fs/coredump.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/slab.h>
   3 #include <linux/file.h>
   4 #include <linux/fdtable.h>
   5 #include <linux/freezer.h>
   6 #include <linux/mm.h>
   7 #include <linux/stat.h>
   8 #include <linux/fcntl.h>
   9 #include <linux/swap.h>
  10 #include <linux/ctype.h>
  11 #include <linux/string.h>
  12 #include <linux/init.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/perf_event.h>
  15 #include <linux/highmem.h>
  16 #include <linux/spinlock.h>
  17 #include <linux/key.h>
  18 #include <linux/personality.h>
  19 #include <linux/binfmts.h>
  20 #include <linux/coredump.h>
  21 #include <linux/sched/coredump.h>
  22 #include <linux/sched/signal.h>
  23 #include <linux/sched/task_stack.h>
  24 #include <linux/utsname.h>
  25 #include <linux/pid_namespace.h>
  26 #include <linux/module.h>
  27 #include <linux/namei.h>
  28 #include <linux/mount.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/tsacct_kern.h>
  32 #include <linux/cn_proc.h>
  33 #include <linux/audit.h>
  34 #include <linux/kmod.h>
  35 #include <linux/fsnotify.h>
  36 #include <linux/fs_struct.h>
  37 #include <linux/pipe_fs_i.h>
  38 #include <linux/oom.h>
  39 #include <linux/compat.h>
  40 #include <linux/fs.h>
  41 #include <linux/path.h>
  42 #include <linux/timekeeping.h>
  43 #include <linux/sysctl.h>
  44 #include <linux/elf.h>
  45
  46 #include <linux/uaccess.h>
  47 #include <asm/mmu_context.h>
  48 #include <asm/tlb.h>
  49 #include <asm/exec.h>
  50
  51 #include <trace/events/task.h>
  52 #include "internal.h"
  53
  54 #include <trace/events/sched.h>
  55
  56 static bool dump_vma_snapshot(struct coredump_params *cprm);
  57 static void free_vma_snapshot(struct coredump_params *cprm);
  58
  59 #define CORE_FILE_NOTE_SIZE_DEFAULT (4*1024*1024)
  60 /* Define a reasonable max cap */
  61 #define CORE_FILE_NOTE_SIZE_MAX (16*1024*1024)
  62
  63 static int core_uses_pid;
  64 static unsigned int core_pipe_limit;
  65 static char core_pattern[CORENAME_MAX_SIZE] = "core";
  66 static int core_name_size = CORENAME_MAX_SIZE;
  67 unsigned int core_file_note_size_limit = CORE_FILE_NOTE_SIZE_DEFAULT;
  68
  69 struct core_name {
  70         char *corename;
  71         int used, size;
  72 };
  73
  74 static int expand_corename(struct core_name *cn, int size)
  75 {
  76         char *corename;
  77
  78         size = kmalloc_size_roundup(size);
  79         corename = krealloc(cn->corename, size, GFP_KERNEL);
  80
  81         if (!corename)
  82                 return -ENOMEM;
  83
  84         if (size > core_name_size) /* racy but harmless */
  85                 core_name_size = size;
  86
  87         cn->size = size;
  88         cn->corename = corename;
  89         return 0;
  90 }
  91
  92 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  93                                      va_list arg)
  94 {
  95         int free, need;
  96         va_list arg_copy;
  97
  98 again:
  99         free = cn->size - cn->used;
 100
 101         va_copy(arg_copy, arg);
 102         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
 103         va_end(arg_copy);
 104
 105         if (need < free) {
 106                 cn->used += need;
 107                 return 0;
 108         }
 109
 110         if (!expand_corename(cn, cn->size + need - free + 1))
 111                 goto again;
 112
 113         return -ENOMEM;
 114 }
 115
 116 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 117 {
 118         va_list arg;
 119         int ret;
 120
 121         va_start(arg, fmt);
 122         ret = cn_vprintf(cn, fmt, arg);
 123         va_end(arg);
 124
 125         return ret;
 126 }
 127
 128 static __printf(2, 3)
 129 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 130 {
 131         int cur = cn->used;
 132         va_list arg;
 133         int ret;
 134
 135         va_start(arg, fmt);
 136         ret = cn_vprintf(cn, fmt, arg);
 137         va_end(arg);
 138
 139         if (ret == 0) {
 140                 /*
 141                  * Ensure that this coredump name component can't cause the
 142                  * resulting corefile path to consist of a ".." or ".".
 143                  */
 144                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 145                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 146                                 && cn->corename[cur+1] == '.'))
 147                         cn->corename[cur] = '!';
 148
 149                 /*
 150                  * Empty names are fishy and could be used to create a "//" in a
 151                  * corefile name, causing the coredump to happen one directory
 152                  * level too high. Enforce that all components of the core
 153                  * pattern are at least one character long.
