fs/coredump.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/slab.h>
   3 #include <linux/file.h>
   4 #include <linux/fdtable.h>
   5 #include <linux/freezer.h>
   6 #include <linux/mm.h>
   7 #include <linux/stat.h>
   8 #include <linux/fcntl.h>
   9 #include <linux/swap.h>
  10 #include <linux/ctype.h>
  11 #include <linux/string.h>
  12 #include <linux/init.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/perf_event.h>
  15 #include <linux/highmem.h>
  16 #include <linux/spinlock.h>
  17 #include <linux/key.h>
  18 #include <linux/personality.h>
  19 #include <linux/binfmts.h>
  20 #include <linux/coredump.h>
  21 #include <linux/sched/coredump.h>
  22 #include <linux/sched/signal.h>
  23 #include <linux/sched/task_stack.h>
  24 #include <linux/utsname.h>
  25 #include <linux/pid_namespace.h>
  26 #include <linux/module.h>
  27 #include <linux/namei.h>
  28 #include <linux/mount.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/tsacct_kern.h>
  32 #include <linux/cn_proc.h>
  33 #include <linux/audit.h>
  34 #include <linux/kmod.h>
  35 #include <linux/fsnotify.h>
  36 #include <linux/fs_struct.h>
  37 #include <linux/pipe_fs_i.h>
  38 #include <linux/oom.h>
  39 #include <linux/compat.h>
  40 #include <linux/fs.h>
  41 #include <linux/path.h>
  42 #include <linux/timekeeping.h>
  43 #include <linux/sysctl.h>
  44 #include <linux/elf.h>
  45
  46 #include <linux/uaccess.h>
  47 #include <asm/mmu_context.h>
  48 #include <asm/tlb.h>
  49 #include <asm/exec.h>
  50
  51 #include <trace/events/task.h>
  52 #include "internal.h"
  53
  54 #include <trace/events/sched.h>
  55
  56 static bool dump_vma_snapshot(struct coredump_params *cprm);
  57 static void free_vma_snapshot(struct coredump_params *cprm);
  58
  59 static int core_uses_pid;
  60 static unsigned int core_pipe_limit;
  61 static char core_pattern[CORENAME_MAX_SIZE] = "core";
  62 static int core_name_size = CORENAME_MAX_SIZE;
  63
  64 struct core_name {
  65         char *corename;
  66         int used, size;
  67 };
  68
  69 static int expand_corename(struct core_name *cn, int size)
  70 {
  71         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  72
  73         if (!corename)
  74                 return -ENOMEM;
  75
  76         if (size > core_name_size) /* racy but harmless */
  77                 core_name_size = size;
  78
  79         cn->size = ksize(corename);
  80         cn->corename = corename;
  81         return 0;
  82 }
  83
  84 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  85                                      va_list arg)
  86 {
  87         int free, need;
  88         va_list arg_copy;
  89
  90 again:
  91         free = cn->size - cn->used;
  92
  93         va_copy(arg_copy, arg);
  94         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  95         va_end(arg_copy);
  96
  97         if (need < free) {
  98                 cn->used += need;
  99                 return 0;
 100         }
 101
 102         if (!expand_corename(cn, cn->size + need - free + 1))
 103                 goto again;
 104
 105         return -ENOMEM;
 106 }
 107
 108 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 109 {
 110         va_list arg;
 111         int ret;
 112
 113         va_start(arg, fmt);
 114         ret = cn_vprintf(cn, fmt, arg);
 115         va_end(arg);
 116
 117         return ret;
 118 }
 119
 120 static __printf(2, 3)
 121 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 122 {
 123         int cur = cn->used;
 124         va_list arg;
 125         int ret;
 126
 127         va_start(arg, fmt);
 128         ret = cn_vprintf(cn, fmt, arg);
 129         va_end(arg);
 130
 131         if (ret == 0) {
 132                 /*
 133                  * Ensure that this coredump name component can't cause the
 134                  * resulting corefile path to consist of a ".." or ".".
 135                  */
 136                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 137                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 138                                 && cn->corename[cur+1] == '.'))
 139                         cn->corename[cur] = '!';
 140
 141                 /*
 142                  * Empty names are fishy and could be used to create a "//" in a
 143                  * corefile name, causing the coredump to happen one directory
 144                  * level too high. Enforce that all components of the core
 145                  * pattern are at least one character long.
