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, struct coredump_params *cprm)
 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                         case 'd':
 198                                 err = cn_printf(cn, "%d",
 199                                         __get_dumpable(cprm->mm_flags));
 200                                 break;
 201                         /* signal that caused the coredump */
 202                         case 's':
 203                                 err = cn_printf(cn, "%ld", cprm->siginfo->si_signo);
 204                                 break;
 205                         /* UNIX time of coredump */
 206                         case 't': {
 207                                 struct timeval tv;
 208                                 do_gettimeofday(&tv);
 209                                 err = cn_printf(cn, "%lu", tv.tv_sec);
 210                                 break;
 211                         }
 212                         /* hostname */
 213                         case 'h': {
 214                                 char *namestart = cn->corename + cn->used;
 215                                 down_read(&uts_sem);
 216                                 err = cn_printf(cn, "%s",
 217                                               utsname()->nodename);
 218                                 up_read(&uts_sem);
 219                                 cn_escape(namestart);
 220                                 break;
 221                         }
 222                         /* executable */
 223                         case 'e': {
 224                                 char *commstart = cn->corename + cn->used;
 225                                 err = cn_printf(cn, "%s", current->comm);
 226                                 cn_escape(commstart);
 227                                 break;
 228                         }
 229                         case 'E':
 230                                 err = cn_print_exe_file(cn);
 231                                 break;
 232                         /* core limit size */
 233                         case 'c':
 234                                 err = cn_printf(cn, "%lu",
 235                                               rlimit(RLIMIT_CORE));
 236                                 break;
 237                         default:
 238                                 break;
 239                         }
 240                         ++pat_ptr;
 241                 }
 242
 243                 if (err)
 244                         return err;
 245         }
 246
 247         /* Backward compatibility with core_uses_pid:
 248          *
 249          * If core_pattern does not include a %p (as is the default)
 250          * and core_uses_pid is set, then .%pid will be appended to
 251          * the filename. Do not do this for piped commands. */
 252         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 253                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 254                 if (err)
 255                         return err;
 256         }
 257 out:
 258         return ispipe;
 259 }
 260
 261 static int zap_process(struct task_struct *start, int exit_code)
 262 {
 263         struct task_struct *t;
 264         int nr = 0;
 265
 266         start->signal->flags = SIGNAL_GROUP_EXIT;
 267         start->signal->group_exit_code = exit_code;
 268         start->signal->group_stop_count = 0;
 269
 270         t = start;
 271         do {
 272                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 273                 if (t != current && t->mm) {
 274                         sigaddset(&t->pending.signal, SIGKILL);
 275                         signal_wake_up(t, 1);
 276                         nr++;
 277                 }
 278         } while_each_thread(start, t);
 279
 280         return nr;
 281 }
 282
 283 static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 284                                 struct core_state *core_state, int exit_code)
 285 {
 286         struct task_struct *g, *p;
 287         unsigned long flags;
 288         int nr = -EAGAIN;
 289
 290         spin_lock_irq(&tsk->sighand->siglock);
 291         if (!signal_group_exit(tsk->signal)) {
 292                 mm->core_state = core_state;
 293                 nr = zap_process(tsk, exit_code);
 294         }
 295         spin_unlock_irq(&tsk->sighand->siglock);
 296         if (unlikely(nr < 0))
 297                 return nr;
 298
 299         if (atomic_read(&mm->mm_users) == nr + 1)
 300                 goto done;
 301         /*
 302          * We should find and kill all tasks which use this mm, and we should
 303          * count them correctly into ->nr_threads. We don't take tasklist
 304          * lock, but this is safe wrt:
 305          *
 306          * fork:
 307          *      None of sub-threads can fork after zap_process(leader). All
 308          *      processes which were created before this point should be
 309          *      visible to zap_threads() because copy_process() adds the new
 310          *      process to the tail of init_task.tasks list, and lock/unlock
 311          *      of ->siglock provides a memory barrier.
 312          *
 313          * do_exit:
 314          *      The caller holds mm->mmap_sem. This means that the task which
 315          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 316          *      its ->mm.
 317          *
 318          * de_thread:
 319          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 320          *      we must see either old or new leader, this does not matter.
 321          *      However, it can change p->sighand, so lock_task_sighand(p)
 322          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
 323          *      it can't fail.
