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/kmod.h>
  35 #include <linux/fsnotify.h>
  36 #include <linux/fs_struct.h>
  37 #include <linux/pipe_fs_i.h>
  38 #include <linux/oom.h>
  39 #include <linux/compat.h>
  40 #include <linux/fs.h>
  41 #include <linux/path.h>
  42 #include <linux/timekeeping.h>
  43 #include <linux/sysctl.h>
  44 #include <linux/elf.h>
  45
  46 #include <linux/uaccess.h>
  47 #include <asm/mmu_context.h>
  48 #include <asm/tlb.h>
  49 #include <asm/exec.h>
  50
  51 #include <trace/events/task.h>
  52 #include "internal.h"
  53
  54 #include <trace/events/sched.h>
  55
  56 static bool dump_vma_snapshot(struct coredump_params *cprm);
  57 static void free_vma_snapshot(struct coredump_params *cprm);
  58
  59 static int core_uses_pid;
  60 static unsigned int core_pipe_limit;
  61 static char core_pattern[CORENAME_MAX_SIZE] = "core";
  62 static int core_name_size = CORENAME_MAX_SIZE;
  63
  64 struct core_name {
  65         char *corename;
  66         int used, size;
  67 };
  68
  69 static int expand_corename(struct core_name *cn, int size)
  70 {
  71         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  72
  73         if (!corename)
  74                 return -ENOMEM;
  75
  76         if (size > core_name_size) /* racy but harmless */
  77                 core_name_size = size;
  78
  79         cn->size = ksize(corename);
  80         cn->corename = corename;
  81         return 0;
  82 }
  83
  84 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  85                                      va_list arg)
  86 {
  87         int free, need;
  88         va_list arg_copy;
  89
  90 again:
  91         free = cn->size - cn->used;
  92
  93         va_copy(arg_copy, arg);
  94         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  95         va_end(arg_copy);
  96
  97         if (need < free) {
  98                 cn->used += need;
  99                 return 0;
 100         }
 101
 102         if (!expand_corename(cn, cn->size + need - free + 1))
 103                 goto again;
 104
 105         return -ENOMEM;
 106 }
 107
 108 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 109 {
 110         va_list arg;
 111         int ret;
 112
 113         va_start(arg, fmt);
 114         ret = cn_vprintf(cn, fmt, arg);
 115         va_end(arg);
 116
 117         return ret;
 118 }
 119
 120 static __printf(2, 3)
 121 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 122 {
 123         int cur = cn->used;
 124         va_list arg;
 125         int ret;
 126
 127         va_start(arg, fmt);
 128         ret = cn_vprintf(cn, fmt, arg);
 129         va_end(arg);
 130
 131         if (ret == 0) {
 132                 /*
 133                  * Ensure that this coredump name component can't cause the
 134                  * resulting corefile path to consist of a ".." or ".".
 135                  */
 136                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 137                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 138                                 && cn->corename[cur+1] == '.'))
 139                         cn->corename[cur] = '!';
 140
 141                 /*
 142                  * Empty names are fishy and could be used to create a "//" in a
 143                  * corefile name, causing the coredump to happen one directory
 144                  * level too high. Enforce that all components of the core
 145                  * pattern are at least one character long.
 146                  */
 147                 if (cn->used == cur)
 148                         ret = cn_printf(cn, "!");
 149         }
 150
 151         for (; cur < cn->used; ++cur) {
 152                 if (cn->corename[cur] == '/')
 153                         cn->corename[cur] = '!';
 154         }
 155         return ret;
 156 }
 157
 158 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 159 {
 160         struct file *exe_file;
 161         char *pathbuf, *path, *ptr;
 162         int ret;
 163
 164         exe_file = get_mm_exe_file(current->mm);
 165         if (!exe_file)
 166                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 167
 168         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 169         if (!pathbuf) {
 170                 ret = -ENOMEM;
 171                 goto put_exe_file;
 172         }
 173
 174         path = file_path(exe_file, pathbuf, PATH_MAX);
 175         if (IS_ERR(path)) {
 176                 ret = PTR_ERR(path);
 177                 goto free_buf;
 178         }
 179
 180         if (name_only) {
 181                 ptr = strrchr(path, '/');
 182                 if (ptr)
 183                         path = ptr + 1;
 184         }
 185         ret = cn_esc_printf(cn, "%s", path);
 186
 187 free_buf:
 188         kfree(pathbuf);
 189 put_exe_file:
 190         fput(exe_file);
 191         return ret;
 192 }
 193
 194 /* format_corename will inspect the pattern parameter, and output a
 195  * name into corename, which must have space for at least
 196  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 197  */
 198 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 199                            size_t **argv, int *argc)
 200 {
 201         const struct cred *cred = current_cred();
 202         const char *pat_ptr = core_pattern;
 203         int ispipe = (*pat_ptr == '|');
 204         bool was_space = false;
 205         int pid_in_pattern = 0;
 206         int err = 0;
 207
 208         cn->used = 0;
 209         cn->corename = NULL;
 210         if (expand_corename(cn, core_name_size))
 211                 return -ENOMEM;
 212         cn->corename[0] = '\0';
 213
 214         if (ispipe) {
 215                 int argvs = sizeof(core_pattern) / 2;
 216                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 217                 if (!(*argv))
 218                         return -ENOMEM;
