GNU Linux-libre 5.10.219-gnu1
[releases.git] / fs / coredump.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/slab.h>
3 #include <linux/file.h>
4 #include <linux/fdtable.h>
5 #include <linux/freezer.h>
6 #include <linux/mm.h>
7 #include <linux/stat.h>
8 #include <linux/fcntl.h>
9 #include <linux/swap.h>
10 #include <linux/ctype.h>
11 #include <linux/string.h>
12 #include <linux/init.h>
13 #include <linux/pagemap.h>
14 #include <linux/perf_event.h>
15 #include <linux/highmem.h>
16 #include <linux/spinlock.h>
17 #include <linux/key.h>
18 #include <linux/personality.h>
19 #include <linux/binfmts.h>
20 #include <linux/coredump.h>
21 #include <linux/sched/coredump.h>
22 #include <linux/sched/signal.h>
23 #include <linux/sched/task_stack.h>
24 #include <linux/utsname.h>
25 #include <linux/pid_namespace.h>
26 #include <linux/module.h>
27 #include <linux/namei.h>
28 #include <linux/mount.h>
29 #include <linux/security.h>
30 #include <linux/syscalls.h>
31 #include <linux/tsacct_kern.h>
32 #include <linux/cn_proc.h>
33 #include <linux/audit.h>
34 #include <linux/tracehook.h>
35 #include <linux/kmod.h>
36 #include <linux/fsnotify.h>
37 #include <linux/fs_struct.h>
38 #include <linux/pipe_fs_i.h>
39 #include <linux/oom.h>
40 #include <linux/compat.h>
41 #include <linux/fs.h>
42 #include <linux/path.h>
43 #include <linux/timekeeping.h>
44 #include <linux/elf.h>
45
46 #include <linux/uaccess.h>
47 #include <asm/mmu_context.h>
48 #include <asm/tlb.h>
49 #include <asm/exec.h>
50
51 #include <trace/events/task.h>
52 #include "internal.h"
53
54 #include <trace/events/sched.h>
55
56 static bool dump_vma_snapshot(struct coredump_params *cprm);
57 static void free_vma_snapshot(struct coredump_params *cprm);
58
59 int core_uses_pid;
60 unsigned int core_pipe_limit;
61 char core_pattern[CORENAME_MAX_SIZE] = "core";
62 static int core_name_size = CORENAME_MAX_SIZE;
63
64 struct core_name {
65         char *corename;
66         int used, size;
67 };
68
69 /* The maximal length of core_pattern is also specified in sysctl.c */
70
71 static int expand_corename(struct core_name *cn, int size)
72 {
73         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
74
75         if (!corename)
76                 return -ENOMEM;
77
78         if (size > core_name_size) /* racy but harmless */
79                 core_name_size = size;
80
81         cn->size = ksize(corename);
82         cn->corename = corename;
83         return 0;
84 }
85
86 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
87                                      va_list arg)
88 {
89         int free, need;
90         va_list arg_copy;
91
92 again:
93         free = cn->size - cn->used;
94
95         va_copy(arg_copy, arg);
96         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
97         va_end(arg_copy);
98
99         if (need < free) {
100                 cn->used += need;
101                 return 0;
102         }
103
104         if (!expand_corename(cn, cn->size + need - free + 1))
105                 goto again;
106
107         return -ENOMEM;
108 }
109
110 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
111 {
112         va_list arg;
113         int ret;
114
115         va_start(arg, fmt);
116         ret = cn_vprintf(cn, fmt, arg);
117         va_end(arg);
118
119         return ret;
120 }
121
122 static __printf(2, 3)
123 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
124 {
125         int cur = cn->used;
126         va_list arg;
127         int ret;
128
129         va_start(arg, fmt);
130         ret = cn_vprintf(cn, fmt, arg);
131         va_end(arg);
132
133         if (ret == 0) {
134                 /*
135                  * Ensure that this coredump name component can't cause the
136                  * resulting corefile path to consist of a ".." or ".".
137                  */
138                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
139                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
140                                 && cn->corename[cur+1] == '.'))
141                         cn->corename[cur] = '!';
142
143                 /*
144                  * Empty names are fishy and could be used to create a "//" in a
145                  * corefile name, causing the coredump to happen one directory
146                  * level too high. Enforce that all components of the core
147                  * pattern are at least one character long.
