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                                 if (cn->used != 0)
 233                                         was_space = true;
 234                                 pat_ptr++;
 235                                 continue;
 236                         } else if (was_space) {
 237                                 was_space = false;
 238                                 err = cn_printf(cn, "%c", '\0');
 239                                 if (err)
 240                                         return err;
 241                                 (*argv)[(*argc)++] = cn->used;
 242                         }
 243                 }
 244                 if (*pat_ptr != '%') {
 245                         err = cn_printf(cn, "%c", *pat_ptr++);
 246                 } else {
 247                         switch (*++pat_ptr) {
 248                         /* single % at the end, drop that */
 249                         case 0:
 250                                 goto out;
 251                         /* Double percent, output one percent */
 252                         case '%':
 253                                 err = cn_printf(cn, "%c", '%');
 254                                 break;
 255                         /* pid */
 256                         case 'p':
 257                                 pid_in_pattern = 1;
 258                                 err = cn_printf(cn, "%d",
 259                                               task_tgid_vnr(current));
 260                                 break;
 261                         /* global pid */
 262                         case 'P':
 263                                 err = cn_printf(cn, "%d",
 264                                               task_tgid_nr(current));
 265                                 break;
 266                         case 'i':
 267                                 err = cn_printf(cn, "%d",
 268                                               task_pid_vnr(current));
 269                                 break;
 270                         case 'I':
 271                                 err = cn_printf(cn, "%d",
 272                                               task_pid_nr(current));
 273                                 break;
 274                         /* uid */
 275                         case 'u':
 276                                 err = cn_printf(cn, "%u",
 277                                                 from_kuid(&init_user_ns,
 278                                                           cred->uid));
 279                                 break;
 280                         /* gid */
 281                         case 'g':
 282                                 err = cn_printf(cn, "%u",
 283                                                 from_kgid(&init_user_ns,
 284                                                           cred->gid));
 285                                 break;
 286                         case 'd':
 287                                 err = cn_printf(cn, "%d",
 288                                         __get_dumpable(cprm->mm_flags));
 289                                 break;
 290                         /* signal that caused the coredump */
 291                         case 's':
 292                                 err = cn_printf(cn, "%d",
 293                                                 cprm->siginfo->si_signo);
 294                                 break;
 295                         /* UNIX time of coredump */
 296                         case 't': {
 297                                 time64_t time;
 298
 299                                 time = ktime_get_real_seconds();
 300                                 err = cn_printf(cn, "%lld", time);
 301                                 break;
 302                         }
 303                         /* hostname */
 304                         case 'h':
 305                                 down_read(&uts_sem);
 306                                 err = cn_esc_printf(cn, "%s",
 307                                               utsname()->nodename);
 308                                 up_read(&uts_sem);
 309                                 break;
 310                         /* executable, could be changed by prctl PR_SET_NAME etc */
 311                         case 'e':
 312                                 err = cn_esc_printf(cn, "%s", current->comm);
 313                                 break;
 314                         /* file name of executable */
 315                         case 'f':
 316                                 err = cn_print_exe_file(cn, true);
 317                                 break;
 318                         case 'E':
 319                                 err = cn_print_exe_file(cn, false);
 320                                 break;
 321                         /* core limit size */
 322                         case 'c':
 323                                 err = cn_printf(cn, "%lu",
 324                                               rlimit(RLIMIT_CORE));
 325                                 break;
 326                         default:
 327                                 break;
 328                         }
 329                         ++pat_ptr;
 330                 }
 331
 332                 if (err)
 333                         return err;
