]> rtime.felk.cvut.cz Git - linux-imx.git/blob - fs/coredump.c
coredump: kill call_count, add core_name_size
[linux-imx.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 unsigned int core_pipe_limit;
49 char core_pattern[CORENAME_MAX_SIZE] = "core";
50 static int core_name_size = CORENAME_MAX_SIZE;
51
52 struct core_name {
53         char *corename;
54         int used, size;
55 };
56
57 /* The maximal length of core_pattern is also specified in sysctl.c */
58
59 static int expand_corename(struct core_name *cn, int size)
60 {
61         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
62
63         if (!corename)
64                 return -ENOMEM;
65
66         if (size > core_name_size) /* racy but harmless */
67                 core_name_size = size;
68
69         cn->size = ksize(corename);
70         cn->corename = corename;
71         return 0;
72 }
73
74 static int cn_vprintf(struct core_name *cn, const char *fmt, va_list arg)
75 {
76         int free, need;
77
78 again:
79         free = cn->size - cn->used;
80         need = vsnprintf(cn->corename + cn->used, free, fmt, arg);
81         if (need < free) {
82                 cn->used += need;
83                 return 0;
84         }
85
86         if (!expand_corename(cn, cn->size + need - free + 1))
87                 goto again;
88
89         return -ENOMEM;
90 }
91
92 static int cn_printf(struct core_name *cn, const char *fmt, ...)
93 {
94         va_list arg;
95         int ret;
96
97         va_start(arg, fmt);
98         ret = cn_vprintf(cn, fmt, arg);
99         va_end(arg);
100
101         return ret;
102 }
103
104 static int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
105 {
106         int cur = cn->used;
107         va_list arg;
108         int ret;
109
110         va_start(arg, fmt);
111         ret = cn_vprintf(cn, fmt, arg);
112         va_end(arg);
113
114         for (; cur < cn->used; ++cur) {
115                 if (cn->corename[cur] == '/')
116                         cn->corename[cur] = '!';
117         }
118         return ret;
119 }
120
121 static int cn_print_exe_file(struct core_name *cn)
122 {
123         struct file *exe_file;
124         char *pathbuf, *path;
125         int ret;
126
127         exe_file = get_mm_exe_file(current->mm);
128         if (!exe_file)
129                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
130
131         pathbuf = kmalloc(PATH_MAX, GFP_TEMPORARY);
132         if (!pathbuf) {
133                 ret = -ENOMEM;
134                 goto put_exe_file;
135         }
136
137         path = d_path(&exe_file->f_path, pathbuf, PATH_MAX);
138         if (IS_ERR(path)) {
139                 ret = PTR_ERR(path);
140                 goto free_buf;
141         }
142
143         ret = cn_esc_printf(cn, "%s", path);
144
145 free_buf:
146         kfree(pathbuf);
147 put_exe_file:
148         fput(exe_file);
149         return ret;
150 }
151
152 /* format_corename will inspect the pattern parameter, and output a
153  * name into corename, which must have space for at least
154  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
155  */
156 static int format_corename(struct core_name *cn, struct coredump_params *cprm)
157 {
158         const struct cred *cred = current_cred();
159         const char *pat_ptr = core_pattern;
160         int ispipe = (*pat_ptr == '|');
161         int pid_in_pattern = 0;
162         int err = 0;
163
164         cn->used = 0;
165         cn->corename = NULL;
166         if (expand_corename(cn, core_name_size))
167                 return -ENOMEM;
168
169         /* Repeat as long as we have more pattern to process and more output
170            space */
171         while (*pat_ptr) {
172                 if (*pat_ptr != '%') {
173                         if (*pat_ptr == 0)
174                                 goto out;
175                         err = cn_printf(cn, "%c", *pat_ptr++);
176                 } else {
177                         switch (*++pat_ptr) {
178                         /* single % at the end, drop that */
179                         case 0:
180                                 goto out;
181                         /* Double percent, output one percent */
182                         case '%':
183                                 err = cn_printf(cn, "%c", '%');
184                                 break;
185                         /* pid */
186                         case 'p':
187                                 pid_in_pattern = 1;
188                                 err = cn_printf(cn, "%d",
189                                               task_tgid_vnr(current));
190                                 break;
191                         /* uid */
192                         case 'u':
193                                 err = cn_printf(cn, "%d", cred->uid);
194                                 break;
195                         /* gid */
196                         case 'g':
197                                 err = cn_printf(cn, "%d", cred->gid);
198                                 break;
199                         case 'd':
200                                 err = cn_printf(cn, "%d",
201                                         __get_dumpable(cprm->mm_flags));
202                                 break;
203                         /* signal that caused the coredump */