148                  */
149                 if (cn->used == cur)
150                         ret = cn_printf(cn, "!");
151         }
152
153         for (; cur < cn->used; ++cur) {
154                 if (cn->corename[cur] == '/')
155                         cn->corename[cur] = '!';
156         }
157         return ret;
158 }
159
160 static int cn_print_exe_file(struct core_name *cn, bool name_only)
161 {
162         struct file *exe_file;
163         char *pathbuf, *path, *ptr;
164         int ret;
165
166         exe_file = get_mm_exe_file(current->mm);
167         if (!exe_file)
168                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
169
170         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
171         if (!pathbuf) {
172                 ret = -ENOMEM;
173                 goto put_exe_file;
174         }
175
176         path = file_path(exe_file, pathbuf, PATH_MAX);
177         if (IS_ERR(path)) {
178                 ret = PTR_ERR(path);
179                 goto free_buf;
180         }
181
182         if (name_only) {
183                 ptr = strrchr(path, '/');
184                 if (ptr)
185                         path = ptr + 1;
186         }
187         ret = cn_esc_printf(cn, "%s", path);
188
189 free_buf:
190         kfree(pathbuf);
191 put_exe_file:
192         fput(exe_file);
193         return ret;
194 }
195
196 /* format_corename will inspect the pattern parameter, and output a
197  * name into corename, which must have space for at least
198  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
199  */
200 static int format_corename(struct core_name *cn, struct coredump_params *cprm,
201                            size_t **argv, int *argc)
202 {
203         const struct cred *cred = current_cred();
204         const char *pat_ptr = core_pattern;
205         int ispipe = (*pat_ptr == '|');
206         bool was_space = false;
207         int pid_in_pattern = 0;
208         int err = 0;
209
210         cn->used = 0;
211         cn->corename = NULL;
212         if (expand_corename(cn, core_name_size))
213                 return -ENOMEM;
214         cn->corename[0] = '\0';
215
216         if (ispipe) {
217                 int argvs = sizeof(core_pattern) / 2;
218                 (*argv) = kmalloc_array(argvs, sizeof(**argv), GFP_KERNEL);
219                 if (!(*argv))
220                         return -ENOMEM;
221                 (*argv)[(*argc)++] = 0;
222                 ++pat_ptr;
223                 if (!(*pat_ptr))
224                         return -ENOMEM;
225         }
226
227         /* Repeat as long as we have more pattern to process and more output
228            space */
229         while (*pat_ptr) {
230                 /*
231                  * Split on spaces before doing template expansion so that
232                  * %e and %E don't get split if they have spaces in them
233                  */
234                 if (ispipe) {
235                         if (isspace(*pat_ptr)) {
236                                 if (cn->used != 0)
237                                         was_space = true;
238                                 pat_ptr++;
239                                 continue;
240                         } else if (was_space) {
241                                 was_space = false;
242                                 err = cn_printf(cn, "%c", '\0');
243                                 if (err)
244                                         return err;
245                                 (*argv)[(*argc)++] = cn->used;
246                         }
247                 }
248                 if (*pat_ptr != '%') {
249                         err = cn_printf(cn, "%c", *pat_ptr++);
250                 } else {
251                         switch (*++pat_ptr) {
252                         /* single % at the end, drop that */
253                         case 0:
254                                 goto out;
255                         /* Double percent, output one percent */
256                         case '%':
257                                 err = cn_printf(cn, "%c", '%');
258                                 break;
259                         /* pid */
260                         case 'p':
261                                 pid_in_pattern = 1;
262                                 err = cn_printf(cn, "%d",
263                                               task_tgid_vnr(current));
264                                 break;
265                         /* global pid */
266                         case 'P':
267                                 err = cn_printf(cn, "%d",
268                                               task_tgid_nr(current));
269                                 break;
270                         case 'i':
271                                 err = cn_printf(cn, "%d",
272                                               task_pid_vnr(current));
273                                 break;
274                         case 'I':
275                                 err = cn_printf(cn, "%d",
276                                               task_pid_nr(current));
277                                 break;
278                         /* uid */
279                         case 'u':
280                                 err = cn_printf(cn, "%u",
281                                                 from_kuid(&init_user_ns,
282                                                           cred->uid));
283                                 break;
284                         /* gid */
285                         case 'g':
286                                 err = cn_printf(cn, "%u",
287                                                 from_kgid(&init_user_ns,
288                                                           cred->gid));
289                                 break;
290                         case 'd':
291                                 err = cn_printf(cn, "%d",
292                                         __get_dumpable(cprm->mm_flags));