 154                  */
 155                 if (cn->used == cur)
 156                         ret = cn_printf(cn, "!");
 157         }
 158
 159         for (; cur < cn->used; ++cur) {
 160                 if (cn->corename[cur] == '/')
 161                         cn->corename[cur] = '!';
 162         }
 163         return ret;
 164 }
 165
 166 static int cn_print_exe_file(struct core_name *cn, bool name_only)
 167 {
 168         struct file *exe_file;
 169         char *pathbuf, *path, *ptr;
 170         int ret;
 171
 172         exe_file = get_mm_exe_file(current->mm);
 173         if (!exe_file)
 174                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 175
 176         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
 177         if (!pathbuf) {
 178                 ret = -ENOMEM;
 179                 goto put_exe_file;
 180         }
 181
 182         path = file_path(exe_file, pathbuf, PATH_MAX);
 183         if (IS_ERR(path)) {
 184                 ret = PTR_ERR(path);
 185                 goto free_buf;
 186         }
 187
 188         if (name_only) {
 189                 ptr = strrchr(path, '/');
 190                 if (ptr)
 191                         path = ptr + 1;
 192         }
 193         ret = cn_esc_printf(cn, "%s", path);
 194
 195 free_buf:
 196         kfree(pathbuf);
 197 put_exe_file:
 198         fput(exe_file);
 199         return ret;
 200 }
 201
 202 /* format_corename will inspect the pattern parameter, and output a
 203  * name into corename, which must have space for at least
 204  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 205  */
 206 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
 207                            size_t **argv, int *argc)
 208 {
 209         const struct cred *cred = current_cred();
 210         const char *pat_ptr = core_pattern;
 211         int ispipe = (*pat_ptr == '|');
 212         bool was_space = false;
 213         int pid_in_pattern = 0;
 214         int err = 0;
 215
 216         cn->used = 0;
 217         cn->corename = NULL;
 218         if (expand_corename(cn, core_name_size))
 219                 return -ENOMEM;
 220         cn->corename[0] = '\0';
 221
 222         if (ispipe) {
 223                 int argvs = sizeof(core_pattern) / 2;
 224                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
 225                 if (!(*argv))
 226                         return -ENOMEM;
 227                 (*argv)[(*argc)++] = 0;
 228                 ++pat_ptr;
 229                 if (!(*pat_ptr))
 230                         return -ENOMEM;
 231         }
 232
 233         /* Repeat as long as we have more pattern to process and more output
 234            space */
 235         while (*pat_ptr) {
 236                 /*
 237                  * Split on spaces before doing template expansion so that
 238                  * %e and %E don't get split if they have spaces in them
 239                  */
 240                 if (ispipe) {
 241                         if (isspace(*pat_ptr)) {
 242                                 if (cn->used != 0)
 243                                         was_space = true;
 244                                 pat_ptr++;
 245                                 continue;
 246                         } else if (was_space) {
 247                                 was_space = false;
 248                                 err = cn_printf(cn, "%c", '\0');
 249                                 if (err)
 250                                         return err;
 251                                 (*argv)[(*argc)++] = cn->used;
 252                         }
 253                 }
 254                 if (*pat_ptr != '%') {
 255                         err = cn_printf(cn, "%c", *pat_ptr++);
 256                 } else {
 257                         switch (*++pat_ptr) {
 258                         /* single % at the end, drop that */
 259                         case 0:
 260                                 goto out;
 261                         /* Double percent, output one percent */
 262                         case '%':
 263                                 err = cn_printf(cn, "%c", '%');
 264                                 break;
 265                         /* pid */
 266                         case 'p':
 267                                 pid_in_pattern = 1;
 268                                 err = cn_printf(cn, "%d",
 269                                               task_tgid_vnr(current));
 270                                 break;
 271                         /* global pid */
 272                         case 'P':
 273                                 err = cn_printf(cn, "%d",
 274                                               task_tgid_nr(current));
 275                                 break;
 276                         case 'i':
 277                                 err = cn_printf(cn, "%d",
 278                                               task_pid_vnr(current));
 279                                 break;
 280                         case 'I':
 281                                 err = cn_printf(cn, "%d",
 282                                               task_pid_nr(current));
 283                                 break;
 284                         /* uid */
 285                         case 'u':
 286                                 err = cn_printf(cn, "%u",
 287                                                 from_kuid(&init_user_ns,
 288                                                           cred->uid));
 289                                 break;
 290                         /* gid */
 291                         case 'g':
 292                                 err = cn_printf(cn, "%u",
 293                                                 from_kgid(&init_user_ns,
 294                                                           cred->gid));
 295                                 break;
 296                         case 'd':
 297                                 err = cn_printf(cn, "%d",
 298                                         __get_dumpable(cprm->mm_flags));
 299                                 break;
 300                         /* signal that caused the coredump */
 301                         case 's':
 302                                 err = cn_printf(cn, "%d",
 303                                                 cprm->siginfo->si_signo);
 304                                 break;
 305                         /* UNIX time of coredump */
 306                         case 't': {
 307                                 time64_t time;
 308
 309                                 time = ktime_get_real_seconds();
 310                                 err = cn_printf(cn, "%lld", time);
 311                                 break;
 312                         }
 313                         /* hostname */
 314                         case 'h':