 334         }
 335
 336 out:
 337         /* Backward compatibility with core_uses_pid:
 338          *
 339          * If core_pattern does not include a %p (as is the default)
 340          * and core_uses_pid is set, then .%pid will be appended to
 341          * the filename. Do not do this for piped commands. */
 342         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 343                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 344                 if (err)
 345                         return err;
 346         }
 347         return ispipe;
 348 }
 349
 350 static int zap_process(struct task_struct *start, int exit_code, int flags)
 351 {
 352         struct task_struct *t;
 353         int nr = 0;
 354
 355         /* ignore all signals except SIGKILL, see prepare_signal() */
 356         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
 357         start->signal->group_exit_code = exit_code;
 358         start->signal->group_stop_count = 0;
 359
 360         for_each_thread(start, t) {
 361                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 362                 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
 363                         sigaddset(&t->pending.signal, SIGKILL);
 364                         signal_wake_up(t, 1);
 365                         nr++;
 366                 }
 367         }
 368
 369         return nr;
 370 }
 371
 372 static int zap_threads(struct task_struct *tsk,
 373                         struct core_state *core_state, int exit_code)
 374 {
 375         int nr = -EAGAIN;
 376
 377         spin_lock_irq(&tsk->sighand->siglock);
 378         if (!signal_group_exit(tsk->signal)) {
 379                 tsk->signal->core_state = core_state;
 380                 tsk->signal->group_exit_task = tsk;
 381                 nr = zap_process(tsk, exit_code, 0);
 382                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 383                 tsk->flags |= PF_DUMPCORE;
 384                 atomic_set(&core_state->nr_threads, nr);
 385         }
 386         spin_unlock_irq(&tsk->sighand->siglock);
 387         return nr;
 388 }
 389
 390 static int coredump_wait(int exit_code, struct core_state *core_state)
 391 {
 392         struct task_struct *tsk = current;
 393         int core_waiters = -EBUSY;
 394
 395         init_completion(&core_state->startup);
 396         core_state->dumper.task = tsk;
 397         core_state->dumper.next = NULL;
 398
 399         core_waiters = zap_threads(tsk, core_state, exit_code);
 400         if (core_waiters > 0) {
 401                 struct core_thread *ptr;
 402
 403                 freezer_do_not_count();
 404                 wait_for_completion(&core_state->startup);
 405                 freezer_count();
 406                 /*
 407                  * Wait for all the threads to become inactive, so that
 408                  * all the thread context (extended register state, like
 409                  * fpu etc) gets copied to the memory.
 410                  */
 411                 ptr = core_state->dumper.next;
 412                 while (ptr != NULL) {
 413                         wait_task_inactive(ptr->task, 0);
 414                         ptr = ptr->next;
 415                 }
 416         }
 417
 418         return core_waiters;
 419 }
 420
 421 static void coredump_finish(bool core_dumped)
 422 {
 423         struct core_thread *curr, *next;
 424         struct task_struct *task;
 425
 426         spin_lock_irq(&current->sighand->siglock);
 427         if (core_dumped && !__fatal_signal_pending(current))
 428                 current->signal->group_exit_code |= 0x80;
 429         current->signal->group_exit_task = NULL;
 430         current->signal->flags = SIGNAL_GROUP_EXIT;
 431         next = current->signal->core_state->dumper.next;
 432         current->signal->core_state = NULL;
 433         spin_unlock_irq(&current->sighand->siglock);
 434
 435         while ((curr = next) != NULL) {
 436                 next = curr->next;
 437                 task = curr->task;
 438                 /*
 439                  * see coredump_task_exit(), curr->task must not see
 440                  * ->task == NULL before we read ->next.
 441                  */
 442                 smp_mb();
 443                 curr->task = NULL;
 444                 wake_up_process(task);
 445         }
 446 }
 447
 448 static bool dump_interrupted(void)
 449 {
 450         /*
 451          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 452          * can do try_to_freeze() and check __fatal_signal_pending(),
 453          * but then we need to teach dump_write() to restart and clear
 454          * TIF_SIGPENDING.
 455          */
 456         return fatal_signal_pending(current) || freezing(current);
 457 }
 458
 459 static void wait_for_dump_helpers(struct file *file)
 460 {
 461         struct pipe_inode_info *pipe = file->private_data;
 462
 463         pipe_lock(pipe);
 464         pipe->readers++;
 465         pipe->writers--;
 466         wake_up_interruptible_sync(&pipe->rd_wait);
 467         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 468         pipe_unlock(pipe);
 469
 470         /*
 471          * We actually want wait_event_freezable() but then we need
 472          * to clear TIF_SIGPENDING and improve dump_interrupted().
