coredump: format_corename() can leak cn->corename
[cascardo/linux.git] / 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         int size = CORENAME_MAX_SIZE * atomic_inc_return(&call_count);
62         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
63
64         if (!corename)
65                 return -ENOMEM;
66
67         cn->size = size;
68         cn->corename = corename;
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, struct coredump_params *cprm)
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->used = 0;
159         cn->size = CORENAME_MAX_SIZE * atomic_read(&call_count);
160         cn->corename = kmalloc(cn->size, GFP_KERNEL);
161         if (!cn->corename)
162                 return -ENOMEM;
163
164         /* Repeat as long as we have more pattern to process and more output
165            space */
166         while (*pat_ptr) {
167                 if (*pat_ptr != '%') {
168                         if (*pat_ptr == 0)
169                                 goto out;
170                         err = cn_printf(cn, "%c", *pat_ptr++);
171                 } else {
172                         switch (*++pat_ptr) {
173                         /* single % at the end, drop that */
174                         case 0:
175                                 goto out;
176                         /* Double percent, output one percent */
177                         case '%':
178                                 err = cn_printf(cn, "%c", '%');
179                                 break;
180                         /* pid */
181                         case 'p':
182                                 pid_in_pattern = 1;
183                                 err = cn_printf(cn, "%d",
184                                               task_tgid_vnr(current));
185                                 break;
186                         /* uid */
187                         case 'u':
188                                 err = cn_printf(cn, "%d", cred->uid);
189                                 break;
190                         /* gid */
191                         case 'g':
192                                 err = cn_printf(cn, "%d", cred->gid);
193                                 break;
194                         case 'd':
195                                 err = cn_printf(cn, "%d",
196                                         __get_dumpable(cprm->mm_flags));
197                                 break;
198                         /* signal that caused the coredump */
199                         case 's':
200                                 err = cn_printf(cn, "%ld", cprm->siginfo->si_signo);
201                                 break;
202                         /* UNIX time of coredump */
203                         case 't': {
204                                 struct timeval tv;
205                                 do_gettimeofday(&tv);
206                                 err = cn_printf(cn, "%lu", tv.tv_sec);
207                                 break;
208                         }
209                         /* hostname */
210                         case 'h': {
211                                 char *namestart = cn->corename + cn->used;
212                                 down_read(&uts_sem);
213                                 err = cn_printf(cn, "%s",
214                                               utsname()->nodename);
215                                 up_read(&uts_sem);
216                                 cn_escape(namestart);
217                                 break;
218                         }
219                         /* executable */
220                         case 'e': {
221                                 char *commstart = cn->corename + cn->used;
222                                 err = cn_printf(cn, "%s", current->comm);
223                                 cn_escape(commstart);
224                                 break;
225                         }
226                         case 'E':
227                                 err = cn_print_exe_file(cn);
228                                 break;
229                         /* core limit size */
230                         case 'c':
231                                 err = cn_printf(cn, "%lu",
232                                               rlimit(RLIMIT_CORE));
233                                 break;
234                         default:
235                                 break;
236                         }
237                         ++pat_ptr;
238                 }
239
240                 if (err)
241                         return err;
242         }
243
244         /* Backward compatibility with core_uses_pid:
245          *
246          * If core_pattern does not include a %p (as is the default)
247          * and core_uses_pid is set, then .%pid will be appended to
248          * the filename. Do not do this for piped commands. */
249         if (!ispipe && !pid_in_pattern && core_uses_pid) {
250                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
251                 if (err)
252                         return err;
253         }
254 out:
255         return ispipe;
256 }
257
258 static int zap_process(struct task_struct *start, int exit_code)
259 {
260         struct task_struct *t;
261         int nr = 0;
262
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 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                 tsk->signal->group_exit_task = tsk;
291                 /* ignore all signals except SIGKILL, see prepare_signal() */
292                 tsk->signal->flags = SIGNAL_GROUP_COREDUMP;
293                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
294         }
295         spin_unlock_irq(&tsk->sighand->siglock);
296         if (unlikely(nr < 0))
297                 return nr;
298
299         tsk->flags = PF_DUMPCORE;
300         if (atomic_read(&mm->mm_users) == nr + 1)
301                 goto done;
302         /*
303          * We should find and kill all tasks which use this mm, and we should
304          * count them correctly into ->nr_threads. We don't take tasklist
305          * lock, but this is safe wrt:
306          *
307          * fork:
308          *      None of sub-threads can fork after zap_process(leader). All
309          *      processes which were created before this point should be
310          *      visible to zap_threads() because copy_process() adds the new
311          *      process to the tail of init_task.tasks list, and lock/unlock
312          *      of ->siglock provides a memory barrier.