293                                 break;
294                         /* signal that caused the coredump */
295                         case 's':
296                                 err = cn_printf(cn, "%d",
297                                                 cprm->siginfo->si_signo);
298                                 break;
299                         /* UNIX time of coredump */
300                         case 't': {
301                                 time64_t time;
302
303                                 time = ktime_get_real_seconds();
304                                 err = cn_printf(cn, "%lld", time);
305                                 break;
306                         }
307                         /* hostname */
308                         case 'h':
309                                 down_read(&uts_sem);
310                                 err = cn_esc_printf(cn, "%s",
311                                               utsname()->nodename);
312                                 up_read(&uts_sem);
313                                 break;
314                         /* executable, could be changed by prctl PR_SET_NAME etc */
315                         case 'e':
316                                 err = cn_esc_printf(cn, "%s", current->comm);
317                                 break;
318                         /* file name of executable */
319                         case 'f':
320                                 err = cn_print_exe_file(cn, true);
321                                 break;
322                         case 'E':
323                                 err = cn_print_exe_file(cn, false);
324                                 break;
325                         /* core limit size */
326                         case 'c':
327                                 err = cn_printf(cn, "%lu",
328                                               rlimit(RLIMIT_CORE));
329                                 break;
330                         default:
331                                 break;
332                         }
333                         ++pat_ptr;
334                 }
335
336                 if (err)
337                         return err;
338         }
339
340 out:
341         /* Backward compatibility with core_uses_pid:
342          *
343          * If core_pattern does not include a %p (as is the default)
344          * and core_uses_pid is set, then .%pid will be appended to
345          * the filename. Do not do this for piped commands. */
346         if (!ispipe && !pid_in_pattern && core_uses_pid) {
347                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
348                 if (err)
349                         return err;
350         }
351         return ispipe;
352 }
353
354 static int zap_process(struct task_struct *start, int exit_code, int flags)
355 {
356         struct task_struct *t;
357         int nr = 0;
358
359         /* ignore all signals except SIGKILL, see prepare_signal() */
360         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
361         start->signal->group_exit_code = exit_code;
362         start->signal->group_stop_count = 0;
363
364         for_each_thread(start, t) {
365                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
366                 if (t != current && t->mm) {
367                         sigaddset(&t->pending.signal, SIGKILL);
368                         signal_wake_up(t, 1);
369                         nr++;
370                 }
371         }
372
373         return nr;
374 }
375
376 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
377                         struct core_state *core_state, int exit_code)
378 {
379         struct task_struct *g, *p;
380         unsigned long flags;
381         int nr = -EAGAIN;
382
383         spin_lock_irq(&tsk->sighand->siglock);
384         if (!signal_group_exit(tsk->signal)) {
385                 mm->core_state = core_state;
386                 tsk->signal->group_exit_task = tsk;
387                 nr = zap_process(tsk, exit_code, 0);
388                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
389         }
390         spin_unlock_irq(&tsk->sighand->siglock);
391         if (unlikely(nr < 0))
392                 return nr;
393
394         tsk->flags |= PF_DUMPCORE;
395         if (atomic_read(&mm->mm_users) == nr + 1)
396                 goto done;
397         /*
398          * We should find and kill all tasks which use this mm, and we should
399          * count them correctly into ->nr_threads. We don't take tasklist
400          * lock, but this is safe wrt:
401          *
402          * fork:
403          *      None of sub-threads can fork after zap_process(leader). All
404          *      processes which were created before this point should be
405          *      visible to zap_threads() because copy_process() adds the new
406          *      process to the tail of init_task.tasks list, and lock/unlock
407          *      of ->siglock provides a memory barrier.
408          *
409          * do_exit:
410          *      The caller holds mm->mmap_lock. This means that the task which
411          *      uses this mm can't pass exit_mm(), so it can't exit or clear
412          *      its ->mm.
413          *
414          * de_thread:
415          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
416          *      we must see either old or new leader, this does not matter.
417          *      However, it can change p->sighand, so lock_task_sighand(p)
418          *      must be used. Since p->mm != NULL and we hold ->mmap_lock
419          *      it can't fail.
420          *
421          *      Note also that "g" can be the old leader with ->mm == NULL
422          *      and already unhashed and thus removed from ->thread_group.
423          *      This is OK, __unhash_process()->list_del_rcu() does not
424          *      clear the ->next pointer, we will find the new leader via
425          *      next_thread().
