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/poll.h>
19 #include <linux/slab.h>
20 #include <linux/seq_file.h>
21 #include <linux/file.h>
22 #include <linux/bug.h>
23 #include <linux/anon_inodes.h>
24 #include <linux/syscalls.h>
25 #include <linux/userfaultfd_k.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ioctl.h>
28 #include <linux/security.h>
29 #include <linux/hugetlb.h>
31 int sysctl_unprivileged_userfaultfd __read_mostly
= 1;
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.
46 * fault_pending_wqh.lock
50 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
51 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
52 * also taken in IRQ context.
54 struct userfaultfd_ctx
{
55 /* waitqueue head for the pending (i.e. not read) userfaults */
56 wait_queue_head_t fault_pending_wqh
;
57 /* waitqueue head for the userfaults */
58 wait_queue_head_t fault_wqh
;
59 /* waitqueue head for the pseudo fd to wakeup poll/read */
60 wait_queue_head_t fd_wqh
;
61 /* waitqueue head for events */
62 wait_queue_head_t event_wqh
;
63 /* a refile sequence protected by fault_pending_wqh lock */
64 seqcount_spinlock_t refile_seq
;
65 /* pseudo fd refcounting */
67 /* userfaultfd syscall flags */
69 /* features requested from the userspace */
70 unsigned int features
;
72 enum userfaultfd_state state
;
75 /* memory mappings are changing because of non-cooperative event */
77 /* mm with one ore more vmas attached to this userfaultfd_ctx */
81 struct userfaultfd_fork_ctx
{
82 struct userfaultfd_ctx
*orig
;
83 struct userfaultfd_ctx
*new;
84 struct list_head list
;
87 struct userfaultfd_unmap_ctx
{
88 struct userfaultfd_ctx
*ctx
;
91 struct list_head list
;
94 struct userfaultfd_wait_queue
{
96 wait_queue_entry_t wq
;
97 struct userfaultfd_ctx
*ctx
;
101 struct userfaultfd_wake_range
{
106 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
107 int wake_flags
, void *key
)
109 struct userfaultfd_wake_range
*range
= key
;
111 struct userfaultfd_wait_queue
*uwq
;
112 unsigned long start
, len
;
114 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
116 /* len == 0 means wake all */
117 start
= range
->start
;
119 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
120 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
122 WRITE_ONCE(uwq
->waken
, true);
124 * The Program-Order guarantees provided by the scheduler
125 * ensure uwq->waken is visible before the task is woken.
127 ret
= wake_up_state(wq
->private, mode
);
130 * Wake only once, autoremove behavior.
132 * After the effect of list_del_init is visible to the other
133 * CPUs, the waitqueue may disappear from under us, see the
134 * !list_empty_careful() in handle_userfault().
136 * try_to_wake_up() has an implicit smp_mb(), and the
137 * wq->private is read before calling the extern function
138 * "wake_up_state" (which in turns calls try_to_wake_up).
140 list_del_init(&wq
->entry
);
147 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
149 * @ctx: [in] Pointer to the userfaultfd context.
151 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
153 refcount_inc(&ctx
->refcount
);
157 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
159 * @ctx: [in] Pointer to userfaultfd context.
161 * The userfaultfd context reference must have been previously acquired either
162 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
164 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
166 if (refcount_dec_and_test(&ctx
->refcount
)) {
167 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
168 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
169 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
170 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
171 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
172 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
173 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
174 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
176 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
180 static inline void msg_init(struct uffd_msg
*msg
)
182 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
184 * Must use memset to zero out the paddings or kernel data is
185 * leaked to userland.
187 memset(msg
, 0, sizeof(struct uffd_msg
));
190 static inline struct uffd_msg
userfault_msg(unsigned long address
,
192 unsigned long reason
,
193 unsigned int features
)
197 msg
.event
= UFFD_EVENT_PAGEFAULT
;
198 msg
.arg
.pagefault
.address
= address
;
199 if (flags
& FAULT_FLAG_WRITE
)
201 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
202 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
203 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
204 * was a read fault, otherwise if set it means it's
207 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
208 if (reason
& VM_UFFD_WP
)
210 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
211 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
212 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
213 * a missing fault, otherwise if set it means it's a
214 * write protect fault.
