fs/coredump.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/slab.h>
   3 #include <linux/file.h>
   4 #include <linux/fdtable.h>
   5 #include <linux/freezer.h>
   6 #include <linux/mm.h>
   7 #include <linux/stat.h>
   8 #include <linux/fcntl.h>
   9 #include <linux/swap.h>
  10 #include <linux/ctype.h>
  11 #include <linux/string.h>
  12 #include <linux/init.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/perf_event.h>
  15 #include <linux/highmem.h>
  16 #include <linux/spinlock.h>
  17 #include <linux/key.h>
  18 #include <linux/personality.h>
  19 #include <linux/binfmts.h>
  20 #include <linux/coredump.h>
  21 #include <linux/sched/coredump.h>
  22 #include <linux/sched/signal.h>
  23 #include <linux/sched/task_stack.h>
  24 #include <linux/utsname.h>
  25 #include <linux/pid_namespace.h>
  26 #include <linux/module.h>
  27 #include <linux/namei.h>
  28 #include <linux/mount.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/tsacct_kern.h>
  32 #include <linux/cn_proc.h>
  33 #include <linux/audit.h>
  34 #include <linux/tracehook.h>
  35 #include <linux/kmod.h>
  36 #include <linux/fsnotify.h>
  37 #include <linux/fs_struct.h>
  38 #include <linux/pipe_fs_i.h>
  39 #include <linux/oom.h>
  40 #include <linux/compat.h>
  41 #include <linux/fs.h>
  42 #include <linux/path.h>
  43 #include <linux/timekeeping.h>
  44 #include <linux/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 int core_uses_pid;
  57 unsigned int core_pipe_limit;
  58 char core_pattern[CORENAME_MAX_SIZE] = "core";
  59 static int core_name_size = CORENAME_MAX_SIZE;
  60
  61 struct core_name {
  62         char *corename;
  63         int used, size;
  64 };
  65
  66 /* The maximal length of core_pattern is also specified in sysctl.c */
  67
  68 static int expand_corename(struct core_name *cn, int size)
  69 {
  70         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  71
  72         if (!corename)
  73                 return -ENOMEM;
  74
  75         if (size > core_name_size) /* racy but harmless */
  76                 core_name_size = size;
  77
  78         cn->size = ksize(corename);
  79         cn->corename = corename;
  80         return 0;
  81 }
  82
  83 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  84                                      va_list arg)
  85 {
  86         int free, need;
  87         va_list arg_copy;
  88
  89 again:
  90         free = cn->size - cn->used;
  91
  92         va_copy(arg_copy, arg);
  93         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  94         va_end(arg_copy);
  95
  96         if (need < free) {
  97                 cn->used += need;
  98                 return 0;
  99         }
 100
 101         if (!expand_corename(cn, cn->size + need - free + 1))
 102                 goto again;
 103
 104         return -ENOMEM;
 105 }
 106
 107 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 108 {
 109         va_list arg;
 110         int ret;
 111
 112         va_start(arg, fmt);
 113         ret = cn_vprintf(cn, fmt, arg);
 114         va_end(arg);
 115
 116         return ret;
 117 }
 118
 119 static __printf(2, 3)
 120 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 121 {
 122         int cur = cn->used;
 123         va_list arg;
 124         int ret;
 125
 126         va_start(arg, fmt);
 127         ret = cn_vprintf(cn, fmt, arg);
 128         va_end(arg);
 129
 130         if (ret == 0) {
 131                 /*
 132                  * Ensure that this coredump name component can't cause the
 133                  * resulting corefile path to consist of a ".." or ".".
 134                  */
 135                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 136                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 137                                 && cn->corename[cur+1] == '.'))
 138                         cn->corename[cur] = '!';
 139
 140                 /*
 141                  * Empty names are fishy and could be used to create a "//" in a
 142                  * corefile name, causing the coredump to happen one directory
 143                  * level too high. Enforce that all components of the core
 144                  * pattern are at least one character long.
