1 // SPDX-License-Identifier: GPL-2.0-only 1
2 /*
3 * fs/userfaultfd.c
4 *
5 * Copyright (C) 2007 Davide Libenzi <davide
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
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 __r
37
38 #ifdef CONFIG_SYSCTL
39 static struct ctl_table vm_userfaultfd_table[]
40 {
41 .procname = "unprivilege
42 .data = &sysctl_unpr
43 .maxlen = sizeof(sysct
44 .mode = 0644,
45 .proc_handler = proc_dointve
46 .extra1 = SYSCTL_ZERO,
47 .extra2 = SYSCTL_ONE,
48 },
49 };
50 #endif
51
52 static struct kmem_cache *userfaultfd_ctx_cach
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
80 #define UFFD_FEATURE_INITIALIZED
81
82 static bool userfaultfd_is_initialized(struct
83 {
84 return ctx->features & UFFD_FEATURE_IN
85 }
86
87 static bool userfaultfd_wp_async_ctx(struct us
88 {
89 return ctx && (ctx->features & UFFD_FE
90 }
91
92 /*
93 * Whether WP_UNPOPULATED is enabled on the uf
94 * meaningful when userfaultfd_wp()==true on t
95 * anonymous.
96 */
97 bool userfaultfd_wp_unpopulated(struct vm_area
98 {
99 struct userfaultfd_ctx *ctx = vma->vm_
100
101 if (!ctx)
102 return false;
103
104 return ctx->features & UFFD_FEATURE_WP
105 }
106
107 static void userfaultfd_set_vm_flags(struct vm
108 vm_flags_
109 {
110 const bool uffd_wp_changed = (vma->vm_
111
112 vm_flags_reset(vma, flags);
113 /*
114 * For shared mappings, we want to ena
115 * userfaultfd-wp is enabled (see vma_
116 * recalculate vma->vm_page_prot whene
117 */
118 if ((vma->vm_flags & VM_SHARED) && uff
119 vma_set_page_prot(vma);
120 }
121
122 static int userfaultfd_wake_function(wait_queu
123 int wake_
124 {
125 struct userfaultfd_wake_range *range =
126 int ret;
127 struct userfaultfd_wait_queue *uwq;
128 unsigned long start, len;
129
130 uwq = container_of(wq, struct userfaul
131 ret = 0;
132 /* len == 0 means wake all */
133 start = range->start;
134 len = range->len;
135 if (len && (start > uwq->msg.arg.pagef
136 start + len <= uwq->msg.ar
137 goto out;
138 WRITE_ONCE(uwq->waken, true);
139 /*
140 * The Program-Order guarantees provid
141 * ensure uwq->waken is visible before
142 */
143 ret = wake_up_state(wq->private, mode)
144 if (ret) {
145 /*
146 * Wake only once, autoremove
147 *
148 * After the effect of list_de
149 * CPUs, the waitqueue may dis
150 * !list_empty_careful() in ha
151 *
152 * try_to_wake_up() has an imp
153 * wq->private is read before
154 * "wake_up_state" (which in t
155 */
156 list_del_init(&wq->entry);
157 }
158 out:
159 return ret;
160 }
161
162 /**
163 * userfaultfd_ctx_get - Acquires a reference
164 * context.
165 * @ctx: [in] Pointer to the userfaultfd conte
166 */
167 static void userfaultfd_ctx_get(struct userfau
168 {
169 refcount_inc(&ctx->refcount);
170 }
171
172 /**
173 * userfaultfd_ctx_put - Releases a reference
174 * context.
175 * @ctx: [in] Pointer to userfaultfd context.
176 *
177 * The userfaultfd context reference must have
178 * with userfaultfd_ctx_get() or userfaultfd_c
179 */
180 static void userfaultfd_ctx_put(struct userfau
181 {
182 if (refcount_dec_and_test(&ctx->refcou
183 VM_BUG_ON(spin_is_locked(&ctx-
184 VM_BUG_ON(waitqueue_active(&ct
185 VM_BUG_ON(spin_is_locked(&ctx-
186 VM_BUG_ON(waitqueue_active(&ct
187 VM_BUG_ON(spin_is_locked(&ctx-
188 VM_BUG_ON(waitqueue_active(&ct
189 VM_BUG_ON(spin_is_locked(&ctx-
190 VM_BUG_ON(waitqueue_active(&ct
191 mmdrop(ctx->mm);
192 kmem_cache_free(userfaultfd_ct
193 }
194 }
195
196 static inline void msg_init(struct uffd_msg *m
197 {
198 BUILD_BUG_ON(sizeof(struct uffd_msg) !
199 /*
200 * Must use memset to zero out the pad
201 * leaked to userland.
202 */
203 memset(msg, 0, sizeof(struct uffd_msg)
204 }
205
206 static inline struct uffd_msg userfault_msg(un
207 un
208 un
209 un
210 un
211 {
212 struct uffd_msg msg;
213
214 msg_init(&msg);
215 msg.event = UFFD_EVENT_PAGEFAULT;
216
217 msg.arg.pagefault.address = (features
218 real_addre
219
220 /*
221 * These flags indicate why the userfa
222 * - UFFD_PAGEFAULT_FLAG_WP indicates
223 * - UFFD_PAGEFAULT_FLAG_MINOR indicat
224 * - Neither of these flags being set
225 *
226 * Separately, UFFD_PAGEFAULT_FLAG_WRI
227 * fault. Otherwise, it was a read fau
228 */
229 if (flags & FAULT_FLAG_WRITE)
230 msg.arg.pagefault.flags |= UFF
231 if (reason & VM_UFFD_WP)
232 msg.arg.pagefault.flags |= UFF
233 if (reason & VM_UFFD_MINOR)
234 msg.arg.pagefault.flags |= UFF
235 if (features & UFFD_FEATURE_THREAD_ID)
236 msg.arg.pagefault.feat.ptid =
237 return msg;
238 }
239
240 #ifdef CONFIG_HUGETLB_PAGE
241 /*
242 * Same functionality as userfaultfd_must_wait
243 * hugepmd ranges.
244 */
245 static inline bool userfaultfd_huge_must_wait(
246
247
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,
256 if (!ptep)
257 goto out;
258
259 ret = false;
260 pte = huge_ptep_get(vma->vm_mm, vmf->a
261
262 /*
263 * Lockless access: we're in a wait_ev
264 * changes under us. PTE markers shou
265 * ptes here.