 324          *
 325          *      Note also that "g" can be the old leader with ->mm == NULL
 326          *      and already unhashed and thus removed from ->thread_group.
 327          *      This is OK, __unhash_process()->list_del_rcu() does not
 328          *      clear the ->next pointer, we will find the new leader via
 329          *      next_thread().
 330          */
 331         rcu_read_lock();
 332         for_each_process(g) {
 333                 if (g == tsk->group_leader)
 334                         continue;
 335                 if (g->flags & PF_KTHREAD)
 336                         continue;
 337                 p = g;
 338                 do {
 339                         if (p->mm) {
 340                                 if (unlikely(p->mm == mm)) {
 341                                         lock_task_sighand(p, &flags);
 342                                         nr += zap_process(p, exit_code);
 343                                         unlock_task_sighand(p, &flags);
 344                                 }
 345                                 break;
 346                         }
 347                 } while_each_thread(g, p);
 348         }
 349         rcu_read_unlock();
 350 done:
 351         atomic_set(&core_state->nr_threads, nr);
 352         return nr;
 353 }
 354
 355 static int coredump_wait(int exit_code, struct core_state *core_state)
 356 {
 357         struct task_struct *tsk = current;
 358         struct mm_struct *mm = tsk->mm;
 359         int core_waiters = -EBUSY;
 360
 361         init_completion(&core_state->startup);
 362         core_state->dumper.task = tsk;
 363         core_state->dumper.next = NULL;
 364
 365         down_write(&mm->mmap_sem);
 366         if (!mm->core_state)
 367                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 368         up_write(&mm->mmap_sem);
 369
 370         if (core_waiters > 0) {
 371                 struct core_thread *ptr;
 372
 373                 wait_for_completion(&core_state->startup);
 374                 /*
 375                  * Wait for all the threads to become inactive, so that
 376                  * all the thread context (extended register state, like
 377                  * fpu etc) gets copied to the memory.
 378                  */
 379                 ptr = core_state->dumper.next;
 380                 while (ptr != NULL) {
 381                         wait_task_inactive(ptr->task, 0);
 382                         ptr = ptr->next;
 383                 }
 384         }
 385
 386         return core_waiters;
 387 }
 388
 389 static void coredump_finish(struct mm_struct *mm)
 390 {
 391         struct core_thread *curr, *next;
 392         struct task_struct *task;
 393
 394         next = mm->core_state->dumper.next;
 395         while ((curr = next) != NULL) {
 396                 next = curr->next;
 397                 task = curr->task;
 398                 /*
 399                  * see exit_mm(), curr->task must not see
 400                  * ->task == NULL before we read ->next.
 401                  */
 402                 smp_mb();
 403                 curr->task = NULL;
 404                 wake_up_process(task);
 405         }
 406
 407         mm->core_state = NULL;
 408 }
 409
 410 static void wait_for_dump_helpers(struct file *file)
 411 {
 412         struct pipe_inode_info *pipe;
 413
 414         pipe = file_inode(file)->i_pipe;
 415
 416         pipe_lock(pipe);
 417         pipe->readers++;
 418         pipe->writers--;
 419
 420         while ((pipe->readers > 1) && (!signal_pending(current))) {
 421                 wake_up_interruptible_sync(&pipe->wait);
 422                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 423                 pipe_wait(pipe);
 424         }
 425
 426         pipe->readers--;
 427         pipe->writers++;
 428         pipe_unlock(pipe);
 429
 430 }
 431
 432 /*
 433  * umh_pipe_setup
 434  * helper function to customize the process used
 435  * to collect the core in userspace.  Specifically
 436  * it sets up a pipe and installs it as fd 0 (stdin)
 437  * for the process.  Returns 0 on success, or
 438  * PTR_ERR on failure.