 315                                 down_read(&uts_sem);
 316                                 err = cn_esc_printf(cn, "%s",
 317                                               utsname()->nodename);
 318                                 up_read(&uts_sem);
 319                                 break;
 320                         /* executable, could be changed by prctl PR_SET_NAME etc */
 321                         case 'e':
 322                                 err = cn_esc_printf(cn, "%s", current->comm);
 323                                 break;
 324                         /* file name of executable */
 325                         case 'f':
 326                                 err = cn_print_exe_file(cn, true);
 327                                 break;
 328                         case 'E':
 329                                 err = cn_print_exe_file(cn, false);
 330                                 break;
 331                         /* core limit size */
 332                         case 'c':
 333                                 err = cn_printf(cn, "%lu",
 334                                               rlimit(RLIMIT_CORE));
 335                                 break;
 336                         /* CPU the task ran on */
 337                         case 'C':
 338                                 err = cn_printf(cn, "%d", cprm->cpu);
 339                                 break;
 340                         default:
 341                                 break;
 342                         }
 343                         ++pat_ptr;
 344                 }
 345
 346                 if (err)
 347                         return err;
 348         }
 349
 350 out:
 351         /* Backward compatibility with core_uses_pid:
 352          *
 353          * If core_pattern does not include a %p (as is the default)
 354          * and core_uses_pid is set, then .%pid will be appended to
 355          * the filename. Do not do this for piped commands. */
 356         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 357                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 358                 if (err)
 359                         return err;
 360         }
 361         return ispipe;
 362 }
 363
 364 static int zap_process(struct signal_struct *signal, int exit_code)
 365 {
 366         struct task_struct *t;
 367         int nr = 0;
 368
 369         signal->flags = SIGNAL_GROUP_EXIT;
 370         signal->group_exit_code = exit_code;
 371         signal->group_stop_count = 0;
 372
 373         __for_each_thread(signal, t) {
 374                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 375                 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
 376                         sigaddset(&t->pending.signal, SIGKILL);
 377                         signal_wake_up(t, 1);
 378                         nr++;
 379                 }
 380         }
 381
 382         return nr;
 383 }
 384
 385 static int zap_threads(struct task_struct *tsk,
 386                         struct core_state *core_state, int exit_code)
 387 {
 388         struct signal_struct *signal = tsk->signal;
 389         int nr = -EAGAIN;
 390
 391         spin_lock_irq(&tsk->sighand->siglock);
 392         if (!(signal->flags & SIGNAL_GROUP_EXIT) && !signal->group_exec_task) {
 393                 /* Allow SIGKILL, see prepare_signal() */
 394                 signal->core_state = core_state;
 395                 nr = zap_process(signal, exit_code);
 396                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 397                 tsk->flags |= PF_DUMPCORE;
 398                 atomic_set(&core_state->nr_threads, nr);
 399         }
 400         spin_unlock_irq(&tsk->sighand->siglock);
 401         return nr;
 402 }
 403
 404 static int coredump_wait(int exit_code, struct core_state *core_state)
 405 {
 406         struct task_struct *tsk = current;
 407         int core_waiters = -EBUSY;
 408
 409         init_completion(&core_state->startup);
 410         core_state->dumper.task = tsk;
 411         core_state->dumper.next = NULL;
 412
 413         core_waiters = zap_threads(tsk, core_state, exit_code);
 414         if (core_waiters > 0) {
 415                 struct core_thread *ptr;
 416
 417                 wait_for_completion_state(&core_state->startup,
 418                                           TASK_UNINTERRUPTIBLE|TASK_FREEZABLE);
 419                 /*
 420                  * Wait for all the threads to become inactive, so that
 421                  * all the thread context (extended register state, like
 422                  * fpu etc) gets copied to the memory.
 423                  */
 424                 ptr = core_state->dumper.next;
 425                 while (ptr != NULL) {
 426                         wait_task_inactive(ptr->task, TASK_ANY);
 427                         ptr = ptr->next;
 428                 }
 429         }
 430
 431         return core_waiters;
 432 }
 433
 434 static void coredump_finish(bool core_dumped)
 435 {
 436         struct core_thread *curr, *next;
 437         struct task_struct *task;
 438
 439         spin_lock_irq(&current->sighand->siglock);
 440         if (core_dumped && !__fatal_signal_pending(current))
 441                 current->signal->group_exit_code |= 0x80;
 442         next = current->signal->core_state->dumper.next;
 443         current->signal->core_state = NULL;
 444         spin_unlock_irq(&current->sighand->siglock);
 445
 446         while ((curr = next) != NULL) {
 447                 next = curr->next;
 448                 task = curr->task;
 449                 /*
 450                  * see coredump_task_exit(), curr->task must not see
 451                  * ->task == NULL before we read ->next.
 452                  */
 453                 smp_mb();
 454                 curr->task = NULL;
 455                 wake_up_process(task);
 456         }
 457 }
 458
 459 static bool dump_interrupted(void)
 460 {
 461         /*
 462          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 463          * can do try_to_freeze() and check __fatal_signal_pending(),
 464          * but then we need to teach dump_write() to restart and clear
 465          * TIF_SIGPENDING.
 466          */
 467         return fatal_signal_pending(current) || freezing(current);
 468 }
 469
 470 static void wait_for_dump_helpers(struct file *file)
 471 {
 472         struct pipe_inode_info *pipe = file->private_data;
 473
 474         pipe_lock(pipe);
 475         pipe->readers++;
 476         pipe->writers--;
 477         wake_up_interruptible_sync(&pipe->rd_wait);
 478         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 479         pipe_unlock(pipe);
 480
 481         /*
 482          * We actually want wait_event_freezable() but then we need
 483          * to clear TIF_SIGPENDING and improve dump_interrupted().