216 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
217 if (features
& UFFD_FEATURE_THREAD_ID
)
218 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
222 #ifdef CONFIG_HUGETLB_PAGE
224 * Same functionality as userfaultfd_must_wait below with modifications for
227 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
228 struct vm_area_struct
*vma
,
229 unsigned long address
,
231 unsigned long reason
)
233 struct mm_struct
*mm
= ctx
->mm
;
237 mmap_assert_locked(mm
);
239 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
245 pte
= huge_ptep_get(ptep
);
248 * Lockless access: we're in a wait_event so it's ok if it
251 if (huge_pte_none(pte
))
253 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
259 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
260 struct vm_area_struct
*vma
,
261 unsigned long address
,
263 unsigned long reason
)
265 return false; /* should never get here */
267 #endif /* CONFIG_HUGETLB_PAGE */
270 * Verify the pagetables are still not ok after having reigstered into
271 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
272 * userfault that has already been resolved, if userfaultfd_read and
273 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
276 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
277 unsigned long address
,
279 unsigned long reason
)
281 struct mm_struct
*mm
= ctx
->mm
;
289 mmap_assert_locked(mm
);
291 pgd
= pgd_offset(mm
, address
);
292 if (!pgd_present(*pgd
))
294 p4d
= p4d_offset(pgd
, address
);
295 if (!p4d_present(*p4d
))
297 pud
= pud_offset(p4d
, address
);
298 if (!pud_present(*pud
))
300 pmd
= pmd_offset(pud
, address
);
302 * READ_ONCE must function as a barrier with narrower scope
303 * and it must be equivalent to:
304 * _pmd = *pmd; barrier();
306 * This is to deal with the instability (as in
307 * pmd_trans_unstable) of the pmd.
309 _pmd
= READ_ONCE(*pmd
);
314 if (!pmd_present(_pmd
))
317 if (pmd_trans_huge(_pmd
)) {
318 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
324 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
325 * and use the standard pte_offset_map() instead of parsing _pmd.
327 pte
= pte_offset_map(pmd
, address
);
329 * Lockless access: we're in a wait_event so it's ok if it
334 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
342 static inline long userfaultfd_get_blocking_state(unsigned int flags
)
344 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
345 return TASK_INTERRUPTIBLE
;
347 if (flags
& FAULT_FLAG_KILLABLE
)
348 return TASK_KILLABLE
;
350 return TASK_UNINTERRUPTIBLE
;
354 * The locking rules involved in returning VM_FAULT_RETRY depending on
355 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
356 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
357 * recommendation in __lock_page_or_retry is not an understatement.
359 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
360 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
363 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
364 * set, VM_FAULT_RETRY can still be returned if and only if there are
365 * fatal_signal_pending()s, and the mmap_lock must be released before
368 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
370 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
371 struct userfaultfd_ctx
*ctx
;
372 struct userfaultfd_wait_queue uwq
;
373 vm_fault_t ret
= VM_FAULT_SIGBUS
;
378 * We don't do userfault handling for the final child pid update.
380 * We also don't do userfault handling during
381 * coredumping. hugetlbfs has the special
382 * follow_hugetlb_page() to skip missing pages in the
383 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
384 * the no_page_table() helper in follow_page_mask(), but the
385 * shmem_vm_ops->fault method is invoked even during
386 * coredumping without mmap_lock and it ends up here.
388 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
392 * Coredumping runs without mmap_lock so we can only check that
393 * the mmap_lock is held, if PF_DUMPCORE was not set.
395 mmap_assert_locked(mm
);
397 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
401 BUG_ON(ctx
->mm
!= mm
);
403 VM_BUG_ON(reason
& ~(VM_UFFD_MISSING
|VM_UFFD_WP
));
404 VM_BUG_ON(!(reason
& VM_UFFD_MISSING
) ^ !!(reason
& VM_UFFD_WP
));
406 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
410 * If it's already released don't get it. This avoids to loop
411 * in __get_user_pages if userfaultfd_release waits on the
412 * caller of handle_userfault to release the mmap_lock.
