projects/linux·blob master · refs

2022 lines · 51.8 KiB

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/sort.h>
22#include <linux/sched/coredump.h>
23#include <linux/sched/signal.h>
24#include <linux/sched/task_stack.h>
25#include <linux/utsname.h>
26#include <linux/pid_namespace.h>
27#include <linux/module.h>
28#include <linux/namei.h>
29#include <linux/mount.h>
30#include <linux/security.h>
31#include <linux/syscalls.h>
32#include <linux/tsacct_kern.h>
33#include <linux/cn_proc.h>
34#include <linux/audit.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/wait_bit.h>
43#include <linux/path.h>
44#include <linux/timekeeping.h>
45#include <linux/sysctl.h>
46#include <linux/elf.h>
47#include <linux/pidfs.h>
48#include <linux/net.h>
49#include <linux/socket.h>
50#include <net/af_unix.h>
51#include <net/net_namespace.h>
52#include <net/sock.h>
53#include <uapi/linux/pidfd.h>
54#include <uapi/linux/un.h>
55
56#include <linux/uaccess.h>
57#include <asm/mmu_context.h>
58#include <asm/tlb.h>
59#include <asm/exec.h>
60
61#include <trace/events/task.h>
62#include "internal.h"
63
64#include <trace/events/sched.h>
65
66#define CREATE_TRACE_POINTS
67#include <trace/events/coredump.h>
68
69static bool dump_vma_snapshot(struct coredump_params *cprm);
70static void free_vma_snapshot(struct coredump_params *cprm);
71static void dump_end_record(struct coredump_params *cprm);
72static bool dump_flush_skip(struct coredump_params *cprm);
73
74#define CORE_FILE_NOTE_SIZE_DEFAULT (4*1024*1024)
75/* Define a reasonable max cap */
76#define CORE_FILE_NOTE_SIZE_MAX (16*1024*1024)
77/*
78 * File descriptor number for the pidfd for the thread-group leader of
79 * the coredumping task installed into the usermode helper's file
80 * descriptor table.
81 */
82#define COREDUMP_PIDFD_NUMBER 3
83
84static int core_uses_pid;
85static unsigned int core_pipe_limit;
86static unsigned int core_sort_vma;
87static char core_pattern[CORENAME_MAX_SIZE] = "core";
88/* Taken around every copy in and out of core_pattern. */
89static DEFINE_SPINLOCK(core_pattern_lock);
90static int core_name_size = CORENAME_MAX_SIZE;
91unsigned int core_file_note_size_limit = CORE_FILE_NOTE_SIZE_DEFAULT;
92static atomic_t core_pipe_count = ATOMIC_INIT(0);
93
94enum coredump_type_t {
95 COREDUMP_FILE = 1,
96 COREDUMP_PIPE = 2,
97 COREDUMP_SOCK = 3,
98 COREDUMP_SOCK_REQ = 4,
99};
100
101struct core_name {
102 char *corename __counted_by_ptr(size);
103 int used, size;
104 unsigned int core_pipe_limit;
105 enum coredump_type_t core_type;
106};
107
108static int expand_corename(struct core_name *cn, int size)
109{
110 char *corename;
111
112 size = kmalloc_size_roundup(size);
113 corename = krealloc(cn->corename, size, GFP_KERNEL);
114 if (!corename)
115 return -ENOMEM;
116
117 cn->corename = corename;
118 cn->size = size;
119
120 if (size > core_name_size) /* racy but harmless */
121 core_name_size = size;
122
123 return 0;
124}
125
126static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
127 va_list arg)
128{
129 int free, need;
130 va_list arg_copy;
131
132again:
133 free = cn->size - cn->used;
134
135 va_copy(arg_copy, arg);
136 need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
137 va_end(arg_copy);
138
139 if (need < free) {
140 cn->used += need;
141 return 0;
142 }
143
144 if (!expand_corename(cn, cn->size + need - free + 1))
145 goto again;
146
147 return -ENOMEM;
148}
149
150static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
151{
152 va_list arg;
153 int ret;
154
155 va_start(arg, fmt);
156 ret = cn_vprintf(cn, fmt, arg);
157 va_end(arg);
158
159 return ret;
160}
161
162static __printf(2, 3)
163int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
164{
165 int cur = cn->used;
166 va_list arg;
167 int ret;
168
169 va_start(arg, fmt);
170 ret = cn_vprintf(cn, fmt, arg);
171 va_end(arg);
172
173 if (ret == 0) {
174 /*
175 * Ensure that this coredump name component can't cause the
176 * resulting corefile path to consist of a ".." or ".".
177 */
178 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
179 (cn->used - cur == 2 && cn->corename[cur] == '.'
180 && cn->corename[cur+1] == '.'))
181 cn->corename[cur] = '!';
182
183 /*
184 * Empty names are fishy and could be used to create a "//" in a
185 * corefile name, causing the coredump to happen one directory
186 * level too high. Enforce that all components of the core
187 * pattern are at least one character long.
188 */
189 if (cn->used == cur)
190 ret = cn_printf(cn, "!");
191 }
192
193 for (; cur < cn->used; ++cur) {
194 if (cn->corename[cur] == '/')
195 cn->corename[cur] = '!';
196 }
197 return ret;
198}
199
200static int cn_print_exe_file(struct core_name *cn, bool name_only)
201{
202 struct file *exe_file;
203 char *pathbuf, *path, *ptr;
204 int ret;
205
206 exe_file = get_mm_exe_file(current->mm);
207 if (!exe_file)
208 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
209
210 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
211 if (!pathbuf) {
212 ret = -ENOMEM;
213 goto put_exe_file;
214 }
215
216 path = file_path(exe_file, pathbuf, PATH_MAX);
217 if (IS_ERR(path)) {
218 ret = PTR_ERR(path);
219 goto free_buf;
220 }
221
222 if (name_only) {
223 ptr = strrchr(path, '/');
224 if (ptr)
225 path = ptr + 1;
226 }
227 ret = cn_esc_printf(cn, "%s", path);
228
229free_buf:
230 kfree(pathbuf);
231put_exe_file:
232 fput(exe_file);
233 return ret;
234}
235
236/*
237 * coredump_parse will inspect the pattern parameter, and output a name
238 * into corename, which must have space for at least CORENAME_MAX_SIZE
239 * bytes plus one byte for the zero terminator.
240 */
241static bool coredump_parse(struct core_name *cn, struct coredump_params *cprm,
242 size_t **argv, int *argc)
243{
244 const struct cred *cred = current_cred();
245 char pattern[CORENAME_MAX_SIZE];
246 const char *pat_ptr = pattern;
247 bool was_space = false;
248 int pid_in_pattern = 0;
249 int err = 0;
250
251 /* The sysctl handler may be publishing a new pattern. */
252 scoped_guard(spinlock, &core_pattern_lock)
253 strscpy(pattern, core_pattern);
254
255 cprm->mask = COREDUMP_KERNEL;
256 if (core_pipe_limit)
257 cprm->mask |= COREDUMP_WAIT;
258 cn->used = 0;
259 cn->corename = NULL;
260 cn->core_pipe_limit = 0;
261 if (*pat_ptr == '|')
262 cn->core_type = COREDUMP_PIPE;
263 else if (*pat_ptr == '@')
264 cn->core_type = COREDUMP_SOCK;
265 else
266 cn->core_type = COREDUMP_FILE;
267 if (expand_corename(cn, core_name_size))
268 return false;
269 cn->corename[0] = '\0';
270
271 switch (cn->core_type) {
272 case COREDUMP_PIPE: {
273 int argvs = sizeof(core_pattern) / 2;
274 (*argv) = kmalloc_objs(**argv, argvs);
275 if (!(*argv))
276 return false;
277 (*argv)[(*argc)++] = 0;
278 ++pat_ptr;
279 if (!(*pat_ptr))
280 return false;
281 break;
282 }
283 case COREDUMP_SOCK: {
284 /* skip the @ */
285 pat_ptr++;
286 if (!(*pat_ptr))
287 return false;
288 if (*pat_ptr == '@') {
289 pat_ptr++;
290 if (!(*pat_ptr))
291 return false;
292
293 cn->core_type = COREDUMP_SOCK_REQ;
294 }
295
296 err = cn_printf(cn, "%s", pat_ptr);
297 if (err)
298 return false;
299
300 /* Require absolute paths. */
301 if (cn->corename[0] != '/')
302 return false;
303
304 /*
305 * Ensure we can uses spaces to indicate additional
306 * parameters in the future.
307 */
308 if (strchr(cn->corename, ' ')) {
309 coredump_report_failure("Coredump socket may not %s contain spaces", cn->corename);
310 return false;
311 }
312
313 /* Must not contain ".." in the path. */
314 if (name_contains_dotdot(cn->corename)) {
315 coredump_report_failure("Coredump socket may not %s contain '..' spaces", cn->corename);
316 return false;
317 }
318
319 if (strlen(cn->corename) >= UNIX_PATH_MAX) {
320 coredump_report_failure("Coredump socket path %s too long", cn->corename);
321 return false;
322 }
323
324 /*
325 * Currently no need to parse any other options.