 145                  */
 146                 if (cn->used == cur)
 147                         ret = cn_printf(cn, "!");
 148         }
 149
 150         for (; cur < cn->used; ++cur) {
 151                 if (cn->corename[cur] == '/')
 152                         cn->corename[cur] = '!';
 153         }
 154         return ret;
 155 }
 156
 157 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 158 {
 159         struct file *exe_file;
 160         char *pathbuf, *path, *ptr;
 161         int ret;
 162
 163         exe_file = get_mm_exe_file(current->mm);
 164         if (!exe_file)
 165                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 166
 167         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 168         if (!pathbuf) {
 169                 ret = -ENOMEM;
 170                 goto put_exe_file;
 171         }
 172
 173         path = file_path(exe_file, pathbuf, PATH_MAX);
 174         if (IS_ERR(path)) {
 175                 ret = PTR_ERR(path);
 176                 goto free_buf;
 177         }
 178
 179         if (name_only) {
 180                 ptr = strrchr(path, '/');
 181                 if (ptr)
 182                         path = ptr + 1;
 183         }
 184         ret = cn_esc_printf(cn, "%s", path);
 185
 186 free_buf:
 187         kfree(pathbuf);
 188 put_exe_file:
 189         fput(exe_file);
 190         return ret;
 191 }
 192
 193 /* format_corename will inspect the pattern parameter, and output a
 194  * name into corename, which must have space for at least
 195  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 196  */
 197 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 198                            size_t **argv, int *argc)
 199 {
 200         const struct cred *cred = current_cred();
 201         const char *pat_ptr = core_pattern;
 202         int ispipe = (*pat_ptr == '|');
 203         bool was_space = false;
 204         int pid_in_pattern = 0;
 205         int err = 0;
 206
 207         cn->used = 0;
 208         cn->corename = NULL;
 209         if (expand_corename(cn, core_name_size))
 210                 return -ENOMEM;
 211         cn->corename[0] = '\0';
 212
 213         if (ispipe) {
 214                 int argvs = sizeof(core_pattern) / 2;
 215                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 216                 if (!(*argv))
 217                         return -ENOMEM;
 218                 (*argv)[(*argc)++] = 0;
 219                 ++pat_ptr;
 220                 if (!(*pat_ptr))
 221                         return -ENOMEM;
 222         }
 223
 224         /* Repeat as long as we have more pattern to process and more output
 225            space */
 226         while (*pat_ptr) {
 227                 /*
 228                  * Split on spaces before doing template expansion so that
 229                  * %e and %E don't get split if they have spaces in them
 230                  */
 231                 if (ispipe) {
 232                         if (isspace(*pat_ptr)) {
 233                                 if (cn->used != 0)
 234                                         was_space = true;
 235                                 pat_ptr++;
 236                                 continue;
 237                         } else if (was_space) {
 238                                 was_space = false;
 239                                 err = cn_printf(cn, "%c", '\0');
 240                                 if (err)
 241                                         return err;
 242                                 (*argv)[(*argc)++] = cn->used;
 243                         }
 244                 }
 245                 if (*pat_ptr != '%') {
 246                         err = cn_printf(cn, "%c", *pat_ptr++);
 247                 } else {
 248                         switch (*++pat_ptr) {
 249                         /* single % at the end, drop that */
 250                         case 0:
 251                                 goto out;
 252                         /* Double percent, output one percent */
 253                         case '%':
 254                                 err = cn_printf(cn, "%c", '%');
 255                                 break;
 256                         /* pid */
 257                         case 'p':
 258                                 pid_in_pattern = 1;
 259                                 err = cn_printf(cn, "%d",
 260                                               task_tgid_vnr(current));
 261                                 break;
 262                         /* global pid */
 263                         case 'P':
 264                                 err = cn_printf(cn, "%d",
 265                                               task_tgid_nr(current));
 266                                 break;
 267                         case 'i':
 268                                 err = cn_printf(cn, "%d",
 269                                               task_pid_vnr(current));
 270                                 break;
 271                         case 'I':
 272                                 err = cn_printf(cn, "%d",
 273                                               task_pid_nr(current));
 274                                 break;
 275                         /* uid */
 276                         case 'u':
 277                                 err = cn_printf(cn, "%u",
 278                                                 from_kuid(&init_user_ns,
 279                                                           cred->uid));
 280                                 break;
 281                         /* gid */
 282                         case 'g':
 283                                 err = cn_printf(cn, "%u",
 284                                                 from_kgid(&init_user_ns,
 285                                                           cred->gid));
 286                                 break;
 287                         case 'd':
 288                                 err = cn_printf(cn, "%d",
 289                                         __get_dumpable(cprm->mm_flags));
 290                                 break;
 291                         /* signal that caused the coredump */
 292                         case 's':
 293                                 err = cn_printf(cn, "%d",
 294                                                 cprm->siginfo->si_signo);
 295                                 break;
 296                         /* UNIX time of coredump */
 297                         case 't': {
 298                                 time64_t time;
 299
 300                                 time = ktime_get_real_seconds();
 301                                 err = cn_printf(cn, "%lld", time);
 302                                 break;
 303                         }
 304                         /* hostname */
 305                         case 'h':
 306                                 down_read(&uts_sem);
 307                                 err = cn_esc_printf(cn, "%s",
 308                                               utsname()->nodename);
 309                                 up_read(&uts_sem);
