4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
5 * Copyright (C) 2008-2009 Red Hat, Inc.
6 * Copyright (C) 2015 Red Hat, Inc.
8 * This work is licensed under the terms of the GNU GPL, version 2. See
9 * the COPYING file in the top-level directory.
11 * Some part derived from fs/eventfd.c (anon inode setup) and
12 * mm/ksm.c (mm hashing).
15 #include <linux/list.h>
16 #include <linux/hashtable.h>
17 #include <linux/sched/signal.h>
18 #include <linux/sched/mm.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>
33 static struct kmem_cache
*userfaultfd_ctx_cachep __read_mostly
;
35 enum userfaultfd_state
{
41 * Start with fault_pending_wqh and fault_wqh so they're more likely
42 * to be in the same cacheline.
44 struct userfaultfd_ctx
{
45 /* waitqueue head for the pending (i.e. not read) userfaults */
46 wait_queue_head_t fault_pending_wqh
;
47 /* waitqueue head for the userfaults */
48 wait_queue_head_t fault_wqh
;
49 /* waitqueue head for the pseudo fd to wakeup poll/read */
50 wait_queue_head_t fd_wqh
;
51 /* waitqueue head for events */
52 wait_queue_head_t event_wqh
;
53 /* a refile sequence protected by fault_pending_wqh lock */
54 struct seqcount refile_seq
;
55 /* pseudo fd refcounting */
57 /* userfaultfd syscall flags */
59 /* features requested from the userspace */
60 unsigned int features
;
62 enum userfaultfd_state state
;
65 /* mm with one ore more vmas attached to this userfaultfd_ctx */
69 struct userfaultfd_fork_ctx
{
70 struct userfaultfd_ctx
*orig
;
71 struct userfaultfd_ctx
*new;
72 struct list_head list
;
75 struct userfaultfd_unmap_ctx
{
76 struct userfaultfd_ctx
*ctx
;
79 struct list_head list
;
82 struct userfaultfd_wait_queue
{
84 wait_queue_entry_t wq
;
85 struct userfaultfd_ctx
*ctx
;
89 struct userfaultfd_wake_range
{
94 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
95 int wake_flags
, void *key
)
97 struct userfaultfd_wake_range
*range
= key
;
99 struct userfaultfd_wait_queue
*uwq
;
100 unsigned long start
, len
;
102 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
104 /* len == 0 means wake all */
105 start
= range
->start
;
107 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
108 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
110 WRITE_ONCE(uwq
->waken
, true);
112 * The implicit smp_mb__before_spinlock in try_to_wake_up()
113 * renders uwq->waken visible to other CPUs before the task is
116 ret
= wake_up_state(wq
->private, mode
);
119 * Wake only once, autoremove behavior.
121 * After the effect of list_del_init is visible to the
122 * other CPUs, the waitqueue may disappear from under
123 * us, see the !list_empty_careful() in
124 * handle_userfault(). try_to_wake_up() has an
125 * implicit smp_mb__before_spinlock, and the
126 * wq->private is read before calling the extern
127 * function "wake_up_state" (which in turns calls
128 * try_to_wake_up). While the spin_lock;spin_unlock;
129 * wouldn't be enough, the smp_mb__before_spinlock is
130 * enough to avoid an explicit smp_mb() here.
132 list_del_init(&wq
->entry
);
138 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
140 * @ctx: [in] Pointer to the userfaultfd context.
142 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
144 if (!atomic_inc_not_zero(&ctx
->refcount
))
149 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
151 * @ctx: [in] Pointer to userfaultfd context.
153 * The userfaultfd context reference must have been previously acquired either
154 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
156 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
158 if (atomic_dec_and_test(&ctx
->refcount
)) {
159 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
160 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
161 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
162 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
163 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
164 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
165 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
166 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
168 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
172 static inline void msg_init(struct uffd_msg
*msg
)
174 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
176 * Must use memset to zero out the paddings or kernel data is
177 * leaked to userland.
179 memset(msg
, 0, sizeof(struct uffd_msg
));
182 static inline struct uffd_msg
userfault_msg(unsigned long address
,
184 unsigned long reason
)
188 msg
.event
= UFFD_EVENT_PAGEFAULT
;
189 msg
.arg
.pagefault
.address
= address
;
190 if (flags
& FAULT_FLAG_WRITE
)
192 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
193 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
194 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
195 * was a read fault, otherwise if set it means it's
198 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
199 if (reason
& VM_UFFD_WP
)
201 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
202 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
203 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
204 * a missing fault, otherwise if set it means it's a
205 * write protect fault.
