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/sort.h>
  22 #include <linux/sched/coredump.h>
  23 #include <linux/sched/signal.h>
  24 #include <linux/sched/task_stack.h>
  25 #include <linux/utsname.h>
  26 #include <linux/pid_namespace.h>
  27 #include <linux/module.h>
  28 #include <linux/namei.h>
  29 #include <linux/mount.h>
  30 #include <linux/security.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/tsacct_kern.h>
  33 #include <linux/cn_proc.h>
  34 #include <linux/audit.h>
  35 #include <linux/kmod.h>
  36 #include <linux/fsnotify.h>
  37 #include <linux/fs_struct.h>
  38 #include <linux/pipe_fs_i.h>
  39 #include <linux/oom.h>
  40 #include <linux/compat.h>
  41 #include <linux/fs.h>
  42 #include <linux/path.h>
  43 #include <linux/timekeeping.h>
  44 #include <linux/sysctl.h>
  45 #include <linux/elf.h>
  46
  47 #include <linux/uaccess.h>
  48 #include <asm/mmu_context.h>
  49 #include <asm/tlb.h>
  50 #include <asm/exec.h>
  51
  52 #include <trace/events/task.h>
  53 #include "internal.h"
  54
  55 #include <trace/events/sched.h>
  56
  57 static bool dump_vma_snapshot(struct coredump_params *cprm);
  58 static void free_vma_snapshot(struct coredump_params *cprm);
  59
  60 #define CORE_FILE_NOTE_SIZE_DEFAULT (4*1024*1024)
  61 /* Define a reasonable max cap */
  62 #define CORE_FILE_NOTE_SIZE_MAX (16*1024*1024)
  63
  64 static int core_uses_pid;
  65 static unsigned int core_pipe_limit;
  66 static char core_pattern[CORENAME_MAX_SIZE] = "core";
  67 static int core_name_size = CORENAME_MAX_SIZE;
  68 unsigned int core_file_note_size_limit = CORE_FILE_NOTE_SIZE_DEFAULT;
  69
  70 struct core_name {
  71         char *corename;
  72         int used, size;
  73 };
  74
  75 static int expand_corename(struct core_name *cn, int size)
  76 {
  77         char *corename;
  78
  79         size = kmalloc_size_roundup(size);
  80         corename = krealloc(cn->corename, size, GFP_KERNEL);
  81
  82         if (!corename)
  83                 return -ENOMEM;
  84
  85         if (size > core_name_size) /* racy but harmless */
  86                 core_name_size = size;
  87
  88         cn->size = size;
  89         cn->corename = corename;
  90         return 0;
  91 }
  92
  93 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  94                                      va_list arg)
  95 {
  96         int free, need;
  97         va_list arg_copy;
  98
  99 again:
 100         free = cn->size - cn->used;
 101
 102         va_copy(arg_copy, arg);
 103         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
 104         va_end(arg_copy);
 105
 106         if (need < free) {
 107                 cn->used += need;
 108                 return 0;
 109         }
 110
 111         if (!expand_corename(cn, cn->size + need - free + 1))
 112                 goto again;
 113
 114         return -ENOMEM;
 115 }
 116
 117 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 118 {
 119         va_list arg;
 120         int ret;
 121
 122         va_start(arg, fmt);
 123         ret = cn_vprintf(cn, fmt, arg);
 124         va_end(arg);
 125
 126         return ret;
 127 }
 128
 129 static __printf(2, 3)
 130 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 131 {
 132         int cur = cn->used;
 133         va_list arg;
 134         int ret;
 135
 136         va_start(arg, fmt);
 137         ret = cn_vprintf(cn, fmt, arg);
 138         va_end(arg);
 139
 140         if (ret == 0) {
 141                 /*
 142                  * Ensure that this coredump name component can't cause the
 143                  * resulting corefile path to consist of a ".." or ".".
 144                  */
 145                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 146                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 147                                 && cn->corename[cur+1] == '.'))
 148                         cn->corename[cur] = '!';
 149
 150                 /*
 151                  * Empty names are fishy and could be used to create a "//" in a
 152                  * corefile name, causing the coredump to happen one directory
 153                  * level too high. Enforce that all components of the core
 154                  * pattern are at least one character long.