 219                 (*argv)[(*argc)++] = 0;
 220                 ++pat_ptr;
 221                 if (!(*pat_ptr))
 222                         return -ENOMEM;
 223         }
 224
 225         /* Repeat as long as we have more pattern to process and more output
 226            space */
 227         while (*pat_ptr) {
 228                 /*
 229                  * Split on spaces before doing template expansion so that
 230                  * %e and %E don't get split if they have spaces in them
 231                  */
 232                 if (ispipe) {
 233                         if (isspace(*pat_ptr)) {
 234                                 if (cn->used != 0)
 235                                         was_space = true;
 236                                 pat_ptr++;
 237                                 continue;
 238                         } else if (was_space) {
 239                                 was_space = false;
 240                                 err = cn_printf(cn, "%c", '\0');
 241                                 if (err)
 242                                         return err;
 243                                 (*argv)[(*argc)++] = cn->used;
 244                         }
 245                 }
 246                 if (*pat_ptr != '%') {
 247                         err = cn_printf(cn, "%c", *pat_ptr++);
 248                 } else {
 249                         switch (*++pat_ptr) {
 250                         /* single % at the end, drop that */
 251                         case 0:
 252                                 goto out;
 253                         /* Double percent, output one percent */
 254                         case '%':
 255                                 err = cn_printf(cn, "%c", '%');
 256                                 break;
 257                         /* pid */
 258                         case 'p':
 259                                 pid_in_pattern = 1;
 260                                 err = cn_printf(cn, "%d",
 261                                               task_tgid_vnr(current));
 262                                 break;
 263                         /* global pid */
 264                         case 'P':
 265                                 err = cn_printf(cn, "%d",
 266                                               task_tgid_nr(current));
 267                                 break;
 268                         case 'i':
 269                                 err = cn_printf(cn, "%d",
 270                                               task_pid_vnr(current));
 271                                 break;
 272                         case 'I':
 273                                 err = cn_printf(cn, "%d",
 274                                               task_pid_nr(current));
 275                                 break;
 276                         /* uid */
 277                         case 'u':
 278                                 err = cn_printf(cn, "%u",
 279                                                 from_kuid(&init_user_ns,
 280                                                           cred->uid));
 281                                 break;
 282                         /* gid */
 283                         case 'g':
 284                                 err = cn_printf(cn, "%u",
 285                                                 from_kgid(&init_user_ns,
 286                                                           cred->gid));
 287                                 break;
 288                         case 'd':
 289                                 err = cn_printf(cn, "%d",
 290                                         __get_dumpable(cprm->mm_flags));
 291                                 break;
 292                         /* signal that caused the coredump */
 293                         case 's':
 294                                 err = cn_printf(cn, "%d",
 295                                                 cprm->siginfo->si_signo);
 296                                 break;
 297                         /* UNIX time of coredump */
 298                         case 't': {
 299                                 time64_t time;
 300
 301                                 time = ktime_get_real_seconds();
 302                                 err = cn_printf(cn, "%lld", time);
 303                                 break;
 304                         }
 305                         /* hostname */
 306                         case 'h':
 307                                 down_read(&uts_sem);
 308                                 err = cn_esc_printf(cn, "%s",
 309                                               utsname()->nodename);
 310                                 up_read(&uts_sem);
 311                                 break;
 312                         /* executable, could be changed by prctl PR_SET_NAME etc */
 313                         case 'e':
 314                                 err = cn_esc_printf(cn, "%s", current->comm);
 315                                 break;
 316                         /* file name of executable */
 317                         case 'f':
 318                                 err = cn_print_exe_file(cn, true);
 319                                 break;
 320                         case 'E':
 321                                 err = cn_print_exe_file(cn, false);
 322                                 break;
 323                         /* core limit size */
 324                         case 'c':
 325                                 err = cn_printf(cn, "%lu",
 326                                               rlimit(RLIMIT_CORE));
 327                                 break;
 328                         /* CPU the task ran on */
 329                         case 'C':
 330                                 err = cn_printf(cn, "%d", cprm->cpu);
 331                                 break;
 332                         default:
 333                                 break;
