timerfd: support CLOCK_BOOTTIME clock
[linux-3.10.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         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->group_exit_code = exit_code;
267         start->signal->group_stop_count = 0;
268
269         t = start;
270         do {
271                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
272                 if (t != current && t->mm) {
273                         sigaddset(&t->pending.signal, SIGKILL);
274                         signal_wake_up(t, 1);
275                         nr++;
276                 }
277         } while_each_thread(start, t);
278
279         return nr;
280 }
281
282 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
283                         struct core_state *core_state, int exit_code)
284 {
285         struct task_struct *g, *p;
286         unsigned long flags;
287         int nr = -EAGAIN;
288
289         spin_lock_irq(&tsk->sighand->siglock);
290         if (!signal_group_exit(tsk->signal)) {
291                 mm->core_state = core_state;
292                 nr = zap_process(tsk, exit_code);
293                 tsk->signal->group_exit_task = tsk;
294                 /* ignore all signals except SIGKILL, see prepare_signal() */
295                 tsk->signal->flags = SIGNAL_GROUP_COREDUMP;
296                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
297         }
298         spin_unlock_irq(&tsk->sighand->siglock);
299         if (unlikely(nr < 0))
300                 return nr;
301
302         tsk->flags = PF_DUMPCORE;
303         if (atomic_read(&mm->mm_users) == nr + 1)
304                 goto done;
305         /*
306          * We should find and kill all tasks which use this mm, and we should
307          * count them correctly into ->nr_threads. We don't take tasklist
308          * lock, but this is safe wrt:
309          *
310          * fork:
311          *      None of sub-threads can fork after zap_process(leader). All
312          *      processes which were created before this point should be
313          *      visible to zap_threads() because copy_process() adds the new
314          *      process to the tail of init_task.tasks list, and lock/unlock
315          *      of ->siglock provides a memory barrier.
316          *
317          * do_exit:
318          *      The caller holds mm->mmap_sem. This means that the task which
319          *      uses this mm can't pass exit_mm(), so it can't exit or clear
320          *      its ->mm.
321          *
322          * de_thread:
323          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
324          *      we must see either old or new leader, this does not matter.
325          *      However, it can change p->sighand, so lock_task_sighand(p)
326          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
327          *      it can't fail.
328          *
329          *      Note also that "g" can be the old leader with ->mm == NULL
330          *      and already unhashed and thus removed from ->thread_group.
331          *      This is OK, __unhash_process()->list_del_rcu() does not
332          *      clear the ->next pointer, we will find the new leader via
333          *      next_thread().
334          */
335         rcu_read_lock();
336         for_each_process(g) {
337                 if (g == tsk->group_leader)
338                         continue;
339                 if (g->flags & PF_KTHREAD)
340                         continue;
341                 p = g;
342                 do {
343                         if (p->mm) {
344                                 if (unlikely(p->mm == mm)) {
345                                         lock_task_sighand(p, &flags);
346                                         nr += zap_process(p, exit_code);
347                                         p->signal->flags = SIGNAL_GROUP_EXIT;
348                                         unlock_task_sighand(p, &flags);
349                                 }
350                                 break;
351                         }
352                 } while_each_thread(g, p);
353         }
354         rcu_read_unlock();
355 done:
356         atomic_set(&core_state->nr_threads, nr);
357         return nr;
358 }
359
360 static int coredump_wait(int exit_code, struct core_state *core_state)
361 {
362         struct task_struct *tsk = current;
363         struct mm_struct *mm = tsk->mm;
364         int core_waiters = -EBUSY;
365
366         init_completion(&core_state->startup);
367         core_state->dumper.task = tsk;
368         core_state->dumper.next = NULL;
369
370         down_write(&mm->mmap_sem);
371         if (!mm->core_state)
372                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
373         up_write(&mm->mmap_sem);
374
375         if (core_waiters > 0) {
376                 struct core_thread *ptr;
377
378                 wait_for_completion(&core_state->startup);
379                 /*
380                  * Wait for all the threads to become inactive, so that
381                  * all the thread context (extended register state, like
382                  * fpu etc) gets copied to the memory.
383                  */
384                 ptr = core_state->dumper.next;
385                 while (ptr != NULL) {
386                         wait_task_inactive(ptr->task, 0);
387                         ptr = ptr->next;
388                 }
389         }
390
391         return core_waiters;
392 }
393
394 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
395 {
396         struct core_thread *curr, *next;
397         struct task_struct *task;
398
399         spin_lock_irq(&current->sighand->siglock);
400         if (core_dumped && !__fatal_signal_pending(current))
401                 current->signal->group_exit_code |= 0x80;
402         current->signal->group_exit_task = NULL;
403         current->signal->flags = SIGNAL_GROUP_EXIT;
404         spin_unlock_irq(&current->sighand->siglock);
405
406         next = mm->core_state->dumper.next;
407         while ((curr = next) != NULL) {
408                 next = curr->next;
409                 task = curr->task;
410                 /*
411                  * see exit_mm(), curr->task must not see
412                  * ->task == NULL before we read ->next.
