1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
6 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc.
9 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing).
13 #include <linux/list.h>
14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/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>
36 static int sysctl_unprivileged_userfaultfd __read_mostly
;
39 static struct ctl_table vm_userfaultfd_table
[] = {
41 .procname
= "unprivileged_userfaultfd",
42 .data
= &sysctl_unprivileged_userfaultfd
,
43 .maxlen
= sizeof(sysctl_unprivileged_userfaultfd
),
45 .proc_handler
= proc_dointvec_minmax
,
46 .extra1
= SYSCTL_ZERO
,
52 static struct kmem_cache
*userfaultfd_ctx_cachep __ro_after_init
;
54 struct userfaultfd_fork_ctx
{
55 struct userfaultfd_ctx
*orig
;
56 struct userfaultfd_ctx
*new;
57 struct list_head list
;
60 struct userfaultfd_unmap_ctx
{
61 struct userfaultfd_ctx
*ctx
;
64 struct list_head list
;
67 struct userfaultfd_wait_queue
{
69 wait_queue_entry_t wq
;
70 struct userfaultfd_ctx
*ctx
;
74 struct userfaultfd_wake_range
{
79 /* internal indication that UFFD_API ioctl was successfully executed */
80 #define UFFD_FEATURE_INITIALIZED (1u << 31)
82 static bool userfaultfd_is_initialized(struct userfaultfd_ctx
*ctx
)
84 return ctx
->features
& UFFD_FEATURE_INITIALIZED
;
87 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx
*ctx
)
89 return ctx
&& (ctx
->features
& UFFD_FEATURE_WP_ASYNC
);
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
97 bool userfaultfd_wp_unpopulated(struct vm_area_struct
*vma
)
99 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
104 return ctx
->features
& UFFD_FEATURE_WP_UNPOPULATED
;
107 static void userfaultfd_set_vm_flags(struct vm_area_struct
*vma
,
110 const bool uffd_wp_changed
= (vma
->vm_flags
^ flags
) & VM_UFFD_WP
;
112 vm_flags_reset(vma
, flags
);
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.
118 if ((vma
->vm_flags
& VM_SHARED
) && uffd_wp_changed
)
119 vma_set_page_prot(vma
);
122 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
123 int wake_flags
, void *key
)
125 struct userfaultfd_wake_range
*range
= key
;
127 struct userfaultfd_wait_queue
*uwq
;
128 unsigned long start
, len
;
130 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
132 /* len == 0 means wake all */
133 start
= range
->start
;
135 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
136 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
138 WRITE_ONCE(uwq
->waken
, true);
140 * The Program-Order guarantees provided by the scheduler
141 * ensure uwq->waken is visible before the task is woken.
143 ret
= wake_up_state(wq
->private, mode
);
146 * Wake only once, autoremove behavior.
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().
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).
156 list_del_init(&wq
->entry
);
163 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
165 * @ctx: [in] Pointer to the userfaultfd context.
167 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
169 refcount_inc(&ctx
->refcount
);
173 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
175 * @ctx: [in] Pointer to userfaultfd context.
177 * The userfaultfd context reference must have been previously acquired either
178 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
180 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
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
));
192 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
196 static inline void msg_init(struct uffd_msg
*msg
)
198 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
200 * Must use memset to zero out the paddings or kernel data is
201 * leaked to userland.
203 memset(msg
, 0, sizeof(struct uffd_msg
));
206 static inline struct uffd_msg
userfault_msg(unsigned long address
,
207 unsigned long real_address
,
209 unsigned long reason
,
210 unsigned int features
)
215 msg
.event
= UFFD_EVENT_PAGEFAULT
;
217 msg
.arg
.pagefault
.address
= (features
& UFFD_FEATURE_EXACT_ADDRESS
) ?
218 real_address
: address
;
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.
226 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
227 * fault. Otherwise, it was a read fault.
