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