 155                  */
 156                 if (cn->used == cur)
 157                         ret = cn_printf(cn, "!");
 158         }
 159
 160         for (; cur < cn->used; ++cur) {
 161                 if (cn->corename[cur] == '/')
 162                         cn->corename[cur] = '!';
 163         }
 164         return ret;
 165 }
 166
 167 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 168 {
 169         struct file *exe_file;
 170         char *pathbuf, *path, *ptr;
 171         int ret;
 172
 173         exe_file = get_mm_exe_file(current->mm);
 174         if (!exe_file)
 175                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 176
 177         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 178         if (!pathbuf) {
 179                 ret = -ENOMEM;
 180                 goto put_exe_file;
 181         }
 182
 183         path = file_path(exe_file, pathbuf, PATH_MAX);
 184         if (IS_ERR(path)) {
 185                 ret = PTR_ERR(path);
 186                 goto free_buf;
 187         }
 188
 189         if (name_only) {
 190                 ptr = strrchr(path, '/');
 191                 if (ptr)
 192                         path = ptr + 1;
 193         }
 194         ret = cn_esc_printf(cn, "%s", path);
 195
 196 free_buf:
 197         kfree(pathbuf);
 198 put_exe_file:
 199         fput(exe_file);
 200         return ret;
 201 }
 202
 203 /* format_corename will inspect the pattern parameter, and output a
 204  * name into corename, which must have space for at least
 205  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 206  */
 207 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 208                            size_t **argv, int *argc)
 209 {
 210         const struct cred *cred = current_cred();
 211         const char *pat_ptr = core_pattern;
 212         int ispipe = (*pat_ptr == '|');
 213         bool was_space = false;
 214         int pid_in_pattern = 0;
 215         int err = 0;
 216
 217         cn->used = 0;
 218         cn->corename = NULL;
 219         if (expand_corename(cn, core_name_size))
 220                 return -ENOMEM;
 221         cn->corename[0] = '\0';
 222
 223         if (ispipe) {
 224                 int argvs = sizeof(core_pattern) / 2;
 225                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 226                 if (!(*argv))
 227                         return -ENOMEM;
 228                 (*argv)[(*argc)++] = 0;
 229                 ++pat_ptr;
 230                 if (!(*pat_ptr))
 231                         return -ENOMEM;
 232         }
 233
 234         /* Repeat as long as we have more pattern to process and more output
 235            space */
 236         while (*pat_ptr) {
 237                 /*
 238                  * Split on spaces before doing template expansion so that
 239                  * %e and %E don't get split if they have spaces in them
 240                  */
 241                 if (ispipe) {
 242                         if (isspace(*pat_ptr)) {
 243                                 if (cn->used != 0)
 244                                         was_space = true;
 245                                 pat_ptr++;
 246                                 continue;
 247                         } else if (was_space) {
 248                                 was_space = false;
 249                                 err = cn_printf(cn, "%c", '\0');
 250                                 if (err)
 251                                         return err;
 252                                 (*argv)[(*argc)++] = cn->used;
 253                         }
 254                 }
 255                 if (*pat_ptr != '%') {
 256                         err = cn_printf(cn, "%c", *pat_ptr++);
 257                 } else {
 258                         switch (*++pat_ptr) {
 259                         /* single % at the end, drop that */
 260                         case 0:
 261                                 goto out;
 262                         /* Double percent, output one percent */
 263                         case '%':
 264                                 err = cn_printf(cn, "%c", '%');
 265                                 break;
 266                         /* pid */
 267                         case 'p':
 268                                 pid_in_pattern = 1;
 269                                 err = cn_printf(cn, "%d",
 270                                               task_tgid_vnr(current));
 271                                 break;
 272                         /* global pid */
 273                         case 'P':
 274                                 err = cn_printf(cn, "%d",
 275                                               task_tgid_nr(current));
 276                                 break;
 277                         case 'i':
 278                                 err = cn_printf(cn, "%d",
 279                                               task_pid_vnr(current));
 280                                 break;
 281                         case 'I':
 282                                 err = cn_printf(cn, "%d",
 283                                               task_pid_nr(current));
 284                                 break;
 285                         /* uid */
 286                         case 'u':
 287                                 err = cn_printf(cn, "%u",
 288                                                 from_kuid(&init_user_ns,
 289                                                           cred->uid));
 290                                 break;
 291                         /* gid */
 292                         case 'g':
 293                                 err = cn_printf(cn, "%u",
 294                                                 from_kgid(&init_user_ns,
 295                                                           cred->gid));
 296                                 break;
 297                         case 'd':
 298                                 err = cn_printf(cn, "%d",
 299                                         __get_dumpable(cprm->mm_flags));
 300                                 break;
 301                         /* signal that caused the coredump */
 302                         case 's':
 303                                 err = cn_printf(cn, "%d",
 304                                                 cprm->siginfo->si_signo);
 305                                 break;
 306                         /* UNIX time of coredump */
 307                         case 't': {
 308                                 time64_t time;
 309
 310                                 time = ktime_get_real_seconds();
 311                                 err = cn_printf(cn, "%lld", time);
 312                                 break;
 313                         }
 314                         /* hostname */
 315                         case 'h':
 316                                 down_read(&uts_sem);
 317                                 err = cn_esc_printf(cn, "%s",
 318                                               utsname()->nodename);
 319                                 up_read(&uts_sem);
 320                                 break;
 321                         /* executable, could be changed by prctl PR_SET_NAME etc */
 322                         case 'e':
 323                                 err = cn_esc_printf(cn, "%s", current->comm);
 324                                 break;
 325                         /* file name of executable */
 326                         case 'f':
 327                                 err = cn_print_exe_file(cn, true);
 328                                 break;
 329                         case 'E':
 330                                 err = cn_print_exe_file(cn, false);
 331                                 break;
 332                         /* core limit size */
 333                         case 'c':
 334                                 err = cn_printf(cn, "%lu",
 335                                               rlimit(RLIMIT_CORE));
 336                                 break;
 337                         /* CPU the task ran on */