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
);
240 #ifdef CONFIG_HUGETLB_PAGE
242 * Same functionality as userfaultfd_must_wait below with modifications for
245 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
246 struct vm_fault
*vmf
,
247 unsigned long reason
)
249 struct vm_area_struct
*vma
= vmf
->vma
;
253 assert_fault_locked(vmf
);
255 ptep
= hugetlb_walk(vma
, vmf
->address
, vma_mmu_pagesize(vma
));
260 pte
= huge_ptep_get(ptep
);
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
267 if (huge_pte_none_mostly(pte
))
269 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
275 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
276 struct vm_fault
*vmf
,
277 unsigned long reason
)
279 return false; /* should never get here */
281 #endif /* CONFIG_HUGETLB_PAGE */
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
290 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
291 struct vm_fault
*vmf
,
292 unsigned long reason
)
294 struct mm_struct
*mm
= ctx
->mm
;
295 unsigned long address
= vmf
->address
;
304 assert_fault_locked(vmf
);
306 pgd
= pgd_offset(mm
, address
);
307 if (!pgd_present(*pgd
))
309 p4d
= p4d_offset(pgd
, address
);
310 if (!p4d_present(*p4d
))
312 pud
= pud_offset(p4d
, address
);
313 if (!pud_present(*pud
))
315 pmd
= pmd_offset(pud
, address
);
317 _pmd
= pmdp_get_lockless(pmd
);
322 if (!pmd_present(_pmd
) || pmd_devmap(_pmd
))
325 if (pmd_trans_huge(_pmd
)) {
326 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
331 pte
= pte_offset_map(pmd
, address
);
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
341 ptent
= ptep_get(pte
);
342 if (pte_none_mostly(ptent
))
344 if (!pte_write(ptent
) && (reason
& VM_UFFD_WP
))
352 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
354 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
355 return TASK_INTERRUPTIBLE
;
357 if (flags
& FAULT_FLAG_KILLABLE
)
358 return TASK_KILLABLE
;
360 return TASK_UNINTERRUPTIBLE
;
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.
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
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
378 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
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
;
386 unsigned int blocking_state
;
389 * We don't do userfault handling for the final child pid update.
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.
399 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
402 assert_fault_locked(vmf
);
404 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
408 BUG_ON(ctx
->mm
!= mm
);
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)));
415 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
417 if (!(vmf
->flags
& FAULT_FLAG_USER
) && (ctx
->flags
& UFFD_USER_MODE_ONLY
))
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.
425 if (unlikely(READ_ONCE(ctx
->released
))) {
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'.
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
442 ret
= VM_FAULT_NOPAGE
;
447 * Check that we can return VM_FAULT_RETRY.
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.
456 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
458 * Validate the invariant that nowait must allow retry
459 * to be sure not to return SIGBUS erroneously on
460 * nowait invocations.
462 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
463 #ifdef CONFIG_DEBUG_VM
464 if (printk_ratelimit()) {
466 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
475 * Handle nowait, not much to do other than tell it to retry
478 ret
= VM_FAULT_RETRY
;
479 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
482 /* take the reference before dropping the mmap_lock */
483 userfaultfd_ctx_get(ctx
);
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
);
492 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
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().
500 if (is_vm_hugetlb_page(vma
))
501 hugetlb_vma_lock_read(vma
);
503 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
505 * After the __add_wait_queue the uwq is visible to userland
506 * through poll/read().
508 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
510 * The smp_mb() after __set_current_state prevents the reads
511 * following the spin_unlock to happen before the list_add in
514 set_current_state(blocking_state
);
515 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
517 if (!is_vm_hugetlb_page(vma
))
518 must_wait
= userfaultfd_must_wait(ctx
, vmf
, reason
);
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
);
525 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
526 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
530 __set_current_state(TASK_RUNNING
);
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.
545 if (!list_empty_careful(&uwq
.wq
.entry
)) {
546 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
548 * No need of list_del_init(), the uwq on the stack
549 * will be freed shortly anyway.
551 list_del(&uwq
.wq
.entry
);
552 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
556 * ctx may go away after this if the userfault pseudo fd is
559 userfaultfd_ctx_put(ctx
);
565 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
566 struct userfaultfd_wait_queue
*ewq
)
568 struct userfaultfd_ctx
*release_new_ctx
;
570 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
574 init_waitqueue_entry(&ewq
->wq
, current
);
575 release_new_ctx
= NULL
;
577 spin_lock_irq(&ctx
->event_wqh
.lock
);
579 * After the __add_wait_queue the uwq is visible to userland
580 * through poll/read().