266 */
267 if (huge_pte_none_mostly(pte))
268 ret = true;
269 if (!huge_pte_write(pte) && (reason &
270 ret = true;
271 out:
272 return ret;
273 }
274 #else
275 static inline bool userfaultfd_huge_must_wait(
276
277
278 {
279 return false; /* should never get he
280 }
281 #endif /* CONFIG_HUGETLB_PAGE */
282
283 /*
284 * Verify the pagetables are still not ok afte
285 * the fault_pending_wqh to avoid userland hav
286 * userfault that has already been resolved, i
287 * UFFDIO_COPY|ZEROPAGE are being run simultan
288 * threads.
289 */
290 static inline bool userfaultfd_must_wait(struc
291 struc
292 unsig
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(_
323 goto out;
324
325 if (pmd_trans_huge(_pmd)) {
326 if (!pmd_write(_pmd) && (reaso
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_ev
338 * changes under us. PTE markers shou
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_
345 ret = true;
346 pte_unmap(pte);
347
348 out:
349 return ret;
350 }
351
352 static inline unsigned int userfaultfd_get_blo
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_
365 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NO
366 * FAULT_FLAG_KILLABLE are not straightforward
367 * recommendation in __lock_page_or_retry is n
368 *
369 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_
370 * before returning VM_FAULT_RETRY only if FAU
371 * not set.
372 *
373 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_
374 * set, VM_FAULT_RETRY can still be returned i
375 * fatal_signal_pending()s, and the mmap_lock
376 * returning it.
377 */
378 vm_fault_t handle_userfault(struct vm_fault *v
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
390 *
391 * We also don't do userfault handling
392 * coredumping. hugetlbfs has the spec
393 * hugetlb_follow_page_mask() to skip
394 * FOLL_DUMP case, anon memory also ch
395 * the no_page_table() helper in follo
396 * shmem_vm_ops->fault method is invok
397 * coredumping and it ends up here.
398 */
399 if (current->flags & (PF_EXITING|PF_DU
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 exp
413 VM_BUG_ON(!reason || (reason & (reason
414
415 if (ctx->features & UFFD_FEATURE_SIGBU
416 goto out;
417 if (!(vmf->flags & FAULT_FLAG_USER) &&
418 goto out;
419
420 /*
421 * If it's already released don't get
422 * in __get_user_pages if userfaultfd_
423 * caller of handle_userfault to relea
424 */
425 if (unlikely(READ_ONCE(ctx->released))
426 /*
427 * Don't return VM_FAULT_SIGBU
428 * cooperative manager can clo
429 * last UFFDIO_COPY, without r
430 * involuntary SIGBUS if the p
431 * userfaultfd while the userf
432 * (but after the last UFFDIO_
433 * wasn't already closed when
434 * this point, that would norm
435 * userfaultfd_must_wait retur
436 *
437 * If we were to return VM_FAU
438 * cooperative manager would b
439 * always call UFFDIO_UNREGIST
440 * close the uffd.
441 */
442 ret = VM_FAULT_NOPAGE;
443 goto out;
444 }
445
446 /*
447 * Check that we can return VM_FAULT_R
448 *
449 * NOTE: it should become possible to
450 * even if FAULT_FLAG_TRIED is set wit
451 * -EBUSY failures, if the userfaultfd
452 * VM_UFFD_WP tracking and we intend t
453 * without first stopping userland acc
454 * VM_UFFD_MISSING userfaults this is
455 */
456 if (unlikely(!(vmf->flags & FAULT_FLAG
457 /*
458 * Validate the invariant that
459 * to be sure not to return SI
460 * nowait invocations.
461 */
462 BUG_ON(vmf->flags & FAULT_FLAG
463 #ifdef CONFIG_DEBUG_VM
464 if (printk_ratelimit()) {
465 printk(KERN_WARNING
466 "FAULT_FLAG_ALL
467 vmf->flags);
468 dump_stack();
469 }
470 #endif
471 goto out;
472 }
473
474 /*
475 * Handle nowait, not much to do other
476 * and wait.
477 */
478 ret = VM_FAULT_RETRY;
479 if (vmf->flags & FAULT_FLAG_RETRY_NOWA
480 goto out;
481
482 /* take the reference before dropping
483 userfaultfd_ctx_get(ctx);
484
485 init_waitqueue_func_entry(&uwq.wq, use
486 uwq.wq.private = current;
487 uwq.msg = userfault_msg(vmf->address,
488 reason, ctx->f
489 uwq.ctx = ctx;
490 uwq.waken = false;
491
492 blocking_state = userfaultfd_get_block
493
494 /*
495 * Take the vma lock now, in order to
496 * userfaultfd_huge_must_wait() later.
497 * (sleepable) vma lock can modify the
498 * must be before explicitly calling s
499 */
500 if (is_vm_hugetlb_page(vma))
501 hugetlb_vma_lock_read(vma);
502
503 spin_lock_irq(&ctx->fault_pending_wqh.
504 /*
505 * After the __add_wait_queue the uwq
506 * through poll/read().
507 */
508 __add_wait_queue(&ctx->fault_pending_w
509 /*
510 * The smp_mb() after __set_current_st
511 * following the spin_unlock to happen
512 * __add_wait_queue.
513 */
514 set_current_state(blocking_state);
515 spin_unlock_irq(&ctx->fault_pending_wq
516
517 if (!is_vm_hugetlb_page(vma))
518 must_wait = userfaultfd_must_w
519 else
520 must_wait = userfaultfd_huge_m
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
526 wake_up_poll(&ctx->fd_wqh, EPO
527 schedule();
528 }
529
530 __set_current_state(TASK_RUNNING);
531
532 /*
533 * Here we race with the list_del; lis
534 * userfaultfd_ctx_read(), however bec
535 * list_del_init() to refile across th
536 * and next pointers will never point
537 * would never let any of the two poin
538 * self. So list_empty_careful won't r
539 * pointing to self at any time during
540 * only case where list_del_init() is
541 * removal in the wake function and th
542 * and it's fine not to block on the s
543 * kernel stack can be released after
544 */
545 if (!list_empty_careful(&uwq.wq.entry)
546 spin_lock_irq(&ctx->fault_pend
547 /*
548 * No need of list_del_init(),
549 * will be freed shortly anywa
550 */
551 list_del(&uwq.wq.entry);
552 spin_unlock_irq(&ctx->fault_pe
553 }
554
555 /*
556 * ctx may go away after this if the u
557 * already released.