 338                         case 'C':
 339                                 err = cn_printf(cn, "%d", cprm->cpu);
 340                                 break;
 341                         default:
 342                                 break;
 343                         }
 344                         ++pat_ptr;
 345                 }
 346
 347                 if (err)
 348                         return err;
 349         }
 350
 351 out:
 352         /* Backward compatibility with core_uses_pid:
 353          *
 354          * If core_pattern does not include a %p (as is the default)
 355          * and core_uses_pid is set, then .%pid will be appended to
 356          * the filename. Do not do this for piped commands. */
 357         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 358                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 359                 if (err)
 360                         return err;
 361         }
 362         return ispipe;
 363 }
 364
 365 static int zap_process(struct signal_struct *signal, int exit_code)
 366 {
 367         struct task_struct *t;
 368         int nr = 0;
 369
 370         signal->flags = SIGNAL_GROUP_EXIT;
 371         signal->group_exit_code = exit_code;
 372         signal->group_stop_count = 0;
 373
 374         __for_each_thread(signal, t) {
 375                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 376                 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
 377                         sigaddset(&t->pending.signal, SIGKILL);
 378                         signal_wake_up(t, 1);
 379                         nr++;
 380                 }
 381         }
 382
 383         return nr;
 384 }
 385
 386 static int zap_threads(struct task_struct *tsk,
 387                         struct core_state *core_state, int exit_code)
 388 {
 389         struct signal_struct *signal = tsk->signal;
 390         int nr = -EAGAIN;
 391
 392         spin_lock_irq(&tsk->sighand->siglock);
 393         if (!(signal->flags & SIGNAL_GROUP_EXIT) && !signal->group_exec_task) {
 394                 /* Allow SIGKILL, see prepare_signal() */
 395                 signal->core_state = core_state;
 396                 nr = zap_process(signal, exit_code);
 397                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 398                 tsk->flags |= PF_DUMPCORE;
 399                 atomic_set(&core_state->nr_threads, nr);
 400         }
 401         spin_unlock_irq(&tsk->sighand->siglock);
 402         return nr;
 403 }
 404
 405 static int coredump_wait(int exit_code, struct core_state *core_state)
 406 {
 407         struct task_struct *tsk = current;
 408         int core_waiters = -EBUSY;
 409
 410         init_completion(&core_state->startup);
 411         core_state->dumper.task = tsk;
 412         core_state->dumper.next = NULL;
 413
 414         core_waiters = zap_threads(tsk, core_state, exit_code);
 415         if (core_waiters > 0) {
 416                 struct core_thread *ptr;
 417
 418                 wait_for_completion_state(&core_state->startup,
 419                                           TASK_UNINTERRUPTIBLE|TASK_FREEZABLE);
 420                 /*
 421                  * Wait for all the threads to become inactive, so that
 422                  * all the thread context (extended register state, like
 423                  * fpu etc) gets copied to the memory.
 424                  */
 425                 ptr = core_state->dumper.next;
 426                 while (ptr != NULL) {
 427                         wait_task_inactive(ptr->task, TASK_ANY);
 428                         ptr = ptr->next;
 429                 }
 430         }
 431
 432         return core_waiters;
 433 }
 434
 435 static void coredump_finish(bool core_dumped)
 436 {
 437         struct core_thread *curr, *next;
 438         struct task_struct *task;
 439
 440         spin_lock_irq(&current->sighand->siglock);
 441         if (core_dumped && !__fatal_signal_pending(current))
 442                 current->signal->group_exit_code |= 0x80;
 443         next = current->signal->core_state->dumper.next;
 444         current->signal->core_state = NULL;
 445         spin_unlock_irq(&current->sighand->siglock);
 446
 447         while ((curr = next) != NULL) {
 448                 next = curr->next;
 449                 task = curr->task;
 450                 /*
 451                  * see coredump_task_exit(), curr->task must not see
 452                  * ->task == NULL before we read ->next.
 453                  */
 454                 smp_mb();
 455                 curr->task = NULL;
 456                 wake_up_process(task);
 457         }
 458 }
 459
 460 static bool dump_interrupted(void)
 461 {
 462         /*
 463          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 464          * can do try_to_freeze() and check __fatal_signal_pending(),
 465          * but then we need to teach dump_write() to restart and clear
 466          * TIF_SIGPENDING.
 467          */
 468         if (fatal_signal_pending(current)) {
 469                 coredump_report_failure("interrupted: fatal signal pending");
 470                 return true;
 471         }
 472
 473         if (freezing(current)) {
 474                 coredump_report_failure("interrupted: freezing");
 475                 return true;
 476         }
 477
 478         return false;
 479 }
 480
 481 static void wait_for_dump_helpers(struct file *file)
 482 {
 483         struct pipe_inode_info *pipe = file->private_data;
 484
 485         pipe_lock(pipe);
 486         pipe->readers++;
 487         pipe->writers--;
 488         wake_up_interruptible_sync(&pipe->rd_wait);
 489         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 490         pipe_unlock(pipe);
 491
 492         /*
 493          * We actually want wait_event_freezable() but then we need
 494          * to clear TIF_SIGPENDING and improve dump_interrupted().
 495          */
 496         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 497
 498         pipe_lock(pipe);
 499         pipe->readers--;
 500         pipe->writers++;
 501         pipe_unlock(pipe);
 502 }
 503
 504 /*
 505  * umh_pipe_setup
 506  * helper function to customize the process used
 507  * to collect the core in userspace.  Specifically
 508  * it sets up a pipe and installs it as fd 0 (stdin)
 509  * for the process.  Returns 0 on success, or
 510  * PTR_ERR on failure.
 511  * Note that it also sets the core limit to 1.  This
 512  * is a special value that we use to trap recursive
 513  * core dumps
 514  */
 515 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 516 {
 517         struct file *files[2];
 518         struct coredump_params *cp = (struct coredump_params *)info->data;
 519         int err = create_pipe_files(files, 0);
 520         if (err)
 521                 return err;
 522
 523         cp->file = files[1];
 524
 525         err = replace_fd(0, files[0], 0);
 526         fput(files[0]);
 527         /* and disallow core files too */
 528         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 529
 530         return err;
 531 }
 532
 533 int do_coredump(const kernel_siginfo_t *siginfo)
 534 {
 535         struct core_state core_state;
 536         struct core_name cn;
 537         struct mm_struct *mm = current->mm;
 538         struct linux_binfmt * binfmt;
 539         const struct cred *old_cred;
 540         struct cred *cred;
 541         int retval;
 542         int ispipe;
 543         size_t *argv = NULL;
 544         int argc = 0;
 545         /* require nonrelative corefile path and be extra careful */
 546         bool need_suid_safe = false;
 547         bool core_dumped = false;
 548         static atomic_t core_dump_count = ATOMIC_INIT(0);
 549         struct coredump_params cprm = {
 550                 .siginfo = siginfo,
 551                 .limit = rlimit(RLIMIT_CORE),
 552                 /*
 553                  * We must use the same mm->flags while dumping core to avoid
 554                  * inconsistency of bit flags, since this flag is not protected
 555                  * by any locks.