326 * Relevant information can be retrieved from the peer
327 * pidfd retrievable via SO_PEERPIDFD by the receiver or
328 * via /proc/<pid>, using the SO_PEERPIDFD to guard
329 * against pid recycling when opening /proc/<pid>.
330 */
331 return true;
332 }
333 case COREDUMP_FILE:
334 break;
335 default:
336 WARN_ON_ONCE(true);
337 return false;
338 }
339
340 /* Repeat as long as we have more pattern to process and more output
341 space */
342 while (*pat_ptr) {
343 /*
344 * Split on spaces before doing template expansion so that
345 * %e and %E don't get split if they have spaces in them
346 */
347 if (cn->core_type == COREDUMP_PIPE) {
348 if (isspace(*pat_ptr)) {
349 if (cn->used != 0)
350 was_space = true;
351 pat_ptr++;
352 continue;
353 } else if (was_space) {
354 was_space = false;
355 err = cn_printf(cn, "%c", '\0');
356 if (err)
357 return false;
358 (*argv)[(*argc)++] = cn->used;
359 }
360 }
361 if (*pat_ptr != '%') {
362 err = cn_printf(cn, "%c", *pat_ptr++);
363 } else {
364 switch (*++pat_ptr) {
365 /* single % at the end, drop that */
366 case 0:
367 goto out;
368 /* Double percent, output one percent */
369 case '%':
370 err = cn_printf(cn, "%c", '%');
371 break;
372 /* pid */
373 case 'p':
374 pid_in_pattern = 1;
375 err = cn_printf(cn, "%d",
376 task_tgid_vnr(current));
377 break;
378 /* global pid */
379 case 'P':
380 err = cn_printf(cn, "%d",
381 task_tgid_nr(current));
382 break;
383 case 'i':
384 err = cn_printf(cn, "%d",
385 task_pid_vnr(current));
386 break;
387 case 'I':
388 err = cn_printf(cn, "%d",
389 task_pid_nr(current));
390 break;
391 /* uid */
392 case 'u':
393 err = cn_printf(cn, "%u",
394 from_kuid(&init_user_ns,
395 cred->uid));
396 break;
397 /* gid */
398 case 'g':
399 err = cn_printf(cn, "%u",
400 from_kgid(&init_user_ns,
401 cred->gid));
402 break;
403 case 'd':
404 err = cn_printf(cn, "%d", cprm->dumpable);
405 break;
406 /* signal that caused the coredump */
407 case 's':
408 err = cn_printf(cn, "%d",
409 cprm->siginfo->si_signo);
410 break;
411 /* UNIX time of coredump */
412 case 't': {
413 time64_t time;
414
415 time = ktime_get_real_seconds();
416 err = cn_printf(cn, "%lld", time);
417 break;
418 }
419 /* hostname */
420 case 'h':
421 down_read(&uts_sem);
422 err = cn_esc_printf(cn, "%s",
423 utsname()->nodename);
424 up_read(&uts_sem);
425 break;
426 /* executable, could be changed by prctl PR_SET_NAME etc */
427 case 'e':
428 err = cn_esc_printf(cn, "%s", current->comm);
429 break;
430 /* file name of executable */
431 case 'f':
432 err = cn_print_exe_file(cn, true);
433 break;
434 case 'E':
435 err = cn_print_exe_file(cn, false);
436 break;
437 /* core limit size */
438 case 'c':
439 err = cn_printf(cn, "%lu",
440 rlimit(RLIMIT_CORE));
441 break;
442 /* CPU the task ran on */
443 case 'C':
444 err = cn_printf(cn, "%d", cprm->cpu);
445 break;
446 /* pidfd number */
447 case 'F': {
448 /*
449 * Installing a pidfd only makes sense if
450 * we actually spawn a usermode helper.
451 */
452 if (cn->core_type != COREDUMP_PIPE)
453 break;
454
455 /*
456 * Note that we'll install a pidfd for the
457 * thread-group leader. We know that task
458 * linkage hasn't been removed yet and even if
459 * this @current isn't the actual thread-group
460 * leader we know that the thread-group leader
461 * cannot be reaped until @current has exited.
462 */
463 task_pids(cprm->pid, current);
464 err = cn_printf(cn, "%d", COREDUMP_PIDFD_NUMBER);
465 break;
466 }
467 default:
468 break;
469 }
470 ++pat_ptr;
471 }
472
473 if (err)
474 return false;
475 }
476
477out:
478 /* Backward compatibility with core_uses_pid:
479 *
480 * If core_pattern does not include a %p (as is the default)
481 * and core_uses_pid is set, then .%pid will be appended to
482 * the filename. Do not do this for piped commands. */
483 if (cn->core_type == COREDUMP_FILE && !pid_in_pattern && core_uses_pid)
484 return cn_printf(cn, ".%d", task_tgid_vnr(current)) == 0;
485
486 return true;
487}
488
489static int zap_process(struct signal_struct *signal, int exit_code)
490{
491 struct task_struct *t;
492 int nr = 0;
493
494 signal->flags = SIGNAL_GROUP_EXIT;
495 signal->group_exit_code = exit_code;
496 signal->group_stop_count = 0;
497
498 __for_each_thread(signal, t) {
499 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
500 if (t != current && !(t->flags & PF_POSTCOREDUMP)) {
501 sigaddset(&t->pending.signal, SIGKILL);
502 signal_wake_up(t, 1);
503 nr++;
504 }
505 }
506
507 return nr;
508}
509
510static int zap_threads(struct task_struct *tsk,
511 struct core_state *core_state, int exit_code)
512{
513 struct signal_struct *signal = tsk->signal;
514 int nr = -EAGAIN;
515
516 spin_lock_irq(&tsk->sighand->siglock);
517 /* A freeze requested before the dump would be lost with TIF_SIGPENDING. */
518 if (!(signal->flags & SIGNAL_GROUP_EXIT) && !signal->group_exec_task &&
519 !freezing(tsk) && !(tsk->jobctl & JOBCTL_TRAP_FREEZE)) {
520 /* Allow SIGKILL, see prepare_signal() */
521 signal->core_state = core_state;
522 nr = zap_process(signal, exit_code);
523 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
524 tsk->flags |= PF_DUMPCORE;
525 atomic_set(&core_state->threads_remaining, nr);
526 }
527 spin_unlock_irq(&tsk->sighand->siglock);
528 return nr;
529}
530
531static void coredump_wait_inactive(struct core_state *core_state)
532{
533 struct core_thread *ptr;
534
535 wait_var_event_state(&core_state->threads_remaining,
536 !atomic_read_acquire(&core_state->threads_remaining),
537 TASK_UNINTERRUPTIBLE | TASK_FREEZABLE);
538 /*
539 * Wait for all the threads to become inactive, so that
540 * all the thread context (extended register state, like
541 * fpu etc) gets copied to the memory.
542 */
543 for (ptr = core_state->tasks; ptr; ptr = ptr->next)
544 wait_task_inactive(ptr->task, TASK_ANY);
545}
546
547static int coredump_wait(int exit_code, struct core_state *core_state)
548{
549 struct task_struct *tsk = current;
550 int core_waiters = -EBUSY;
551
552 core_state->tasks = NULL;
553
554 core_waiters = zap_threads(tsk, core_state, exit_code);
555 if (core_waiters > 0)
556 coredump_wait_inactive(core_state);
557
558 return core_waiters;
559}
560
561static void coredump_finish(enum coredump_state state)
562{
563 struct core_thread *curr, *next;
564 struct task_struct *task;
565
566 spin_lock_irq(&current->sighand->siglock);
567 if ((state & COREDUMP_STATE_STARTED) && !__fatal_signal_pending(current))
568 current->signal->group_exit_code |= 0x80;
569 next = current->signal->core_state->tasks;
570 current->signal->core_state = NULL;
571 spin_unlock_irq(&current->sighand->siglock);
572
573 /* A released thread may exit and be freed before it is woken. */
574 guard(rcu)();
575 while ((curr = next) != NULL) {
576 next = curr->next;
577 task = curr->task;
578 /*
579 * see coredump_task_exit(), curr->task must not see
580 * ->task == NULL before we read ->next.
