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;
  72
  73         size = kmalloc_size_roundup(size);
  74         corename = krealloc(cn->corename, size, GFP_KERNEL);
  75
  76         if (!corename)
  77                 return -ENOMEM;
  78
  79         if (size > core_name_size) /* racy but harmless */
  80                 core_name_size = size;
  81
  82         cn->size = size;
  83         cn->corename = corename;
  84         return 0;
  85 }
  86
  87 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  88                                      va_list arg)
  89 {
  90         int free, need;
  91         va_list arg_copy;
  92
  93 again:
  94         free = cn->size - cn->used;
  95
  96         va_copy(arg_copy, arg);
  97         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  98         va_end(arg_copy);
  99
 100         if (need < free) {
 101                 cn->used += need;
 102                 return 0;
 103         }
 104
 105         if (!expand_corename(cn, cn->size + need - free + 1))
 106                 goto again;
 107
 108         return -ENOMEM;
 109 }
 110
 111 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 112 {
 113         va_list arg;
 114         int ret;
 115
 116         va_start(arg, fmt);
 117         ret = cn_vprintf(cn, fmt, arg);
 118         va_end(arg);
 119
 120         return ret;
 121 }
 122
 123 static __printf(2, 3)
 124 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 125 {
 126         int cur = cn->used;
 127         va_list arg;
 128         int ret;
 129
 130         va_start(arg, fmt);
 131         ret = cn_vprintf(cn, fmt, arg);
 132         va_end(arg);
 133
 134         if (ret == 0) {
 135                 /*
 136                  * Ensure that this coredump name component can't cause the
 137                  * resulting corefile path to consist of a ".." or ".".
 138                  */
 139                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 140                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 141                                 && cn->corename[cur+1] == '.'))
 142                         cn->corename[cur] = '!';
 143
 144                 /*
 145                  * Empty names are fishy and could be used to create a "//" in a
 146                  * corefile name, causing the coredump to happen one directory
 147                  * level too high. Enforce that all components of the core
 148                  * pattern are at least one character long.
 149                  */
 150                 if (cn->used == cur)
 151                         ret = cn_printf(cn, "!");
 152         }
 153
 154         for (; cur < cn->used; ++cur) {
 155                 if (cn->corename[cur] == '/')
 156                         cn->corename[cur] = '!';
 157         }
 158         return ret;
 159 }
 160
 161 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 162 {
 163         struct file *exe_file;
 164         char *pathbuf, *path, *ptr;
 165         int ret;
 166
 167         exe_file = get_mm_exe_file(current->mm);
 168         if (!exe_file)
 169                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 170
 171         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 172         if (!pathbuf) {
 173                 ret = -ENOMEM;
 174                 goto put_exe_file;
 175         }
 176
 177         path = file_path(exe_file, pathbuf, PATH_MAX);
 178         if (IS_ERR(path)) {
 179                 ret = PTR_ERR(path);
 180                 goto free_buf;
 181         }
 182
 183         if (name_only) {
 184                 ptr = strrchr(path, '/');
 185                 if (ptr)
 186                         path = ptr + 1;
 187         }
 188         ret = cn_esc_printf(cn, "%s", path);
 189
 190 free_buf:
 191         kfree(pathbuf);
 192 put_exe_file:
 193         fput(exe_file);
 194         return ret;
 195 }
 196
 197 /* format_corename will inspect the pattern parameter, and output a
 198  * name into corename, which must have space for at least
 199  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 200  */
 201 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 202                            size_t **argv, int *argc)
 203 {
 204         const struct cred *cred = current_cred();
 205         const char *pat_ptr = core_pattern;
 206         int ispipe = (*pat_ptr == '|');
 207         bool was_space = false;
 208         int pid_in_pattern = 0;
 209         int err = 0;
 210
 211         cn->used = 0;
 212         cn->corename = NULL;
 213         if (expand_corename(cn, core_name_size))
 214                 return -ENOMEM;
 215         cn->corename[0] = '\0';
 216
 217         if (ispipe) {
 218                 int argvs = sizeof(core_pattern) / 2;
 219                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 220                 if (!(*argv))
 221                         return -ENOMEM;
 222                 (*argv)[(*argc)++] = 0;
 223                 ++pat_ptr;
 224                 if (!(*pat_ptr))
 225                         return -ENOMEM;
 226         }
 227
 228         /* Repeat as long as we have more pattern to process and more output
 229            space */
 230         while (*pat_ptr) {
 231                 /*
 232                  * Split on spaces before doing template expansion so that
 233                  * %e and %E don't get split if they have spaces in them
 234                  */
 235                 if (ispipe) {
 236                         if (isspace(*pat_ptr)) {
 237                                 if (cn->used != 0)
 238                                         was_space = true;
 239                                 pat_ptr++;
 240                                 continue;
 241                         } else if (was_space) {
 242                                 was_space = false;
 243                                 err = cn_printf(cn, "%c", '\0');
 244                                 if (err)
 245                                         return err;
 246                                 (*argv)[(*argc)++] = cn->used;
 247                         }
 248                 }
 249                 if (*pat_ptr != '%') {
 250                         err = cn_printf(cn, "%c", *pat_ptr++);
 251                 } else {
 252                         switch (*++pat_ptr) {
 253                         /* single % at the end, drop that */
 254                         case 0:
 255                                 goto out;
 256                         /* Double percent, output one percent */
 257                         case '%':
 258                                 err = cn_printf(cn, "%c", '%');
 259                                 break;
 260                         /* pid */
 261                         case 'p':
 262                                 pid_in_pattern = 1;
 263                                 err = cn_printf(cn, "%d",
 264                                               task_tgid_vnr(current));
 265                                 break;
 266                         /* global pid */
 267                         case 'P':
