fs/coredump.c

   1 #include <linux/slab.h>
   2 #include <linux/file.h>
   3 #include <linux/fdtable.h>
   4 #include <linux/mm.h>
   5 #include <linux/stat.h>
   6 #include <linux/fcntl.h>
   7 #include <linux/swap.h>
   8 #include <linux/string.h>
   9 #include <linux/init.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/perf_event.h>
  12 #include <linux/highmem.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/key.h>
  15 #include <linux/personality.h>
  16 #include <linux/binfmts.h>
  17 #include <linux/coredump.h>
  18 #include <linux/utsname.h>
  19 #include <linux/pid_namespace.h>
  20 #include <linux/module.h>
  21 #include <linux/namei.h>
  22 #include <linux/mount.h>
  23 #include <linux/security.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/tsacct_kern.h>
  26 #include <linux/cn_proc.h>
  27 #include <linux/audit.h>
  28 #include <linux/tracehook.h>
  29 #include <linux/kmod.h>
  30 #include <linux/fsnotify.h>
  31 #include <linux/fs_struct.h>
  32 #include <linux/pipe_fs_i.h>
  33 #include <linux/oom.h>
  34 #include <linux/compat.h>
  35
  36 #include <asm/uaccess.h>
  37 #include <asm/mmu_context.h>
  38 #include <asm/tlb.h>
  39 #include <asm/exec.h>
  40
  41 #include <trace/events/task.h>
  42 #include "internal.h"
  43 #include "coredump.h"
  44
  45 #include <trace/events/sched.h>
  46
  47 int core_uses_pid;
  48 char core_pattern[CORENAME_MAX_SIZE] = "core";
  49 unsigned int core_pipe_limit;
  50
  51 struct core_name {
  52         char *corename;
  53         int used, size;
  54 };
  55 static atomic_t call_count = ATOMIC_INIT(1);
  56
  57 /* The maximal length of core_pattern is also specified in sysctl.c */
  58
  59 static int expand_corename(struct core_name *cn)
  60 {
  61         char *old_corename = cn->corename;
  62
  63         cn->size = CORENAME_MAX_SIZE * atomic_inc_return(&call_count);
  64         cn->corename = krealloc(old_corename, cn->size, GFP_KERNEL);
  65
  66         if (!cn->corename) {
  67                 kfree(old_corename);
  68                 return -ENOMEM;
  69         }
  70
  71         return 0;
  72 }
  73
  74 static int cn_printf(struct core_name *cn, const char *fmt, ...)
  75 {
  76         char *cur;
  77         int need;
  78         int ret;
  79         va_list arg;
  80
  81         va_start(arg, fmt);
  82         need = vsnprintf(NULL, 0, fmt, arg);
  83         va_end(arg);
  84
  85         if (likely(need < cn->size - cn->used - 1))
  86                 goto out_printf;
  87
  88         ret = expand_corename(cn);
  89         if (ret)
  90                 goto expand_fail;
  91
  92 out_printf:
  93         cur = cn->corename + cn->used;
  94         va_start(arg, fmt);
  95         vsnprintf(cur, need + 1, fmt, arg);
  96         va_end(arg);
  97         cn->used += need;
  98         return 0;
  99
 100 expand_fail:
 101         return ret;
 102 }
 103
 104 static void cn_escape(char *str)
 105 {
 106         for (; *str; str++)
 107                 if (*str == '/')
 108                         *str = '!';
 109 }
 110
 111 static int cn_print_exe_file(struct core_name *cn)
 112 {
 113         struct file *exe_file;
 114         char *pathbuf, *path;
 115         int ret;
 116
 117         exe_file = get_mm_exe_file(current->mm);
 118         if (!exe_file) {
 119                 char *commstart = cn->corename + cn->used;
 120                 ret = cn_printf(cn, "%s (path unknown)", current->comm);
 121                 cn_escape(commstart);
 122                 return ret;
 123         }
 124
 125         pathbuf = kmalloc(PATH_MAX, GFP_TEMPORARY);
 126         if (!pathbuf) {
 127                 ret = -ENOMEM;
 128                 goto put_exe_file;
 129         }
 130
 131         path = d_path(&exe_file->f_path, pathbuf, PATH_MAX);
 132         if (IS_ERR(path)) {
 133                 ret = PTR_ERR(path);
 134                 goto free_buf;
 135         }
 136
 137         cn_escape(path);
 138
 139         ret = cn_printf(cn, "%s", path);
 140
 141 free_buf:
 142         kfree(pathbuf);
 143 put_exe_file:
 144         fput(exe_file);
 145         return ret;
 146 }
 147
 148 /* format_corename will inspect the pattern parameter, and output a
 149  * name into corename, which must have space for at least
 150  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 151  */
 152 static int format_corename(struct core_name *cn, long signr)
 153 {
 154         const struct cred *cred = current_cred();
 155         const char *pat_ptr = core_pattern;
 156         int ispipe = (*pat_ptr == '|');
 157         int pid_in_pattern = 0;
 158         int err = 0;
 159
 160         cn->size = CORENAME_MAX_SIZE * atomic_read(&call_count);