 310                                 break;
 311                         /* executable, could be changed by prctl PR_SET_NAME etc */
 312                         case 'e':
 313                                 err = cn_esc_printf(cn, "%s", current->comm);
 314                                 break;
 315                         /* file name of executable */
 316                         case 'f':
 317                                 err = cn_print_exe_file(cn, true);
 318                                 break;
 319                         case 'E':
 320                                 err = cn_print_exe_file(cn, false);
 321                                 break;
 322                         /* core limit size */
 323                         case 'c':
 324                                 err = cn_printf(cn, "%lu",
 325                                               rlimit(RLIMIT_CORE));
 326                                 break;
 327                         default:
 328                                 break;
 329                         }
 330                         ++pat_ptr;
 331                 }
 332
 333                 if (err)
 334                         return err;
 335         }
 336
 337 out:
 338         /* Backward compatibility with core_uses_pid:
 339          *
 340          * If core_pattern does not include a %p (as is the default)
 341          * and core_uses_pid is set, then .%pid will be appended to
 342          * the filename. Do not do this for piped commands. */
 343         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 344                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 345                 if (err)
 346                         return err;
 347         }
 348         return ispipe;
 349 }
 350
 351 static int zap_process(struct task_struct *start, int exit_code, int flags)
 352 {
 353         struct task_struct *t;
 354         int nr = 0;
 355
 356         /* ignore all signals except SIGKILL, see prepare_signal() */
 357         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
 358         start->signal->group_exit_code = exit_code;
 359         start->signal->group_stop_count = 0;
 360
 361         for_each_thread(start, t) {
 362                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 363                 if (t != current && t->mm) {
 364                         sigaddset(&t->pending.signal, SIGKILL);
 365                         signal_wake_up(t, 1);
 366                         nr++;
 367                 }
 368         }
 369
 370         return nr;
 371 }
 372
 373 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 374                         struct core_state *core_state, int exit_code)
 375 {
 376         struct task_struct *g, *p;
 377         unsigned long flags;
 378         int nr = -EAGAIN;
 379
 380         spin_lock_irq(&tsk->sighand->siglock);
 381         if (!signal_group_exit(tsk->signal)) {
 382                 mm->core_state = core_state;
 383                 tsk->signal->group_exit_task = tsk;
 384                 nr = zap_process(tsk, exit_code, 0);
 385                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 386         }
 387         spin_unlock_irq(&tsk->sighand->siglock);
 388         if (unlikely(nr < 0))
 389                 return nr;
 390
 391         tsk->flags |= PF_DUMPCORE;
 392         if (atomic_read(&mm->mm_users) == nr + 1)
 393                 goto done;
 394         /*
 395          * We should find and kill all tasks which use this mm, and we should
 396          * count them correctly into ->nr_threads. We don't take tasklist
 397          * lock, but this is safe wrt:
 398          *
 399          * fork:
 400          *      None of sub-threads can fork after zap_process(leader). All
 401          *      processes which were created before this point should be
 402          *      visible to zap_threads() because copy_process() adds the new
 403          *      process to the tail of init_task.tasks list, and lock/unlock
 404          *      of ->siglock provides a memory barrier.
 405          *
 406          * do_exit:
 407          *      The caller holds mm->mmap_lock. This means that the task which
 408          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 409          *      its ->mm.
 410          *
 411          * de_thread:
 412          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 413          *      we must see either old or new leader, this does not matter.
 414          *      However, it can change p->sighand, so lock_task_sighand(p)
 415          *      must be used. Since p->mm != NULL and we hold ->mmap_lock
 416          *      it can't fail.
 417          *
 418          *      Note also that "g" can be the old leader with ->mm == NULL
 419          *      and already unhashed and thus removed from ->thread_group.
 420          *      This is OK, __unhash_process()->list_del_rcu() does not
 421          *      clear the ->next pointer, we will find the new leader via
 422          *      next_thread().
 423          */
 424         rcu_read_lock();
 425         for_each_process(g) {
 426                 if (g == tsk->group_leader)
 427                         continue;
 428                 if (g->flags & PF_KTHREAD)
 429                         continue;
 430
 431                 for_each_thread(g, p) {
 432                         if (unlikely(!p->mm))
 433                                 continue;
 434                         if (unlikely(p->mm == mm)) {
 435                                 lock_task_sighand(p, &flags);
 436                                 nr += zap_process(p, exit_code,
 437                                                         SIGNAL_GROUP_EXIT);
 438                                 unlock_task_sighand(p, &flags);
 439                         }
 440                         break;
 441                 }
 442         }
 443         rcu_read_unlock();
 444 done:
 445         atomic_set(&core_state->nr_threads, nr);
 446         return nr;
 447 }
 448
 449 static int coredump_wait(int exit_code, struct core_state *core_state)
 450 {
 451         struct task_struct *tsk = current;
 452         struct mm_struct *mm = tsk->mm;
 453         int core_waiters = -EBUSY;
 454
 455         init_completion(&core_state->startup);
 456         core_state->dumper.task = tsk;
 457         core_state->dumper.next = NULL;
 458
 459         if (mmap_write_lock_killable(mm))
 460                 return -EINTR;
 461
 462         if (!mm->core_state)
 463                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 464         mmap_write_unlock(mm);
 465
 466         if (core_waiters > 0) {
 467                 struct core_thread *ptr;
 468
 469                 freezer_do_not_count();
 470                 wait_for_completion(&core_state->startup);
 471                 freezer_count();
 472                 /*
 473                  * Wait for all the threads to become inactive, so that
 474                  * all the thread context (extended register state, like
 475                  * fpu etc) gets copied to the memory.
