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