 473          */
 474         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 475
 476         pipe_lock(pipe);
 477         pipe->readers--;
 478         pipe->writers++;
 479         pipe_unlock(pipe);
 480 }
 481
 482 /*
 483  * umh_pipe_setup
 484  * helper function to customize the process used
 485  * to collect the core in userspace.  Specifically
 486  * it sets up a pipe and installs it as fd 0 (stdin)
 487  * for the process.  Returns 0 on success, or
 488  * PTR_ERR on failure.
 489  * Note that it also sets the core limit to 1.  This
 490  * is a special value that we use to trap recursive
 491  * core dumps
 492  */
 493 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 494 {
 495         struct file *files[2];
 496         struct coredump_params *cp = (struct coredump_params *)info->data;
 497         int err = create_pipe_files(files, 0);
 498         if (err)
 499                 return err;
 500
 501         cp->file = files[1];
 502
 503         err = replace_fd(0, files[0], 0);
 504         fput(files[0]);
 505         /* and disallow core files too */
 506         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 507
 508         return err;
 509 }
 510
 511 void do_coredump(const kernel_siginfo_t *siginfo)
 512 {
 513         struct core_state core_state;
 514         struct core_name cn;
 515         struct mm_struct *mm = current->mm;
 516         struct linux_binfmt * binfmt;
 517         const struct cred *old_cred;
 518         struct cred *cred;
 519         int retval = 0;
 520         int ispipe;
 521         size_t *argv = NULL;
 522         int argc = 0;
 523         /* require nonrelative corefile path and be extra careful */
 524         bool need_suid_safe = false;
 525         bool core_dumped = false;
 526         static atomic_t core_dump_count = ATOMIC_INIT(0);
 527         struct coredump_params cprm = {
 528                 .siginfo = siginfo,
 529                 .regs = signal_pt_regs(),
 530                 .limit = rlimit(RLIMIT_CORE),
 531                 /*
 532                  * We must use the same mm->flags while dumping core to avoid
 533                  * inconsistency of bit flags, since this flag is not protected
 534                  * by any locks.
 535                  */
 536                 .mm_flags = mm->flags,
 537         };
 538
 539         audit_core_dumps(siginfo->si_signo);
 540
 541         binfmt = mm->binfmt;
 542         if (!binfmt || !binfmt->core_dump)
 543                 goto fail;
 544         if (!__get_dumpable(cprm.mm_flags))
 545                 goto fail;
 546
 547         cred = prepare_creds();
 548         if (!cred)
 549                 goto fail;
 550         /*
 551          * We cannot trust fsuid as being the "true" uid of the process
 552          * nor do we know its entire history. We only know it was tainted
 553          * so we dump it as root in mode 2, and only into a controlled
 554          * environment (pipe handler or fully qualified path).
 555          */
 556         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 557                 /* Setuid core dump mode */
 558                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 559                 need_suid_safe = true;
 560         }
 561
 562         retval = coredump_wait(siginfo->si_signo, &core_state);
 563         if (retval < 0)
 564                 goto fail_creds;
 565
 566         old_cred = override_creds(cred);
 567
 568         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 569
 570         if (ispipe) {
 571                 int argi;
 572                 int dump_count;
 573                 char **helper_argv;
 574                 struct subprocess_info *sub_info;
 575
 576                 if (ispipe < 0) {
 577                         printk(KERN_WARNING "format_corename failed\n");
 578                         printk(KERN_WARNING "Aborting core\n");
 579                         goto fail_unlock;
 580                 }
 581
 582                 if (cprm.limit == 1) {
 583                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 584                          *
 585                          * Normally core limits are irrelevant to pipes, since
 586                          * we're not writing to the file system, but we use
 587                          * cprm.limit of 1 here as a special value, this is a
 588                          * consistent way to catch recursive crashes.
 589                          * We can still crash if the core_pattern binary sets
 590                          * RLIM_CORE = !1, but it runs as root, and can do
 591                          * lots of stupid things.