 268                                 err = cn_printf(cn, "%d",
 269                                               task_tgid_nr(current));
 270                                 break;
 271                         case 'i':
 272                                 err = cn_printf(cn, "%d",
 273                                               task_pid_vnr(current));
 274                                 break;
 275                         case 'I':
 276                                 err = cn_printf(cn, "%d",
 277                                               task_pid_nr(current));
 278                                 break;
 279                         /* uid */
 280                         case 'u':
 281                                 err = cn_printf(cn, "%u",
 282                                                 from_kuid(&init_user_ns,
 283                                                           cred->uid));
 284                                 break;
 285                         /* gid */
 286                         case 'g':
 287                                 err = cn_printf(cn, "%u",
 288                                                 from_kgid(&init_user_ns,
 289                                                           cred->gid));
 290                                 break;
 291                         case 'd':
 292                                 err = cn_printf(cn, "%d",
 293                                         __get_dumpable(cprm->mm_flags));
 294                                 break;
 295                         /* signal that caused the coredump */
 296                         case 's':
 297                                 err = cn_printf(cn, "%d",
 298                                                 cprm->siginfo->si_signo);
 299                                 break;
 300                         /* UNIX time of coredump */
 301                         case 't': {
 302                                 time64_t time;
 303
 304                                 time = ktime_get_real_seconds();
 305                                 err = cn_printf(cn, "%lld", time);
 306                                 break;
 307                         }
 308                         /* hostname */
 309                         case 'h':
 310                                 down_read(&uts_sem);
 311                                 err = cn_esc_printf(cn, "%s",
 312                                               utsname()->nodename);
 313                                 up_read(&uts_sem);
 314                                 break;
 315                         /* executable, could be changed by prctl PR_SET_NAME etc */
 316                         case 'e':
 317                                 err = cn_esc_printf(cn, "%s", current->comm);
 318                                 break;
 319                         /* file name of executable */
 320                         case 'f':
 321                                 err = cn_print_exe_file(cn, true);
 322                                 break;
 323                         case 'E':
 324                                 err = cn_print_exe_file(cn, false);
 325                                 break;
 326                         /* core limit size */
 327                         case 'c':
 328                                 err = cn_printf(cn, "%lu",
 329                                               rlimit(RLIMIT_CORE));
 330                                 break;
 331                         /* CPU the task ran on */
 332                         case 'C':
 333                                 err = cn_printf(cn, "%d", cprm->cpu);
 334                                 break;
 335                         default:
 336                                 break;
 337                         }
 338                         ++pat_ptr;
 339                 }
 340
 341                 if (err)
 342                         return err;
 343         }
 344
 345 out:
 346         /* Backward compatibility with core_uses_pid:
 347          *
 348          * If core_pattern does not include a %p (as is the default)
 349          * and core_uses_pid is set, then .%pid will be appended to
 350          * the filename. Do not do this for piped commands. */
 351         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 352                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 353                 if (err)
 354                         return err;
 355         }
 356         return ispipe;
 357 }
 358
 359 static int zap_process(struct task_struct *start, int exit_code)
 360 {
 361         struct task_struct *t;
 362         int nr = 0;
 363
 364         /* Allow SIGKILL, see prepare_signal() */
 365         start->signal->flags = SIGNAL_GROUP_EXIT;
 366         start->signal->group_exit_code = exit_code;
 367         start->signal->group_stop_count = 0;
 368
 369         for_each_thread(start, t) {
 370                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 371                 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
 372                         sigaddset(&t->pending.signal, SIGKILL);
 373                         signal_wake_up(t, 1);
 374                         /* The vhost_worker does not particpate in coredumps */
 375                         if ((t->flags & (PF_USER_WORKER | PF_IO_WORKER)) != PF_USER_WORKER)
 376                                 nr++;
 377                 }
 378         }
 379
 380         return nr;
 381 }
 382
 383 static int zap_threads(struct task_struct *tsk,
 384                         struct core_state *core_state, int exit_code)
 385 {
 386         struct signal_struct *signal = tsk->signal;
 387         int nr = -EAGAIN;
 388
 389         spin_lock_irq(&tsk->sighand->siglock);
 390         if (!(signal->flags & SIGNAL_GROUP_EXIT) && !signal->group_exec_task) {
 391                 signal->core_state = core_state;
 392                 nr = zap_process(tsk, exit_code);
 393                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 394                 tsk->flags |= PF_DUMPCORE;
 395                 atomic_set(&core_state->nr_threads, nr);
 396         }
 397         spin_unlock_irq(&tsk->sighand->siglock);
 398         return nr;
 399 }
 400
 401 static int coredump_wait(int exit_code, struct core_state *core_state)
 402 {
 403         struct task_struct *tsk = current;
 404         int core_waiters = -EBUSY;
 405
 406         init_completion(&core_state->startup);
 407         core_state->dumper.task = tsk;
 408         core_state->dumper.next = NULL;
 409
 410         core_waiters = zap_threads(tsk, core_state, exit_code);
 411         if (core_waiters > 0) {
 412                 struct core_thread *ptr;
 413
 414                 wait_for_completion_state(&core_state->startup,
 415                                           TASK_UNINTERRUPTIBLE|TASK_FREEZABLE);
 416                 /*
 417                  * Wait for all the threads to become inactive, so that
 418                  * all the thread context (extended register state, like
 419                  * fpu etc) gets copied to the memory.