558 */
559 userfaultfd_ctx_put(ctx);
560
561 out:
562 return ret;
563 }
564
565 static void userfaultfd_event_wait_completion(
566
567 {
568 struct userfaultfd_ctx *release_new_ct
569
570 if (WARN_ON_ONCE(current->flags & PF_E
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
580 * through poll/read().
581 */
582 __add_wait_queue(&ctx->event_wqh, &ewq
583 for (;;) {
584 set_current_state(TASK_KILLABL
585 if (ewq->msg.event == 0)
586 break;
587 if (READ_ONCE(ctx->released) |
588 fatal_signal_pending(curre
589 /*
590 * &ewq->wq may be que
591 * __remove_wait_queue
592 * parameter. It would
593 * didn't.
594 */
595 __remove_wait_queue(&c
596 if (ewq->msg.event ==
597 struct userfau
598
599 new = (struct
600 (unsig
601 ewq->m
602 release_new_ct
603 }
604 break;
605 }
606
607 spin_unlock_irq(&ctx->event_wq
608
609 wake_up_poll(&ctx->fd_wqh, EPO
610 schedule();
611
612 spin_lock_irq(&ctx->event_wqh.
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
620 VMA_ITERATOR(vmi, mm, 0);
621
622 /* the various vma->vm_userfau
623 mmap_write_lock(mm);
624 for_each_vma(vmi, vma) {
625 if (vma->vm_userfaultf
626 vma_start_writ
627 vma->vm_userfa
628 userfaultfd_se
629
630 }
631 }
632 mmap_write_unlock(mm);
633
634 userfaultfd_ctx_put(release_ne
635 }
636
637 /*
638 * ctx may go away after this if the u
639 * already released.
640 */
641 out:
642 atomic_dec(&ctx->mmap_changing);
643 VM_BUG_ON(atomic_read(&ctx->mmap_chang
644 userfaultfd_ctx_put(ctx);
645 }
646
647 static void userfaultfd_event_complete(struct
648 struct
649 {
650 ewq->msg.event = 0;
651 wake_up_locked(&ctx->event_wqh);
652 __remove_wait_queue(&ctx->event_wqh, &
653 }
654
655 int dup_userfaultfd(struct vm_area_struct *vma
656 {
657 struct userfaultfd_ctx *ctx = NULL, *o
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_EV
665 vma_start_write(vma);
666 vma->vm_userfaultfd_ctx = NULL
667 userfaultfd_set_vm_flags(vma,
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),
679 if (!fctx)
680 return -ENOMEM;
681
682 ctx = kmem_cache_alloc(userfau
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_
693 atomic_set(&ctx->mmap_changing
694 ctx->mm = vma->vm_mm;
695 mmgrab(ctx->mm);
696
697 userfaultfd_ctx_get(octx);
698 down_write(&octx->map_changing
699 atomic_inc(&octx->mmap_changin
700 up_write(&octx->map_changing_l
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_c
711 {
712 struct userfaultfd_ctx *ctx = fctx->or
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 = (unsi
719
720 userfaultfd_event_wait_completion(ctx,
721 }
722
723 void dup_userfaultfd_complete(struct list_head
724 {
725 struct userfaultfd_fork_ctx *fctx, *n;
726
727 list_for_each_entry_safe(fctx, n, fcs,
728 dup_fctx(fctx);
729 list_del(&fctx->list);
730 kfree(fctx);
731 }
732 }
733
734 void mremap_userfaultfd_prep(struct vm_area_st
735 struct vm_userfau
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
745 vm_ctx->ctx = ctx;
746 userfaultfd_ctx_get(ctx);
747 down_write(&ctx->map_changing_
748 atomic_inc(&ctx->mmap_changing
749 up_write(&ctx->map_changing_lo
750 } else {
751 /* Drop uffd context if remap
752 vma_start_write(vma);
753 vma->vm_userfaultfd_ctx = NULL
754 userfaultfd_set_vm_flags(vma,
755 }
756 }
757
758 void mremap_userfaultfd_complete(struct vm_use
759 unsigned long
760 unsigned long
761 {
762 struct userfaultfd_ctx *ctx = vm_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,
781 }
782
783 bool userfaultfd_remove(struct vm_area_struct
784 unsigned long start, u
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_FEA
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,
807
808 return false;
809 }
810
811 static bool has_unmap_ctx(struct userfaultfd_c
812 unsigned long start,
813 {
814 struct userfaultfd_unmap_ctx *unmap_ct
815
816 list_for_each_entry(unmap_ctx, unmaps,
817 if (unmap_ctx->ctx == ctx && u
818 unmap_ctx->end == end)
819 return true;
820
821 return false;
822 }
823
824 int userfaultfd_unmap_prep(struct vm_area_stru
825 unsigned long end,
826 {
827 struct userfaultfd_unmap_ctx *unmap_ct
828 struct userfaultfd_ctx *ctx = vma->vm_
829
830 if (!ctx || !(ctx->features & UFFD_FEA
831 has_unmap_ctx(ctx, unmaps, start,
832 return 0;
833
834 unmap_ctx = kzalloc(sizeof(*unmap_ctx)
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_stru
851 {
852 struct userfaultfd_unmap_ctx *ctx, *n;
853 struct userfaultfd_wait_queue ewq;
854
855 list_for_each_entry_safe(ctx, n, uf, l
856 msg_init(&ewq.msg);
857
858 ewq.msg.event = UFFD_EVENT_UNM
859 ewq.msg.arg.remove.start = ctx
860 ewq.msg.arg.remove.end = ctx->
861
862 userfaultfd_event_wait_complet
863
864 list_del(&ctx->list);
865 kfree(ctx);
866 }
867 }
868
869 static int userfaultfd_release(struct inode *i
870 {
871 struct userfaultfd_ctx *ctx = file->pr
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 =
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.