426          */
427         rcu_read_lock();
428         for_each_process(g) {
429                 if (g == tsk->group_leader)
430                         continue;
431                 if (g->flags & PF_KTHREAD)
432                         continue;
433
434                 for_each_thread(g, p) {
435                         if (unlikely(!p->mm))
436                                 continue;
437                         if (unlikely(p->mm == mm)) {
438                                 lock_task_sighand(p, &flags);
439                                 nr += zap_process(p, exit_code,
440                                                         SIGNAL_GROUP_EXIT);
441                                 unlock_task_sighand(p, &flags);
442                         }
443                         break;
444                 }
445         }
446         rcu_read_unlock();
447 done:
448         atomic_set(&core_state->nr_threads, nr);
449         return nr;
450 }
451
452 static int coredump_wait(int exit_code, struct core_state *core_state)
453 {
454         struct task_struct *tsk = current;
455         struct mm_struct *mm = tsk->mm;
456         int core_waiters = -EBUSY;
457
458         init_completion(&core_state->startup);
459         core_state->dumper.task = tsk;
460         core_state->dumper.next = NULL;
461
462         if (mmap_write_lock_killable(mm))
463                 return -EINTR;
464
465         if (!mm->core_state)
466                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
467         mmap_write_unlock(mm);
468
469         if (core_waiters > 0) {
470                 struct core_thread *ptr;
471
472                 freezer_do_not_count();
473                 wait_for_completion(&core_state->startup);
474                 freezer_count();
475                 /*
476                  * Wait for all the threads to become inactive, so that
477                  * all the thread context (extended register state, like
478                  * fpu etc) gets copied to the memory.
479                  */
480                 ptr = core_state->dumper.next;
481                 while (ptr != NULL) {
482                         wait_task_inactive(ptr->task, 0);
483                         ptr = ptr->next;
484                 }
485         }
486
487         return core_waiters;
488 }
489
490 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
491 {
492         struct core_thread *curr, *next;
493         struct task_struct *task;
494
495         spin_lock_irq(&current->sighand->siglock);
496         if (core_dumped && !__fatal_signal_pending(current))
497                 current->signal->group_exit_code |= 0x80;
498         current->signal->group_exit_task = NULL;
499         current->signal->flags = SIGNAL_GROUP_EXIT;
500         spin_unlock_irq(&current->sighand->siglock);
501
502         next = mm->core_state->dumper.next;
503         while ((curr = next) != NULL) {
504                 next = curr->next;
505                 task = curr->task;
506                 /*
507                  * see exit_mm(), curr->task must not see
508                  * ->task == NULL before we read ->next.
509                  */
510                 smp_mb();
511                 curr->task = NULL;
512                 wake_up_process(task);
513         }
514
515         mm->core_state = NULL;
516 }
517
518 static bool dump_interrupted(void)
519 {
520         /*
521          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
522          * can do try_to_freeze() and check __fatal_signal_pending(),
523          * but then we need to teach dump_write() to restart and clear
524          * TIF_SIGPENDING.
525          */
526         return fatal_signal_pending(current) || freezing(current);
527 }
528
529 static void wait_for_dump_helpers(struct file *file)
530 {
531         struct pipe_inode_info *pipe = file->private_data;
532
533         pipe_lock(pipe);
534         pipe->readers++;
535         pipe->writers--;
536         wake_up_interruptible_sync(&pipe->rd_wait);
537         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
538         pipe_unlock(pipe);
539
540         /*
541          * We actually want wait_event_freezable() but then we need
542          * to clear TIF_SIGPENDING and improve dump_interrupted().
543          */
544         wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
545
546         pipe_lock(pipe);
547         pipe->readers--;
548         pipe->writers++;
549         pipe_unlock(pipe);
550 }
551
552 /*
553  * umh_pipe_setup
554  * helper function to customize the process used
555  * to collect the core in userspace.  Specifically
556  * it sets up a pipe and installs it as fd 0 (stdin)
557  * for the process.  Returns 0 on success, or
558  * PTR_ERR on failure.
559  * Note that it also sets the core limit to 1.  This
560  * is a special value that we use to trap recursive
561  * core dumps
562  */
563 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
564 {
565         struct file *files[2];
566         struct coredump_params *cp = (struct coredump_params *)info->data;
567         int err = create_pipe_files(files, 0);
568         if (err)
569                 return err;
570
571         cp->file = files[1];
572
573         err = replace_fd(0, files[0], 0);
574         fput(files[0]);
575         /* and disallow core files too */
576         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
577
578         return err;
579 }
580
581 void do_coredump(const kernel_siginfo_t *siginfo)
582 {
583         struct core_state core_state;
584         struct core_name cn;
585         struct mm_struct *mm = current->mm;
586         struct linux_binfmt * binfmt;
587         const struct cred *old_cred;
588         struct cred *cred;
589         int retval = 0;
590         int ispipe;
591         size_t *argv = NULL;
592         int argc = 0;
593         struct files_struct *displaced;
594         /* require nonrelative corefile path and be extra careful */
595         bool need_suid_safe = false;
596         bool core_dumped = false;
597         static atomic_t core_dump_count = ATOMIC_INIT(0);
598         struct coredump_params cprm = {
599                 .siginfo = siginfo,
600                 .regs = signal_pt_regs(),
601                 .limit = rlimit(RLIMIT_CORE),
602                 /*
603                  * We must use the same mm->flags while dumping core to avoid
604                  * inconsistency of bit flags, since this flag is not protected
605                  * by any locks.
