1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * fs/userfaultfd.c
4 *
5 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
6 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc.
8 *
9 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing).
11 */
12
13 #include <linux/list.h>
14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/mm.h>
17 #include <linux/mm.h>
18 #include <linux/mm_inline.h>
19 #include <linux/mmu_notifier.h>
20 #include <linux/poll.h>
21 #include <linux/slab.h>
22 #include <linux/seq_file.h>
23 #include <linux/file.h>
24 #include <linux/bug.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h>
30 #include <linux/security.h>
31 #include <linux/hugetlb.h>
32 #include <linux/swapops.h>
33 #include <linux/miscdevice.h>
34 #include <linux/uio.h>
35
36 static int sysctl_unprivileged_userfaultfd __read_mostly;
37
38 #ifdef CONFIG_SYSCTL
39 static struct ctl_table vm_userfaultfd_table[] = {
40 {
41 .procname = "unprivileged_userfaultfd",
42 .data = &sysctl_unprivileged_userfaultfd,
43 .maxlen = sizeof(sysctl_unprivileged_userfaultfd),
44 .mode = 0644,
45 .proc_handler = proc_dointvec_minmax,
46 .extra1 = SYSCTL_ZERO,
47 .extra2 = SYSCTL_ONE,
48 },
49 };
50 #endif
51
52 static struct kmem_cache *userfaultfd_ctx_cachep __ro_after_init;
53
54 struct userfaultfd_fork_ctx {
55 struct userfaultfd_ctx *orig;
56 struct userfaultfd_ctx *new;
57 struct list_head list;
58 };
59
60 struct userfaultfd_unmap_ctx {
61 struct userfaultfd_ctx *ctx;
62 unsigned long start;
63 unsigned long end;
64 struct list_head list;
65 };
66
67 struct userfaultfd_wait_queue {
68 struct uffd_msg msg;
69 wait_queue_entry_t wq;
70 struct userfaultfd_ctx *ctx;
71 bool waken;
72 };
73
74 struct userfaultfd_wake_range {
75 unsigned long start;
76 unsigned long len;
77 };
78
79 /* internal indication that UFFD_API ioctl was successfully executed */
80 #define UFFD_FEATURE_INITIALIZED (1u << 31)
81
82 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
83 {
84 return ctx->features & UFFD_FEATURE_INITIALIZED;
85 }
86
87 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx *ctx)
88 {
89 return ctx && (ctx->features & UFFD_FEATURE_WP_ASYNC);
90 }
91
92 /*
93 * Whether WP_UNPOPULATED is enabled on the uffd context. It is only
94 * meaningful when userfaultfd_wp()==true on the vma and when it's
95 * anonymous.
96 */
97 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
98 {
99 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
100
101 if (!ctx)
102 return false;
103
104 return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
105 }
106
107 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
108 vm_flags_t flags)
109 {
110 const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP;
111
112 vm_flags_reset(vma, flags);
113 /*
114 * For shared mappings, we want to enable writenotify while
115 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
116 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
117 */
118 if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
119 vma_set_page_prot(vma);
120 }
121
122 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
123 int wake_flags, void *key)
124 {
125 struct userfaultfd_wake_range *range = key;
126 int ret;
127 struct userfaultfd_wait_queue *uwq;
128 unsigned long start, len;
129
130 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
131 ret = 0;
132 /* len == 0 means wake all */
133 start = range->start;
134 len = range->len;
135 if (len && (start > uwq->msg.arg.pagefault.address ||
136 start + len <= uwq->msg.arg.pagefault.address))
137 goto out;
138 WRITE_ONCE(uwq->waken, true);
139 /*
140 * The Program-Order guarantees provided by the scheduler
141 * ensure uwq->waken is visible before the task is woken.
142 */
143 ret = wake_up_state(wq->private, mode);
144 if (ret) {
145 /*
146 * Wake only once, autoremove behavior.
147 *
148 * After the effect of list_del_init is visible to the other
149 * CPUs, the waitqueue may disappear from under us, see the
150 * !list_empty_careful() in handle_userfault().
151 *
152 * try_to_wake_up() has an implicit smp_mb(), and the
153 * wq->private is read before calling the extern function
154 * "wake_up_state" (which in turns calls try_to_wake_up).
155 */
156 list_del_init(&wq->entry);
157 }
158 out:
159 return ret;
160 }
161
162 /**
163 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
164 * context.
165 * @ctx: [in] Pointer to the userfaultfd context.
166 */
167 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
168 {
169 refcount_inc(&ctx->refcount);
170 }
171
172 /**
173 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
174 * context.
175 * @ctx: [in] Pointer to userfaultfd context.
176 *
177 * The userfaultfd context reference must have been previously acquired either
178 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
179 */
180 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
181 {
182 if (refcount_dec_and_test(&ctx->refcount)) {
183 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
184 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
185 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
186 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
187 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
188 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
189 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
190 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
191 mmdrop(ctx->mm);
192 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
193 }
194 }
195
196 static inline void msg_init(struct uffd_msg *msg)
197 {
198 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
199 /*
200 * Must use memset to zero out the paddings or kernel data is
201 * leaked to userland.
202 */
203 memset(msg, 0, sizeof(struct uffd_msg));
204 }
205
206 static inline struct uffd_msg userfault_msg(unsigned long address,
207 unsigned long real_address,
208 unsigned int flags,
209 unsigned long reason,
210 unsigned int features)
211 {
212 struct uffd_msg msg;
213
214 msg_init(&msg);
215 msg.event = UFFD_EVENT_PAGEFAULT;
216
217 msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
218 real_address : address;
219
220 /*
221 * These flags indicate why the userfault occurred:
222 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
223 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
224 * - Neither of these flags being set indicates a MISSING fault.
225 *
226 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
227 * fault. Otherwise, it was a read fault.
228 */
229 if (flags & FAULT_FLAG_WRITE)
230 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
231 if (reason & VM_UFFD_WP)
232 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
233 if (reason & VM_UFFD_MINOR)
234 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
235 if (features & UFFD_FEATURE_THREAD_ID)
236 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
237 return msg;
238 }
239
240 #ifdef CONFIG_HUGETLB_PAGE
241 /*
242 * Same functionality as userfaultfd_must_wait below with modifications for
243 * hugepmd ranges.
244 */
245 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
246 struct vm_fault *vmf,
247 unsigned long reason)
248 {
249 struct vm_area_struct *vma = vmf->vma;
250 pte_t *ptep, pte;
251 bool ret = true;
252
253 assert_fault_locked(vmf);
254
255 ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
256 if (!ptep)
257 goto out;
258
259 ret = false;
260 pte = huge_ptep_get(vma->vm_mm, vmf->address, ptep);
261
262 /*
263 * Lockless access: we're in a wait_event so it's ok if it
264 * changes under us. PTE markers should be handled the same as none
265 * ptes here.
266 */
267 if (huge_pte_none_mostly(pte))
268 ret = true;
269 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
270 ret = true;
271 out:
272 return ret;
273 }
274 #else
275 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
276 struct vm_fault *vmf,
277 unsigned long reason)
278 {
279 return false; /* should never get here */
280 }
281 #endif /* CONFIG_HUGETLB_PAGE */
282
283 /*
284 * Verify the pagetables are still not ok after having reigstered into
285 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
286 * userfault that has already been resolved, if userfaultfd_read_iter and
287 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
288 * threads.
289 */
290 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
291 struct vm_fault *vmf,
292 unsigned long reason)
293 {
294 struct mm_struct *mm = ctx->mm;
295 unsigned long address = vmf->address;
296 pgd_t *pgd;
297 p4d_t *p4d;
298 pud_t *pud;
299 pmd_t *pmd, _pmd;
300 pte_t *pte;
301 pte_t ptent;
302 bool ret = true;
303
304 assert_fault_locked(vmf);
305
306 pgd = pgd_offset(mm, address);
307 if (!pgd_present(*pgd))
308 goto out;
309 p4d = p4d_offset(pgd, address);
310 if (!p4d_present(*p4d))
311 goto out;
312 pud = pud_offset(p4d, address);
313 if (!pud_present(*pud))
314 goto out;
315 pmd = pmd_offset(pud, address);
316 again:
317 _pmd = pmdp_get_lockless(pmd);
318 if (pmd_none(_pmd))
319 goto out;
320
321 ret = false;
322 if (!pmd_present(_pmd) || pmd_devmap(_pmd))
323 goto out;
324
325 if (pmd_trans_huge(_pmd)) {
326 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
327 ret = true;
328 goto out;
329 }
330
331 pte = pte_offset_map(pmd, address);
332 if (!pte) {
333 ret = true;
334 goto again;
335 }
336 /*
337 * Lockless access: we're in a wait_event so it's ok if it
338 * changes under us. PTE markers should be handled the same as none
339 * ptes here.
