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
{
85 struct userfaultfd_ctx
*ctx
;
89 struct userfaultfd_wake_range
{
94 static int userfaultfd_wake_function(wait_queue_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
->task_list
);
138 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
140 * @ctx: [in] Pointer to the userfaultfd context.
142 * Returns: In case of success, returns not zero.
144 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
146 if (!atomic_inc_not_zero(&ctx
->refcount
))
151 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
153 * @ctx: [in] Pointer to userfaultfd context.
155 * The userfaultfd context reference must have been previously acquired either
156 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
158 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
160 if (atomic_dec_and_test(&ctx
->refcount
)) {
161 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
162 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
163 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
164 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
165 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
166 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
167 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
168 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
170 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
174 static inline void msg_init(struct uffd_msg
*msg
)
176 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
178 * Must use memset to zero out the paddings or kernel data is
179 * leaked to userland.
181 memset(msg
, 0, sizeof(struct uffd_msg
));
184 static inline struct uffd_msg
userfault_msg(unsigned long address
,
186 unsigned long reason
)
190 msg
.event
= UFFD_EVENT_PAGEFAULT
;
191 msg
.arg
.pagefault
.address
= address
;
192 if (flags
& FAULT_FLAG_WRITE
)
194 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
195 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
196 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
197 * was a read fault, otherwise if set it means it's
200 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
201 if (reason
& VM_UFFD_WP
)
203 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
204 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
205 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
206 * a missing fault, otherwise if set it means it's a
207 * write protect fault.
209 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
213 #ifdef CONFIG_HUGETLB_PAGE
215 * Same functionality as userfaultfd_must_wait below with modifications for
218 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
219 unsigned long address
,
221 unsigned long reason
)
223 struct mm_struct
*mm
= ctx
->mm
;
227 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
229 pte
= huge_pte_offset(mm
, address
);
236 * Lockless access: we're in a wait_event so it's ok if it
239 if (huge_pte_none(*pte
))
241 if (!huge_pte_write(*pte
) && (reason
& VM_UFFD_WP
))
247 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
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
;
275 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
277 pgd
= pgd_offset(mm
, address
);
278 if (!pgd_present(*pgd
))
280 pud
= pud_offset(pgd
, address
);
281 if (!pud_present(*pud
))
283 pmd
= pmd_offset(pud
, address
);
285 * READ_ONCE must function as a barrier with narrower scope
286 * and it must be equivalent to:
287 * _pmd = *pmd; barrier();
289 * This is to deal with the instability (as in
290 * pmd_trans_unstable) of the pmd.
292 _pmd
= READ_ONCE(*pmd
);
293 if (!pmd_present(_pmd
))
297 if (pmd_trans_huge(_pmd
))
301 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
302 * and use the standard pte_offset_map() instead of parsing _pmd.
304 pte
= pte_offset_map(pmd
, address
);
306 * Lockless access: we're in a wait_event so it's ok if it
318 * The locking rules involved in returning VM_FAULT_RETRY depending on
319 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
320 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
321 * recommendation in __lock_page_or_retry is not an understatement.
323 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
324 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
327 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
328 * set, VM_FAULT_RETRY can still be returned if and only if there are
329 * fatal_signal_pending()s, and the mmap_sem must be released before
332 int handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
334 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
335 struct userfaultfd_ctx
*ctx
;
336 struct userfaultfd_wait_queue uwq
;
338 bool must_wait
, return_to_userland
;
341 BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
343 ret
= VM_FAULT_SIGBUS
;
344 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
348 BUG_ON(ctx
->mm
!= mm
);
350 VM_BUG_ON(reason
& ~(VM_UFFD_MISSING
|VM_UFFD_WP
));
351 VM_BUG_ON(!(reason
& VM_UFFD_MISSING
) ^ !!(reason
& VM_UFFD_WP
));
354 * If it's already released don't get it. This avoids to loop
355 * in __get_user_pages if userfaultfd_release waits on the
356 * caller of handle_userfault to release the mmap_sem.
358 if (unlikely(ACCESS_ONCE(ctx
->released
)))
362 * We don't do userfault handling for the final child pid update.
364 if (current
->flags
& PF_EXITING
)
368 * Check that we can return VM_FAULT_RETRY.
370 * NOTE: it should become possible to return VM_FAULT_RETRY
371 * even if FAULT_FLAG_TRIED is set without leading to gup()
372 * -EBUSY failures, if the userfaultfd is to be extended for
373 * VM_UFFD_WP tracking and we intend to arm the userfault
374 * without first stopping userland access to the memory. For
375 * VM_UFFD_MISSING userfaults this is enough for now.
377 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
379 * Validate the invariant that nowait must allow retry
380 * to be sure not to return SIGBUS erroneously on
381 * nowait invocations.