 161         cn->corename = kmalloc(cn->size, GFP_KERNEL);
 162         cn->used = 0;
 163
 164         if (!cn->corename)
 165                 return -ENOMEM;
 166
 167         /* Repeat as long as we have more pattern to process and more output
 168            space */
 169         while (*pat_ptr) {
 170                 if (*pat_ptr != '%') {
 171                         if (*pat_ptr == 0)
 172                                 goto out;
 173                         err = cn_printf(cn, "%c", *pat_ptr++);
 174                 } else {
 175                         switch (*++pat_ptr) {
 176                         /* single % at the end, drop that */
 177                         case 0:
 178                                 goto out;
 179                         /* Double percent, output one percent */
 180                         case '%':
 181                                 err = cn_printf(cn, "%c", '%');
 182                                 break;
 183                         /* pid */
 184                         case 'p':
 185                                 pid_in_pattern = 1;
 186                                 err = cn_printf(cn, "%d",
 187                                               task_tgid_vnr(current));
 188                                 break;
 189                         /* uid */
 190                         case 'u':
 191                                 err = cn_printf(cn, "%d", cred->uid);
 192                                 break;
 193                         /* gid */
 194                         case 'g':
 195                                 err = cn_printf(cn, "%d", cred->gid);
 196                                 break;
 197                         /* signal that caused the coredump */
 198                         case 's':
 199                                 err = cn_printf(cn, "%ld", signr);
 200                                 break;
 201                         /* UNIX time of coredump */
 202                         case 't': {
 203                                 struct timeval tv;
 204                                 do_gettimeofday(&tv);
 205                                 err = cn_printf(cn, "%lu", tv.tv_sec);
 206                                 break;
 207                         }
 208                         /* hostname */
 209                         case 'h': {
 210                                 char *namestart = cn->corename + cn->used;
 211                                 down_read(&uts_sem);
 212                                 err = cn_printf(cn, "%s",
 213                                               utsname()->nodename);
 214                                 up_read(&uts_sem);
 215                                 cn_escape(namestart);
 216                                 break;
 217                         }
 218                         /* executable */
 219                         case 'e': {
 220                                 char *commstart = cn->corename + cn->used;
 221                                 err = cn_printf(cn, "%s", current->comm);
 222                                 cn_escape(commstart);
 223                                 break;
 224                         }
 225                         case 'E':
 226                                 err = cn_print_exe_file(cn);
 227                                 break;
 228                         /* core limit size */
 229                         case 'c':
 230                                 err = cn_printf(cn, "%lu",
 231                                               rlimit(RLIMIT_CORE));
 232                                 break;
 233                         default:
 234                                 break;
 235                         }
 236                         ++pat_ptr;
 237                 }
 238
 239                 if (err)
 240                         return err;
 241         }
 242
 243         /* Backward compatibility with core_uses_pid:
 244          *
 245          * If core_pattern does not include a %p (as is the default)
 246          * and core_uses_pid is set, then .%pid will be appended to
 247          * the filename. Do not do this for piped commands. */
 248         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 249                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 250                 if (err)
 251                         return err;
 252         }
 253 out:
 254         return ispipe;
 255 }
 256
 257 static int zap_process(struct task_struct *start, int exit_code)
 258 {
 259         struct task_struct *t;
 260         int nr = 0;
 261
 262         start->signal->flags = SIGNAL_GROUP_EXIT;
 263         start->signal->group_exit_code = exit_code;