207 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
211 #ifdef CONFIG_HUGETLB_PAGE
213 * Same functionality as userfaultfd_must_wait below with modifications for
216 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
217 struct vm_area_struct
*vma
,
218 unsigned long address
,
220 unsigned long reason
)
222 struct mm_struct
*mm
= ctx
->mm
;
226 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
228 pte
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
235 * Lockless access: we're in a wait_event so it's ok if it
238 if (huge_pte_none(*pte
))
240 if (!huge_pte_write(*pte
) && (reason
& VM_UFFD_WP
))
246 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
247 struct vm_area_struct
*vma
,
248 unsigned long address
,
250 unsigned long reason
)
252 return false; /* should never get here */
254 #endif /* CONFIG_HUGETLB_PAGE */
257 * Verify the pagetables are still not ok after having reigstered into
258 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
259 * userfault that has already been resolved, if userfaultfd_read and
260 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
263 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
264 unsigned long address
,
266 unsigned long reason
)
268 struct mm_struct
*mm
= ctx
->mm
;
276 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
278 pgd
= pgd_offset(mm
, address
);
279 if (!pgd_present(*pgd
))
281 p4d
= p4d_offset(pgd
, address
);
282 if (!p4d_present(*p4d
))
284 pud
= pud_offset(p4d
, address
);
285 if (!pud_present(*pud
))
287 pmd
= pmd_offset(pud
, address
);
289 * READ_ONCE must function as a barrier with narrower scope
290 * and it must be equivalent to:
291 * _pmd = *pmd; barrier();
293 * This is to deal with the instability (as in
294 * pmd_trans_unstable) of the pmd.
296 _pmd
= READ_ONCE(*pmd
);
297 if (!pmd_present(_pmd
))
301 if (pmd_trans_huge(_pmd
))
305 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
306 * and use the standard pte_offset_map() instead of parsing _pmd.
308 pte
= pte_offset_map(pmd
, address
);
310 * Lockless access: we're in a wait_event so it's ok if it
322 * The locking rules involved in returning VM_FAULT_RETRY depending on
323 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
324 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
325 * recommendation in __lock_page_or_retry is not an understatement.
327 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
328 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
331 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
332 * set, VM_FAULT_RETRY can still be returned if and only if there are
333 * fatal_signal_pending()s, and the mmap_sem must be released before
336 int handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
338 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
339 struct userfaultfd_ctx
*ctx
;
340 struct userfaultfd_wait_queue uwq
;
342 bool must_wait
, return_to_userland
;
345 ret
= VM_FAULT_SIGBUS
;
348 * We don't do userfault handling for the final child pid update.
350 * We also don't do userfault handling during
351 * coredumping. hugetlbfs has the special
352 * follow_hugetlb_page() to skip missing pages in the
353 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
354 * the no_page_table() helper in follow_page_mask(), but the
355 * shmem_vm_ops->fault method is invoked even during
356 * coredumping without mmap_sem and it ends up here.
358 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
362 * Coredumping runs without mmap_sem so we can only check that
363 * the mmap_sem is held, if PF_DUMPCORE was not set.
365 WARN_ON_ONCE(!rwsem_is_locked(&mm
->mmap_sem
));
367 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
371 BUG_ON(ctx
->mm
!= mm
);
373 VM_BUG_ON(reason
& ~(VM_UFFD_MISSING
|VM_UFFD_WP
));
374 VM_BUG_ON(!(reason
& VM_UFFD_MISSING
) ^ !!(reason
& VM_UFFD_WP
));
377 * If it's already released don't get it. This avoids to loop
378 * in __get_user_pages if userfaultfd_release waits on the
379 * caller of handle_userfault to release the mmap_sem.
381 if (unlikely(ACCESS_ONCE(ctx
->released
)))
385 * Check that we can return VM_FAULT_RETRY.
387 * NOTE: it should become possible to return VM_FAULT_RETRY
388 * even if FAULT_FLAG_TRIED is set without leading to gup()
389 * -EBUSY failures, if the userfaultfd is to be extended for
390 * VM_UFFD_WP tracking and we intend to arm the userfault
391 * without first stopping userland access to the memory. For
392 * VM_UFFD_MISSING userfaults this is enough for now.
394 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
396 * Validate the invariant that nowait must allow retry
397 * to be sure not to return SIGBUS erroneously on
398 * nowait invocations.
400 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
401 #ifdef CONFIG_DEBUG_VM
402 if (printk_ratelimit()) {
404 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
413 * Handle nowait, not much to do other than tell it to retry
416 ret
= VM_FAULT_RETRY
;
417 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
420 /* take the reference before dropping the mmap_sem */
421 userfaultfd_ctx_get(ctx
);
423 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
424 uwq
.wq
.private = current
;
425 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
);
430 (vmf
->flags
& (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
)) ==
431 (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
);
432 blocking_state
= return_to_userland
? TASK_INTERRUPTIBLE
:
435 spin_lock(&ctx
->fault_pending_wqh
.lock
);
437 * After the __add_wait_queue the uwq is visible to userland
438 * through poll/read().
440 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
442 * The smp_mb() after __set_current_state prevents the reads
443 * following the spin_unlock to happen before the list_add in
446 set_current_state(blocking_state
);
447 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
449 if (!is_vm_hugetlb_page(vmf
->vma
))
450 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
453 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
456 up_read(&mm
->mmap_sem
);
458 if (likely(must_wait
&& !ACCESS_ONCE(ctx
->released
) &&
459 (return_to_userland
? !signal_pending(current
) :
460 !fatal_signal_pending(current
)))) {
461 wake_up_poll(&ctx
->fd_wqh
, POLLIN
);
463 ret
|= VM_FAULT_MAJOR
;
466 * False wakeups can orginate even from rwsem before
467 * up_read() however userfaults will wait either for a
468 * targeted wakeup on the specific uwq waitqueue from
469 * wake_userfault() or for signals or for uffd
472 while (!READ_ONCE(uwq
.waken
)) {
474 * This needs the full smp_store_mb()
475 * guarantee as the state write must be
476 * visible to other CPUs before reading
477 * uwq.waken from other CPUs.