 592                          *
 593                          * Note that we use task_tgid_vnr here to grab the pid
 594                          * of the process group leader.  That way we get the
 595                          * right pid if a thread in a multi-threaded
 596                          * core_pattern process dies.
 597                          */
 598                         printk(KERN_WARNING
 599                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 600                                 task_tgid_vnr(current), current->comm);
 601                         printk(KERN_WARNING "Aborting core\n");
 602                         goto fail_unlock;
 603                 }
 604                 cprm.limit = RLIM_INFINITY;
 605
 606                 dump_count = atomic_inc_return(&core_dump_count);
 607                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 608                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 609                                task_tgid_vnr(current), current->comm);
 610                         printk(KERN_WARNING "Skipping core dump\n");
 611                         goto fail_dropcount;
 612                 }
 613
 614                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 615                                             GFP_KERNEL);
 616                 if (!helper_argv) {
 617                         printk(KERN_WARNING "%s failed to allocate memory\n",
 618                                __func__);
 619                         goto fail_dropcount;
 620                 }
 621                 for (argi = 0; argi < argc; argi++)
 622                         helper_argv[argi] = cn.corename + argv[argi];
 623                 helper_argv[argi] = NULL;
 624
 625                 retval = -ENOMEM;
 626                 sub_info = call_usermodehelper_setup(helper_argv[0],
 627                                                 helper_argv, NULL, GFP_KERNEL,
 628                                                 umh_pipe_setup, NULL, &cprm);
 629                 if (sub_info)
 630                         retval = call_usermodehelper_exec(sub_info,
 631                                                           UMH_WAIT_EXEC);
 632
 633                 kfree(helper_argv);
 634                 if (retval) {
 635                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 636                                cn.corename);
 637                         goto close_fail;
 638                 }
 639         } else {
 640                 struct user_namespace *mnt_userns;
 641                 struct inode *inode;
 642                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 643                                  O_LARGEFILE | O_EXCL;
 644
 645                 if (cprm.limit < binfmt->min_coredump)
 646                         goto fail_unlock;
 647
 648                 if (need_suid_safe && cn.corename[0] != '/') {
 649                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 650                                 "to fully qualified path!\n",
 651                                 task_tgid_vnr(current), current->comm);
 652                         printk(KERN_WARNING "Skipping core dump\n");
 653                         goto fail_unlock;
 654                 }
 655
 656                 /*
 657                  * Unlink the file if it exists unless this is a SUID
 658                  * binary - in that case, we're running around with root
 659                  * privs and don't want to unlink another user's coredump.
 660                  */
 661                 if (!need_suid_safe) {
 662                         /*
 663                          * If it doesn't exist, that's fine. If there's some
 664                          * other problem, we'll catch it at the filp_open().
 665                          */
 666                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 667                 }
 668
 669                 /*
 670                  * There is a race between unlinking and creating the
 671                  * file, but if that causes an EEXIST here, that's
 672                  * fine - another process raced with us while creating
 673                  * the corefile, and the other process won. To userspace,
 674                  * what matters is that at least one of the two processes
 675                  * writes its coredump successfully, not which one.
 676                  */
 677                 if (need_suid_safe) {
 678                         /*
 679                          * Using user namespaces, normal user tasks can change
 680                          * their current->fs->root to point to arbitrary
 681                          * directories. Since the intention of the "only dump
 682                          * with a fully qualified path" rule is to control where
 683                          * coredumps may be placed using root privileges,
 684                          * current->fs->root must not be used. Instead, use the
 685                          * root directory of init_task.
 686                          */
 687                         struct path root;
 688
 689                         task_lock(&init_task);
 690                         get_fs_root(init_task.fs, &root);
 691                         task_unlock(&init_task);
 692                         cprm.file = file_open_root(&root, cn.corename,
 693                                                    open_flags, 0600);
 694                         path_put(&root);
 695                 } else {
 696                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 697                 }
 698                 if (IS_ERR(cprm.file))
 699                         goto fail_unlock;
 700
 701                 inode = file_inode(cprm.file);
 702                 if (inode->i_nlink > 1)
 703                         goto close_fail;
 704                 if (d_unhashed(cprm.file->f_path.dentry))
 705                         goto close_fail;
 706                 /*
 707                  * AK: actually i see no reason to not allow this for named
 708                  * pipes etc, but keep the previous behaviour for now.