383 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
384 #ifdef CONFIG_DEBUG_VM
385 if (printk_ratelimit()) {
387 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
396 * Handle nowait, not much to do other than tell it to retry
399 ret
= VM_FAULT_RETRY
;
400 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
403 /* take the reference before dropping the mmap_sem */
404 userfaultfd_ctx_get(ctx
);
406 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
407 uwq
.wq
.private = current
;
408 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
);
413 (vmf
->flags
& (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
)) ==
414 (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
);
415 blocking_state
= return_to_userland
? TASK_INTERRUPTIBLE
:
418 spin_lock(&ctx
->fault_pending_wqh
.lock
);
420 * After the __add_wait_queue the uwq is visible to userland
421 * through poll/read().
423 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
425 * The smp_mb() after __set_current_state prevents the reads
426 * following the spin_unlock to happen before the list_add in
429 set_current_state(blocking_state
);
430 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
432 if (!is_vm_hugetlb_page(vmf
->vma
))
433 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
436 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->address
,
438 up_read(&mm
->mmap_sem
);
440 if (likely(must_wait
&& !ACCESS_ONCE(ctx
->released
) &&
441 (return_to_userland
? !signal_pending(current
) :
442 !fatal_signal_pending(current
)))) {
443 wake_up_poll(&ctx
->fd_wqh
, POLLIN
);
445 ret
|= VM_FAULT_MAJOR
;
448 * False wakeups can orginate even from rwsem before
449 * up_read() however userfaults will wait either for a
450 * targeted wakeup on the specific uwq waitqueue from
451 * wake_userfault() or for signals or for uffd
454 while (!READ_ONCE(uwq
.waken
)) {
456 * This needs the full smp_store_mb()
457 * guarantee as the state write must be
458 * visible to other CPUs before reading
459 * uwq.waken from other CPUs.
461 set_current_state(blocking_state
);
462 if (READ_ONCE(uwq
.waken
) ||
463 READ_ONCE(ctx
->released
) ||
464 (return_to_userland
? signal_pending(current
) :
465 fatal_signal_pending(current
)))
471 __set_current_state(TASK_RUNNING
);
473 if (return_to_userland
) {
474 if (signal_pending(current
) &&
475 !fatal_signal_pending(current
)) {
477 * If we got a SIGSTOP or SIGCONT and this is
478 * a normal userland page fault, just let
479 * userland return so the signal will be
480 * handled and gdb debugging works. The page
481 * fault code immediately after we return from
482 * this function is going to release the
483 * mmap_sem and it's not depending on it
484 * (unlike gup would if we were not to return
487 * If a fatal signal is pending we still take
488 * the streamlined VM_FAULT_RETRY failure path
489 * and there's no need to retake the mmap_sem
492 down_read(&mm
->mmap_sem
);
498 * Here we race with the list_del; list_add in
499 * userfaultfd_ctx_read(), however because we don't ever run
500 * list_del_init() to refile across the two lists, the prev
501 * and next pointers will never point to self. list_add also
502 * would never let any of the two pointers to point to
503 * self. So list_empty_careful won't risk to see both pointers
504 * pointing to self at any time during the list refile. The
505 * only case where list_del_init() is called is the full
506 * removal in the wake function and there we don't re-list_add
507 * and it's fine not to block on the spinlock. The uwq on this
508 * kernel stack can be released after the list_del_init.
510 if (!list_empty_careful(&uwq
.wq
.task_list
)) {
511 spin_lock(&ctx
->fault_pending_wqh
.lock
);
513 * No need of list_del_init(), the uwq on the stack
514 * will be freed shortly anyway.
516 list_del(&uwq
.wq
.task_list
);
517 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
521 * ctx may go away after this if the userfault pseudo fd is
524 userfaultfd_ctx_put(ctx
);
530 static int userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
531 struct userfaultfd_wait_queue
*ewq
)
536 init_waitqueue_entry(&ewq
->wq
, current
);
538 spin_lock(&ctx
->event_wqh
.lock
);
540 * After the __add_wait_queue the uwq is visible to userland
541 * through poll/read().