581 */
582 smp_mb();
583 /* Any wakeup now lets the thread exit, rcu keeps it alive. */
584 curr->task = NULL;
585 wake_up_process(task);
586 }
587}
588
589static bool dump_interrupted(void)
590{
591 /* Only SIGKILL and the freezers set it after zap_threads(). */
592 return task_sigpending(current);
593}
594
595static void wait_for_dump_helpers(struct file *file)
596{
597 struct pipe_inode_info *pipe = file->private_data;
598
599 pipe_lock(pipe);
600 pipe->readers++;
601 pipe->writers--;
602 wake_up_interruptible_sync(&pipe->rd_wait);
603 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
604 pipe_unlock(pipe);
605
606 /*
607 * We actually want wait_event_freezable() but then we need
608 * to clear TIF_SIGPENDING and improve dump_interrupted().
609 */
610 wait_event_interruptible(pipe->rd_wait, pipe->readers == 1);
611
612 pipe_lock(pipe);
613 pipe->readers--;
614 pipe->writers++;
615 pipe_unlock(pipe);
616}
617
618/*
619 * umh_coredump_setup
620 * helper function to customize the process used
621 * to collect the core in userspace. Specifically
622 * it sets up a pipe and installs it as fd 0 (stdin)
623 * for the process. Returns 0 on success, or
624 * PTR_ERR on failure.
625 * Note that it also sets the core limit to 1. This
626 * is a special value that we use to trap recursive
627 * core dumps
628 */
629static int umh_coredump_setup(struct subprocess_info *info, struct cred *new)
630{
631 struct file *files[2];
632 struct coredump_params *cp = (struct coredump_params *)info->data;
633 int err;
634
635 if (cp->pid[PIDTYPE_TGID]) {
636 struct file *pidfs_file __free(fput) = NULL;
637
638 pidfs_file = pidfs_alloc_file(cp->pid[PIDTYPE_TGID], 0);
639 if (IS_ERR(pidfs_file))
640 return PTR_ERR(pidfs_file);
641
642 err = pidfs_register_pids(cp->pid);
643 if (err)
644 return err;
645
646 pidfs_coredump(cp);
647
648 /*
649 * Usermode helpers are childen of either
650 * system_dfl_wq or of kthreadd. So we know that
651 * we're starting off with a clean file descriptor
652 * table. So we should always be able to use
653 * COREDUMP_PIDFD_NUMBER as our file descriptor value.
654 */
655 err = replace_fd(COREDUMP_PIDFD_NUMBER, pidfs_file, 0);
656 if (err < 0)
657 return err;
658 }
659
660 err = create_pipe_files(files, 0);
661 if (err)
662 return err;
663
664 cp->file = files[1];
665
666 err = replace_fd(0, files[0], 0);
667 fput(files[0]);
668 if (err < 0)
669 return err;
670
671 /* and disallow core files too */
672 current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
673
674 return 0;
675}
676
677static_assert(sizeof(struct coredump_record_header) == COREDUMP_RECORD_HEADER_SIZE_VER0);
678
679#ifdef CONFIG_UNIX
680/* af_unix halves the send buffer to size a single skb. */
681#define COREDUMP_SOCK_SNDBUF_MIN (3 * PAGE_SIZE)
682
683static bool coredump_sock_connect(struct core_name *cn, struct coredump_params *cprm)
684{
685 struct file *file __free(fput) = NULL;
686 struct sockaddr_un addr = {
687 .sun_family = AF_UNIX,
688 };
689 ssize_t addr_len;
690 int retval;
691 struct socket *socket;
692
693 addr_len = strscpy(addr.sun_path, cn->corename);
694 if (addr_len < 0)
695 return false;
696 addr_len += offsetof(struct sockaddr_un, sun_path) + 1;
697
698 /*
699 * It is possible that the userspace process which is supposed
700 * to handle the coredump and is listening on the AF_UNIX socket
701 * coredumps. Userspace should just mark itself non dumpable.
702 */
703
704 retval = sock_create_kern(&init_net, AF_UNIX, SOCK_STREAM, 0, &socket);
705 if (retval < 0)
706 return false;
707
708 /* Don't let a page-sized write split into several skbs. */
709 socket->sk->sk_sndbuf = max_t(int, socket->sk->sk_sndbuf,
710 COREDUMP_SOCK_SNDBUF_MIN);
711
712 file = sock_alloc_file(socket, 0, NULL);
713 if (IS_ERR(file))
714 return false;
715
716 /*
717 * Set the pids of the dumping thread and its thread-group leader
718 * which are used for the peer credentials during connect() below.
719 * Then immediately register them in pidfs...
720 */
721 task_pids(cprm->pid, current);
722 retval = pidfs_register_pids(cprm->pid);
723 if (retval)
724 return false;
725
726 /*
727 * ... and set the coredump information so userspace has it
728 * available after connect()...
729 */
730 pidfs_coredump(cprm);
731
732 retval = kernel_connect(socket, (struct sockaddr_unsized *)(&addr), addr_len,
733 O_NONBLOCK | SOCK_COREDUMP);
734
735 if (retval) {
736 if (retval == -EAGAIN)
737 coredump_report_failure("Coredump socket %s receive queue full", addr.sun_path);
738 else
739 coredump_report_failure("Coredump socket connection %s failed %d", addr.sun_path, retval);
740 return false;
741 }
742
743 /* ... and validate that @sk_peer_pid matches @cprm.pid. */
744 if (WARN_ON_ONCE(!pids_equal(unix_peer(socket->sk)->sk_peer_pid, cprm->pid)))
745 return false;
746
747 cprm->limit = RLIM_INFINITY;
748 cprm->file = no_free_ptr(file);
749
750 return true;
751}
752
753static inline bool coredump_sock_recv(struct file *file, struct coredump_ack *ack, size_t size, int flags)
754{
755 struct msghdr msg = {};
756 struct kvec iov = { .iov_base = ack, .iov_len = size };
757 ssize_t ret;
758
759 memset(ack, 0, size);
760 ret = kernel_recvmsg(sock_from_file(file), &msg, &iov, 1, size, flags);
761 return ret == size;
762}
763
764static inline bool coredump_sock_send(struct file *file, struct coredump_req *req)
765{
766 struct msghdr msg = { .msg_flags = MSG_NOSIGNAL };
767 struct kvec iov = { .iov_base = req, .iov_len = sizeof(*req) };
768 ssize_t ret;
769
770 ret = kernel_sendmsg(sock_from_file(file), &msg, &iov, 1, sizeof(*req));
771 return ret == sizeof(*req);
772}
773
774static_assert(sizeof(struct coredump_req) == COREDUMP_REQ_SIZE_VER1);
775static_assert(sizeof(struct coredump_ack) == COREDUMP_ACK_SIZE_VER1);
776static_assert(sizeof(enum coredump_mark) == sizeof(__u32));
777
778/* Every memory type this kernel knows. */
779#define COREDUMP_MEMORY_ALL \
780 (COREDUMP_MEMORY_ANON_PRIVATE | COREDUMP_MEMORY_ANON_SHARED | \
781 COREDUMP_MEMORY_FILE_PRIVATE | COREDUMP_MEMORY_FILE_SHARED | \
782 COREDUMP_MEMORY_ELF_HEADERS | \
783 COREDUMP_MEMORY_HUGETLB_PRIVATE | COREDUMP_MEMORY_HUGETLB_SHARED | \
784 COREDUMP_MEMORY_DAX_PRIVATE | COREDUMP_MEMORY_DAX_SHARED)
785
786#define COREDUMP_MEMORY_TYPE_BIT(mmf) BIT((mmf) - MMF_DUMP_FILTER_SHIFT)
787static_assert(COREDUMP_MEMORY_ALL == (MMF_DUMP_FILTER_MASK >> MMF_DUMP_FILTER_SHIFT));
788static_assert(COREDUMP_MEMORY_ANON_PRIVATE ==
789 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_ANON_PRIVATE));
790static_assert(COREDUMP_MEMORY_ANON_SHARED ==
791 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_ANON_SHARED));
792static_assert(COREDUMP_MEMORY_FILE_PRIVATE ==
793 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_MAPPED_PRIVATE));
794static_assert(COREDUMP_MEMORY_FILE_SHARED ==
795 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_MAPPED_SHARED));
796static_assert(COREDUMP_MEMORY_ELF_HEADERS ==
797 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_ELF_HEADERS));
798static_assert(COREDUMP_MEMORY_HUGETLB_PRIVATE ==
799 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_HUGETLB_PRIVATE));
800static_assert(COREDUMP_MEMORY_HUGETLB_SHARED ==
801 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_HUGETLB_SHARED));
802static_assert(COREDUMP_MEMORY_DAX_PRIVATE ==
803 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_DAX_PRIVATE));
804static_assert(COREDUMP_MEMORY_DAX_SHARED ==
805 COREDUMP_MEMORY_TYPE_BIT(MMF_DUMP_DAX_SHARED));
806
807static inline bool coredump_sock_mark(struct file *file, enum coredump_mark mark)
808{
809 struct msghdr msg = { .msg_flags = MSG_NOSIGNAL };
810 struct kvec iov = { .iov_base = &mark, .iov_len = sizeof(mark) };
811 ssize_t ret;
812
813 ret = kernel_sendmsg(sock_from_file(file), &msg, &iov, 1, sizeof(mark));
814 return ret == sizeof(mark);
815}
816
817static inline void coredump_sock_wait(struct file *file)
818{
819 ssize_t n;
820
821 /*
822 * We use a simple read to wait for the coredump processing to
823 * finish. Either the socket is closed or we get sent unexpected
824 * data. In both cases, we're done.