413                  */
414                 smp_mb();
415                 curr->task = NULL;
416                 wake_up_process(task);
417         }
418
419         mm->core_state = NULL;
420 }
421
422 static bool dump_interrupted(void)
423 {
424         /*
425          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
426          * can do try_to_freeze() and check __fatal_signal_pending(),
427          * but then we need to teach dump_write() to restart and clear
428          * TIF_SIGPENDING.
429          */
430         return signal_pending(current);
431 }
432
433 static void wait_for_dump_helpers(struct file *file)
434 {
435         struct pipe_inode_info *pipe = file->private_data;
436
437         pipe_lock(pipe);
438         pipe->readers++;
439         pipe->writers--;
440         wake_up_interruptible_sync(&pipe->wait);
441         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
442         pipe_unlock(pipe);
443
444         /*
445          * We actually want wait_event_freezable() but then we need
446          * to clear TIF_SIGPENDING and improve dump_interrupted().
447          */
448         wait_event_interruptible(pipe->wait, pipe->readers == 1);
449
450         pipe_lock(pipe);
451         pipe->readers--;
452         pipe->writers++;
453         pipe_unlock(pipe);
454 }
455
456 /*
457  * umh_pipe_setup
458  * helper function to customize the process used
459  * to collect the core in userspace.  Specifically
460  * it sets up a pipe and installs it as fd 0 (stdin)
461  * for the process.  Returns 0 on success, or
462  * PTR_ERR on failure.
463  * Note that it also sets the core limit to 1.  This
464  * is a special value that we use to trap recursive
465  * core dumps
466  */
467 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
468 {
469         struct file *files[2];
470         struct coredump_params *cp = (struct coredump_params *)info->data;
471         int err = create_pipe_files(files, 0);
472         if (err)
473                 return err;
474
475         cp->file = files[1];
476
477         err = replace_fd(0, files[0], 0);
478         fput(files[0]);
479         /* and disallow core files too */
480         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
481
482         return err;
483 }
484
485 void do_coredump(siginfo_t *siginfo)
486 {
487         struct core_state core_state;
488         struct core_name cn;
489         struct mm_struct *mm = current->mm;
490         struct linux_binfmt * binfmt;
491         const struct cred *old_cred;
492         struct cred *cred;
493         int retval = 0;
494         int flag = 0;
495         int ispipe;
496         struct files_struct *displaced;
497         bool need_nonrelative = false;
498         bool core_dumped = false;
499         static atomic_t core_dump_count = ATOMIC_INIT(0);
500         struct coredump_params cprm = {
501                 .siginfo = siginfo,
502                 .regs = signal_pt_regs(),
503                 .limit = rlimit(RLIMIT_CORE),
504                 /*
505                  * We must use the same mm->flags while dumping core to avoid
506                  * inconsistency of bit flags, since this flag is not protected
507                  * by any locks.
508                  */
509                 .mm_flags = mm->flags,
510         };
511
512         audit_core_dumps(siginfo->si_signo);
513
514         binfmt = mm->binfmt;
515         if (!binfmt || !binfmt->core_dump)
516                 goto fail;
517         if (!__get_dumpable(cprm.mm_flags))
518                 goto fail;
519
520         cred = prepare_creds();
521         if (!cred)
522                 goto fail;
523         /*
524          * We cannot trust fsuid as being the "true" uid of the process
525          * nor do we know its entire history. We only know it was tainted
526          * so we dump it as root in mode 2, and only into a controlled
527          * environment (pipe handler or fully qualified path).
528          */
529         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
530                 /* Setuid core dump mode */
531                 flag = O_EXCL;          /* Stop rewrite attacks */
532                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
533                 need_nonrelative = true;
534         }
535
536         retval = coredump_wait(siginfo->si_signo, &core_state);
537         if (retval < 0)
538                 goto fail_creds;
539
540         old_cred = override_creds(cred);
541
542         ispipe = format_corename(&cn, &cprm);
543
544         if (ispipe) {
545                 int dump_count;
546                 char **helper_argv;
547                 struct subprocess_info *sub_info;
548
549                 if (ispipe < 0) {
550                         printk(KERN_WARNING "format_corename failed\n");
551                         printk(KERN_WARNING "Aborting core\n");
552                         goto fail_corename;
553                 }
554
555                 if (cprm.limit == 1) {
556                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
557                          *
558                          * Normally core limits are irrelevant to pipes, since
559                          * we're not writing to the file system, but we use
560                          * cprm.limit of 1 here as a speacial value, this is a
561                          * consistent way to catch recursive crashes.
562                          * We can still crash if the core_pattern binary sets
563                          * RLIM_CORE = !1, but it runs as root, and can do
564                          * lots of stupid things.
565                          *
566                          * Note that we use task_tgid_vnr here to grab the pid
567                          * of the process group leader.  That way we get the
568                          * right pid if a thread in a multi-threaded
569                          * core_pattern process dies.