 146                  */
 147                 if (cn->used == cur)
 148                         ret = cn_printf(cn, "!");
 149         }
 150
 151         for (; cur < cn->used; ++cur) {
 152                 if (cn->corename[cur] == '/')
 153                         cn->corename[cur] = '!';
 154         }
 155         return ret;
 156 }
 157
 158 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 159 {
 160         struct file *exe_file;
 161         char *pathbuf, *path, *ptr;
 162         int ret;
 163
 164         exe_file = get_mm_exe_file(current->mm);
 165         if (!exe_file)
 166                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 167
 168         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 169         if (!pathbuf) {
 170                 ret = -ENOMEM;
 171                 goto put_exe_file;
 172         }
 173
 174         path = file_path(exe_file, pathbuf, PATH_MAX);
 175         if (IS_ERR(path)) {
 176                 ret = PTR_ERR(path);
 177                 goto free_buf;
 178         }
 179
 180         if (name_only) {
 181                 ptr = strrchr(path, '/');
 182                 if (ptr)
 183                         path = ptr + 1;
 184         }
 185         ret = cn_esc_printf(cn, "%s", path);
 186
 187 free_buf:
 188         kfree(pathbuf);
 189 put_exe_file:
 190         fput(exe_file);
 191         return ret;
 192 }
 193
 194 /* format_corename will inspect the pattern parameter, and output a
 195  * name into corename, which must have space for at least
 196  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 197  */
 198 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 199                            size_t **argv, int *argc)
 200 {
 201         const struct cred *cred = current_cred();
 202         const char *pat_ptr = core_pattern;
 203         int ispipe = (*pat_ptr == '|');
 204         bool was_space = false;
 205         int pid_in_pattern = 0;
 206         int err = 0;
 207
 208         cn->used = 0;
 209         cn->corename = NULL;
 210         if (expand_corename(cn, core_name_size))
 211                 return -ENOMEM;
 212         cn->corename[0] = '\0';
 213
 214         if (ispipe) {
 215                 int argvs = sizeof(core_pattern) / 2;
 216                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 217                 if (!(*argv))
 218                         return -ENOMEM;
 219                 (*argv)[(*argc)++] = 0;
 220                 ++pat_ptr;
 221                 if (!(*pat_ptr))
 222                         return -ENOMEM;
 223         }
 224
 225         /* Repeat as long as we have more pattern to process and more output
 226            space */
 227         while (*pat_ptr) {
 228                 /*
 229                  * Split on spaces before doing template expansion so that
 230                  * %e and %E don't get split if they have spaces in them
 231                  */
 232                 if (ispipe) {
 233                         if (isspace(*pat_ptr)) {
 234                                 if (cn->used != 0)
 235                                         was_space = true;
 236                                 pat_ptr++;
 237                                 continue;
 238                         } else if (was_space) {
 239                                 was_space = false;
 240                                 err = cn_printf(cn, "%c", '\0');
 241                                 if (err)
 242                                         return err;
 243                                 (*argv)[(*argc)++] = cn->used;
 244                         }
 245                 }
 246                 if (*pat_ptr != '%') {
 247                         err = cn_printf(cn, "%c", *pat_ptr++);
 248                 } else {
 249                         switch (*++pat_ptr) {
 250                         /* single % at the end, drop that */
 251                         case 0:
 252                                 goto out;
 253                         /* Double percent, output one percent */
 254                         case '%':
 255                                 err = cn_printf(cn, "%c", '%');
 256                                 break;
 257                         /* pid */
 258                         case 'p':
 259                                 pid_in_pattern = 1;
 260                                 err = cn_printf(cn, "%d",
 261                                               task_tgid_vnr(current));
 262                                 break;
 263                         /* global pid */
 264                         case 'P':
 265                                 err = cn_printf(cn, "%d",
 266                                               task_tgid_nr(current));
 267                                 break;
 268                         case 'i':
 269                                 err = cn_printf(cn, "%d",
 270                                               task_pid_vnr(current));
 271                                 break;
 272                         case 'I':
 273                                 err = cn_printf(cn, "%d",
 274                                               task_pid_nr(current));
 275                                 break;
 276                         /* uid */
 277                         case 'u':
 278                                 err = cn_printf(cn, "%u",
 279                                                 from_kuid(&init_user_ns,
 280                                                           cred->uid));
 281                                 break;
 282                         /* gid */
 283                         case 'g':
 284                                 err = cn_printf(cn, "%u",
 285                                                 from_kgid(&init_user_ns,
 286                                                           cred->gid));
 287                                 break;
 288                         case 'd':
 289                                 err = cn_printf(cn, "%d",
 290                                         __get_dumpable(cprm->mm_flags));
 291                                 break;
 292                         /* signal that caused the coredump */
 293                         case 's':
 294                                 err = cn_printf(cn, "%d",
 295                                                 cprm->siginfo->si_signo);
 296                                 break;
 297                         /* UNIX time of coredump */