582 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
584 set_current_state(TASK_KILLABLE
);
585 if (ewq
->msg
.event
== 0)
587 if (READ_ONCE(ctx
->released
) ||
588 fatal_signal_pending(current
)) {
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
595 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
596 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
597 struct userfaultfd_ctx
*new;
599 new = (struct userfaultfd_ctx
*)
601 ewq
->msg
.arg
.reserved
.reserved1
;
602 release_new_ctx
= new;
607 spin_unlock_irq(&ctx
->event_wqh
.lock
);
609 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
612 spin_lock_irq(&ctx
->event_wqh
.lock
);
614 __set_current_state(TASK_RUNNING
);
615 spin_unlock_irq(&ctx
->event_wqh
.lock
);
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);
622 /* the various vma->vm_userfaultfd_ctx still points to it */
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
);
632 mmap_write_unlock(mm
);
634 userfaultfd_ctx_put(release_new_ctx
);
638 * ctx may go away after this if the userfault pseudo fd is
642 atomic_dec(&ctx
->mmap_changing
);
643 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
644 userfaultfd_ctx_put(ctx
);
647 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
648 struct userfaultfd_wait_queue
*ewq
)
651 wake_up_locked(&ctx
->event_wqh
);
652 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
655 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
657 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
658 struct userfaultfd_fork_ctx
*fctx
;
660 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
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
);
671 list_for_each_entry(fctx
, fcs
, list
)
672 if (fctx
->orig
== octx
) {
678 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
682 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
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
;
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
);
703 list_add_tail(&fctx
->list
, fcs
);
706 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
710 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
712 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
713 struct userfaultfd_wait_queue ewq
;
717 ewq
.msg
.event
= UFFD_EVENT_FORK
;
718 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
720 userfaultfd_event_wait_completion(ctx
, &ewq
);
723 void dup_userfaultfd_complete(struct list_head
*fcs
)
725 struct userfaultfd_fork_ctx
*fctx
, *n
;
727 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
729 list_del(&fctx
->list
);
734 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
735 struct vm_userfaultfd_ctx
*vm_ctx
)
737 struct userfaultfd_ctx
*ctx
;
739 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
744 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
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
);
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
);
758 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
759 unsigned long from
, unsigned long to
,
762 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
763 struct userfaultfd_wait_queue ewq
;
768 if (to
& ~PAGE_MASK
) {
769 userfaultfd_ctx_put(ctx
);
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
;
780 userfaultfd_event_wait_completion(ctx
, &ewq
);
783 bool userfaultfd_remove(struct vm_area_struct
*vma
,
784 unsigned long start
, unsigned long end
)
786 struct mm_struct
*mm
= vma
->vm_mm
;
787 struct userfaultfd_ctx
*ctx
;
788 struct userfaultfd_wait_queue ewq
;
790 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
791 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
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
);
802 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
803 ewq
.msg
.arg
.remove
.start
= start
;
804 ewq
.msg
.arg
.remove
.end
= end
;
806 userfaultfd_event_wait_completion(ctx
, &ewq
);
811 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
812 unsigned long start
, unsigned long end
)
814 struct userfaultfd_unmap_ctx
*unmap_ctx
;
816 list_for_each_entry(unmap_ctx
, unmaps
, list
)
817 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
818 unmap_ctx
->end
== end
)
824 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
, unsigned long start
,
825 unsigned long end
, struct list_head
*unmaps
)
827 struct userfaultfd_unmap_ctx
*unmap_ctx
;
828 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
830 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
831 has_unmap_ctx(ctx
, unmaps
, start
, end
))
834 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
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
);
850 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
852 struct userfaultfd_unmap_ctx
*ctx
, *n
;
853 struct userfaultfd_wait_queue ewq
;
855 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
858 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
859 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
860 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
862 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
864 list_del(&ctx
->list
);
869 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
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);
879 WRITE_ONCE(ctx
->released
, true);
881 if (!mmget_not_zero(mm
))
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.
894 for_each_vma(vmi
, vma
) {
896 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
897 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
898 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
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
,
911 vma_start_write(vma
);
912 userfaultfd_set_vm_flags(vma
, new_flags
);
913 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
917 mmap_write_unlock(mm
);
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
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
);
930 /* Flush pending events that may still wait on event_wqh */
931 wake_up_all(&ctx
->event_wqh
);
933 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
934 userfaultfd_ctx_put(ctx
);
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
)
942 wait_queue_entry_t
*wq
;
943 struct userfaultfd_wait_queue
*uwq
;
945 lockdep_assert_held(&wqh
->lock
);
948 if (!waitqueue_active(wqh
))
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
);
957 static inline struct userfaultfd_wait_queue
*find_userfault(
958 struct userfaultfd_ctx
*ctx
)
960 return find_userfault_in(&ctx
->fault_pending_wqh
);
963 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
964 struct userfaultfd_ctx
*ctx
)
966 return find_userfault_in(&ctx
->event_wqh
);
969 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
971 struct userfaultfd_ctx
*ctx
= file
->private_data
;
974 poll_wait(file
, &ctx
->fd_wqh
, wait
);
976 if (!userfaultfd_is_initialized(ctx
))
980 * poll() never guarantees that read won't block.