 439  * Note that it also sets the core limit to 1.  This
 440  * is a special value that we use to trap recursive
 441  * core dumps
 442  */
 443 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 444 {
 445         struct file *files[2];
 446         struct coredump_params *cp = (struct coredump_params *)info->data;
 447         int err = create_pipe_files(files, 0);
 448         if (err)
 449                 return err;
 450
 451         cp->file = files[1];
 452
 453         err = replace_fd(0, files[0], 0);
 454         fput(files[0]);
 455         /* and disallow core files too */
 456         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 457
 458         return err;
 459 }
 460
 461 void do_coredump(siginfo_t *siginfo)
 462 {
 463         struct core_state core_state;
 464         struct core_name cn;
 465         struct mm_struct *mm = current->mm;
 466         struct linux_binfmt * binfmt;
 467         const struct cred *old_cred;
 468         struct cred *cred;
 469         int retval = 0;
 470         int flag = 0;
 471         int ispipe;
 472         struct files_struct *displaced;
 473         bool need_nonrelative = false;
 474         static atomic_t core_dump_count = ATOMIC_INIT(0);
 475         struct coredump_params cprm = {
 476                 .siginfo = siginfo,
 477                 .regs = signal_pt_regs(),
 478                 .limit = rlimit(RLIMIT_CORE),
 479                 /*
 480                  * We must use the same mm->flags while dumping core to avoid
 481                  * inconsistency of bit flags, since this flag is not protected
 482                  * by any locks.
 483                  */
 484                 .mm_flags = mm->flags,
 485         };
 486
 487         audit_core_dumps(siginfo->si_signo);
 488
 489         binfmt = mm->binfmt;
 490         if (!binfmt || !binfmt->core_dump)
 491                 goto fail;
 492         if (!__get_dumpable(cprm.mm_flags))
 493                 goto fail;
 494
 495         cred = prepare_creds();
 496         if (!cred)
 497                 goto fail;
 498         /*
 499          * We cannot trust fsuid as being the "true" uid of the process
 500          * nor do we know its entire history. We only know it was tainted
 501          * so we dump it as root in mode 2, and only into a controlled
 502          * environment (pipe handler or fully qualified path).
 503          */
 504         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 505                 /* Setuid core dump mode */
 506                 flag = O_EXCL;          /* Stop rewrite attacks */
 507                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 508                 need_nonrelative = true;
 509         }
 510
 511         retval = coredump_wait(siginfo->si_signo, &core_state);
 512         if (retval < 0)
 513                 goto fail_creds;
 514
 515         old_cred = override_creds(cred);
 516
 517         /*
 518          * Clear any false indication of pending signals that might
 519          * be seen by the filesystem code called to write the core file.
 520          */
 521         clear_thread_flag(TIF_SIGPENDING);
 522
 523         ispipe = format_corename(&cn, &cprm);
 524
 525         if (ispipe) {
 526                 int dump_count;
 527                 char **helper_argv;
 528
 529                 if (ispipe < 0) {
 530                         printk(KERN_WARNING "format_corename failed\n");
 531                         printk(KERN_WARNING "Aborting core\n");
 532                         goto fail_corename;
 533                 }
 534
 535                 if (cprm.limit == 1) {
 536                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 537                          *
 538                          * Normally core limits are irrelevant to pipes, since
 539                          * we're not writing to the file system, but we use
 540                          * cprm.limit of 1 here as a speacial value, this is a
 541                          * consistent way to catch recursive crashes.
 542                          * We can still crash if the core_pattern binary sets
 543                          * RLIM_CORE = !1, but it runs as root, and can do
 544                          * lots of stupid things.
 545                          *
 546                          * Note that we use task_tgid_vnr here to grab the pid
 547                          * of the process group leader.  That way we get the
 548                          * right pid if a thread in a multi-threaded
 549                          * core_pattern process dies.