313          *
314          * do_exit:
315          *      The caller holds mm->mmap_sem. This means that the task which
316          *      uses this mm can't pass exit_mm(), so it can't exit or clear
317          *      its ->mm.
318          *
319          * de_thread:
320          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
321          *      we must see either old or new leader, this does not matter.
322          *      However, it can change p->sighand, so lock_task_sighand(p)
323          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
324          *      it can't fail.
325          *
326          *      Note also that "g" can be the old leader with ->mm == NULL
327          *      and already unhashed and thus removed from ->thread_group.
328          *      This is OK, __unhash_process()->list_del_rcu() does not
329          *      clear the ->next pointer, we will find the new leader via
330          *      next_thread().
331          */
332         rcu_read_lock();
333         for_each_process(g) {
334                 if (g == tsk->group_leader)
335                         continue;
336                 if (g->flags & PF_KTHREAD)
337                         continue;
338                 p = g;
339                 do {
340                         if (p->mm) {
341                                 if (unlikely(p->mm == mm)) {
342                                         lock_task_sighand(p, &flags);
343                                         nr += zap_process(p, exit_code);
344                                         p->signal->flags = SIGNAL_GROUP_EXIT;
345                                         unlock_task_sighand(p, &flags);
346                                 }
347                                 break;
348                         }
349                 } while_each_thread(g, p);
350         }
351         rcu_read_unlock();
352 done:
353         atomic_set(&core_state->nr_threads, nr);
354         return nr;
355 }
356
357 static int coredump_wait(int exit_code, struct core_state *core_state)
358 {
359         struct task_struct *tsk = current;
360         struct mm_struct *mm = tsk->mm;
361         int core_waiters = -EBUSY;
362
363         init_completion(&core_state->startup);
364         core_state->dumper.task = tsk;
365         core_state->dumper.next = NULL;
366
367         down_write(&mm->mmap_sem);
368         if (!mm->core_state)
369                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
370         up_write(&mm->mmap_sem);
371
372         if (core_waiters > 0) {
373                 struct core_thread *ptr;
374
375                 wait_for_completion(&core_state->startup);
376                 /*
377                  * Wait for all the threads to become inactive, so that
378                  * all the thread context (extended register state, like
379                  * fpu etc) gets copied to the memory.
380                  */
381                 ptr = core_state->dumper.next;
382                 while (ptr != NULL) {
383                         wait_task_inactive(ptr->task, 0);
384                         ptr = ptr->next;
385                 }
386         }
387
388         return core_waiters;
389 }
390
391 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
392 {
393         struct core_thread *curr, *next;
394         struct task_struct *task;
395
396         spin_lock_irq(&current->sighand->siglock);
397         if (core_dumped && !__fatal_signal_pending(current))
398                 current->signal->group_exit_code |= 0x80;
399         current->signal->group_exit_task = NULL;
400         current->signal->flags = SIGNAL_GROUP_EXIT;
401         spin_unlock_irq(&current->sighand->siglock);
402
403         next = mm->core_state->dumper.next;
404         while ((curr = next) != NULL) {
405                 next = curr->next;
406                 task = curr->task;
407                 /*
408                  * see exit_mm(), curr->task must not see
409                  * ->task == NULL before we read ->next.
410                  */
411                 smp_mb();
412                 curr->task = NULL;
413                 wake_up_process(task);
414         }
415
416         mm->core_state = NULL;
417 }
418
419 static bool dump_interrupted(void)
420 {
421         /*
422          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
423          * can do try_to_freeze() and check __fatal_signal_pending(),
424          * but then we need to teach dump_write() to restart and clear
425          * TIF_SIGPENDING.
426          */
427         return signal_pending(current);
428 }
429
430 static void wait_for_dump_helpers(struct file *file)
431 {
432         struct pipe_inode_info *pipe = file->private_data;
433
434         pipe_lock(pipe);
435         pipe->readers++;
436         pipe->writers--;
437         wake_up_interruptible_sync(&pipe->wait);
438         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
439         pipe_unlock(pipe);
440
441         /*
442          * We actually want wait_event_freezable() but then we need
443          * to clear TIF_SIGPENDING and improve dump_interrupted().