 420                  */
 421                 ptr = core_state->dumper.next;
 422                 while (ptr != NULL) {
 423                         wait_task_inactive(ptr->task, TASK_ANY);
 424                         ptr = ptr->next;
 425                 }
 426         }
 427
 428         return core_waiters;
 429 }
 430
 431 static void coredump_finish(bool core_dumped)
 432 {
 433         struct core_thread *curr, *next;
 434         struct task_struct *task;
 435
 436         spin_lock_irq(&current->sighand->siglock);
 437         if (core_dumped && !__fatal_signal_pending(current))
 438                 current->signal->group_exit_code |= 0x80;
 439         next = current->signal->core_state->dumper.next;
 440         current->signal->core_state = NULL;
 441         spin_unlock_irq(&current->sighand->siglock);
 442
 443         while ((curr = next) != NULL) {
 444                 next = curr->next;
 445                 task = curr->task;
 446                 /*
 447                  * see coredump_task_exit(), curr->task must not see
 448                  * ->task == NULL before we read ->next.
 449                  */
 450                 smp_mb();
 451                 curr->task = NULL;
 452                 wake_up_process(task);
 453         }
 454 }
 455
 456 static bool dump_interrupted(void)
 457 {
 458         /*
 459          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 460          * can do try_to_freeze() and check __fatal_signal_pending(),
 461          * but then we need to teach dump_write() to restart and clear
 462          * TIF_SIGPENDING.
 463          */
 464         return fatal_signal_pending(current) || freezing(current);
 465 }
 466
 467 static void wait_for_dump_helpers(struct file *file)
 468 {
 469         struct pipe_inode_info *pipe = file->private_data;
 470
 471         pipe_lock(pipe);
 472         pipe->readers++;
 473         pipe->writers--;
 474         wake_up_interruptible_sync(&pipe->rd_wait);
 475         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 476         pipe_unlock(pipe);
 477
 478         /*
 479          * We actually want wait_event_freezable() but then we need
 480          * to clear TIF_SIGPENDING and improve dump_interrupted().
 481          */
 482         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 483
 484         pipe_lock(pipe);
 485         pipe->readers--;
 486         pipe->writers++;
 487         pipe_unlock(pipe);
 488 }
 489
 490 /*
 491  * umh_pipe_setup
 492  * helper function to customize the process used
 493  * to collect the core in userspace.  Specifically
 494  * it sets up a pipe and installs it as fd 0 (stdin)
 495  * for the process.  Returns 0 on success, or
 496  * PTR_ERR on failure.
 497  * Note that it also sets the core limit to 1.  This
 498  * is a special value that we use to trap recursive
 499  * core dumps
 500  */
 501 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 502 {
 503         struct file *files[2];
 504         struct coredump_params *cp = (struct coredump_params *)info->data;
 505         int err = create_pipe_files(files, 0);
 506         if (err)
 507                 return err;
 508
 509         cp->file = files[1];
 510
 511         err = replace_fd(0, files[0], 0);
 512         fput(files[0]);
 513         /* and disallow core files too */
 514         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 515
 516         return err;
 517 }
 518
 519 void do_coredump(const kernel_siginfo_t *siginfo)
 520 {
 521         struct core_state core_state;
 522         struct core_name cn;
 523         struct mm_struct *mm = current->mm;
 524         struct linux_binfmt * binfmt;
 525         const struct cred *old_cred;
 526         struct cred *cred;
 527         int retval = 0;
 528         int ispipe;
 529         size_t *argv = NULL;
 530         int argc = 0;
 531         /* require nonrelative corefile path and be extra careful */
 532         bool need_suid_safe = false;
 533         bool core_dumped = false;
 534         static atomic_t core_dump_count = ATOMIC_INIT(0);
 535         struct coredump_params cprm = {
 536                 .siginfo = siginfo,
 537                 .limit = rlimit(RLIMIT_CORE),
 538                 /*
 539                  * We must use the same mm->flags while dumping core to avoid
 540                  * inconsistency of bit flags, since this flag is not protected
 541                  * by any locks.