 709                  */
 710                 if (!S_ISREG(inode->i_mode))
 711                         goto close_fail;
 712                 /*
 713                  * Don't dump core if the filesystem changed owner or mode
 714                  * of the file during file creation. This is an issue when
 715                  * a process dumps core while its cwd is e.g. on a vfat
 716                  * filesystem.
 717                  */
 718                 mnt_userns = file_mnt_user_ns(cprm.file);
 719                 if (!uid_eq(i_uid_into_mnt(mnt_userns, inode),
 720                             current_fsuid())) {
 721                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file owner\n",
 722                                             cn.corename);
 723                         goto close_fail;
 724                 }
 725                 if ((inode->i_mode & 0677) != 0600) {
 726                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file permissions\n",
 727                                             cn.corename);
 728                         goto close_fail;
 729                 }
 730                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 731                         goto close_fail;
 732                 if (do_truncate(mnt_userns, cprm.file->f_path.dentry,
 733                                 0, 0, cprm.file))
 734                         goto close_fail;
 735         }
 736
 737         /* get us an unshared descriptor table; almost always a no-op */
 738         /* The cell spufs coredump code reads the file descriptor tables */
 739         retval = unshare_files();
 740         if (retval)
 741                 goto close_fail;
 742         if (!dump_interrupted()) {
 743                 /*
 744                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 745                  * have this set to NULL.
 746                  */
 747                 if (!cprm.file) {
 748                         pr_info("Core dump to |%s disabled\n", cn.corename);
 749                         goto close_fail;
 750                 }
 751                 file_start_write(cprm.file);
 752                 core_dumped = binfmt->core_dump(&cprm);
 753                 /*
 754                  * Ensures that file size is big enough to contain the current
 755                  * file postion. This prevents gdb from complaining about
 756                  * a truncated file if the last "write" to the file was
 757                  * dump_skip.
 758                  */
 759                 if (cprm.to_skip) {
 760                         cprm.to_skip--;
 761                         dump_emit(&cprm, "", 1);
 762                 }
 763                 file_end_write(cprm.file);
 764         }
 765         if (ispipe && core_pipe_limit)
 766                 wait_for_dump_helpers(cprm.file);
 767 close_fail:
 768         if (cprm.file)
 769                 filp_close(cprm.file, NULL);
 770 fail_dropcount:
 771         if (ispipe)
 772                 atomic_dec(&core_dump_count);
 773 fail_unlock:
 774         kfree(argv);
 775         kfree(cn.corename);
 776         coredump_finish(core_dumped);
 777         revert_creds(old_cred);
 778 fail_creds:
 779         put_cred(cred);
 780 fail:
 781         return;
 782 }
 783
 784 /*
 785  * Core dumping helper functions.  These are the only things you should
 786  * do on a core-file: use only these functions to write out all the
 787  * necessary info.