340 */
341 ptent = ptep_get(pte);
342 if (pte_none_mostly(ptent))
343 ret = true;
344 if (!pte_write(ptent) && (reason & VM_UFFD_WP))
345 ret = true;
346 pte_unmap(pte);
347
348 out:
349 return ret;
350 }
351
352 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
353 {
354 if (flags & FAULT_FLAG_INTERRUPTIBLE)
355 return TASK_INTERRUPTIBLE;
356
357 if (flags & FAULT_FLAG_KILLABLE)
358 return TASK_KILLABLE;
359
360 return TASK_UNINTERRUPTIBLE;
361 }
362
363 /*
364 * The locking rules involved in returning VM_FAULT_RETRY depending on
365 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
366 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
367 * recommendation in __lock_page_or_retry is not an understatement.
368 *
369 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
370 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
371 * not set.
372 *
373 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
374 * set, VM_FAULT_RETRY can still be returned if and only if there are
375 * fatal_signal_pending()s, and the mmap_lock must be released before
376 * returning it.
377 */
378 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
379 {
380 struct vm_area_struct *vma = vmf->vma;
381 struct mm_struct *mm = vma->vm_mm;
382 struct userfaultfd_ctx *ctx;
383 struct userfaultfd_wait_queue uwq;
384 vm_fault_t ret = VM_FAULT_SIGBUS;
385 bool must_wait;
386 unsigned int blocking_state;
387
388 /*
389 * We don't do userfault handling for the final child pid update.
390 *
391 * We also don't do userfault handling during
392 * coredumping. hugetlbfs has the special
393 * hugetlb_follow_page_mask() to skip missing pages in the
394 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
395 * the no_page_table() helper in follow_page_mask(), but the
396 * shmem_vm_ops->fault method is invoked even during
397 * coredumping and it ends up here.
398 */
399 if (current->flags & (PF_EXITING|PF_DUMPCORE))
400 goto out;
401
402 assert_fault_locked(vmf);
403
404 ctx = vma->vm_userfaultfd_ctx.ctx;
405 if (!ctx)
406 goto out;
407
408 BUG_ON(ctx->mm != mm);
409
410 /* Any unrecognized flag is a bug. */
411 VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
412 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
413 VM_BUG_ON(!reason || (reason & (reason - 1)));
414
415 if (ctx->features & UFFD_FEATURE_SIGBUS)
416 goto out;
417 if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY))
418 goto out;
419
420 /*
421 * If it's already released don't get it. This avoids to loop
422 * in __get_user_pages if userfaultfd_release waits on the
423 * caller of handle_userfault to release the mmap_lock.
424 */
425 if (unlikely(READ_ONCE(ctx->released))) {
426 /*
427 * Don't return VM_FAULT_SIGBUS in this case, so a non
428 * cooperative manager can close the uffd after the
429 * last UFFDIO_COPY, without risking to trigger an
430 * involuntary SIGBUS if the process was starting the
431 * userfaultfd while the userfaultfd was still armed
432 * (but after the last UFFDIO_COPY). If the uffd
433 * wasn't already closed when the userfault reached
434 * this point, that would normally be solved by
435 * userfaultfd_must_wait returning 'false'.
436 *
437 * If we were to return VM_FAULT_SIGBUS here, the non
438 * cooperative manager would be instead forced to
439 * always call UFFDIO_UNREGISTER before it can safely
440 * close the uffd.
441 */
442 ret = VM_FAULT_NOPAGE;
443 goto out;
444 }
445
446 /*
447 * Check that we can return VM_FAULT_RETRY.
448 *
449 * NOTE: it should become possible to return VM_FAULT_RETRY
450 * even if FAULT_FLAG_TRIED is set without leading to gup()
451 * -EBUSY failures, if the userfaultfd is to be extended for
452 * VM_UFFD_WP tracking and we intend to arm the userfault
453 * without first stopping userland access to the memory. For
454 * VM_UFFD_MISSING userfaults this is enough for now.
455 */
456 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
457 /*
458 * Validate the invariant that nowait must allow retry
459 * to be sure not to return SIGBUS erroneously on
460 * nowait invocations.
461 */
462 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
463 #ifdef CONFIG_DEBUG_VM
464 if (printk_ratelimit()) {
465 printk(KERN_WARNING
466 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
467 vmf->flags);
468 dump_stack();
469 }
470 #endif
471 goto out;
472 }
473
474 /*
475 * Handle nowait, not much to do other than tell it to retry
476 * and wait.
477 */
478 ret = VM_FAULT_RETRY;
479 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
480 goto out;
481
482 /* take the reference before dropping the mmap_lock */
483 userfaultfd_ctx_get(ctx);
484
485 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
486 uwq.wq.private = current;
487 uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
488 reason, ctx->features);
489 uwq.ctx = ctx;
490 uwq.waken = false;
491
492 blocking_state = userfaultfd_get_blocking_state(vmf->flags);
493
494 /*
495 * Take the vma lock now, in order to safely call
496 * userfaultfd_huge_must_wait() later. Since acquiring the
497 * (sleepable) vma lock can modify the current task state, that
498 * must be before explicitly calling set_current_state().
499 */
500 if (is_vm_hugetlb_page(vma))
501 hugetlb_vma_lock_read(vma);
502
503 spin_lock_irq(&ctx->fault_pending_wqh.lock);
504 /*
505 * After the __add_wait_queue the uwq is visible to userland
506 * through poll/read().
507 */
508 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
509 /*
510 * The smp_mb() after __set_current_state prevents the reads
511 * following the spin_unlock to happen before the list_add in
512 * __add_wait_queue.
513 */
514 set_current_state(blocking_state);
515 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
516
517 if (!is_vm_hugetlb_page(vma))
518 must_wait = userfaultfd_must_wait(ctx, vmf, reason);
519 else
520 must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
521 if (is_vm_hugetlb_page(vma))
522 hugetlb_vma_unlock_read(vma);
523 release_fault_lock(vmf);
524
525 if (likely(must_wait && !READ_ONCE(ctx->released))) {
526 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
527 schedule();
528 }
529
530 __set_current_state(TASK_RUNNING);
531
532 /*
533 * Here we race with the list_del; list_add in
534 * userfaultfd_ctx_read(), however because we don't ever run
535 * list_del_init() to refile across the two lists, the prev
536 * and next pointers will never point to self. list_add also
537 * would never let any of the two pointers to point to
538 * self. So list_empty_careful won't risk to see both pointers
539 * pointing to self at any time during the list refile. The
540 * only case where list_del_init() is called is the full
541 * removal in the wake function and there we don't re-list_add
542 * and it's fine not to block on the spinlock. The uwq on this
543 * kernel stack can be released after the list_del_init.
544 */
545 if (!list_empty_careful(&uwq.wq.entry)) {
546 spin_lock_irq(&ctx->fault_pending_wqh.lock);
547 /*
548 * No need of list_del_init(), the uwq on the stack
549 * will be freed shortly anyway.
550 */
551 list_del(&uwq.wq.entry);
552 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
553 }
554
555 /*
556 * ctx may go away after this if the userfault pseudo fd is
557 * already released.
558 */
559 userfaultfd_ctx_put(ctx);
560
561 out:
562 return ret;
563 }
564
565 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
566 struct userfaultfd_wait_queue *ewq)
567 {
568 struct userfaultfd_ctx *release_new_ctx;
569
570 if (WARN_ON_ONCE(current->flags & PF_EXITING))
571 goto out;
572
573 ewq->ctx = ctx;
574 init_waitqueue_entry(&ewq->wq, current);
575 release_new_ctx = NULL;
576
577 spin_lock_irq(&ctx->event_wqh.lock);
578 /*
579 * After the __add_wait_queue the uwq is visible to userland
580 * through poll/read().
581 */
582 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
583 for (;;) {
584 set_current_state(TASK_KILLABLE);
585 if (ewq->msg.event == 0)
586 break;
587 if (READ_ONCE(ctx->released) ||
588 fatal_signal_pending(current)) {
589 /*
590 * &ewq->wq may be queued in fork_event, but
591 * __remove_wait_queue ignores the head
592 * parameter. It would be a problem if it
593 * didn't.