 334                         }
 335                         ++pat_ptr;
 336                 }
 337
 338                 if (err)
 339                         return err;
 340         }
 341
 342 out:
 343         /* Backward compatibility with core_uses_pid:
 344          *
 345          * If core_pattern does not include a %p (as is the default)
 346          * and core_uses_pid is set, then .%pid will be appended to
 347          * the filename. Do not do this for piped commands. */
 348         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 349                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 350                 if (err)
 351                         return err;
 352         }
 353         return ispipe;
 354 }
 355
 356 static int zap_process(struct task_struct *start, int exit_code)
 357 {
 358         struct task_struct *t;
 359         int nr = 0;
 360
 361         /* Allow SIGKILL, see prepare_signal() */
 362         start->signal->flags = SIGNAL_GROUP_EXIT;
 363         start->signal->group_exit_code = exit_code;
 364         start->signal->group_stop_count = 0;
 365
 366         for_each_thread(start, t) {
 367                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 368                 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
 369                         sigaddset(&t->pending.signal, SIGKILL);
 370                         signal_wake_up(t, 1);
 371                         nr++;
 372                 }
 373         }
 374
 375         return nr;
 376 }
 377
 378 static int zap_threads(struct task_struct *tsk,
 379                         struct core_state *core_state, int exit_code)
 380 {
 381         struct signal_struct *signal = tsk->signal;
 382         int nr = -EAGAIN;
 383
 384         spin_lock_irq(&tsk->sighand->siglock);
 385         if (!(signal->flags & SIGNAL_GROUP_EXIT) && !signal->group_exec_task) {
 386                 signal->core_state = core_state;
 387                 nr = zap_process(tsk, exit_code);
 388                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 389                 tsk->flags |= PF_DUMPCORE;
 390                 atomic_set(&core_state->nr_threads, nr);
 391         }
 392         spin_unlock_irq(&tsk->sighand->siglock);
 393         return nr;
 394 }
 395
 396 static int coredump_wait(int exit_code, struct core_state *core_state)
 397 {
 398         struct task_struct *tsk = current;
 399         int core_waiters = -EBUSY;
 400
 401         init_completion(&core_state->startup);
 402         core_state->dumper.task = tsk;
 403         core_state->dumper.next = NULL;
 404
 405         core_waiters = zap_threads(tsk, core_state, exit_code);
 406         if (core_waiters > 0) {
 407                 struct core_thread *ptr;
 408
 409                 wait_for_completion_state(&core_state->startup,
 410                                           TASK_UNINTERRUPTIBLE|TASK_FREEZABLE);
 411                 /*
 412                  * Wait for all the threads to become inactive, so that
 413                  * all the thread context (extended register state, like
 414                  * fpu etc) gets copied to the memory.
 415                  */
 416                 ptr = core_state->dumper.next;
 417                 while (ptr != NULL) {
 418                         wait_task_inactive(ptr->task, TASK_ANY);
 419                         ptr = ptr->next;
 420                 }
 421         }
 422
 423         return core_waiters;
 424 }
 425
 426 static void coredump_finish(bool core_dumped)
 427 {
 428         struct core_thread *curr, *next;
 429         struct task_struct *task;
 430
 431         spin_lock_irq(&current->sighand->siglock);
 432         if (core_dumped && !__fatal_signal_pending(current))
 433                 current->signal->group_exit_code |= 0x80;
 434         next = current->signal->core_state->dumper.next;
 435         current->signal->core_state = NULL;
 436         spin_unlock_irq(&current->sighand->siglock);
 437
 438         while ((curr = next) != NULL) {
 439                 next = curr->next;
 440                 task = curr->task;
 441                 /*
 442                  * see coredump_task_exit(), curr->task must not see
 443                  * ->task == NULL before we read ->next.
 444                  */
 445                 smp_mb();
 446                 curr->task = NULL;
 447                 wake_up_process(task);
 448         }
 449 }
 450
 451 static bool dump_interrupted(void)
 452 {
 453         /*
 454          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 455          * can do try_to_freeze() and check __fatal_signal_pending(),
 456          * but then we need to teach dump_write() to restart and clear
 457          * TIF_SIGPENDING.
 458          */
 459         return fatal_signal_pending(current) || freezing(current);
 460 }
 461
 462 static void wait_for_dump_helpers(struct file *file)
 463 {
 464         struct pipe_inode_info *pipe = file->private_data;
 465
 466         pipe_lock(pipe);
 467         pipe->readers++;
 468         pipe->writers--;
 469         wake_up_interruptible_sync(&pipe->rd_wait);
 470         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 471         pipe_unlock(pipe);
 472
 473         /*
 474          * We actually want wait_event_freezable() but then we need
 475          * to clear TIF_SIGPENDING and improve dump_interrupted().
 476          */
 477         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 478
 479         pipe_lock(pipe);
 480         pipe->readers--;
 481         pipe->writers++;
 482         pipe_unlock(pipe);
 483 }
 484
 485 /*
 486  * umh_pipe_setup
 487  * helper function to customize the process used
 488  * to collect the core in userspace.  Specifically
 489  * it sets up a pipe and installs it as fd 0 (stdin)
 490  * for the process.  Returns 0 on success, or
 491  * PTR_ERR on failure.