 542                  */
 543                 .mm_flags = mm->flags,
 544                 .vma_meta = NULL,
 545                 .cpu = raw_smp_processor_id(),
 546         };
 547
 548         audit_core_dumps(siginfo->si_signo);
 549
 550         binfmt = mm->binfmt;
 551         if (!binfmt || !binfmt->core_dump)
 552                 goto fail;
 553         if (!__get_dumpable(cprm.mm_flags))
 554                 goto fail;
 555
 556         cred = prepare_creds();
 557         if (!cred)
 558                 goto fail;
 559         /*
 560          * We cannot trust fsuid as being the "true" uid of the process
 561          * nor do we know its entire history. We only know it was tainted
 562          * so we dump it as root in mode 2, and only into a controlled
 563          * environment (pipe handler or fully qualified path).
 564          */
 565         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 566                 /* Setuid core dump mode */
 567                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 568                 need_suid_safe = true;
 569         }
 570
 571         retval = coredump_wait(siginfo->si_signo, &core_state);
 572         if (retval < 0)
 573                 goto fail_creds;
 574
 575         old_cred = override_creds(cred);
 576
 577         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 578
 579         if (ispipe) {
 580                 int argi;
 581                 int dump_count;
 582                 char **helper_argv;
 583                 struct subprocess_info *sub_info;
 584
 585                 if (ispipe < 0) {
 586                         printk(KERN_WARNING "format_corename failed\n");
 587                         printk(KERN_WARNING "Aborting core\n");
 588                         goto fail_unlock;
 589                 }
 590
 591                 if (cprm.limit == 1) {
 592                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 593                          *
 594                          * Normally core limits are irrelevant to pipes, since
 595                          * we're not writing to the file system, but we use
 596                          * cprm.limit of 1 here as a special value, this is a
 597                          * consistent way to catch recursive crashes.
 598                          * We can still crash if the core_pattern binary sets
 599                          * RLIM_CORE = !1, but it runs as root, and can do
 600                          * lots of stupid things.
 601                          *
 602                          * Note that we use task_tgid_vnr here to grab the pid
 603                          * of the process group leader.  That way we get the
 604                          * right pid if a thread in a multi-threaded
 605                          * core_pattern process dies.
 606                          */
 607                         printk(KERN_WARNING
 608                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 609                                 task_tgid_vnr(current), current->comm);
 610                         printk(KERN_WARNING "Aborting core\n");
 611                         goto fail_unlock;
 612                 }
 613                 cprm.limit = RLIM_INFINITY;
 614
 615                 dump_count = atomic_inc_return(&core_dump_count);
 616                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 617                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 618                                task_tgid_vnr(current), current->comm);
 619                         printk(KERN_WARNING "Skipping core dump\n");
 620                         goto fail_dropcount;
 621                 }
 622
 623                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 624                                             GFP_KERNEL);
 625                 if (!helper_argv) {
 626                         printk(KERN_WARNING "%s failed to allocate memory\n",
 627                                __func__);
 628                         goto fail_dropcount;
 629                 }
 630                 for (argi = 0; argi < argc; argi++)
 631                         helper_argv[argi] = cn.corename + argv[argi];
 632                 helper_argv[argi] = NULL;
 633
 634                 retval = -ENOMEM;
 635                 sub_info = call_usermodehelper_setup(helper_argv[0],
 636                                                 helper_argv, NULL, GFP_KERNEL,
 637                                                 umh_pipe_setup, NULL, &cprm);
 638                 if (sub_info)
 639                         retval = call_usermodehelper_exec(sub_info,
 640                                                           UMH_WAIT_EXEC);
 641
 642                 kfree(helper_argv);
 643                 if (retval) {
 644                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 645                                cn.corename);
 646                         goto close_fail;
 647                 }
 648         } else {
 649                 struct mnt_idmap *idmap;
 650                 struct inode *inode;
 651                 int open_flags = O_CREAT | O_WRONLY | O_NOFOLLOW |
 652                                  O_LARGEFILE | O_EXCL;
 653
 654                 if (cprm.limit < binfmt->min_coredump)
 655                         goto fail_unlock;
 656
 657                 if (need_suid_safe && cn.corename[0] != '/') {
 658                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 659                                 "to fully qualified path!\n",
 660                                 task_tgid_vnr(current), current->comm);
 661                         printk(KERN_WARNING "Skipping core dump\n");
 662                         goto fail_unlock;
 663                 }
 664
 665                 /*
 666                  * Unlink the file if it exists unless this is a SUID
 667                  * binary - in that case, we're running around with root
 668                  * privs and don't want to unlink another user's coredump.