444          */
445         wait_event_interruptible(pipe->wait, pipe->readers == 1);
446
447         pipe_lock(pipe);
448         pipe->readers--;
449         pipe->writers++;
450         pipe_unlock(pipe);
451 }
452
453 /*
454  * umh_pipe_setup
455  * helper function to customize the process used
456  * to collect the core in userspace.  Specifically
457  * it sets up a pipe and installs it as fd 0 (stdin)
458  * for the process.  Returns 0 on success, or
459  * PTR_ERR on failure.
460  * Note that it also sets the core limit to 1.  This
461  * is a special value that we use to trap recursive
462  * core dumps
463  */
464 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
465 {
466         struct file *files[2];
467         struct coredump_params *cp = (struct coredump_params *)info->data;
468         int err = create_pipe_files(files, 0);
469         if (err)
470                 return err;
471
472         cp->file = files[1];
473
474         err = replace_fd(0, files[0], 0);
475         fput(files[0]);
476         /* and disallow core files too */
477         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
478
479         return err;
480 }
481
482 void do_coredump(siginfo_t *siginfo)
483 {
484         struct core_state core_state;
485         struct core_name cn;
486         struct mm_struct *mm = current->mm;
487         struct linux_binfmt * binfmt;
488         const struct cred *old_cred;
489         struct cred *cred;
490         int retval = 0;
491         int flag = 0;
492         int ispipe;
493         struct files_struct *displaced;
494         bool need_nonrelative = false;
495         bool core_dumped = false;
496         static atomic_t core_dump_count = ATOMIC_INIT(0);
497         struct coredump_params cprm = {
498                 .siginfo = siginfo,
499                 .regs = signal_pt_regs(),
500                 .limit = rlimit(RLIMIT_CORE),
501                 /*
502                  * We must use the same mm->flags while dumping core to avoid
503                  * inconsistency of bit flags, since this flag is not protected
504                  * by any locks.
505                  */
506                 .mm_flags = mm->flags,
507         };
508
509         audit_core_dumps(siginfo->si_signo);
510
511         binfmt = mm->binfmt;
512         if (!binfmt || !binfmt->core_dump)
513                 goto fail;
514         if (!__get_dumpable(cprm.mm_flags))
515                 goto fail;
516
517         cred = prepare_creds();
518         if (!cred)
519                 goto fail;
520         /*
521          * We cannot trust fsuid as being the "true" uid of the process
522          * nor do we know its entire history. We only know it was tainted
523          * so we dump it as root in mode 2, and only into a controlled
524          * environment (pipe handler or fully qualified path).
525          */
526         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
527                 /* Setuid core dump mode */
528                 flag = O_EXCL;          /* Stop rewrite attacks */
529                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
530                 need_nonrelative = true;
531         }
532
533         retval = coredump_wait(siginfo->si_signo, &core_state);
534         if (retval < 0)
535                 goto fail_creds;
536
537         old_cred = override_creds(cred);
538
539         ispipe = format_corename(&cn, &cprm);
540
541         if (ispipe) {
542                 int dump_count;
543                 char **helper_argv;
544                 struct subprocess_info *sub_info;
545
546                 if (ispipe < 0) {
547                         printk(KERN_WARNING "format_corename failed\n");
548                         printk(KERN_WARNING "Aborting core\n");
549                         goto fail_unlock;
550                 }
551
552                 if (cprm.limit == 1) {
553                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
554                          *
555                          * Normally core limits are irrelevant to pipes, since
556                          * we're not writing to the file system, but we use
557                          * cprm.limit of 1 here as a speacial value, this is a
558                          * consistent way to catch recursive crashes.
559                          * We can still crash if the core_pattern binary sets
560                          * RLIM_CORE = !1, but it runs as root, and can do
561                          * lots of stupid things.
562                          *
563                          * Note that we use task_tgid_vnr here to grab the pid
564                          * of the process group leader.  That way we get the
565                          * right pid if a thread in a multi-threaded
566                          * core_pattern process dies.