414 if (unlikely(READ_ONCE(ctx
->released
))) {
416 * Don't return VM_FAULT_SIGBUS in this case, so a non
417 * cooperative manager can close the uffd after the
418 * last UFFDIO_COPY, without risking to trigger an
419 * involuntary SIGBUS if the process was starting the
420 * userfaultfd while the userfaultfd was still armed
421 * (but after the last UFFDIO_COPY). If the uffd
422 * wasn't already closed when the userfault reached
423 * this point, that would normally be solved by
424 * userfaultfd_must_wait returning 'false'.
426 * If we were to return VM_FAULT_SIGBUS here, the non
427 * cooperative manager would be instead forced to
428 * always call UFFDIO_UNREGISTER before it can safely
431 ret
= VM_FAULT_NOPAGE
;
436 * Check that we can return VM_FAULT_RETRY.
438 * NOTE: it should become possible to return VM_FAULT_RETRY
439 * even if FAULT_FLAG_TRIED is set without leading to gup()
440 * -EBUSY failures, if the userfaultfd is to be extended for
441 * VM_UFFD_WP tracking and we intend to arm the userfault
442 * without first stopping userland access to the memory. For
443 * VM_UFFD_MISSING userfaults this is enough for now.
445 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
447 * Validate the invariant that nowait must allow retry
448 * to be sure not to return SIGBUS erroneously on
449 * nowait invocations.
451 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
452 #ifdef CONFIG_DEBUG_VM
453 if (printk_ratelimit()) {
455 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
464 * Handle nowait, not much to do other than tell it to retry
467 ret
= VM_FAULT_RETRY
;
468 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
471 /* take the reference before dropping the mmap_lock */
472 userfaultfd_ctx_get(ctx
);
474 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
475 uwq
.wq
.private = current
;
476 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
,
481 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
483 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
485 * After the __add_wait_queue the uwq is visible to userland
486 * through poll/read().
488 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
490 * The smp_mb() after __set_current_state prevents the reads
491 * following the spin_unlock to happen before the list_add in
494 set_current_state(blocking_state
);
495 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
497 if (!is_vm_hugetlb_page(vmf
->vma
))
498 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
501 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
504 mmap_read_unlock(mm
);
506 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
507 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
511 __set_current_state(TASK_RUNNING
);
514 * Here we race with the list_del; list_add in
515 * userfaultfd_ctx_read(), however because we don't ever run
516 * list_del_init() to refile across the two lists, the prev
517 * and next pointers will never point to self. list_add also
518 * would never let any of the two pointers to point to
519 * self. So list_empty_careful won't risk to see both pointers
520 * pointing to self at any time during the list refile. The
521 * only case where list_del_init() is called is the full
522 * removal in the wake function and there we don't re-list_add
523 * and it's fine not to block on the spinlock. The uwq on this
524 * kernel stack can be released after the list_del_init.
526 if (!list_empty_careful(&uwq
.wq
.entry
)) {
527 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
529 * No need of list_del_init(), the uwq on the stack
530 * will be freed shortly anyway.
532 list_del(&uwq
.wq
.entry
);
533 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
537 * ctx may go away after this if the userfault pseudo fd is
540 userfaultfd_ctx_put(ctx
);
546 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
547 struct userfaultfd_wait_queue
*ewq
)
549 struct userfaultfd_ctx
*release_new_ctx
;
551 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
555 init_waitqueue_entry(&ewq
->wq
, current
);
556 release_new_ctx
= NULL
;
558 spin_lock_irq(&ctx
->event_wqh
.lock
);
560 * After the __add_wait_queue the uwq is visible to userland
561 * through poll/read().