 484          */
 485         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
 486
 487         pipe_lock(pipe);
 488         pipe->readers--;
 489         pipe->writers++;
 490         pipe_unlock(pipe);
 491 }
 492
 493 /*
 494  * umh_pipe_setup
 495  * helper function to customize the process used
 496  * to collect the core in userspace.  Specifically
 497  * it sets up a pipe and installs it as fd 0 (stdin)
 498  * for the process.  Returns 0 on success, or
 499  * PTR_ERR on failure.
 500  * Note that it also sets the core limit to 1.  This
 501  * is a special value that we use to trap recursive
 502  * core dumps
 503  */
 504 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 505 {
 506         struct file *files[2];
 507         struct coredump_params *cp = (struct coredump_params *)info->data;
 508         int err = create_pipe_files(files, 0);
 509         if (err)
 510                 return err;
 511
 512         cp->file = files[1];
 513
 514         err = replace_fd(0, files[0], 0);
 515         fput(files[0]);
 516         /* and disallow core files too */
 517         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 518
 519         return err;
 520 }
 521
 522 void do_coredump(const kernel_siginfo_t *siginfo)
 523 {
 524         struct core_state core_state;
 525         struct core_name cn;
 526         struct mm_struct *mm = current->mm;
 527         struct linux_binfmt * binfmt;
 528         const struct cred *old_cred;
 529         struct cred *cred;
 530         int retval = 0;
 531         int ispipe;
 532         size_t *argv = NULL;
 533         int argc = 0;
 534         /* require nonrelative corefile path and be extra careful */
 535         bool need_suid_safe = false;
 536         bool core_dumped = false;
 537         static atomic_t core_dump_count = ATOMIC_INIT(0);
 538         struct coredump_params cprm = {
 539                 .siginfo = siginfo,
 540                 .limit = rlimit(RLIMIT_CORE),
 541                 /*
 542                  * We must use the same mm->flags while dumping core to avoid
 543                  * inconsistency of bit flags, since this flag is not protected
 544                  * by any locks.
 545                  */
 546                 .mm_flags = mm->flags,
 547                 .vma_meta = NULL,
 548                 .cpu = raw_smp_processor_id(),
 549         };
 550
 551         audit_core_dumps(siginfo->si_signo);
 552
 553         binfmt = mm->binfmt;
 554         if (!binfmt || !binfmt->core_dump)
 555                 goto fail;
 556         if (!__get_dumpable(cprm.mm_flags))
 557                 goto fail;
 558
 559         cred = prepare_creds();
 560         if (!cred)
 561                 goto fail;
 562         /*
 563          * We cannot trust fsuid as being the "true" uid of the process
 564          * nor do we know its entire history. We only know it was tainted
 565          * so we dump it as root in mode 2, and only into a controlled
 566          * environment (pipe handler or fully qualified path).
 567          */
 568         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 569                 /* Setuid core dump mode */
 570                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 571                 need_suid_safe = true;
 572         }
 573
 574         retval = coredump_wait(siginfo->si_signo, &core_state);
 575         if (retval < 0)
 576                 goto fail_creds;
 577
 578         old_cred = override_creds(cred);
 579
 580         ispipe = format_corename(&cn, &cprm, &argv, &argc);
 581
 582         if (ispipe) {
 583                 int argi;
 584                 int dump_count;
 585                 char **helper_argv;
 586                 struct subprocess_info *sub_info;
 587
 588                 if (ispipe < 0) {
 589                         printk(KERN_WARNING "format_corename failed\n");
 590                         printk(KERN_WARNING "Aborting core\n");
 591                         goto fail_unlock;
 592                 }
 593
 594                 if (cprm.limit == 1) {
 595                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 596                          *
 597                          * Normally core limits are irrelevant to pipes, since
 598                          * we're not writing to the file system, but we use
 599                          * cprm.limit of 1 here as a special value, this is a
 600                          * consistent way to catch recursive crashes.
 601                          * We can still crash if the core_pattern binary sets
 602                          * RLIM_CORE = !1, but it runs as root, and can do
 603                          * lots of stupid things.
 604                          *
 605                          * Note that we use task_tgid_vnr here to grab the pid
 606                          * of the process group leader.  That way we get the
 607                          * right pid if a thread in a multi-threaded
 608                          * core_pattern process dies.