 788  */
 789 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 790 {
 791         struct file *file = cprm->file;
 792         loff_t pos = file->f_pos;
 793         ssize_t n;
 794         if (cprm->written + nr > cprm->limit)
 795                 return 0;
 796
 797
 798         if (dump_interrupted())
 799                 return 0;
 800         n = __kernel_write(file, addr, nr, &pos);
 801         if (n != nr)
 802                 return 0;
 803         file->f_pos = pos;
 804         cprm->written += n;
 805         cprm->pos += n;
 806
 807         return 1;
 808 }
 809
 810 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 811 {
 812         static char zeroes[PAGE_SIZE];
 813         struct file *file = cprm->file;
 814         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 815                 if (dump_interrupted() ||
 816                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
 817                         return 0;
 818                 cprm->pos += nr;
 819                 return 1;
 820         } else {
 821                 while (nr > PAGE_SIZE) {
 822                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 823                                 return 0;
 824                         nr -= PAGE_SIZE;
 825                 }
 826                 return __dump_emit(cprm, zeroes, nr);
 827         }
 828 }
 829
 830 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 831 {
 832         if (cprm->to_skip) {
 833                 if (!__dump_skip(cprm, cprm->to_skip))
 834                         return 0;
 835                 cprm->to_skip = 0;
 836         }
 837         return __dump_emit(cprm, addr, nr);
 838 }
 839 EXPORT_SYMBOL(dump_emit);
 840
 841 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 842 {
 843         cprm->to_skip = pos - cprm->pos;
 844 }
 845 EXPORT_SYMBOL(dump_skip_to);
 846
 847 void dump_skip(struct coredump_params *cprm, size_t nr)
 848 {
 849         cprm->to_skip += nr;
 850 }
 851 EXPORT_SYMBOL(dump_skip);
 852
 853 #ifdef CONFIG_ELF_CORE
 854 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 855                     unsigned long len)
 856 {
 857         unsigned long addr;
 858
 859         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 860                 struct page *page;
 861                 int stop;
 862
 863                 /*
 864                  * To avoid having to allocate page tables for virtual address
 865                  * ranges that have never been used yet, and also to make it
 866                  * easy to generate sparse core files, use a helper that returns
 867                  * NULL when encountering an empty page table entry that would
 868                  * otherwise have been filled with the zero page.
 869                  */
 870                 page = get_dump_page(addr);
 871                 if (page) {
 872                         void *kaddr = kmap_local_page(page);
 873
 874                         stop = !dump_emit(cprm, kaddr, PAGE_SIZE);
 875                         kunmap_local(kaddr);
 876                         put_page(page);
 877                         if (stop)
 878                                 return 0;
 879                 } else {
 880                         dump_skip(cprm, PAGE_SIZE);
 881                 }
 882         }
 883         return 1;
 884 }
 885 #endif
 886
 887 int dump_align(struct coredump_params *cprm, int align)
 888 {
 889         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
 890         if (align & (align - 1))
 891                 return 0;
 892         if (mod)
 893                 cprm->to_skip += align - mod;
 894         return 1;
 895 }
 896 EXPORT_SYMBOL(dump_align);
 897
 898 /*
 899  * The purpose of always_dump_vma() is to make sure that special kernel mappings
 900  * that are useful for post-mortem analysis are included in every core dump.
 901  * In that way we ensure that the core dump is fully interpretable later
 902  * without matching up the same kernel and hardware config to see what PC values
 903  * meant. These special mappings include - vDSO, vsyscall, and other
 904  * architecture specific mappings
 905  */
 906 static bool always_dump_vma(struct vm_area_struct *vma)
 907 {
 908         /* Any vsyscall mappings? */
 909         if (vma == get_gate_vma(vma->vm_mm))
 910                 return true;
 911
 912         /*
 913          * Assume that all vmas with a .name op should always be dumped.
 914          * If this changes, a new vm_ops field can easily be added.
 915          */
 916         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
 917                 return true;
 918
 919         /*
 920          * arch_vma_name() returns non-NULL for special architecture mappings,
 921          * such as vDSO sections.
 922          */
 923         if (arch_vma_name(vma))
 924                 return true;
 925
 926         return false;
 927 }
 928
 929 /*
 930  * Decide how much of @vma's contents should be included in a core dump.