563 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
565 set_current_state(TASK_KILLABLE
);
566 if (ewq
->msg
.event
== 0)
568 if (READ_ONCE(ctx
->released
) ||
569 fatal_signal_pending(current
)) {
571 * &ewq->wq may be queued in fork_event, but
572 * __remove_wait_queue ignores the head
573 * parameter. It would be a problem if it
576 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
577 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
578 struct userfaultfd_ctx
*new;
580 new = (struct userfaultfd_ctx
*)
582 ewq
->msg
.arg
.reserved
.reserved1
;
583 release_new_ctx
= new;
588 spin_unlock_irq(&ctx
->event_wqh
.lock
);
590 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
593 spin_lock_irq(&ctx
->event_wqh
.lock
);
595 __set_current_state(TASK_RUNNING
);
596 spin_unlock_irq(&ctx
->event_wqh
.lock
);
598 if (release_new_ctx
) {
599 struct vm_area_struct
*vma
;
600 struct mm_struct
*mm
= release_new_ctx
->mm
;
602 /* the various vma->vm_userfaultfd_ctx still points to it */
604 /* no task can run (and in turn coredump) yet */
605 VM_WARN_ON(!mmget_still_valid(mm
));
606 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
607 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
608 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
609 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
611 mmap_write_unlock(mm
);
613 userfaultfd_ctx_put(release_new_ctx
);
617 * ctx may go away after this if the userfault pseudo fd is
621 WRITE_ONCE(ctx
->mmap_changing
, false);
622 userfaultfd_ctx_put(ctx
);
625 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
626 struct userfaultfd_wait_queue
*ewq
)
629 wake_up_locked(&ctx
->event_wqh
);
630 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
633 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
635 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
636 struct userfaultfd_fork_ctx
*fctx
;
638 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
639 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
640 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
641 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
645 list_for_each_entry(fctx
, fcs
, list
)
646 if (fctx
->orig
== octx
) {
652 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
656 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
662 refcount_set(&ctx
->refcount
, 1);
663 ctx
->flags
= octx
->flags
;
664 ctx
->state
= UFFD_STATE_RUNNING
;
665 ctx
->features
= octx
->features
;
666 ctx
->released
= false;
667 ctx
->mmap_changing
= false;
668 ctx
->mm
= vma
->vm_mm
;
671 userfaultfd_ctx_get(octx
);
672 WRITE_ONCE(octx
->mmap_changing
, true);
675 list_add_tail(&fctx
->list
, fcs
);
678 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
682 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
684 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
685 struct userfaultfd_wait_queue ewq
;
689 ewq
.msg
.event
= UFFD_EVENT_FORK
;
690 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
692 userfaultfd_event_wait_completion(ctx
, &ewq
);
695 void dup_userfaultfd_complete(struct list_head
*fcs
)
697 struct userfaultfd_fork_ctx
*fctx
, *n
;
699 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
701 list_del(&fctx
->list
);
706 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
707 struct vm_userfaultfd_ctx
*vm_ctx
)
709 struct userfaultfd_ctx
*ctx
;
711 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
716 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
718 userfaultfd_ctx_get(ctx
);
719 WRITE_ONCE(ctx
->mmap_changing
, true);
721 /* Drop uffd context if remap feature not enabled */
722 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
723 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
727 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
728 unsigned long from
, unsigned long to
,
731 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
732 struct userfaultfd_wait_queue ewq
;
737 if (to
& ~PAGE_MASK
) {
738 userfaultfd_ctx_put(ctx
);
744 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
745 ewq
.msg
.arg
.remap
.from
= from
;
746 ewq
.msg
.arg
.remap
.to
= to
;
747 ewq
.msg
.arg
.remap
.len
= len
;
749 userfaultfd_event_wait_completion(ctx
, &ewq
);
752 bool userfaultfd_remove(struct vm_area_struct
*vma
,
753 unsigned long start
, unsigned long end
)
755 struct mm_struct
*mm
= vma
->vm_mm
;
756 struct userfaultfd_ctx
*ctx
;
757 struct userfaultfd_wait_queue ewq
;
759 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
760 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
763 userfaultfd_ctx_get(ctx
);
764 WRITE_ONCE(ctx
->mmap_changing
, true);
765 mmap_read_unlock(mm
);
769 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
770 ewq
.msg
.arg
.remove
.start