 264         start->signal->group_stop_count = 0;
 265
 266         t = start;
 267         do {
 268                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 269                 if (t != current && t->mm) {
 270                         sigaddset(&t->pending.signal, SIGKILL);
 271                         signal_wake_up(t, 1);
 272                         nr++;
 273                 }
 274         } while_each_thread(start, t);
 275
 276         return nr;
 277 }
 278
 279 static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 280                                 struct core_state *core_state, int exit_code)
 281 {
 282         struct task_struct *g, *p;
 283         unsigned long flags;
 284         int nr = -EAGAIN;
 285
 286         spin_lock_irq(&tsk->sighand->siglock);
 287         if (!signal_group_exit(tsk->signal)) {
 288                 mm->core_state = core_state;
 289                 nr = zap_process(tsk, exit_code);
 290         }
 291         spin_unlock_irq(&tsk->sighand->siglock);
 292         if (unlikely(nr < 0))
 293                 return nr;
 294
 295         if (atomic_read(&mm->mm_users) == nr + 1)
 296                 goto done;
 297         /*
 298          * We should find and kill all tasks which use this mm, and we should
 299          * count them correctly into ->nr_threads. We don't take tasklist
 300          * lock, but this is safe wrt:
 301          *
 302          * fork:
 303          *      None of sub-threads can fork after zap_process(leader). All
 304          *      processes which were created before this point should be
 305          *      visible to zap_threads() because copy_process() adds the new
 306          *      process to the tail of init_task.tasks list, and lock/unlock
 307          *      of ->siglock provides a memory barrier.
 308          *
 309          * do_exit:
 310          *      The caller holds mm->mmap_sem. This means that the task which
 311          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 312          *      its ->mm.
 313          *
 314          * de_thread:
 315          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 316          *      we must see either old or new leader, this does not matter.
 317          *      However, it can change p->sighand, so lock_task_sighand(p)
 318          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
 319          *      it can't fail.
 320          *
 321          *      Note also that "g" can be the old leader with ->mm == NULL
 322          *      and already unhashed and thus removed from ->thread_group.
 323          *      This is OK, __unhash_process()->list_del_rcu() does not
 324          *      clear the ->next pointer, we will find the new leader via
 325          *      next_thread().
 326          */
 327         rcu_read_lock();
 328         for_each_process(g) {
 329                 if (g == tsk->group_leader)
 330                         continue;
 331                 if (g->flags & PF_KTHREAD)
 332                         continue;
 333                 p = g;
 334                 do {
 335                         if (p->mm) {
 336                                 if (unlikely(p->mm == mm)) {
 337                                         lock_task_sighand(p, &flags);
 338                                         nr += zap_process(p, exit_code);
 339                                         unlock_task_sighand(p, &flags);
 340                                 }
 341                                 break;
 342                         }
 343                 } while_each_thread(g, p);
 344         }
 345         rcu_read_unlock();
 346 done:
 347         atomic_set(&core_state->nr_threads, nr);
 348         return nr;
 349 }
 350
 351 static int coredump_wait(int exit_code, struct core_state *core_state)
 352 {
 353         struct task_struct *tsk = current;
 354         struct mm_struct *mm = tsk->mm;
 355         int core_waiters = -EBUSY;
 356
 357         init_completion(&core_state->startup);
 358         core_state->dumper.task = tsk;
 359         core_state->dumper.next = NULL;
 360
 361         down_write(&mm->mmap_sem);
 362         if (!mm->core_state)
 363                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 364         up_write(&mm->mmap_sem);
 365
 366         if (core_waiters > 0) {
 367                 struct core_thread *ptr;
 368
 369                 wait_for_completion(&core_state->startup);
 370                 /*
 371                  * Wait for all the threads to become inactive, so that
 372                  * all the thread context (extended register state, like
 373                  * fpu etc) gets copied to the memory.