 609                          */
 610                         printk(KERN_WARNING
 611                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 612                                 task_tgid_vnr(current), current->comm);
 613                         printk(KERN_WARNING "Aborting core\n");
 614                         goto fail_unlock;
 615                 }
 616                 cprm.limit = RLIM_INFINITY;
 617
 618                 dump_count = atomic_inc_return(&core_dump_count);
 619                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 620                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 621                                task_tgid_vnr(current), current->comm);
 622                         printk(KERN_WARNING "Skipping core dump\n");
 623                         goto fail_dropcount;
 624                 }
 625
 626                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
 627                                             GFP_KERNEL);
 628                 if (!helper_argv) {
 629                         printk(KERN_WARNING "%s failed to allocate memory\n",
 630                                __func__);
 631                         goto fail_dropcount;
 632                 }
 633                 for (argi = 0; argi < argc; argi++)
 634                         helper_argv[argi] = cn.corename + argv[argi];
 635                 helper_argv[argi] = NULL;
 636
 637                 retval = -ENOMEM;
 638                 sub_info = call_usermodehelper_setup(helper_argv[0],
 639                                                 helper_argv, NULL, GFP_KERNEL,
 640                                                 umh_pipe_setup, NULL, &cprm);
 641                 if (sub_info)
 642                         retval = call_usermodehelper_exec(sub_info,
 643                                                           UMH_WAIT_EXEC);
 644
 645                 kfree(helper_argv);
 646                 if (retval) {
 647                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 648                                cn.corename);
 649                         goto close_fail;
 650                 }
 651         } else {
 652                 struct mnt_idmap *idmap;
 653                 struct inode *inode;
 654                 int open_flags = O_CREAT | O_WRONLY | O_NOFOLLOW |
 655                                  O_LARGEFILE | O_EXCL;
 656
 657                 if (cprm.limit < binfmt->min_coredump)
 658                         goto fail_unlock;
 659
 660                 if (need_suid_safe && cn.corename[0] != '/') {
 661                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 662                                 "to fully qualified path!\n",
 663                                 task_tgid_vnr(current), current->comm);
 664                         printk(KERN_WARNING "Skipping core dump\n");
 665                         goto fail_unlock;
 666                 }
 667
 668                 /*
 669                  * Unlink the file if it exists unless this is a SUID
 670                  * binary - in that case, we're running around with root
 671                  * privs and don't want to unlink another user's coredump.
 672                  */
 673                 if (!need_suid_safe) {
 674                         /*
 675                          * If it doesn't exist, that's fine. If there's some
 676                          * other problem, we'll catch it at the filp_open().
 677                          */
 678                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
 679                 }
 680
 681                 /*
 682                  * There is a race between unlinking and creating the
 683                  * file, but if that causes an EEXIST here, that's
 684                  * fine - another process raced with us while creating
 685                  * the corefile, and the other process won. To userspace,
 686                  * what matters is that at least one of the two processes
 687                  * writes its coredump successfully, not which one.
 688                  */
 689                 if (need_suid_safe) {
 690                         /*
 691                          * Using user namespaces, normal user tasks can change
 692                          * their current->fs->root to point to arbitrary
 693                          * directories. Since the intention of the "only dump
 694                          * with a fully qualified path" rule is to control where
 695                          * coredumps may be placed using root privileges,
 696                          * current->fs->root must not be used. Instead, use the
 697                          * root directory of init_task.
 698                          */
 699                         struct path root;
 700
 701                         task_lock(&init_task);
 702                         get_fs_root(init_task.fs, &root);
 703                         task_unlock(&init_task);
 704                         cprm.file = file_open_root(&root, cn.corename,
 705                                                    open_flags, 0600);
 706                         path_put(&root);
 707                 } else {
 708                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 709                 }
 710                 if (IS_ERR(cprm.file))
 711                         goto fail_unlock;
 712
 713                 inode = file_inode(cprm.file);
 714                 if (inode->i_nlink > 1)
 715                         goto close_fail;
 716                 if (d_unhashed(cprm.file->f_path.dentry))
 717                         goto close_fail;
 718                 /*
 719                  * AK: actually i see no reason to not allow this for named
 720                  * pipes etc, but keep the previous behaviour for now.
 721                  */
 722                 if (!S_ISREG(inode->i_mode))
 723                         goto close_fail;
 724                 /*
 725                  * Don't dump core if the filesystem changed owner or mode
 726                  * of the file during file creation. This is an issue when
 727                  * a process dumps core while its cwd is e.g. on a vfat
 728                  * filesystem.
 729                  */
 730                 idmap = file_mnt_idmap(cprm.file);
 731                 if (!vfsuid_eq_kuid(i_uid_into_vfsuid(idmap, inode),
 732                                     current_fsuid())) {
 733                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file owner\n",
 734                                             cn.corename);
 735                         goto close_fail;
 736                 }
 737                 if ((inode->i_mode & 0677) != 0600) {
 738                         pr_info_ratelimited("Core dump to %s aborted: cannot preserve file permissions\n",
 739                                             cn.corename);
 740                         goto close_fail;
 741                 }
 742                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 743                         goto close_fail;
 744                 if (do_truncate(idmap, cprm.file->f_path.dentry,
 745                                 0, 0, cprm.file))
 746                         goto close_fail;
 747         }
 748
 749         /* get us an unshared descriptor table; almost always a no-op */
 750         /* The cell spufs coredump code reads the file descriptor tables */
 751         retval = unshare_files();
 752         if (retval)
 753                 goto close_fail;
 754         if (!dump_interrupted()) {
 755                 /*
 756                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
 757                  * have this set to NULL.
 758                  */
 759                 if (!cprm.file) {
 760                         pr_info("Core dump to |%s disabled\n", cn.corename);
 761                         goto close_fail;
 762                 }
 763                 if (!dump_vma_snapshot(&cprm))
 764                         goto close_fail;
 765
 766                 file_start_write(cprm.file);
 767                 core_dumped = binfmt->core_dump(&cprm);
 768                 /*
 769                  * Ensures that file size is big enough to contain the current
 770                  * file postion. This prevents gdb from complaining about
 771                  * a truncated file if the last "write" to the file was
 772                  * dump_skip.
