 't': {
 296                                 time64_t time;
 297
 298                                 time = ktime_get_real_seconds();
 299                                 err = cn_printf(cn, "%lld", time);
 300                                 break;
 301                         }
 302                         /* hostname */
 303                         case 'h':
 304                                 down_read(&uts_sem);
 305                                 err = cn_esc_printf(cn, "%s",
 306                                               utsname()->nodename);
 307                                 up_read(&uts_sem);
 308                                 break;
 309                         /* executable, could be changed by prctl PR_SET_NAME etc */
 310                         case 'e':
 311                                 err = cn_esc_printf(cn, "%s", current->comm);
 312                                 break;
 313                         /* file name of executable */
 314                         case 'f':
 315                                 err = cn_print_exe_file(cn, true);
 316                                 break;
 317                         case 'E':
 318                                 err = cn_print_exe_file(cn, false);
 319                                 break;
 320                         /* core limit size */
 321                         case 'c':
 322                                 err = cn_printf(cn, "%lu",
 323                                               rlimit(RLIMIT_CORE));
 324                                 break;
 325                         default:
 326                                 break;
 327                         }
 328                         ++pat_ptr;
 329                 }
 330
 331                 if (err)
 332                         return err;
 333         }
 334
 335 out:
 336         /* Backward compatibility with core_uses_pid:
 337          *
 338          * If core_pattern does not include a %p (as is the default)
 339          * and core_uses_pid is set, then .%pid will be appended to
 340          * the filename. Do not do this for piped commands. */
 341         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 342                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 343                 if (err)
 344                         return err;
 345         }
 346         return ispipe;
 347 }
 348
 349 static int zap_process(struct task_struct *start, int exit_code, int flags)
 350 {
 351         struct task_struct *t;
 352         int nr = 0;
 353
 354         /* ignore all signals except SIGKILL, see prepare_signal() */
 355         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
 356         start->signal->group_exit_code = exit_code;
 357         start->signal->group_stop_count = 0;
 358
 359         for_each_thread(start, t) {
 360                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 361                 if (t != current && t->mm) {
 362                         sigaddset(&t->pending.signal, SIGKILL);
 363                         signal_wake_up(t, 1);
 364                         nr++;
 365                 }
 366         }
 367
 368         return nr;
 369 }
 370
 371 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 372                         struct core_state *core_state, int exit_code)
 373 {
 374         struct task_struct *g, *p;
 375         unsigned long flags;
 376         int nr = -EAGAIN;
 377
 378         spin_lock_irq(&tsk->sighand->siglock);
 379         if (!signal_group_exit(tsk->signal)) {
 380                 mm->core_state = core_state;
 381                 tsk->signal->group_exit_task = tsk;
 382                 nr = zap_process(tsk, exit_code, 0);
 383                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 384         }
 385         spin_unlock_irq(&tsk->sighand->siglock);
 386         if (unlikely(nr < 0))
 387                 return nr;
 388
 389         tsk->flags |= PF_DUMPCORE;
 390         if (atomic_read(&mm->mm_users) == nr + 1)
 391                 goto done;
 392         /*
 393          * We should find and kill all tasks which use this mm, and we should
 394          * count them correctly into ->nr_threads. We don't take tasklist
 395          * lock, but this is safe wrt:
 396          *
 397          * fork:
 398          *      None of sub-threads can fork after zap_process(leader). All
 399          *      processes which were created before this point should be
 400          *      visible to zap_threads() because copy_process() adds the new
 401          *      process to the tail of init_task.tasks list, and lock/unlock
 402          *      of ->siglock provides a memory barrier.
 403          *
 404          * do_exit:
 405          *      The caller holds mm->mmap_lock. This means that the task which
 406          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 407          *      its ->mm.
 408          *
 409          * de_thread:
 410          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 411          *      we must see either old or new leader, this does not matter.
 412          *      However, it can change p->sighand, so lock_task_sighand(p)
 413          *      must be used. Since p->mm != NULL and we hold ->mmap_lock
 414          *      it can't fail.