 556                  */
 557                 .mm_flags = mm->flags,
 558                 .vma_meta = NULL,
 559                 .cpu = raw_smp_processor_id(),
 560         };
 561
 562         audit_core_dumps(siginfo->si_signo);
 563
 564         binfmt = mm->binfmt;
 565         if (!binfmt || !binfmt->core_dump) {
 566                 retval = -ENOEXEC;
 567                 goto fail;
 568         }
 569         if (!__get_dumpable(cprm.mm_flags)) {
 570                 retval = -EACCES;
 571                 goto fail;
 572         }
 573
 574         cred = prepare_creds();
 575         if (!cred) {
 576                 retval = -EPERM;
 577                 goto fail;
 578         }
 579         /*
 580          * We cannot trust fsuid as being the "true" uid of the process
 581          * nor do we know its entire history. We only know it was tainted
 582          * so we dump it as root in mode 2, and only into a controlled
 583          * environment (pipe handler or fully qualified path).
 584          */
 585         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 586                 /* Setuid core dump mode */
 587                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 588                 need_suid_safe = true;
 589         }
 590
 591         retval = coredump_wait(siginfo->si_signo, &core_state);
 592         if (retval < 0)
 593                 goto fail_creds;
 594
 595         old_cred = override_creds(cred);
 596
 597         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 598
 599         if (ispipe) {
 600                 int argi;
 601                 int dump_count;
 602                 char **helper_argv;
 603                 struct subprocess_info *sub_info;
 604
 605                 if (ispipe < 0) {
 606                         coredump_report_failure("format_corename failed, aborting core");
 607                         retval = ispipe;
 608                         goto fail_unlock;
 609                 }
 610
 611                 if (cprm.limit == 1) {
 612                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 613                          *
 614                          * Normally core limits are irrelevant to pipes, since
 615                          * we're not writing to the file system, but we use
 616                          * cprm.limit of 1 here as a special value, this is a
 617                          * consistent way to catch recursive crashes.
 618                          * We can still crash if the core_pattern binary sets
 619                          * RLIM_CORE = !1, but it runs as root, and can do
 620                          * lots of stupid things.
 621                          *
 622                          * Note that we use task_tgid_vnr here to grab the pid
 623                          * of the process group leader.  That way we get the
 624                          * right pid if a thread in a multi-threaded
 625                          * core_pattern process dies.
 626                          */
 627                         coredump_report_failure("RLIMIT_CORE is set to 1, aborting core");
 628                         retval = -EPERM;
 629                         goto fail_unlock;
 630                 }
 631                 cprm.limit = RLIM_INFINITY;
 632
 633                 dump_count = atomic_inc_return(&core_dump_count);
 634                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 635                         coredump_report_failure("over core_pipe_limit, skipping core dump");
 636                         retval = -E2BIG;
 637                         goto fail_dropcount;
 638                 }
 639
 640                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 641                                             GFP_KERNEL);
 642                 if (!helper_argv) {
 643                         coredump_report_failure("%s failed to allocate memory", __func__);
 644                         retval = -ENOMEM;
 645                         goto fail_dropcount;
 646                 }
 647                 for (argi = 0; argi < argc; argi++)
 648                         helper_argv[argi] = cn.corename + argv[argi];
 649                 helper_argv[argi] = NULL;
 650
 651                 retval = -ENOMEM;
 652                 sub_info = call_usermodehelper_setup(helper_argv[0],
 653                                                 helper_argv, NULL, GFP_KERNEL,
 654                                                 umh_pipe_setup, NULL, &cprm);
 655                 if (sub_info)
 656                         retval = call_usermodehelper_exec(sub_info,
 657                                                           UMH_WAIT_EXEC);
 658
 659                 kfree(helper_argv);
 660                 if (retval) {
 661                         coredump_report_failure("|%s pipe failed", cn.corename);
 662                         goto close_fail;
 663                 }
 664         } else {
 665                 struct mnt_idmap *idmap;
 666                 struct inode *inode;
 667                 int open_flags = O_CREAT | O_WRONLY | O_NOFOLLOW |
 668                                  O_LARGEFILE | O_EXCL;
 669
 670                 if (cprm.limit < binfmt->min_coredump) {
 671                         coredump_report_failure("over coredump resource limit, skipping core dump");
 672                         retval = -E2BIG;
 673                         goto fail_unlock;
 674                 }
 675
 676                 if (need_suid_safe && cn.corename[0] != '/') {
 677                         coredump_report_failure(
 678                                 "this process can only dump core to a fully qualified path, skipping core dump");
 679                         retval = -EPERM;
 680                         goto fail_unlock;
 681                 }
 682
 683                 /*
 684                  * Unlink the file if it exists unless this is a SUID
 685                  * binary - in that case, we're running around with root
 686                  * privs and don't want to unlink another user's coredump.