 298                         case 't': {
 299                                 time64_t time;
 300
 301                                 time = ktime_get_real_seconds();
 302                                 err = cn_printf(cn, "%lld", time);
 303                                 break;
 304                         }
 305                         /* hostname */
 306                         case 'h':
 307                                 down_read(&uts_sem);
 308                                 err = cn_esc_printf(cn, "%s",
 309                                               utsname()->nodename);
 310                                 up_read(&uts_sem);
 311                                 break;
 312                         /* executable, could be changed by prctl PR_SET_NAME etc */
 313                         case 'e':
 314                                 err = cn_esc_printf(cn, "%s", current->comm);
 315                                 break;
 316                         /* file name of executable */
 317                         case 'f':
 318                                 err = cn_print_exe_file(cn, true);
 319                                 break;
 320                         case 'E':
 321                                 err = cn_print_exe_file(cn, false);
 322                                 break;
 323                         /* core limit size */
 324                         case 'c':
 325                                 err = cn_printf(cn, "%lu",
 326                                               rlimit(RLIMIT_CORE));
 327                                 break;
 328                         default:
 329                                 break;
 330                         }
 331                         ++pat_ptr;
 332                 }
 333
 334                 if (err)
 335                         return err;
 336         }
 337
 338 out:
 339         /* Backward compatibility with core_uses_pid:
 340          *
 341          * If core_pattern does not include a %p (as is the default)
 342          * and core_uses_pid is set, then .%pid will be appended to
 343          * the filename. Do not do this for piped commands. */
 344         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 345                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 346                 if (err)
 347                         return err;
 348         }
 349         return ispipe;
 350 }
 351
 352 static int zap_process(struct task_struct *start, int exit_code)
 353 {
 354         struct task_struct *t;
 355         int nr = 0;
 356
 357         /* Allow SIGKILL, see prepare_signal() */
 358         start->signal->flags = SIGNAL_GROUP_EXIT;
 359         start->signal->group_exit_code = exit_code;
 360         start->signal->group_stop_count = 0;
 361
 362         for_each_thread(start, t) {
 363                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 364                 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
 365                         sigaddset(&t->pending.signal, SIGKILL);
 366                         signal_wake_up(t, 1);
 367                         nr++;
 368                 }
 369         }
 370
 371         return nr;
 372 }
 373
 374 static int zap_threads(struct task_struct *tsk,
 375                         struct core_state *core_state, int exit_code)
 376 {
 377         struct signal_struct *signal = tsk->signal;
 378         int nr = -EAGAIN;
 379
 380         spin_lock_irq(&tsk->sighand->siglock);
 381         if (!(signal->flags & SIGNAL_GROUP_EXIT) && !signal->group_exec_task) {
 382                 signal->core_state = core_state;
 383                 nr = zap_process(tsk, exit_code);
 384                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 385                 tsk->flags |= PF_DUMPCORE;
 386                 atomic_set(&core_state->nr_threads, nr);
 387         }
 388         spin_unlock_irq(&tsk->sighand->siglock);
 389         return nr;
 390 }
 391
 392 static int coredump_wait(int exit_code, struct core_state *core_state)
 393 {
 394         struct task_struct *tsk = current;
 395         int core_waiters = -EBUSY;
 396
 397         init_completion(&core_state->startup);
 398         core_state->dumper.task = tsk;
 399         core_state->dumper.next = NULL;
 400
 401         core_waiters = zap_threads(tsk, core_state, exit_code);
 402         if (core_waiters > 0) {
 403                 struct core_thread *ptr;
 404
 405                 wait_for_completion_state(&core_state->startup,
 406                                           TASK_UNINTERRUPTIBLE|TASK_FREEZABLE);
 407                 /*
 408                  * Wait for all the threads to become inactive, so that
 409                  * all the thread context (extended register state, like
 410                  * fpu etc) gets copied to the memory.
 411                  */
 412                 ptr = core_state->dumper.next;
 413                 while (ptr != NULL) {
 414                         wait_task_inactive(ptr->task, TASK_ANY);
 415                         ptr = ptr->next;
 416                 }
 417         }
 418
 419         return core_waiters;
 420 }
 421
 422 static void coredump_finish(bool core_dumped)
 423 {
 424         struct core_thread *curr, *next;
 425         struct task_struct *task;
 426
 427         spin_lock_irq(&current->sighand->siglock);
 428         if (core_dumped && !__fatal_signal_pending(current))
 429                 current->signal->group_exit_code |= 0x80;
 430         next = current->signal->core_state->dumper.next;
 431         current->signal->core_state = NULL;
 432         spin_unlock_irq(&current->sighand->siglock);
 433
 434         while ((curr = next) != NULL) {
 435                 next = curr->next;
 436                 task = curr->task;
 437                 /*
 438                  * see coredump_task_exit(), curr->task must not see
 439                  * ->task == NULL before we read ->next.
 440                  */
 441                 smp_mb();
 442                 curr->task = NULL;
 443                 wake_up_process(task);
 444         }
 445 }
 446
 447 static bool dump_interrupted(void)
 448 {
 449         /*
 450          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 451          * can do try_to_freeze() and check __fatal_signal_pending(),
 452          * but then we need to teach dump_write() to restart and clear
 453          * TIF_SIGPENDING.