543 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
545 set_current_state(TASK_KILLABLE
);
546 if (ewq
->msg
.event
== 0)
548 if (ACCESS_ONCE(ctx
->released
) ||
549 fatal_signal_pending(current
)) {
551 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
555 spin_unlock(&ctx
->event_wqh
.lock
);
557 wake_up_poll(&ctx
->fd_wqh
, POLLIN
);
560 spin_lock(&ctx
->event_wqh
.lock
);
562 __set_current_state(TASK_RUNNING
);
563 spin_unlock(&ctx
->event_wqh
.lock
);
566 * ctx may go away after this if the userfault pseudo fd is
570 userfaultfd_ctx_put(ctx
);
574 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
575 struct userfaultfd_wait_queue
*ewq
)
578 wake_up_locked(&ctx
->event_wqh
);
579 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
582 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
584 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
585 struct userfaultfd_fork_ctx
*fctx
;
587 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
588 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
589 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
590 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
594 list_for_each_entry(fctx
, fcs
, list
)
595 if (fctx
->orig
== octx
) {
601 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
605 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
611 atomic_set(&ctx
->refcount
, 1);
612 ctx
->flags
= octx
->flags
;
613 ctx
->state
= UFFD_STATE_RUNNING
;
614 ctx
->features
= octx
->features
;
615 ctx
->released
= false;
616 ctx
->mm
= vma
->vm_mm
;
617 atomic_inc(&ctx
->mm
->mm_count
);
619 userfaultfd_ctx_get(octx
);
622 list_add_tail(&fctx
->list
, fcs
);
625 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
629 static int dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
631 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
632 struct userfaultfd_wait_queue ewq
;
636 ewq
.msg
.event
= UFFD_EVENT_FORK
;
637 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
639 return userfaultfd_event_wait_completion(ctx
, &ewq
);
642 void dup_userfaultfd_complete(struct list_head
*fcs
)
645 struct userfaultfd_fork_ctx
*fctx
, *n
;
647 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
649 ret
= dup_fctx(fctx
);
650 list_del(&fctx
->list
);
655 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
656 struct vm_userfaultfd_ctx
*vm_ctx
)
658 struct userfaultfd_ctx
*ctx
;
660 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
661 if (ctx
&& (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
)) {
663 userfaultfd_ctx_get(ctx
);
667 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
668 unsigned long from
, unsigned long to
,
671 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
672 struct userfaultfd_wait_queue ewq
;
677 if (to
& ~PAGE_MASK
) {
678 userfaultfd_ctx_put(ctx
);
684 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
685 ewq
.msg
.arg
.remap
.from
= from
;
686 ewq
.msg
.arg
.remap
.to
= to
;
687 ewq
.msg
.arg
.remap
.len
= len
;
689 userfaultfd_event_wait_completion(ctx
, &ewq
);
692 void userfaultfd_remove(struct vm_area_struct
*vma
,
693 struct vm_area_struct
**prev
,
694 unsigned long start
, unsigned long end
)
696 struct mm_struct
*mm
= vma
->vm_mm
;
697 struct userfaultfd_ctx
*ctx
;
698 struct userfaultfd_wait_queue ewq
;
700 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
701 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
704 userfaultfd_ctx_get(ctx
);
705 up_read(&mm
->mmap_sem
);
707 *prev
= NULL
; /* We wait for ACK w/o the mmap semaphore */
711 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
712 ewq
.msg
.arg
.remove
.start
= start
;
713 ewq
.msg
.arg
.remove
.end
= end
;
715 userfaultfd_event_wait_completion(ctx
, &ewq
);
717 down_read(&mm
->mmap_sem
);
720 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
721 unsigned long start
, unsigned long end
)
723 struct userfaultfd_unmap_ctx
*unmap_ctx
;
725 list_for_each_entry(unmap_ctx
, unmaps
, list
)
726 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
727 unmap_ctx
->end
== end
)
733 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
734 unsigned long start
, unsigned long end
,
735 struct list_head
*unmaps
)
737 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
738 struct userfaultfd_unmap_ctx
*unmap_ctx
;
739 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
741 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
742 has_unmap_ctx(ctx
, unmaps
, start
, end
))
745 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
749 userfaultfd_ctx_get(ctx
);
750 unmap_ctx
->ctx
= ctx
;
751 unmap_ctx
->start
= start
;
752 unmap_ctx
->end
= end
;
753 list_add_tail(&unmap_ctx
->list
, unmaps
);
759 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
761 struct userfaultfd_unmap_ctx
*ctx
, *n
;
762 struct userfaultfd_wait_queue ewq
;
764 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
767 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
768 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
769 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
771 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
773 list_del(&ctx
->list
);
778 void userfaultfd_exit(struct mm_struct
*mm
)
780 struct vm_area_struct
*vma
= mm
->mmap
;
783 * We can do the vma walk without locking because the caller
784 * (exit_mm) knows it now has exclusive access
787 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
789 if (ctx
&& (ctx
->features
& UFFD_FEATURE_EVENT_EXIT
)) {
790 struct userfaultfd_wait_queue ewq
;
792 userfaultfd_ctx_get(ctx
);
795 ewq
.msg
.event
= UFFD_EVENT_EXIT
;
797 userfaultfd_event_wait_completion(ctx
, &ewq
);
799 ctx
->features
&= ~UFFD_FEATURE_EVENT_EXIT
;
806 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
808 struct userfaultfd_ctx
*ctx
= file
->private_data
;
809 struct mm_struct
*mm
= ctx
->mm
;
810 struct vm_area_struct
*vma
, *prev
;
811 /* len == 0 means wake all */
812 struct userfaultfd_wake_range range
= { .len
= 0, };
813 unsigned long new_flags
;
815 ACCESS_ONCE(ctx
->released
) = true;
817 if (!mmget_not_zero(mm
))
821 * Flush page faults out of all CPUs. NOTE: all page faults
822 * must be retried without returning VM_FAULT_SIGBUS if
823 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
824 * changes while handle_userfault released the mmap_sem. So
825 * it's critical that released is set to true (above), before
826 * taking the mmap_sem for writing.