 773                  */
 774                 if (cprm.to_skip) {
 775                         cprm.to_skip--;
 776                         dump_emit(&cprm, "", 1);
 777                 }
 778                 file_end_write(cprm.file);
 779                 free_vma_snapshot(&cprm);
 780         }
 781         if (ispipe && core_pipe_limit)
 782                 wait_for_dump_helpers(cprm.file);
 783 close_fail:
 784         if (cprm.file)
 785                 filp_close(cprm.file, NULL);
 786 fail_dropcount:
 787         if (ispipe)
 788                 atomic_dec(&core_dump_count);
 789 fail_unlock:
 790         kfree(argv);
 791         kfree(cn.corename);
 792         coredump_finish(core_dumped);
 793         revert_creds(old_cred);
 794 fail_creds:
 795         put_cred(cred);
 796 fail:
 797         return;
 798 }
 799
 800 /*
 801  * Core dumping helper functions.  These are the only things you should
 802  * do on a core-file: use only these functions to write out all the
 803  * necessary info.
 804  */
 805 static int __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 806 {
 807         struct file *file = cprm->file;
 808         loff_t pos = file->f_pos;
 809         ssize_t n;
 810         if (cprm->written + nr > cprm->limit)
 811                 return 0;
 812
 813
 814         if (dump_interrupted())
 815                 return 0;
 816         n = __kernel_write(file, addr, nr, &pos);
 817         if (n != nr)
 818                 return 0;
 819         file->f_pos = pos;
 820         cprm->written += n;
 821         cprm->pos += n;
 822
 823         return 1;
 824 }
 825
 826 static int __dump_skip(struct coredump_params *cprm, size_t nr)
 827 {
 828         static char zeroes[PAGE_SIZE];
 829         struct file *file = cprm->file;
 830         if (file->f_mode & FMODE_LSEEK) {
 831                 if (dump_interrupted() ||
 832                     vfs_llseek(file, nr, SEEK_CUR) < 0)
 833                         return 0;
 834                 cprm->pos += nr;
 835                 return 1;
 836         } else {
 837                 while (nr > PAGE_SIZE) {
 838                         if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
 839                                 return 0;
 840                         nr -= PAGE_SIZE;
 841                 }
 842                 return __dump_emit(cprm, zeroes, nr);
 843         }
 844 }
 845
 846 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 847 {
 848         if (cprm->to_skip) {
 849                 if (!__dump_skip(cprm, cprm->to_skip))
 850                         return 0;
 851                 cprm->to_skip = 0;
 852         }
 853         return __dump_emit(cprm, addr, nr);
 854 }
 855 EXPORT_SYMBOL(dump_emit);
 856
 857 void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
 858 {
 859         cprm->to_skip = pos - cprm->pos;
 860 }
 861 EXPORT_SYMBOL(dump_skip_to);
 862
 863 void dump_skip(struct coredump_params *cprm, size_t nr)
 864 {
 865         cprm->to_skip += nr;
 866 }
 867 EXPORT_SYMBOL(dump_skip);
 868
 869 #ifdef CONFIG_ELF_CORE
 870 static int dump_emit_page(struct coredump_params *cprm, struct page *page)
 871 {
 872         struct bio_vec bvec;
 873         struct iov_iter iter;
 874         struct file *file = cprm->file;
 875         loff_t pos;
 876         ssize_t n;
 877
 878         if (!page)
 879                 return 0;
 880
 881         if (cprm->to_skip) {
 882                 if (!__dump_skip(cprm, cprm->to_skip))
 883                         return 0;
 884                 cprm->to_skip = 0;
 885         }
 886         if (cprm->written + PAGE_SIZE > cprm->limit)
 887                 return 0;
 888         if (dump_interrupted())
 889                 return 0;
 890         pos = file->f_pos;
 891         bvec_set_page(&bvec, page, PAGE_SIZE, 0);
 892         iov_iter_bvec(&iter, ITER_SOURCE, &bvec, 1, PAGE_SIZE);
 893         n = __kernel_write_iter(cprm->file, &iter, &pos);
 894         if (n != PAGE_SIZE)
 895                 return 0;
 896         file->f_pos = pos;
 897         cprm->written += PAGE_SIZE;
 898         cprm->pos += PAGE_SIZE;
 899
 900         return 1;
 901 }
 902
 903 /*
 904  * If we might get machine checks from kernel accesses during the
 905  * core dump, let's get those errors early rather than during the
 906  * IO. This is not performance-critical enough to warrant having
 907  * all the machine check logic in the iovec paths.
 908  */
 909 #ifdef copy_mc_to_kernel
 910
 911 #define dump_page_alloc() alloc_page(GFP_KERNEL)
 912 #define dump_page_free(x) __free_page(x)
 913 static struct page *dump_page_copy(struct page *src, struct page *dst)
 914 {
 915         void *buf = kmap_local_page(src);
 916         size_t left = copy_mc_to_kernel(page_address(dst), buf, PAGE_SIZE);
 917         kunmap_local(buf);
 918         return left ? NULL : dst;
 919 }
 920
 921 #else
 922
 923 /* We just want to return non-NULL; it's never used. */
 924 #define dump_page_alloc() ERR_PTR(-EINVAL)
 925 #define dump_page_free(x) ((void)(x))
 926 static inline struct page *dump_page_copy(struct page *src, struct page *dst)
 927 {
 928         return src;
 929 }
 930 #endif
 931
 932 int dump_user_range(struct coredump_params *cprm, unsigned long start,
 933                     unsigned long len)
 934 {
 935         unsigned long addr;
 936         struct page *dump_page;
 937
 938         dump_page = dump_page_alloc();
 939         if (!dump_page)
 940                 return 0;
 941
 942         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
 943                 struct page *page;
 944
 945                 /*
 946                  * To avoid having to allocate page tables for virtual address
 947                  * ranges that have never been used yet, and also to make it
 948                  * easy to generate sparse core files, use a helper that returns
 949                  * NULL when encountering an empty page table entry that would
 950                  * otherwise have been filled with the zero page.