594 */
595 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
596 if (ewq->msg.event == UFFD_EVENT_FORK) {
597 struct userfaultfd_ctx *new;
598
599 new = (struct userfaultfd_ctx *)
600 (unsigned long)
601 ewq->msg.arg.reserved.reserved1;
602 release_new_ctx = new;
603 }
604 break;
605 }
606
607 spin_unlock_irq(&ctx->event_wqh.lock);
608
609 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
610 schedule();
611
612 spin_lock_irq(&ctx->event_wqh.lock);
613 }
614 __set_current_state(TASK_RUNNING);
615 spin_unlock_irq(&ctx->event_wqh.lock);
616
617 if (release_new_ctx) {
618 struct vm_area_struct *vma;
619 struct mm_struct *mm = release_new_ctx->mm;
620 VMA_ITERATOR(vmi, mm, 0);
621
622 /* the various vma->vm_userfaultfd_ctx still points to it */
623 mmap_write_lock(mm);
624 for_each_vma(vmi, vma) {
625 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
626 vma_start_write(vma);
627 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
628 userfaultfd_set_vm_flags(vma,
629 vma->vm_flags & ~__VM_UFFD_FLAGS);
630 }
631 }
632 mmap_write_unlock(mm);
633
634 userfaultfd_ctx_put(release_new_ctx);
635 }
636
637 /*
638 * ctx may go away after this if the userfault pseudo fd is
639 * already released.
640 */
641 out:
642 atomic_dec(&ctx->mmap_changing);
643 VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
644 userfaultfd_ctx_put(ctx);
645 }
646
647 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
648 struct userfaultfd_wait_queue *ewq)
649 {
650 ewq->msg.event = 0;
651 wake_up_locked(&ctx->event_wqh);
652 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
653 }
654
655 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
656 {
657 struct userfaultfd_ctx *ctx = NULL, *octx;
658 struct userfaultfd_fork_ctx *fctx;
659
660 octx = vma->vm_userfaultfd_ctx.ctx;
661 if (!octx)
662 return 0;
663
664 if (!(octx->features & UFFD_FEATURE_EVENT_FORK)) {
665 vma_start_write(vma);
666 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
667 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
668 return 0;
669 }
670
671 list_for_each_entry(fctx, fcs, list)
672 if (fctx->orig == octx) {
673 ctx = fctx->new;
674 break;
675 }
676
677 if (!ctx) {
678 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
679 if (!fctx)
680 return -ENOMEM;
681
682 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
683 if (!ctx) {
684 kfree(fctx);
685 return -ENOMEM;
686 }
687
688 refcount_set(&ctx->refcount, 1);
689 ctx->flags = octx->flags;
690 ctx->features = octx->features;
691 ctx->released = false;
692 init_rwsem(&ctx->map_changing_lock);
693 atomic_set(&ctx->mmap_changing, 0);
694 ctx->mm = vma->vm_mm;
695 mmgrab(ctx->mm);
696
697 userfaultfd_ctx_get(octx);
698 down_write(&octx->map_changing_lock);
699 atomic_inc(&octx->mmap_changing);
700 up_write(&octx->map_changing_lock);
701 fctx->orig = octx;
702 fctx->new = ctx;
703 list_add_tail(&fctx->list, fcs);
704 }
705
706 vma->vm_userfaultfd_ctx.ctx = ctx;
707 return 0;
708 }
709
710 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
711 {
712 struct userfaultfd_ctx *ctx = fctx->orig;
713 struct userfaultfd_wait_queue ewq;
714
715 msg_init(&ewq.msg);
716
717 ewq.msg.event = UFFD_EVENT_FORK;
718 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
719
720 userfaultfd_event_wait_completion(ctx, &ewq);
721 }
722
723 void dup_userfaultfd_complete(struct list_head *fcs)
724 {
725 struct userfaultfd_fork_ctx *fctx, *n;
726
727 list_for_each_entry_safe(fctx, n, fcs, list) {
728 dup_fctx(fctx);
729 list_del(&fctx->list);
730 kfree(fctx);
731 }
732 }
733
734 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
735 struct vm_userfaultfd_ctx *vm_ctx)
736 {
737 struct userfaultfd_ctx *ctx;
738
739 ctx = vma->vm_userfaultfd_ctx.ctx;
740
741 if (!ctx)
742 return;
743
744 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
745 vm_ctx->ctx = ctx;
746 userfaultfd_ctx_get(ctx);
747 down_write(&ctx->map_changing_lock);
748 atomic_inc(&ctx->mmap_changing);
749 up_write(&ctx->map_changing_lock);
750 } else {
751 /* Drop uffd context if remap feature not enabled */
752 vma_start_write(vma);
753 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
754 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
755 }
756 }
757
758 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
759 unsigned long from, unsigned long to,
760 unsigned long len)
761 {
762 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
763 struct userfaultfd_wait_queue ewq;
764
765 if (!ctx)
766 return;
767
768 if (to & ~PAGE_MASK) {
769 userfaultfd_ctx_put(ctx);
770 return;
771 }
772
773 msg_init(&ewq.msg);
774
775 ewq.msg.event = UFFD_EVENT_REMAP;
776 ewq.msg.arg.remap.from = from;
777 ewq.msg.arg.remap.to = to;
778 ewq.msg.arg.remap.len = len;
779
780 userfaultfd_event_wait_completion(ctx, &ewq);
781 }
782
783 bool userfaultfd_remove(struct vm_area_struct *vma,
784 unsigned long start, unsigned long end)
785 {
786 struct mm_struct *mm = vma->vm_mm;
787 struct userfaultfd_ctx *ctx;
788 struct userfaultfd_wait_queue ewq;
789
790 ctx = vma->vm_userfaultfd_ctx.ctx;
791 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
792 return true;
793
794 userfaultfd_ctx_get(ctx);
795 down_write(&ctx->map_changing_lock);
796 atomic_inc(&ctx->mmap_changing);
797 up_write(&ctx->map_changing_lock);
798 mmap_read_unlock(mm);
799
800 msg_init(&ewq.msg);
801
802 ewq.msg.event = UFFD_EVENT_REMOVE;
803 ewq.msg.arg.remove.start = start;
804 ewq.msg.arg.remove.end = end;
805
806 userfaultfd_event_wait_completion(ctx, &ewq);
807
808 return false;
809 }
810
811 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
812 unsigned long start, unsigned long end)
813 {
814 struct userfaultfd_unmap_ctx *unmap_ctx;
815
816 list_for_each_entry(unmap_ctx, unmaps, list)
817 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
818 unmap_ctx->end == end)
819 return true;
820
821 return false;
822 }
823
824 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
825 unsigned long end, struct list_head *unmaps)
826 {
827 struct userfaultfd_unmap_ctx *unmap_ctx;
828 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
829
830 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
831 has_unmap_ctx(ctx, unmaps, start, end))
832 return 0;
833
834 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
835 if (!unmap_ctx)
836 return -ENOMEM;
837
838 userfaultfd_ctx_get(ctx);
839 down_write(&ctx->map_changing_lock);
840 atomic_inc(&ctx->mmap_changing);
841 up_write(&ctx->map_changing_lock);
842 unmap_ctx->ctx = ctx;
843 unmap_ctx->start = start;
844 unmap_ctx->end = end;
845 list_add_tail(&unmap_ctx->list, unmaps);
846
847 return 0;
848 }
849
850 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
851 {
852 struct userfaultfd_unmap_ctx *ctx, *n;
853 struct userfaultfd_wait_queue ewq;
854
855 list_for_each_entry_safe(ctx, n, uf, list) {
856 msg_init(&ewq.msg);
857
858 ewq.msg.event = UFFD_EVENT_UNMAP;
859 ewq.msg.arg.remove.start = ctx->start;
860 ewq.msg.arg.remove.end = ctx->end;
861
862 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
863
864 list_del(&ctx->list);
865 kfree(ctx);
866 }
867 }
868
869 static int userfaultfd_release(struct inode *inode, struct file *file)
870 {
871 struct userfaultfd_ctx *ctx = file->private_data;
872 struct mm_struct *mm = ctx->mm;
873 struct vm_area_struct *vma, *prev;
874 /* len == 0 means wake all */
875 struct userfaultfd_wake_range range = { .len = 0, };
876 unsigned long new_flags;
877 VMA_ITERATOR(vmi, mm, 0);
878
879 WRITE_ONCE(ctx->released, true);
880
881 if (!mmget_not_zero(mm))
882 goto wakeup;
883
884 /*
885 * Flush page faults out of all CPUs. NOTE: all page faults
886 * must be retried without returning VM_FAULT_SIGBUS if
887 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
888 * changes while handle_userfault released the mmap_lock. So
889 * it's critical that released is set to true (above), before
890 * taking the mmap_lock for writing.