981 * userfaults can be waken before they're read().
983 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
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
997 if (waitqueue_active(&ctx
->fault_pending_wqh
))
999 else if (waitqueue_active(&ctx
->event_wqh
))
1005 static const struct file_operations userfaultfd_fops
;
1007 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
1008 struct inode
*inode
,
1009 struct uffd_msg
*msg
)
1013 fd
= anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops
, new,
1014 O_RDONLY
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
1018 msg
->arg
.reserved
.reserved1
= 0;
1019 msg
->arg
.fork
.ufd
= fd
;
1023 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1024 struct uffd_msg
*msg
, struct inode
*inode
)
1027 DECLARE_WAITQUEUE(wait
, current
);
1028 struct userfaultfd_wait_queue
*uwq
;
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
1036 LIST_HEAD(fork_event
);
1037 struct userfaultfd_ctx
*fork_nctx
= NULL
;
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
);
1043 set_current_state(TASK_INTERRUPTIBLE
);
1044 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1045 uwq
= find_userfault(ctx
);
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
1054 write_seqcount_begin(&ctx
->refile_seq
);
1057 * The fault_pending_wqh.lock prevents the uwq
1058 * to disappear from under us.
1060 * Refile this userfault from
1061 * fault_pending_wqh to fault_wqh, it's not
1062 * pending anymore after we read it.
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
1077 list_del(&uwq
->wq
.entry
);
1078 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1080 write_seqcount_end(&ctx
->refile_seq
);
1082 /* careful to always initialize msg if ret == 0 */
1084 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1088 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1090 spin_lock(&ctx
->event_wqh
.lock
);
1091 uwq
= find_userfault_evt(ctx
);
1095 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1096 fork_nctx
= (struct userfaultfd_ctx
*)
1098 uwq
->msg
.arg
.reserved
.reserved1
;
1099 list_move(&uwq
->wq
.entry
, &fork_event
);
1101 * fork_nctx can be freed as soon as
1102 * we drop the lock, unless we take a
1105 userfaultfd_ctx_get(fork_nctx
);
1106 spin_unlock(&ctx
->event_wqh
.lock
);
1111 userfaultfd_event_complete(ctx
, uwq
);
1112 spin_unlock(&ctx
->event_wqh
.lock
);
1116 spin_unlock(&ctx
->event_wqh
.lock
);
1118 if (signal_pending(current
)) {
1126 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1128 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1130 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1131 __set_current_state(TASK_RUNNING
);
1132 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
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
)) {
1139 * The fork thread didn't abort, so we can
1140 * drop the temporary refcount.
1142 userfaultfd_ctx_put(fork_nctx
);
1144 uwq
= list_first_entry(&fork_event
,
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
1157 list_del(&uwq
->wq
.entry
);
1158 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1161 * Leave the event in the waitqueue and report
1162 * error to userland if we failed to resolve
1163 * the userfault fork.
1166 userfaultfd_event_complete(ctx
, uwq
);
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.
1180 userfaultfd_ctx_put(fork_nctx
);
1182 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1188 static ssize_t
userfaultfd_read_iter(struct kiocb
*iocb
, struct iov_iter
*to
)
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
);
1197 if (!userfaultfd_is_initialized(ctx
))
1200 no_wait
= file
->f_flags
& O_NONBLOCK
|| iocb
->ki_flags
& IOCB_NOWAIT
;
1202 if (iov_iter_count(to
) < sizeof(msg
))
1203 return ret
? ret
: -EINVAL
;
1204 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1206 return ret
? ret
: _ret
;
1207 _ret
= !copy_to_iter_full(&msg
, sizeof(msg
), to
);
1209 return ret
? ret
: -EFAULT
;
1212 * Allow to read more than one fault at time but only
1213 * block if waiting for the very first one.
1219 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1220 struct userfaultfd_wake_range
*range
)
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
,
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
);
1232 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1233 struct userfaultfd_wake_range
*range
)
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.
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.