 951                  */
 952                 page = get_dump_page(addr);
 953                 if (page) {
 954                         int stop = !dump_emit_page(cprm, dump_page_copy(page, dump_page));
 955                         put_page(page);
 956                         if (stop) {
 957                                 dump_page_free(dump_page);
 958                                 return 0;
 959                         }
 960                 } else {
 961                         dump_skip(cprm, PAGE_SIZE);
 962                 }
 963         }
 964         dump_page_free(dump_page);
 965         return 1;
 966 }
 967 #endif
 968
 969 int dump_align(struct coredump_params *cprm, int align)
 970 {
 971         unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
 972         if (align & (align - 1))
 973                 return 0;
 974         if (mod)
 975                 cprm->to_skip += align - mod;
 976         return 1;
 977 }
 978 EXPORT_SYMBOL(dump_align);
 979
 980 #ifdef CONFIG_SYSCTL
 981
 982 void validate_coredump_safety(void)
 983 {
 984         if (suid_dumpable == SUID_DUMP_ROOT &&
 985             core_pattern[0] != '/' && core_pattern[0] != '|') {
 986                 pr_warn(
 987 "Unsafe core_pattern used with fs.suid_dumpable=2.\n"
 988 "Pipe handler or fully qualified core dump path required.\n"
 989 "Set kernel.core_pattern before fs.suid_dumpable.\n"
 990                 );
 991         }
 992 }
 993
 994 static int proc_dostring_coredump(const struct ctl_table *table, int write,
 995                   void *buffer, size_t *lenp, loff_t *ppos)
 996 {
 997         int error = proc_dostring(table, write, buffer, lenp, ppos);
 998
 999         if (!error)
1000                 validate_coredump_safety();
1001         return error;
1002 }
1003
1004 static const unsigned int core_file_note_size_min = CORE_FILE_NOTE_SIZE_DEFAULT;
1005 static const unsigned int core_file_note_size_max = CORE_FILE_NOTE_SIZE_MAX;
1006
1007 static struct ctl_table coredump_sysctls[] = {
1008         {
1009                 .procname       = "core_uses_pid",
1010                 .data           = &core_uses_pid,
1011                 .maxlen         = sizeof(int),
1012                 .mode           = 0644,
1013                 .proc_handler   = proc_dointvec,
1014         },
1015         {
1016                 .procname       = "core_pattern",
1017                 .data           = core_pattern,
1018                 .maxlen         = CORENAME_MAX_SIZE,
1019                 .mode           = 0644,
1020                 .proc_handler   = proc_dostring_coredump,
1021         },
1022         {
1023                 .procname       = "core_pipe_limit",
1024                 .data           = &core_pipe_limit,
1025                 .maxlen         = sizeof(unsigned int),
1026                 .mode           = 0644,
1027                 .proc_handler   = proc_dointvec,
1028         },
1029         {
1030                 .procname       = "core_file_note_size_limit",
1031                 .data           = &core_file_note_size_limit,
1032                 .maxlen         = sizeof(unsigned int),
1033                 .mode           = 0644,
1034                 .proc_handler   = proc_douintvec_minmax,
1035                 .extra1         = (unsigned int *)&core_file_note_size_min,
1036                 .extra2         = (unsigned int *)&core_file_note_size_max,
1037         },
1038 };
1039
1040 static int __init init_fs_coredump_sysctls(void)
1041 {
1042         register_sysctl_init("kernel", coredump_sysctls);
1043         return 0;
1044 }
1045 fs_initcall(init_fs_coredump_sysctls);
1046 #endif /* CONFIG_SYSCTL */
1047
1048 /*
1049  * The purpose of always_dump_vma() is to make sure that special kernel mappings
1050  * that are useful for post-mortem analysis are included in every core dump.
1051  * In that way we ensure that the core dump is fully interpretable later
1052  * without matching up the same kernel and hardware config to see what PC values
1053  * meant. These special mappings include - vDSO, vsyscall, and other
1054  * architecture specific mappings
1055  */
1056 static bool always_dump_vma(struct vm_area_struct *vma)
1057 {
1058         /* Any vsyscall mappings? */
1059         if (vma == get_gate_vma(vma->vm_mm))
1060                 return true;
1061
1062         /*
1063          * Assume that all vmas with a .name op should always be dumped.
1064          * If this changes, a new vm_ops field can easily be added.
1065          */
1066         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
1067                 return true;
1068
1069         /*
1070          * arch_vma_name() returns non-NULL for special architecture mappings,
1071          * such as vDSO sections.
1072          */
1073         if (arch_vma_name(vma))
1074                 return true;
1075
1076         return false;
1077 }
1078
1079 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
1080
1081 /*
1082  * Decide how much of @vma's contents should be included in a core dump.