 374                  */
 375                 ptr = core_state->dumper.next;
 376                 while (ptr != NULL) {
 377                         wait_task_inactive(ptr->task, 0);
 378                         ptr = ptr->next;
 379                 }
 380         }
 381
 382         return core_waiters;
 383 }
 384
 385 static void coredump_finish(struct mm_struct *mm)
 386 {
 387         struct core_thread *curr, *next;
 388         struct task_struct *task;
 389
 390         next = mm->core_state->dumper.next;
 391         while ((curr = next) != NULL) {
 392                 next = curr->next;
 393                 task = curr->task;
 394                 /*
 395                  * see exit_mm(), curr->task must not see
 396                  * ->task == NULL before we read ->next.
 397                  */
 398                 smp_mb();
 399                 curr->task = NULL;
 400                 wake_up_process(task);
 401         }
 402
 403         mm->core_state = NULL;
 404 }
 405
 406 static void wait_for_dump_helpers(struct file *file)
 407 {
 408         struct pipe_inode_info *pipe;
 409
 410         pipe = file->f_path.dentry->d_inode->i_pipe;
 411
 412         pipe_lock(pipe);
 413         pipe->readers++;
 414         pipe->writers--;
 415
 416         while ((pipe->readers > 1) && (!signal_pending(current))) {
 417                 wake_up_interruptible_sync(&pipe->wait);
 418                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 419                 pipe_wait(pipe);
 420         }
 421
 422         pipe->readers--;
 423         pipe->writers++;
 424         pipe_unlock(pipe);
 425
 426 }
 427
 428 /*
 429  * umh_pipe_setup
 430  * helper function to customize the process used
 431  * to collect the core in userspace.  Specifically
 432  * it sets up a pipe and installs it as fd 0 (stdin)
 433  * for the process.  Returns 0 on success, or
 434  * PTR_ERR on failure.
 435  * Note that it also sets the core limit to 1.  This
 436  * is a special value that we use to trap recursive
 437  * core dumps
 438  */
 439 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 440 {
 441         struct file *files[2];
 442         struct coredump_params *cp = (struct coredump_params *)info->data;
 443         int err = create_pipe_files(files, 0);
 444         if (err)
 445                 return err;
 446
 447         cp->file = files[1];
 448
 449         replace_fd(0, files[0], 0);
 450         /* and disallow core files too */
 451         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 452
 453         return 0;
 454 }
 455
 456 void do_coredump(long signr, int exit_code, struct pt_regs *regs)
 457 {
 458         struct core_state core_state;
 459         struct core_name cn;
 460         struct mm_struct *mm = current->mm;
 461         struct linux_binfmt * binfmt;
 462         const struct cred *old_cred;
 463         struct cred *cred;
 464         int retval = 0;
 465         int flag = 0;
 466         int ispipe;
 467         struct files_struct *displaced;
 468         bool need_nonrelative = false;
 469         static atomic_t core_dump_count = ATOMIC_INIT(0);
 470         struct coredump_params cprm = {
 471                 .signr = signr,
 472                 .regs = regs,
 473                 .limit = rlimit(RLIMIT_CORE),
 474                 /*
 475                  * We must use the same mm->flags while dumping core to avoid
 476                  * inconsistency of bit flags, since this flag is not protected
 477                  * by any locks.