 492  * Note that it also sets the core limit to 1.  This
 493  * is a special value that we use to trap recursive
 494  * core dumps
 495  */
 496 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 497 {
 498         struct file *files[2];
 499         struct coredump_params *cp = (struct coredump_params *)info->data;
 500         int err = create_pipe_files(files, 0);
 501         if (err)
 502                 return err;
 503
 504         cp->file = files[1];
 505
 506         err = replace_fd(0, files[0], 0);
 507         fput(files[0]);
 508         /* and disallow core files too */
 509         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 510
 511         return err;
 512 }
 513
 514 void do_coredump(const kernel_siginfo_t *siginfo)
 515 {
 516         struct core_state core_state;
 517         struct core_name cn;
 518         struct mm_struct *mm = current->mm;
 519         struct linux_binfmt * binfmt;
 520         const struct cred *old_cred;
 521         struct cred *cred;
 522         int retval = 0;
 523         int ispipe;
 524         size_t *argv = NULL;
 525         int argc = 0;
 526         /* require nonrelative corefile path and be extra careful */
 527         bool need_suid_safe = false;
 528         bool core_dumped = false;
 529         static atomic_t core_dump_count = ATOMIC_INIT(0);
 530         struct coredump_params cprm = {
 531                 .siginfo = siginfo,
 532                 .limit = rlimit(RLIMIT_CORE),
 533                 /*
 534                  * We must use the same mm->flags while dumping core to avoid
 535                  * inconsistency of bit flags, since this flag is not protected
 536                  * by any locks.
 537                  */
 538                 .mm_flags = mm->flags,
 539                 .vma_meta = NULL,
 540                 .cpu = raw_smp_processor_id(),
 541         };
 542
 543         audit_core_dumps(siginfo->si_signo);
 544
 545         binfmt = mm->binfmt;
 546         if (!binfmt || !binfmt->core_dump)
 547                 goto fail;
 548         if (!__get_dumpable(cprm.mm_flags))
 549                 goto fail;
 550
 551         cred = prepare_creds();
 552         if (!cred)
 553                 goto fail;
 554         /*
 555          * We cannot trust fsuid as being the "true" uid of the process
 556          * nor do we know its entire history. We only know it was tainted
 557          * so we dump it as root in mode 2, and only into a controlled
 558          * environment (pipe handler or fully qualified path).
 559          */
 560         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 561                 /* Setuid core dump mode */
 562                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 563                 need_suid_safe = true;
 564         }
 565
 566         retval = coredump_wait(siginfo->si_signo, &core_state);
 567         if (retval < 0)
 568                 goto fail_creds;
 569
 570         old_cred = override_creds(cred);
 571
 572         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 573
 574         if (ispipe) {
 575                 int argi;
 576                 int dump_count;
 577                 char **helper_argv;
 578                 struct subprocess_info *sub_info;
 579
 580                 if (ispipe < 0) {
 581                         printk(KERN_WARNING "format_corename failed\n");
 582                         printk(KERN_WARNING "Aborting core\n");
 583                         goto fail_unlock;
 584                 }
 585
 586                 if (cprm.limit == 1) {
 587                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 588                          *
 589                          * Normally core limits are irrelevant to pipes, since
 590                          * we're not writing to the file system, but we use
 591                          * cprm.limit of 1 here as a special value, this is a
 592                          * consistent way to catch recursive crashes.
 593                          * We can still crash if the core_pattern binary sets
 594                          * RLIM_CORE = !1, but it runs as root, and can do
 595                          * lots of stupid things.
 596                          *
 597                          * Note that we use task_tgid_vnr here to grab the pid
 598                          * of the process group leader.  That way we get the
 599                          * right pid if a thread in a multi-threaded
 600                          * core_pattern process dies.