204                         case 's':
205                                 err = cn_printf(cn, "%ld", cprm->siginfo->si_signo);
206                                 break;
207                         /* UNIX time of coredump */
208                         case 't': {
209                                 struct timeval tv;
210                                 do_gettimeofday(&tv);
211                                 err = cn_printf(cn, "%lu", tv.tv_sec);
212                                 break;
213                         }
214                         /* hostname */
215                         case 'h':
216                                 down_read(&uts_sem);
217                                 err = cn_esc_printf(cn, "%s",
218                                               utsname()->nodename);
219                                 up_read(&uts_sem);
220                                 break;
221                         /* executable */
222                         case 'e':
223                                 err = cn_esc_printf(cn, "%s", current->comm);
224                                 break;
225                         case 'E':
226                                 err = cn_print_exe_file(cn);
227                                 break;
228                         /* core limit size */
229                         case 'c':
230                                 err = cn_printf(cn, "%lu",
231                                               rlimit(RLIMIT_CORE));
232                                 break;
233                         default:
234                                 break;
235                         }
236                         ++pat_ptr;
237                 }
238
239                 if (err)
240                         return err;
241         }
242
243         /* Backward compatibility with core_uses_pid:
244          *
245          * If core_pattern does not include a %p (as is the default)
246          * and core_uses_pid is set, then .%pid will be appended to
247          * the filename. Do not do this for piped commands. */
248         if (!ispipe && !pid_in_pattern && core_uses_pid) {
249                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
250                 if (err)
251                         return err;
252         }
253 out:
254         return ispipe;
255 }
256
257 static int zap_process(struct task_struct *start, int exit_code)
258 {
259         struct task_struct *t;
260         int nr = 0;
261
262         start->signal->group_exit_code = exit_code;
263         start->signal->group_stop_count = 0;
264
265         t = start;
266         do {
267                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
268                 if (t != current && t->mm) {
269                         sigaddset(&t->pending.signal, SIGKILL);
270                         signal_wake_up(t, 1);
271                         nr++;
272                 }
273         } while_each_thread(start, t);
274
275         return nr;
276 }
277
278 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
279                         struct core_state *core_state, int exit_code)
280 {
281         struct task_struct *g, *p;
282         unsigned long flags;
283         int nr = -EAGAIN;
284
285         spin_lock_irq(&tsk->sighand->siglock);
286         if (!signal_group_exit(tsk->signal)) {
287                 mm->core_state = core_state;
288                 nr = zap_process(tsk, exit_code);
289                 tsk->signal->group_exit_task = tsk;
290                 /* ignore all signals except SIGKILL, see prepare_signal() */
291                 tsk->signal->flags = SIGNAL_GROUP_COREDUMP;
292                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
293         }
294         spin_unlock_irq(&tsk->sighand->siglock);
295         if (unlikely(nr < 0))
296                 return nr;
297
298         tsk->flags = PF_DUMPCORE;
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                                         p->signal->flags = SIGNAL_GROUP_EXIT;
344                                         unlock_task_sighand(p, &flags);
345                                 }
346                                 break;
347                         }
348                 } while_each_thread(g, p);
349         }
350         rcu_read_unlock();
351 done:
352         atomic_set(&core_state->nr_threads, nr);
353         return nr;
354 }
355
356 static int coredump_wait(int exit_code, struct core_state *core_state)
357 {
358         struct task_struct *tsk = current;
359         struct mm_struct *mm = tsk->mm;
360         int core_waiters = -EBUSY;
361
362         init_completion(&core_state->startup);
363         core_state->dumper.task = tsk;
364         core_state->dumper.next = NULL;
365
366         down_write(&mm->mmap_sem);
367         if (!mm->core_state)
368                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
369         up_write(&mm->mmap_sem);
370
371         if (core_waiters > 0) {
372                 struct core_thread *ptr;
373
374                 wait_for_completion(&core_state->startup);
375                 /*
376                  * Wait for all the threads to become inactive, so that
377                  * all the thread context (extended register state, like
378                  * fpu etc) gets copied to the memory.