886 * must be retried without returning V
887 * userfaultfd_ctx_get() succeeds but
888 * changes while handle_userfault rele
889 * it's critical that released is set
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_c
897 !!(vma->vm_flags & __VM
898 if (vma->vm_userfaultfd_ctx.ct
899 prev = vma;
900 continue;
901 }
902 /* Reset ptes for the whole vm
903 if (userfaultfd_wp(vma))
904 uffd_wp_range(vma, vma
905 vma->vm_
906 new_flags = vma->vm_flags & ~_
907 vma = vma_modify_flags_uffd(&v
908 vm
909 NU
910
911 vma_start_write(vma);
912 userfaultfd_set_vm_flags(vma,
913 vma->vm_userfaultfd_ctx = NULL
914
915 prev = vma;
916 }
917 mmap_write_unlock(mm);
918 mmput(mm);
919 wakeup:
920 /*
921 * After no new page faults can wait o
922 * the last page faults that may have
923 * the fault_*wqh.
924 */
925 spin_lock_irq(&ctx->fault_pending_wqh.
926 __wake_up_locked_key(&ctx->fault_pendi
927 __wake_up(&ctx->fault_wqh, TASK_NORMAL
928 spin_unlock_irq(&ctx->fault_pending_wq
929
930 /* Flush pending events that may still
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
939 static inline struct userfaultfd_wait_queue *f
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 beh
951 wq = list_last_entry(&wqh->head, typeo
952 uwq = container_of(wq, struct userfaul
953 out:
954 return uwq;
955 }
956
957 static inline struct userfaultfd_wait_queue *f
958 struct userfaultfd_ctx *ctx)
959 {
960 return find_userfault_in(&ctx->fault_p
961 }
962
963 static inline struct userfaultfd_wait_queue *f
964 struct userfaultfd_ctx *ctx)
965 {
966 return find_userfault_in(&ctx->event_w
967 }
968
969 static __poll_t userfaultfd_poll(struct file *
970 {
971 struct userfaultfd_ctx *ctx = file->pr
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 w
981 * userfaults can be waken before they
982 */
983 if (unlikely(!(file->f_flags & O_NONBL
984 return EPOLLERR;
985 /*
986 * lockless access to see if there are
987 * __pollwait last action is the add_w
988 * the spin_unlock would allow the wai
989 * pass above the actual list_add insi
990 * add_wait_queue critical section. So
991 * memory barrier to serialize the lis
992 * add_wait_queue() with the waitqueue
993 * below.
994 */
995 ret = 0;
996 smp_mb();
997 if (waitqueue_active(&ctx->fault_pendi
998 ret = EPOLLIN;
999 else if (waitqueue_active(&ctx->event_
1000 ret = EPOLLIN;
1001
1002 return ret;
1003 }
1004
1005 static const struct file_operations userfault
1006
1007 static int resolve_userfault_fork(struct user
1008 struct inod
1009 struct uffd
1010 {
1011 int fd;
1012
1013 fd = anon_inode_create_getfd("[userfa
1014 O_RDONLY | (new->flag
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 us
1024 struct uf
1025 {
1026 ssize_t ret;
1027 DECLARE_WAITQUEUE(wait, current);
1028 struct userfaultfd_wait_queue *uwq;
1029 /*
1030 * Handling fork event requires sleep
1031 * we drop the event_wqh lock, then d
1032 * lock it back and wake up the waite
1033 * dropped the ewq may go away so we
1034 * carefully.
1035 */
1036 LIST_HEAD(fork_event);
1037 struct userfaultfd_ctx *fork_nctx = N
1038
1039 /* always take the fd_wqh lock before
1040 spin_lock_irq(&ctx->fd_wqh.lock);
1041 __add_wait_queue(&ctx->fd_wqh, &wait)
1042 for (;;) {
1043 set_current_state(TASK_INTERR
1044 spin_lock(&ctx->fault_pending
1045 uwq = find_userfault(ctx);
1046 if (uwq) {
1047 /*
1048 * Use a seqcount to
1049 * in wake_userfault(
1050 * wakeups because du
1051 * waitqueue could be
1052 * only userfault.
1053 */
1054 write_seqcount_begin(
1055
1056 /*
1057 * The fault_pending_
1058 * to disappear from
1059 *
1060 * Refile this userfa
1061 * fault_pending_wqh
1062 * pending anymore af
1063 *
1064 * Use list_del() by
1065 * userfaultfd_wake_f
1066 * list_del_init() by
1067 * changes __remove_w
1068 * list_del_init() in
1069 * !list_empty_carefu
1070 * handle_userfault()
1071 * must never be empt
1072 * refile, or the wai
1073 * from under us. The
1074 * parameter of __rem
1075 * anyway.
1076 */
1077 list_del(&uwq->wq.ent
1078 add_wait_queue(&ctx->
1079
1080 write_seqcount_end(&c
1081
1082 /* careful to always
1083 *msg = uwq->msg;
1084 spin_unlock(&ctx->fau
1085 ret = 0;
1086 break;
1087 }
1088 spin_unlock(&ctx->fault_pendi
1089
1090 spin_lock(&ctx->event_wqh.loc
1091 uwq = find_userfault_evt(ctx)
1092 if (uwq) {
1093 *msg = uwq->msg;
1094
1095 if (uwq->msg.event ==
1096 fork_nctx = (
1097 (unsi
1098 uwq->
1099 list_move(&uw
1100 /*
1101 * fork_nctx
1102 * we drop th
1103 * reference
1104 */
1105 userfaultfd_c
1106 spin_unlock(&
1107 ret = 0;
1108 break;
1109 }
1110
1111 userfaultfd_event_com
1112 spin_unlock(&ctx->eve
1113 ret = 0;
1114 break;
1115 }
1116 spin_unlock(&ctx->event_wqh.l
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.
1127 schedule();
1128 spin_lock_irq(&ctx->fd_wqh.lo
1129 }
1130 __remove_wait_queue(&ctx->fd_wqh, &wa
1131 __set_current_state(TASK_RUNNING);
1132 spin_unlock_irq(&ctx->fd_wqh.lock);
1133
1134 if (!ret && msg->event == UFFD_EVENT_
1135 ret = resolve_userfault_fork(
1136 spin_lock_irq(&ctx->event_wqh
1137 if (!list_empty(&fork_event))
1138 /*
1139 * The fork thread di
1140 * drop the temporary
1141 */
1142 userfaultfd_ctx_put(f
1143
1144 uwq = list_first_entr
1145
1146
1147 /*
1148 * If fork_event list
1149 * the event wasn't a
1150 * (the event is allo
1151 * stack), put the ev
1152 * the event_wq. fork
1153 * as soon as we retu
1154 * stay queued there
1155 * "ret" value.