479 set_current_state(blocking_state
);
480 if (READ_ONCE(uwq
.waken
) ||
481 READ_ONCE(ctx
->released
) ||
482 (return_to_userland
? signal_pending(current
) :
483 fatal_signal_pending(current
)))
489 __set_current_state(TASK_RUNNING
);
491 if (return_to_userland
) {
492 if (signal_pending(current
) &&
493 !fatal_signal_pending(current
)) {
495 * If we got a SIGSTOP or SIGCONT and this is
496 * a normal userland page fault, just let
497 * userland return so the signal will be
498 * handled and gdb debugging works. The page
499 * fault code immediately after we return from
500 * this function is going to release the
501 * mmap_sem and it's not depending on it
502 * (unlike gup would if we were not to return
505 * If a fatal signal is pending we still take
506 * the streamlined VM_FAULT_RETRY failure path
507 * and there's no need to retake the mmap_sem
510 down_read(&mm
->mmap_sem
);
511 ret
= VM_FAULT_NOPAGE
;
516 * Here we race with the list_del; list_add in
517 * userfaultfd_ctx_read(), however because we don't ever run
518 * list_del_init() to refile across the two lists, the prev
519 * and next pointers will never point to self. list_add also
520 * would never let any of the two pointers to point to
521 * self. So list_empty_careful won't risk to see both pointers
522 * pointing to self at any time during the list refile. The
523 * only case where list_del_init() is called is the full
524 * removal in the wake function and there we don't re-list_add
525 * and it's fine not to block on the spinlock. The uwq on this
526 * kernel stack can be released after the list_del_init.
528 if (!list_empty_careful(&uwq
.wq
.entry
)) {
529 spin_lock(&ctx
->fault_pending_wqh
.lock
);
531 * No need of list_del_init(), the uwq on the stack
532 * will be freed shortly anyway.
534 list_del(&uwq
.wq
.entry
);
535 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
539 * ctx may go away after this if the userfault pseudo fd is
542 userfaultfd_ctx_put(ctx
);
548 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
549 struct userfaultfd_wait_queue
*ewq
)
551 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
555 init_waitqueue_entry(&ewq
->wq
, current
);
557 spin_lock(&ctx
->event_wqh
.lock
);
559 * After the __add_wait_queue the uwq is visible to userland
560 * through poll/read().
562 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
564 set_current_state(TASK_KILLABLE
);
565 if (ewq
->msg
.event
== 0)
567 if (ACCESS_ONCE(ctx
->released
) ||
568 fatal_signal_pending(current
)) {
570 * &ewq->wq may be queued in fork_event, but
571 * __remove_wait_queue ignores the head
572 * parameter. It would be a problem if it
575 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
576 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
577 struct userfaultfd_ctx
*new;
579 new = (struct userfaultfd_ctx
*)
581 ewq
->msg
.arg
.reserved
.reserved1
;
583 userfaultfd_ctx_put(new);
588 spin_unlock(&ctx
->event_wqh
.lock
);
590 wake_up_poll(&ctx
->fd_wqh
, POLLIN
);
593 spin_lock(&ctx
->event_wqh
.lock
);
595 __set_current_state(TASK_RUNNING
);
596 spin_unlock(&ctx
->event_wqh
.lock
);
599 * ctx may go away after this if the userfault pseudo fd is
603 userfaultfd_ctx_put(ctx
);
606 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
607 struct userfaultfd_wait_queue
*ewq
)
610 wake_up_locked(&ctx
->event_wqh
);
611 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
614 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
616 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
617 struct userfaultfd_fork_ctx
*fctx
;
619 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
620 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
621 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
622 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
626 list_for_each_entry(fctx
, fcs
, list
)
627 if (fctx
->orig
== octx
) {
633 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
637 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
643 atomic_set(&ctx
->refcount
, 1);
644 ctx
->flags
= octx
->flags
;
645 ctx
->state
= UFFD_STATE_RUNNING
;
646 ctx
->features
= octx
->features
;
647 ctx
->released
= false;
648 ctx
->mm
= vma
->vm_mm
;
649 atomic_inc(&ctx
->mm
->mm_count
);
651 userfaultfd_ctx_get(octx
);
654 list_add_tail(&fctx
->list
, fcs
);
657 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
661 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
663 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
664 struct userfaultfd_wait_queue ewq
;
668 ewq
.msg
.event
= UFFD_EVENT_FORK
;
669 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
671 userfaultfd_event_wait_completion(ctx
, &ewq
);
674 void dup_userfaultfd_complete(struct list_head
*fcs
)
676 struct userfaultfd_fork_ctx
*fctx
, *n
;
678 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
680 list_del(&fctx
->list
);
685 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
686 struct vm_userfaultfd_ctx
*vm_ctx
)
688 struct userfaultfd_ctx
*ctx
;
690 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
691 if (ctx
&& (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
)) {
693 userfaultfd_ctx_get(ctx
);
697 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
698 unsigned long from
, unsigned long to
,
701 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