606                  */
607                 .mm_flags = mm->flags,
608                 .vma_meta = NULL,
609         };
610
611         audit_core_dumps(siginfo->si_signo);
612
613         binfmt = mm->binfmt;
614         if (!binfmt || !binfmt->core_dump)
615                 goto fail;
616         if (!__get_dumpable(cprm.mm_flags))
617                 goto fail;
618
619         cred = prepare_creds();
620         if (!cred)
621                 goto fail;
622         /*
623          * We cannot trust fsuid as being the "true" uid of the process
624          * nor do we know its entire history. We only know it was tainted
625          * so we dump it as root in mode 2, and only into a controlled
626          * environment (pipe handler or fully qualified path).
627          */
628         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
629                 /* Setuid core dump mode */
630                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
631                 need_suid_safe = true;
632         }
633
634         retval = coredump_wait(siginfo->si_signo, &core_state);
635         if (retval < 0)
636                 goto fail_creds;
637
638         old_cred = override_creds(cred);
639
640         ispipe = format_corename(&cn, &cprm, &argv, &argc);
641
642         if (ispipe) {
643                 int argi;
644                 int dump_count;
645                 char **helper_argv;
646                 struct subprocess_info *sub_info;
647
648                 if (ispipe < 0) {
649                         printk(KERN_WARNING "format_corename failed\n");
650                         printk(KERN_WARNING "Aborting core\n");
651                         goto fail_unlock;
652                 }
653
654                 if (cprm.limit == 1) {
655                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
656                          *
657                          * Normally core limits are irrelevant to pipes, since
658                          * we're not writing to the file system, but we use
659                          * cprm.limit of 1 here as a special value, this is a
660                          * consistent way to catch recursive crashes.
661                          * We can still crash if the core_pattern binary sets
662                          * RLIM_CORE = !1, but it runs as root, and can do
663                          * lots of stupid things.
664                          *
665                          * Note that we use task_tgid_vnr here to grab the pid
666                          * of the process group leader.  That way we get the
667                          * right pid if a thread in a multi-threaded
668                          * core_pattern process dies.
669                          */
670                         printk(KERN_WARNING
671                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
672                                 task_tgid_vnr(current), current->comm);
673                         printk(KERN_WARNING "Aborting core\n");
674                         goto fail_unlock;
675                 }
676                 cprm.limit = RLIM_INFINITY;
677
678                 dump_count = atomic_inc_return(&core_dump_count);
679                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
680                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
681                                task_tgid_vnr(current), current->comm);
682                         printk(KERN_WARNING "Skipping core dump\n");
683                         goto fail_dropcount;
684                 }
685
686                 helper_argv = kmalloc_array(argc + 1, sizeof(*helper_argv),
687                                             GFP_KERNEL);
688                 if (!helper_argv) {
689                         printk(KERN_WARNING "%s failed to allocate memory\n",
690                                __func__);
691                         goto fail_dropcount;
692                 }
693                 for (argi = 0; argi < argc; argi++)
694                         helper_argv[argi] = cn.corename + argv[argi];
695                 helper_argv[argi] = NULL;
696
697                 retval = -ENOMEM;
698                 sub_info = call_usermodehelper_setup(helper_argv[0],
699                                                 helper_argv, NULL, GFP_KERNEL,
700                                                 umh_pipe_setup, NULL, &cprm);
701                 if (sub_info)
702                         retval = call_usermodehelper_exec(sub_info,
703                                                           UMH_WAIT_EXEC);
704
705                 kfree(helper_argv);
706                 if (retval) {
707                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
708                                cn.corename);
709                         goto close_fail;
710                 }
711         } else {
712                 struct inode *inode;
713                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
714                                  O_LARGEFILE | O_EXCL;
715
716                 if (cprm.limit < binfmt->min_coredump)
717                         goto fail_unlock;
718
719                 if (need_suid_safe && cn.corename[0] != '/') {
720                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
721                                 "to fully qualified path!\n",
722                                 task_tgid_vnr(current), current->comm);
723                         printk(KERN_WARNING "Skipping core dump\n");
724                         goto fail_unlock;
725                 }
726
727                 /*
728                  * Unlink the file if it exists unless this is a SUID
729                  * binary - in that case, we're running around with root
730                  * privs and don't want to unlink another user's coredump.
731                  */
732                 if (!need_suid_safe) {
733                         /*
734                          * If it doesn't exist, that's fine. If there's some
735                          * other problem, we'll catch it at the filp_open().