825 */
826 n = __kernel_read(file, &(char){ 0 }, 1, NULL);
827 if (n > 0)
828 coredump_report_failure("Coredump socket had unexpected data");
829 else if (n < 0)
830 coredump_report_failure("Coredump socket failed");
831}
832
833static inline void coredump_sock_shutdown(struct file *file)
834{
835 struct socket *socket;
836
837 socket = sock_from_file(file);
838 if (!socket)
839 return;
840
841 /* Let userspace know we're done processing the coredump. */
842 kernel_sock_shutdown(socket, SHUT_WR);
843}
844
845static bool coredump_sock_request(struct core_name *cn, struct coredump_params *cprm)
846{
847 struct coredump_req req = {
848 .size = sizeof(struct coredump_req),
849 .mask = COREDUMP_KERNEL | COREDUMP_USERSPACE |
850 COREDUMP_REJECT | COREDUMP_WAIT |
851 COREDUMP_RECORDS | COREDUMP_SPARSE |
852 COREDUMP_MEMORY_TYPES,
853 .size_ack = sizeof(struct coredump_ack),
854 .memory_types = cprm->memory_types,
855 .memory_types_mask = COREDUMP_MEMORY_ALL,
856 };
857 struct coredump_ack ack = {};
858 ssize_t usize;
859
860 if (cn->core_type != COREDUMP_SOCK_REQ)
861 return true;
862
863 /* Let userspace know what we support. */
864 if (!coredump_sock_send(cprm->file, &req))
865 return false;
866
867 /* Peek the size of the coredump_ack. */
868 if (!coredump_sock_recv(cprm->file, &ack, sizeof(ack.size),
869 MSG_PEEK | MSG_WAITALL))
870 return false;
871
872 /* Refuse unknown coredump_ack sizes. */
873 usize = ack.size;
874 if (usize < COREDUMP_ACK_SIZE_VER0) {
875 coredump_sock_mark(cprm->file, COREDUMP_MARK_MINSIZE);
876 return false;
877 }
878
879 if (usize > sizeof(ack)) {
880 coredump_sock_mark(cprm->file, COREDUMP_MARK_MAXSIZE);
881 return false;
882 }
883
884 /* Now retrieve the coredump_ack. */
885 if (!coredump_sock_recv(cprm->file, &ack, usize, MSG_WAITALL))
886 return false;
887 if (ack.size != usize)
888 return false;
889
890 /* Refuse unknown coredump_ack flags. */
891 if (ack.mask & ~req.mask) {
892 coredump_sock_mark(cprm->file, COREDUMP_MARK_UNSUPPORTED);
893 return false;
894 }
895
896 /* Refuse mutually exclusive options. */
897 if (hweight64(ack.mask & (COREDUMP_USERSPACE | COREDUMP_KERNEL |
898 COREDUMP_REJECT)) != 1) {
899 coredump_sock_mark(cprm->file, COREDUMP_MARK_CONFLICTING);
900 return false;
901 }
902
903 if (ack.spare) {
904 coredump_sock_mark(cprm->file, COREDUMP_MARK_UNSUPPORTED);
905 return false;
906 }
907
908 /* Records only describe a coredump the kernel writes. */
909 if ((ack.mask & COREDUMP_RECORDS) && !(ack.mask & COREDUMP_KERNEL)) {
910 coredump_sock_mark(cprm->file, COREDUMP_MARK_CONFLICTING);
911 return false;
912 }
913
914 /* Zero records only exist inside a record stream. */
915 if ((ack.mask & COREDUMP_SPARSE) && !(ack.mask & COREDUMP_RECORDS)) {
916 coredump_sock_mark(cprm->file, COREDUMP_MARK_CONFLICTING);
917 return false;
918 }
919
920 if (ack.mask & COREDUMP_MEMORY_TYPES) {
921 /* The memory types need the whole field. */
922 if (usize < COREDUMP_ACK_SIZE_VER1) {
923 coredump_sock_mark(cprm->file, COREDUMP_MARK_MINSIZE);
924 return false;
925 }
926
927 /* The memory types only select what the kernel writes. */
928 if (!(ack.mask & COREDUMP_KERNEL)) {
929 coredump_sock_mark(cprm->file, COREDUMP_MARK_CONFLICTING);
930 return false;
931 }
932
933 /* Refuse unknown memory types. */
934 if (ack.memory_types & ~req.memory_types_mask) {
935 coredump_sock_mark(cprm->file, COREDUMP_MARK_UNSUPPORTED);
936 return false;
937 }
938 } else if (ack.memory_types) {
939 /* Like @spare the field must be zero when it isn't used. */
940 coredump_sock_mark(cprm->file, COREDUMP_MARK_UNSUPPORTED);
941 return false;
942 }
943
944 /* Record header scratch; a bvec can't point at the stack. */
945 if (ack.mask & COREDUMP_RECORDS) {
946 cprm->record_hdr = kmalloc_obj(*cprm->record_hdr);
947 if (!cprm->record_hdr)
948 return false;
949 }
950
951 /* The server's selection replaces the task's entirely. */
952 if (ack.mask & COREDUMP_MEMORY_TYPES)
953 cprm->memory_types = ack.memory_types;
954
955 cprm->mask = ack.mask;
956 return coredump_sock_mark(cprm->file, COREDUMP_MARK_REQACK);
957}
958
959static bool coredump_socket(struct core_name *cn, struct coredump_params *cprm)
960{
961 if (!coredump_sock_connect(cn, cprm))
962 return false;
963
964 return coredump_sock_request(cn, cprm);
965}
966#else
967static inline void coredump_sock_wait(struct file *file) { }
968static inline void coredump_sock_shutdown(struct file *file) { }
969static inline bool coredump_socket(struct core_name *cn, struct coredump_params *cprm) { return false; }
970#endif
971
972/* cprm->dumpable is the snapshot of task dumpability at dump start. */
973static inline bool coredump_force_suid_safe(const struct coredump_params *cprm)
974{
975 /* Require nonrelative corefile path and be extra careful. */
976 return cprm->dumpable == TASK_DUMPABLE_ROOT;
977}
978
979static bool coredump_file(struct core_name *cn, struct coredump_params *cprm,
980 const struct linux_binfmt *binfmt)
981{
982 const struct mnt_idmap *idmap;
983 struct inode *inode;
984 struct file *file __free(fput) = NULL;
985 int open_flags = O_CREAT | O_WRONLY | O_NOFOLLOW | O_LARGEFILE | O_EXCL;
986
987 if (cprm->limit < binfmt->min_coredump)
988 return false;
989
990 if (coredump_force_suid_safe(cprm) && cn->corename[0] != '/') {
991 coredump_report_failure("this process can only dump core to a fully qualified path, skipping core dump");
992 return false;
993 }
994
995 /*
996 * Unlink the file if it exists unless this is a SUID
997 * binary - in that case, we're running around with root
998 * privs and don't want to unlink another user's coredump.
999 */
1000 if (!coredump_force_suid_safe(cprm)) {
1001 CLASS(filename_kernel, name)(cn->corename);
1002 /*
1003 * If it doesn't exist, that's fine. If there's some
1004 * other problem, we'll catch it at the filp_open().
1005 */
1006 filename_unlinkat(AT_FDCWD, name);
1007 }
1008
1009 /*
1010 * There is a race between unlinking and creating the
1011 * file, but if that causes an EEXIST here, that's
1012 * fine - another process raced with us while creating
1013 * the corefile, and the other process won. To userspace,
1014 * what matters is that at least one of the two processes
1015 * writes its coredump successfully, not which one.
1016 */
1017 if (coredump_force_suid_safe(cprm)) {
1018 /*
1019 * Using user namespaces, normal user tasks can change
1020 * their current->fs->root to point to arbitrary
1021 * directories. Since the intention of the "only dump
1022 * with a fully qualified path" rule is to control where
1023 * coredumps may be placed using root privileges,
1024 * current->fs->root must not be used. Instead, use the
1025 * root directory of PID 1.
1026 */
1027 scoped_with_init_fs()
1028 file = filp_open(cn->corename, open_flags, 0600);
1029 } else {
1030 file = filp_open(cn->corename, open_flags, 0600);
1031 }
1032 if (IS_ERR(file))
1033 return false;
1034
1035 inode = file_inode(file);
1036 if (inode->i_nlink > 1)
1037 return false;
1038 if (d_unhashed(file->f_path.dentry))
1039 return false;
1040 /*
1041 * AK: actually i see no reason to not allow this for named
1042 * pipes etc, but keep the previous behaviour for now.