379                  */
380                 ptr = core_state->dumper.next;
381                 while (ptr != NULL) {
382                         wait_task_inactive(ptr->task, 0);
383                         ptr = ptr->next;
384                 }
385         }
386
387         return core_waiters;
388 }
389
390 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
391 {
392         struct core_thread *curr, *next;
393         struct task_struct *task;
394
395         spin_lock_irq(&current->sighand->siglock);
396         if (core_dumped && !__fatal_signal_pending(current))
397                 current->signal->group_exit_code |= 0x80;
398         current->signal->group_exit_task = NULL;
399         current->signal->flags = SIGNAL_GROUP_EXIT;
400         spin_unlock_irq(&current->sighand->siglock);
401
402         next = mm->core_state->dumper.next;
403         while ((curr = next) != NULL) {
404                 next = curr->next;
405                 task = curr->task;
406                 /*
407                  * see exit_mm(), curr->task must not see
408                  * ->task == NULL before we read ->next.
409                  */
410                 smp_mb();
411                 curr->task = NULL;
412                 wake_up_process(task);
413         }
414
415         mm->core_state = NULL;
416 }
417
418 static bool dump_interrupted(void)
419 {
420         /*
421          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
422          * can do try_to_freeze() and check __fatal_signal_pending(),
423          * but then we need to teach dump_write() to restart and clear
424          * TIF_SIGPENDING.
425          */
426         return signal_pending(current);
427 }
428
429 static void wait_for_dump_helpers(struct file *file)
430 {
431         struct pipe_inode_info *pipe = file->private_data;
432
433         pipe_lock(pipe);
434         pipe->readers++;
435         pipe->writers--;
436         wake_up_interruptible_sync(&pipe->wait);
437         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
438         pipe_unlock(pipe);
439
440         /*
441          * We actually want wait_event_freezable() but then we need
442          * to clear TIF_SIGPENDING and improve dump_interrupted().
443          */
444         wait_event_interruptible(pipe->wait, pipe->readers == 1);
445
446         pipe_lock(pipe);
447         pipe->readers--;
448         pipe->writers++;
449         pipe_unlock(pipe);
450 }
451
452 /*
453  * umh_pipe_setup
454  * helper function to customize the process used
455  * to collect the core in userspace.  Specifically
456  * it sets up a pipe and installs it as fd 0 (stdin)
457  * for the process.  Returns 0 on success, or
458  * PTR_ERR on failure.