 550                          */
 551                         printk(KERN_WARNING
 552                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 553                                 task_tgid_vnr(current), current->comm);
 554                         printk(KERN_WARNING "Aborting core\n");
 555                         goto fail_unlock;
 556                 }
 557                 cprm.limit = RLIM_INFINITY;
 558
 559                 dump_count = atomic_inc_return(&core_dump_count);
 560                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 561                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 562                                task_tgid_vnr(current), current->comm);
 563                         printk(KERN_WARNING "Skipping core dump\n");
 564                         goto fail_dropcount;
 565                 }
 566
 567                 helper_argv = argv_split(GFP_KERNEL, cn.corename+1, NULL);
 568                 if (!helper_argv) {
 569                         printk(KERN_WARNING "%s failed to allocate memory\n",
 570                                __func__);
 571                         goto fail_dropcount;
 572                 }
 573
 574                 retval = call_usermodehelper_fns(helper_argv[0], helper_argv,
 575                                         NULL, UMH_WAIT_EXEC, umh_pipe_setup,
 576                                         NULL, &cprm);
 577                 argv_free(helper_argv);
 578                 if (retval) {
 579                         printk(KERN_INFO "Core dump to %s pipe failed\n",
 580                                cn.corename);
 581                         goto close_fail;
 582                 }
 583         } else {
 584                 struct inode *inode;
 585
 586                 if (cprm.limit < binfmt->min_coredump)
 587                         goto fail_unlock;
 588
 589                 if (need_nonrelative && cn.corename[0] != '/') {
 590                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 591                                 "to fully qualified path!\n",
 592                                 task_tgid_vnr(current), current->comm);
 593                         printk(KERN_WARNING "Skipping core dump\n");
 594                         goto fail_unlock;
 595                 }
 596
 597                 cprm.file = filp_open(cn.corename,
 598                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
 599                                  0600);
 600                 if (IS_ERR(cprm.file))
 601                         goto fail_unlock;
 602
 603                 inode = file_inode(cprm.file);
 604                 if (inode->i_nlink > 1)
 605                         goto close_fail;
 606                 if (d_unhashed(cprm.file->f_path.dentry))
 607                         goto close_fail;
 608                 /*
 609                  * AK: actually i see no reason to not allow this for named
 610                  * pipes etc, but keep the previous behaviour for now.
 611                  */
 612                 if (!S_ISREG(inode->i_mode))
 613                         goto close_fail;
 614                 /*
 615                  * Dont allow local users get cute and trick others to coredump
 616                  * into their pre-created files.
 617                  */
 618                 if (!uid_eq(inode->i_uid, current_fsuid()))
 619                         goto close_fail;
 620                 if (!cprm.file->f_op || !cprm.file->f_op->write)
 621                         goto close_fail;
 622                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
 623                         goto close_fail;
 624         }
 625
 626         /* get us an unshared descriptor table; almost always a no-op */
 627         retval = unshare_files(&displaced);
 628         if (retval)
 629                 goto close_fail;
 630         if (displaced)
 631                 put_files_struct(displaced);
 632         retval = binfmt->core_dump(&cprm);
 633         if (retval)
 634                 current->signal->group_exit_code |= 0x80;
 635
 636         if (ispipe && core_pipe_limit)
 637                 wait_for_dump_helpers(cprm.file);
 638 close_fail:
 639         if (cprm.file)
 640                 filp_close(cprm.file, NULL);
 641 fail_dropcount:
 642         if (ispipe)
 643                 atomic_dec(&core_dump_count);
 644 fail_unlock:
 645         kfree(cn.corename);
 646 fail_corename:
 647         coredump_finish(mm);
 648         revert_creds(old_cred);
 649 fail_creds:
 650         put_cred(cred);
 651 fail:
 652         return;
 653 }
 654
 655 /*
 656  * Core dumping helper functions.  These are the only things you should
 657  * do on a core-file: use only these functions to write out all the
 658  * necessary info.
 659  */
 660 int dump_write(struct file *file, const void *addr, int nr)
 661 {
 662         return access_ok(VERIFY_READ, addr, nr) && file->f_op->write(file, addr, nr, &file->f_pos) == nr;
 663 }
 664 EXPORT_SYMBOL(dump_write);
 665
 666 int dump_seek(struct file *file, loff_t off)
 667 {
 668         int ret = 1;
 669
 670         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 671                 if (file->f_op->llseek(file, off, SEEK_CUR) < 0)
 672                         return 0;
 673         } else {
 674                 char *buf = (char *)get_zeroed_page(GFP_KERNEL);
 675
 676                 if (!buf)
 677                         return 0;
 678                 while (off > 0) {
 679                         unsigned long n = off;
 680
 681                         if (n > PAGE_SIZE)
 682                                 n = PAGE_SIZE;
 683                         if (!dump_write(file, buf, n)) {
 684                                 ret = 0;
 685                                 break;
 686                         }
 687                         off -= n;
 688                 }
 689                 free_page((unsigned long)buf);
 690         }
 691         return ret;
 692 }
 693 EXPORT_SYMBOL(dump_seek);