 454          */
 455         return fatal_signal_pending(current) || freezing(current);
 456 }
 457
 458 static void wait_for_dump_helpers(struct file *file)
 459 {
 460         struct pipe_inode_info *pipe = file->private_data;
 461
 462         pipe_lock(pipe);
 463         pipe->readers++;
 464         pipe->writers--;
 465         wake_up_interruptible_sync(&pipe->rd_wait);
 466         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 467         pipe_unlock(pipe);
 468
 469         /*
 470          * We actually want wait_event_freezable() but then we need
 471          * to clear TIF_SIGPENDING and improve dump_interrupted().
 472          */
 473         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 474
 475         pipe_lock(pipe);
 476         pipe->readers--;
 477         pipe->writers++;
 478         pipe_unlock(pipe);
 479 }
 480
 481 /*
 482  * umh_pipe_setup
 483  * helper function to customize the process used
 484  * to collect the core in userspace.  Specifically
 485  * it sets up a pipe and installs it as fd 0 (stdin)
 486  * for the process.  Returns 0 on success, or
 487  * PTR_ERR on failure.
 488  * Note that it also sets the core limit to 1.  This
 489  * is a special value that we use to trap recursive
 490  * core dumps
 491  */
 492 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 493 {
 494         struct file *files[2];
 495         struct coredump_params *cp = (struct coredump_params *)info->data;
 496         int err = create_pipe_files(files, 0);
 497         if (err)
 498                 return err;
 499
 500         cp->file = files[1];
 501
 502         err = replace_fd(0, files[0], 0);
 503         fput(files[0]);
 504         /* and disallow core files too */
 505         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 506
 507         return err;
 508 }
 509
 510 void do_coredump(const kernel_siginfo_t *siginfo)
 511 {
 512         struct core_state core_state;
 513         struct core_name cn;
 514         struct mm_struct *mm = current->mm;
 515         struct linux_binfmt * binfmt;
 516         const struct cred *old_cred;
 517         struct cred *cred;
 518         int retval = 0;
 519         int ispipe;
 520         size_t *argv = NULL;
 521         int argc = 0;
 522         /* require nonrelative corefile path and be extra careful */
 523         bool need_suid_safe = false;
 524         bool core_dumped = false;
 525         static atomic_t core_dump_count = ATOMIC_INIT(0);
 526         struct coredump_params cprm = {
 527                 .siginfo = siginfo,
 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                 .vma_meta = NULL,
 536         };
 537
 538         audit_core_dumps(siginfo->si_signo);
 539
 540         binfmt = mm->binfmt;
 541         if (!binfmt || !binfmt->core_dump)
 542                 goto fail;
 543         if (!__get_dumpable(cprm.mm_flags))
 544                 goto fail;
 545
 546         cred = prepare_creds();
 547         if (!cred)
 548                 goto fail;
 549         /*
 550          * We cannot trust fsuid as being the "true" uid of the process
 551          * nor do we know its entire history. We only know it was tainted
 552          * so we dump it as root in mode 2, and only into a controlled
 553          * environment (pipe handler or fully qualified path).
 554          */
 555         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 556                 /* Setuid core dump mode */
 557                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 558                 need_suid_safe = true;
 559         }
 560
 561         retval = coredump_wait(siginfo->si_signo, &core_state);
 562         if (retval < 0)
 563                 goto fail_creds;
 564
 565         old_cred = override_creds(cred);
 566
 567         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 568
 569         if (ispipe) {
 570                 int argi;
 571                 int dump_count;
 572                 char **helper_argv;
 573                 struct subprocess_info *sub_info;
 574
 575                 if (ispipe < 0) {
 576                         printk(KERN_WARNING "format_corename failed\n");
 577                         printk(KERN_WARNING "Aborting core\n");
 578                         goto fail_unlock;
 579                 }
 580
 581                 if (cprm.limit == 1) {
 582                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 583                          *
 584                          * Normally core limits are irrelevant to pipes, since
 585                          * we're not writing to the file system, but we use
 586                          * cprm.limit of 1 here as a special value, this is a
 587                          * consistent way to catch recursive crashes.
 588                          * We can still crash if the core_pattern binary sets
 589                          * RLIM_CORE = !1, but it runs as root, and can do
 590                          * lots of stupid things.
 591                          *
 592                          * Note that we use task_tgid_vnr here to grab the pid
 593                          * of the process group leader.  That way we get the
 594                          * right pid if a thread in a multi-threaded
 595                          * core_pattern process dies.