1043 */
1044 if (!S_ISREG(inode->i_mode))
1045 return false;
1046 /*
1047 * Don't dump core if the filesystem changed owner or mode
1048 * of the file during file creation. This is an issue when
1049 * a process dumps core while its cwd is e.g. on a vfat
1050 * filesystem.
1051 */
1052 idmap = file_mnt_idmap(file);
1053 if (!vfsuid_eq_kuid(i_uid_into_vfsuid(idmap, inode), current_fsuid())) {
1054 coredump_report_failure("Core dump to %s aborted: cannot preserve file owner", cn->corename);
1055 return false;
1056 }
1057 if ((inode->i_mode & 0677) != 0600) {
1058 coredump_report_failure("Core dump to %s aborted: cannot preserve file permissions", cn->corename);
1059 return false;
1060 }
1061 if (!(file->f_mode & FMODE_CAN_WRITE))
1062 return false;
1063 if (do_truncate(idmap, file->f_path.dentry, 0, 0, file))
1064 return false;
1065
1066 cprm->file = no_free_ptr(file);
1067 return true;
1068}
1069
1070static bool coredump_pipe(struct core_name *cn, struct coredump_params *cprm,
1071 size_t *argv, int argc)
1072{
1073 int argi;
1074 char **helper_argv __free(kfree) = NULL;
1075 struct subprocess_info *sub_info;
1076
1077 if (cprm->limit == 1) {
1078 /* See umh_coredump_setup() which sets RLIMIT_CORE = 1.
1079 *
1080 * Normally core limits are irrelevant to pipes, since
1081 * we're not writing to the file system, but we use
1082 * cprm.limit of 1 here as a special value, this is a
1083 * consistent way to catch recursive crashes.
1084 * We can still crash if the core_pattern binary sets
1085 * RLIM_CORE = !1, but it runs as root, and can do
1086 * lots of stupid things.
1087 *
1088 * Note that we use task_tgid_vnr here to grab the pid
1089 * of the process group leader. That way we get the
1090 * right pid if a thread in a multi-threaded
1091 * core_pattern process dies.
1092 */
1093 coredump_report_failure("RLIMIT_CORE is set to 1, aborting core");
1094 return false;
1095 }
1096 cprm->limit = RLIM_INFINITY;
1097
1098 cn->core_pipe_limit = atomic_inc_return(&core_pipe_count);
1099 if (core_pipe_limit && (core_pipe_limit < cn->core_pipe_limit)) {
1100 coredump_report_failure("over core_pipe_limit, skipping core dump");
1101 return false;
1102 }
1103
1104 helper_argv = kmalloc_objs(*helper_argv, argc + 1);
1105 if (!helper_argv) {
1106 coredump_report_failure("%s failed to allocate memory", __func__);
1107 return false;
1108 }
1109 for (argi = 0; argi < argc; argi++)
1110 helper_argv[argi] = cn->corename + argv[argi];
1111 helper_argv[argi] = NULL;
1112
1113 sub_info = call_usermodehelper_setup(helper_argv[0], helper_argv, NULL,
1114 GFP_KERNEL, umh_coredump_setup,
1115 NULL, cprm);
1116 if (!sub_info)
1117 return false;
1118
1119 if (call_usermodehelper_exec(sub_info, UMH_WAIT_EXEC)) {
1120 coredump_report_failure("|%s pipe failed", cn->corename);
1121 return false;
1122 }
1123
1124 /*
1125 * umh disabled with CONFIG_STATIC_USERMODEHELPER_PATH="" would
1126 * have this set to NULL.
1127 */
1128 if (!cprm->file) {
1129 coredump_report_failure("Core dump to |%s disabled", cn->corename);
1130 return false;
1131 }
1132
1133 return true;
1134}
1135
1136static bool coredump_write(struct coredump_params *cprm,
1137 const struct linux_binfmt *binfmt)
1138{
1139 if (dump_interrupted()) {
1140 cprm->state |= COREDUMP_STATE_TRUNCATED;
1141 return true;
1142 }
1143
1144 if (!dump_vma_snapshot(cprm)) {
1145 cprm->state |= COREDUMP_STATE_TRUNCATED;
1146 return false;
1147 }
1148
1149 file_start_write(cprm->file);
1150 if (!binfmt->core_dump(cprm))
1151 cprm->state |= COREDUMP_STATE_TRUNCATED;
1152 /*
1153 * A trailing hole still has to land in the coredump. Seeking over
1154 * it doesn't grow the file, so the last byte of it is written
1155 * instead and gdb doesn't see a truncated file. Everything else
1156 * puts the hole on the wire as it flushes it.
1157 */
1158 if (cprm->to_skip) {
1159 bool flushed;
1160
1161 if (cprm->file->f_mode & FMODE_LSEEK) {
1162 cprm->to_skip--;
1163 flushed = dump_emit(cprm, "", 1);
1164 } else {
1165 flushed = dump_flush_skip(cprm);
1166 }
1167 if (!flushed)
1168 cprm->state |= COREDUMP_STATE_TRUNCATED;
1169 }
1170 dump_end_record(cprm);
1171 file_end_write(cprm->file);
1172 free_vma_snapshot(cprm);
1173 return true;
1174}
1175
1176static void coredump_cleanup(struct core_name *cn, struct coredump_params *cprm)
1177{
1178 if (cprm->file)
1179 filp_close(cprm->file, NULL);
1180 if (cn->core_pipe_limit) {
1181 VFS_WARN_ON_ONCE(cn->core_type != COREDUMP_PIPE);
1182 atomic_dec(&core_pipe_count);
1183 }
1184 kfree(cn->corename);
1185 kfree(cprm->record_hdr);
1186 coredump_finish(cprm->state);
1187}
1188
1189static inline bool coredump_skip(const struct coredump_params *cprm,
1190 const struct linux_binfmt *binfmt)
1191{
1192 if (!binfmt)
1193 return true;
1194 if (!binfmt->core_dump)
1195 return true;
1196 if (cprm->dumpable == TASK_DUMPABLE_OFF)
1197 return true;
1198 return false;
1199}
1200
1201static void do_coredump(struct core_name *cn, struct coredump_params *cprm,
1202 size_t **argv, int *argc, const struct linux_binfmt *binfmt)
1203{
1204 trace_coredump(cprm->siginfo->si_signo);
1205
1206 if (!coredump_parse(cn, cprm, argv, argc)) {
1207 coredump_report_failure("format_corename failed, aborting core");
1208 return;
1209 }
1210
1211 switch (cn->core_type) {
1212 case COREDUMP_FILE:
1213 if (!coredump_file(cn, cprm, binfmt))
1214 return;
1215 break;
1216 case COREDUMP_PIPE:
1217 if (!coredump_pipe(cn, cprm, *argv, *argc))
1218 return;
1219 break;
1220 case COREDUMP_SOCK_REQ:
1221 fallthrough;
1222 case COREDUMP_SOCK:
1223 if (!coredump_socket(cn, cprm))
1224 return;
1225 break;
1226 default:
1227 WARN_ON_ONCE(true);
1228 return;
1229 }
1230
1231 /* Don't even generate the coredump. */
1232 if (cprm->mask & COREDUMP_REJECT)
1233 return;
1234
1235 if ((cprm->mask & COREDUMP_KERNEL) && !coredump_write(cprm, binfmt))
1236 return;
1237
1238 coredump_sock_shutdown(cprm->file);
1239
1240 /* Let the parent know that a coredump was generated. */
1241 if (cprm->mask & COREDUMP_USERSPACE)
1242 cprm->state |= COREDUMP_STATE_STARTED;
1243
1244 /*
1245 * When core_pipe_limit is set we wait for the coredump server
1246 * or usermodehelper to finish before exiting so it can e.g.,
1247 * inspect /proc/<pid>.