 478                  */
 479                 .mm_flags = mm->flags,
 480         };
 481
 482         audit_core_dumps(signr);
 483
 484         binfmt = mm->binfmt;
 485         if (!binfmt || !binfmt->core_dump)
 486                 goto fail;
 487         if (!__get_dumpable(cprm.mm_flags))
 488                 goto fail;
 489
 490         cred = prepare_creds();
 491         if (!cred)
 492                 goto fail;
 493         /*
 494          * We cannot trust fsuid as being the "true" uid of the process
 495          * nor do we know its entire history. We only know it was tainted
 496          * so we dump it as root in mode 2, and only into a controlled
 497          * environment (pipe handler or fully qualified path).
 498          */
 499         if (__get_dumpable(cprm.mm_flags) == SUID_DUMPABLE_SAFE) {
 500                 /* Setuid core dump mode */
 501                 flag = O_EXCL;          /* Stop rewrite attacks */
 502                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 503                 need_nonrelative = true;
 504         }
 505
 506         retval = coredump_wait(exit_code, &core_state);
 507         if (retval < 0)
 508                 goto fail_creds;
 509
 510         old_cred = override_creds(cred);
 511
 512         /*
 513          * Clear any false indication of pending signals that might
 514          * be seen by the filesystem code called to write the core file.
 515          */
 516         clear_thread_flag(TIF_SIGPENDING);
 517
 518         ispipe = format_corename(&cn, signr);
 519
 520         if (ispipe) {
 521                 int dump_count;
 522                 char **helper_argv;
 523
 524                 if (ispipe < 0) {
 525                         printk(KERN_WARNING "format_corename failed\n");
 526                         printk(KERN_WARNING "Aborting core\n");
 527                         goto fail_corename;
 528                 }
 529
 530                 if (cprm.limit == 1) {
 531                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 532                          *
 533                          * Normally core limits are irrelevant to pipes, since
 534                          * we're not writing to the file system, but we use
 535                          * cprm.limit of 1 here as a speacial value, this is a
 536                          * consistent way to catch recursive crashes.
 537                          * We can still crash if the core_pattern binary sets
 538                          * RLIM_CORE = !1, but it runs as root, and can do
 539                          * lots of stupid things.
 540                          *
 541                          * Note that we use task_tgid_vnr here to grab the pid
 542                          * of the process group leader.  That way we get the
 543                          * right pid if a thread in a multi-threaded
 544                          * core_pattern process dies.
 545                          */
 546                         printk(KERN_WARNING
 547                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 548                                 task_tgid_vnr(current), current->comm);
 549                         printk(KERN_WARNING "Aborting core\n");
 550                         goto fail_unlock;
 551                 }
 552                 cprm.limit = RLIM_INFINITY;
 553
 554                 dump_count = atomic_inc_return(&core_dump_count);
 555                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 556                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 557                                task_tgid_vnr(current), current->comm);
 558                         printk(KERN_WARNING "Skipping core dump\n");
 559                         goto fail_dropcount;
 560                 }
 561
 562                 helper_argv = argv_split(GFP_KERNEL, cn.corename+1, NULL);
 563                 if (!helper_argv) {
 564                         printk(KERN_WARNING "%s failed to allocate memory\n",
 565                                __func__);
 566                         goto fail_dropcount;
 567                 }
 568
 569                 retval = call_usermodehelper_fns(helper_argv[0], helper_argv,
 570                                         NULL, UMH_WAIT_EXEC, umh_pipe_setup,
 571                                         NULL, &cprm);
 572                 argv_free(helper_argv);
 573                 if (retval) {
 574                         printk(KERN_INFO "Core dump to %s pipe failed\n",
 575                                cn.corename);
 576                         goto close_fail;
 577                 }
 578         } else {
 579                 struct inode *inode;
 580
 581                 if (cprm.limit < binfmt->min_coredump)
 582                         goto fail_unlock;
 583
 584                 if (need_nonrelative && cn.corename[0] != '/') {
 585                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 586                                 "to fully qualified path!\n",
 587                                 task_tgid_vnr(current), current->comm);
 588                         printk(KERN_WARNING "Skipping core dump\n");
 589                         goto fail_unlock;
 590                 }
 591
 592                 cprm.file = filp_open(cn.corename,
 593                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
 594                                  0600);
 595                 if (IS_ERR(cprm.file))
 596                         goto fail_unlock;
 597
 598                 inode = cprm.file->f_path.dentry->d_inode;
 599                 if (inode->i_nlink > 1)
 600                         goto close_fail;
 601                 if (d_unhashed(cprm.file->f_path.dentry))
 602                         goto close_fail;
 603                 /*
 604                  * AK: actually i see no reason to not allow this for named
 605                  * pipes etc, but keep the previous behaviour for now.