 669                  */
 670                 if (!need_suid_safe) {
 671                         /*
 672                          * If it doesn't exist, that's fine. If there's some
 673                          * other problem, we'll catch it at the filp_open().
 674                          */
 675                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 676                 }
 677
 678                 /*
 679                  * There is a race between unlinking and creating the
 680                  * file, but if that causes an EEXIST here, that's
 681                  * fine - another process raced with us while creating
 682                  * the corefile, and the other process won. To userspace,
 683                  * what matters is that at least one of the two processes
 684                  * writes its coredump successfully, not which one.
 685                  */
 686                 if (need_suid_safe) {
 687                         /*
 688                          * Using user namespaces, normal user tasks can change
 689                          * their current->fs->root to point to arbitrary
 690                          * directories. Since the intention of the "only dump
 691                          * with a fully qualified path" rule is to control where
 692                          * coredumps may be placed using root privileges,
 693                          * current->fs->root must not be used. Instead, use the
 694                          * root directory of init_task.
 695                          */
 696                         struct path root;
 697
 698                         task_lock(&init_task);
 699                         get_fs_root(init_task.fs, &root);
 700                         task_unlock(&init_task);
 701                         cprm.file = file_open_root(&root, cn.corename,
 702                                                    open_flags, 0600);
 703                         path_put(&root);
 704                 } else {
 705                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 706                 }
 707                 if (IS_ERR(cprm.file))
 708                         goto fail_unlock;
 709
 710                 inode = file_inode(cprm.file);
 711                 if (inode->i_nlink > 1)
 712                         goto close_fail;
 713                 if (d_unhashed(cprm.file->f_path.dentry))
 714                         goto close_fail;
 715                 /*
 716                  * AK: actually i see no reason to not allow this for named
 717                  * pipes etc, but keep the previous behaviour for now.
 718                  */
 719                 if (!S_ISREG(inode->i_mode))
 720                         goto close_fail;
 721                 /*
 722                  * Don't dump core if the filesystem changed owner or mode
 723                  * of the file during file creation. This is an issue when
 724                  * a process dumps core while its cwd is e.g. on a vfat
 725                  * filesystem.
 726                  */
 727                 idmap = file_mnt_idmap(cprm.file);
 728                 if (!vfsuid_eq_kuid(i_uid_into_vfsuid(idmap, inode),
 729                                     current_fsuid())) {
 730                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file owner\n",
 731                                             cn.corename);
 732                         goto close_fail;
 733                 }
 734                 if ((inode->i_mode & 0677) != 0600) {
 735                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file permissions\n",
 736                                             cn.corename);
 737                         goto close_fail;
 738                 }
 739                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 740                         goto close_fail;
 741                 if (do_truncate(idmap, cprm.file->f_path.dentry,
 742                                 0, 0, cprm.file))
 743                         goto close_fail;
 744         }
 745
 746         /* get us an unshared descriptor table; almost always a no-op */
 747         /* The cell spufs coredump code reads the file descriptor tables */
 748         retval = unshare_files();
 749         if (retval)
 750                 goto close_fail;
 751         if (!dump_interrupted()) {
 752                 /*
 753                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 754                  * have this set to NULL.
 755                  */
 756                 if (!cprm.file) {
 757                         pr_info("Core dump to |%s disabled\n", cn.corename);
 758                         goto close_fail;
 759                 }
 760                 if (!dump_vma_snapshot(&cprm))
 761                         goto close_fail;
 762
 763                 file_start_write(cprm.file);
 764                 core_dumped = binfmt->core_dump(&cprm);
 765                 /*
 766                  * Ensures that file size is big enough to contain the current
 767                  * file postion. This prevents gdb from complaining about
 768                  * a truncated file if the last "write" to the file was
 769                  * dump_skip.
 770                  */
 771                 if (cprm.to_skip) {
 772                         cprm.to_skip--;
 773                         dump_emit(&cprm, "", 1);
 774                 }
 775                 file_end_write(cprm.file);
 776                 free_vma_snapshot(&cprm);
 777         }
 778         if (ispipe && core_pipe_limit)
 779                 wait_for_dump_helpers(cprm.file);
 780 close_fail:
 781         if (cprm.file)
 782                 filp_close(cprm.file, NULL);
 783 fail_dropcount:
 784         if (ispipe)
 785                 atomic_dec(&core_dump_count);
 786 fail_unlock:
 787         kfree(argv);
 788         kfree(cn.corename);
 789         coredump_finish(core_dumped);
 790         revert_creds(old_cred);
 791 fail_creds:
 792         put_cred(cred);
 793 fail:
 794         return;
 795 }
 796
 797 /*
 798  * Core dumping helper functions.  These are the only things you should
 799  * do on a core-file: use only these functions to write out all the
 800  * necessary info.