 476                  */
 477                 ptr = core_state->dumper.next;
 478                 while (ptr != NULL) {
 479                         wait_task_inactive(ptr->task, 0);
 480                         ptr = ptr->next;
 481                 }
 482         }
 483
 484         return core_waiters;
 485 }
 486
 487 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
 488 {
 489         struct core_thread *curr, *next;
 490         struct task_struct *task;
 491
 492         spin_lock_irq(&current->sighand->siglock);
 493         if (core_dumped && !__fatal_signal_pending(current))
 494                 current->signal->group_exit_code |= 0x80;
 495         current->signal->group_exit_task = NULL;
 496         current->signal->flags = SIGNAL_GROUP_EXIT;
 497         spin_unlock_irq(&current->sighand->siglock);
 498
 499         next = mm->core_state->dumper.next;
 500         while ((curr = next) != NULL) {
 501                 next = curr->next;
 502                 task = curr->task;
 503                 /*
 504                  * see exit_mm(), curr->task must not see
 505                  * ->task == NULL before we read ->next.
 506                  */
 507                 smp_mb();
 508                 curr->task = NULL;
 509                 wake_up_process(task);
 510         }
 511
 512         mm->core_state = NULL;
 513 }
 514
 515 static bool dump_interrupted(void)
 516 {
 517         /*
 518          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 519          * can do try_to_freeze() and check __fatal_signal_pending(),
 520          * but then we need to teach dump_write() to restart and clear
 521          * TIF_SIGPENDING.
 522          */
 523         return fatal_signal_pending(current) || freezing(current);
 524 }
 525
 526 static void wait_for_dump_helpers(struct file *file)
 527 {
 528         struct pipe_inode_info *pipe = file->private_data;
 529
 530         pipe_lock(pipe);
 531         pipe->readers++;
 532         pipe->writers--;
 533         wake_up_interruptible_sync(&pipe->rd_wait);
 534         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 535         pipe_unlock(pipe);
 536
 537         /*
 538          * We actually want wait_event_freezable() but then we need
 539          * to clear TIF_SIGPENDING and improve dump_interrupted().
 540          */
 541         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 542
 543         pipe_lock(pipe);
 544         pipe->readers--;
 545         pipe->writers++;
 546         pipe_unlock(pipe);
 547 }
 548
 549 /*
 550  * umh_pipe_setup
 551  * helper function to customize the process used
 552  * to collect the core in userspace.  Specifically
 553  * it sets up a pipe and installs it as fd 0 (stdin)
 554  * for the process.  Returns 0 on success, or
 555  * PTR_ERR on failure.
 556  * Note that it also sets the core limit to 1.  This
 557  * is a special value that we use to trap recursive
 558  * core dumps
 559  */
 560 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 561 {
 562         struct file *files[2];
 563         struct coredump_params *cp = (struct coredump_params *)info->data;
 564         int err = create_pipe_files(files, 0);
 565         if (err)
 566                 return err;
 567
 568         cp->file = files[1];
 569
 570         err = replace_fd(0, files[0], 0);
 571         fput(files[0]);
 572         /* and disallow core files too */
 573         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 574
 575         return err;
 576 }
 577
 578 void do_coredump(const kernel_siginfo_t *siginfo)
 579 {
 580         struct core_state core_state;
 581         struct core_name cn;
 582         struct mm_struct *mm = current->mm;
 583         struct linux_binfmt * binfmt;
 584         const struct cred *old_cred;
 585         struct cred *cred;
 586         int retval = 0;
 587         int ispipe;
 588         size_t *argv = NULL;
 589         int argc = 0;
 590         /* require nonrelative corefile path and be extra careful */
 591         bool need_suid_safe = false;
 592         bool core_dumped = false;
 593         static atomic_t core_dump_count = ATOMIC_INIT(0);
 594         struct coredump_params cprm = {
 595                 .siginfo = siginfo,
 596                 .regs = signal_pt_regs(),
 597                 .limit = rlimit(RLIMIT_CORE),
 598                 /*
 599                  * We must use the same mm->flags while dumping core to avoid
 600                  * inconsistency of bit flags, since this flag is not protected
 601                  * by any locks.