702 struct userfaultfd_wait_queue ewq
;
707 if (to
& ~PAGE_MASK
) {
708 userfaultfd_ctx_put(ctx
);
714 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
715 ewq
.msg
.arg
.remap
.from
= from
;
716 ewq
.msg
.arg
.remap
.to
= to
;
717 ewq
.msg
.arg
.remap
.len
= len
;
719 userfaultfd_event_wait_completion(ctx
, &ewq
);
722 bool userfaultfd_remove(struct vm_area_struct
*vma
,
723 unsigned long start
, unsigned long end
)
725 struct mm_struct
*mm
= vma
->vm_mm
;
726 struct userfaultfd_ctx
*ctx
;
727 struct userfaultfd_wait_queue ewq
;
729 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
730 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
733 userfaultfd_ctx_get(ctx
);
734 up_read(&mm
->mmap_sem
);
738 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
739 ewq
.msg
.arg
.remove
.start
= start
;
740 ewq
.msg
.arg
.remove
.end
= end
;
742 userfaultfd_event_wait_completion(ctx
, &ewq
);
747 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
748 unsigned long start
, unsigned long end
)
750 struct userfaultfd_unmap_ctx
*unmap_ctx
;
752 list_for_each_entry(unmap_ctx
, unmaps
, list
)
753 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
754 unmap_ctx
->end
== end
)
760 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
761 unsigned long start
, unsigned long end
,
762 struct list_head
*unmaps
)
764 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
765 struct userfaultfd_unmap_ctx
*unmap_ctx
;
766 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
768 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
769 has_unmap_ctx(ctx
, unmaps
, start
, end
))
772 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
776 userfaultfd_ctx_get(ctx
);
777 unmap_ctx
->ctx
= ctx
;
778 unmap_ctx
->start
= start
;
779 unmap_ctx
->end
= end
;
780 list_add_tail(&unmap_ctx
->list
, unmaps
);
786 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
788 struct userfaultfd_unmap_ctx
*ctx
, *n
;
789 struct userfaultfd_wait_queue ewq
;
791 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
794 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
795 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
796 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
798 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
800 list_del(&ctx
->list
);
805 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
807 struct userfaultfd_ctx
*ctx
= file
->private_data
;
808 struct mm_struct
*mm
= ctx
->mm
;
809 struct vm_area_struct
*vma
, *prev
;
810 /* len == 0 means wake all */
811 struct userfaultfd_wake_range range
= { .len
= 0, };
812 unsigned long new_flags
;
814 ACCESS_ONCE(ctx
->released
) = true;
816 if (!mmget_not_zero(mm
))
820 * Flush page faults out of all CPUs. NOTE: all page faults
821 * must be retried without returning VM_FAULT_SIGBUS if
822 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
823 * changes while handle_userfault released the mmap_sem. So
824 * it's critical that released is set to true (above), before
825 * taking the mmap_sem for writing.
827 down_write(&mm
->mmap_sem
);
829 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
831 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
832 !!(vma
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
833 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
837 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
838 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
839 new_flags
, vma
->anon_vma
,
840 vma
->vm_file
, vma
->vm_pgoff
,
847 vma
->vm_flags
= new_flags
;
848 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
850 up_write(&mm
->mmap_sem
);
854 * After no new page faults can wait on this fault_*wqh, flush
855 * the last page faults that may have been already waiting on
858 spin_lock(&ctx
->fault_pending_wqh
.lock
);
859 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
860 __wake_up_locked_key(&ctx
->fault_wqh
, TASK_NORMAL
, &range
);
861 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
863 /* Flush pending events that may still wait on event_wqh */
864 wake_up_all(&ctx
->event_wqh
);
866 wake_up_poll(&ctx
->fd_wqh
, POLLHUP
);
867 userfaultfd_ctx_put(ctx
);
871 /* fault_pending_wqh.lock must be hold by the caller */
872 static inline struct userfaultfd_wait_queue
*find_userfault_in(
873 wait_queue_head_t
*wqh
)
875 wait_queue_entry_t
*wq
;
876 struct userfaultfd_wait_queue
*uwq
;
878 VM_BUG_ON(!spin_is_locked(&wqh
->lock
));
881 if (!waitqueue_active(wqh
))
883 /* walk in reverse to provide FIFO behavior to read userfaults */
884 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
885 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
890 static inline struct userfaultfd_wait_queue
*find_userfault(
891 struct userfaultfd_ctx
*ctx
)
893 return find_userfault_in(&ctx
->fault_pending_wqh
);
896 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
897 struct userfaultfd_ctx
*ctx
)
899 return find_userfault_in(&ctx
->event_wqh
);
902 static unsigned int userfaultfd_poll(struct file
*file
, poll_table
*wait
)
904 struct userfaultfd_ctx
*ctx
= file
->private_data
;
907 poll_wait(file
, &ctx
->fd_wqh
, wait
);
909 switch (ctx
->state
) {
910 case UFFD_STATE_WAIT_API
:
912 case UFFD_STATE_RUNNING
:
914 * poll() never guarantees that read won't block.
915 * userfaults can be waken before they're read().