 801  */
 802 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 803 {
 804         struct file *file = cprm->file;
 805         loff_t pos = file->f_pos;
 806         ssize_t n;
 807         if (cprm->written + nr > cprm->limit)
 808                 return 0;
 809
 810
 811         if (dump_interrupted())
 812                 return 0;
 813         n = __kernel_write(file, addr, nr, &pos);
 814         if (n != nr)
 815                 return 0;
 816         file->f_pos = pos;
 817         cprm->written += n;
 818         cprm->pos += n;
 819
 820         return 1;
 821 }
 822
 823 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 824 {
 825         static char zeroes[PAGE_SIZE];
 826         struct file *file = cprm->file;
 827         if (file->f_mode & FMODE_LSEEK) {
 828                 if (dump_interrupted() ||
 829                     vfs_llseek(file, nr, SEEK_CUR) < 0)
 830                         return 0;
 831                 cprm->pos += nr;
 832                 return 1;
 833         } else {
 834                 while (nr > PAGE_SIZE) {
 835                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 836                                 return 0;
 837                         nr -= PAGE_SIZE;
 838                 }
 839                 return __dump_emit(cprm, zeroes, nr);
 840         }
 841 }
 842
 843 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 844 {
 845         if (cprm->to_skip) {
 846                 if (!__dump_skip(cprm, cprm->to_skip))
 847                         return 0;
 848                 cprm->to_skip = 0;
 849         }
 850         return __dump_emit(cprm, addr, nr);
 851 }
 852 EXPORT_SYMBOL(dump_emit);
 853
 854 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 855 {
 856         cprm->to_skip = pos - cprm->pos;
 857 }
 858 EXPORT_SYMBOL(dump_skip_to);
 859
 860 void dump_skip(struct coredump_params *cprm, size_t nr)
 861 {
 862         cprm->to_skip += nr;
 863 }
 864 EXPORT_SYMBOL(dump_skip);
 865
 866 #ifdef CONFIG_ELF_CORE
 867 static int dump_emit_page(struct coredump_params *cprm, struct page *page)
 868 {
 869         struct bio_vec bvec;
 870         struct iov_iter iter;
 871         struct file *file = cprm->file;
 872         loff_t pos;
 873         ssize_t n;
 874
 875         if (!page)
 876                 return 0;
 877
 878         if (cprm->to_skip) {
 879                 if (!__dump_skip(cprm, cprm->to_skip))
 880                         return 0;
 881                 cprm->to_skip = 0;
 882         }
 883         if (cprm->written + PAGE_SIZE > cprm->limit)
 884                 return 0;
 885         if (dump_interrupted())
 886                 return 0;
 887         pos = file->f_pos;
 888         bvec_set_page(&bvec, page, PAGE_SIZE, 0);
 889         iov_iter_bvec(&iter, ITER_SOURCE, &bvec, 1, PAGE_SIZE);
 890         n = __kernel_write_iter(cprm->file, &iter, &pos);
 891         if (n != PAGE_SIZE)
 892                 return 0;
 893         file->f_pos = pos;
 894         cprm->written += PAGE_SIZE;
 895         cprm->pos += PAGE_SIZE;
 896
 897         return 1;
 898 }
 899
 900 /*
 901  * If we might get machine checks from kernel accesses during the
 902  * core dump, let's get those errors early rather than during the
 903  * IO. This is not performance-critical enough to warrant having
 904  * all the machine check logic in the iovec paths.
 905  */
 906 #ifdef copy_mc_to_kernel
 907
 908 #define dump_page_alloc() alloc_page(GFP_KERNEL)
 909 #define dump_page_free(x) __free_page(x)
 910 static struct page *dump_page_copy(struct page *src, struct page *dst)
 911 {
 912         void *buf = kmap_local_page(src);
 913         size_t left = copy_mc_to_kernel(page_address(dst), buf, PAGE_SIZE);
 914         kunmap_local(buf);
 915         return left ? NULL : dst;
 916 }
 917
 918 #else
 919
 920 /* We just want to return non-NULL; it's never used. */
 921 #define dump_page_alloc() ERR_PTR(-EINVAL)
 922 #define dump_page_free(x) ((void)(x))
 923 static inline struct page *dump_page_copy(struct page *src, struct page *dst)
 924 {
 925         return src;
 926 }
 927 #endif
 928
 929 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 930                     unsigned long len)
 931 {
 932         unsigned long addr;
 933         struct page *dump_page;
 934
 935         dump_page = dump_page_alloc();
 936         if (!dump_page)
 937                 return 0;
 938
 939         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 940                 struct page *page;
 941
 942                 /*
 943                  * To avoid having to allocate page tables for virtual address
 944                  * ranges that have never been used yet, and also to make it
 945                  * easy to generate sparse core files, use a helper that returns
 946                  * NULL when encountering an empty page table entry that would
 947                  * otherwise have been filled with the zero page.