 596                          */
 597                         printk(KERN_WARNING
 598                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 599                                 task_tgid_vnr(current), current->comm);
 600                         printk(KERN_WARNING "Aborting core\n");
 601                         goto fail_unlock;
 602                 }
 603                 cprm.limit = RLIM_INFINITY;
 604
 605                 dump_count = atomic_inc_return(&core_dump_count);
 606                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 607                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 608                                task_tgid_vnr(current), current->comm);
 609                         printk(KERN_WARNING "Skipping core dump\n");
 610                         goto fail_dropcount;
 611                 }
 612
 613                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 614                                             GFP_KERNEL);
 615                 if (!helper_argv) {
 616                         printk(KERN_WARNING "%s failed to allocate memory\n",
 617                                __func__);
 618                         goto fail_dropcount;
 619                 }
 620                 for (argi = 0; argi < argc; argi++)
 621                         helper_argv[argi] = cn.corename + argv[argi];
 622                 helper_argv[argi] = NULL;
 623
 624                 retval = -ENOMEM;
 625                 sub_info = call_usermodehelper_setup(helper_argv[0],
 626                                                 helper_argv, NULL, GFP_KERNEL,
 627                                                 umh_pipe_setup, NULL, &cprm);
 628                 if (sub_info)
 629                         retval = call_usermodehelper_exec(sub_info,
 630                                                           UMH_WAIT_EXEC);
 631
 632                 kfree(helper_argv);
 633                 if (retval) {
 634                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 635                                cn.corename);
 636                         goto close_fail;
 637                 }
 638         } else {
 639                 struct user_namespace *mnt_userns;
 640                 struct inode *inode;
 641                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 642                                  O_LARGEFILE | O_EXCL;
 643
 644                 if (cprm.limit < binfmt->min_coredump)
 645                         goto fail_unlock;
 646
 647                 if (need_suid_safe && cn.corename[0] != '/') {
 648                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 649                                 "to fully qualified path!\n",
 650                                 task_tgid_vnr(current), current->comm);
 651                         printk(KERN_WARNING "Skipping core dump\n");
 652                         goto fail_unlock;
 653                 }
 654
 655                 /*
 656                  * Unlink the file if it exists unless this is a SUID
 657                  * binary - in that case, we're running around with root
 658                  * privs and don't want to unlink another user's coredump.
 659                  */
 660                 if (!need_suid_safe) {
 661                         /*
 662                          * If it doesn't exist, that's fine. If there's some
 663                          * other problem, we'll catch it at the filp_open().
 664                          */
 665                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 666                 }
 667
 668                 /*
 669                  * There is a race between unlinking and creating the
 670                  * file, but if that causes an EEXIST here, that's
 671                  * fine - another process raced with us while creating
 672                  * the corefile, and the other process won. To userspace,
 673                  * what matters is that at least one of the two processes
 674                  * writes its coredump successfully, not which one.
 675                  */
 676                 if (need_suid_safe) {
 677                         /*
 678                          * Using user namespaces, normal user tasks can change
 679                          * their current->fs->root to point to arbitrary
 680                          * directories. Since the intention of the "only dump
 681                          * with a fully qualified path" rule is to control where
 682                          * coredumps may be placed using root privileges,
 683                          * current->fs->root must not be used. Instead, use the
 684                          * root directory of init_task.
 685                          */
 686                         struct path root;
 687
 688                         task_lock(&init_task);
 689                         get_fs_root(init_task.fs, &root);
 690                         task_unlock(&init_task);
 691                         cprm.file = file_open_root(&root, cn.corename,
 692                                                    open_flags, 0600);
 693                         path_put(&root);
 694                 } else {
 695                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 696                 }
 697                 if (IS_ERR(cprm.file))
 698                         goto fail_unlock;
 699
 700                 inode = file_inode(cprm.file);
 701                 if (inode->i_nlink > 1)
 702                         goto close_fail;
 703                 if (d_unhashed(cprm.file->f_path.dentry))
 704                         goto close_fail;
 705                 /*
 706                  * AK: actually i see no reason to not allow this for named
 707                  * pipes etc, but keep the previous behaviour for now.
 708                  */
 709                 if (!S_ISREG(inode->i_mode))
 710                         goto close_fail;
 711                 /*
 712                  * Don't dump core if the filesystem changed owner or mode
 713                  * of the file during file creation. This is an issue when
 714                  * a process dumps core while its cwd is e.g. on a vfat
 715                  * filesystem.
 716                  */
 717                 mnt_userns = file_mnt_user_ns(cprm.file);
 718                 if (!uid_eq(i_uid_into_mnt(mnt_userns, inode),
 719                             current_fsuid())) {
 720                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file owner\n",
 721                                             cn.corename);
 722                         goto close_fail;
 723                 }
 724                 if ((inode->i_mode & 0677) != 0600) {
 725                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file permissions\n",
 726                                             cn.corename);
 727                         goto close_fail;
 728                 }
 729                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 730                         goto close_fail;
 731                 if (do_truncate(mnt_userns, cprm.file->f_path.dentry,
 732                                 0, 0, cprm.file))
 733                         goto close_fail;
 734         }
 735
 736         /* get us an unshared descriptor table; almost always a no-op */
 737         /* The cell spufs coredump code reads the file descriptor tables */
 738         retval = unshare_files();
 739         if (retval)
 740                 goto close_fail;
 741         if (!dump_interrupted()) {
 742                 /*
 743                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 744                  * have this set to NULL.