891 */
892 mmap_write_lock(mm);
893 prev = NULL;
894 for_each_vma(vmi, vma) {
895 cond_resched();
896 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
897 !!(vma->vm_flags & __VM_UFFD_FLAGS));
898 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
899 prev = vma;
900 continue;
901 }
902 /* Reset ptes for the whole vma range if wr-protected */
903 if (userfaultfd_wp(vma))
904 uffd_wp_range(vma, vma->vm_start,
905 vma->vm_end - vma->vm_start, false);
906 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
907 vma = vma_modify_flags_uffd(&vmi, prev, vma, vma->vm_start,
908 vma->vm_end, new_flags,
909 NULL_VM_UFFD_CTX);
910
911 vma_start_write(vma);
912 userfaultfd_set_vm_flags(vma, new_flags);
913 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
914
915 prev = vma;
916 }
917 mmap_write_unlock(mm);
918 mmput(mm);
919 wakeup:
920 /*
921 * After no new page faults can wait on this fault_*wqh, flush
922 * the last page faults that may have been already waiting on
923 * the fault_*wqh.
924 */
925 spin_lock_irq(&ctx->fault_pending_wqh.lock);
926 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
927 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
928 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
929
930 /* Flush pending events that may still wait on event_wqh */
931 wake_up_all(&ctx->event_wqh);
932
933 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
934 userfaultfd_ctx_put(ctx);
935 return 0;
936 }
937
938 /* fault_pending_wqh.lock must be hold by the caller */
939 static inline struct userfaultfd_wait_queue *find_userfault_in(
940 wait_queue_head_t *wqh)
941 {
942 wait_queue_entry_t *wq;
943 struct userfaultfd_wait_queue *uwq;
944
945 lockdep_assert_held(&wqh->lock);
946
947 uwq = NULL;
948 if (!waitqueue_active(wqh))
949 goto out;
950 /* walk in reverse to provide FIFO behavior to read userfaults */
951 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
952 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
953 out:
954 return uwq;
955 }
956
957 static inline struct userfaultfd_wait_queue *find_userfault(
958 struct userfaultfd_ctx *ctx)
959 {
960 return find_userfault_in(&ctx->fault_pending_wqh);
961 }
962
963 static inline struct userfaultfd_wait_queue *find_userfault_evt(
964 struct userfaultfd_ctx *ctx)
965 {
966 return find_userfault_in(&ctx->event_wqh);
967 }
968
969 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
970 {
971 struct userfaultfd_ctx *ctx = file->private_data;
972 __poll_t ret;
973
974 poll_wait(file, &ctx->fd_wqh, wait);
975
976 if (!userfaultfd_is_initialized(ctx))
977 return EPOLLERR;
978
979 /*
980 * poll() never guarantees that read won't block.
981 * userfaults can be waken before they're read().
982 */
983 if (unlikely(!(file->f_flags & O_NONBLOCK)))
984 return EPOLLERR;
985 /*
986 * lockless access to see if there are pending faults
987 * __pollwait last action is the add_wait_queue but
988 * the spin_unlock would allow the waitqueue_active to
989 * pass above the actual list_add inside
990 * add_wait_queue critical section. So use a full
991 * memory barrier to serialize the list_add write of
992 * add_wait_queue() with the waitqueue_active read
993 * below.
994 */
995 ret = 0;
996 smp_mb();
997 if (waitqueue_active(&ctx->fault_pending_wqh))
998 ret = EPOLLIN;
999 else if (waitqueue_active(&ctx->event_wqh))
1000 ret = EPOLLIN;
1001
1002 return ret;
1003 }
1004
1005 static const struct file_operations userfaultfd_fops;
1006
1007 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
1008 struct inode *inode,
1009 struct uffd_msg *msg)
1010 {
1011 int fd;
1012
1013 fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, new,
1014 O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1015 if (fd < 0)
1016 return fd;
1017
1018 msg->arg.reserved.reserved1 = 0;
1019 msg->arg.fork.ufd = fd;
1020 return 0;
1021 }
1022
1023 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1024 struct uffd_msg *msg, struct inode *inode)
1025 {
1026 ssize_t ret;
1027 DECLARE_WAITQUEUE(wait, current);
1028 struct userfaultfd_wait_queue *uwq;
1029 /*
1030 * Handling fork event requires sleeping operations, so
1031 * we drop the event_wqh lock, then do these ops, then
1032 * lock it back and wake up the waiter. While the lock is
1033 * dropped the ewq may go away so we keep track of it
1034 * carefully.
1035 */
1036 LIST_HEAD(fork_event);
1037 struct userfaultfd_ctx *fork_nctx = NULL;
1038
1039 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1040 spin_lock_irq(&ctx->fd_wqh.lock);
1041 __add_wait_queue(&ctx->fd_wqh, &wait);
1042 for (;;) {
1043 set_current_state(TASK_INTERRUPTIBLE);
1044 spin_lock(&ctx->fault_pending_wqh.lock);
1045 uwq = find_userfault(ctx);
1046 if (uwq) {
1047 /*
1048 * Use a seqcount to repeat the lockless check
1049 * in wake_userfault() to avoid missing
1050 * wakeups because during the refile both
1051 * waitqueue could become empty if this is the
1052 * only userfault.
1053 */
1054 write_seqcount_begin(&ctx->refile_seq);
1055
1056 /*
1057 * The fault_pending_wqh.lock prevents the uwq
1058 * to disappear from under us.
1059 *
1060 * Refile this userfault from
1061 * fault_pending_wqh to fault_wqh, it's not
1062 * pending anymore after we read it.
1063 *
1064 * Use list_del() by hand (as
1065 * userfaultfd_wake_function also uses
1066 * list_del_init() by hand) to be sure nobody
1067 * changes __remove_wait_queue() to use
1068 * list_del_init() in turn breaking the
1069 * !list_empty_careful() check in
1070 * handle_userfault(). The uwq->wq.head list
1071 * must never be empty at any time during the
1072 * refile, or the waitqueue could disappear
1073 * from under us. The "wait_queue_head_t"
1074 * parameter of __remove_wait_queue() is unused
1075 * anyway.
1076 */
1077 list_del(&uwq->wq.entry);
1078 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1079
1080 write_seqcount_end(&ctx->refile_seq);
1081
1082 /* careful to always initialize msg if ret == 0 */
1083 *msg = uwq->msg;
1084 spin_unlock(&ctx->fault_pending_wqh.lock);
1085 ret = 0;
1086 break;
1087 }
1088 spin_unlock(&ctx->fault_pending_wqh.lock);
1089
1090 spin_lock(&ctx->event_wqh.lock);
1091 uwq = find_userfault_evt(ctx);
1092 if (uwq) {
1093 *msg = uwq->msg;
1094
1095 if (uwq->msg.event == UFFD_EVENT_FORK) {
1096 fork_nctx = (struct userfaultfd_ctx *)
1097 (unsigned long)
1098 uwq->msg.arg.reserved.reserved1;
1099 list_move(&uwq->wq.entry, &fork_event);
1100 /*
1101 * fork_nctx can be freed as soon as
1102 * we drop the lock, unless we take a
1103 * reference on it.
1104 */
1105 userfaultfd_ctx_get(fork_nctx);
1106 spin_unlock(&ctx->event_wqh.lock);
1107 ret = 0;
1108 break;
1109 }
1110
1111 userfaultfd_event_complete(ctx, uwq);
1112 spin_unlock(&ctx->event_wqh.lock);
1113 ret = 0;
1114 break;
1115 }
1116 spin_unlock(&ctx->event_wqh.lock);
1117
1118 if (signal_pending(current)) {
1119 ret = -ERESTARTSYS;
1120 break;
1121 }
1122 if (no_wait) {
1123 ret = -EAGAIN;
1124 break;
1125 }
1126 spin_unlock_irq(&ctx->fd_wqh.lock);
1127 schedule();
1128 spin_lock_irq(&ctx->fd_wqh.lock);
1129 }
1130 __remove_wait_queue(&ctx->fd_wqh, &wait);
1131 __set_current_state(TASK_RUNNING);
1132 spin_unlock_irq(&ctx->fd_wqh.lock);
1133
1134 if (!ret && msg->event == UFFD_EVENT_FORK) {
1135 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1136 spin_lock_irq(&ctx->event_wqh.lock);
1137 if (!list_empty(&fork_event)) {
1138 /*
1139 * The fork thread didn't abort, so we can
1140 * drop the temporary refcount.
1141 */
1142 userfaultfd_ctx_put(fork_nctx);
1143
1144 uwq = list_first_entry(&fork_event,
1145 typeof(*uwq),
1146 wq.entry);
1147 /*
1148 * If fork_event list wasn't empty and in turn
1149 * the event wasn't already released by fork
1150 * (the event is allocated on fork kernel
1151 * stack), put the event back to its place in
1152 * the event_wq. fork_event head will be freed
1153 * as soon as we return so the event cannot
1154 * stay queued there no matter the current
1155 * "ret" value.
1156 */
1157 list_del(&uwq->wq.entry);
1158 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1159
1160 /*
1161 * Leave the event in the waitqueue and report
1162 * error to userland if we failed to resolve
1163 * the userfault fork.