 948                  */
 949                 page = get_dump_page(addr);
 950                 if (page) {
 951                         int stop = !dump_emit_page(cprm, dump_page_copy(page, dump_page));
 952                         put_page(page);
 953                         if (stop) {
 954                                 dump_page_free(dump_page);
 955                                 return 0;
 956                         }
 957                 } else {
 958                         dump_skip(cprm, PAGE_SIZE);
 959                 }
 960         }
 961         dump_page_free(dump_page);
 962         return 1;
 963 }
 964 #endif
 965
 966 int dump_align(struct coredump_params *cprm, int align)
 967 {
 968         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
 969         if (align & (align - 1))
 970                 return 0;
 971         if (mod)
 972                 cprm->to_skip += align - mod;
 973         return 1;
 974 }
 975 EXPORT_SYMBOL(dump_align);
 976
 977 #ifdef CONFIG_SYSCTL
 978
 979 void validate_coredump_safety(void)
 980 {
 981         if (suid_dumpable == SUID_DUMP_ROOT &&
 982             core_pattern[0] != '/' && core_pattern[0] != '|') {
 983                 pr_warn(
 984 "Unsafe core_pattern used with fs.suid_dumpable=2.\n"
 985 "Pipe handler or fully qualified core dump path required.\n"
 986 "Set kernel.core_pattern before fs.suid_dumpable.\n"
 987                 );
 988         }
 989 }
 990
 991 static int proc_dostring_coredump(struct ctl_table *table, int write,
 992                   void *buffer, size_t *lenp, loff_t *ppos)
 993 {
 994         int error = proc_dostring(table, write, buffer, lenp, ppos);
 995
 996         if (!error)
 997                 validate_coredump_safety();
 998         return error;
 999 }
1000
1001 static struct ctl_table coredump_sysctls[] = {
1002         {
1003                 .procname       = "core_uses_pid",
1004                 .data           = &core_uses_pid,
1005                 .maxlen         = sizeof(int),
1006                 .mode           = 0644,
1007                 .proc_handler   = proc_dointvec,
1008         },
1009         {
1010                 .procname       = "core_pattern",
1011                 .data           = core_pattern,
1012                 .maxlen         = CORENAME_MAX_SIZE,
1013                 .mode           = 0644,
1014                 .proc_handler   = proc_dostring_coredump,
1015         },
1016         {
1017                 .procname       = "core_pipe_limit",
1018                 .data           = &core_pipe_limit,
1019                 .maxlen         = sizeof(unsigned int),
1020                 .mode           = 0644,
1021                 .proc_handler   = proc_dointvec,
1022         },
1023 };
1024
1025 static int __init init_fs_coredump_sysctls(void)
1026 {
1027         register_sysctl_init("kernel", coredump_sysctls);
1028         return 0;
1029 }
1030 fs_initcall(init_fs_coredump_sysctls);
1031 #endif /* CONFIG_SYSCTL */
1032
1033 /*
1034  * The purpose of always_dump_vma() is to make sure that special kernel mappings
1035  * that are useful for post-mortem analysis are included in every core dump.
1036  * In that way we ensure that the core dump is fully interpretable later
1037  * without matching up the same kernel and hardware config to see what PC values
1038  * meant. These special mappings include - vDSO, vsyscall, and other
1039  * architecture specific mappings
1040  */
1041 static bool always_dump_vma(struct vm_area_struct *vma)
1042 {
1043         /* Any vsyscall mappings? */
1044         if (vma == get_gate_vma(vma->vm_mm))
1045                 return true;
1046
1047         /*
1048          * Assume that all vmas with a .name op should always be dumped.
1049          * If this changes, a new vm_ops field can easily be added.
1050          */
1051         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
1052                 return true;
1053
1054         /*
1055          * arch_vma_name() returns non-NULL for special architecture mappings,
1056          * such as vDSO sections.
1057          */
1058         if (arch_vma_name(vma))
1059                 return true;
1060
1061         return false;
1062 }
1063
1064 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
1065
1066 /*
1067  * Decide how much of @vma's contents should be included in a core dump.