 687                  */
 688                 if (!need_suid_safe) {
 689                         /*
 690                          * If it doesn't exist, that's fine. If there's some
 691                          * other problem, we'll catch it at the filp_open().
 692                          */
 693                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 694                 }
 695
 696                 /*
 697                  * There is a race between unlinking and creating the
 698                  * file, but if that causes an EEXIST here, that's
 699                  * fine - another process raced with us while creating
 700                  * the corefile, and the other process won. To userspace,
 701                  * what matters is that at least one of the two processes
 702                  * writes its coredump successfully, not which one.
 703                  */
 704                 if (need_suid_safe) {
 705                         /*
 706                          * Using user namespaces, normal user tasks can change
 707                          * their current->fs->root to point to arbitrary
 708                          * directories. Since the intention of the "only dump
 709                          * with a fully qualified path" rule is to control where
 710                          * coredumps may be placed using root privileges,
 711                          * current->fs->root must not be used. Instead, use the
 712                          * root directory of init_task.
 713                          */
 714                         struct path root;
 715
 716                         task_lock(&init_task);
 717                         get_fs_root(init_task.fs, &root);
 718                         task_unlock(&init_task);
 719                         cprm.file = file_open_root(&root, cn.corename,
 720                                                    open_flags, 0600);
 721                         path_put(&root);
 722                 } else {
 723                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 724                 }
 725                 if (IS_ERR(cprm.file)) {
 726                         retval = PTR_ERR(cprm.file);
 727                         goto fail_unlock;
 728                 }
 729
 730                 inode = file_inode(cprm.file);
 731                 if (inode->i_nlink > 1) {
 732                         retval = -EMLINK;
 733                         goto close_fail;
 734                 }
 735                 if (d_unhashed(cprm.file->f_path.dentry)) {
 736                         retval = -EEXIST;
 737                         goto close_fail;
 738                 }
 739                 /*
 740                  * AK: actually i see no reason to not allow this for named
 741                  * pipes etc, but keep the previous behaviour for now.
 742                  */
 743                 if (!S_ISREG(inode->i_mode)) {
 744                         retval = -EISDIR;
 745                         goto close_fail;
 746                 }
 747                 /*
 748                  * Don't dump core if the filesystem changed owner or mode
 749                  * of the file during file creation. This is an issue when
 750                  * a process dumps core while its cwd is e.g. on a vfat
 751                  * filesystem.
 752                  */
 753                 idmap = file_mnt_idmap(cprm.file);
 754                 if (!vfsuid_eq_kuid(i_uid_into_vfsuid(idmap, inode),
 755                                     current_fsuid())) {
 756                         coredump_report_failure("Core dump to %s aborted: "
 757                                 "cannot preserve file owner", cn.corename);
 758                         retval = -EPERM;
 759                         goto close_fail;
 760                 }
 761                 if ((inode->i_mode & 0677) != 0600) {
 762                         coredump_report_failure("Core dump to %s aborted: "
 763                                 "cannot preserve file permissions", cn.corename);
 764                         retval = -EPERM;
 765                         goto close_fail;
 766                 }
 767                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE)) {
 768                         retval = -EACCES;
 769                         goto close_fail;
 770                 }
 771                 retval = do_truncate(idmap, cprm.file->f_path.dentry,
 772                                 0, 0, cprm.file);
 773                 if (retval)
 774                         goto close_fail;
 775         }
 776
 777         /* get us an unshared descriptor table; almost always a no-op */
 778         /* The cell spufs coredump code reads the file descriptor tables */
 779         retval = unshare_files();
 780         if (retval)
 781                 goto close_fail;
 782         if (!dump_interrupted()) {
 783                 /*
 784                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 785                  * have this set to NULL.
 786                  */
 787                 if (!cprm.file) {
 788                         coredump_report_failure("Core dump to |%s disabled", cn.corename);
 789                         retval = -EPERM;
 790                         goto close_fail;
 791                 }
 792                 if (!dump_vma_snapshot(&cprm)) {
 793                         coredump_report_failure("Can't get VMA snapshot for core dump |%s",
 794                                 cn.corename);
 795                         retval = -EACCES;
 796                         goto close_fail;
 797                 }
 798
 799                 file_start_write(cprm.file);
 800                 core_dumped = binfmt->core_dump(&cprm);
 801                 /*
 802                  * Ensures that file size is big enough to contain the current
 803                  * file postion. This prevents gdb from complaining about
 804                  * a truncated file if the last "write" to the file was
 805                  * dump_skip.
 806                  */
 807                 if (cprm.to_skip) {
 808                         cprm.to_skip--;
 809                         dump_emit(&cprm, "", 1);
 810                 }
 811                 file_end_write(cprm.file);
 812                 free_vma_snapshot(&cprm);
 813         } else {
 814                 coredump_report_failure("Core dump to %s%s has been interrupted",
 815                         ispipe ? "|" : "", cn.corename);
 816                 retval = -EAGAIN;
 817                 goto fail;
 818         }
 819         coredump_report(
 820                 "written to %s%s: VMAs: %d, size %zu; core: %lld bytes, pos %lld",
 821                 ispipe ? "|" : "", cn.corename,
 822                 cprm.vma_count, cprm.vma_data_size, cprm.written, cprm.pos);
 823         if (ispipe && core_pipe_limit)
 824                 wait_for_dump_helpers(cprm.file);
 825
 826         retval = 0;
 827
 828 close_fail:
 829         if (cprm.file)
 830                 filp_close(cprm.file, NULL);
 831 fail_dropcount:
 832         if (ispipe)
 833                 atomic_dec(&core_dump_count);
 834 fail_unlock:
 835         kfree(argv);
 836         kfree(cn.corename);
 837         coredump_finish(core_dumped);
 838         revert_creds(old_cred);
 839 fail_creds:
 840         put_cred(cred);
 841 fail:
 842         return retval;
 843 }
 844
 845 /*
 846  * Core dumping helper functions.  These are the only things you should
 847  * do on a core-file: use only these functions to write out all the
 848  * necessary info.