 745                  */
 746                 if (!cprm.file) {
 747                         pr_info("Core dump to |%s disabled\n", cn.corename);
 748                         goto close_fail;
 749                 }
 750                 if (!dump_vma_snapshot(&cprm))
 751                         goto close_fail;
 752
 753                 file_start_write(cprm.file);
 754                 core_dumped = binfmt->core_dump(&cprm);
 755                 /*
 756                  * Ensures that file size is big enough to contain the current
 757                  * file postion. This prevents gdb from complaining about
 758                  * a truncated file if the last "write" to the file was
 759                  * dump_skip.
 760                  */
 761                 if (cprm.to_skip) {
 762                         cprm.to_skip--;
 763                         dump_emit(&cprm, "", 1);
 764                 }
 765                 file_end_write(cprm.file);
 766                 free_vma_snapshot(&cprm);
 767         }
 768         if (ispipe && core_pipe_limit)
 769                 wait_for_dump_helpers(cprm.file);
 770 close_fail:
 771         if (cprm.file)
 772                 filp_close(cprm.file, NULL);
 773 fail_dropcount:
 774         if (ispipe)
 775                 atomic_dec(&core_dump_count);
 776 fail_unlock:
 777         kfree(argv);
 778         kfree(cn.corename);
 779         coredump_finish(core_dumped);
 780         revert_creds(old_cred);
 781 fail_creds:
 782         put_cred(cred);
 783 fail:
 784         return;
 785 }
 786
 787 /*
 788  * Core dumping helper functions.  These are the only things you should
 789  * do on a core-file: use only these functions to write out all the
 790  * necessary info.
 791  */
 792 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 793 {
 794         struct file *file = cprm->file;
 795         loff_t pos = file->f_pos;
 796         ssize_t n;
 797         if (cprm->written + nr > cprm->limit)
 798                 return 0;
 799
 800
 801         if (dump_interrupted())
 802                 return 0;
 803         n = __kernel_write(file, addr, nr, &pos);
 804         if (n != nr)
 805                 return 0;
 806         file->f_pos = pos;
 807         cprm->written += n;
 808         cprm->pos += n;
 809
 810         return 1;
 811 }
 812
 813 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 814 {
 815         static char zeroes[PAGE_SIZE];
 816         struct file *file = cprm->file;
 817         if (file->f_mode & FMODE_LSEEK) {
 818                 if (dump_interrupted() ||
 819                     vfs_llseek(file, nr, SEEK_CUR) < 0)
 820                         return 0;
 821                 cprm->pos += nr;
 822                 return 1;
 823         } else {
 824                 while (nr > PAGE_SIZE) {
 825                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 826                                 return 0;
 827                         nr -= PAGE_SIZE;
 828                 }
 829                 return __dump_emit(cprm, zeroes, nr);
 830         }
 831 }
 832
 833 static int dump_emit_page(struct coredump_params *cprm, struct page *page)
 834 {
 835         struct bio_vec bvec = {
 836                 .bv_page        = page,
 837                 .bv_offset      = 0,
 838                 .bv_len         = PAGE_SIZE,
 839         };
 840         struct iov_iter iter;
 841         struct file *file = cprm->file;
 842         loff_t pos;
 843         ssize_t n;
 844
 845         if (cprm->to_skip) {
 846                 if (!__dump_skip(cprm, cprm->to_skip))
 847                         return 0;
 848                 cprm->to_skip = 0;
 849         }
 850         if (cprm->written + PAGE_SIZE > cprm->limit)
 851                 return 0;
 852         if (dump_interrupted())
 853                 return 0;
 854         pos = file->f_pos;
 855         iov_iter_bvec(&iter, WRITE, &bvec, 1, PAGE_SIZE);
 856         n = __kernel_write_iter(cprm->file, &iter, &pos);
 857         if (n != PAGE_SIZE)
 858                 return 0;
 859         file->f_pos = pos;
 860         cprm->written += PAGE_SIZE;
 861         cprm->pos += PAGE_SIZE;
 862
 863         return 1;
 864 }
 865
 866 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 867 {
 868         if (cprm->to_skip) {
 869                 if (!__dump_skip(cprm, cprm->to_skip))
 870                         return 0;
 871                 cprm->to_skip = 0;
 872         }
 873         return __dump_emit(cprm, addr, nr);
 874 }
 875 EXPORT_SYMBOL(dump_emit);
 876
 877 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 878 {
 879         cprm->to_skip = pos - cprm->pos;
 880 }
 881 EXPORT_SYMBOL(dump_skip_to);
 882
 883 void dump_skip(struct coredump_params *cprm, size_t nr)
 884 {
 885         cprm->to_skip += nr;
 886 }
 887 EXPORT_SYMBOL(dump_skip);
 888
 889 #ifdef CONFIG_ELF_CORE
 890 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 891                     unsigned long len)
 892 {
 893         unsigned long addr;
 894
 895         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 896                 struct page *page;
 897
 898                 /*
 899                  * To avoid having to allocate page tables for virtual address
 900                  * ranges that have never been used yet, and also to make it
 901                  * easy to generate sparse core files, use a helper that returns
 902                  * NULL when encountering an empty page table entry that would
 903                  * otherwise have been filled with the zero page.