570                          */
571                         printk(KERN_WARNING
572                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
573                                 task_tgid_vnr(current), current->comm);
574                         printk(KERN_WARNING "Aborting core\n");
575                         goto fail_unlock;
576                 }
577                 cprm.limit = RLIM_INFINITY;
578
579                 dump_count = atomic_inc_return(&core_dump_count);
580                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
581                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
582                                task_tgid_vnr(current), current->comm);
583                         printk(KERN_WARNING "Skipping core dump\n");
584                         goto fail_dropcount;
585                 }
586
587                 helper_argv = argv_split(GFP_KERNEL, cn.corename+1, NULL);
588                 if (!helper_argv) {
589                         printk(KERN_WARNING "%s failed to allocate memory\n",
590                                __func__);
591                         goto fail_dropcount;
592                 }
593
594                 retval = -ENOMEM;
595                 sub_info = call_usermodehelper_setup(helper_argv[0],
596                                                 helper_argv, NULL, GFP_KERNEL,
597                                                 umh_pipe_setup, NULL, &cprm);
598                 if (sub_info)
599                         retval = call_usermodehelper_exec(sub_info,
600                                                           UMH_WAIT_EXEC);
601
602                 argv_free(helper_argv);
603                 if (retval) {
604                         printk(KERN_INFO "Core dump to %s pipe failed\n",
605                                cn.corename);
606                         goto close_fail;
607                 }
608         } else {
609                 struct inode *inode;
610
611                 if (cprm.limit < binfmt->min_coredump)
612                         goto fail_unlock;
613
614                 if (need_nonrelative && cn.corename[0] != '/') {
615                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
616                                 "to fully qualified path!\n",
617                                 task_tgid_vnr(current), current->comm);
618                         printk(KERN_WARNING "Skipping core dump\n");
619                         goto fail_unlock;
620                 }
621
622                 cprm.file = filp_open(cn.corename,
623                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
624                                  0600);
625                 if (IS_ERR(cprm.file))
626                         goto fail_unlock;
627
628                 inode = file_inode(cprm.file);
629                 if (inode->i_nlink > 1)
630                         goto close_fail;
631                 if (d_unhashed(cprm.file->f_path.dentry))
632                         goto close_fail;
633                 /*
634                  * AK: actually i see no reason to not allow this for named
635                  * pipes etc, but keep the previous behaviour for now.
636                  */
637                 if (!S_ISREG(inode->i_mode))
638                         goto close_fail;
639                 /*
640                  * Dont allow local users get cute and trick others to coredump
641                  * into their pre-created files.
642                  */
643                 if (!uid_eq(inode->i_uid, current_fsuid()))
644                         goto close_fail;
645                 if (!cprm.file->f_op || !cprm.file->f_op->write)
646                         goto close_fail;
647                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
648                         goto close_fail;
649         }
650
651         /* get us an unshared descriptor table; almost always a no-op */
652         retval = unshare_files(&displaced);
653         if (retval)
654                 goto close_fail;
655         if (displaced)
656                 put_files_struct(displaced);
657         if (!dump_interrupted()) {
658                 file_start_write(cprm.file);
659                 core_dumped = binfmt->core_dump(&cprm);
660                 file_end_write(cprm.file);
661         }
662         if (ispipe && core_pipe_limit)
663                 wait_for_dump_helpers(cprm.file);
664 close_fail:
665         if (cprm.file)
666                 filp_close(cprm.file, NULL);
667 fail_dropcount:
668         if (ispipe)
669                 atomic_dec(&core_dump_count);
670 fail_unlock:
671         kfree(cn.corename);
672 fail_corename:
673         coredump_finish(mm, core_dumped);
674         revert_creds(old_cred);
675 fail_creds:
676         put_cred(cred);
677 fail:
678         return;
679 }
680
681 /*
682  * Core dumping helper functions.  These are the only things you should
683  * do on a core-file: use only these functions to write out all the
684  * necessary info.
685  */
686 int dump_write(struct file *file, const void *addr, int nr)
687 {
688         return !dump_interrupted() &&
689                 access_ok(VERIFY_READ, addr, nr) &&
690                 file->f_op->write(file, addr, nr, &file->f_pos) == nr;
691 }
692 EXPORT_SYMBOL(dump_write);
693
694 int dump_seek(struct file *file, loff_t off)
695 {
696         int ret = 1;
697
698         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
699                 if (dump_interrupted() ||
700                     file->f_op->llseek(file, off, SEEK_CUR) < 0)
701                         return 0;
702         } else {
703                 char *buf = (char *)get_zeroed_page(GFP_KERNEL);
704
705                 if (!buf)
706                         return 0;
707                 while (off > 0) {
708                         unsigned long n = off;
709
710                         if (n > PAGE_SIZE)
711                                 n = PAGE_SIZE;
712                         if (!dump_write(file, buf, n)) {
713                                 ret = 0;
714                                 break;
715                         }
716                         off -= n;
717                 }
718                 free_page((unsigned long)buf);
719         }
720         return ret;
721 }
722 EXPORT_SYMBOL(dump_seek);