828 down_write(&mm
->mmap_sem
);
830 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
832 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
833 !!(vma
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
834 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
838 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
839 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
840 new_flags
, vma
->anon_vma
,
841 vma
->vm_file
, vma
->vm_pgoff
,
848 vma
->vm_flags
= new_flags
;
849 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
851 up_write(&mm
->mmap_sem
);
855 * After no new page faults can wait on this fault_*wqh, flush
856 * the last page faults that may have been already waiting on
859 spin_lock(&ctx
->fault_pending_wqh
.lock
);
860 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
861 __wake_up_locked_key(&ctx
->fault_wqh
, TASK_NORMAL
, &range
);
862 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
864 wake_up_poll(&ctx
->fd_wqh
, POLLHUP
);
865 userfaultfd_ctx_put(ctx
);
869 /* fault_pending_wqh.lock must be hold by the caller */
870 static inline struct userfaultfd_wait_queue
*find_userfault_in(
871 wait_queue_head_t
*wqh
)
874 struct userfaultfd_wait_queue
*uwq
;
876 VM_BUG_ON(!spin_is_locked(&wqh
->lock
));
879 if (!waitqueue_active(wqh
))
881 /* walk in reverse to provide FIFO behavior to read userfaults */
882 wq
= list_last_entry(&wqh
->task_list
, typeof(*wq
), task_list
);
883 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
888 static inline struct userfaultfd_wait_queue
*find_userfault(
889 struct userfaultfd_ctx
*ctx
)
891 return find_userfault_in(&ctx
->fault_pending_wqh
);
894 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
895 struct userfaultfd_ctx
*ctx
)
897 return find_userfault_in(&ctx
->event_wqh
);
900 static unsigned int userfaultfd_poll(struct file
*file
, poll_table
*wait
)
902 struct userfaultfd_ctx
*ctx
= file
->private_data
;
905 poll_wait(file
, &ctx
->fd_wqh
, wait
);
907 switch (ctx
->state
) {
908 case UFFD_STATE_WAIT_API
:
910 case UFFD_STATE_RUNNING
:
912 * poll() never guarantees that read won't block.
913 * userfaults can be waken before they're read().
915 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
918 * lockless access to see if there are pending faults
919 * __pollwait last action is the add_wait_queue but
920 * the spin_unlock would allow the waitqueue_active to
921 * pass above the actual list_add inside
922 * add_wait_queue critical section. So use a full
923 * memory barrier to serialize the list_add write of
924 * add_wait_queue() with the waitqueue_active read
929 if (waitqueue_active(&ctx
->fault_pending_wqh
))
931 else if (waitqueue_active(&ctx
->event_wqh
))
941 static const struct file_operations userfaultfd_fops
;
943 static int resolve_userfault_fork(struct userfaultfd_ctx
*ctx
,
944 struct userfaultfd_ctx
*new,
945 struct uffd_msg
*msg
)
949 unsigned int flags
= new->flags
& UFFD_SHARED_FCNTL_FLAGS
;
951 fd
= get_unused_fd_flags(flags
);
955 file
= anon_inode_getfile("[userfaultfd]", &userfaultfd_fops
, new,
959 return PTR_ERR(file
);
962 fd_install(fd
, file
);
963 msg
->arg
.reserved
.reserved1
= 0;
964 msg
->arg
.fork
.ufd
= fd
;
969 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
970 struct uffd_msg
*msg
)
973 DECLARE_WAITQUEUE(wait
, current
);
974 struct userfaultfd_wait_queue
*uwq
;
976 * Handling fork event requires sleeping operations, so
977 * we drop the event_wqh lock, then do these ops, then
978 * lock it back and wake up the waiter. While the lock is
979 * dropped the ewq may go away so we keep track of it
982 LIST_HEAD(fork_event
);
983 struct userfaultfd_ctx
*fork_nctx
= NULL
;
985 /* always take the fd_wqh lock before the fault_pending_wqh lock */
986 spin_lock(&ctx
->fd_wqh
.lock
);
987 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
989 set_current_state(TASK_INTERRUPTIBLE
);
990 spin_lock(&ctx
->fault_pending_wqh
.lock
);
991 uwq
= find_userfault(ctx
);
994 * Use a seqcount to repeat the lockless check
995 * in wake_userfault() to avoid missing
996 * wakeups because during the refile both
997 * waitqueue could become empty if this is the
1000 write_seqcount_begin(&ctx
->refile_seq
);
1003 * The fault_pending_wqh.lock prevents the uwq
1004 * to disappear from under us.