 415          *
 416          *      Note also that "g" can be the old leader with ->mm == NULL
 417          *      and already unhashed and thus removed from ->thread_group.
 418          *      This is OK, __unhash_process()->list_del_rcu() does not
 419          *      clear the ->next pointer, we will find the new leader via
 420          *      next_thread().
 421          */
 422         rcu_read_lock();
 423         for_each_process(g) {
 424                 if (g == tsk->group_leader)
 425                         continue;
 426                 if (g->flags & PF_KTHREAD)
 427                         continue;
 428
 429                 for_each_thread(g, p) {
 430                         if (unlikely(!p->mm))
 431                                 continue;
 432                         if (unlikely(p->mm == mm)) {
 433                                 lock_task_sighand(p, &flags);
 434                                 nr += zap_process(p, exit_code,
 435                                                         SIGNAL_GROUP_EXIT);
 436                                 unlock_task_sighand(p, &flags);
 437                         }
 438                         break;
 439                 }
 440         }
 441         rcu_read_unlock();
 442 done:
 443         atomic_set(&core_state->nr_threads, nr);
 444         return nr;
 445 }
 446
 447 static int coredump_wait(int exit_code, struct core_state *core_state)
 448 {
 449         struct task_struct *tsk = current;
 450         struct mm_struct *mm = tsk->mm;
 451         int core_waiters = -EBUSY;
 452
 453         init_completion(&core_state->startup);
 454         core_state->dumper.task = tsk;
 455         core_state->dumper.next = NULL;
 456
 457         if (mmap_write_lock_killable(mm))
 458                 return -EINTR;
 459
 460         if (!mm->core_state)
 461                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 462         mmap_write_unlock(mm);
 463
 464         if (core_waiters > 0) {
 465                 struct core_thread *ptr;
 466
 467                 freezer_do_not_count();
 468                 wait_for_completion(&core_state->startup);
 469                 freezer_count();
 470                 /*
 471                  * Wait for all the threads to become inactive, so that
 472                  * all the thread context (extended register state, like
 473                  * fpu etc) gets copied to the memory.
 474                  */
 475                 ptr = core_state->dumper.next;
 476                 while (ptr != NULL) {
 477                         wait_task_inactive(ptr->task, 0);
 478                         ptr = ptr->next;
 479                 }
 480         }
 481
 482         return core_waiters;
 483 }
 484
 485 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
 486 {
 487         struct core_thread *curr, *next;
 488         struct task_struct *task;
 489
 490         spin_lock_irq(&current->sighand->siglock);
 491         if (core_dumped && !__fatal_signal_pending(current))
 492                 current->signal->group_exit_code |= 0x80;
 493         current->signal->group_exit_task = NULL;
 494         current->signal->flags = SIGNAL_GROUP_EXIT;
 495         spin_unlock_irq(&current->sighand->siglock);
 496
 497         next = mm->core_state->dumper.next;
 498         while ((curr = next) != NULL) {
 499                 next = curr->next;
 500                 task = curr->task;
 501                 /*
 502                  * see exit_mm(), curr->task must not see
 503                  * ->task == NULL before we read ->next.
 504                  */
 505                 smp_mb();
 506                 curr->task = NULL;
 507                 wake_up_process(task);
 508         }
 509
 510         mm->core_state = NULL;
 511 }
 512
 513 static bool dump_interrupted(void)
 514 {
 515         /*
 516          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 517          * can do try_to_freeze() and check __fatal_signal_pending(),
 518          * but then we need to teach dump_write() to restart and clear
 519          * TIF_SIGPENDING.
 520          */
 521         return signal_pending(current);
 522 }
 523
 524 static void wait_for_dump_helpers(struct file *file)
 525 {
 526         struct pipe_inode_info *pipe = file->private_data;
 527
 528         pipe_lock(pipe);
 529         pipe->readers++;
 530         pipe->writers--;
 531         wake_up_interruptible_sync(&pipe->rd_wait);
 532         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 533         pipe_unlock(pipe);
 534
 535         /*
 536          * We actually want wait_event_freezable() but then we need
 537          * to clear TIF_SIGPENDING and improve dump_interrupted().