1254 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1255 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1256 waitqueue_active(&ctx
->fault_wqh
);
1258 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1260 __wake_userfault(ctx
, range
);
1263 static __always_inline
int validate_unaligned_range(
1264 struct mm_struct
*mm
, __u64 start
, __u64 len
)
1266 __u64 task_size
= mm
->task_size
;
1268 if (len
& ~PAGE_MASK
)
1272 if (start
< mmap_min_addr
)
1274 if (start
>= task_size
)
1276 if (len
> task_size
- start
)
1278 if (start
+ len
<= start
)
1283 static __always_inline
int validate_range(struct mm_struct
*mm
,
1284 __u64 start
, __u64 len
)
1286 if (start
& ~PAGE_MASK
)
1289 return validate_unaligned_range(mm
, start
, len
);
1292 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1295 struct mm_struct
*mm
= ctx
->mm
;
1296 struct vm_area_struct
*vma
, *prev
, *cur
;
1298 struct uffdio_register uffdio_register
;
1299 struct uffdio_register __user
*user_uffdio_register
;
1300 unsigned long vm_flags
, new_flags
;
1303 unsigned long start
, end
, vma_end
;
1304 struct vma_iterator vmi
;
1305 bool wp_async
= userfaultfd_wp_async_ctx(ctx
);
1307 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1310 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1311 sizeof(uffdio_register
)-sizeof(__u64
)))
1315 if (!uffdio_register
.mode
)
1317 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
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
1326 vm_flags
|= VM_UFFD_WP
;
1328 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1329 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1332 vm_flags
|= VM_UFFD_MINOR
;
1335 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1336 uffdio_register
.range
.len
);
1340 start
= uffdio_register
.range
.start
;
1341 end
= start
+ uffdio_register
.range
.len
;
1344 if (!mmget_not_zero(mm
))
1348 mmap_write_lock(mm
);
1349 vma_iter_init(&vmi
, mm
, start
);
1350 vma
= vma_find(&vmi
, end
);
1355 * If the first vma contains huge pages, make sure start address
1356 * is aligned to huge page size.
1358 if (is_vm_hugetlb_page(vma
)) {
1359 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1361 if (start
& (vma_hpagesize
- 1))
1366 * Search for not compatible vmas.
1369 basic_ioctls
= false;
1374 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1375 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1377 /* check not compatible vmas */
1379 if (!vma_can_userfault(cur
, vm_flags
, wp_async
))
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.
1391 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1395 * If this vma contains ending address, and huge pages
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
);
1404 if (end
& (vma_hpagesize
- 1))
1407 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
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.
1417 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1418 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1422 * Note vmas containing huge pages
1424 if (is_vm_hugetlb_page(cur
))
1425 basic_ioctls
= true;
1428 } for_each_vma_range(vmi
, cur
, end
);
1431 vma_iter_set(&vmi
, start
);
1432 prev
= vma_prev(&vmi
);
1433 if (vma
->vm_start
< start
)
1437 for_each_vma_range(vmi
, vma
, end
) {
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
));
1446 * Nothing to do: this vma is already registered into this
1447 * userfaultfd and with the right tracking mode too.
1449 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1450 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1453 if (vma
->vm_start
> start
)
1454 start
= vma
->vm_start
;
1455 vma_end
= min(end
, vma
->vm_end
);
1457 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1458 vma
= vma_modify_flags_uffd(&vmi
, prev
, vma
, start
, vma_end
,
1460 (struct vm_userfaultfd_ctx
){ctx
});
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.
1471 vma_start_write(vma
);
1472 userfaultfd_set_vm_flags(vma
, new_flags
);
1473 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1475 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1476 hugetlb_unshare_all_pmds(vma
);
1480 start
= vma
->vm_end
;
1484 mmap_write_unlock(mm
);
1489 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1490 UFFD_API_RANGE_IOCTLS
;
1493 * Declare the WP ioctl only if the WP mode is
1494 * specified and all checks passed with the range
1496 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1497 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
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
);
1504 * Now that we scanned all vmas we can already tell
1505 * userland which ioctls methods are guaranteed to
1506 * succeed on this range.
1508 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1515 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1518 struct mm_struct
*mm
= ctx
->mm
;
1519 struct vm_area_struct
*vma
, *prev
, *cur
;
1521 struct uffdio_range uffdio_unregister
;
1522 unsigned long new_flags
;
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
);
1530 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1533 ret
= validate_range(mm
, uffdio_unregister
.start
,
1534 uffdio_unregister
.len
);
1538 start
= uffdio_unregister
.start
;
1539 end
= start
+ uffdio_unregister
.len
;
1542 if (!mmget_not_zero(mm
))
1545 mmap_write_lock(mm
);
1547 vma_iter_init(&vmi
, mm
, start
);
1548 vma
= vma_find(&vmi
, end
);
1553 * If the first vma contains huge pages, make sure start address
1554 * is aligned to huge page size.
1556 if (is_vm_hugetlb_page(vma
)) {
1557 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1559 if (start
& (vma_hpagesize
- 1))
1564 * Search for not compatible vmas.