1006 * Refile this userfault from
1007 * fault_pending_wqh to fault_wqh, it's not
1008 * pending anymore after we read it.
1010 * Use list_del() by hand (as
1011 * userfaultfd_wake_function also uses
1012 * list_del_init() by hand) to be sure nobody
1013 * changes __remove_wait_queue() to use
1014 * list_del_init() in turn breaking the
1015 * !list_empty_careful() check in
1016 * handle_userfault(). The uwq->wq.task_list
1017 * must never be empty at any time during the
1018 * refile, or the waitqueue could disappear
1019 * from under us. The "wait_queue_head_t"
1020 * parameter of __remove_wait_queue() is unused
1023 list_del(&uwq
->wq
.task_list
);
1024 __add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1026 write_seqcount_end(&ctx
->refile_seq
);
1028 /* careful to always initialize msg if ret == 0 */
1030 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1034 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1036 spin_lock(&ctx
->event_wqh
.lock
);
1037 uwq
= find_userfault_evt(ctx
);
1041 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1042 fork_nctx
= (struct userfaultfd_ctx
*)
1044 uwq
->msg
.arg
.reserved
.reserved1
;
1045 list_move(&uwq
->wq
.task_list
, &fork_event
);
1046 spin_unlock(&ctx
->event_wqh
.lock
);
1051 userfaultfd_event_complete(ctx
, uwq
);
1052 spin_unlock(&ctx
->event_wqh
.lock
);
1056 spin_unlock(&ctx
->event_wqh
.lock
);
1058 if (signal_pending(current
)) {
1066 spin_unlock(&ctx
->fd_wqh
.lock
);
1068 spin_lock(&ctx
->fd_wqh
.lock
);
1070 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1071 __set_current_state(TASK_RUNNING
);
1072 spin_unlock(&ctx
->fd_wqh
.lock
);
1074 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1075 ret
= resolve_userfault_fork(ctx
, fork_nctx
, msg
);
1078 spin_lock(&ctx
->event_wqh
.lock
);
1079 if (!list_empty(&fork_event
)) {
1080 uwq
= list_first_entry(&fork_event
,
1083 list_del(&uwq
->wq
.task_list
);
1084 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1085 userfaultfd_event_complete(ctx
, uwq
);
1087 spin_unlock(&ctx
->event_wqh
.lock
);
1094 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1095 size_t count
, loff_t
*ppos
)
1097 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1098 ssize_t _ret
, ret
= 0;
1099 struct uffd_msg msg
;
1100 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1102 if (ctx
->state
== UFFD_STATE_WAIT_API
)
1106 if (count
< sizeof(msg
))
1107 return ret
? ret
: -EINVAL
;
1108 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
);
1110 return ret
? ret
: _ret
;
1111 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1112 return ret
? ret
: -EFAULT
;
1115 count
-= sizeof(msg
);
1117 * Allow to read more than one fault at time but only
1118 * block if waiting for the very first one.
1120 no_wait
= O_NONBLOCK
;
1124 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1125 struct userfaultfd_wake_range
*range
)
1127 unsigned long start
, end
;
1129 start
= range
->start
;
1130 end
= range
->start
+ range
->len
;
1132 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1133 /* wake all in the range and autoremove */
1134 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1135 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1137 if (waitqueue_active(&ctx
->fault_wqh
))
1138 __wake_up_locked_key(&ctx
->fault_wqh
, TASK_NORMAL
, range
);
1139 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1142 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1143 struct userfaultfd_wake_range
*range
)
1149 * To be sure waitqueue_active() is not reordered by the CPU
1150 * before the pagetable update, use an explicit SMP memory
1151 * barrier here. PT lock release or up_read(mmap_sem) still
1152 * have release semantics that can allow the
1153 * waitqueue_active() to be reordered before the pte update.
1158 * Use waitqueue_active because it's very frequent to
1159 * change the address space atomically even if there are no
1160 * userfaults yet. So we take the spinlock only when we're
1161 * sure we've userfaults to wake.
