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