917 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
920 * lockless access to see if there are pending faults
921 * __pollwait last action is the add_wait_queue but
922 * the spin_unlock would allow the waitqueue_active to
923 * pass above the actual list_add inside
924 * add_wait_queue critical section. So use a full
925 * memory barrier to serialize the list_add write of
926 * add_wait_queue() with the waitqueue_active read
931 if (waitqueue_active(&ctx
->fault_pending_wqh
))
933 else if (waitqueue_active(&ctx
->event_wqh
))
943 static const struct file_operations userfaultfd_fops
;
945 static int resolve_userfault_fork(struct userfaultfd_ctx
*ctx
,
946 struct userfaultfd_ctx
*new,
947 struct uffd_msg
*msg
)
951 unsigned int flags
= new->flags
& UFFD_SHARED_FCNTL_FLAGS
;
953 fd
= get_unused_fd_flags(flags
);
957 file
= anon_inode_getfile("[userfaultfd]", &userfaultfd_fops
, new,
961 return PTR_ERR(file
);
964 fd_install(fd
, file
);
965 msg
->arg
.reserved
.reserved1
= 0;
966 msg
->arg
.fork
.ufd
= fd
;
971 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
972 struct uffd_msg
*msg
)
975 DECLARE_WAITQUEUE(wait
, current
);
976 struct userfaultfd_wait_queue
*uwq
;
978 * Handling fork event requires sleeping operations, so
979 * we drop the event_wqh lock, then do these ops, then
980 * lock it back and wake up the waiter. While the lock is
981 * dropped the ewq may go away so we keep track of it
984 LIST_HEAD(fork_event
);
985 struct userfaultfd_ctx
*fork_nctx
= NULL
;
987 /* always take the fd_wqh lock before the fault_pending_wqh lock */
988 spin_lock(&ctx
->fd_wqh
.lock
);
989 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
991 set_current_state(TASK_INTERRUPTIBLE
);
992 spin_lock(&ctx
->fault_pending_wqh
.lock
);
993 uwq
= find_userfault(ctx
);
996 * Use a seqcount to repeat the lockless check
997 * in wake_userfault() to avoid missing
998 * wakeups because during the refile both
999 * waitqueue could become empty if this is the
1002 write_seqcount_begin(&ctx
->refile_seq
);
1005 * The fault_pending_wqh.lock prevents the uwq
1006 * to disappear from under us.
1008 * Refile this userfault from
1009 * fault_pending_wqh to fault_wqh, it's not
1010 * pending anymore after we read it.
1012 * Use list_del() by hand (as
1013 * userfaultfd_wake_function also uses
1014 * list_del_init() by hand) to be sure nobody
1015 * changes __remove_wait_queue() to use
1016 * list_del_init() in turn breaking the
1017 * !list_empty_careful() check in
1018 * handle_userfault(). The uwq->wq.head list
1019 * must never be empty at any time during the
1020 * refile, or the waitqueue could disappear
1021 * from under us. The "wait_queue_head_t"
1022 * parameter of __remove_wait_queue() is unused
1025 list_del(&uwq
->wq
.entry
);
1026 __add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1028 write_seqcount_end(&ctx
->refile_seq
);
1030 /* careful to always initialize msg if ret == 0 */
1032 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1036 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1038 spin_lock(&ctx
->event_wqh
.lock
);
1039 uwq
= find_userfault_evt(ctx
);
1043 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1044 fork_nctx
= (struct userfaultfd_ctx
*)
1046 uwq
->msg
.arg
.reserved
.reserved1
;
1047 list_move(&uwq
->wq
.entry
, &fork_event
);
1049 * fork_nctx can be freed as soon as
1050 * we drop the lock, unless we take a
1053 userfaultfd_ctx_get(fork_nctx
);
1054 spin_unlock(&ctx
->event_wqh
.lock
);
1059 userfaultfd_event_complete(ctx
, uwq
);
1060 spin_unlock(&ctx
->event_wqh
.lock
);
1064 spin_unlock(&ctx
->event_wqh
.lock
);
1066 if (signal_pending(current
)) {
1074 spin_unlock(&ctx
->fd_wqh
.lock
);
1076 spin_lock(&ctx
->fd_wqh
.lock
);
1078 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1079 __set_current_state(TASK_RUNNING
);
1080 spin_unlock(&ctx
->fd_wqh
.lock
);
1082 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1083 ret
= resolve_userfault_fork(ctx
, fork_nctx
, msg
);
1084 spin_lock(&ctx
->event_wqh
.lock
);
1085 if (!list_empty(&fork_event
)) {
1087 * The fork thread didn't abort, so we can
1088 * drop the temporary refcount.
1090 userfaultfd_ctx_put(fork_nctx
);
1092 uwq
= list_first_entry(&fork_event
,
1096 * If fork_event list wasn't empty and in turn
1097 * the event wasn't already released by fork
1098 * (the event is allocated on fork kernel
1099 * stack), put the event back to its place in
1100 * the event_wq. fork_event head will be freed
1101 * as soon as we return so the event cannot
1102 * stay queued there no matter the current
1105 list_del(&uwq
->wq
.entry
);
1106 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1109 * Leave the event in the waitqueue and report
1110 * error to userland if we failed to resolve
1111 * the userfault fork.
1114 userfaultfd_event_complete(ctx
, uwq
);
1117 * Here the fork thread aborted and the
1118 * refcount from the fork thread on fork_nctx
1119 * has already been released. We still hold
1120 * the reference we took before releasing the
1121 * lock above. If resolve_userfault_fork
1122 * failed we've to drop it because the
1123 * fork_nctx has to be freed in such case. If
1124 * it succeeded we'll hold it because the new
1125 * uffd references it.