 538          */
 539         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 540
 541         pipe_lock(pipe);
 542         pipe->readers--;
 543         pipe->writers++;
 544         pipe_unlock(pipe);
 545 }
 546
 547 /*
 548  * umh_pipe_setup
 549  * helper function to customize the process used
 550  * to collect the core in userspace.  Specifically
 551  * it sets up a pipe and installs it as fd 0 (stdin)
 552  * for the process.  Returns 0 on success, or
 553  * PTR_ERR on failure.
 554  * Note that it also sets the core limit to 1.  This
 555  * is a special value that we use to trap recursive
 556  * core dumps
 557  */
 558 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 559 {
 560         struct file *files[2];
 561         struct coredump_params *cp = (struct coredump_params *)info->data;
 562         int err = create_pipe_files(files, 0);
 563         if (err)
 564                 return err;
 565
 566         cp->file = files[1];
 567
 568         err = replace_fd(0, files[0], 0);
 569         fput(files[0]);
 570         /* and disallow core files too */
 571         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 572
 573         return err;
 574 }
 575
 576 void do_coredump(const kernel_siginfo_t *siginfo)
 577 {
 578         struct core_state core_state;
 579         struct core_name cn;
 580         struct mm_struct *mm = current->mm;
 581         struct linux_binfmt * binfmt;
 582         const struct cred *old_cred;
 583         struct cred *cred;
 584         int retval = 0;
 585         int ispipe;
 586         size_t *argv = NULL;
 587         int argc = 0;
 588         struct files_struct *displaced;
 589         /* require nonrelative corefile path and be extra careful */
 590         bool need_suid_safe = false;
 591         bool core_dumped = false;
 592         static atomic_t core_dump_count = ATOMIC_INIT(0);
 593         struct coredump_params cprm = {
 594                 .siginfo = siginfo,
 595                 .regs = signal_pt_regs(),
 596                 .limit = rlimit(RLIMIT_CORE),
 597                 /*
 598                  * We must use the same mm->flags while dumping core to avoid
 599                  * inconsistency of bit flags, since this flag is not protected
 600                  * by any locks.
 601                  */
 602                 .mm_flags = mm->flags,
 603         };
 604
 605         audit_core_dumps(siginfo->si_signo);
 606
 607         binfmt = mm->binfmt;
 608         if (!binfmt || !binfmt->core_dump)
 609                 goto fail;
 610         if (!__get_dumpable(cprm.mm_flags))
 611                 goto fail;
 612
 613         cred = prepare_creds();
 614         if (!cred)
 615                 goto fail;
 616         /*
 617          * We cannot trust fsuid as being the "true" uid of the process
 618          * nor do we know its entire history. We only know it was tainted
 619          * so we dump it as root in mode 2, and only into a controlled
 620          * environment (pipe handler or fully qualified path).
 621          */
 622         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 623                 /* Setuid core dump mode */
 624                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 625                 need_suid_safe = true;
 626         }
 627
 628         retval = coredump_wait(siginfo->si_signo, &core_state);
 629         if (retval < 0)
 630                 goto fail_creds;
 631
 632         old_cred = override_creds(cred);
 633
 634         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 635
 636         if (ispipe) {
 637                 int argi;
 638                 int dump_count;
 639                 char **helper_argv;
 640                 struct subprocess_info *sub_info;
 641
 642                 if (ispipe < 0) {
 643                         printk(KERN_WARNING "format_corename failed\n");
 644                         printk(KERN_WARNING "Aborting core\n");
 645                         goto fail_unlock;
 646                 }
 647
 648                 if (cprm.limit == 1) {
 649                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 650                          *
 651                          * Normally core limits are irrelevant to pipes, since
 652                          * we're not writing to the file system, but we use
 653                          * cprm.limit of 1 here as a special value, this is a
 654                          * consistent way to catch recursive crashes.