1164 */
1165 if (likely(!ret))
1166 userfaultfd_event_complete(ctx, uwq);
1167 } else {
1168 /*
1169 * Here the fork thread aborted and the
1170 * refcount from the fork thread on fork_nctx
1171 * has already been released. We still hold
1172 * the reference we took before releasing the
1173 * lock above. If resolve_userfault_fork
1174 * failed we've to drop it because the
1175 * fork_nctx has to be freed in such case. If
1176 * it succeeded we'll hold it because the new
1177 * uffd references it.
1178 */
1179 if (ret)
1180 userfaultfd_ctx_put(fork_nctx);
1181 }
1182 spin_unlock_irq(&ctx->event_wqh.lock);
1183 }
1184
1185 return ret;
1186 }
1187
1188 static ssize_t userfaultfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
1189 {
1190 struct file *file = iocb->ki_filp;
1191 struct userfaultfd_ctx *ctx = file->private_data;
1192 ssize_t _ret, ret = 0;
1193 struct uffd_msg msg;
1194 struct inode *inode = file_inode(file);
1195 bool no_wait;
1196
1197 if (!userfaultfd_is_initialized(ctx))
1198 return -EINVAL;
1199
1200 no_wait = file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT;
1201 for (;;) {
1202 if (iov_iter_count(to) < sizeof(msg))
1203 return ret ? ret : -EINVAL;
1204 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1205 if (_ret < 0)
1206 return ret ? ret : _ret;
1207 _ret = !copy_to_iter_full(&msg, sizeof(msg), to);
1208 if (_ret)
1209 return ret ? ret : -EFAULT;
1210 ret += sizeof(msg);
1211 /*
1212 * Allow to read more than one fault at time but only
1213 * block if waiting for the very first one.
1214 */
1215 no_wait = true;
1216 }
1217 }
1218
1219 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1220 struct userfaultfd_wake_range *range)
1221 {
1222 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1223 /* wake all in the range and autoremove */
1224 if (waitqueue_active(&ctx->fault_pending_wqh))
1225 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1226 range);
1227 if (waitqueue_active(&ctx->fault_wqh))
1228 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1229 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1230 }
1231
1232 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1233 struct userfaultfd_wake_range *range)
1234 {
1235 unsigned seq;
1236 bool need_wakeup;
1237
1238 /*
1239 * To be sure waitqueue_active() is not reordered by the CPU
1240 * before the pagetable update, use an explicit SMP memory
1241 * barrier here. PT lock release or mmap_read_unlock(mm) still
1242 * have release semantics that can allow the
1243 * waitqueue_active() to be reordered before the pte update.
1244 */
1245 smp_mb();
1246
1247 /*
1248 * Use waitqueue_active because it's very frequent to
1249 * change the address space atomically even if there are no
1250 * userfaults yet. So we take the spinlock only when we're
1251 * sure we've userfaults to wake.
1252 */
1253 do {
1254 seq = read_seqcount_begin(&ctx->refile_seq);
1255 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1256 waitqueue_active(&ctx->fault_wqh);
1257 cond_resched();
1258 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1259 if (need_wakeup)
1260 __wake_userfault(ctx, range);
1261 }
1262
1263 static __always_inline int validate_unaligned_range(
1264 struct mm_struct *mm, __u64 start, __u64 len)
1265 {
1266 __u64 task_size = mm->task_size;
1267
1268 if (len & ~PAGE_MASK)
1269 return -EINVAL;
1270 if (!len)
1271 return -EINVAL;
1272 if (start < mmap_min_addr)
1273 return -EINVAL;
1274 if (start >= task_size)
1275 return -EINVAL;
1276 if (len > task_size - start)
1277 return -EINVAL;
1278 if (start + len <= start)
1279 return -EINVAL;
1280 return 0;
1281 }
1282
1283 static __always_inline int validate_range(struct mm_struct *mm,
1284 __u64 start, __u64 len)
1285 {
1286 if (start & ~PAGE_MASK)
1287 return -EINVAL;
1288
1289 return validate_unaligned_range(mm, start, len);
1290 }
1291
1292 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1293 unsigned long arg)
1294 {
1295 struct mm_struct *mm = ctx->mm;
1296 struct vm_area_struct *vma, *prev, *cur;
1297 int ret;
1298 struct uffdio_register uffdio_register;
1299 struct uffdio_register __user *user_uffdio_register;
1300 unsigned long vm_flags, new_flags;
1301 bool found;
1302 bool basic_ioctls;
1303 unsigned long start, end, vma_end;
1304 struct vma_iterator vmi;
1305 bool wp_async = userfaultfd_wp_async_ctx(ctx);
1306
1307 user_uffdio_register = (struct uffdio_register __user *) arg;
1308
1309 ret = -EFAULT;
1310 if (copy_from_user(&uffdio_register, user_uffdio_register,
1311 sizeof(uffdio_register)-sizeof(__u64)))
1312 goto out;
1313
1314 ret = -EINVAL;
1315 if (!uffdio_register.mode)
1316 goto out;
1317 if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1318 goto out;
1319 vm_flags = 0;
1320 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1321 vm_flags |= VM_UFFD_MISSING;
1322 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1323 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1324 goto out;
1325 #endif
1326 vm_flags |= VM_UFFD_WP;
1327 }
1328 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1329 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1330 goto out;
1331 #endif
1332 vm_flags |= VM_UFFD_MINOR;
1333 }
1334
1335 ret = validate_range(mm, uffdio_register.range.start,
1336 uffdio_register.range.len);
1337 if (ret)
1338 goto out;
1339
1340 start = uffdio_register.range.start;
1341 end = start + uffdio_register.range.len;
1342
1343 ret = -ENOMEM;
1344 if (!mmget_not_zero(mm))
1345 goto out;
1346
1347 ret = -EINVAL;
1348 mmap_write_lock(mm);
1349 vma_iter_init(&vmi, mm, start);
1350 vma = vma_find(&vmi, end);
1351 if (!vma)
1352 goto out_unlock;
1353
1354 /*
1355 * If the first vma contains huge pages, make sure start address
1356 * is aligned to huge page size.
1357 */
1358 if (is_vm_hugetlb_page(vma)) {
1359 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1360
1361 if (start & (vma_hpagesize - 1))
1362 goto out_unlock;
1363 }
1364
1365 /*
1366 * Search for not compatible vmas.
1367 */
1368 found = false;
1369 basic_ioctls = false;
1370 cur = vma;
1371 do {
1372 cond_resched();
1373
1374 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1375 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1376
1377 /* check not compatible vmas */
1378 ret = -EINVAL;
1379 if (!vma_can_userfault(cur, vm_flags, wp_async))
1380 goto out_unlock;
1381
1382 /*
1383 * UFFDIO_COPY will fill file holes even without
1384 * PROT_WRITE. This check enforces that if this is a
1385 * MAP_SHARED, the process has write permission to the backing
1386 * file. If VM_MAYWRITE is set it also enforces that on a
1387 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1388 * F_WRITE_SEAL can be taken until the vma is destroyed.
1389 */
1390 ret = -EPERM;
1391 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1392 goto out_unlock;
1393
1394 /*
1395 * If this vma contains ending address, and huge pages
1396 * check alignment.
1397 */
1398 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1399 end > cur->vm_start) {
1400 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1401
1402 ret = -EINVAL;
1403
1404 if (end & (vma_hpagesize - 1))
1405 goto out_unlock;
1406 }
1407 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1408 goto out_unlock;
1409
1410 /*
1411 * Check that this vma isn't already owned by a
1412 * different userfaultfd. We can't allow more than one
1413 * userfaultfd to own a single vma simultaneously or we
1414 * wouldn't know which one to deliver the userfaults to.
1415 */
1416 ret = -EBUSY;
1417 if (cur->vm_userfaultfd_ctx.ctx &&
1418 cur->vm_userfaultfd_ctx.ctx != ctx)
1419 goto out_unlock;
1420
1421 /*
1422 * Note vmas containing huge pages
1423 */
1424 if (is_vm_hugetlb_page(cur))
1425 basic_ioctls = true;
1426
1427 found = true;
1428 } for_each_vma_range(vmi, cur, end);
1429 BUG_ON(!found);
1430
1431 vma_iter_set(&vmi, start);
1432 prev = vma_prev(&vmi);
1433 if (vma->vm_start < start)
1434 prev = vma;
1435
1436 ret = 0;
1437 for_each_vma_range(vmi, vma, end) {
1438 cond_resched();
1439
1440 BUG_ON(!vma_can_userfault(vma, vm_flags, wp_async));
1441 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1442 vma->vm_userfaultfd_ctx.ctx != ctx);
1443 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1444
1445 /*
1446 * Nothing to do: this vma is already registered into this
1447 * userfaultfd and with the right tracking mode too.