1571 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1572 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
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.
1581 if (!vma_can_userfault(cur
, cur
->vm_flags
, wp_async
))
1585 } for_each_vma_range(vmi
, cur
, end
);
1588 vma_iter_set(&vmi
, start
);
1589 prev
= vma_prev(&vmi
);
1590 if (vma
->vm_start
< start
)
1594 for_each_vma_range(vmi
, vma
, end
) {
1597 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
, wp_async
));
1600 * Nothing to do: this vma is already registered into this
1601 * userfaultfd and with the right tracking mode too.
1603 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1606 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1608 if (vma
->vm_start
> start
)
1609 start
= vma
->vm_start
;
1610 vma_end
= min(end
, vma
->vm_end
);
1612 if (userfaultfd_missing(vma
)) {
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.
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
);
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);
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
);
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.
1642 vma_start_write(vma
);
1643 userfaultfd_set_vm_flags(vma
, new_flags
);
1644 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1648 start
= vma
->vm_end
;
1652 mmap_write_unlock(mm
);
1659 * userfaultfd_wake may be used in combination with the
1660 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1662 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1666 struct uffdio_range uffdio_wake
;
1667 struct userfaultfd_wake_range range
;
1668 const void __user
*buf
= (void __user
*)arg
;
1671 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1674 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1678 range
.start
= uffdio_wake
.start
;
1679 range
.len
= uffdio_wake
.len
;
1682 * len == 0 means wake all and we don't want to wake all here,
1683 * so check it again to be sure.
1685 VM_BUG_ON(!range
.len
);
1687 wake_userfault(ctx
, &range
);
1694 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
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;
1703 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1706 if (atomic_read(&ctx
->mmap_changing
))
1710 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1711 /* don't copy "copy" last field */
1712 sizeof(uffdio_copy
)-sizeof(__s64
)))
1715 ret
= validate_unaligned_range(ctx
->mm
, uffdio_copy
.src
,
1719 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1724 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
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
);
1735 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1740 /* len == 0 would wake all */
1742 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1743 range
.start
= uffdio_copy
.dst
;
1744 wake_userfault(ctx
, &range
);
1746 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1751 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1755 struct uffdio_zeropage uffdio_zeropage
;
1756 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1757 struct userfaultfd_wake_range range
;
1759 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1762 if (atomic_read(&ctx
->mmap_changing
))
1766 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1767 /* don't copy "zeropage" last field */
1768 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1771 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1772 uffdio_zeropage
.range
.len
);
1776 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1779 if (mmget_not_zero(ctx
->mm
)) {
1780 ret
= mfill_atomic_zeropage(ctx
, uffdio_zeropage
.range
.start
,
1781 uffdio_zeropage
.range
.len
);
1786 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1790 /* len == 0 would wake all */
1793 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1794 range
.start
= uffdio_zeropage
.range
.start
;
1795 wake_userfault(ctx
, &range
);
1797 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1802 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
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
;
1811 if (atomic_read(&ctx
->mmap_changing
))
1814 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1816 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1817 sizeof(struct uffdio_writeprotect
)))
1820 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1821 uffdio_wp
.range
.len
);
1825 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1826 UFFDIO_WRITEPROTECT_MODE_WP
))
1829 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1830 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1832 if (mode_wp
&& mode_dontwake
)
1835 if (mmget_not_zero(ctx
->mm
)) {
1836 ret
= mwriteprotect_range(ctx
, uffdio_wp
.range
.start
,
1837 uffdio_wp
.range
.len
, mode_wp
);
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
);
1854 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
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;
1862 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1865 if (atomic_read(&ctx
->mmap_changing
))
1869 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1870 /* don't copy the output fields */
1871 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1874 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1875 uffdio_continue
.range
.len
);
1880 if (uffdio_continue
.mode
& ~(UFFDIO_CONTINUE_MODE_DONTWAKE
|
1881 UFFDIO_CONTINUE_MODE_WP
))
1883 if (uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_WP
)
1884 flags
|= MFILL_ATOMIC_WP
;
1886 if (mmget_not_zero(ctx
->mm
)) {
1887 ret
= mfill_atomic_continue(ctx
, uffdio_continue