1248 */
1249 if (cprm->mask & COREDUMP_WAIT) {
1250 switch (cn->core_type) {
1251 case COREDUMP_PIPE:
1252 wait_for_dump_helpers(cprm->file);
1253 break;
1254 case COREDUMP_SOCK_REQ:
1255 fallthrough;
1256 case COREDUMP_SOCK:
1257 coredump_sock_wait(cprm->file);
1258 break;
1259 default:
1260 break;
1261 }
1262 }
1263}
1264
1265#define COREDUMP_TASK_MEMORY_TYPES(mm) \
1266 ((__mm_flags_get_word((mm)) & MMF_DUMP_FILTER_MASK) >> \
1267 MMF_DUMP_FILTER_SHIFT)
1268
1269void vfs_coredump(const kernel_siginfo_t *siginfo)
1270{
1271 size_t *argv __free(kfree) = NULL;
1272 struct core_state core_state;
1273 struct core_name cn;
1274 const struct mm_struct *mm = current->mm;
1275 const struct linux_binfmt *binfmt = mm->binfmt;
1276 int argc = 0;
1277 struct coredump_params cprm = {
1278 .siginfo = siginfo,
1279 .limit = rlimit(RLIMIT_CORE),
1280 /* Snapshot the memory types (unlocked) and dumpable for the dump. */
1281 .memory_types = COREDUMP_TASK_MEMORY_TYPES(mm),
1282 .dumpable = task_exec_state_get_dumpable(current),
1283 .vma_meta = NULL,
1284 .cpu = raw_smp_processor_id(),
1285 };
1286
1287 audit_core_dumps(siginfo->si_signo);
1288
1289 if (coredump_skip(&cprm, binfmt))
1290 return;
1291
1292 CLASS(prepare_creds, cred)();
1293 if (!cred)
1294 return;
1295 /*
1296 * We cannot trust fsuid as being the "true" uid of the process
1297 * nor do we know its entire history. We only know it was tainted
1298 * so we dump it as root in mode 2, and only into a controlled
1299 * environment (pipe handler or fully qualified path).
1300 */
1301 if (coredump_force_suid_safe(&cprm))
1302 cred->fsuid = GLOBAL_ROOT_UID;
1303
1304 if (coredump_wait(siginfo->si_signo, &core_state) < 0)
1305 return;
1306
1307 /* Task work must not cut the dump short, see signal_pending(). */
1308 guard(no_notify_signal)();
1309 scoped_with_creds(cred)
1310 do_coredump(&cn, &cprm, &argv, &argc, binfmt);
1311 coredump_cleanup(&cn, &cprm);
1312 return;
1313}
1314
1315/*
1316 * Core dumping helper functions. These are the only things you should
1317 * do on a core-file: use only these functions to write out all the
1318 * necessary info.
1319 */
1320static bool dump_records(const struct coredump_params *cprm)
1321{
1322 return cprm->mask & COREDUMP_RECORDS;
1323}
1324
1325static bool dump_sparse(const struct coredump_params *cprm)
1326{
1327 return cprm->mask & COREDUMP_SPARSE;
1328}
1329
1330/* Describe the next @len bytes of the coredump. Returns the header size. */
1331static size_t dump_record_init(struct coredump_params *cprm,
1332 enum coredump_record_type type, u64 flags,
1333 u64 len)
1334{
1335 if (!dump_records(cprm))
1336 return 0;
1337
1338 *cprm->record_hdr = (struct coredump_record_header) {
1339 .size = sizeof(*cprm->record_hdr),
1340 .type = type,
1341 .flags = flags,
1342 .offset = cprm->pos,
1343 .len = len,
1344 };
1345
1346 return sizeof(*cprm->record_hdr);
1347}
1348
1349/* Write @iter whole or fail. @len is what it advances the coredump by. */
1350static bool dump_write_iter(struct coredump_params *cprm, struct iov_iter *iter,
1351 size_t len)
1352{
1353 struct file *file = cprm->file;
1354 size_t count = iov_iter_count(iter);
1355 loff_t pos = file->f_pos;
1356 ssize_t n;
1357
1358 n = __kernel_write_iter(file, iter, &pos);
1359 if (n != (ssize_t)count)
1360 return false;
1361 file->f_pos = pos;
1362 cprm->written += count;
1363 cprm->pos += len;
1364
1365 return true;
1366}
1367
1368/* One record, never more than a page. See __dump_emit(). */
1369static bool dump_emit_chunk(struct coredump_params *cprm, const void *addr,
1370 int nr)
1371{
1372 struct kvec kvec[2];
1373 struct iov_iter iter;
1374 unsigned int nseg = 0;
1375 size_t hdrlen;
1376
1377 if (dump_interrupted())
1378 return false;
1379
1380 hdrlen = dump_record_init(cprm, COREDUMP_RECORD_DATA, 0, nr);
1381 if (hdrlen) {
1382 kvec[nseg].iov_base = cprm->record_hdr;
1383 kvec[nseg].iov_len = hdrlen;
1384 nseg++;
1385 }
1386 kvec[nseg].iov_base = (void *)addr;
1387 kvec[nseg].iov_len = nr;
1388 nseg++;
1389
1390 iov_iter_kvec(&iter, ITER_SOURCE, kvec, nseg, hdrlen + nr);
1391
1392 return dump_write_iter(cprm, &iter, nr);
1393}
1394
1395static bool __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
1396{
1397 if (cprm->written + nr > cprm->limit)
1398 return false;
1399
1400 while (nr) {
1401 int chunk = min_t(int, nr, PAGE_SIZE);
1402
1403 if (!dump_emit_chunk(cprm, addr, chunk))
1404 return false;
1405
1406 addr += chunk;
1407 nr -= chunk;
1408 }
1409
1410 return true;
1411}
1412
1413/* Send a record that stands on its own: a header and nothing else. */
1414static bool dump_emit_record(struct coredump_params *cprm,
1415 enum coredump_record_type type, u64 flags, u64 len)
1416{
1417 struct kvec kvec;
1418 struct iov_iter iter;
1419 size_t hdrlen;
1420
1421 hdrlen = dump_record_init(cprm, type, flags, len);
1422 if (!hdrlen)
1423 return false;
1424
1425 kvec.iov_base = cprm->record_hdr;
1426 kvec.iov_len = hdrlen;
1427 iov_iter_kvec(&iter, ITER_SOURCE, &kvec, 1, hdrlen);
1428
1429 return dump_write_iter(cprm, &iter, len);
1430}
1431
1432/* Close the record stream. Only a whole coredump gets an end record. */
1433static void dump_end_record(struct coredump_params *cprm)
1434{
1435 if (cprm->state & COREDUMP_STATE_TRUNCATED)
1436 return;
1437
1438 dump_emit_record(cprm, COREDUMP_RECORD_END, 0, 0);
1439}
1440
1441static bool __dump_skip(struct coredump_params *cprm, size_t nr)
1442{
1443 static char zeroes[PAGE_SIZE];
1444 struct file *file = cprm->file;
1445
1446 if (dump_sparse(cprm)) {
1447 /* Hand the server the length of the hole instead of the hole itself. */
1448 if (dump_interrupted())
1449 return false;
1450 return dump_emit_record(cprm, COREDUMP_RECORD_ZERO, 0, nr);
1451 }
1452
1453 if (file->f_mode & FMODE_LSEEK) {
1454 if (dump_interrupted() || vfs_llseek(file, nr, SEEK_CUR) < 0)
1455 return false;
1456 cprm->pos += nr;
1457 return true;
1458 }
1459
1460 while (nr) {
1461 size_t chunk = min_t(size_t, nr, PAGE_SIZE);
1462
1463 if (!__dump_emit(cprm, zeroes, chunk))
1464 return false;
1465
1466 nr -= chunk;
1467 }
1468
1469 return true;
1470}
1471
1472/* Flush the accumulated hole before writing data. */
1473static bool dump_flush_skip(struct coredump_params *cprm)
1474{
1475 if (cprm->to_skip) {
1476 if (!__dump_skip(cprm, cprm->to_skip))
1477 return false;
1478 cprm->to_skip = 0;
1479 }
1480 return true;
1481}
1482
1483bool dump_emit(struct coredump_params *cprm, const void *addr, int nr)
1484{
1485 if (!dump_flush_skip(cprm))
1486 return false;
1487 return __dump_emit(cprm, addr, nr);
1488}
1489EXPORT_SYMBOL(dump_emit);
1490
1491void dump_skip_to(struct coredump_params *cprm, unsigned long pos)
1492{
1493 if (WARN_ON_ONCE(pos < cprm->pos))
1494 return;
1495 cprm->to_skip = pos - cprm->pos;
1496}
1497EXPORT_SYMBOL(dump_skip_to);
1498
1499void dump_skip(struct coredump_params *cprm, size_t nr)
1500{
1501 cprm->to_skip += nr;
1502}
1503EXPORT_SYMBOL(dump_skip);
1504
1505#ifdef CONFIG_ELF_CORE
1506static bool dump_emit_page(struct coredump_params *cprm, struct page *page)
1507{
1508 struct bio_vec bvec[2];
1509 struct iov_iter iter;
1510 unsigned int nseg = 0;
1511 size_t hdrlen;
1512
1513 if (!page)
1514 return false;
1515
1516 if (!dump_flush_skip(cprm))
1517 return false;
1518 if (cprm->written + PAGE_SIZE > cprm->limit)
1519 return false;
1520 if (dump_interrupted())
1521 return false;
1522
1523 /* Hand the record header to the same write as the page it describes. */
1524 hdrlen = dump_record_init(cprm, COREDUMP_RECORD_DATA, 0, PAGE_SIZE);
1525 if (hdrlen)
1526 bvec_set_virt(&bvec[nseg++], cprm->record_hdr, hdrlen);
1527 bvec_set_page(&bvec[nseg++], page, PAGE_SIZE, 0);
1528
1529 iov_iter_bvec(&iter, ITER_SOURCE, bvec, nseg, hdrlen + PAGE_SIZE);
1530
1531 return dump_write_iter(cprm, &iter, PAGE_SIZE);
1532}
1533
1534/*
1535 * If we might get machine checks from kernel accesses during the
1536 * core dump, let's get those errors early rather than during the
1537 * IO. This is not performance-critical enough to warrant having
1538 * all the machine check logic in the iovec paths.