1448 */
1449 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1450 (vma->vm_flags & vm_flags) == vm_flags)
1451 goto skip;
1452
1453 if (vma->vm_start > start)
1454 start = vma->vm_start;
1455 vma_end = min(end, vma->vm_end);
1456
1457 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1458 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1459 new_flags,
1460 (struct vm_userfaultfd_ctx){ctx});
1461 if (IS_ERR(vma)) {
1462 ret = PTR_ERR(vma);
1463 break;
1464 }
1465
1466 /*
1467 * In the vma_merge() successful mprotect-like case 8:
1468 * the next vma was merged into the current one and
1469 * the current one has not been updated yet.
1470 */
1471 vma_start_write(vma);
1472 userfaultfd_set_vm_flags(vma, new_flags);
1473 vma->vm_userfaultfd_ctx.ctx = ctx;
1474
1475 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1476 hugetlb_unshare_all_pmds(vma);
1477
1478 skip:
1479 prev = vma;
1480 start = vma->vm_end;
1481 }
1482
1483 out_unlock:
1484 mmap_write_unlock(mm);
1485 mmput(mm);
1486 if (!ret) {
1487 __u64 ioctls_out;
1488
1489 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1490 UFFD_API_RANGE_IOCTLS;
1491
1492 /*
1493 * Declare the WP ioctl only if the WP mode is
1494 * specified and all checks passed with the range
1495 */
1496 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1497 ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1498
1499 /* CONTINUE ioctl is only supported for MINOR ranges. */
1500 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1501 ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1502
1503 /*
1504 * Now that we scanned all vmas we can already tell
1505 * userland which ioctls methods are guaranteed to
1506 * succeed on this range.
1507 */
1508 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1509 ret = -EFAULT;
1510 }
1511 out:
1512 return ret;
1513 }
1514
1515 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1516 unsigned long arg)
1517 {
1518 struct mm_struct *mm = ctx->mm;
1519 struct vm_area_struct *vma, *prev, *cur;
1520 int ret;
1521 struct uffdio_range uffdio_unregister;
1522 unsigned long new_flags;
1523 bool found;
1524 unsigned long start, end, vma_end;
1525 const void __user *buf = (void __user *)arg;
1526 struct vma_iterator vmi;
1527 bool wp_async = userfaultfd_wp_async_ctx(ctx);
1528
1529 ret = -EFAULT;
1530 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1531 goto out;
1532
1533 ret = validate_range(mm, uffdio_unregister.start,
1534 uffdio_unregister.len);
1535 if (ret)
1536 goto out;
1537
1538 start = uffdio_unregister.start;
1539 end = start + uffdio_unregister.len;
1540
1541 ret = -ENOMEM;
1542 if (!mmget_not_zero(mm))
1543 goto out;
1544
1545 mmap_write_lock(mm);
1546 ret = -EINVAL;
1547 vma_iter_init(&vmi, mm, start);
1548 vma = vma_find(&vmi, end);
1549 if (!vma)
1550 goto out_unlock;
1551
1552 /*
1553 * If the first vma contains huge pages, make sure start address
1554 * is aligned to huge page size.
1555 */
1556 if (is_vm_hugetlb_page(vma)) {
1557 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1558
1559 if (start & (vma_hpagesize - 1))
1560 goto out_unlock;
1561 }
1562
1563 /*
1564 * Search for not compatible vmas.
1565 */
1566 found = false;
1567 cur = vma;
1568 do {
1569 cond_resched();
1570
1571 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1572 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1573
1574 /*
1575 * Check not compatible vmas, not strictly required
1576 * here as not compatible vmas cannot have an
1577 * userfaultfd_ctx registered on them, but this
1578 * provides for more strict behavior to notice
1579 * unregistration errors.
1580 */
1581 if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
1582 goto out_unlock;
1583
1584 found = true;
1585 } for_each_vma_range(vmi, cur, end);
1586 BUG_ON(!found);
1587
1588 vma_iter_set(&vmi, start);
1589 prev = vma_prev(&vmi);
1590 if (vma->vm_start < start)
1591 prev = vma;
1592
1593 ret = 0;
1594 for_each_vma_range(vmi, vma, end) {
1595 cond_resched();
1596
1597 BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
1598
1599 /*
1600 * Nothing to do: this vma is already registered into this
1601 * userfaultfd and with the right tracking mode too.
1602 */
1603 if (!vma->vm_userfaultfd_ctx.ctx)
1604 goto skip;
1605
1606 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1607
1608 if (vma->vm_start > start)
1609 start = vma->vm_start;
1610 vma_end = min(end, vma->vm_end);
1611
1612 if (userfaultfd_missing(vma)) {
1613 /*
1614 * Wake any concurrent pending userfault while
1615 * we unregister, so they will not hang
1616 * permanently and it avoids userland to call
1617 * UFFDIO_WAKE explicitly.
1618 */
1619 struct userfaultfd_wake_range range;
1620 range.start = start;
1621 range.len = vma_end - start;
1622 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1623 }
1624
1625 /* Reset ptes for the whole vma range if wr-protected */
1626 if (userfaultfd_wp(vma))
1627 uffd_wp_range(vma, start, vma_end - start, false);
1628
1629 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1630 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1631 new_flags, NULL_VM_UFFD_CTX);
1632 if (IS_ERR(vma)) {
1633 ret = PTR_ERR(vma);
1634 break;
1635 }
1636
1637 /*
1638 * In the vma_merge() successful mprotect-like case 8:
1639 * the next vma was merged into the current one and
1640 * the current one has not been updated yet.
1641 */
1642 vma_start_write(vma);
1643 userfaultfd_set_vm_flags(vma, new_flags);
1644 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1645
1646 skip:
1647 prev = vma;
1648 start = vma->vm_end;
1649 }
1650
1651 out_unlock:
1652 mmap_write_unlock(mm);
1653 mmput(mm);
1654 out:
1655 return ret;
1656 }
1657
1658 /*
1659 * userfaultfd_wake may be used in combination with the
1660 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1661 */
1662 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1663 unsigned long arg)
1664 {
1665 int ret;
1666 struct uffdio_range uffdio_wake;
1667 struct userfaultfd_wake_range range;
1668 const void __user *buf = (void __user *)arg;
1669
1670 ret = -EFAULT;
1671 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1672 goto out;
1673
1674 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1675 if (ret)
1676 goto out;
1677
1678 range.start = uffdio_wake.start;
1679 range.len = uffdio_wake.len;
1680
1681 /*
1682 * len == 0 means wake all and we don't want to wake all here,
1683 * so check it again to be sure.