 655                          * We can still crash if the core_pattern binary sets
 656                          * RLIM_CORE = !1, but it runs as root, and can do
 657                          * lots of stupid things.
 658                          *
 659                          * Note that we use task_tgid_vnr here to grab the pid
 660                          * of the process group leader.  That way we get the
 661                          * right pid if a thread in a multi-threaded
 662                          * core_pattern process dies.
 663                          */
 664                         printk(KERN_WARNING
 665                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 666                                 task_tgid_vnr(current), current->comm);
 667                         printk(KERN_WARNING "Aborting core\n");
 668                         goto fail_unlock;
 669                 }
 670                 cprm.limit = RLIM_INFINITY;
 671
 672                 dump_count = atomic_inc_return(&core_dump_count);
 673                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 674                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 675                                task_tgid_vnr(current), current->comm);
 676                         printk(KERN_WARNING "Skipping core dump\n");
 677                         goto fail_dropcount;
 678                 }
 679
 680                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 681                                             GFP_KERNEL);
 682                 if (!helper_argv) {
 683                         printk(KERN_WARNING "%s failed to allocate memory\n",
 684                                __func__);
 685                         goto fail_dropcount;
 686                 }
 687                 for (argi = 0; argi < argc; argi++)
 688                         helper_argv[argi] = cn.corename + argv[argi];
 689                 helper_argv[argi] = NULL;
 690
 691                 retval = -ENOMEM;
 692                 sub_info = call_usermodehelper_setup(helper_argv[0],
 693                                                 helper_argv, NULL, GFP_KERNEL,
 694                                                 umh_pipe_setup, NULL, &cprm);
 695                 if (sub_info)
 696                         retval = call_usermodehelper_exec(sub_info,
 697                                                           UMH_WAIT_EXEC);
 698
 699                 kfree(helper_argv);
 700                 if (retval) {
 701                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 702                                cn.corename);
 703                         goto close_fail;
 704                 }
 705         } else {
 706                 struct inode *inode;
 707                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 708                                  O_LARGEFILE | O_EXCL;
 709
 710                 if (cprm.limit < binfmt->min_coredump)
 711                         goto fail_unlock;
 712
 713                 if (need_suid_safe && cn.corename[0] != '/') {
 714                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 715                                 "to fully qualified path!\n",
 716                                 task_tgid_vnr(current), current->comm);
 717                         printk(KERN_WARNING "Skipping core dump\n");
 718                         goto fail_unlock;
 719                 }
 720
 721                 /*
 722                  * Unlink the file if it exists unless this is a SUID
 723                  * binary - in that case, we're running around with root
 724                  * privs and don't want to unlink another user's coredump.
 725                  */
 726                 if (!need_suid_safe) {
 727                         /*
 728                          * If it doesn't exist, that's fine. If there's some
 729                          * other problem, we'll catch it at the filp_open().
 730                          */
 731                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 732                 }
 733
 734                 /*
 735                  * There is a race between unlinking and creating the
 736                  * file, but if that causes an EEXIST here, that's
 737                  * fine - another process raced with us while creating
 738                  * the corefile, and the other process won. To userspace,
 739                  * what matters is that at least one of the two processes
 740                  * writes its coredump successfully, not which one.
 741                  */
 742                 if (need_suid_safe) {
 743                         /*
 744                          * Using user namespaces, normal user tasks can change
 745                          * their current->fs->root to point to arbitrary
 746                          * directories. Since the intention of the "only dump
 747                          * with a fully qualified path" rule is to control where
 748                          * coredumps may be placed using root privileges,
 749                          * current->fs->root must not be used. Instead, use the
 750                          * root directory of init_task.