736                          */
737                         do_unlinkat(AT_FDCWD, getname_kernel(cn.corename));
738                 }
739
740                 /*
741                  * There is a race between unlinking and creating the
742                  * file, but if that causes an EEXIST here, that's
743                  * fine - another process raced with us while creating
744                  * the corefile, and the other process won. To userspace,
745                  * what matters is that at least one of the two processes
746                  * writes its coredump successfully, not which one.
747                  */
748                 if (need_suid_safe) {
749                         /*
750                          * Using user namespaces, normal user tasks can change
751                          * their current->fs->root to point to arbitrary
752                          * directories. Since the intention of the "only dump
753                          * with a fully qualified path" rule is to control where
754                          * coredumps may be placed using root privileges,
755                          * current->fs->root must not be used. Instead, use the
756                          * root directory of init_task.
757                          */
758                         struct path root;
759
760                         task_lock(&init_task);
761                         get_fs_root(init_task.fs, &root);
762                         task_unlock(&init_task);
763                         cprm.file = file_open_root(root.dentry, root.mnt,
764                                 cn.corename, open_flags, 0600);
765                         path_put(&root);
766                 } else {
767                         cprm.file = filp_open(cn.corename, open_flags, 0600);
768                 }
769                 if (IS_ERR(cprm.file))
770                         goto fail_unlock;
771
772                 inode = file_inode(cprm.file);
773                 if (inode->i_nlink > 1)
774                         goto close_fail;
775                 if (d_unhashed(cprm.file->f_path.dentry))
776                         goto close_fail;
777                 /*
778                  * AK: actually i see no reason to not allow this for named
779                  * pipes etc, but keep the previous behaviour for now.
780                  */
781                 if (!S_ISREG(inode->i_mode))
782                         goto close_fail;
783                 /*
784                  * Don't dump core if the filesystem changed owner or mode
785                  * of the file during file creation. This is an issue when
786                  * a process dumps core while its cwd is e.g. on a vfat
787                  * filesystem.
788                  */
789                 if (!uid_eq(inode->i_uid, current_fsuid()))
790                         goto close_fail;
791                 if ((inode->i_mode & 0677) != 0600)
792                         goto close_fail;
793                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
794                         goto close_fail;
795                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
796                         goto close_fail;
797         }
798
799         /* get us an unshared descriptor table; almost always a no-op */
800         retval = unshare_files(&displaced);
801         if (retval)
802                 goto close_fail;
803         if (displaced)
804                 put_files_struct(displaced);
805         if (!dump_interrupted()) {
806                 /*
807                  * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
808                  * have this set to NULL.
809                  */
810                 if (!cprm.file) {
811                         pr_info("Core dump to |%s disabled\n", cn.corename);
812                         goto close_fail;
813                 }
814                 if (!dump_vma_snapshot(&cprm))
815                         goto close_fail;
816
817                 file_start_write(cprm.file);
818                 core_dumped = binfmt->core_dump(&cprm);
819                 file_end_write(cprm.file);
820                 free_vma_snapshot(&cprm);
821         }
822         if (ispipe && core_pipe_limit)
823                 wait_for_dump_helpers(cprm.file);
824 close_fail:
825         if (cprm.file)
826                 filp_close(cprm.file, NULL);
827 fail_dropcount:
828         if (ispipe)
829                 atomic_dec(&core_dump_count);
830 fail_unlock:
831         kfree(argv);
832         kfree(cn.corename);
833         coredump_finish(mm, core_dumped);
834         revert_creds(old_cred);
835 fail_creds:
836         put_cred(cred);
837 fail:
838         return;
839 }
840
841 /*
842  * Core dumping helper functions.  These are the only things you should
843  * do on a core-file: use only these functions to write out all the
844  * necessary info.
845  */
846 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
847 {
848         struct file *file = cprm->file;
849         loff_t pos = file->f_pos;
850         ssize_t n;
851         if (cprm->written + nr > cprm->limit)
852                 return 0;
853
854
855         if (dump_interrupted())
856                 return 0;
857         n = __kernel_write(file, addr, nr, &pos);
858         if (n != nr)
859                 return 0;
860         file->f_pos = pos;
861         cprm->written += n;
862         cprm->pos += n;
863
864         return 1;
865 }
866 EXPORT_SYMBOL(dump_emit);
867
868 int dump_skip(struct coredump_params *cprm, size_t nr)
869 {
870         static char zeroes[PAGE_SIZE];
871         struct file *file = cprm->file;
872         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
873                 if (dump_interrupted() ||
874                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
875                         return 0;
876                 cprm->pos += nr;
877                 return 1;
878         } else {
879                 while (nr > PAGE_SIZE) {
880                         if (!dump_emit(cprm, zeroes, PAGE_SIZE))
881                                 return 0;
882                         nr -= PAGE_SIZE;
883                 }
884                 return dump_emit(cprm, zeroes, nr);
885         }
886 }
887 EXPORT_SYMBOL(dump_skip);
888
889 #ifdef CONFIG_ELF_CORE
890 int dump_user_range(struct coredump_params *cprm, unsigned long start,
891                     unsigned long len)
892 {
893         unsigned long addr;
894
895         for (addr = start; addr < start + len; addr += PAGE_SIZE) {
896                 struct page *page;
897                 int stop;
898
899                 /*
900                  * To avoid having to allocate page tables for virtual address
901                  * ranges that have never been used yet, and also to make it
902                  * easy to generate sparse core files, use a helper that returns
903                  * NULL when encountering an empty page table entry that would
904                  * otherwise have been filled with the zero page.