459  * Note that it also sets the core limit to 1.  This
460  * is a special value that we use to trap recursive
461  * core dumps
462  */
463 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
464 {
465         struct file *files[2];
466         struct coredump_params *cp = (struct coredump_params *)info->data;
467         int err = create_pipe_files(files, 0);
468         if (err)
469                 return err;
470
471         cp->file = files[1];
472
473         err = replace_fd(0, files[0], 0);
474         fput(files[0]);
475         /* and disallow core files too */
476         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
477
478         return err;
479 }
480
481 void do_coredump(siginfo_t *siginfo)
482 {
483         struct core_state core_state;
484         struct core_name cn;
485         struct mm_struct *mm = current->mm;
486         struct linux_binfmt * binfmt;
487         const struct cred *old_cred;
488         struct cred *cred;
489         int retval = 0;
490         int flag = 0;
491         int ispipe;
492         struct files_struct *displaced;
493         bool need_nonrelative = false;
494         bool core_dumped = false;
495         static atomic_t core_dump_count = ATOMIC_INIT(0);
496         struct coredump_params cprm = {
497                 .siginfo = siginfo,
498                 .regs = signal_pt_regs(),
499                 .limit = rlimit(RLIMIT_CORE),
500                 /*
501                  * We must use the same mm->flags while dumping core to avoid
502                  * inconsistency of bit flags, since this flag is not protected
503                  * by any locks.
504                  */
505                 .mm_flags = mm->flags,
506         };
507
508         audit_core_dumps(siginfo->si_signo);
509
510         binfmt = mm->binfmt;
511         if (!binfmt || !binfmt->core_dump)
512                 goto fail;
513         if (!__get_dumpable(cprm.mm_flags))
514                 goto fail;
515
516         cred = prepare_creds();
517         if (!cred)
518                 goto fail;
519         /*
520          * We cannot trust fsuid as being the "true" uid of the process
521          * nor do we know its entire history. We only know it was tainted
522          * so we dump it as root in mode 2, and only into a controlled
523          * environment (pipe handler or fully qualified path).
524          */
525         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
526                 /* Setuid core dump mode */
527                 flag = O_EXCL;          /* Stop rewrite attacks */
528                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
529                 need_nonrelative = true;
530         }
531
532         retval = coredump_wait(siginfo->si_signo, &core_state);
533         if (retval < 0)
534                 goto fail_creds;
535
536         old_cred = override_creds(cred);
537
538         ispipe = format_corename(&cn, &cprm);
539
540         if (ispipe) {
541                 int dump_count;
542                 char **helper_argv;
543                 struct subprocess_info *sub_info;
544
545                 if (ispipe < 0) {
546                         printk(KERN_WARNING "format_corename failed\n");
547                         printk(KERN_WARNING "Aborting core\n");
548                         goto fail_unlock;
549                 }
550
551                 if (cprm.limit == 1) {
552                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
553                          *
554                          * Normally core limits are irrelevant to pipes, since
555                          * we're not writing to the file system, but we use
556                          * cprm.limit of 1 here as a speacial value, this is a
557                          * consistent way to catch recursive crashes.
558                          * We can still crash if the core_pattern binary sets
559                          * RLIM_CORE = !1, but it runs as root, and can do
560                          * lots of stupid things.
561                          *
562                          * Note that we use task_tgid_vnr here to grab the pid
563                          * of the process group leader.  That way we get the
564                          * right pid if a thread in a multi-threaded
565                          * core_pattern process dies.