 904                  */
 905                 page = get_dump_page(addr);
 906                 if (page) {
 907                         int stop = !dump_emit_page(cprm, page);
 908                         put_page(page);
 909                         if (stop)
 910                                 return 0;
 911                 } else {
 912                         dump_skip(cprm, PAGE_SIZE);
 913                 }
 914         }
 915         return 1;
 916 }
 917 #endif
 918
 919 int dump_align(struct coredump_params *cprm, int align)
 920 {
 921         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
 922         if (align & (align - 1))
 923                 return 0;
 924         if (mod)
 925                 cprm->to_skip += align - mod;
 926         return 1;
 927 }
 928 EXPORT_SYMBOL(dump_align);
 929
 930 #ifdef CONFIG_SYSCTL
 931
 932 void validate_coredump_safety(void)
 933 {
 934         if (suid_dumpable == SUID_DUMP_ROOT &&
 935             core_pattern[0] != '/' && core_pattern[0] != '|') {
 936                 pr_warn(
 937 "Unsafe core_pattern used with fs.suid_dumpable=2.\n"
 938 "Pipe handler or fully qualified core dump path required.\n"
 939 "Set kernel.core_pattern before fs.suid_dumpable.\n"
 940                 );
 941         }
 942 }
 943
 944 static int proc_dostring_coredump(struct ctl_table *table, int write,
 945                   void *buffer, size_t *lenp, loff_t *ppos)
 946 {
 947         int error = proc_dostring(table, write, buffer, lenp, ppos);
 948
 949         if (!error)
 950                 validate_coredump_safety();
 951         return error;
 952 }
 953
 954 static struct ctl_table coredump_sysctls[] = {
 955         {
 956                 .procname       = "core_uses_pid",
 957                 .data           = &core_uses_pid,
 958                 .maxlen         = sizeof(int),
 959                 .mode           = 0644,
 960                 .proc_handler   = proc_dointvec,
 961         },
 962         {
 963                 .procname       = "core_pattern",
 964                 .data           = core_pattern,
 965                 .maxlen         = CORENAME_MAX_SIZE,
 966                 .mode           = 0644,
 967                 .proc_handler   = proc_dostring_coredump,
 968         },
 969         {
 970                 .procname       = "core_pipe_limit",
 971                 .data           = &core_pipe_limit,
 972                 .maxlen         = sizeof(unsigned int),
 973                 .mode           = 0644,
 974                 .proc_handler   = proc_dointvec,
 975         },
 976         { }
 977 };
 978
 979 static int __init init_fs_coredump_sysctls(void)
 980 {
 981         register_sysctl_init("kernel", coredump_sysctls);
 982         return 0;
 983 }
 984 fs_initcall(init_fs_coredump_sysctls);
 985 #endif /* CONFIG_SYSCTL */
 986
 987 /*
 988  * The purpose of always_dump_vma() is to make sure that special kernel mappings
 989  * that are useful for post-mortem analysis are included in every core dump.
 990  * In that way we ensure that the core dump is fully interpretable later
 991  * without matching up the same kernel and hardware config to see what PC values
 992  * meant. These special mappings include - vDSO, vsyscall, and other
 993  * architecture specific mappings
 994  */
 995 static bool always_dump_vma(struct vm_area_struct *vma)
 996 {
 997         /* Any vsyscall mappings? */
 998         if (vma == get_gate_vma(vma->vm_mm))
 999                 return true;
1000
1001         /*
1002          * Assume that all vmas with a .name op should always be dumped.
1003          * If this changes, a new vm_ops field can easily be added.
1004          */
1005         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
1006                 return true;
1007
1008         /*
1009          * arch_vma_name() returns non-NULL for special architecture mappings,
1010          * such as vDSO sections.
1011          */
1012         if (arch_vma_name(vma))
1013                 return true;
1014
1015         return false;
1016 }
1017
1018 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
1019
1020 /*
1021  * Decide how much of @vma's contents should be included in a core dump.