1539 */
1540#ifdef copy_mc_to_kernel
1541
1542#define dump_page_alloc() alloc_page(GFP_KERNEL)
1543#define dump_page_free(x) __free_page(x)
1544static struct page *dump_page_copy(struct page *src, struct page *dst)
1545{
1546 void *buf = kmap_local_page(src);
1547 size_t left = copy_mc_to_kernel(page_address(dst), buf, PAGE_SIZE);
1548 kunmap_local(buf);
1549 return left ? NULL : dst;
1550}
1551
1552#else
1553
1554/* We just want to return non-NULL; it's never used. */
1555#define dump_page_alloc() ERR_PTR(-EINVAL)
1556#define dump_page_free(x) ((void)(x))
1557static inline struct page *dump_page_copy(struct page *src, struct page *dst)
1558{
1559 return src;
1560}
1561#endif
1562
1563bool dump_user_range(struct coredump_params *cprm, unsigned long start,
1564 unsigned long len)
1565{
1566 unsigned long addr;
1567 struct page *dump_page;
1568 int locked;
1569 bool ret;
1570
1571 dump_page = dump_page_alloc();
1572 if (!dump_page)
1573 return false;
1574
1575 ret = false;
1576 locked = 0;
1577 for (addr = start; addr < start + len; addr += PAGE_SIZE) {
1578 struct page *page;
1579
1580 if (!locked) {
1581 if (mmap_read_lock_killable(current->mm))
1582 goto out;
1583 locked = 1;
1584 }
1585
1586 /*
1587 * To avoid having to allocate page tables for virtual address
1588 * ranges that have never been used yet, and also to make it
1589 * easy to generate sparse core files, use a helper that returns
1590 * NULL when encountering an empty page table entry that would
1591 * otherwise have been filled with the zero page.
1592 */
1593 page = get_dump_page(addr, &locked);
1594 if (page) {
1595 if (locked) {
1596 mmap_read_unlock(current->mm);
1597 locked = 0;
1598 }
1599 bool stop = !dump_emit_page(cprm, dump_page_copy(page, dump_page));
1600 put_page(page);
1601 if (stop)
1602 goto out;
1603 } else {
1604 dump_skip(cprm, PAGE_SIZE);
1605 }
1606
1607 if (dump_interrupted())
1608 goto out;
1609
1610 if (!need_resched())
1611 continue;
1612 if (locked) {
1613 mmap_read_unlock(current->mm);
1614 locked = 0;
1615 }
1616 cond_resched();
1617 }
1618 ret = true;
1619out:
1620 if (locked)
1621 mmap_read_unlock(current->mm);
1622
1623 dump_page_free(dump_page);
1624 return ret;
1625}
1626#endif
1627
1628bool dump_align(struct coredump_params *cprm, int align)
1629{
1630 unsigned mod = (cprm->pos + cprm->to_skip) & (align - 1);
1631 if (align & (align - 1))
1632 return false;
1633 if (mod)
1634 cprm->to_skip += align - mod;
1635 return true;
1636}
1637EXPORT_SYMBOL(dump_align);
1638
1639#ifdef CONFIG_SYSCTL
1640
1641void validate_coredump_safety(void)
1642{
1643 if (suid_dumpable == TASK_DUMPABLE_ROOT &&
1644 core_pattern[0] != '/' && core_pattern[0] != '|' && core_pattern[0] != '@') {
1645
1646 coredump_report_failure("Unsafe core_pattern used with fs.suid_dumpable=2: "
1647 "pipe handler or fully qualified core dump path required. "
1648 "Set kernel.core_pattern before fs.suid_dumpable.");
1649 }
1650}
1651
1652static inline bool check_coredump_socket(const char *pattern)
1653{
1654 const char *p;
1655
1656 if (pattern[0] != '@')
1657 return true;
1658
1659 /*
1660 * Coredump socket must be located in the initial mount
1661 * namespace. Don't give the impression that anything else is
1662 * supported right now.
1663 */
1664 if (current->nsproxy->mnt_ns != init_task.nsproxy->mnt_ns)
1665 return false;
1666
1667 /* Must be an absolute path... */
1668 if (pattern[1] != '/') {
1669 /* ... or the socket request protocol... */
1670 if (pattern[1] != '@')
1671 return false;
1672 /* ... and if so must be an absolute path. */
1673 if (pattern[2] != '/')
1674 return false;
1675 p = &pattern[2];
1676 } else {
1677 p = &pattern[1];
1678 }
1679
1680 /* The path obviously cannot exceed UNIX_PATH_MAX. */
1681 if (strlen(p) >= UNIX_PATH_MAX)
1682 return false;
1683
1684 /* Must not contain ".." in the path. */
1685 if (name_contains_dotdot(pattern))
1686 return false;
1687
1688 return true;
1689}
1690
1691static int proc_dostring_coredump(const struct ctl_table *table, int write,
1692 void *buffer, size_t *lenp, loff_t *ppos)
1693{
1694 char pattern[CORENAME_MAX_SIZE];
1695 const struct ctl_table tmp = {
1696 .procname = table->procname,
1697 .data = pattern,
1698 .maxlen = sizeof(pattern),
1699 };
1700 bool changed = false;
1701 int error;
1702
1703 /* Work on a copy, proc_dostring() appends at *ppos. */
1704 scoped_guard(spinlock, &core_pattern_lock)
1705 strscpy(pattern, core_pattern);
1706
1707 error = proc_dostring(&tmp, write, buffer, lenp, ppos);
1708 if (error || !write)
1709 return error;
1710
1711 if (!check_coredump_socket(pattern))
1712 return -EINVAL;
1713
1714 /* Publish the validated pattern whole. */
1715 scoped_guard(spinlock, &core_pattern_lock) {
1716 changed = strncmp(pattern, core_pattern, CORENAME_MAX_SIZE);
1717 if (changed)
1718 strscpy(core_pattern, pattern);
1719 }
1720 if (changed)
1721 validate_coredump_safety();
1722 return 0;
1723}
1724
1725static const unsigned int core_file_note_size_min = CORE_FILE_NOTE_SIZE_DEFAULT;
1726static const unsigned int core_file_note_size_max = CORE_FILE_NOTE_SIZE_MAX;
1727static char core_modes[] = {
1728 "file\npipe"
1729#ifdef CONFIG_UNIX
1730 "\nsocket"
1731#endif
1732};
1733
1734static const struct ctl_table coredump_sysctls[] = {
1735 {
1736 .procname = "core_uses_pid",
1737 .data = &core_uses_pid,
1738 .maxlen = sizeof(int),
1739 .mode = 0644,
1740 .proc_handler = proc_dointvec,
1741 },
1742 {
1743 .procname = "core_pattern",
1744 .data = core_pattern,
1745 .maxlen = CORENAME_MAX_SIZE,
1746 .mode = 0644,
1747 .proc_handler = proc_dostring_coredump,
1748 },
1749 {
1750 .procname = "core_pipe_limit",
1751 .data = &core_pipe_limit,
1752 .maxlen = sizeof(unsigned int),
1753 .mode = 0644,
1754 .proc_handler = proc_dointvec_minmax,
1755 .extra1 = SYSCTL_ZERO,
1756 .extra2 = SYSCTL_INT_MAX,
1757 },
1758 {
1759 .procname = "core_file_note_size_limit",
1760 .data = &core_file_note_size_limit,
1761 .maxlen = sizeof(unsigned int),
1762 .mode = 0644,
1763 .proc_handler = proc_douintvec_minmax,
1764 .extra1 = (unsigned int *)&core_file_note_size_min,
1765 .extra2 = (unsigned int *)&core_file_note_size_max,
1766 },
1767 {
1768 .procname = "core_sort_vma",
1769 .data = &core_sort_vma,
1770 .maxlen = sizeof(int),
1771 .mode = 0644,
1772 .proc_handler = proc_douintvec_minmax,
1773 .extra1 = SYSCTL_ZERO,
1774 .extra2 = SYSCTL_ONE,
1775 },
1776 {
1777 .procname = "core_modes",
1778 .data = core_modes,
1779 .maxlen = sizeof(core_modes) - 1,
1780 .mode = 0444,
1781 .proc_handler = proc_dostring,
1782 },
1783};
1784
1785static int __init init_fs_coredump_sysctls(void)
1786{
1787 register_sysctl_init("kernel", coredump_sysctls);
1788 return 0;
1789}
1790fs_initcall(init_fs_coredump_sysctls);
1791#endif /* CONFIG_SYSCTL */
1792
1793/*
1794 * The purpose of always_dump_vma() is to make sure that special kernel mappings
1795 * that are useful for post-mortem analysis are included in every core dump.