1684 */
1685 VM_BUG_ON(!range.len);
1686
1687 wake_userfault(ctx, &range);
1688 ret = 0;
1689
1690 out:
1691 return ret;
1692 }
1693
1694 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1695 unsigned long arg)
1696 {
1697 __s64 ret;
1698 struct uffdio_copy uffdio_copy;
1699 struct uffdio_copy __user *user_uffdio_copy;
1700 struct userfaultfd_wake_range range;
1701 uffd_flags_t flags = 0;
1702
1703 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1704
1705 ret = -EAGAIN;
1706 if (atomic_read(&ctx->mmap_changing))
1707 goto out;
1708
1709 ret = -EFAULT;
1710 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1711 /* don't copy "copy" last field */
1712 sizeof(uffdio_copy)-sizeof(__s64)))
1713 goto out;
1714
1715 ret = validate_unaligned_range(ctx->mm, uffdio_copy.src,
1716 uffdio_copy.len);
1717 if (ret)
1718 goto out;
1719 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1720 if (ret)
1721 goto out;
1722
1723 ret = -EINVAL;
1724 if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1725 goto out;
1726 if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1727 flags |= MFILL_ATOMIC_WP;
1728 if (mmget_not_zero(ctx->mm)) {
1729 ret = mfill_atomic_copy(ctx, uffdio_copy.dst, uffdio_copy.src,
1730 uffdio_copy.len, flags);
1731 mmput(ctx->mm);
1732 } else {
1733 return -ESRCH;
1734 }
1735 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1736 return -EFAULT;
1737 if (ret < 0)
1738 goto out;
1739 BUG_ON(!ret);
1740 /* len == 0 would wake all */
1741 range.len = ret;
1742 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1743 range.start = uffdio_copy.dst;
1744 wake_userfault(ctx, &range);
1745 }
1746 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1747 out:
1748 return ret;
1749 }
1750
1751 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1752 unsigned long arg)
1753 {
1754 __s64 ret;
1755 struct uffdio_zeropage uffdio_zeropage;
1756 struct uffdio_zeropage __user *user_uffdio_zeropage;
1757 struct userfaultfd_wake_range range;
1758
1759 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1760
1761 ret = -EAGAIN;
1762 if (atomic_read(&ctx->mmap_changing))
1763 goto out;
1764
1765 ret = -EFAULT;
1766 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1767 /* don't copy "zeropage" last field */
1768 sizeof(uffdio_zeropage)-sizeof(__s64)))
1769 goto out;
1770
1771 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1772 uffdio_zeropage.range.len);
1773 if (ret)
1774 goto out;
1775 ret = -EINVAL;
1776 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1777 goto out;
1778
1779 if (mmget_not_zero(ctx->mm)) {
1780 ret = mfill_atomic_zeropage(ctx, uffdio_zeropage.range.start,
1781 uffdio_zeropage.range.len);
1782 mmput(ctx->mm);
1783 } else {
1784 return -ESRCH;
1785 }
1786 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1787 return -EFAULT;
1788 if (ret < 0)
1789 goto out;
1790 /* len == 0 would wake all */
1791 BUG_ON(!ret);
1792 range.len = ret;
1793 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1794 range.start = uffdio_zeropage.range.start;
1795 wake_userfault(ctx, &range);
1796 }
1797 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1798 out:
1799 return ret;
1800 }
1801
1802 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1803 unsigned long arg)
1804 {
1805 int ret;
1806 struct uffdio_writeprotect uffdio_wp;
1807 struct uffdio_writeprotect __user *user_uffdio_wp;
1808 struct userfaultfd_wake_range range;
1809 bool mode_wp, mode_dontwake;
1810
1811 if (atomic_read(&ctx->mmap_changing))
1812 return -EAGAIN;
1813
1814 user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1815
1816 if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1817 sizeof(struct uffdio_writeprotect)))
1818 return -EFAULT;
1819
1820 ret = validate_range(ctx->mm, uffdio_wp.range.start,
1821 uffdio_wp.range.len);
1822 if (ret)
1823 return ret;
1824
1825 if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1826 UFFDIO_WRITEPROTECT_MODE_WP))
1827 return -EINVAL;
1828
1829 mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1830 mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1831
1832 if (mode_wp && mode_dontwake)
1833 return -EINVAL;
1834
1835 if (mmget_not_zero(ctx->mm)) {
1836 ret = mwriteprotect_range(ctx, uffdio_wp.range.start,
1837 uffdio_wp.range.len, mode_wp);
1838 mmput(ctx->mm);
1839 } else {
1840 return -ESRCH;
1841 }
1842
1843 if (ret)
1844 return ret;
1845
1846 if (!mode_wp && !mode_dontwake) {
1847 range.start = uffdio_wp.range.start;
1848 range.len = uffdio_wp.range.len;
1849 wake_userfault(ctx, &range);
1850 }
1851 return ret;
1852 }
1853
1854 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1855 {
1856 __s64 ret;
1857 struct uffdio_continue uffdio_continue;
1858 struct uffdio_continue __user *user_uffdio_continue;
1859 struct userfaultfd_wake_range range;
1860 uffd_flags_t flags = 0;
1861
1862 user_uffdio_continue = (struct uffdio_continue __user *)arg;
1863
1864 ret = -EAGAIN;
1865 if (atomic_read(&ctx->mmap_changing))
1866 goto out;
1867
1868 ret = -EFAULT;
1869 if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1870 /* don't copy the output fields */
1871 sizeof(uffdio_continue) - (sizeof(__s64))))
1872 goto out;
1873
1874 ret = validate_range(ctx->mm, uffdio_continue.range.start,
1875 uffdio_continue.range.len);
1876 if (ret)
1877 goto out;
1878
1879 ret = -EINVAL;
1880 if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1881 UFFDIO_CONTINUE_MODE_WP))
1882 goto out;
1883 if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1884 flags |= MFILL_ATOMIC_WP;
1885
1886 if (mmget_not_zero(ctx->mm)) {
1887 ret = mfill_atomic_continue(ctx, uffdio_continue.range.start,
1888 uffdio_continue.range.len, flags);
1889 mmput(ctx->mm);
1890 } else {
1891 return -ESRCH;
1892 }
1893
1894 if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1895 return -EFAULT;
1896 if (ret < 0)
1897 goto out;
1898
1899 /* len == 0 would wake all */
1900 BUG_ON(!ret);
1901 range.len = ret;
1902 if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1903 range.start = uffdio_continue.range.start;
1904 wake_userfault(ctx, &range);
1905 }
1906 ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1907
1908 out:
1909 return ret;
1910 }
1911
1912 static inline int userfaultfd_poison(struct userfaultfd_ctx *ctx, unsigned long arg)
1913 {
1914 __s64 ret;
1915 struct uffdio_poison uffdio_poison;
1916 struct uffdio_poison __user *user_uffdio_poison;
1917 struct userfaultfd_wake_range range;
1918
1919 user_uffdio_poison = (struct uffdio_poison __user *)arg;
1920
1921 ret = -EAGAIN;
1922 if (atomic_read(&ctx->mmap_changing))
1923 goto out;
1924
1925 ret = -EFAULT;
1926 if (copy_from_user(&uffdio_poison, user_uffdio_poison,
1927 /* don't copy the output fields */
1928 sizeof(uffdio_poison) - (sizeof(__s64))))
1929 goto out;
1930
1931 ret = validate_range(ctx->mm, uffdio_poison.range.start,
1932 uffdio_poison.range.len);
1933 if (ret)
1934 goto out;
1935
1936 ret = -EINVAL;
1937 if (uffdio_poison.mode & ~UFFDIO_POISON_MODE_DONTWAKE)
1938 goto out;
1939
1940 if (mmget_not_zero(ctx->mm)) {
1941 ret = mfill_atomic_poison(ctx, uffdio_poison.range.start,
1942 uffdio_poison.range.len, 0);
1943 mmput(ctx->mm);
1944 } else {
1945 return -ESRCH;
1946 }
1947
1948 if (unlikely(put_user(ret, &user_uffdio_poison->updated)))
1949 return -EFAULT;
1950 if (ret < 0)
1951 goto out;
1952
1953 /* len == 0 would wake all */
1954 BUG_ON(!ret);
1955 range.len = ret;
1956 if (!(uffdio_poison.mode & UFFDIO_POISON_MODE_DONTWAKE)) {
1957 range.start = uffdio_poison.range.start;
1958 wake_userfault(ctx, &range);
1959 }
1960 ret = range.len == uffdio_poison.range.len ? 0 : -EAGAIN;
1961
1962 out:
1963 return ret;
1964 }
1965
1966 bool userfaultfd_wp_async(struct vm_area_struct *vma)
1967 {
1968 return userfaultfd_wp_async_ctx(vma->vm_userfaultfd_ctx.ctx);
1969 }
1970
1971 static inline unsigned int uffd_ctx_features(__u64 user_features)
1972 {
1973 /*
1974 * For the current set of features the bits just coincide. Set
1975 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1976 */
1977 return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1978 }
1979
1980 static int userfaultfd_move(struct userfaultfd_ctx *ctx,
1981 unsigned long arg)
1982 {
1983 __s64 ret;
1984 struct uffdio_move uffdio_move;
1985 struct uffdio_move __user *user_uffdio_move;
1986 struct userfaultfd_wake_range range;
1987 struct mm_struct *mm = ctx->mm;
1988
1989 user_uffdio_move = (struct uffdio_move __user *) arg;
1990
1991 if (atomic_read(&ctx->mmap_changing))
1992 return -EAGAIN;
1993
1994 if (copy_from_user(&uffdio_move, user_uffdio_move,
1995 /* don't copy "move" last field */
1996 sizeof(uffdio_move)-sizeof(__s64)))
1997 return -EFAULT;
1998
1999 /* Do not allow cross-mm moves. */
2000 if (mm != current->mm)
2001 return -EINVAL;
2002
2003 ret = validate_range(mm, uffdio_move.dst, uffdio_move.len);
2004 if (ret)
2005 return ret;
2006
2007 ret = validate_range(mm, uffdio_move.src, uffdio_move.len);
2008 if (ret)
2009 return ret;
2010
2011 if (uffdio_move.mode & ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES|
2012 UFFDIO_MOVE_MODE_DONTWAKE))
2013 return -EINVAL;
2014
2015 if (mmget_not_zero(mm)) {
2016 ret = move_pages(ctx, uffdio_move.dst, uffdio_move.src,
2017 uffdio_move.len, uffdio_move.mode);
2018 mmput(mm);
2019 } else {
2020 return -ESRCH;
2021 }
2022
2023 if (unlikely(put_user(ret, &user_uffdio_move->move)))
2024 return -EFAULT;
2025 if (ret < 0)
2026 goto out;
2027
2028 /* len == 0 would wake all */
2029 VM_WARN_ON(!ret);
2030 range.len = ret;
2031 if (!(uffdio_move.mode & UFFDIO_MOVE_MODE_DONTWAKE)) {
2032 range.start = uffdio_move.dst;
2033 wake_userfault(ctx, &range);
2034 }
2035 ret = range.len == uffdio_move.len ? 0 : -EAGAIN;
2036
2037 out:
2038 return ret;
2039 }
2040
2041 /*
2042 * userland asks for a certain API version and we return which bits
2043 * and ioctl commands are implemented in this kernel for such API
2044 * version or -EINVAL if unknown.