1128 userfaultfd_ctx_put(fork_nctx
);
1130 spin_unlock(&ctx
->event_wqh
.lock
);
1136 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1137 size_t count
, loff_t
*ppos
)
1139 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1140 ssize_t _ret
, ret
= 0;
1141 struct uffd_msg msg
;
1142 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1144 if (ctx
->state
== UFFD_STATE_WAIT_API
)
1148 if (count
< sizeof(msg
))
1149 return ret
? ret
: -EINVAL
;
1150 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
);
1152 return ret
? ret
: _ret
;
1153 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1154 return ret
? ret
: -EFAULT
;
1157 count
-= sizeof(msg
);
1159 * Allow to read more than one fault at time but only
1160 * block if waiting for the very first one.
1162 no_wait
= O_NONBLOCK
;
1166 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1167 struct userfaultfd_wake_range
*range
)
1169 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1170 /* wake all in the range and autoremove */
1171 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1172 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1174 if (waitqueue_active(&ctx
->fault_wqh
))
1175 __wake_up_locked_key(&ctx
->fault_wqh
, TASK_NORMAL
, range
);
1176 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1179 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1180 struct userfaultfd_wake_range
*range
)
1186 * To be sure waitqueue_active() is not reordered by the CPU
1187 * before the pagetable update, use an explicit SMP memory
1188 * barrier here. PT lock release or up_read(mmap_sem) still
1189 * have release semantics that can allow the
1190 * waitqueue_active() to be reordered before the pte update.
1195 * Use waitqueue_active because it's very frequent to
1196 * change the address space atomically even if there are no
1197 * userfaults yet. So we take the spinlock only when we're
1198 * sure we've userfaults to wake.
1201 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1202 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1203 waitqueue_active(&ctx
->fault_wqh
);
1205 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1207 __wake_userfault(ctx
, range
);
1210 static __always_inline
int validate_range(struct mm_struct
*mm
,
1211 __u64 start
, __u64 len
)
1213 __u64 task_size
= mm
->task_size
;
1215 if (start
& ~PAGE_MASK
)
1217 if (len
& ~PAGE_MASK
)
1221 if (start
< mmap_min_addr
)
1223 if (start
>= task_size
)
1225 if (len
> task_size
- start
)
1230 static inline bool vma_can_userfault(struct vm_area_struct
*vma
)
1232 return vma_is_anonymous(vma
) || is_vm_hugetlb_page(vma
) ||
1236 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1239 struct mm_struct
*mm
= ctx
->mm
;
1240 struct vm_area_struct
*vma
, *prev
, *cur
;
1242 struct uffdio_register uffdio_register
;
1243 struct uffdio_register __user
*user_uffdio_register
;
1244 unsigned long vm_flags
, new_flags
;
1246 bool non_anon_pages
;
1247 unsigned long start
, end
, vma_end
;
1249 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1252 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1253 sizeof(uffdio_register
)-sizeof(__u64
)))
1257 if (!uffdio_register
.mode
)
1259 if (uffdio_register
.mode
& ~(UFFDIO_REGISTER_MODE_MISSING
|
1260 UFFDIO_REGISTER_MODE_WP
))
1263 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1264 vm_flags
|= VM_UFFD_MISSING
;
1265 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1266 vm_flags
|= VM_UFFD_WP
;
1268 * FIXME: remove the below error constraint by
1269 * implementing the wprotect tracking mode.
1275 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1276 uffdio_register
.range
.len
);
1280 start
= uffdio_register
.range
.start
;
1281 end
= start
+ uffdio_register
.range
.len
;
1284 if (!mmget_not_zero(mm
))
1287 down_write(&mm
->mmap_sem
);
1288 vma
= find_vma_prev(mm
, start
, &prev
);
1292 /* check that there's at least one vma in the range */
1294 if (vma
->vm_start
>= end
)
1298 * If the first vma contains huge pages, make sure start address
1299 * is aligned to huge page size.
1301 if (is_vm_hugetlb_page(vma
)) {
1302 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1304 if (start
& (vma_hpagesize
- 1))
1309 * Search for not compatible vmas.
1312 non_anon_pages
= false;
1313 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1316 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1317 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1319 /* check not compatible vmas */
1321 if (!vma_can_userfault(cur
))
1324 * If this vma contains ending address, and huge pages
1327 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1328 end
> cur
->vm_start
) {
1329 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1333 if (end
& (vma_hpagesize
- 1))
1338 * Check that this vma isn't already owned by a
1339 * different userfaultfd. We can't allow more than one
1340 * userfaultfd to own a single vma simultaneously or we
1341 * wouldn't know which one to deliver the userfaults to.
1344 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1345 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1349 * Note vmas containing huge pages
1351 if (is_vm_hugetlb_page(cur
) || vma_is_shmem(cur
))
1352 non_anon_pages
= true;
1358 if (vma
->vm_start
< start
)
1365 BUG_ON(!vma_can_userfault(vma
));
1366 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1367 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1370 * Nothing to do: this vma is already registered into this
1371 * userfaultfd and with the right tracking mode too.
1373 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1374 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1377 if (vma
->vm_start
> start
)
1378 start
= vma
->vm_start
;
1379 vma_end
= min(end
, vma
->vm_end
);
1381 new_flags
= (vma
->vm_flags
& ~vm_flags
) | vm_flags
;
1382 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1383 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1385 ((struct vm_userfaultfd_ctx
){ ctx
}));
1390 if (vma
->vm_start
< start
) {
1391 ret
= split_vma(mm
, vma
, start
, 1);
1395 if (vma
->vm_end
> end
) {
1396 ret
= split_vma(mm
, vma
, end
, 0);
1402 * In the vma_merge() successful mprotect-like case 8:
1403 * the next vma was merged into the current one and
1404 * the current one has not been updated yet.