 602                  */
 603                 .mm_flags = mm->flags,
 604         };
 605
 606         audit_core_dumps(siginfo->si_signo);
 607
 608         binfmt = mm->binfmt;
 609         if (!binfmt || !binfmt->core_dump)
 610                 goto fail;
 611         if (!__get_dumpable(cprm.mm_flags))
 612                 goto fail;
 613
 614         cred = prepare_creds();
 615         if (!cred)
 616                 goto fail;
 617         /*
 618          * We cannot trust fsuid as being the "true" uid of the process
 619          * nor do we know its entire history. We only know it was tainted
 620          * so we dump it as root in mode 2, and only into a controlled
 621          * environment (pipe handler or fully qualified path).
 622          */
 623         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 624                 /* Setuid core dump mode */
 625                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 626                 need_suid_safe = true;
 627         }
 628
 629         retval = coredump_wait(siginfo->si_signo, &core_state);
 630         if (retval < 0)
 631                 goto fail_creds;
 632
 633         old_cred = override_creds(cred);
 634
 635         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 636
 637         if (ispipe) {
 638                 int argi;
 639                 int dump_count;
 640                 char **helper_argv;
 641                 struct subprocess_info *sub_info;
 642
 643                 if (ispipe < 0) {
 644                         printk(KERN_WARNING "format_corename failed\n");
 645                         printk(KERN_WARNING "Aborting core\n");
 646                         goto fail_unlock;
 647                 }
 648
 649                 if (cprm.limit == 1) {
 650                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 651                          *
 652                          * Normally core limits are irrelevant to pipes, since
 653                          * we're not writing to the file system, but we use
 654                          * cprm.limit of 1 here as a special value, this is a
 655                          * consistent way to catch recursive crashes.
 656                          * We can still crash if the core_pattern binary sets
 657                          * RLIM_CORE = !1, but it runs as root, and can do
 658                          * lots of stupid things.
 659                          *
 660                          * Note that we use task_tgid_vnr here to grab the pid
 661                          * of the process group leader.  That way we get the
 662                          * right pid if a thread in a multi-threaded
 663                          * core_pattern process dies.
 664                          */
 665                         printk(KERN_WARNING
 666                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 667                                 task_tgid_vnr(current), current->comm);
 668                         printk(KERN_WARNING "Aborting core\n");
 669                         goto fail_unlock;
 670                 }
 671                 cprm.limit = RLIM_INFINITY;
 672
 673                 dump_count = atomic_inc_return(&core_dump_count);
 674                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 675                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 676                                task_tgid_vnr(current), current->comm);
 677                         printk(KERN_WARNING "Skipping core dump\n");
 678                         goto fail_dropcount;
 679                 }
 680
 681                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 682                                             GFP_KERNEL);
 683                 if (!helper_argv) {
 684                         printk(KERN_WARNING "%s failed to allocate memory\n",
 685                                __func__);
 686                         goto fail_dropcount;
 687                 }
 688                 for (argi = 0; argi < argc; argi++)
 689                         helper_argv[argi] = cn.corename + argv[argi];
 690                 helper_argv[argi] = NULL;
 691
 692                 retval = -ENOMEM;
 693                 sub_info = call_usermodehelper_setup(helper_argv[0],
 694                                                 helper_argv, NULL, GFP_KERNEL,
 695                                                 umh_pipe_setup, NULL, &cprm);
 696                 if (sub_info)
 697                         retval = call_usermodehelper_exec(sub_info,
 698                                                           UMH_WAIT_EXEC);
 699
 700                 kfree(helper_argv);
 701                 if (retval) {
 702                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 703                                cn.corename);
 704                         goto close_fail;
 705                 }
 706         } else {
 707                 struct user_namespace *mnt_userns;
 708                 struct inode *inode;
 709                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 710                                  O_LARGEFILE | O_EXCL;
 711
 712                 if (cprm.limit < binfmt->min_coredump)
 713                         goto fail_unlock;
 714
 715                 if (need_suid_safe && cn.corename[0] != '/') {
 716                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 717                                 "to fully qualified path!\n",
 718                                 task_tgid_vnr(current), current->comm);
 719                         printk(KERN_WARNING "Skipping core dump\n");
 720                         goto fail_unlock;
 721                 }
 722
 723                 /*
 724                  * Unlink the file if it exists unless this is a SUID
 725                  * binary - in that case, we're running around with root
 726                  * privs and don't want to unlink another user's coredump.
 727                  */
 728                 if (!need_suid_safe) {
 729                         /*
 730                          * If it doesn't exist, that's fine. If there's some
 731                          * other problem, we'll catch it at the filp_open().