1164 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1165 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1166 waitqueue_active(&ctx
->fault_wqh
);
1168 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1170 __wake_userfault(ctx
, range
);
1173 static __always_inline
int validate_range(struct mm_struct
*mm
,
1174 __u64 start
, __u64 len
)
1176 __u64 task_size
= mm
->task_size
;
1178 if (start
& ~PAGE_MASK
)
1180 if (len
& ~PAGE_MASK
)
1184 if (start
< mmap_min_addr
)
1186 if (start
>= task_size
)
1188 if (len
> task_size
- start
)
1193 static inline bool vma_can_userfault(struct vm_area_struct
*vma
)
1195 return vma_is_anonymous(vma
) || is_vm_hugetlb_page(vma
) ||
1199 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1202 struct mm_struct
*mm
= ctx
->mm
;
1203 struct vm_area_struct
*vma
, *prev
, *cur
;
1205 struct uffdio_register uffdio_register
;
1206 struct uffdio_register __user
*user_uffdio_register
;
1207 unsigned long vm_flags
, new_flags
;
1209 bool non_anon_pages
;
1210 unsigned long start
, end
, vma_end
;
1212 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1215 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1216 sizeof(uffdio_register
)-sizeof(__u64
)))
1220 if (!uffdio_register
.mode
)
1222 if (uffdio_register
.mode
& ~(UFFDIO_REGISTER_MODE_MISSING
|
1223 UFFDIO_REGISTER_MODE_WP
))
1226 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1227 vm_flags
|= VM_UFFD_MISSING
;
1228 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1229 vm_flags
|= VM_UFFD_WP
;
1231 * FIXME: remove the below error constraint by
1232 * implementing the wprotect tracking mode.
1238 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1239 uffdio_register
.range
.len
);
1243 start
= uffdio_register
.range
.start
;
1244 end
= start
+ uffdio_register
.range
.len
;
1247 if (!mmget_not_zero(mm
))
1250 down_write(&mm
->mmap_sem
);
1251 vma
= find_vma_prev(mm
, start
, &prev
);
1255 /* check that there's at least one vma in the range */
1257 if (vma
->vm_start
>= end
)
1261 * If the first vma contains huge pages, make sure start address
1262 * is aligned to huge page size.
1264 if (is_vm_hugetlb_page(vma
)) {
1265 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1267 if (start
& (vma_hpagesize
- 1))
1272 * Search for not compatible vmas.
1275 non_anon_pages
= false;
1276 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1279 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1280 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1282 /* check not compatible vmas */
1284 if (!vma_can_userfault(cur
))
1287 * If this vma contains ending address, and huge pages
1290 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1291 end
> cur
->vm_start
) {
1292 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1296 if (end
& (vma_hpagesize
- 1))
1301 * Check that this vma isn't already owned by a
1302 * different userfaultfd. We can't allow more than one
1303 * userfaultfd to own a single vma simultaneously or we
1304 * wouldn't know which one to deliver the userfaults to.
1307 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1308 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1312 * Note vmas containing huge pages
1314 if (is_vm_hugetlb_page(cur
) || vma_is_shmem(cur
))
1315 non_anon_pages
= true;
1321 if (vma
->vm_start
< start
)
1328 BUG_ON(!vma_can_userfault(vma
));
1329 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1330 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1333 * Nothing to do: this vma is already registered into this
1334 * userfaultfd and with the right tracking mode too.
1336 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1337 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1340 if (vma
->vm_start
> start
)
1341 start
= vma
->vm_start
;
1342 vma_end
= min(end
, vma
->vm_end
);
1344 new_flags
= (vma
->vm_flags
& ~vm_flags
) | vm_flags
;
1345 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1346 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1348 ((struct vm_userfaultfd_ctx
){ ctx
}));
1353 if (vma
->vm_start
< start
) {
1354 ret
= split_vma(mm
, vma
, start
, 1);
1358 if (vma
->vm_end
> end
) {
1359 ret
= split_vma(mm
, vma
, end
, 0);
1365 * In the vma_merge() successful mprotect-like case 8:
1366 * the next vma was merged into the current one and
1367 * the current one has not been updated yet.
1369 vma
->vm_flags
= new_flags
;
1370 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1374 start
= vma
->vm_end
;
1376 } while (vma
&& vma
->vm_start
< end
);
1378 up_write(&mm
->mmap_sem
);
1382 * Now that we scanned all vmas we can already tell
1383 * userland which ioctls methods are guaranteed to
1384 * succeed on this range.
1386 if (put_user(non_anon_pages
? UFFD_API_RANGE_IOCTLS_BASIC
:
1387 UFFD_API_RANGE_IOCTLS
,
1388 &user_uffdio_register
->ioctls
))
1395 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1398 struct mm_struct
*mm
= ctx
->mm
;
1399 struct vm_area_struct
*vma
, *prev
, *cur
;
1401 struct uffdio_range uffdio_unregister
;
1402 unsigned long new_flags
;
1404 unsigned long start
, end
, vma_end
;
1405 const void __user
*buf
= (void __user
*)arg
;
1408 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1411 ret
= validate_range(mm
, uffdio_unregister
.start
,
1412 uffdio_unregister
.len
);
1416 start
= uffdio_unregister
.start
;
1417 end
= start
+ uffdio_unregister
.len
;
1420 if (!mmget_not_zero(mm
))
1423 down_write(&mm
->mmap_sem
);
1424 vma
= find_vma_prev(mm
, start
, &prev
);
1428 /* check that there's at least one vma in the range */
1430 if (vma
->vm_start
>= end
)
1434 * If the first vma contains huge pages, make sure start address
1435 * is aligned to huge page size.