 751                          */
 752                         struct path root;
 753
 754                         task_lock(&init_task);
 755                         get_fs_root(init_task.fs, &root);
 756                         task_unlock(&init_task);
 757                         cprm.file = file_open_root(root.dentry, root.mnt,
 758                                 cn.corename, open_flags, 0600);
 759                         path_put(&root);
 760                 } else {
 761                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 762                 }
 763                 if (IS_ERR(cprm.file))
 764                         goto fail_unlock;
 765
 766                 inode = file_inode(cprm.file);
 767                 if (inode->i_nlink > 1)
 768                         goto close_fail;
 769                 if (d_unhashed(cprm.file->f_path.dentry))
 770                         goto close_fail;
 771                 /*
 772                  * AK: actually i see no reason to not allow this for named
 773                  * pipes etc, but keep the previous behaviour for now.
 774                  */
 775                 if (!S_ISREG(inode->i_mode))
 776                         goto close_fail;
 777                 /*
 778                  * Don't dump core if the filesystem changed owner or mode
 779                  * of the file during file creation. This is an issue when
 780                  * a process dumps core while its cwd is e.g. on a vfat
 781                  * filesystem.
 782                  */
 783                 if (!uid_eq(inode->i_uid, current_fsuid()))
 784                         goto close_fail;
 785                 if ((inode->i_mode & 0677) != 0600)
 786                         goto close_fail;
 787                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 788                         goto close_fail;
 789                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
 790                         goto close_fail;
 791         }
 792
 793         /* get us an unshared descriptor table; almost always a no-op */
 794         retval = unshare_files(&displaced);
 795         if (retval)
 796                 goto close_fail;
 797         if (displaced)
 798                 put_files_struct(displaced);
 799         if (!dump_interrupted()) {
 800                 /*
 801                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 802                  * have this set to NULL.
 803                  */
 804                 if (!cprm.file) {
 805                         pr_info("Core dump to |%s disabled\n", cn.corename);
 806                         goto close_fail;
 807                 }
 808                 file_start_write(cprm.file);
 809                 core_dumped = binfmt->core_dump(&cprm);
 810                 file_end_write(cprm.file);
 811         }
 812         if (ispipe && core_pipe_limit)
 813                 wait_for_dump_helpers(cprm.file);
 814 close_fail:
 815         if (cprm.file)
 816                 filp_close(cprm.file, NULL);
 817 fail_dropcount:
 818         if (ispipe)
 819                 atomic_dec(&core_dump_count);
 820 fail_unlock:
 821         kfree(argv);
 822         kfree(cn.corename);
 823         coredump_finish(mm, core_dumped);
 824         revert_creds(old_cred);
 825 fail_creds:
 826         put_cred(cred);
 827 fail:
 828         return;
 829 }
 830
 831 /*
 832  * Core dumping helper functions.  These are the only things you should
 833  * do on a core-file: use only these functions to write out all the
 834  * necessary info.
 835  */
 836 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 837 {
 838         struct file *file = cprm->file;
 839         loff_t pos = file->f_pos;
 840         ssize_t n;
 841         if (cprm->written + nr > cprm->limit)
 842                 return 0;
 843
 844
 845         if (dump_interrupted())
 846                 return 0;
 847         n = __kernel_write(file, addr, nr, &pos);
 848         if (n != nr)
 849                 return 0;
 850         file->f_pos = pos;
 851         cprm->written += n;
 852         cprm->pos += n;
 853
 854         return 1;
 855 }
 856 EXPORT_SYMBOL(dump_emit);
 857
 858 int dump_skip(struct coredump_params *cprm, size_t nr)
 859 {
 860         static char zeroes[PAGE_SIZE];
 861         struct file *file = cprm->file;
 862         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 863                 if (dump_interrupted() ||
 864                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
 865                         return 0;
 866                 cprm->pos += nr;
 867                 return 1;
 868         } else {
 869                 while (nr > PAGE_SIZE) {
 870                         if (!dump_emit(cprm, zeroes, PAGE_SIZE))
 871                                 return 0;
 872                         nr -= PAGE_SIZE;
 873                 }
 874                 return dump_emit(cprm, zeroes, nr);
 875         }
 876 }
 877 EXPORT_SYMBOL(dump_skip);
 878
 879 #ifdef CONFIG_ELF_CORE
 880 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 881                     unsigned long len)
 882 {
 883         unsigned long addr;
 884
 885         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 886                 struct page *page;
 887                 int stop;
 888
 889                 /*
 890                  * To avoid having to allocate page tables for virtual address
 891                  * ranges that have never been used yet, and also to make it
 892                  * easy to generate sparse core files, use a helper that returns
 893                  * NULL when encountering an empty page table entry that would
 894                  * otherwise have been filled with the zero page.