 732                          */
 733                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 734                 }
 735
 736                 /*
 737                  * There is a race between unlinking and creating the
 738                  * file, but if that causes an EEXIST here, that's
 739                  * fine - another process raced with us while creating
 740                  * the corefile, and the other process won. To userspace,
 741                  * what matters is that at least one of the two processes
 742                  * writes its coredump successfully, not which one.
 743                  */
 744                 if (need_suid_safe) {
 745                         /*
 746                          * Using user namespaces, normal user tasks can change
 747                          * their current->fs->root to point to arbitrary
 748                          * directories. Since the intention of the "only dump
 749                          * with a fully qualified path" rule is to control where
 750                          * coredumps may be placed using root privileges,
 751                          * current->fs->root must not be used. Instead, use the
 752                          * root directory of init_task.
 753                          */
 754                         struct path root;
 755
 756                         task_lock(&init_task);
 757                         get_fs_root(init_task.fs, &root);
 758                         task_unlock(&init_task);
 759                         cprm.file = file_open_root(&root, cn.corename,
 760                                                    open_flags, 0600);
 761                         path_put(&root);
 762                 } else {
 763                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 764                 }
 765                 if (IS_ERR(cprm.file))
 766                         goto fail_unlock;
 767
 768                 inode = file_inode(cprm.file);
 769                 if (inode->i_nlink > 1)
 770                         goto close_fail;
 771                 if (d_unhashed(cprm.file->f_path.dentry))
 772                         goto close_fail;
 773                 /*
 774                  * AK: actually i see no reason to not allow this for named
 775                  * pipes etc, but keep the previous behaviour for now.
 776                  */
 777                 if (!S_ISREG(inode->i_mode))
 778                         goto close_fail;
 779                 /*
 780                  * Don't dump core if the filesystem changed owner or mode
 781                  * of the file during file creation. This is an issue when
 782                  * a process dumps core while its cwd is e.g. on a vfat
 783                  * filesystem.
 784                  */
 785                 mnt_userns = file_mnt_user_ns(cprm.file);
 786                 if (!uid_eq(i_uid_into_mnt(mnt_userns, inode),
 787                             current_fsuid())) {
 788                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file owner\n",
 789                                             cn.corename);
 790                         goto close_fail;
 791                 }
 792                 if ((inode->i_mode & 0677) != 0600) {
 793                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file permissions\n",
 794                                             cn.corename);
 795                         goto close_fail;
 796                 }
 797                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 798                         goto close_fail;
 799                 if (do_truncate(mnt_userns, cprm.file->f_path.dentry,
 800                                 0, 0, cprm.file))
 801                         goto close_fail;
 802         }
 803
 804         /* get us an unshared descriptor table; almost always a no-op */
 805         /* The cell spufs coredump code reads the file descriptor tables */
 806         retval = unshare_files();
 807         if (retval)
 808                 goto close_fail;
 809         if (!dump_interrupted()) {
 810                 /*
 811                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 812                  * have this set to NULL.
 813                  */
 814                 if (!cprm.file) {
 815                         pr_info("Core dump to |%s disabled\n", cn.corename);
 816                         goto close_fail;
 817                 }
 818                 file_start_write(cprm.file);
 819                 core_dumped = binfmt->core_dump(&cprm);
 820                 /*
 821                  * Ensures that file size is big enough to contain the current
 822                  * file postion. This prevents gdb from complaining about
 823                  * a truncated file if the last "write" to the file was
 824                  * dump_skip.
 825                  */
 826                 if (cprm.to_skip) {
 827                         cprm.to_skip--;
 828                         dump_emit(&cprm, "", 1);
 829                 }
 830                 file_end_write(cprm.file);
 831         }
 832         if (ispipe && core_pipe_limit)
 833                 wait_for_dump_helpers(cprm.file);
 834 close_fail:
 835         if (cprm.file)
 836                 filp_close(cprm.file, NULL);
 837 fail_dropcount:
 838         if (ispipe)
 839                 atomic_dec(&core_dump_count);
 840 fail_unlock:
 841         kfree(argv);
 842         kfree(cn.corename);
 843         coredump_finish(mm, core_dumped);
 844         revert_creds(old_cred);
 845 fail_creds:
 846         put_cred(cred);
 847 fail:
 848         return;
 849 }
 850
 851 /*
 852  * Core dumping helper functions.  These are the only things you should
 853  * do on a core-file: use only these functions to write out all the
 854  * necessary info.