1796 * In that way we ensure that the core dump is fully interpretable later
1797 * without matching up the same kernel and hardware config to see what PC values
1798 * meant. These special mappings include - vDSO, vsyscall, and other
1799 * architecture specific mappings
1800 */
1801static bool always_dump_vma(struct vm_area_struct *vma)
1802{
1803 /* Any vsyscall mappings? */
1804 if (vma == get_gate_vma(vma->vm_mm))
1805 return true;
1806
1807 /*
1808 * Assume that all vmas with a .name op should always be dumped.
1809 * If this changes, a new vm_ops field can easily be added.
1810 */
1811 if (vma->vm_ops && vma->vm_ops->name && vma->vm_ops->name(vma))
1812 return true;
1813
1814 /*
1815 * arch_vma_name() returns non-NULL for special architecture mappings,
1816 * such as vDSO sections.
1817 */
1818 if (arch_vma_name(vma))
1819 return true;
1820
1821 return false;
1822}
1823
1824#define DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER 1
1825#define COREDUMP_MEMORY_TYPE_INCLUDE(types, type) \
1826 ((types) & COREDUMP_MEMORY_##type)
1827
1828/*
1829 * Decide how much of @vma's contents should be included in a core dump.
1830 */
1831static unsigned long vma_dump_size(struct vm_area_struct *vma,
1832 u64 memory_types)
1833{
1834 /* always dump the vdso and vsyscall sections */
1835 if (always_dump_vma(vma))
1836 goto whole;
1837
1838 if (vma->vm_flags & VM_DONTDUMP)
1839 return 0;
1840
1841 /* support for DAX */
1842 if (vma_is_dax(vma)) {
1843 if ((vma->vm_flags & VM_SHARED) &&
1844 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, DAX_SHARED))
1845 goto whole;
1846 if (!(vma->vm_flags & VM_SHARED) &&
1847 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, DAX_PRIVATE))
1848 goto whole;
1849 return 0;
1850 }
1851
1852 /* Hugetlb memory check */
1853 if (is_vm_hugetlb_page(vma)) {
1854 if ((vma->vm_flags & VM_SHARED) &&
1855 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, HUGETLB_SHARED))
1856 goto whole;
1857 if (!(vma->vm_flags & VM_SHARED) &&
1858 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, HUGETLB_PRIVATE))
1859 goto whole;
1860 return 0;
1861 }
1862
1863 /* Do not dump I/O mapped devices or special mappings */
1864 if (vma->vm_flags & VM_IO)
1865 return 0;
1866
1867 /* By default, dump shared memory if mapped from an anonymous file. */
1868 if (vma->vm_flags & VM_SHARED) {
1869 if (file_inode(vma->vm_file)->i_nlink == 0 ?
1870 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, ANON_SHARED) :
1871 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, FILE_SHARED))
1872 goto whole;
1873 return 0;
1874 }
1875
1876 /* Dump segments that have been written to. */
1877 if ((!IS_ENABLED(CONFIG_MMU) || vma->anon_vma) &&
1878 COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, ANON_PRIVATE))
1879 goto whole;
1880 if (vma->vm_file == NULL)
1881 return 0;
1882
1883 if (COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, FILE_PRIVATE))
1884 goto whole;
1885
1886 /*
1887 * If this is the beginning of an executable file mapping,
1888 * dump the first page to aid in determining what was mapped here.
1889 */
1890 if (COREDUMP_MEMORY_TYPE_INCLUDE(memory_types, ELF_HEADERS) &&
1891 vma->vm_pgoff == 0 && (vma->vm_flags & VM_READ)) {
1892 if ((READ_ONCE(file_inode(vma->vm_file)->i_mode) & 0111) != 0)
1893 return PAGE_SIZE;
1894
1895 /*
1896 * ELF libraries aren't always executable.
1897 * We'll want to check whether the mapping starts with the ELF
1898 * magic, but not now - we're holding the mmap lock,
1899 * so copy_from_user() doesn't work here.
1900 * Use a placeholder instead, and fix it up later in
1901 * dump_vma_snapshot().
1902 */
1903 return DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER;
1904 }
1905
1906 return 0;
1907
1908whole:
1909 return vma->vm_end - vma->vm_start;
1910}
1911
1912/*
1913 * Helper function for iterating across a vma list. It ensures that the caller
1914 * will visit `gate_vma' prior to terminating the search.
1915 */
1916static struct vm_area_struct *coredump_next_vma(struct vma_iterator *vmi,
1917 struct vm_area_struct *vma,
1918 struct vm_area_struct *gate_vma)
1919{
1920 if (gate_vma && (vma == gate_vma))
1921 return NULL;
1922
1923 vma = vma_next(vmi);
1924 if (vma)
1925 return vma;
1926 return gate_vma;
1927}
1928
1929static void free_vma_snapshot(struct coredump_params *cprm)
1930{
1931 if (cprm->vma_meta) {
1932 int i;
1933 for (i = 0; i < cprm->vma_count; i++) {
1934 struct file *file = cprm->vma_meta[i].file;
1935 if (file)
1936 fput(file);
1937 }
1938 kvfree(cprm->vma_meta);
1939 cprm->vma_meta = NULL;
1940 }
1941}
1942
1943static int cmp_vma_size(const void *vma_meta_lhs_ptr, const void *vma_meta_rhs_ptr)
1944{
1945 const struct core_vma_metadata *vma_meta_lhs = vma_meta_lhs_ptr;
1946 const struct core_vma_metadata *vma_meta_rhs = vma_meta_rhs_ptr;
1947
1948 if (vma_meta_lhs->dump_size < vma_meta_rhs->dump_size)
1949 return -1;
1950 if (vma_meta_lhs->dump_size > vma_meta_rhs->dump_size)
1951 return 1;
1952 return 0;
1953}
1954
1955/*
1956 * Under the mmap_lock, take a snapshot of relevant information about the task's
1957 * VMAs.
1958 */
1959static bool dump_vma_snapshot(struct coredump_params *cprm)
1960{
1961 struct vm_area_struct *gate_vma, *vma = NULL;
1962 struct mm_struct *mm = current->mm;
1963 VMA_ITERATOR(vmi, mm, 0);
1964 int i = 0;
1965
1966 /*
1967 * Once the stack expansion code is fixed to not change VMA bounds
1968 * under mmap_lock in read mode, this can be changed to take the
1969 * mmap_lock in read mode.
1970 */
1971 if (mmap_write_lock_killable(mm))
1972 return false;
1973
1974 cprm->vma_data_size = 0;
1975 gate_vma = get_gate_vma(mm);
1976 cprm->vma_count = mm->map_count + (gate_vma ? 1 : 0);
1977
1978 cprm->vma_meta = kvmalloc_objs(*cprm->vma_meta, cprm->vma_count);
1979 if (!cprm->vma_meta) {
1980 mmap_write_unlock(mm);
1981 return false;
1982 }
1983
1984 while ((vma = coredump_next_vma(&vmi, vma, gate_vma)) != NULL) {
1985 struct core_vma_metadata *m = cprm->vma_meta + i;
1986
1987 m->start = vma->vm_start;
1988 m->end = vma->vm_end;
1989 m->flags = vma->vm_flags;
1990 m->dump_size = vma_dump_size(vma, cprm->memory_types);
1991 m->pgoff = vma->vm_pgoff;
1992 m->file = vma->vm_file;
1993 if (m->file)
1994 get_file(m->file);
1995 i++;
1996 }
1997
1998 mmap_write_unlock(mm);
1999
2000 for (i = 0; i < cprm->vma_count; i++) {
2001 struct core_vma_metadata *m = cprm->vma_meta + i;
2002
2003 if (m->dump_size == DUMP_SIZE_MAYBE_ELFHDR_PLACEHOLDER) {
2004 char elfmag[SELFMAG];
2005
2006 if (copy_from_user(elfmag, (void __user *)m->start, SELFMAG) ||
2007 memcmp(elfmag, ELFMAG, SELFMAG) != 0) {
2008 m->dump_size = 0;
2009 } else {
2010 m->dump_size = PAGE_SIZE;
2011 }
2012 }
2013
2014 cprm->vma_data_size += m->dump_size;
2015 }
2016
2017 if (core_sort_vma)
2018 sort(cprm->vma_meta, cprm->vma_count, sizeof(*cprm->vma_meta),
2019 cmp_vma_size, NULL);
2020
2021 return true;
2022}