1406 vma
->vm_flags
= new_flags
;
1407 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1411 start
= vma
->vm_end
;
1413 } while (vma
&& vma
->vm_start
< end
);
1415 up_write(&mm
->mmap_sem
);
1419 * Now that we scanned all vmas we can already tell
1420 * userland which ioctls methods are guaranteed to
1421 * succeed on this range.
1423 if (put_user(non_anon_pages
? UFFD_API_RANGE_IOCTLS_BASIC
:
1424 UFFD_API_RANGE_IOCTLS
,
1425 &user_uffdio_register
->ioctls
))
1432 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1435 struct mm_struct
*mm
= ctx
->mm
;
1436 struct vm_area_struct
*vma
, *prev
, *cur
;
1438 struct uffdio_range uffdio_unregister
;
1439 unsigned long new_flags
;
1441 unsigned long start
, end
, vma_end
;
1442 const void __user
*buf
= (void __user
*)arg
;
1445 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1448 ret
= validate_range(mm
, uffdio_unregister
.start
,
1449 uffdio_unregister
.len
);
1453 start
= uffdio_unregister
.start
;
1454 end
= start
+ uffdio_unregister
.len
;
1457 if (!mmget_not_zero(mm
))
1460 down_write(&mm
->mmap_sem
);
1461 vma
= find_vma_prev(mm
, start
, &prev
);
1465 /* check that there's at least one vma in the range */
1467 if (vma
->vm_start
>= end
)
1471 * If the first vma contains huge pages, make sure start address
1472 * is aligned to huge page size.
1474 if (is_vm_hugetlb_page(vma
)) {
1475 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1477 if (start
& (vma_hpagesize
- 1))
1482 * Search for not compatible vmas.
1486 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1489 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1490 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1493 * Check not compatible vmas, not strictly required
1494 * here as not compatible vmas cannot have an
1495 * userfaultfd_ctx registered on them, but this
1496 * provides for more strict behavior to notice
1497 * unregistration errors.
1499 if (!vma_can_userfault(cur
))
1506 if (vma
->vm_start
< start
)
1513 BUG_ON(!vma_can_userfault(vma
));
1516 * Nothing to do: this vma is already registered into this
1517 * userfaultfd and with the right tracking mode too.
1519 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1522 if (vma
->vm_start
> start
)
1523 start
= vma
->vm_start
;
1524 vma_end
= min(end
, vma
->vm_end
);
1526 if (userfaultfd_missing(vma
)) {
1528 * Wake any concurrent pending userfault while
1529 * we unregister, so they will not hang
1530 * permanently and it avoids userland to call
1531 * UFFDIO_WAKE explicitly.
1533 struct userfaultfd_wake_range range
;
1534 range
.start
= start
;
1535 range
.len
= vma_end
- start
;
1536 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1539 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
1540 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1541 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1548 if (vma
->vm_start
< start
) {
1549 ret
= split_vma(mm
, vma
, start
, 1);
1553 if (vma
->vm_end
> end
) {
1554 ret
= split_vma(mm
, vma
, end
, 0);
1560 * In the vma_merge() successful mprotect-like case 8:
1561 * the next vma was merged into the current one and
1562 * the current one has not been updated yet.
1564 vma
->vm_flags
= new_flags
;
1565 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1569 start
= vma
->vm_end
;
1571 } while (vma
&& vma
->vm_start
< end
);
1573 up_write(&mm
->mmap_sem
);
1580 * userfaultfd_wake may be used in combination with the
1581 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1583 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1587 struct uffdio_range uffdio_wake
;
1588 struct userfaultfd_wake_range range
;
1589 const void __user
*buf
= (void __user
*)arg
;
1592 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1595 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1599 range
.start
= uffdio_wake
.start
;
1600 range
.len
= uffdio_wake
.len
;
1603 * len == 0 means wake all and we don't want to wake all here,
1604 * so check it again to be sure.
1606 VM_BUG_ON(!range
.len
);
1608 wake_userfault(ctx
, &range
);
1615 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1619 struct uffdio_copy uffdio_copy
;
1620 struct uffdio_copy __user
*user_uffdio_copy
;
1621 struct userfaultfd_wake_range range
;
1623 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1626 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1627 /* don't copy "copy" last field */
1628 sizeof(uffdio_copy
)-sizeof(__s64
)))
1631 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1635 * double check for wraparound just in case. copy_from_user()
1636 * will later check uffdio_copy.src + uffdio_copy.len to fit
1637 * in the userland range.
1640 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1642 if (uffdio_copy
.mode
& ~UFFDIO_COPY_MODE_DONTWAKE
)
1644 if (mmget_not_zero(ctx
->mm
)) {
1645 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1651 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1656 /* len == 0 would wake all */
1658 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1659 range
.start
= uffdio_copy
.dst
;
1660 wake_userfault(ctx
, &range
);
1662 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1667 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1671 struct uffdio_zeropage uffdio_zeropage
;
1672 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1673 struct userfaultfd_wake_range range
;
1675 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1678 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1679 /* don't copy "zeropage" last field */
1680 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1683 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1684 uffdio_zeropage
.range
.len
);
1688 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1691 if (mmget_not_zero(ctx
->mm
)) {
1692 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1693 uffdio_zeropage
.range
.len
);
1698 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1702 /* len == 0 would wake all */
1705 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1706 range
.start
= uffdio_zeropage
.range
.start
;
1707 wake_userfault(ctx
, &range
);
1709 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1714 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1717 * For the current set of features the bits just coincide
1719 return (unsigned int)user_features
;
1723 * userland asks for a certain API version and we return which bits
1724 * and ioctl commands are implemented in this kernel for such API
1725 * version or -EINVAL if unknown.