1022  */
1023 static unsigned long vma_dump_size(struct vm_area_struct *vma,
1024                                    unsigned long mm_flags)
1025 {
1026 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
1027
1028         /* always dump the vdso and vsyscall sections */
1029         if (always_dump_vma(vma))
1030                 goto whole;
1031
1032         if (vma->vm_flags & VM_DONTDUMP)
1033                 return 0;
1034
1035         /* support for DAX */
1036         if (vma_is_dax(vma)) {
1037                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1038                         goto whole;
1039                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1040                         goto whole;
1041                 return 0;
1042         }
1043
1044         /* Hugetlb memory check */
1045         if (is_vm_hugetlb_page(vma)) {
1046                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1047                         goto whole;
1048                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1049                         goto whole;
1050                 return 0;
1051         }
1052
1053         /* Do not dump I/O mapped devices or special mappings */
1054         if (vma->vm_flags & VM_IO)
1055                 return 0;
1056
1057         /* By default, dump shared memory if mapped from an anonymous file. */
1058         if (vma->vm_flags & VM_SHARED) {
1059                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1060                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1061                         goto whole;
1062                 return 0;
1063         }
1064
1065         /* Dump segments that have been written to.  */
1066         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1067                 goto whole;
1068         if (vma->vm_file == NULL)
1069                 return 0;
1070
1071         if (FILTER(MAPPED_PRIVATE))
1072                 goto whole;
1073
1074         /*
1075          * If this is the beginning of an executable file mapping,
1076          * dump the first page to aid in determining what was mapped here.
1077          */
1078         if (FILTER(ELF_HEADERS) &&
1079             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1080                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1081                         return PAGE_SIZE;
1082
1083                 /*
1084                  * ELF libraries aren't always executable.
1085                  * We'll want to check whether the mapping starts with the ELF
1086                  * magic, but not now - we're holding the mmap lock,
1087                  * so copy_from_user() doesn't work here.
1088                  * Use a placeholder instead, and fix it up later in
1089                  * dump_vma_snapshot().
1090                  */
1091                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1092         }
1093
1094 #undef  FILTER
1095
1096         return 0;
1097
1098 whole:
1099         return vma->vm_end - vma->vm_start;
1100 }
1101
1102 /*
1103  * Helper function for iterating across a vma list.  It ensures that the caller
1104  * will visit `gate_vma' prior to terminating the search.
1105  */
1106 static struct vm_area_struct *coredump_next_vma(struct ma_state *mas,
1107                                        struct vm_area_struct *vma,
1108                                        struct vm_area_struct *gate_vma)
1109 {
1110         if (gate_vma && (vma == gate_vma))
1111                 return NULL;
1112
1113         vma = mas_next(mas, ULONG_MAX);
1114         if (vma)
1115                 return vma;
1116         return gate_vma;
1117 }
1118
1119 static void free_vma_snapshot(struct coredump_params *cprm)
1120 {
1121         if (cprm->vma_meta) {
1122                 int i;
1123                 for (i = 0; i < cprm->vma_count; i++) {
1124                         struct file *file = cprm->vma_meta[i].file;
1125                         if (file)
1126                                 fput(file);
1127                 }
1128                 kvfree(cprm->vma_meta);
1129                 cprm->vma_meta = NULL;
1130         }
1131 }
1132
1133 /*
1134  * Under the mmap_lock, take a snapshot of relevant information about the task's
1135  * VMAs.
1136  */
1137 static bool dump_vma_snapshot(struct coredump_params *cprm)
1138 {
1139         struct vm_area_struct *gate_vma, *vma = NULL;
1140         struct mm_struct *mm = current->mm;
1141         MA_STATE(mas, &mm->mm_mt, 0, 0);
1142         int i = 0;
1143
1144         /*
1145          * Once the stack expansion code is fixed to not change VMA bounds
1146          * under mmap_lock in read mode, this can be changed to take the
1147          * mmap_lock in read mode.
1148          */
1149         if (mmap_write_lock_killable(mm))
1150                 return false;
1151
1152         cprm->vma_data_size = 0;
1153         gate_vma = get_gate_vma(mm);
1154         cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1155
1156         cprm->vma_meta = kvmalloc_array(cprm->vma_count, sizeof(*cprm->vma_meta), GFP_KERNEL);
1157         if (!cprm->vma_meta) {
1158                 mmap_write_unlock(mm);
1159                 return false;
1160         }
1161
1162         while ((vma = coredump_next_vma(&mas, vma, gate_vma)) != NULL) {
1163                 struct core_vma_metadata *m = cprm->vma_meta + i;
1164
1165                 m->start = vma->vm_start;
1166                 m->end = vma->vm_end;
1167                 m->flags = vma->vm_flags;
1168                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1169                 m->pgoff = vma->vm_pgoff;
1170                 m->file = vma->vm_file;
1171                 if (m->file)
1172                         get_file(m->file);
1173                 i++;
1174         }
1175
1176         mmap_write_unlock(mm);
1177
1178         for (i = 0; i < cprm->vma_count; i++) {
1179                 struct core_vma_metadata *m = cprm->vma_meta + i;
1180
1181                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1182                         char elfmag[SELFMAG];
1183
1184                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1185                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1186                                 m->dump_size = 0;
1187                         } else {
1188                                 m->dump_size = PAGE_SIZE;
1189                         }
1190                 }
1191
1192                 cprm->vma_data_size += m->dump_size;
1193         }
1194
1195         return true;
1196 }