1068  */
1069 static unsigned long vma_dump_size(struct vm_area_struct *vma,
1070                                    unsigned long mm_flags)
1071 {
1072 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
1073
1074         /* always dump the vdso and vsyscall sections */
1075         if (always_dump_vma(vma))
1076                 goto whole;
1077
1078         if (vma->vm_flags & VM_DONTDUMP)
1079                 return 0;
1080
1081         /* support for DAX */
1082         if (vma_is_dax(vma)) {
1083                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1084                         goto whole;
1085                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1086                         goto whole;
1087                 return 0;
1088         }
1089
1090         /* Hugetlb memory check */
1091         if (is_vm_hugetlb_page(vma)) {
1092                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1093                         goto whole;
1094                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1095                         goto whole;
1096                 return 0;
1097         }
1098
1099         /* Do not dump I/O mapped devices or special mappings */
1100         if (vma->vm_flags & VM_IO)
1101                 return 0;
1102
1103         /* By default, dump shared memory if mapped from an anonymous file. */
1104         if (vma->vm_flags & VM_SHARED) {
1105                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1106                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1107                         goto whole;
1108                 return 0;
1109         }
1110
1111         /* Dump segments that have been written to.  */
1112         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1113                 goto whole;
1114         if (vma->vm_file == NULL)
1115                 return 0;
1116
1117         if (FILTER(MAPPED_PRIVATE))
1118                 goto whole;
1119
1120         /*
1121          * If this is the beginning of an executable file mapping,
1122          * dump the first page to aid in determining what was mapped here.
1123          */
1124         if (FILTER(ELF_HEADERS) &&
1125             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1126                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1127                         return PAGE_SIZE;
1128
1129                 /*
1130                  * ELF libraries aren't always executable.
1131                  * We'll want to check whether the mapping starts with the ELF
1132                  * magic, but not now - we're holding the mmap lock,
1133                  * so copy_from_user() doesn't work here.
1134                  * Use a placeholder instead, and fix it up later in
1135                  * dump_vma_snapshot().
1136                  */
1137                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1138         }
1139
1140 #undef  FILTER
1141
1142         return 0;
1143
1144 whole:
1145         return vma->vm_end - vma->vm_start;
1146 }
1147
1148 /*
1149  * Helper function for iterating across a vma list.  It ensures that the caller
1150  * will visit `gate_vma' prior to terminating the search.
1151  */
1152 static struct vm_area_struct *coredump_next_vma(struct vma_iterator *vmi,
1153                                        struct vm_area_struct *vma,
1154                                        struct vm_area_struct *gate_vma)
1155 {
1156         if (gate_vma && (vma == gate_vma))
1157                 return NULL;
1158
1159         vma = vma_next(vmi);
1160         if (vma)
1161                 return vma;
1162         return gate_vma;
1163 }
1164
1165 static void free_vma_snapshot(struct coredump_params *cprm)
1166 {
1167         if (cprm->vma_meta) {
1168                 int i;
1169                 for (i = 0; i < cprm->vma_count; i++) {
1170                         struct file *file = cprm->vma_meta[i].file;
1171                         if (file)
1172                                 fput(file);
1173                 }
1174                 kvfree(cprm->vma_meta);
1175                 cprm->vma_meta = NULL;
1176         }
1177 }
1178
1179 /*
1180  * Under the mmap_lock, take a snapshot of relevant information about the task's
1181  * VMAs.
1182  */
1183 static bool dump_vma_snapshot(struct coredump_params *cprm)
1184 {
1185         struct vm_area_struct *gate_vma, *vma = NULL;
1186         struct mm_struct *mm = current->mm;
1187         VMA_ITERATOR(vmi, mm, 0);
1188         int i = 0;
1189
1190         /*
1191          * Once the stack expansion code is fixed to not change VMA bounds
1192          * under mmap_lock in read mode, this can be changed to take the
1193          * mmap_lock in read mode.
1194          */
1195         if (mmap_write_lock_killable(mm))
1196                 return false;
1197
1198         cprm->vma_data_size = 0;
1199         gate_vma = get_gate_vma(mm);
1200         cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1201
1202         cprm->vma_meta = kvmalloc_array(cprm->vma_count, sizeof(*cprm->vma_meta), GFP_KERNEL);
1203         if (!cprm->vma_meta) {
1204                 mmap_write_unlock(mm);
1205                 return false;
1206         }
1207
1208         while ((vma = coredump_next_vma(&vmi, vma, gate_vma)) != NULL) {
1209                 struct core_vma_metadata *m = cprm->vma_meta + i;
1210
1211                 m->start = vma->vm_start;
1212                 m->end = vma->vm_end;
1213                 m->flags = vma->vm_flags;
1214                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1215                 m->pgoff = vma->vm_pgoff;
1216                 m->file = vma->vm_file;
1217                 if (m->file)
1218                         get_file(m->file);
1219                 i++;
1220         }
1221
1222         mmap_write_unlock(mm);
1223
1224         for (i = 0; i < cprm->vma_count; i++) {
1225                 struct core_vma_metadata *m = cprm->vma_meta + i;
1226
1227                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1228                         char elfmag[SELFMAG];
1229
1230                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1231                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1232                                 m->dump_size = 0;
1233                         } else {
1234                                 m->dump_size = PAGE_SIZE;
1235                         }
1236                 }
1237
1238                 cprm->vma_data_size += m->dump_size;
1239         }
1240
1241         return true;
1242 }