 855  */
 856 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 857 {
 858         struct file *file = cprm->file;
 859         loff_t pos = file->f_pos;
 860         ssize_t n;
 861         if (cprm->written + nr > cprm->limit)
 862                 return 0;
 863
 864
 865         if (dump_interrupted())
 866                 return 0;
 867         n = __kernel_write(file, addr, nr, &pos);
 868         if (n != nr)
 869                 return 0;
 870         file->f_pos = pos;
 871         cprm->written += n;
 872         cprm->pos += n;
 873
 874         return 1;
 875 }
 876
 877 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 878 {
 879         static char zeroes[PAGE_SIZE];
 880         struct file *file = cprm->file;
 881         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 882                 if (dump_interrupted() ||
 883                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
 884                         return 0;
 885                 cprm->pos += nr;
 886                 return 1;
 887         } else {
 888                 while (nr > PAGE_SIZE) {
 889                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 890                                 return 0;
 891                         nr -= PAGE_SIZE;
 892                 }
 893                 return __dump_emit(cprm, zeroes, nr);
 894         }
 895 }
 896
 897 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 898 {
 899         if (cprm->to_skip) {
 900                 if (!__dump_skip(cprm, cprm->to_skip))
 901                         return 0;
 902                 cprm->to_skip = 0;
 903         }
 904         return __dump_emit(cprm, addr, nr);
 905 }
 906 EXPORT_SYMBOL(dump_emit);
 907
 908 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 909 {
 910         cprm->to_skip = pos - cprm->pos;
 911 }
 912 EXPORT_SYMBOL(dump_skip_to);
 913
 914 void dump_skip(struct coredump_params *cprm, size_t nr)
 915 {
 916         cprm->to_skip += nr;
 917 }
 918 EXPORT_SYMBOL(dump_skip);
 919
 920 #ifdef CONFIG_ELF_CORE
 921 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 922                     unsigned long len)
 923 {
 924         unsigned long addr;
 925
 926         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 927                 struct page *page;
 928                 int stop;
 929
 930                 /*
 931                  * To avoid having to allocate page tables for virtual address
 932                  * ranges that have never been used yet, and also to make it
 933                  * easy to generate sparse core files, use a helper that returns
 934                  * NULL when encountering an empty page table entry that would
 935                  * otherwise have been filled with the zero page.
 936                  */
 937                 page = get_dump_page(addr);
 938                 if (page) {
 939                         void *kaddr = kmap_local_page(page);
 940
 941                         stop = !dump_emit(cprm, kaddr, PAGE_SIZE);
 942                         kunmap_local(kaddr);
 943                         put_page(page);
 944                         if (stop)
 945                                 return 0;
 946                 } else {
 947                         dump_skip(cprm, PAGE_SIZE);
 948                 }
 949         }
 950         return 1;
 951 }
 952 #endif
 953
 954 int dump_align(struct coredump_params *cprm, int align)
 955 {
 956         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
 957         if (align & (align - 1))
 958                 return 0;
 959         if (mod)
 960                 cprm->to_skip += align - mod;
 961         return 1;
 962 }
 963 EXPORT_SYMBOL(dump_align);
 964
 965 /*
 966  * The purpose of always_dump_vma() is to make sure that special kernel mappings
 967  * that are useful for post-mortem analysis are included in every core dump.
 968  * In that way we ensure that the core dump is fully interpretable later
 969  * without matching up the same kernel and hardware config to see what PC values
 970  * meant. These special mappings include - vDSO, vsyscall, and other
 971  * architecture specific mappings
 972  */
 973 static bool always_dump_vma(struct vm_area_struct *vma)
 974 {
 975         /* Any vsyscall mappings? */
 976         if (vma == get_gate_vma(vma->vm_mm))
 977                 return true;
 978
 979         /*
 980          * Assume that all vmas with a .name op should always be dumped.
 981          * If this changes, a new vm_ops field can easily be added.
 982          */
 983         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
 984                 return true;
 985
 986         /*
 987          * arch_vma_name() returns non-NULL for special architecture mappings,
 988          * such as vDSO sections.
 989          */
 990         if (arch_vma_name(vma))
 991                 return true;
 992
 993         return false;
 994 }
 995
 996 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
 997
 998 /*
 999  * Decide how much of @vma's contents should be included in a core dump.