1156 */
1157 list_del(&uwq->wq.ent
1158 __add_wait_queue(&ctx
1159
1160 /*
1161 * Leave the event in
1162 * error to userland
1163 * the userfault fork
1164 */
1165 if (likely(!ret))
1166 userfaultfd_e
1167 } else {
1168 /*
1169 * Here the fork thre
1170 * refcount from the
1171 * has already been r
1172 * the reference we t
1173 * lock above. If res
1174 * failed we've to dr
1175 * fork_nctx has to b
1176 * it succeeded we'll
1177 * uffd references it
1178 */
1179 if (ret)
1180 userfaultfd_c
1181 }
1182 spin_unlock_irq(&ctx->event_w
1183 }
1184
1185 return ret;
1186 }
1187
1188 static ssize_t userfaultfd_read_iter(struct k
1189 {
1190 struct file *file = iocb->ki_filp;
1191 struct userfaultfd_ctx *ctx = file->p
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
1201 for (;;) {
1202 if (iov_iter_count(to) < size
1203 return ret ? ret : -E
1204 _ret = userfaultfd_ctx_read(c
1205 if (_ret < 0)
1206 return ret ? ret : _r
1207 _ret = !copy_to_iter_full(&ms
1208 if (_ret)
1209 return ret ? ret : -E
1210 ret += sizeof(msg);
1211 /*
1212 * Allow to read more than on
1213 * block if waiting for the v
1214 */
1215 no_wait = true;
1216 }
1217 }
1218
1219 static void __wake_userfault(struct userfault
1220 struct userfault
1221 {
1222 spin_lock_irq(&ctx->fault_pending_wqh
1223 /* wake all in the range and autoremo
1224 if (waitqueue_active(&ctx->fault_pend
1225 __wake_up_locked_key(&ctx->fa
1226 range);
1227 if (waitqueue_active(&ctx->fault_wqh)
1228 __wake_up(&ctx->fault_wqh, TA
1229 spin_unlock_irq(&ctx->fault_pending_w
1230 }
1231
1232 static __always_inline void wake_userfault(st
1233 st
1234 {
1235 unsigned seq;
1236 bool need_wakeup;
1237
1238 /*
1239 * To be sure waitqueue_active() is n
1240 * before the pagetable update, use a
1241 * barrier here. PT lock release or m
1242 * have release semantics that can al
1243 * waitqueue_active() to be reordered
1244 */
1245 smp_mb();
1246
1247 /*
1248 * Use waitqueue_active because it's
1249 * change the address space atomicall
1250 * userfaults yet. So we take the spi
1251 * sure we've userfaults to wake.
1252 */
1253 do {
1254 seq = read_seqcount_begin(&ct
1255 need_wakeup = waitqueue_activ
1256 waitqueue_active(&ctx
1257 cond_resched();
1258 } while (read_seqcount_retry(&ctx->re
1259 if (need_wakeup)
1260 __wake_userfault(ctx, range);
1261 }
1262
1263 static __always_inline int validate_unaligned
1264 struct mm_struct *mm, __u64 start, __
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(str
1284 __u
1285 {
1286 if (start & ~PAGE_MASK)
1287 return -EINVAL;
1288
1289 return validate_unaligned_range(mm, s
1290 }
1291
1292 static int userfaultfd_register(struct userfa
1293 unsigned long
1294 {
1295 struct mm_struct *mm = ctx->mm;
1296 struct vm_area_struct *vma, *prev, *c
1297 int ret;
1298 struct uffdio_register uffdio_registe
1299 struct uffdio_register __user *user_u
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_
1306
1307 user_uffdio_register = (struct uffdio
1308
1309 ret = -EFAULT;
1310 if (copy_from_user(&uffdio_register,
1311 sizeof(uffdio_regi
1312 goto out;
1313
1314 ret = -EINVAL;
1315 if (!uffdio_register.mode)
1316 goto out;
1317 if (uffdio_register.mode & ~UFFD_API_
1318 goto out;
1319 vm_flags = 0;
1320 if (uffdio_register.mode & UFFDIO_REG
1321 vm_flags |= VM_UFFD_MISSING;
1322 if (uffdio_register.mode & UFFDIO_REG
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_REG
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_regis
1336 uffdio_register.
1337 if (ret)
1338 goto out;
1339
1340 start = uffdio_register.range.start;
1341 end = start + uffdio_register.range.l
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 pag
1356 * is aligned to huge page size.
1357 */
1358 if (is_vm_hugetlb_page(vma)) {
1359 unsigned long vma_hpagesize =
1360
1361 if (start & (vma_hpagesize -
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_
1375 !!(cur->vm_flags & __V
1376
1377 /* check not compatible vmas
1378 ret = -EINVAL;
1379 if (!vma_can_userfault(cur, v
1380 goto out_unlock;
1381
1382 /*
1383 * UFFDIO_COPY will fill file
1384 * PROT_WRITE. This check enf
1385 * MAP_SHARED, the process ha
1386 * file. If VM_MAYWRITE is se
1387 * MAP_SHARED vma: there is n
1388 * F_WRITE_SEAL can be taken
1389 */
1390 ret = -EPERM;
1391 if (unlikely(!(cur->vm_flags
1392 goto out_unlock;
1393
1394 /*
1395 * If this vma contains endin
1396 * check alignment.