 895                  */
 896                 page = get_dump_page(addr);
 897                 if (page) {
 898                         void *kaddr = kmap(page);
 899
 900                         stop = !dump_emit(cprm, kaddr, PAGE_SIZE);
 901                         kunmap(page);
 902                         put_page(page);
 903                 } else {
 904                         stop = !dump_skip(cprm, PAGE_SIZE);
 905                 }
 906                 if (stop)
 907                         return 0;
 908         }
 909         return 1;
 910 }
 911 #endif
 912
 913 int dump_align(struct coredump_params *cprm, int align)
 914 {
 915         unsigned mod = cprm->pos & (align - 1);
 916         if (align & (align - 1))
 917                 return 0;
 918         return mod ? dump_skip(cprm, align - mod) : 1;
 919 }
 920 EXPORT_SYMBOL(dump_align);
 921
 922 /*
 923  * Ensures that file size is big enough to contain the current file
 924  * postion. This prevents gdb from complaining about a truncated file
 925  * if the last "write" to the file was dump_skip.
 926  */
 927 void dump_truncate(struct coredump_params *cprm)
 928 {
 929         struct file *file = cprm->file;
 930         loff_t offset;
 931
 932         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 933                 offset = file->f_op->llseek(file, 0, SEEK_CUR);
 934                 if (i_size_read(file->f_mapping->host) < offset)
 935                         do_truncate(file->f_path.dentry, offset, 0, file);
 936         }
 937 }
 938 EXPORT_SYMBOL(dump_truncate);
 939
 940 /*
 941  * The purpose of always_dump_vma() is to make sure that special kernel mappings
 942  * that are useful for post-mortem analysis are included in every core dump.
 943  * In that way we ensure that the core dump is fully interpretable later
 944  * without matching up the same kernel and hardware config to see what PC values
 945  * meant. These special mappings include - vDSO, vsyscall, and other
 946  * architecture specific mappings
 947  */
 948 static bool always_dump_vma(struct vm_area_struct *vma)
 949 {
 950         /* Any vsyscall mappings? */
 951         if (vma == get_gate_vma(vma->vm_mm))
 952                 return true;
 953
 954         /*
 955          * Assume that all vmas with a .name op should always be dumped.
 956          * If this changes, a new vm_ops field can easily be added.
 957          */
 958         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
 959                 return true;
 960
 961         /*
 962          * arch_vma_name() returns non-NULL for special architecture mappings,
 963          * such as vDSO sections.
 964          */
 965         if (arch_vma_name(vma))
 966                 return true;
 967
 968         return false;
 969 }
 970
 971 /*
 972  * Decide how much of @vma's contents should be included in a core dump.