567                          */
568                         printk(KERN_WARNING
569                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
570                                 task_tgid_vnr(current), current->comm);
571                         printk(KERN_WARNING "Aborting core\n");
572                         goto fail_unlock;
573                 }
574                 cprm.limit = RLIM_INFINITY;
575
576                 dump_count = atomic_inc_return(&core_dump_count);
577                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
578                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
579                                task_tgid_vnr(current), current->comm);
580                         printk(KERN_WARNING "Skipping core dump\n");
581                         goto fail_dropcount;
582                 }
583
584                 helper_argv = argv_split(GFP_KERNEL, cn.corename+1, NULL);
585                 if (!helper_argv) {
586                         printk(KERN_WARNING "%s failed to allocate memory\n",
587                                __func__);
588                         goto fail_dropcount;
589                 }
590
591                 retval = -ENOMEM;
592                 sub_info = call_usermodehelper_setup(helper_argv[0],
593                                                 helper_argv, NULL, GFP_KERNEL,
594                                                 umh_pipe_setup, NULL, &cprm);
595                 if (sub_info)
596                         retval = call_usermodehelper_exec(sub_info,
597                                                           UMH_WAIT_EXEC);
598
599                 argv_free(helper_argv);
600                 if (retval) {
601                         printk(KERN_INFO "Core dump to %s pipe failed\n",
602                                cn.corename);
603                         goto close_fail;
604                 }
605         } else {
606                 struct inode *inode;
607
608                 if (cprm.limit < binfmt->min_coredump)
609                         goto fail_unlock;
610
611                 if (need_nonrelative && cn.corename[0] != '/') {
612                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
613                                 "to fully qualified path!\n",
614                                 task_tgid_vnr(current), current->comm);
615                         printk(KERN_WARNING "Skipping core dump\n");
616                         goto fail_unlock;
617                 }
618
619                 cprm.file = filp_open(cn.corename,
620                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
621                                  0600);
622                 if (IS_ERR(cprm.file))
623                         goto fail_unlock;
624
625                 inode = file_inode(cprm.file);
626                 if (inode->i_nlink > 1)
627                         goto close_fail;
628                 if (d_unhashed(cprm.file->f_path.dentry))
629                         goto close_fail;
630                 /*
631                  * AK: actually i see no reason to not allow this for named
632                  * pipes etc, but keep the previous behaviour for now.
633                  */
634                 if (!S_ISREG(inode->i_mode))
635                         goto close_fail;
636                 /*
637                  * Dont allow local users get cute and trick others to coredump
638                  * into their pre-created files.
639                  */
640                 if (!uid_eq(inode->i_uid, current_fsuid()))
641                         goto close_fail;
642                 if (!cprm.file->f_op || !cprm.file->f_op->write)
643                         goto close_fail;
644                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
645                         goto close_fail;
646         }
647
648         /* get us an unshared descriptor table; almost always a no-op */
649         retval = unshare_files(&displaced);
650         if (retval)
651                 goto close_fail;
652         if (displaced)
653                 put_files_struct(displaced);
654         if (!dump_interrupted()) {
655                 file_start_write(cprm.file);
656                 core_dumped = binfmt->core_dump(&cprm);
657                 file_end_write(cprm.file);
658         }
659         if (ispipe && core_pipe_limit)
660                 wait_for_dump_helpers(cprm.file);
661 close_fail:
662         if (cprm.file)
663                 filp_close(cprm.file, NULL);
664 fail_dropcount:
665         if (ispipe)
666                 atomic_dec(&core_dump_count);
667 fail_unlock:
668         kfree(cn.corename);
669         coredump_finish(mm, core_dumped);
670         revert_creds(old_cred);
671 fail_creds:
672         put_cred(cred);
673 fail:
674         return;
675 }
676
677 /*
678  * Core dumping helper functions.  These are the only things you should
679  * do on a core-file: use only these functions to write out all the
680  * necessary info.
681  */
682 int dump_write(struct file *file, const void *addr, int nr)
683 {
684         return !dump_interrupted() &&
685                 access_ok(VERIFY_READ, addr, nr) &&
686                 file->f_op->write(file, addr, nr, &file->f_pos) == nr;
687 }
688 EXPORT_SYMBOL(dump_write);
689
690 int dump_seek(struct file *file, loff_t off)
691 {
692         int ret = 1;
693
694         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
695                 if (dump_interrupted() ||
696                     file->f_op->llseek(file, off, SEEK_CUR) < 0)
697                         return 0;
698         } else {
699                 char *buf = (char *)get_zeroed_page(GFP_KERNEL);
700
701                 if (!buf)
702                         return 0;
703                 while (off > 0) {
704                         unsigned long n = off;
705
706                         if (n > PAGE_SIZE)
707                                 n = PAGE_SIZE;
708                         if (!dump_write(file, buf, n)) {
709                                 ret = 0;
710                                 break;
711                         }
712                         off -= n;
713                 }
714                 free_page((unsigned long)buf);
715         }
716         return ret;
717 }
718 EXPORT_SYMBOL(dump_seek);