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