1397 */
1398 if (is_vm_hugetlb_page(cur) &
1399 end > cur->vm_start) {
1400 unsigned long vma_hpa
1401
1402 ret = -EINVAL;
1403
1404 if (end & (vma_hpages
1405 goto out_unlo
1406 }
1407 if ((vm_flags & VM_UFFD_WP) &
1408 goto out_unlock;
1409
1410 /*
1411 * Check that this vma isn't
1412 * different userfaultfd. We
1413 * userfaultfd to own a singl
1414 * wouldn't know which one to
1415 */
1416 ret = -EBUSY;
1417 if (cur->vm_userfaultfd_ctx.c
1418 cur->vm_userfaultfd_ctx.c
1419 goto out_unlock;
1420
1421 /*
1422 * Note vmas containing huge
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
1441 BUG_ON(vma->vm_userfaultfd_ct
1442 vma->vm_userfaultfd_ct
1443 WARN_ON(!(vma->vm_flags & VM_
1444
1445 /*
1446 * Nothing to do: this vma is
1447 * userfaultfd and with the r
1448 */
1449 if (vma->vm_userfaultfd_ctx.c
1450 (vma->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_en
1456
1457 new_flags = (vma->vm_flags &
1458 vma = vma_modify_flags_uffd(&
1459 n
1460 (
1461 if (IS_ERR(vma)) {
1462 ret = PTR_ERR(vma);
1463 break;
1464 }
1465
1466 /*
1467 * In the vma_merge() success
1468 * the next vma was merged in
1469 * the current one has not be
1470 */
1471 vma_start_write(vma);
1472 userfaultfd_set_vm_flags(vma,
1473 vma->vm_userfaultfd_ctx.ctx =
1474
1475 if (is_vm_hugetlb_page(vma) &
1476 hugetlb_unshare_all_p
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 ? U
1490 UFFD_API_RANGE_IOCTLS;
1491
1492 /*
1493 * Declare the WP ioctl only
1494 * specified and all checks p
1495 */
1496 if (!(uffdio_register.mode &
1497 ioctls_out &= ~((__u6
1498
1499 /* CONTINUE ioctl is only sup
1500 if (!(uffdio_register.mode &
1501 ioctls_out &= ~((__u6
1502
1503 /*
1504 * Now that we scanned all vm
1505 * userland which ioctls meth
1506 * succeed on this range.
1507 */
1508 if (put_user(ioctls_out, &use
1509 ret = -EFAULT;
1510 }
1511 out:
1512 return ret;
1513 }
1514
1515 static int userfaultfd_unregister(struct user
1516 unsigned lo
1517 {
1518 struct mm_struct *mm = ctx->mm;
1519 struct vm_area_struct *vma, *prev, *c
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
1526 struct vma_iterator vmi;
1527 bool wp_async = userfaultfd_wp_async_
1528
1529 ret = -EFAULT;
1530 if (copy_from_user(&uffdio_unregister
1531 goto out;
1532
1533 ret = validate_range(mm, uffdio_unreg
1534 uffdio_unregiste
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 pag
1554 * is aligned to huge page size.
1555 */
1556 if (is_vm_hugetlb_page(vma)) {
1557 unsigned long vma_hpagesize =
1558
1559 if (start & (vma_hpagesize -
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_
1572 !!(cur->vm_flags & __V
1573
1574 /*
1575 * Check not compatible vmas,
1576 * here as not compatible vma
1577 * userfaultfd_ctx registered
1578 * provides for more strict b
1579 * unregistration errors.
1580 */
1581 if (!vma_can_userfault(cur, c
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
1598
1599 /*
1600 * Nothing to do: this vma is
1601 * userfaultfd and with the r
1602 */
1603 if (!vma->vm_userfaultfd_ctx.
1604 goto skip;
1605
1606 WARN_ON(!(vma->vm_flags & VM_
1607
1608 if (vma->vm_start > start)
1609 start = vma->vm_start
1610 vma_end = min(end, vma->vm_en
1611
1612 if (userfaultfd_missing(vma))
1613 /*
1614 * Wake any concurren
1615 * we unregister, so
1616 * permanently and it
1617 * UFFDIO_WAKE explic
1618 */
1619 struct userfaultfd_wa
1620 range.start = start;
1621 range.len = vma_end -
1622 wake_userfault(vma->v
1623 }
1624
1625 /* Reset ptes for the whole v
1626 if (userfaultfd_wp(vma))
1627 uffd_wp_range(vma, st
1628
1629 new_flags = vma->vm_flags & ~
1630 vma = vma_modify_flags_uffd(&
1631 n
1632 if (IS_ERR(vma)) {
1633 ret = PTR_ERR(vma);
1634 break;
1635 }
1636
1637 /*
1638 * In the vma_merge() success
1639 * the next vma was merged in
1640 * the current one has not be
1641 */
1642 vma_start_write(vma);
1643 userfaultfd_set_vm_flags(vma,
1644 vma->vm_userfaultfd_ctx = NUL
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 combinatio
1660 * UFFDIO_*_MODE_DONTWAKE to wakeup userfault
1661 */
1662 static int userfaultfd_wake(struct userfaultf
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
1669
1670 ret = -EFAULT;
1671 if (copy_from_user(&uffdio_wake, buf,
1672 goto out;
1673
1674 ret = validate_range(ctx->mm, uffdio_
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
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 userfaultf
1695 unsigned long arg
1696 {
1697 __s64 ret;
1698 struct uffdio_copy uffdio_copy;
1699 struct uffdio_copy __user *user_uffdi
1700 struct userfaultfd_wake_range range;
1701 uffd_flags_t flags = 0;
1702
1703 user_uffdio_copy = (struct uffdio_cop
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
1711 /* don't copy "cop
1712 sizeof(uffdio_copy
1713 goto out;
1714
1715 ret = validate_unaligned_range(ctx->m
1716 uffdio
1717 if (ret)
1718 goto out;
1719 ret = validate_range(ctx->mm, uffdio_
1720 if (ret)
1721 goto out;
1722
1723 ret = -EINVAL;
1724 if (uffdio_copy.mode & ~(UFFDIO_COPY_
1725 goto out;
1726 if (uffdio_copy.mode & UFFDIO_COPY_MO
1727 flags |= MFILL_ATOMIC_WP;
1728 if (mmget_not_zero(ctx->mm)) {
1729 ret = mfill_atomic_copy(ctx,
1730 uffdi
1731 mmput(ctx->mm);
1732 } else {
1733 return -ESRCH;
1734 }
1735 if (unlikely(put_user(ret, &user_uffd
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_
1743 range.start = uffdio_copy.dst
1744 wake_userfault(ctx, &range);
1745 }
1746 ret = range.len == uffdio_copy.len ?
1747 out:
1748 return ret;
1749 }
1750
1751 static int userfaultfd_zeropage(struct userfa
1752 unsigned long
1753 {
1754 __s64 ret;
1755 struct uffdio_zeropage uffdio_zeropag
1756 struct uffdio_zeropage __user *user_u
1757 struct userfaultfd_wake_range range;
1758
1759 user_uffdio_zeropage = (struct uffdio
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,
1767 /* don't copy "zer
1768 sizeof(uffdio_zero
1769 goto out;
1770
1771 ret = validate_range(ctx->mm, uffdio_
1772 uffdio_zeropage.