.range
.start
,
1888 uffdio_continue
.range
.len
, flags
);
1894 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1899 /* len == 0 would wake all */
1902 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1903 range
.start
= uffdio_continue
.range
.start
;
1904 wake_userfault(ctx
, &range
);
1906 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1912 static inline int userfaultfd_poison(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1915 struct uffdio_poison uffdio_poison
;
1916 struct uffdio_poison __user
*user_uffdio_poison
;
1917 struct userfaultfd_wake_range range
;
1919 user_uffdio_poison
= (struct uffdio_poison __user
*)arg
;
1922 if (atomic_read(&ctx
->mmap_changing
))
1926 if (copy_from_user(&uffdio_poison
, user_uffdio_poison
,
1927 /* don't copy the output fields */
1928 sizeof(uffdio_poison
) - (sizeof(__s64
))))
1931 ret
= validate_range(ctx
->mm
, uffdio_poison
.range
.start
,
1932 uffdio_poison
.range
.len
);
1937 if (uffdio_poison
.mode
& ~UFFDIO_POISON_MODE_DONTWAKE
)
1940 if (mmget_not_zero(ctx
->mm
)) {
1941 ret
= mfill_atomic_poison(ctx
, uffdio_poison
.range
.start
,
1942 uffdio_poison
.range
.len
, 0);
1948 if (unlikely(put_user(ret
, &user_uffdio_poison
->updated
)))
1953 /* len == 0 would wake all */
1956 if (!(uffdio_poison
.mode
& UFFDIO_POISON_MODE_DONTWAKE
)) {
1957 range
.start
= uffdio_poison
.range
.start
;
1958 wake_userfault(ctx
, &range
);
1960 ret
= range
.len
== uffdio_poison
.range
.len
? 0 : -EAGAIN
;
1966 bool userfaultfd_wp_async(struct vm_area_struct
*vma
)
1968 return userfaultfd_wp_async_ctx(vma
->vm_userfaultfd_ctx
.ctx
);
1971 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1974 * For the current set of features the bits just coincide. Set
1975 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1977 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1980 static int userfaultfd_move(struct userfaultfd_ctx
*ctx
,
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
;
1989 user_uffdio_move
= (struct uffdio_move __user
*) arg
;
1991 if (atomic_read(&ctx
->mmap_changing
))
1994 if (copy_from_user(&uffdio_move
, user_uffdio_move
,
1995 /* don't copy "move" last field */
1996 sizeof(uffdio_move
)-sizeof(__s64
)))
1999 /* Do not allow cross-mm moves. */
2000 if (mm
!= current
->mm
)
2003 ret
= validate_range(mm
, uffdio_move
.dst
, uffdio_move
.len
);
2007 ret
= validate_range(mm
, uffdio_move
.src
, uffdio_move
.len
);
2011 if (uffdio_move
.mode
& ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES
|
2012 UFFDIO_MOVE_MODE_DONTWAKE
))
2015 if (mmget_not_zero(mm
)) {
2016 ret
= move_pages(ctx
, uffdio_move
.dst
, uffdio_move
.src
,
2017 uffdio_move
.len
, uffdio_move
.mode
);
2023 if (unlikely(put_user(ret
, &user_uffdio_move
->move
)))
2028 /* len == 0 would wake all */
2031 if (!(uffdio_move
.mode
& UFFDIO_MOVE_MODE_DONTWAKE
)) {
2032 range
.start
= uffdio_move
.dst
;
2033 wake_userfault(ctx
, &range
);
2035 ret
= range
.len
== uffdio_move
.len
? 0 : -EAGAIN
;
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.
2046 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
2049 struct uffdio_api uffdio_api
;
2050 void __user
*buf
= (void __user
*)arg
;
2051 unsigned int ctx_features
;
2056 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
2058 features
= uffdio_api
.features
;
2060 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
2063 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
2066 /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2067 if (features
& UFFD_FEATURE_WP_ASYNC
)
2068 features
|= UFFD_FEATURE_WP_UNPOPULATED
;
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
);
2076 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2077 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
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
;
2084 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
2086 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
2089 /* only enable the requested features for this uffd context */
2090 ctx_features
= uffd_ctx_features(features
);
2092 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
2099 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
2100 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
2105 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
2109 struct userfaultfd_ctx
*ctx
= file
->private_data
;
2111 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
2116 ret
= userfaultfd_api(ctx
, arg
);
2118 case UFFDIO_REGISTER
:
2119 ret
= userfaultfd_register(ctx
, arg
);
2121 case UFFDIO_UNREGISTER
:
2122 ret
= userfaultfd_unregister(ctx
, arg
);
2125 ret
= userfaultfd_wake(ctx
, arg
);
2128 ret
= userfaultfd_copy(ctx
, arg
);
2130 case UFFDIO_ZEROPAGE
:
2131 ret
= userfaultfd_zeropage(ctx
, arg
);
2134 ret
= userfaultfd_move(ctx
, arg
);
2136 case UFFDIO_WRITEPROTECT
:
2137 ret
= userfaultfd_writeprotect(ctx
, arg
);
2139 case UFFDIO_CONTINUE
:
2140 ret
= userfaultfd_continue(ctx
, arg
);
2143 ret
= userfaultfd_poison(ctx
, arg
);
2149 #ifdef CONFIG_PROC_FS
2150 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2152 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2153 wait_queue_entry_t
*wq
;
2154 unsigned long pending
= 0, total
= 0;
2156 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2157 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2161 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2164 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2167 * If more protocols will be added, there will be all shown
2168 * separated by a space. Like this:
2169 * protocols: aa:... bb:...