 606                  */
 607                 if (!S_ISREG(inode->i_mode))
 608                         goto close_fail;
 609                 /*
 610                  * Dont allow local users get cute and trick others to coredump
 611                  * into their pre-created files.
 612                  */
 613                 if (!uid_eq(inode->i_uid, current_fsuid()))
 614                         goto close_fail;
 615                 if (!cprm.file->f_op || !cprm.file->f_op->write)
 616                         goto close_fail;
 617                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
 618                         goto close_fail;
 619         }
 620
 621         /* get us an unshared descriptor table; almost always a no-op */
 622         retval = unshare_files(&displaced);
 623         if (retval)
 624                 goto close_fail;
 625         if (displaced)
 626                 put_files_struct(displaced);
 627         retval = binfmt->core_dump(&cprm);
 628         if (retval)
 629                 current->signal->group_exit_code |= 0x80;
 630
 631         if (ispipe && core_pipe_limit)
 632                 wait_for_dump_helpers(cprm.file);
 633 close_fail:
 634         if (cprm.file)
 635                 filp_close(cprm.file, NULL);
 636 fail_dropcount:
 637         if (ispipe)
 638                 atomic_dec(&core_dump_count);
 639 fail_unlock:
 640         kfree(cn.corename);
 641 fail_corename:
 642         coredump_finish(mm);
 643         revert_creds(old_cred);
 644 fail_creds:
 645         put_cred(cred);
 646 fail:
 647         return;
 648 }
 649
 650 /*
 651  * Core dumping helper functions.  These are the only things you should
 652  * do on a core-file: use only these functions to write out all the
 653  * necessary info.
 654  */
 655 int dump_write(struct file *file, const void *addr, int nr)
 656 {
 657         return access_ok(VERIFY_READ, addr, nr) && file->f_op->write(file, addr, nr, &file->f_pos) == nr;
 658 }
 659 EXPORT_SYMBOL(dump_write);
 660
 661 int dump_seek(struct file *file, loff_t off)
 662 {
 663         int ret = 1;
 664
 665         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 666                 if (file->f_op->llseek(file, off, SEEK_CUR) < 0)
 667                         return 0;
 668         } else {
 669                 char *buf = (char *)get_zeroed_page(GFP_KERNEL);
 670
 671                 if (!buf)
 672                         return 0;
 673                 while (off > 0) {
 674                         unsigned long n = off;
 675
 676                         if (n > PAGE_SIZE)
 677                                 n = PAGE_SIZE;
 678                         if (!dump_write(file, buf, n)) {
 679                                 ret = 0;
 680                                 break;
 681                         }
 682                         off -= n;
 683                 }
 684                 free_page((unsigned long)buf);
 685         }
 686         return ret;
 687 }
 688 EXPORT_SYMBOL(dump_seek);