 601                          */
 602                         printk(KERN_WARNING
 603                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 604                                 task_tgid_vnr(current), current->comm);
 605                         printk(KERN_WARNING "Aborting core\n");
 606                         goto fail_unlock;
 607                 }
 608                 cprm.limit = RLIM_INFINITY;
 609
 610                 dump_count = atomic_inc_return(&core_dump_count);
 611                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 612                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 613                                task_tgid_vnr(current), current->comm);
 614                         printk(KERN_WARNING "Skipping core dump\n");
 615                         goto fail_dropcount;
 616                 }
 617
 618                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 619                                             GFP_KERNEL);
 620                 if (!helper_argv) {
 621                         printk(KERN_WARNING "%s failed to allocate memory\n",
 622                                __func__);
 623                         goto fail_dropcount;
 624                 }
 625                 for (argi = 0; argi < argc; argi++)
 626                         helper_argv[argi] = cn.corename + argv[argi];
 627                 helper_argv[argi] = NULL;
 628
 629                 retval = -ENOMEM;
 630                 sub_info = call_usermodehelper_setup(helper_argv[0],
 631                                                 helper_argv, NULL, GFP_KERNEL,
 632                                                 umh_pipe_setup, NULL, &cprm);
 633                 if (sub_info)
 634                         retval = call_usermodehelper_exec(sub_info,
 635                                                           UMH_WAIT_EXEC);
 636
 637                 kfree(helper_argv);
 638                 if (retval) {
 639                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 640                                cn.corename);
 641                         goto close_fail;
 642                 }
 643         } else {
 644                 struct user_namespace *mnt_userns;
 645                 struct inode *inode;
 646                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 647                                  O_LARGEFILE | O_EXCL;
 648
 649                 if (cprm.limit < binfmt->min_coredump)
 650                         goto fail_unlock;
 651
 652                 if (need_suid_safe && cn.corename[0] != '/') {
 653                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 654                                 "to fully qualified path!\n",
 655                                 task_tgid_vnr(current), current->comm);
 656                         printk(KERN_WARNING "Skipping core dump\n");
 657                         goto fail_unlock;
 658                 }
 659
 660                 /*
 661                  * Unlink the file if it exists unless this is a SUID
 662                  * binary - in that case, we're running around with root
 663                  * privs and don't want to unlink another user's coredump.
 664                  */
 665                 if (!need_suid_safe) {
 666                         /*
 667                          * If it doesn't exist, that's fine. If there's some
 668                          * other problem, we'll catch it at the filp_open().
 669                          */
 670                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 671                 }
 672
 673                 /*
 674                  * There is a race between unlinking and creating the
 675                  * file, but if that causes an EEXIST here, that's
 676                  * fine - another process raced with us while creating
 677                  * the corefile, and the other process won. To userspace,
 678                  * what matters is that at least one of the two processes
 679                  * writes its coredump successfully, not which one.
 680                  */
 681                 if (need_suid_safe) {
 682                         /*
 683                          * Using user namespaces, normal user tasks can change
 684                          * their current->fs->root to point to arbitrary
 685                          * directories. Since the intention of the "only dump
 686                          * with a fully qualified path" rule is to control where
 687                          * coredumps may be placed using root privileges,
 688                          * current->fs->root must not be used. Instead, use the
 689                          * root directory of init_task.
 690                          */
 691                         struct path root;
 692
 693                         task_lock(&init_task);
 694                         get_fs_root(init_task.fs, &root);
 695                         task_unlock(&init_task);
 696                         cprm.file = file_open_root(&root, cn.corename,
 697                                                    open_flags, 0600);
 698                         path_put(&root);
 699                 } else {
 700                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 701                 }
 702                 if (IS_ERR(cprm.file))
 703                         goto fail_unlock;
 704
 705                 inode = file_inode(cprm.file);
 706                 if (inode->i_nlink > 1)
 707                         goto close_fail;
 708                 if (d_unhashed(cprm.file->f_path.dentry))
 709                         goto close_fail;
 710                 /*
 711                  * AK: actually i see no reason to not allow this for named
 712                  * pipes etc, but keep the previous behaviour for now.
 713                  */
 714                 if (!S_ISREG(inode->i_mode))
 715                         goto close_fail;
 716                 /*
 717                  * Don't dump core if the filesystem changed owner or mode
 718                  * of the file during file creation. This is an issue when
 719                  * a process dumps core while its cwd is e.g. on a vfat
 720                  * filesystem.
 721                  */
 722                 mnt_userns = file_mnt_user_ns(cprm.file);
 723                 if (!vfsuid_eq_kuid(i_uid_into_vfsuid(mnt_userns, inode),
 724                                     current_fsuid())) {
 725                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file owner\n",
 726                                             cn.corename);
 727                         goto close_fail;
 728                 }
 729                 if ((inode->i_mode & 0677) != 0600) {
 730                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file permissions\n",
 731                                             cn.corename);
 732                         goto close_fail;
 733                 }
 734                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 735                         goto close_fail;
 736                 if (do_truncate(mnt_userns, cprm.file->f_path.dentry,
 737                                 0, 0, cprm.file))
 738                         goto close_fail;
 739         }
 740
 741         /* get us an unshared descriptor table; almost always a no-op */
 742         /* The cell spufs coredump code reads the file descriptor tables */
 743         retval = unshare_files();
 744         if (retval)
 745                 goto close_fail;
 746         if (!dump_interrupted()) {
 747                 /*
 748                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 749                  * have this set to NULL.