905                  */
906                 page = get_dump_page(addr);
907                 if (page) {
908                         void *kaddr = kmap(page);
909
910                         stop = !dump_emit(cprm, kaddr, PAGE_SIZE);
911                         kunmap(page);
912                         put_page(page);
913                 } else {
914                         stop = !dump_skip(cprm, PAGE_SIZE);
915                 }
916                 if (stop)
917                         return 0;
918         }
919         return 1;
920 }
921 #endif
922
923 int dump_align(struct coredump_params *cprm, int align)
924 {
925         unsigned mod = cprm->pos & (align - 1);
926         if (align & (align - 1))
927                 return 0;
928         return mod ? dump_skip(cprm, align - mod) : 1;
929 }
930 EXPORT_SYMBOL(dump_align);
931
932 /*
933  * Ensures that file size is big enough to contain the current file
934  * postion. This prevents gdb from complaining about a truncated file
935  * if the last "write" to the file was dump_skip.
936  */
937 void dump_truncate(struct coredump_params *cprm)
938 {
939         struct file *file = cprm->file;
940         loff_t offset;
941
942         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
943                 offset = file->f_op->llseek(file, 0, SEEK_CUR);
944                 if (i_size_read(file->f_mapping->host) < offset)
945                         do_truncate(file->f_path.dentry, offset, 0, file);
946         }
947 }
948 EXPORT_SYMBOL(dump_truncate);
949
950 /*
951  * The purpose of always_dump_vma() is to make sure that special kernel mappings
952  * that are useful for post-mortem analysis are included in every core dump.
953  * In that way we ensure that the core dump is fully interpretable later
954  * without matching up the same kernel and hardware config to see what PC values
955  * meant. These special mappings include - vDSO, vsyscall, and other
956  * architecture specific mappings
957  */
958 static bool always_dump_vma(struct vm_area_struct *vma)
959 {
960         /* Any vsyscall mappings? */
961         if (vma == get_gate_vma(vma->vm_mm))
962                 return true;
963
964         /*
965          * Assume that all vmas with a .name op should always be dumped.
966          * If this changes, a new vm_ops field can easily be added.
967          */
968         if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
969                 return true;
970
971         /*
972          * arch_vma_name() returns non-NULL for special architecture mappings,
973          * such as vDSO sections.
974          */
975         if (arch_vma_name(vma))
976                 return true;
977
978         return false;
979 }
980
981 #define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
982
983 /*
984  * Decide how much of @vma's contents should be included in a core dump.
985  */
986 static unsigned long vma_dump_size(struct vm_area_struct *vma,
987                                    unsigned long mm_flags)
988 {
989 #define FILTER(type)    (mm_flags & (1UL << MMF_DUMP_##type))
990
991         /* always dump the vdso and vsyscall sections */
992         if (always_dump_vma(vma))
993                 goto whole;
994
995         if (vma->vm_flags & VM_DONTDUMP)
996                 return 0;
997
998         /* support for DAX */
999         if (vma_is_dax(vma)) {
1000                 if ((vma->vm_flags & VM_SHARED) && FILTER(DAX_SHARED))
1001                         goto whole;
1002                 if (!(vma->vm_flags & VM_SHARED) && FILTER(DAX_PRIVATE))
1003                         goto whole;
1004                 return 0;
1005         }
1006
1007         /* Hugetlb memory check */
1008         if (is_vm_hugetlb_page(vma)) {
1009                 if ((vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_SHARED))
1010                         goto whole;
1011                 if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
1012                         goto whole;
1013                 return 0;
1014         }
1015
1016         /* Do not dump I/O mapped devices or special mappings */
1017         if (vma->vm_flags & VM_IO)
1018                 return 0;
1019
1020         /* By default, dump shared memory if mapped from an anonymous file. */
1021         if (vma->vm_flags & VM_SHARED) {
1022                 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1023                     FILTER(ANON_SHARED) : FILTER(MAPPED_SHARED))
1024                         goto whole;
1025                 return 0;
1026         }
1027
1028         /* Dump segments that have been written to.  */
1029         if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) && FILTER(ANON_PRIVATE))
1030                 goto whole;
1031         if (vma->vm_file == NULL)
1032                 return 0;
1033
1034         if (FILTER(MAPPED_PRIVATE))
1035                 goto whole;
1036
1037         /*
1038          * If this is the beginning of an executable file mapping,
1039          * dump the first page to aid in determining what was mapped here.
