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