 849  */
 850 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 851 {
 852         struct file *file = cprm->file;
 853         loff_t pos = file->f_pos;
 854         ssize_t n;
 855         if (cprm->written + nr > cprm->limit)
 856                 return 0;
 857
 858
 859         if (dump_interrupted())
 860                 return 0;
 861         n = __kernel_write(file, addr, nr, &pos);
 862         if (n != nr) {
 863                 if (n < 0)
 864                         coredump_report_failure("failed when writing out, error %zd", n);
 865                 else
 866                         coredump_report_failure(
 867                                 "partially written out, only %zd(of %d) bytes written",
 868                                 n, nr);
 869
 870                 return 0;
 871         }
 872         file->f_pos = pos;
 873         cprm->written += n;
 874         cprm->pos += n;
 875
 876         return 1;
 877 }
 878
 879 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 880 {
 881         static char zeroes[PAGE_SIZE];
 882         struct file *file = cprm->file;
 883         if (file->f_mode & FMODE_LSEEK) {
 884                 int ret;
 885
 886                 if (dump_interrupted())
 887                         return 0;
 888
 889                 ret = vfs_llseek(file, nr, SEEK_CUR);
 890                 if (ret < 0) {
 891                         coredump_report_failure("failed when seeking, error %d", ret);
 892                         return 0;
 893                 }
 894                 cprm->pos += nr;
 895                 return 1;
 896         } else {
 897                 while (nr > PAGE_SIZE) {
 898                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 899                                 return 0;
 900                         nr -= PAGE_SIZE;
 901                 }
 902                 return __dump_emit(cprm, zeroes, nr);
 903         }
 904 }
 905
 906 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 907 {
 908         if (cprm->to_skip) {
 909                 if (!__dump_skip(cprm, cprm->to_skip))
 910                         return 0;
 911                 cprm->to_skip = 0;
 912         }
 913         return __dump_emit(cprm, addr, nr);
 914 }
 915 EXPORT_SYMBOL(dump_emit);
 916
 917 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 918 {
 919         cprm->to_skip = pos - cprm->pos;
 920 }
 921 EXPORT_SYMBOL(dump_skip_to);
 922
 923 void dump_skip(struct coredump_params *cprm, size_t nr)
 924 {
 925         cprm->to_skip += nr;
 926 }
 927 EXPORT_SYMBOL(dump_skip);
 928
 929 #ifdef CONFIG_ELF_CORE
 930 static int dump_emit_page(struct coredump_params *cprm, struct page *page)
 931 {
 932         struct bio_vec bvec;
 933         struct iov_iter iter;
 934         struct file *file = cprm->file;
 935         loff_t pos;
 936         ssize_t n;
 937
 938         if (!page)
 939                 return 0;
 940
 941         if (cprm->to_skip) {
 942                 if (!__dump_skip(cprm, cprm->to_skip))
 943                         return 0;
 944                 cprm->to_skip = 0;
 945         }
 946         if (cprm->written + PAGE_SIZE > cprm->limit)
 947                 return 0;
 948         if (dump_interrupted())
 949                 return 0;
 950         pos = file->f_pos;
 951         bvec_set_page(&bvec, page, PAGE_SIZE, 0);
 952         iov_iter_bvec(&iter, ITER_SOURCE, &bvec, 1, PAGE_SIZE);
 953         n = __kernel_write_iter(cprm->file, &iter, &pos);
 954         if (n != PAGE_SIZE)
 955                 return 0;
 956         file->f_pos = pos;
 957         cprm->written += PAGE_SIZE;
 958         cprm->pos += PAGE_SIZE;
 959
 960         return 1;
 961 }
 962
 963 /*
 964  * If we might get machine checks from kernel accesses during the
 965  * core dump, let's get those errors early rather than during the
 966  * IO. This is not performance-critical enough to warrant having
 967  * all the machine check logic in the iovec paths.
 968  */
 969 #ifdef copy_mc_to_kernel
 970
 971 #define dump_page_alloc() alloc_page(GFP_KERNEL)
 972 #define dump_page_free(x) __free_page(x)
 973 static struct page *dump_page_copy(struct page *src, struct page *dst)
 974 {
 975         void *buf = kmap_local_page(src);
 976         size_t left = copy_mc_to_kernel(page_address(dst), buf, PAGE_SIZE);
 977         kunmap_local(buf);
 978         return left ? NULL : dst;
 979 }
 980
 981 #else
 982
 983 /* We just want to return non-NULL; it's never used. */
 984 #define dump_page_alloc() ERR_PTR(-EINVAL)
 985 #define dump_page_free(x) ((void)(x))
 986 static inline struct page *dump_page_copy(struct page *src, struct page *dst)
 987 {
 988         return src;
 989 }
 990 #endif
 991
 992 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 993                     unsigned long len)
 994 {
 995         unsigned long addr;
 996         struct page *dump_page;
 997
 998         dump_page = dump_page_alloc();
 999         if (!dump_page)
1000                 return 0;
1001
1002         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
1003                 struct page *page;
1004
1005                 /*
1006                  * To avoid having to allocate page tables for virtual address
1007                  * ranges that have never been used yet, and also to make it
1008                  * easy to generate sparse core files, use a helper that returns
1009                  * NULL when encountering an empty page table entry that would
1010                  * otherwise have been filled with the zero page.