1773 if (ret)
1774 goto out;
1775 ret = -EINVAL;
1776 if (uffdio_zeropage.mode & ~UFFDIO_ZE
1777 goto out;
1778
1779 if (mmget_not_zero(ctx->mm)) {
1780 ret = mfill_atomic_zeropage(c
1781 uf
1782 mmput(ctx->mm);
1783 } else {
1784 return -ESRCH;
1785 }
1786 if (unlikely(put_user(ret, &user_uffd
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_Z
1794 range.start = uffdio_zeropage
1795 wake_userfault(ctx, &range);
1796 }
1797 ret = range.len == uffdio_zeropage.ra
1798 out:
1799 return ret;
1800 }
1801
1802 static int userfaultfd_writeprotect(struct us
1803 unsigned
1804 {
1805 int ret;
1806 struct uffdio_writeprotect uffdio_wp;
1807 struct uffdio_writeprotect __user *us
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_write
1815
1816 if (copy_from_user(&uffdio_wp, user_u
1817 sizeof(struct uffd
1818 return -EFAULT;
1819
1820 ret = validate_range(ctx->mm, uffdio_
1821 uffdio_wp.range.
1822 if (ret)
1823 return ret;
1824
1825 if (uffdio_wp.mode & ~(UFFDIO_WRITEPR
1826 UFFDIO_WRITEPR
1827 return -EINVAL;
1828
1829 mode_wp = uffdio_wp.mode & UFFDIO_WRI
1830 mode_dontwake = uffdio_wp.mode & UFFD
1831
1832 if (mode_wp && mode_dontwake)
1833 return -EINVAL;
1834
1835 if (mmget_not_zero(ctx->mm)) {
1836 ret = mwriteprotect_range(ctx
1837 uff
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
1848 range.len = uffdio_wp.range.l
1849 wake_userfault(ctx, &range);
1850 }
1851 return ret;
1852 }
1853
1854 static int userfaultfd_continue(struct userfa
1855 {
1856 __s64 ret;
1857 struct uffdio_continue uffdio_continu
1858 struct uffdio_continue __user *user_u
1859 struct userfaultfd_wake_range range;
1860 uffd_flags_t flags = 0;
1861
1862 user_uffdio_continue = (struct uffdio
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,
1870 /* don't copy the
1871 sizeof(uffdio_cont
1872 goto out;
1873
1874 ret = validate_range(ctx->mm, uffdio_
1875 uffdio_continue.
1876 if (ret)
1877 goto out;
1878
1879 ret = -EINVAL;
1880 if (uffdio_continue.mode & ~(UFFDIO_C
1881 UFFDIO_C
1882 goto out;
1883 if (uffdio_continue.mode & UFFDIO_CON
1884 flags |= MFILL_ATOMIC_WP;
1885
1886 if (mmget_not_zero(ctx->mm)) {
1887 ret = mfill_atomic_continue(c
1888 u
1889 mmput(ctx->mm);
1890 } else {
1891 return -ESRCH;
1892 }
1893
1894 if (unlikely(put_user(ret, &user_uffd
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_C
1903 range.start = uffdio_continue
1904 wake_userfault(ctx, &range);
1905 }
1906 ret = range.len == uffdio_continue.ra
1907
1908 out:
1909 return ret;
1910 }
1911
1912 static inline int userfaultfd_poison(struct u
1913 {
1914 __s64 ret;
1915 struct uffdio_poison uffdio_poison;
1916 struct uffdio_poison __user *user_uff
1917 struct userfaultfd_wake_range range;
1918
1919 user_uffdio_poison = (struct uffdio_p
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, us
1927 /* don't copy the
1928 sizeof(uffdio_pois
1929 goto out;
1930
1931 ret = validate_range(ctx->mm, uffdio_
1932 uffdio_poison.ra
1933 if (ret)
1934 goto out;
1935
1936 ret = -EINVAL;
1937 if (uffdio_poison.mode & ~UFFDIO_POIS
1938 goto out;
1939
1940 if (mmget_not_zero(ctx->mm)) {
1941 ret = mfill_atomic_poison(ctx
1942 uff
1943 mmput(ctx->mm);
1944 } else {
1945 return -ESRCH;
1946 }
1947
1948 if (unlikely(put_user(ret, &user_uffd
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_POI
1957 range.start = uffdio_poison.r
1958 wake_userfault(ctx, &range);
1959 }
1960 ret = range.len == uffdio_poison.rang
1961
1962 out:
1963 return ret;
1964 }
1965
1966 bool userfaultfd_wp_async(struct vm_area_stru
1967 {
1968 return userfaultfd_wp_async_ctx(vma->
1969 }
1970
1971 static inline unsigned int uffd_ctx_features(
1972 {
1973 /*
1974 * For the current set of features th
1975 * UFFD_FEATURE_INITIALIZED to mark t
1976 */
1977 return (unsigned int)user_features |
1978 }
1979
1980 static int userfaultfd_move(struct userfaultf
1981 unsigned long arg
1982 {
1983 __s64 ret;
1984 struct uffdio_move uffdio_move;
1985 struct uffdio_move __user *user_uffdi
1986 struct userfaultfd_wake_range range;
1987 struct mm_struct *mm = ctx->mm;
1988
1989 user_uffdio_move = (struct uffdio_mov
1990
1991 if (atomic_read(&ctx->mmap_changing))
1992 return -EAGAIN;
1993
1994 if (copy_from_user(&uffdio_move, user
1995 /* don't copy "mov
1996 sizeof(uffdio_move
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.
2004 if (ret)
2005 return ret;
2006
2007 ret = validate_range(mm, uffdio_move.
2008 if (ret)
2009 return ret;
2010
2011 if (uffdio_move.mode & ~(UFFDIO_MOVE_
2012 UFFDIO_MOVE
2013 return -EINVAL;
2014
2015 if (mmget_not_zero(mm)) {
2016 ret = move_pages(ctx, uffdio_
2017 uffdio_move.
2018 mmput(mm);
2019 } else {
2020 return -ESRCH;
2021 }
2022
2023 if (unlikely(put_user(ret, &user_uffd
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_
2032 range.start = uffdio_move.dst
2033 wake_userfault(ctx, &range);
2034 }
2035 ret = range.len == uffdio_move.len ?