1083  */
1084 static unsigned long vma_dump_size(struct vm_area_struct *vma,
1085                                    unsigned long mm_flags)
1086 {
1087 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
1088
1089         /* always dump the vdso and vsyscall sections */
1090         if (always_dump_vma(vma))
1091                 goto whole;
1092
1093         if (vma->vm_flags & VM_DONTDUMP)
1094                 return 0;
1095
1096         /* support for DAX */
1097         if (vma_is_dax(vma)) {
1098                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1099                         goto whole;
1100                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1101                         goto whole;
1102                 return 0;
1103         }
1104
1105         /* Hugetlb memory check */
1106         if (is_vm_hugetlb_page(vma)) {
1107                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1108                         goto whole;
1109                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1110                         goto whole;
1111                 return 0;
1112         }
1113
1114         /* Do not dump I/O mapped devices or special mappings */
1115         if (vma->vm_flags & VM_IO)
1116                 return 0;
1117
1118         /* By default, dump shared memory if mapped from an anonymous file. */
1119         if (vma->vm_flags & VM_SHARED) {
1120                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1121                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1122                         goto whole;
1123                 return 0;
1124         }
1125
1126         /* Dump segments that have been written to.  */
1127         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1128                 goto whole;
1129         if (vma->vm_file == NULL)
1130                 return 0;
1131
1132         if (FILTER(MAPPED_PRIVATE))
1133                 goto whole;
1134
1135         /*
1136          * If this is the beginning of an executable file mapping,
1137          * dump the first page to aid in determining what was mapped here.
1138          */
1139         if (FILTER(ELF_HEADERS) &&
1140             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1141                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1142                         return PAGE_SIZE;
1143
1144                 /*
1145                  * ELF libraries aren't always executable.
1146                  * We'll want to check whether the mapping starts with the ELF
1147                  * magic, but not now - we're holding the mmap lock,
1148                  * so copy_from_user() doesn't work here.
1149                  * Use a placeholder instead, and fix it up later in
1150                  * dump_vma_snapshot().
1151                  */
1152                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1153         }
1154
1155 #undef  FILTER
1156
1157         return 0;
1158
1159 whole:
1160         return vma->vm_end - vma->vm_start;
1161 }
1162
1163 /*
1164  * Helper function for iterating across a vma list.  It ensures that the caller
1165  * will visit `gate_vma' prior to terminating the search.
1166  */
1167 static struct vm_area_struct *coredump_next_vma(struct vma_iterator *vmi,
1168                                        struct vm_area_struct *vma,
1169                                        struct vm_area_struct *gate_vma)
1170 {
1171         if (gate_vma && (vma == gate_vma))
1172                 return NULL;
1173
1174         vma = vma_next(vmi);
1175         if (vma)
1176                 return vma;
1177         return gate_vma;
1178 }
1179
1180 static void free_vma_snapshot(struct coredump_params *cprm)
1181 {
1182         if (cprm->vma_meta) {
1183                 int i;
1184                 for (i = 0; i < cprm->vma_count; i++) {
1185                         struct file *file = cprm->vma_meta[i].file;
1186                         if (file)
1187                                 fput(file);
1188                 }
1189                 kvfree(cprm->vma_meta);
1190                 cprm->vma_meta = NULL;
1191         }
1192 }
1193
1194 /*
1195  * Under the mmap_lock, take a snapshot of relevant information about the task's
1196  * VMAs.
1197  */
1198 static bool dump_vma_snapshot(struct coredump_params *cprm)
1199 {
1200         struct vm_area_struct *gate_vma, *vma = NULL;
1201         struct mm_struct *mm = current->mm;
1202         VMA_ITERATOR(vmi, mm, 0);
1203         int i = 0;
1204
1205         /*
1206          * Once the stack expansion code is fixed to not change VMA bounds
1207          * under mmap_lock in read mode, this can be changed to take the
1208          * mmap_lock in read mode.
1209          */
1210         if (mmap_write_lock_killable(mm))
1211                 return false;
1212
1213         cprm->vma_data_size = 0;
1214         gate_vma = get_gate_vma(mm);
1215         cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1216
1217         cprm->vma_meta = kvmalloc_array(cprm->vma_count, sizeof(*cprm->vma_meta), GFP_KERNEL);
1218         if (!cprm->vma_meta) {
1219                 mmap_write_unlock(mm);
1220                 return false;
1221         }
1222
1223         while ((vma = coredump_next_vma(&vmi, vma, gate_vma)) != NULL) {
1224                 struct core_vma_metadata *m = cprm->vma_meta + i;
1225
1226                 m->start = vma->vm_start;
1227                 m->end = vma->vm_end;
1228                 m->flags = vma->vm_flags;
1229                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1230                 m->pgoff = vma->vm_pgoff;
1231                 m->file = vma->vm_file;
1232                 if (m->file)
1233                         get_file(m->file);
1234                 i++;
1235         }
1236
1237         mmap_write_unlock(mm);
1238
1239         for (i = 0; i < cprm->vma_count; i++) {
1240                 struct core_vma_metadata *m = cprm->vma_meta + i;
1241
1242                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1243                         char elfmag[SELFMAG];
1244
1245                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1246                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1247                                 m->dump_size = 0;
1248                         } else {
1249                                 m->dump_size = PAGE_SIZE;
1250                         }
1251                 }
1252
1253                 cprm->vma_data_size += m->dump_size;
1254         }
1255
1256         return true;
1257 }