1011                  */
1012                 page = get_dump_page(addr);
1013                 if (page) {
1014                         int stop = !dump_emit_page(cprm, dump_page_copy(page, dump_page));
1015                         put_page(page);
1016                         if (stop) {
1017                                 dump_page_free(dump_page);
1018                                 return 0;
1019                         }
1020                 } else {
1021                         dump_skip(cprm, PAGE_SIZE);
1022                 }
1023         }
1024         dump_page_free(dump_page);
1025         return 1;
1026 }
1027 #endif
1028
1029 int dump_align(struct coredump_params *cprm, int align)
1030 {
1031         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
1032         if (align & (align - 1))
1033                 return 0;
1034         if (mod)
1035                 cprm->to_skip += align - mod;
1036         return 1;
1037 }
1038 EXPORT_SYMBOL(dump_align);
1039
1040 #ifdef CONFIG_SYSCTL
1041
1042 void validate_coredump_safety(void)
1043 {
1044         if (suid_dumpable == SUID_DUMP_ROOT &&
1045             core_pattern[0] != '/' && core_pattern[0] != '|') {
1046
1047                 coredump_report_failure("Unsafe core_pattern used with fs.suid_dumpable=2: "
1048                         "pipe handler or fully qualified core dump path required. "
1049                         "Set kernel.core_pattern before fs.suid_dumpable.");
1050         }
1051 }
1052
1053 static int proc_dostring_coredump(const struct ctl_table *table, int write,
1054                   void *buffer, size_t *lenp, loff_t *ppos)
1055 {
1056         int error = proc_dostring(table, write, buffer, lenp, ppos);
1057
1058         if (!error)
1059                 validate_coredump_safety();
1060         return error;
1061 }
1062
1063 static const unsigned int core_file_note_size_min = CORE_FILE_NOTE_SIZE_DEFAULT;
1064 static const unsigned int core_file_note_size_max = CORE_FILE_NOTE_SIZE_MAX;
1065
1066 static struct ctl_table coredump_sysctls[] = {
1067         {
1068                 .procname       = "core_uses_pid",
1069                 .data           = &core_uses_pid,
1070                 .maxlen         = sizeof(int),
1071                 .mode           = 0644,
1072                 .proc_handler   = proc_dointvec,
1073         },
1074         {
1075                 .procname       = "core_pattern",
1076                 .data           = core_pattern,
1077                 .maxlen         = CORENAME_MAX_SIZE,
1078                 .mode           = 0644,
1079                 .proc_handler   = proc_dostring_coredump,
1080         },
1081         {
1082                 .procname       = "core_pipe_limit",
1083                 .data           = &core_pipe_limit,
1084                 .maxlen         = sizeof(unsigned int),
1085                 .mode           = 0644,
1086                 .proc_handler   = proc_dointvec,
1087         },
1088         {
1089                 .procname       = "core_file_note_size_limit",
1090                 .data           = &core_file_note_size_limit,
1091                 .maxlen         = sizeof(unsigned int),
1092                 .mode           = 0644,
1093                 .proc_handler   = proc_douintvec_minmax,
1094                 .extra1         = (unsigned int *)&core_file_note_size_min,
1095                 .extra2         = (unsigned int *)&core_file_note_size_max,
1096         },
1097 };
1098
1099 static int __init init_fs_coredump_sysctls(void)
1100 {
1101         register_sysctl_init("kernel", coredump_sysctls);
1102         return 0;
1103 }
1104 fs_initcall(init_fs_coredump_sysctls);
1105 #endif /* CONFIG_SYSCTL */
1106
1107 /*
1108  * The purpose of always_dump_vma() is to make sure that special kernel mappings
1109  * that are useful for post-mortem analysis are included in every core dump.
1110  * In that way we ensure that the core dump is fully interpretable later
1111  * without matching up the same kernel and hardware config to see what PC values
1112  * meant. These special mappings include - vDSO, vsyscall, and other
1113  * architecture specific mappings
1114  */
1115 static bool always_dump_vma(struct vm_area_struct *vma)
1116 {
1117         /* Any vsyscall mappings? */
1118         if (vma == get_gate_vma(vma->vm_mm))
1119                 return true;
1120
1121         /*
1122          * Assume that all vmas with a .name op should always be dumped.
1123          * If this changes, a new vm_ops field can easily be added.
1124          */
1125         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
1126                 return true;
1127
1128         /*
1129          * arch_vma_name() returns non-NULL for special architecture mappings,
1130          * such as vDSO sections.
1131          */
1132         if (arch_vma_name(vma))
1133                 return true;
1134
1135         return false;
1136 }
1137
1138 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
1139
1140 /*
1141  * Decide how much of @vma's contents should be included in a core dump.