1437 if (is_vm_hugetlb_page(vma
)) {
1438 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1440 if (start
& (vma_hpagesize
- 1))
1445 * Search for not compatible vmas.
1449 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1452 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1453 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1456 * Check not compatible vmas, not strictly required
1457 * here as not compatible vmas cannot have an
1458 * userfaultfd_ctx registered on them, but this
1459 * provides for more strict behavior to notice
1460 * unregistration errors.
1462 if (!vma_can_userfault(cur
))
1469 if (vma
->vm_start
< start
)
1476 BUG_ON(!vma_can_userfault(vma
));
1479 * Nothing to do: this vma is already registered into this
1480 * userfaultfd and with the right tracking mode too.
1482 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1485 if (vma
->vm_start
> start
)
1486 start
= vma
->vm_start
;
1487 vma_end
= min(end
, vma
->vm_end
);
1489 if (userfaultfd_missing(vma
)) {
1491 * Wake any concurrent pending userfault while
1492 * we unregister, so they will not hang
1493 * permanently and it avoids userland to call
1494 * UFFDIO_WAKE explicitly.
1496 struct userfaultfd_wake_range range
;
1497 range
.start
= start
;
1498 range
.len
= vma_end
- start
;
1499 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1502 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
1503 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1504 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1511 if (vma
->vm_start
< start
) {
1512 ret
= split_vma(mm
, vma
, start
, 1);
1516 if (vma
->vm_end
> end
) {
1517 ret
= split_vma(mm
, vma
, end
, 0);
1523 * In the vma_merge() successful mprotect-like case 8:
1524 * the next vma was merged into the current one and
1525 * the current one has not been updated yet.
1527 vma
->vm_flags
= new_flags
;
1528 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1532 start
= vma
->vm_end
;
1534 } while (vma
&& vma
->vm_start
< end
);
1536 up_write(&mm
->mmap_sem
);
1543 * userfaultfd_wake may be used in combination with the
1544 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1546 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1550 struct uffdio_range uffdio_wake
;
1551 struct userfaultfd_wake_range range
;
1552 const void __user
*buf
= (void __user
*)arg
;
1555 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1558 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1562 range
.start
= uffdio_wake
.start
;
1563 range
.len
= uffdio_wake
.len
;
1566 * len == 0 means wake all and we don't want to wake all here,
1567 * so check it again to be sure.
1569 VM_BUG_ON(!range
.len
);
1571 wake_userfault(ctx
, &range
);
1578 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1582 struct uffdio_copy uffdio_copy
;
1583 struct uffdio_copy __user
*user_uffdio_copy
;
1584 struct userfaultfd_wake_range range
;
1586 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1589 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1590 /* don't copy "copy" last field */
1591 sizeof(uffdio_copy
)-sizeof(__s64
)))
1594 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1598 * double check for wraparound just in case. copy_from_user()
1599 * will later check uffdio_copy.src + uffdio_copy.len to fit
1600 * in the userland range.
1603 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1605 if (uffdio_copy
.mode
& ~UFFDIO_COPY_MODE_DONTWAKE
)
1607 if (mmget_not_zero(ctx
->mm
)) {
1608 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1614 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1619 /* len == 0 would wake all */
1621 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1622 range
.start
= uffdio_copy
.dst
;
1623 wake_userfault(ctx
, &range
);
1625 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1630 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1634 struct uffdio_zeropage uffdio_zeropage
;
1635 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1636 struct userfaultfd_wake_range range
;
1638 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1641 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1642 /* don't copy "zeropage" last field */
1643 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1646 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1647 uffdio_zeropage
.range
.len
);
1651 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1654 if (mmget_not_zero(ctx
->mm
)) {
1655 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1656 uffdio_zeropage
.range
.len
);
1659 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1663 /* len == 0 would wake all */
1666 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1667 range
.start
= uffdio_zeropage
.range
.start
;
1668 wake_userfault(ctx
, &range
);
1670 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1675 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1678 * For the current set of features the bits just coincide
1680 return (unsigned int)user_features
;
1684 * userland asks for a certain API version and we return which bits
1685 * and ioctl commands are implemented in this kernel for such API
1686 * version or -EINVAL if unknown.