1727 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1730 struct uffdio_api uffdio_api
;
1731 void __user
*buf
= (void __user
*)arg
;
1736 if (ctx
->state
!= UFFD_STATE_WAIT_API
)
1739 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1741 features
= uffdio_api
.features
;
1742 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
)) {
1743 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1744 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1749 /* report all available features and ioctls to userland */
1750 uffdio_api
.features
= UFFD_API_FEATURES
;
1751 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1753 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1755 ctx
->state
= UFFD_STATE_RUNNING
;
1756 /* only enable the requested features for this uffd context */
1757 ctx
->features
= uffd_ctx_features(features
);
1763 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1767 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1769 if (cmd
!= UFFDIO_API
&& ctx
->state
== UFFD_STATE_WAIT_API
)
1774 ret
= userfaultfd_api(ctx
, arg
);
1776 case UFFDIO_REGISTER
:
1777 ret
= userfaultfd_register(ctx
, arg
);
1779 case UFFDIO_UNREGISTER
:
1780 ret
= userfaultfd_unregister(ctx
, arg
);
1783 ret
= userfaultfd_wake(ctx
, arg
);
1786 ret
= userfaultfd_copy(ctx
, arg
);
1788 case UFFDIO_ZEROPAGE
:
1789 ret
= userfaultfd_zeropage(ctx
, arg
);
1795 #ifdef CONFIG_PROC_FS
1796 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
1798 struct userfaultfd_ctx
*ctx
= f
->private_data
;
1799 wait_queue_entry_t
*wq
;
1800 struct userfaultfd_wait_queue
*uwq
;
1801 unsigned long pending
= 0, total
= 0;
1803 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1804 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
1805 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
1809 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
1810 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
1813 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1816 * If more protocols will be added, there will be all shown
1817 * separated by a space. Like this:
1818 * protocols: aa:... bb:...
1820 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1821 pending
, total
, UFFD_API
, ctx
->features
,
1822 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
1826 static const struct file_operations userfaultfd_fops
= {
1827 #ifdef CONFIG_PROC_FS
1828 .show_fdinfo
= userfaultfd_show_fdinfo
,
1830 .release
= userfaultfd_release
,
1831 .poll
= userfaultfd_poll
,
1832 .read
= userfaultfd_read
,
1833 .unlocked_ioctl
= userfaultfd_ioctl
,
1834 .compat_ioctl
= userfaultfd_ioctl
,
1835 .llseek
= noop_llseek
,
1838 static void init_once_userfaultfd_ctx(void *mem
)
1840 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
1842 init_waitqueue_head(&ctx
->fault_pending_wqh
);
1843 init_waitqueue_head(&ctx
->fault_wqh
);
1844 init_waitqueue_head(&ctx
->event_wqh
);
1845 init_waitqueue_head(&ctx
->fd_wqh
);
1846 seqcount_init(&ctx
->refile_seq
);
1850 * userfaultfd_file_create - Creates a userfaultfd file pointer.
1851 * @flags: Flags for the userfaultfd file.
1853 * This function creates a userfaultfd file pointer, w/out installing
1854 * it into the fd table. This is useful when the userfaultfd file is
1855 * used during the initialization of data structures that require
1856 * extra setup after the userfaultfd creation. So the userfaultfd
1857 * creation is split into the file pointer creation phase, and the
1858 * file descriptor installation phase. In this way races with
1859 * userspace closing the newly installed file descriptor can be
1860 * avoided. Returns a userfaultfd file pointer, or a proper error
1863 static struct file
*userfaultfd_file_create(int flags
)
1866 struct userfaultfd_ctx
*ctx
;
1868 BUG_ON(!current
->mm
);
1870 /* Check the UFFD_* constants for consistency. */
1871 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
1872 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
1874 file
= ERR_PTR(-EINVAL
);
1875 if (flags
& ~UFFD_SHARED_FCNTL_FLAGS
)
1878 file
= ERR_PTR(-ENOMEM
);
1879 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
1883 atomic_set(&ctx
->refcount
, 1);
1886 ctx
->state
= UFFD_STATE_WAIT_API
;
1887 ctx
->released
= false;
1888 ctx
->mm
= current
->mm
;
1889 /* prevent the mm struct to be freed */
1892 file
= anon_inode_getfile("[userfaultfd]", &userfaultfd_fops
, ctx
,
1893 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
));
1896 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
1902 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
1907 error
= get_unused_fd_flags(flags
& UFFD_SHARED_FCNTL_FLAGS
);
1912 file
= userfaultfd_file_create(flags
);
1914 error
= PTR_ERR(file
);
1915 goto err_put_unused_fd
;
1917 fd_install(fd
, file
);
1927 static int __init
userfaultfd_init(void)
1929 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
1930 sizeof(struct userfaultfd_ctx
),
1932 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
1933 init_once_userfaultfd_ctx
);
1936 __initcall(userfaultfd_init
);