 973  */
 974 static unsigned long vma_dump_size(struct vm_area_struct *vma,
 975                                    unsigned long mm_flags)
 976 {
 977 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
 978
 979         /* always dump the vdso and vsyscall sections */
 980         if (always_dump_vma(vma))
 981                 goto whole;
 982
 983         if (vma->vm_flags & VM_DONTDUMP)
 984                 return 0;
 985
 986         /* support for DAX */
 987         if (vma_is_dax(vma)) {
 988                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
 989                         goto whole;
 990                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
 991                         goto whole;
 992                 return 0;
 993         }
 994
 995         /* Hugetlb memory check */
 996         if (is_vm_hugetlb_page(vma)) {
 997                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
 998                         goto whole;
 999                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1000                         goto whole;
1001                 return 0;
1002         }
1003
1004         /* Do not dump I/O mapped devices or special mappings */
1005         if (vma->vm_flags & VM_IO)
1006                 return 0;
1007
1008         /* By default, dump shared memory if mapped from an anonymous file. */
1009         if (vma->vm_flags & VM_SHARED) {
1010                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1011                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1012                         goto whole;
1013                 return 0;
1014         }
1015
1016         /* Dump segments that have been written to.  */
1017         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1018                 goto whole;
1019         if (vma->vm_file == NULL)
1020                 return 0;
1021
1022         if (FILTER(MAPPED_PRIVATE))
1023                 goto whole;
1024
1025         /*
1026          * If this is the beginning of an executable file mapping,
1027          * dump the first page to aid in determining what was mapped here.
1028          */
1029         if (FILTER(ELF_HEADERS) &&
1030             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ) &&
1031             (READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1032                 return PAGE_SIZE;
1033
1034 #undef  FILTER
1035
1036         return 0;
1037
1038 whole:
1039         return vma->vm_end - vma->vm_start;
1040 }
1041
1042 static struct vm_area_struct *first_vma(struct task_struct *tsk,
1043                                         struct vm_area_struct *gate_vma)
1044 {
1045         struct vm_area_struct *ret = tsk->mm->mmap;
1046
1047         if (ret)
1048                 return ret;
1049         return gate_vma;
1050 }
1051
1052 /*
1053  * Helper function for iterating across a vma list.  It ensures that the caller
1054  * will visit `gate_vma' prior to terminating the search.
1055  */
1056 static struct vm_area_struct *next_vma(struct vm_area_struct *this_vma,
1057                                        struct vm_area_struct *gate_vma)
1058 {
1059         struct vm_area_struct *ret;
1060
1061         ret = this_vma->vm_next;
1062         if (ret)
1063                 return ret;
1064         if (this_vma == gate_vma)
1065                 return NULL;
1066         return gate_vma;
1067 }
1068
1069 /*
1070  * Under the mmap_lock, take a snapshot of relevant information about the task's
1071  * VMAs.
1072  */
1073 int dump_vma_snapshot(struct coredump_params *cprm, int *vma_count,
1074                       struct core_vma_metadata **vma_meta,
1075                       size_t *vma_data_size_ptr)
1076 {
1077         struct vm_area_struct *vma, *gate_vma;
1078         struct mm_struct *mm = current->mm;
1079         int i;
1080         size_t vma_data_size = 0;
1081
1082         /*
1083          * Once the stack expansion code is fixed to not change VMA bounds
1084          * under mmap_lock in read mode, this can be changed to take the
1085          * mmap_lock in read mode.
1086          */
1087         if (mmap_write_lock_killable(mm))
1088                 return -EINTR;
1089
1090         gate_vma = get_gate_vma(mm);
1091         *vma_count = mm->map_count + (gate_vma ? 1 : 0);
1092
1093         *vma_meta = kvmalloc_array(*vma_count, sizeof(**vma_meta), GFP_KERNEL);
1094         if (!*vma_meta) {
1095                 mmap_write_unlock(mm);
1096                 return -ENOMEM;
1097         }
1098
1099         for (i = 0, vma = first_vma(current, gate_vma); vma != NULL;
1100                         vma = next_vma(vma, gate_vma), i++) {
1101                 struct core_vma_metadata *m = (*vma_meta) + i;
1102
1103                 m->start = vma->vm_start;
1104                 m->end = vma->vm_end;
1105                 m->flags = vma->vm_flags;
1106                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1107
1108                 vma_data_size += m->dump_size;
1109         }
1110
1111         mmap_write_unlock(mm);
1112
1113         if (WARN_ON(i != *vma_count))
1114                 return -EFAULT;
1115
1116         *vma_data_size_ptr = vma_data_size;
1117         return 0;
1118 }