 750                  */
 751                 if (!cprm.file) {
 752                         pr_info("Core dump to |%s disabled\n", cn.corename);
 753                         goto close_fail;
 754                 }
 755                 if (!dump_vma_snapshot(&cprm))
 756                         goto close_fail;
 757
 758                 file_start_write(cprm.file);
 759                 core_dumped = binfmt->core_dump(&cprm);
 760                 /*
 761                  * Ensures that file size is big enough to contain the current
 762                  * file postion. This prevents gdb from complaining about
 763                  * a truncated file if the last "write" to the file was
 764                  * dump_skip.
 765                  */
 766                 if (cprm.to_skip) {
 767                         cprm.to_skip--;
 768                         dump_emit(&cprm, "", 1);
 769                 }
 770                 file_end_write(cprm.file);
 771                 free_vma_snapshot(&cprm);
 772         }
 773         if (ispipe && core_pipe_limit)
 774                 wait_for_dump_helpers(cprm.file);
 775 close_fail:
 776         if (cprm.file)
 777                 filp_close(cprm.file, NULL);
 778 fail_dropcount:
 779         if (ispipe)
 780                 atomic_dec(&core_dump_count);
 781 fail_unlock:
 782         kfree(argv);
 783         kfree(cn.corename);
 784         coredump_finish(core_dumped);
 785         revert_creds(old_cred);
 786 fail_creds:
 787         put_cred(cred);
 788 fail:
 789         return;
 790 }
 791
 792 /*
 793  * Core dumping helper functions.  These are the only things you should
 794  * do on a core-file: use only these functions to write out all the
 795  * necessary info.
 796  */
 797 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 798 {
 799         struct file *file = cprm->file;
 800         loff_t pos = file->f_pos;
 801         ssize_t n;
 802         if (cprm->written + nr > cprm->limit)
 803                 return 0;
 804
 805
 806         if (dump_interrupted())
 807                 return 0;
 808         n = __kernel_write(file, addr, nr, &pos);
 809         if (n != nr)
 810                 return 0;
 811         file->f_pos = pos;
 812         cprm->written += n;
 813         cprm->pos += n;
 814
 815         return 1;
 816 }
 817
 818 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 819 {
 820         static char zeroes[PAGE_SIZE];
 821         struct file *file = cprm->file;
 822         if (file->f_mode & FMODE_LSEEK) {
 823                 if (dump_interrupted() ||
 824                     vfs_llseek(file, nr, SEEK_CUR) < 0)
 825                         return 0;
 826                 cprm->pos += nr;
 827                 return 1;
 828         } else {
 829                 while (nr > PAGE_SIZE) {
 830                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 831                                 return 0;
 832                         nr -= PAGE_SIZE;
 833                 }
 834                 return __dump_emit(cprm, zeroes, nr);
 835         }
 836 }
 837
 838 static int dump_emit_page(struct coredump_params *cprm, struct page *page)
 839 {
 840         struct bio_vec bvec = {
 841                 .bv_page        = page,
 842                 .bv_offset      = 0,
 843                 .bv_len         = PAGE_SIZE,
 844         };
 845         struct iov_iter iter;
 846         struct file *file = cprm->file;
 847         loff_t pos;
 848         ssize_t n;
 849
 850         if (cprm->to_skip) {
 851                 if (!__dump_skip(cprm, cprm->to_skip))
 852                         return 0;
 853                 cprm->to_skip = 0;
 854         }
 855         if (cprm->written + PAGE_SIZE > cprm->limit)
 856                 return 0;
 857         if (dump_interrupted())
 858                 return 0;
 859         pos = file->f_pos;
 860         iov_iter_bvec(&iter, ITER_SOURCE, &bvec, 1, PAGE_SIZE);
 861         n = __kernel_write_iter(cprm->file, &iter, &pos);
 862         if (n != PAGE_SIZE)
 863                 return 0;
 864         file->f_pos = pos;
 865         cprm->written += PAGE_SIZE;
 866         cprm->pos += PAGE_SIZE;
 867
 868         return 1;
 869 }
 870
 871 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 872 {
 873         if (cprm->to_skip) {
 874                 if (!__dump_skip(cprm, cprm->to_skip))
 875                         return 0;
 876                 cprm->to_skip = 0;
 877         }
 878         return __dump_emit(cprm, addr, nr);
 879 }
 880 EXPORT_SYMBOL(dump_emit);
 881
 882 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 883 {
 884         cprm->to_skip = pos - cprm->pos;
 885 }
 886 EXPORT_SYMBOL(dump_skip_to);
 887
 888 void dump_skip(struct coredump_params *cprm, size_t nr)
 889 {
 890         cprm->to_skip += nr;
 891 }
 892 EXPORT_SYMBOL(dump_skip);
 893
 894 #ifdef CONFIG_ELF_CORE
 895 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 896                     unsigned long len)
 897 {
 898         unsigned long addr;
 899
 900         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 901                 struct page *page;
 902
 903                 /*
 904                  * To avoid having to allocate page tables for virtual address
 905                  * ranges that have never been used yet, and also to make it
 906                  * easy to generate sparse core files, use a helper that returns
 907                  * NULL when encountering an empty page table entry that would
 908                  * otherwise have been filled with the zero page.