2045 */
2046 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
2047 unsigned long arg)
2048 {
2049 struct uffdio_api uffdio_api;
2050 void __user *buf = (void __user *)arg;
2051 unsigned int ctx_features;
2052 int ret;
2053 __u64 features;
2054
2055 ret = -EFAULT;
2056 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
2057 goto out;
2058 features = uffdio_api.features;
2059 ret = -EINVAL;
2060 if (uffdio_api.api != UFFD_API)
2061 goto err_out;
2062 ret = -EPERM;
2063 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
2064 goto err_out;
2065
2066 /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2067 if (features & UFFD_FEATURE_WP_ASYNC)
2068 features |= UFFD_FEATURE_WP_UNPOPULATED;
2069
2070 /* report all available features and ioctls to userland */
2071 uffdio_api.features = UFFD_API_FEATURES;
2072 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2073 uffdio_api.features &=
2074 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
2075 #endif
2076 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2077 uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
2078 #endif
2079 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2080 uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
2081 uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
2082 uffdio_api.features &= ~UFFD_FEATURE_WP_ASYNC;
2083 #endif
2084
2085 ret = -EINVAL;
2086 if (features & ~uffdio_api.features)
2087 goto err_out;
2088
2089 uffdio_api.ioctls = UFFD_API_IOCTLS;
2090 ret = -EFAULT;
2091 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2092 goto out;
2093
2094 /* only enable the requested features for this uffd context */
2095 ctx_features = uffd_ctx_features(features);
2096 ret = -EINVAL;
2097 if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
2098 goto err_out;
2099
2100 ret = 0;
2101 out:
2102 return ret;
2103 err_out:
2104 memset(&uffdio_api, 0, sizeof(uffdio_api));
2105 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2106 ret = -EFAULT;
2107 goto out;
2108 }
2109
2110 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2111 unsigned long arg)
2112 {
2113 int ret = -EINVAL;
2114 struct userfaultfd_ctx *ctx = file->private_data;
2115
2116 if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2117 return -EINVAL;
2118
2119 switch(cmd) {
2120 case UFFDIO_API:
2121 ret = userfaultfd_api(ctx, arg);
2122 break;
2123 case UFFDIO_REGISTER:
2124 ret = userfaultfd_register(ctx, arg);
2125 break;
2126 case UFFDIO_UNREGISTER:
2127 ret = userfaultfd_unregister(ctx, arg);
2128 break;
2129 case UFFDIO_WAKE:
2130 ret = userfaultfd_wake(ctx, arg);
2131 break;
2132 case UFFDIO_COPY:
2133 ret = userfaultfd_copy(ctx, arg);
2134 break;
2135 case UFFDIO_ZEROPAGE:
2136 ret = userfaultfd_zeropage(ctx, arg);
2137 break;
2138 case UFFDIO_MOVE:
2139 ret = userfaultfd_move(ctx, arg);
2140 break;
2141 case UFFDIO_WRITEPROTECT:
2142 ret = userfaultfd_writeprotect(ctx, arg);
2143 break;
2144 case UFFDIO_CONTINUE:
2145 ret = userfaultfd_continue(ctx, arg);
2146 break;
2147 case UFFDIO_POISON:
2148 ret = userfaultfd_poison(ctx, arg);
2149 break;
2150 }
2151 return ret;
2152 }
2153
2154 #ifdef CONFIG_PROC_FS
2155 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2156 {
2157 struct userfaultfd_ctx *ctx = f->private_data;
2158 wait_queue_entry_t *wq;
2159 unsigned long pending = 0, total = 0;
2160
2161 spin_lock_irq(&ctx->fault_pending_wqh.lock);
2162 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2163 pending++;
2164 total++;
2165 }
2166 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2167 total++;
2168 }
2169 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2170
2171 /*
2172 * If more protocols will be added, there will be all shown
2173 * separated by a space. Like this:
2174 * protocols: aa:... bb:...
2175 */
2176 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2177 pending, total, UFFD_API, ctx->features,
2178 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2179 }
2180 #endif
2181
2182 static const struct file_operations userfaultfd_fops = {
2183 #ifdef CONFIG_PROC_FS
2184 .show_fdinfo = userfaultfd_show_fdinfo,
2185 #endif
2186 .release = userfaultfd_release,
2187 .poll = userfaultfd_poll,
2188 .read_iter = userfaultfd_read_iter,
2189 .unlocked_ioctl = userfaultfd_ioctl,
2190 .compat_ioctl = compat_ptr_ioctl,
2191 .llseek = noop_llseek,
2192 };
2193
2194 static void init_once_userfaultfd_ctx(void *mem)
2195 {
2196 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2197
2198 init_waitqueue_head(&ctx->fault_pending_wqh);
2199 init_waitqueue_head(&ctx->fault_wqh);
2200 init_waitqueue_head(&ctx->event_wqh);
2201 init_waitqueue_head(&ctx->fd_wqh);
2202 seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2203 }
2204
2205 static int new_userfaultfd(int flags)
2206 {
2207 struct userfaultfd_ctx *ctx;
2208 struct file *file;
2209 int fd;
2210
2211 BUG_ON(!current->mm);
2212
2213 /* Check the UFFD_* constants for consistency. */
2214 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2215 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2216 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2217
2218 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2219 return -EINVAL;
2220
2221 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2222 if (!ctx)
2223 return -ENOMEM;
2224
2225 refcount_set(&ctx->refcount, 1);
2226 ctx->flags = flags;
2227 ctx->features = 0;
2228 ctx->released = false;
2229 init_rwsem(&ctx->map_changing_lock);
2230 atomic_set(&ctx->mmap_changing, 0);
2231 ctx->mm = current->mm;
2232
2233 fd = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
2234 if (fd < 0)
2235 goto err_out;
2236
2237 /* Create a new inode so that the LSM can block the creation. */
2238 file = anon_inode_create_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
2239 O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2240 if (IS_ERR(file)) {
2241 put_unused_fd(fd);
2242 fd = PTR_ERR(file);
2243 goto err_out;
2244 }
2245 /* prevent the mm struct to be freed */
2246 mmgrab(ctx->mm);
2247 file->f_mode |= FMODE_NOWAIT;
2248 fd_install(fd, file);
2249 return fd;
2250 err_out:
2251 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2252 return fd;
2253 }
2254
2255 static inline bool userfaultfd_syscall_allowed(int flags)
2256 {
2257 /* Userspace-only page faults are always allowed */
2258 if (flags & UFFD_USER_MODE_ONLY)
2259 return true;
2260
2261 /*
2262 * The user is requesting a userfaultfd which can handle kernel faults.
2263 * Privileged users are always allowed to do this.
2264 */
2265 if (capable(CAP_SYS_PTRACE))
2266 return true;
2267
2268 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2269 return sysctl_unprivileged_userfaultfd;
2270 }
2271
2272 SYSCALL_DEFINE1(userfaultfd, int, flags)
2273 {
2274 if (!userfaultfd_syscall_allowed(flags))
2275 return -EPERM;
2276
2277 return new_userfaultfd(flags);
2278 }
2279
2280 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2281 {
2282 if (cmd != USERFAULTFD_IOC_NEW)
2283 return -EINVAL;
2284
2285 return new_userfaultfd(flags);
2286 }
2287
2288 static const struct file_operations userfaultfd_dev_fops = {
2289 .unlocked_ioctl = userfaultfd_dev_ioctl,
2290 .compat_ioctl = userfaultfd_dev_ioctl,
2291 .owner = THIS_MODULE,
2292 .llseek = noop_llseek,
2293 };
2294
2295 static struct miscdevice userfaultfd_misc = {
2296 .minor = MISC_DYNAMIC_MINOR,
2297 .name = "userfaultfd",
2298 .fops = &userfaultfd_dev_fops
2299 };
2300
2301 static int __init userfaultfd_init(void)
2302 {
2303 int ret;
2304
2305 ret = misc_register(&userfaultfd_misc);
2306 if (ret)
2307 return ret;
2308
2309 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2310 sizeof(struct userfaultfd_ctx),
2311 0,
2312 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2313 init_once_userfaultfd_ctx);
2314 #ifdef CONFIG_SYSCTL
2315 register_sysctl_init("vm", vm_userfaultfd_table);
2316 #endif
2317 return 0;
2318 }
2319 __initcall(userfaultfd_init);
2320
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.