2171 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2172 pending
, total
, UFFD_API
, ctx
->features
,
2173 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2177 static const struct file_operations userfaultfd_fops
= {
2178 #ifdef CONFIG_PROC_FS
2179 .show_fdinfo
= userfaultfd_show_fdinfo
,
2181 .release
= userfaultfd_release
,
2182 .poll
= userfaultfd_poll
,
2183 .read_iter
= userfaultfd_read_iter
,
2184 .unlocked_ioctl
= userfaultfd_ioctl
,
2185 .compat_ioctl
= compat_ptr_ioctl
,
2186 .llseek
= noop_llseek
,
2189 static void init_once_userfaultfd_ctx(void *mem
)
2191 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2193 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2194 init_waitqueue_head(&ctx
->fault_wqh
);
2195 init_waitqueue_head(&ctx
->event_wqh
);
2196 init_waitqueue_head(&ctx
->fd_wqh
);
2197 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2200 static int new_userfaultfd(int flags
)
2202 struct userfaultfd_ctx
*ctx
;
2206 BUG_ON(!current
->mm
);
2208 /* Check the UFFD_* constants for consistency. */
2209 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2210 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2211 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2213 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2216 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2220 refcount_set(&ctx
->refcount
, 1);
2223 ctx
->released
= false;
2224 init_rwsem(&ctx
->map_changing_lock
);
2225 atomic_set(&ctx
->mmap_changing
, 0);
2226 ctx
->mm
= current
->mm
;
2228 fd
= get_unused_fd_flags(flags
& UFFD_SHARED_FCNTL_FLAGS
);
2232 /* Create a new inode so that the LSM can block the creation. */
2233 file
= anon_inode_create_getfile("[userfaultfd]", &userfaultfd_fops
, ctx
,
2234 O_RDONLY
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2240 /* prevent the mm struct to be freed */
2242 file
->f_mode
|= FMODE_NOWAIT
;
2243 fd_install(fd
, file
);
2246 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2250 static inline bool userfaultfd_syscall_allowed(int flags
)
2252 /* Userspace-only page faults are always allowed */
2253 if (flags
& UFFD_USER_MODE_ONLY
)
2257 * The user is requesting a userfaultfd which can handle kernel faults.
2258 * Privileged users are always allowed to do this.
2260 if (capable(CAP_SYS_PTRACE
))
2263 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2264 return sysctl_unprivileged_userfaultfd
;
2267 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2269 if (!userfaultfd_syscall_allowed(flags
))
2272 return new_userfaultfd(flags
);
2275 static long userfaultfd_dev_ioctl(struct file
*file
, unsigned int cmd
, unsigned long flags
)
2277 if (cmd
!= USERFAULTFD_IOC_NEW
)
2280 return new_userfaultfd(flags
);
2283 static const struct file_operations userfaultfd_dev_fops
= {
2284 .unlocked_ioctl
= userfaultfd_dev_ioctl
,
2285 .compat_ioctl
= userfaultfd_dev_ioctl
,
2286 .owner
= THIS_MODULE
,
2287 .llseek
= noop_llseek
,
2290 static struct miscdevice userfaultfd_misc
= {
2291 .minor
= MISC_DYNAMIC_MINOR
,
2292 .name
= "userfaultfd",
2293 .fops
= &userfaultfd_dev_fops
2296 static int __init
userfaultfd_init(void)
2300 ret
= misc_register(&userfaultfd_misc
);
2304 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2305 sizeof(struct userfaultfd_ctx
),
2307 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2308 init_once_userfaultfd_ctx
);
2309 #ifdef CONFIG_SYSCTL
2310 register_sysctl_init("vm", vm_userfaultfd_table
);
2314 __initcall(userfaultfd_init
);