566                          */
567                         printk(KERN_WARNING
568                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
569                                 task_tgid_vnr(current), current->comm);
570                         printk(KERN_WARNING "Aborting core\n");
571                         goto fail_unlock;
572                 }
573                 cprm.limit = RLIM_INFINITY;
574
575                 dump_count = atomic_inc_return(&core_dump_count);
576                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
577                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
578                                task_tgid_vnr(current), current->comm);
579                         printk(KERN_WARNING "Skipping core dump\n");
580                         goto fail_dropcount;
581                 }
582
583                 helper_argv = argv_split(GFP_KERNEL, cn.corename+1, NULL);
584                 if (!helper_argv) {
585                         printk(KERN_WARNING "%s failed to allocate memory\n",
586                                __func__);
587                         goto fail_dropcount;
588                 }
589
590                 retval = -ENOMEM;
591                 sub_info = call_usermodehelper_setup(helper_argv[0],
592                                                 helper_argv, NULL, GFP_KERNEL,
593                                                 umh_pipe_setup, NULL, &cprm);
594                 if (sub_info)
595                         retval = call_usermodehelper_exec(sub_info,
596                                                           UMH_WAIT_EXEC);
597
598                 argv_free(helper_argv);
599                 if (retval) {
600                         printk(KERN_INFO "Core dump to %s pipe failed\n",
601                                cn.corename);
602                         goto close_fail;
603                 }
604         } else {
605                 struct inode *inode;
606
607                 if (cprm.limit < binfmt->min_coredump)
608                         goto fail_unlock;
609
610                 if (need_nonrelative && cn.corename[0] != '/') {
611                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
612                                 "to fully qualified path!\n",
613                                 task_tgid_vnr(current), current->comm);
614                         printk(KERN_WARNING "Skipping core dump\n");
615                         goto fail_unlock;
616                 }
617
618                 cprm.file = filp_open(cn.corename,
619                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
620                                  0600);
621                 if (IS_ERR(cprm.file))
622                         goto fail_unlock;
623
624                 inode = file_inode(cprm.file);
625                 if (inode->i_nlink > 1)
626                         goto close_fail;
627                 if (d_unhashed(cprm.file->f_path.dentry))
628                         goto close_fail;
629                 /*
630                  * AK: actually i see no reason to not allow this for named
631                  * pipes etc, but keep the previous behaviour for now.
632                  */
633                 if (!S_ISREG(inode->i_mode))
634                         goto close_fail;
635                 /*
636                  * Dont allow local users get cute and trick others to coredump
637                  * into their pre-created files.
638                  */
639                 if (!uid_eq(inode->i_uid, current_fsuid()))
640                         goto close_fail;
641                 if (!cprm.file->f_op || !cprm.file->f_op->write)
642                         goto close_fail;
643                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
644                         goto close_fail;
645         }
646
647         /* get us an unshared descriptor table; almost always a no-op */
648         retval = unshare_files(&displaced);
649         if (retval)
650                 goto close_fail;
651         if (displaced)
652                 put_files_struct(displaced);
653         if (!dump_interrupted()) {
654                 file_start_write(cprm.file);
655                 core_dumped = binfmt->core_dump(&cprm);
656                 file_end_write(cprm.file);
657         }
658         if (ispipe && core_pipe_limit)
659                 wait_for_dump_helpers(cprm.file);
660 close_fail:
661         if (cprm.file)
662                 filp_close(cprm.file, NULL);
663 fail_dropcount:
664         if (ispipe)
665                 atomic_dec(&core_dump_count);
666 fail_unlock:
667         kfree(cn.corename);
668         coredump_finish(mm, core_dumped);
669         revert_creds(old_cred);
670 fail_creds:
671         put_cred(cred);
672 fail:
673         return;
674 }
675
676 /*
677  * Core dumping helper functions.  These are the only things you should
678  * do on a core-file: use only these functions to write out all the
679  * necessary info.
680  */
681 int dump_write(struct file *file, const void *addr, int nr)
682 {
683         return !dump_interrupted() &&
684                 access_ok(VERIFY_READ, addr, nr) &&
685                 file->f_op->write(file, addr, nr, &file->f_pos) == nr;
686 }
687 EXPORT_SYMBOL(dump_write);
688
689 int dump_seek(struct file *file, loff_t off)
690 {
691         int ret = 1;
692
693         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
694                 if (dump_interrupted() ||
695                     file->f_op->llseek(file, off, SEEK_CUR) < 0)
696                         return 0;
697         } else {
698                 char *buf = (char *)get_zeroed_page(GFP_KERNEL);
699
700                 if (!buf)
701                         return 0;
702                 while (off > 0) {
703                         unsigned long n = off;
704
705                         if (n > PAGE_SIZE)
706                                 n = PAGE_SIZE;
707                         if (!dump_write(file, buf, n)) {
708                                 ret = 0;
709                                 break;
710                         }
711                         off -= n;
712                 }
713                 free_page((unsigned long)buf);
714         }
715         return ret;
716 }
717 EXPORT_SYMBOL(dump_seek);