 931  */
 932 static unsigned long vma_dump_size(struct vm_area_struct *vma,
 933                                    unsigned long mm_flags)
 934 {
 935 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
 936
 937         /* always dump the vdso and vsyscall sections */
 938         if (always_dump_vma(vma))
 939                 goto whole;
 940
 941         if (vma->vm_flags & VM_DONTDUMP)
 942                 return 0;
 943
 944         /* support for DAX */
 945         if (vma_is_dax(vma)) {
 946                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
 947                         goto whole;
 948                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
 949                         goto whole;
 950                 return 0;
 951         }
 952
 953         /* Hugetlb memory check */
 954         if (is_vm_hugetlb_page(vma)) {
 955                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
 956                         goto whole;
 957                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
 958                         goto whole;
 959                 return 0;
 960         }
 961
 962         /* Do not dump I/O mapped devices or special mappings */
 963         if (vma->vm_flags & VM_IO)
 964                 return 0;
 965
 966         /* By default, dump shared memory if mapped from an anonymous file. */
 967         if (vma->vm_flags & VM_SHARED) {
 968                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
 969                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
 970                         goto whole;
 971                 return 0;
 972         }
 973
 974         /* Dump segments that have been written to.  */
 975         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
 976                 goto whole;
 977         if (vma->vm_file == NULL)
 978                 return 0;
 979
 980         if (FILTER(MAPPED_PRIVATE))
 981                 goto whole;
 982
 983         /*
 984          * If this is the beginning of an executable file mapping,
 985          * dump the first page to aid in determining what was mapped here.
 986          */
 987         if (FILTER(ELF_HEADERS) &&
 988             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ) &&
 989             (READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
 990                 return PAGE_SIZE;
 991
 992 #undef  FILTER
 993
 994         return 0;
 995
 996 whole:
 997         return vma->vm_end - vma->vm_start;
 998 }
 999
1000 static struct vm_area_struct *first_vma(struct task_struct *tsk,
1001                                         struct vm_area_struct *gate_vma)
1002 {
1003         struct vm_area_struct *ret = tsk->mm->mmap;
1004
1005         if (ret)
1006                 return ret;
1007         return gate_vma;
1008 }
1009
1010 /*
1011  * Helper function for iterating across a vma list.  It ensures that the caller
1012  * will visit `gate_vma' prior to terminating the search.
1013  */
1014 static struct vm_area_struct *next_vma(struct vm_area_struct *this_vma,
1015                                        struct vm_area_struct *gate_vma)
1016 {
1017         struct vm_area_struct *ret;
1018
1019         ret = this_vma->vm_next;
1020         if (ret)
1021                 return ret;
1022         if (this_vma == gate_vma)
1023                 return NULL;
1024         return gate_vma;
1025 }
1026
1027 /*
1028  * Under the mmap_lock, take a snapshot of relevant information about the task's
1029  * VMAs.
1030  */
1031 int dump_vma_snapshot(struct coredump_params *cprm, int *vma_count,
1032                       struct core_vma_metadata **vma_meta,
1033                       size_t *vma_data_size_ptr)
1034 {
1035         struct vm_area_struct *vma, *gate_vma;
1036         struct mm_struct *mm = current->mm;
1037         int i;
1038         size_t vma_data_size = 0;
1039
1040         /*
1041          * Once the stack expansion code is fixed to not change VMA bounds
1042          * under mmap_lock in read mode, this can be changed to take the
1043          * mmap_lock in read mode.
1044          */
1045         if (mmap_write_lock_killable(mm))
1046                 return -EINTR;
1047
1048         gate_vma = get_gate_vma(mm);
1049         *vma_count = mm->map_count + (gate_vma ? 1 : 0);
1050
1051         *vma_meta = kvmalloc_array(*vma_count, sizeof(**vma_meta), GFP_KERNEL);
1052         if (!*vma_meta) {
1053                 mmap_write_unlock(mm);
1054                 return -ENOMEM;
1055         }
1056
1057         for (i = 0, vma = first_vma(current, gate_vma); vma != NULL;
1058                         vma = next_vma(vma, gate_vma), i++) {
1059                 struct core_vma_metadata *m = (*vma_meta) + i;
1060
1061                 m->start = vma->vm_start;
1062                 m->end = vma->vm_end;
1063                 m->flags = vma->vm_flags;
1064                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1065
1066                 vma_data_size += m->dump_size;
1067         }
1068
1069         mmap_write_unlock(mm);
1070
1071         if (WARN_ON(i != *vma_count)) {
1072                 kvfree(*vma_meta);
1073                 return -EFAULT;
1074         }
1075
1076         *vma_data_size_ptr = vma_data_size;
1077         return 0;
1078 }