2036
2037 out:
2038 return ret;
2039 }
2040
2041 /*
2042 * userland asks for a certain API version an
2043 * and ioctl commands are implemented in this
2044 * version or -EINVAL if unknown.
2045 */
2046 static int userfaultfd_api(struct userfaultfd
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,
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_FO
2064 goto err_out;
2065
2066 /* WP_ASYNC relies on WP_UNPOPULATED,
2067 if (features & UFFD_FEATURE_WP_ASYNC)
2068 features |= UFFD_FEATURE_WP_U
2069
2070 /* report all available features and
2071 uffdio_api.features = UFFD_API_FEATUR
2072 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2073 uffdio_api.features &=
2074 ~(UFFD_FEATURE_MINOR_HUGETLBF
2075 #endif
2076 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2077 uffdio_api.features &= ~UFFD_FEATURE_
2078 #endif
2079 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2080 uffdio_api.features &= ~UFFD_FEATURE_
2081 uffdio_api.features &= ~UFFD_FEATURE_
2082 uffdio_api.features &= ~UFFD_FEATURE_
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, si
2092 goto out;
2093
2094 /* only enable the requested features
2095 ctx_features = uffd_ctx_features(feat
2096 ret = -EINVAL;
2097 if (cmpxchg(&ctx->features, 0, ctx_fe
2098 goto err_out;
2099
2100 ret = 0;
2101 out:
2102 return ret;
2103 err_out:
2104 memset(&uffdio_api, 0, sizeof(uffdio_
2105 if (copy_to_user(buf, &uffdio_api, si
2106 ret = -EFAULT;
2107 goto out;
2108 }
2109
2110 static long userfaultfd_ioctl(struct file *fi
2111 unsigned long a
2112 {
2113 int ret = -EINVAL;
2114 struct userfaultfd_ctx *ctx = file->p
2115
2116 if (cmd != UFFDIO_API && !userfaultfd
2117 return -EINVAL;
2118
2119 switch(cmd) {
2120 case UFFDIO_API:
2121 ret = userfaultfd_api(ctx, ar
2122 break;
2123 case UFFDIO_REGISTER:
2124 ret = userfaultfd_register(ct
2125 break;
2126 case UFFDIO_UNREGISTER:
2127 ret = userfaultfd_unregister(
2128 break;
2129 case UFFDIO_WAKE:
2130 ret = userfaultfd_wake(ctx, a
2131 break;
2132 case UFFDIO_COPY:
2133 ret = userfaultfd_copy(ctx, a
2134 break;
2135 case UFFDIO_ZEROPAGE:
2136 ret = userfaultfd_zeropage(ct
2137 break;
2138 case UFFDIO_MOVE:
2139 ret = userfaultfd_move(ctx, a
2140 break;
2141 case UFFDIO_WRITEPROTECT:
2142 ret = userfaultfd_writeprotec
2143 break;
2144 case UFFDIO_CONTINUE:
2145 ret = userfaultfd_continue(ct
2146 break;
2147 case UFFDIO_POISON:
2148 ret = userfaultfd_poison(ctx,
2149 break;
2150 }
2151 return ret;
2152 }
2153
2154 #ifdef CONFIG_PROC_FS
2155 static void userfaultfd_show_fdinfo(struct se
2156 {
2157 struct userfaultfd_ctx *ctx = f->priv
2158 wait_queue_entry_t *wq;
2159 unsigned long pending = 0, total = 0;
2160
2161 spin_lock_irq(&ctx->fault_pending_wqh
2162 list_for_each_entry(wq, &ctx->fault_p
2163 pending++;
2164 total++;
2165 }
2166 list_for_each_entry(wq, &ctx->fault_w
2167 total++;
2168 }
2169 spin_unlock_irq(&ctx->fault_pending_w
2170
2171 /*
2172 * If more protocols will be added, t
2173 * separated by a space. Like this:
2174 * protocols: aa:... bb:...
2175 */
2176 seq_printf(m, "pending:\t%lu\ntotal:\
2177 pending, total, UFFD_API,
2178 UFFD_API_IOCTLS|UFFD_API_R
2179 }
2180 #endif
2181
2182 static const struct file_operations userfault
2183 #ifdef CONFIG_PROC_FS
2184 .show_fdinfo = userfaultfd_show_fd
2185 #endif
2186 .release = userfaultfd_release
2187 .poll = userfaultfd_poll,
2188 .read_iter = userfaultfd_read_it
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 *m
2195 {
2196 struct userfaultfd_ctx *ctx = (struct
2197
2198 init_waitqueue_head(&ctx->fault_pendi
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_s
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 con
2214 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UF
2215 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXE
2216 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBL
2217
2218 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS
2219 return -EINVAL;
2220
2221 ctx = kmem_cache_alloc(userfaultfd_ct
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
2234 if (fd < 0)
2235 goto err_out;
2236
2237 /* Create a new inode so that the LSM
2238 file = anon_inode_create_getfile("[us
2239 O_RDONLY | (flags & U
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_cache
2252 return fd;
2253 }
2254
2255 static inline bool userfaultfd_syscall_allowe
2256 {
2257 /* Userspace-only page faults are alw
2258 if (flags & UFFD_USER_MODE_ONLY)
2259 return true;
2260
2261 /*
2262 * The user is requesting a userfault
2263 * Privileged users are always allowe
2264 */
2265 if (capable(CAP_SYS_PTRACE))
2266 return true;
2267
2268 /* Otherwise, access to kernel fault
2269 return sysctl_unprivileged_userfaultf
2270 }
2271
2272 SYSCALL_DEFINE1(userfaultfd, int, flags)
2273 {
2274 if (!userfaultfd_syscall_allowed(flag
2275 return -EPERM;
2276
2277 return new_userfaultfd(flags);
2278 }
2279
2280 static long userfaultfd_dev_ioctl(struct file
2281 {
2282 if (cmd != USERFAULTFD_IOC_NEW)
2283 return -EINVAL;
2284
2285 return new_userfaultfd(flags);
2286 }
2287
2288 static const struct file_operations userfault
2289 .unlocked_ioctl = userfaultfd_dev_ioc
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_c
2310
2311
2312
2313
2314 #ifdef CONFIG_SYSCTL
2315 register_sysctl_init("vm", vm_userfau
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.