1000  */
1001 static unsigned long vma_dump_size(struct vm_area_struct *vma,
1002                                    unsigned long mm_flags)
1003 {
1004 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
1005
1006         /* always dump the vdso and vsyscall sections */
1007         if (always_dump_vma(vma))
1008                 goto whole;
1009
1010         if (vma->vm_flags & VM_DONTDUMP)
1011                 return 0;
1012
1013         /* support for DAX */
1014         if (vma_is_dax(vma)) {
1015                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1016                         goto whole;
1017                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1018                         goto whole;
1019                 return 0;
1020         }
1021
1022         /* Hugetlb memory check */
1023         if (is_vm_hugetlb_page(vma)) {
1024                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1025                         goto whole;
1026                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1027                         goto whole;
1028                 return 0;
1029         }
1030
1031         /* Do not dump I/O mapped devices or special mappings */
1032         if (vma->vm_flags & VM_IO)
1033                 return 0;
1034
1035         /* By default, dump shared memory if mapped from an anonymous file. */
1036         if (vma->vm_flags & VM_SHARED) {
1037                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1038                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1039                         goto whole;
1040                 return 0;
1041         }
1042
1043         /* Dump segments that have been written to.  */
1044         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1045                 goto whole;
1046         if (vma->vm_file == NULL)
1047                 return 0;
1048
1049         if (FILTER(MAPPED_PRIVATE))
1050                 goto whole;
1051
1052         /*
1053          * If this is the beginning of an executable file mapping,
1054          * dump the first page to aid in determining what was mapped here.
1055          */
1056         if (FILTER(ELF_HEADERS) &&
1057             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1058                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1059                         return PAGE_SIZE;
1060
1061                 /*
1062                  * ELF libraries aren't always executable.
1063                  * We'll want to check whether the mapping starts with the ELF
1064                  * magic, but not now - we're holding the mmap lock,
1065                  * so copy_from_user() doesn't work here.
1066                  * Use a placeholder instead, and fix it up later in
1067                  * dump_vma_snapshot().
1068                  */
1069                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1070         }
1071
1072 #undef  FILTER
1073
1074         return 0;
1075
1076 whole:
1077         return vma->vm_end - vma->vm_start;
1078 }
1079
1080 static struct vm_area_struct *first_vma(struct task_struct *tsk,
1081                                         struct vm_area_struct *gate_vma)
1082 {
1083         struct vm_area_struct *ret = tsk->mm->mmap;
1084
1085         if (ret)
1086                 return ret;
1087         return gate_vma;
1088 }
1089
1090 /*
1091  * Helper function for iterating across a vma list.  It ensures that the caller
1092  * will visit `gate_vma' prior to terminating the search.
1093  */
1094 static struct vm_area_struct *next_vma(struct vm_area_struct *this_vma,
1095                                        struct vm_area_struct *gate_vma)
1096 {
1097         struct vm_area_struct *ret;
1098
1099         ret = this_vma->vm_next;
1100         if (ret)
1101                 return ret;
1102         if (this_vma == gate_vma)
1103                 return NULL;
1104         return gate_vma;
1105 }
1106
1107 /*
1108  * Under the mmap_lock, take a snapshot of relevant information about the task's
1109  * VMAs.
1110  */
1111 int dump_vma_snapshot(struct coredump_params *cprm, int *vma_count,
1112                       struct core_vma_metadata **vma_meta,
1113                       size_t *vma_data_size_ptr)
1114 {
1115         struct vm_area_struct *vma, *gate_vma;
1116         struct mm_struct *mm = current->mm;
1117         int i;
1118         size_t vma_data_size = 0;
1119
1120         /*
1121          * Once the stack expansion code is fixed to not change VMA bounds
1122          * under mmap_lock in read mode, this can be changed to take the
1123          * mmap_lock in read mode.
1124          */
1125         if (mmap_write_lock_killable(mm))
1126                 return -EINTR;
1127
1128         gate_vma = get_gate_vma(mm);
1129         *vma_count = mm->map_count + (gate_vma ? 1 : 0);
1130
1131         *vma_meta = kvmalloc_array(*vma_count, sizeof(**vma_meta), GFP_KERNEL);
1132         if (!*vma_meta) {
1133                 mmap_write_unlock(mm);
1134                 return -ENOMEM;
1135         }
1136
1137         for (i = 0, vma = first_vma(current, gate_vma); vma != NULL;
1138                         vma = next_vma(vma, gate_vma), i++) {
1139                 struct core_vma_metadata *m = (*vma_meta) + i;
1140
1141                 m->start = vma->vm_start;
1142                 m->end = vma->vm_end;
1143                 m->flags = vma->vm_flags;
1144                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1145         }
1146
1147         mmap_write_unlock(mm);
1148
1149         if (WARN_ON(i != *vma_count)) {
1150                 kvfree(*vma_meta);
1151                 return -EFAULT;
1152         }
1153
1154         for (i = 0; i < *vma_count; i++) {
1155                 struct core_vma_metadata *m = (*vma_meta) + i;
1156
1157                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1158                         char elfmag[SELFMAG];
1159
1160                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1161                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1162                                 m->dump_size = 0;
1163                         } else {
1164                                 m->dump_size = PAGE_SIZE;
1165                         }
1166                 }
1167
1168                 vma_data_size += m->dump_size;
1169         }
1170
1171         *vma_data_size_ptr = vma_data_size;
1172         return 0;
1173 }