1688 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1691 struct uffdio_api uffdio_api
;
1692 void __user
*buf
= (void __user
*)arg
;
1697 if (ctx
->state
!= UFFD_STATE_WAIT_API
)
1700 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1702 features
= uffdio_api
.features
;
1703 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
)) {
1704 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1705 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1710 /* report all available features and ioctls to userland */
1711 uffdio_api
.features
= UFFD_API_FEATURES
;
1712 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1714 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1716 ctx
->state
= UFFD_STATE_RUNNING
;
1717 /* only enable the requested features for this uffd context */
1718 ctx
->features
= uffd_ctx_features(features
);
1724 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1728 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1730 if (cmd
!= UFFDIO_API
&& ctx
->state
== UFFD_STATE_WAIT_API
)
1735 ret
= userfaultfd_api(ctx
, arg
);
1737 case UFFDIO_REGISTER
:
1738 ret
= userfaultfd_register(ctx
, arg
);
1740 case UFFDIO_UNREGISTER
:
1741 ret
= userfaultfd_unregister(ctx
, arg
);
1744 ret
= userfaultfd_wake(ctx
, arg
);
1747 ret
= userfaultfd_copy(ctx
, arg
);
1749 case UFFDIO_ZEROPAGE
:
1750 ret
= userfaultfd_zeropage(ctx
, arg
);
1756 #ifdef CONFIG_PROC_FS
1757 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
1759 struct userfaultfd_ctx
*ctx
= f
->private_data
;
1761 struct userfaultfd_wait_queue
*uwq
;
1762 unsigned long pending
= 0, total
= 0;
1764 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1765 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.task_list
, task_list
) {
1766 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
1770 list_for_each_entry(wq
, &ctx
->fault_wqh
.task_list
, task_list
) {
1771 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
1774 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1777 * If more protocols will be added, there will be all shown
1778 * separated by a space. Like this:
1779 * protocols: aa:... bb:...
1781 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1782 pending
, total
, UFFD_API
, UFFD_API_FEATURES
,
1783 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
1787 static const struct file_operations userfaultfd_fops
= {
1788 #ifdef CONFIG_PROC_FS
1789 .show_fdinfo
= userfaultfd_show_fdinfo
,
1791 .release
= userfaultfd_release
,
1792 .poll
= userfaultfd_poll
,
1793 .read
= userfaultfd_read
,
1794 .unlocked_ioctl
= userfaultfd_ioctl
,
1795 .compat_ioctl
= userfaultfd_ioctl
,
1796 .llseek
= noop_llseek
,
1799 static void init_once_userfaultfd_ctx(void *mem
)
1801 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
1803 init_waitqueue_head(&ctx
->fault_pending_wqh
);
1804 init_waitqueue_head(&ctx
->fault_wqh
);
1805 init_waitqueue_head(&ctx
->event_wqh
);
1806 init_waitqueue_head(&ctx
->fd_wqh
);
1807 seqcount_init(&ctx
->refile_seq
);
1811 * userfaultfd_file_create - Creates a userfaultfd file pointer.
1812 * @flags: Flags for the userfaultfd file.
1814 * This function creates a userfaultfd file pointer, w/out installing
1815 * it into the fd table. This is useful when the userfaultfd file is
1816 * used during the initialization of data structures that require
1817 * extra setup after the userfaultfd creation. So the userfaultfd
1818 * creation is split into the file pointer creation phase, and the
1819 * file descriptor installation phase. In this way races with
1820 * userspace closing the newly installed file descriptor can be
1821 * avoided. Returns a userfaultfd file pointer, or a proper error
1824 static struct file
*userfaultfd_file_create(int flags
)
1827 struct userfaultfd_ctx
*ctx
;
1829 BUG_ON(!current
->mm
);
1831 /* Check the UFFD_* constants for consistency. */
1832 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
1833 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
1835 file
= ERR_PTR(-EINVAL
);
1836 if (flags
& ~UFFD_SHARED_FCNTL_FLAGS
)
1839 file
= ERR_PTR(-ENOMEM
);
1840 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
1844 atomic_set(&ctx
->refcount
, 1);
1847 ctx
->state
= UFFD_STATE_WAIT_API
;
1848 ctx
->released
= false;
1849 ctx
->mm
= current
->mm
;
1850 /* prevent the mm struct to be freed */
1853 file
= anon_inode_getfile("[userfaultfd]", &userfaultfd_fops
, ctx
,
1854 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
));
1857 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
1863 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
1868 error
= get_unused_fd_flags(flags
& UFFD_SHARED_FCNTL_FLAGS
);
1873 file
= userfaultfd_file_create(flags
);
1875 error
= PTR_ERR(file
);
1876 goto err_put_unused_fd
;
1878 fd_install(fd
, file
);
1888 static int __init
userfaultfd_init(void)
1890 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
1891 sizeof(struct userfaultfd_ctx
),
1893 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
1894 init_once_userfaultfd_ctx
);
1897 __initcall(userfaultfd_init
);