1142  */
1143 static unsigned long vma_dump_size(struct vm_area_struct *vma,
1144                                    unsigned long mm_flags)
1145 {
1146 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
1147
1148         /* always dump the vdso and vsyscall sections */
1149         if (always_dump_vma(vma))
1150                 goto whole;
1151
1152         if (vma->vm_flags & VM_DONTDUMP)
1153                 return 0;
1154
1155         /* support for DAX */
1156         if (vma_is_dax(vma)) {
1157                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1158                         goto whole;
1159                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1160                         goto whole;
1161                 return 0;
1162         }
1163
1164         /* Hugetlb memory check */
1165         if (is_vm_hugetlb_page(vma)) {
1166                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1167                         goto whole;
1168                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1169                         goto whole;
1170                 return 0;
1171         }
1172
1173         /* Do not dump I/O mapped devices or special mappings */
1174         if (vma->vm_flags & VM_IO)
1175                 return 0;
1176
1177         /* By default, dump shared memory if mapped from an anonymous file. */
1178         if (vma->vm_flags & VM_SHARED) {
1179                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1180                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1181                         goto whole;
1182                 return 0;
1183         }
1184
1185         /* Dump segments that have been written to.  */
1186         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1187                 goto whole;
1188         if (vma->vm_file == NULL)
1189                 return 0;
1190
1191         if (FILTER(MAPPED_PRIVATE))
1192                 goto whole;
1193
1194         /*
1195          * If this is the beginning of an executable file mapping,
1196          * dump the first page to aid in determining what was mapped here.
1197          */
1198         if (FILTER(ELF_HEADERS) &&
1199             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1200                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1201                         return PAGE_SIZE;
1202
1203                 /*
1204                  * ELF libraries aren't always executable.
1205                  * We'll want to check whether the mapping starts with the ELF
1206                  * magic, but not now - we're holding the mmap lock,
1207                  * so copy_from_user() doesn't work here.
1208                  * Use a placeholder instead, and fix it up later in
1209                  * dump_vma_snapshot().
1210                  */
1211                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1212         }
1213
1214 #undef  FILTER
1215
1216         return 0;
1217
1218 whole:
1219         return vma->vm_end - vma->vm_start;
1220 }
1221
1222 /*
1223  * Helper function for iterating across a vma list.  It ensures that the caller
1224  * will visit `gate_vma' prior to terminating the search.
1225  */
1226 static struct vm_area_struct *coredump_next_vma(struct vma_iterator *vmi,
1227                                        struct vm_area_struct *vma,
1228                                        struct vm_area_struct *gate_vma)
1229 {
1230         if (gate_vma && (vma == gate_vma))
1231                 return NULL;
1232
1233         vma = vma_next(vmi);
1234         if (vma)
1235                 return vma;
1236         return gate_vma;
1237 }
1238
1239 static void free_vma_snapshot(struct coredump_params *cprm)
1240 {
1241         if (cprm->vma_meta) {
1242                 int i;
1243                 for (i = 0; i < cprm->vma_count; i++) {
1244                         struct file *file = cprm->vma_meta[i].file;
1245                         if (file)
1246                                 fput(file);
1247                 }
1248                 kvfree(cprm->vma_meta);
1249                 cprm->vma_meta = NULL;
1250         }
1251 }
1252
1253 static int cmp_vma_size(const void *vma_meta_lhs_ptr, const void *vma_meta_rhs_ptr)
1254 {
1255         const struct core_vma_metadata *vma_meta_lhs = vma_meta_lhs_ptr;
1256         const struct core_vma_metadata *vma_meta_rhs = vma_meta_rhs_ptr;
1257
1258         if (vma_meta_lhs->dump_size < vma_meta_rhs->dump_size)
1259                 return -1;
1260         if (vma_meta_lhs->dump_size > vma_meta_rhs->dump_size)
1261                 return 1;
1262         return 0;
1263 }
1264
1265 /*
1266  * Under the mmap_lock, take a snapshot of relevant information about the task's
1267  * VMAs.
1268  */
1269 static bool dump_vma_snapshot(struct coredump_params *cprm)
1270 {
1271         struct vm_area_struct *gate_vma, *vma = NULL;
1272         struct mm_struct *mm = current->mm;
1273         VMA_ITERATOR(vmi, mm, 0);
1274         int i = 0;
1275
1276         /*
1277          * Once the stack expansion code is fixed to not change VMA bounds
1278          * under mmap_lock in read mode, this can be changed to take the
1279          * mmap_lock in read mode.
1280          */
1281         if (mmap_write_lock_killable(mm))
1282                 return false;
1283
1284         cprm->vma_data_size = 0;
1285         gate_vma = get_gate_vma(mm);
1286         cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1287
1288         cprm->vma_meta = kvmalloc_array(cprm->vma_count, sizeof(*cprm->vma_meta), GFP_KERNEL);
1289         if (!cprm->vma_meta) {
1290                 mmap_write_unlock(mm);
1291                 return false;
1292         }
1293
1294         while ((vma = coredump_next_vma(&vmi, vma, gate_vma)) != NULL) {
1295                 struct core_vma_metadata *m = cprm->vma_meta + i;
1296
1297                 m->start = vma->vm_start;
1298                 m->end = vma->vm_end;
1299                 m->flags = vma->vm_flags;
1300                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1301                 m->pgoff = vma->vm_pgoff;
1302                 m->file = vma->vm_file;
1303                 if (m->file)
1304                         get_file(m->file);
1305                 i++;
1306         }
1307
1308         mmap_write_unlock(mm);
1309
1310         for (i = 0; i < cprm->vma_count; i++) {
1311                 struct core_vma_metadata *m = cprm->vma_meta + i;
1312
1313                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1314                         char elfmag[SELFMAG];
1315
1316                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1317                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1318                                 m->dump_size = 0;
1319                         } else {
1320                                 m->dump_size = PAGE_SIZE;
1321                         }
1322                 }
1323
1324                 cprm->vma_data_size += m->dump_size;
1325         }
1326
1327         sort(cprm->vma_meta, cprm->vma_count, sizeof(*cprm->vma_meta),
1328                 cmp_vma_size, NULL);
1329
1330         return true;
1331 }