 909                  */
 910                 page = get_dump_page(addr);
 911                 if (page) {
 912                         int stop = !dump_emit_page(cprm, page);
 913                         put_page(page);
 914                         if (stop)
 915                                 return 0;
 916                 } else {
 917                         dump_skip(cprm, PAGE_SIZE);
 918                 }
 919         }
 920         return 1;
 921 }
 922 #endif
 923
 924 int dump_align(struct coredump_params *cprm, int align)
 925 {
 926         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
 927         if (align & (align - 1))
 928                 return 0;
 929         if (mod)
 930                 cprm->to_skip += align - mod;
 931         return 1;
 932 }
 933 EXPORT_SYMBOL(dump_align);
 934
 935 #ifdef CONFIG_SYSCTL
 936
 937 void validate_coredump_safety(void)
 938 {
 939         if (suid_dumpable == SUID_DUMP_ROOT &&
 940             core_pattern[0] != '/' && core_pattern[0] != '|') {
 941                 pr_warn(
 942 "Unsafe core_pattern used with fs.suid_dumpable=2.\n"
 943 "Pipe handler or fully qualified core dump path required.\n"
 944 "Set kernel.core_pattern before fs.suid_dumpable.\n"
 945                 );
 946         }
 947 }
 948
 949 static int proc_dostring_coredump(struct ctl_table *table, int write,
 950                   void *buffer, size_t *lenp, loff_t *ppos)
 951 {
 952         int error = proc_dostring(table, write, buffer, lenp, ppos);
 953
 954         if (!error)
 955                 validate_coredump_safety();
 956         return error;
 957 }
 958
 959 static struct ctl_table coredump_sysctls[] = {
 960         {
 961                 .procname       = "core_uses_pid",
 962                 .data           = &core_uses_pid,
 963                 .maxlen         = sizeof(int),
 964                 .mode           = 0644,
 965                 .proc_handler   = proc_dointvec,
 966         },
 967         {
 968                 .procname       = "core_pattern",
 969                 .data           = core_pattern,
 970                 .maxlen         = CORENAME_MAX_SIZE,
 971                 .mode           = 0644,
 972                 .proc_handler   = proc_dostring_coredump,
 973         },
 974         {
 975                 .procname       = "core_pipe_limit",
 976                 .data           = &core_pipe_limit,
 977                 .maxlen         = sizeof(unsigned int),
 978                 .mode           = 0644,
 979                 .proc_handler   = proc_dointvec,
 980         },
 981         { }
 982 };
 983
 984 static int __init init_fs_coredump_sysctls(void)
 985 {
 986         register_sysctl_init("kernel", coredump_sysctls);
 987         return 0;
 988 }
 989 fs_initcall(init_fs_coredump_sysctls);
 990 #endif /* CONFIG_SYSCTL */
 991
 992 /*
 993  * The purpose of always_dump_vma() is to make sure that special kernel mappings
 994  * that are useful for post-mortem analysis are included in every core dump.
 995  * In that way we ensure that the core dump is fully interpretable later
 996  * without matching up the same kernel and hardware config to see what PC values
 997  * meant. These special mappings include - vDSO, vsyscall, and other
 998  * architecture specific mappings
 999  */
1000 static bool always_dump_vma(struct vm_area_struct *vma)
1001 {
1002         /* Any vsyscall mappings? */
1003         if (vma == get_gate_vma(vma->vm_mm))
1004                 return true;
1005
1006         /*
1007          * Assume that all vmas with a .name op should always be dumped.
1008          * If this changes, a new vm_ops field can easily be added.
1009          */
1010         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
1011                 return true;
1012
1013         /*
1014          * arch_vma_name() returns non-NULL for special architecture mappings,
1015          * such as vDSO sections.
1016          */
1017         if (arch_vma_name(vma))
1018                 return true;
1019
1020         return false;
1021 }
1022
1023 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
1024
1025 /*
1026  * Decide how much of @vma's contents should be included in a core dump.