1040          */
1041         if (FILTER(ELF_HEADERS) &&
1042             vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1043                 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1044                         return PAGE_SIZE;
1045
1046                 /*
1047                  * ELF libraries aren't always executable.
1048                  * We'll want to check whether the mapping starts with the ELF
1049                  * magic, but not now - we're holding the mmap lock,
1050                  * so copy_from_user() doesn't work here.
1051                  * Use a placeholder instead, and fix it up later in
1052                  * dump_vma_snapshot().
1053                  */
1054                 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1055         }
1056
1057 #undef  FILTER
1058
1059         return 0;
1060
1061 whole:
1062         return vma->vm_end - vma->vm_start;
1063 }
1064
1065 static struct vm_area_struct *first_vma(struct task_struct *tsk,
1066                                         struct vm_area_struct *gate_vma)
1067 {
1068         struct vm_area_struct *ret = tsk->mm->mmap;
1069
1070         if (ret)
1071                 return ret;
1072         return gate_vma;
1073 }
1074
1075 /*
1076  * Helper function for iterating across a vma list.  It ensures that the caller
1077  * will visit `gate_vma' prior to terminating the search.
1078  */
1079 static struct vm_area_struct *next_vma(struct vm_area_struct *this_vma,
1080                                        struct vm_area_struct *gate_vma)
1081 {
1082         struct vm_area_struct *ret;
1083
1084         ret = this_vma->vm_next;
1085         if (ret)
1086                 return ret;
1087         if (this_vma == gate_vma)
1088                 return NULL;
1089         return gate_vma;
1090 }
1091
1092 static void free_vma_snapshot(struct coredump_params *cprm)
1093 {
1094         if (cprm->vma_meta) {
1095                 int i;
1096                 for (i = 0; i < cprm->vma_count; i++) {
1097                         struct file *file = cprm->vma_meta[i].file;
1098                         if (file)
1099                                 fput(file);
1100                 }
1101                 kvfree(cprm->vma_meta);
1102                 cprm->vma_meta = NULL;
1103         }
1104 }
1105
1106 /*
1107  * Under the mmap_lock, take a snapshot of relevant information about the task's
1108  * VMAs.
1109  */
1110 static bool dump_vma_snapshot(struct coredump_params *cprm)
1111 {
1112         struct vm_area_struct *vma, *gate_vma;
1113         struct mm_struct *mm = current->mm;
1114         int i;
1115
1116         /*
1117          * Once the stack expansion code is fixed to not change VMA bounds
1118          * under mmap_lock in read mode, this can be changed to take the
1119          * mmap_lock in read mode.
1120          */
1121         if (mmap_write_lock_killable(mm))
1122                 return false;
1123
1124         cprm->vma_data_size = 0;
1125         gate_vma = get_gate_vma(mm);
1126         cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1127
1128         cprm->vma_meta = kvmalloc_array(cprm->vma_count, sizeof(*cprm->vma_meta), GFP_KERNEL);
1129         if (!cprm->vma_meta) {
1130                 mmap_write_unlock(mm);
1131                 return false;
1132         }
1133
1134         for (i = 0, vma = first_vma(current, gate_vma); vma != NULL;
1135                         vma = next_vma(vma, gate_vma), i++) {
1136                 struct core_vma_metadata *m = cprm->vma_meta + i;
1137
1138                 m->start = vma->vm_start;
1139                 m->end = vma->vm_end;
1140                 m->flags = vma->vm_flags;
1141                 m->dump_size = vma_dump_size(vma, cprm->mm_flags);
1142                 m->pgoff = vma->vm_pgoff;
1143
1144                 m->file = vma->vm_file;
1145                 if (m->file)
1146                         get_file(m->file);
1147         }
1148
1149         mmap_write_unlock(mm);
1150
1151         for (i = 0; i < cprm->vma_count; i++) {
1152                 struct core_vma_metadata *m = cprm->vma_meta + i;
1153
1154                 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
1155                         char elfmag[SELFMAG];
1156
1157                         if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
1158                                         memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
1159                                 m->dump_size = 0;
1160                         } else {
1161                                 m->dump_size = PAGE_SIZE;
1162                         }
1163                 }
1164
1165                 cprm->vma_data_size += m->dump_size;
1166         }
1167
1168         return true;
1169 }