1027  */
1028 static unsigned long vma_dump_size(struct vm_area_struct *vma,
1029                                    unsigned long mm_flags)
1030 {
1031 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
1032
1033         /* always dump the vdso and vsyscall sections */
1034         if (always_dump_vma(vma))
1035                 goto whole;
1036
1037         if (vma->vm_flags & VM_DONTDUMP)
1038                 return 0;
1039
1040         /* support for DAX */
1041         if (vma_is_dax(vma)) {
1042                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1043                         goto whole;
1044                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1045                         goto whole;
1046                 return 0;
1047         }
1048
1049         /* Hugetlb memory check */
1050         if (is_vm_hugetlb_page(vma)) {
1051                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1052                         goto whole;
1053                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1054                         goto whole;
1055                 return 0;
1056         }
1057
1058         /* Do not dump I/O mapped devices or special mappings */
1059         if (vma->vm_flags & VM_IO)
1060                 return 0;
1061
1062         /* By default, dump shared memory if mapped from an anonymous file. */
1063         if (vma->vm_flags & VM_SHARED) {
1064                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1065                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1066                         goto whole;
1067                 return 0;
1068         }
1069
1070         /* Dump segments that have been written to.  */
1071         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1072                 goto whole;
1073         if (vma->vm_file == NULL)
1074                 return 0;
1075
1076         if (FILTER(MAPPED_PRIVATE))
1077                 goto whole;
1078
1079         /*
1080          * If this is the beginning of an executable file mapping,
1081          * dump the first page to aid in determining what was mapped here.
1082          */
1083         if (FILTER(ELF_HEADERS) &&
1084             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1085                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1086                         return PAGE_SIZE;
1087
1088                 /*
1089                  * ELF libraries aren't always executable.
1090                  * We'll want to check whether the mapping starts with the ELF
1091                  * magic, but not now - we're holding the mmap lock,
1092                  * so copy_from_user() doesn't work here.
1093                  * Use a placeholder instead, and fix it up later in
1094                  * dump_vma_snapshot().
1095                  */
1096                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1097         }
1098
1099 #undef  FILTER
1100
1101         return 0;
1102
1103 whole:
1104         return vma->vm_end - vma->vm_start;
1105 }
1106
1107 /*
1108  * Helper function for iterating across a vma list.  It ensures that the caller
1109  * will visit `gate_vma' prior to terminating the search.
1110  */
1111 static struct vm_area_struct *coredump_next_vma(struct ma_state *mas,
1112                                        struct vm_area_struct *vma,
1113                                        struct vm_area_struct *gate_vma)
1114 {
1115         if (gate_vma && (vma == gate_vma))
1116                 return NULL;
1117
1118         vma = mas_next(mas, ULONG_MAX);
1119         if (vma)
1120                 return vma;
1121         return gate_vma;
1122 }
1123
1124 static void free_vma_snapshot(struct coredump_params *cprm)
1125 {
1126         if (cprm->vma_meta) {
1127                 int i;
1128                 for (i = 0; i < cprm->vma_count; i++) {
1129                         struct file *file = cprm->vma_meta[i].file;
1130                         if (file)
1131                                 fput(file);
1132                 }
1133                 kvfree(cprm->vma_meta);
1134                 cprm->vma_meta = NULL;
1135         }
1136 }
1137
1138 /*
1139  * Under the mmap_lock, take a snapshot of relevant information about the task's
1140  * VMAs.
1141  */
1142 static bool dump_vma_snapshot(struct coredump_params *cprm)
1143 {
1144         struct vm_area_struct *gate_vma, *vma = NULL;
1145         struct mm_struct *mm = current->mm;
1146         MA_STATE(mas, &mm->mm_mt, 0, 0);
1147         int i = 0;
1148
1149         /*
1150          * Once the stack expansion code is fixed to not change VMA bounds
1151          * under mmap_lock in read mode, this can be changed to take the
1152          * mmap_lock in read mode.
1153          */
1154         if (mmap_write_lock_killable(mm))
1155                 return false;
1156
1157         cprm->vma_data_size = 0;
1158         gate_vma = get_gate_vma(mm);
1159         cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1160
1161         cprm->vma_meta = kvmalloc_array(cprm->vma_count, sizeof(*cprm->vma_meta), GFP_KERNEL);
1162         if (!cprm->vma_meta) {
1163                 mmap_write_unlock(mm);
1164                 return false;
1165         }
1166
1167         while ((vma = coredump_next_vma(&mas, vma, gate_vma)) != NULL) {
1168                 struct core_vma_metadata *m = cprm->vma_meta + i;
1169
1170                 m->start = vma->vm_start;
1171                 m->end = vma->vm_end;
1172                 m->flags = vma->vm_flags;
1173                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1174                 m->pgoff = vma->vm_pgoff;
1175                 m->file = vma->vm_file;
1176                 if (m->file)
1177                         get_file(m->file);
1178                 i++;
1179         }
1180
1181         mmap_write_unlock(mm);
1182
1183         for (i = 0; i < cprm->vma_count; i++) {
1184                 struct core_vma_metadata *m = cprm->vma_meta + i;
1185
1186                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1187                         char elfmag[SELFMAG];
1188
1189                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1190                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1191                                 m->dump_size = 0;
1192                         } else {
1193                                 m->dump_size = PAGE_SIZE;
1194                         }
1195                 }
1196
1197                 cprm->vma_data_size += m->dump_size;
1198         }
1199
1200         return true;
1201 }