= start
;
771 ewq
.msg
.arg
.remove
.end
= end
;
773 userfaultfd_event_wait_completion(ctx
, &ewq
);
778 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
779 unsigned long start
, unsigned long end
)
781 struct userfaultfd_unmap_ctx
*unmap_ctx
;
783 list_for_each_entry(unmap_ctx
, unmaps
, list
)
784 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
785 unmap_ctx
->end
== end
)
791 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
792 unsigned long start
, unsigned long end
,
793 struct list_head
*unmaps
)
795 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
796 struct userfaultfd_unmap_ctx
*unmap_ctx
;
797 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
799 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
800 has_unmap_ctx(ctx
, unmaps
, start
, end
))
803 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
807 userfaultfd_ctx_get(ctx
);
808 WRITE_ONCE(ctx
->mmap_changing
, true);
809 unmap_ctx
->ctx
= ctx
;
810 unmap_ctx
->start
= start
;
811 unmap_ctx
->end
= end
;
812 list_add_tail(&unmap_ctx
->list
, unmaps
);
818 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
820 struct userfaultfd_unmap_ctx
*ctx
, *n
;
821 struct userfaultfd_wait_queue ewq
;
823 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
826 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
827 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
828 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
830 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
832 list_del(&ctx
->list
);
837 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
839 struct userfaultfd_ctx
*ctx
= file
->private_data
;
840 struct mm_struct
*mm
= ctx
->mm
;
841 struct vm_area_struct
*vma
, *prev
;
842 /* len == 0 means wake all */
843 struct userfaultfd_wake_range range
= { .len
= 0, };
844 unsigned long new_flags
;
847 WRITE_ONCE(ctx
->released
, true);
849 if (!mmget_not_zero(mm
))
853 * Flush page faults out of all CPUs. NOTE: all page faults
854 * must be retried without returning VM_FAULT_SIGBUS if
855 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
856 * changes while handle_userfault released the mmap_lock. So
857 * it's critical that released is set to true (above), before
858 * taking the mmap_lock for writing.
861 still_valid
= mmget_still_valid(mm
);
863 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
865 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
866 !!(vma
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
867 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
871 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
873 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
874 new_flags
, vma
->anon_vma
,
875 vma
->vm_file
, vma
->vm_pgoff
,
883 vma
->vm_flags
= new_flags
;
884 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
886 mmap_write_unlock(mm
);
890 * After no new page faults can wait on this fault_*wqh, flush
891 * the last page faults that may have been already waiting on
894 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
895 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
896 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
897 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
899 /* Flush pending events that may still wait on event_wqh */
900 wake_up_all(&ctx
->event_wqh
);
902 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
903 userfaultfd_ctx_put(ctx
);
907 /* fault_pending_wqh.lock must be hold by the caller */
908 static inline struct userfaultfd_wait_queue
*find_userfault_in(
909 wait_queue_head_t
*wqh
)
911 wait_queue_entry_t
*wq
;
912 struct userfaultfd_wait_queue
*uwq
;
914 lockdep_assert_held(&wqh
->lock
);
917 if (!waitqueue_active(wqh
))
919 /* walk in reverse to provide FIFO behavior to read userfaults */
920 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
921 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
926 static inline struct userfaultfd_wait_queue
*find_userfault(
927 struct userfaultfd_ctx
*ctx
)
929 return find_userfault_in(&ctx
->fault_pending_wqh
);
932 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
933 struct userfaultfd_ctx
*ctx
)
935 return find_userfault_in(&ctx
->event_wqh
);
938 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
940 struct userfaultfd_ctx
*ctx
= file
->private_data
;
943 poll_wait(file
, &ctx
->fd_wqh
, wait
);
945 switch (ctx
->state
) {
946 case UFFD_STATE_WAIT_API
:
948 case UFFD_STATE_RUNNING
:
950 * poll() never guarantees that read won't block.
951 * userfaults can be waken before they're read().
953 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
956 * lockless access to see if there are pending faults
957 * __pollwait last action is the add_wait_queue but
958 * the spin_unlock would allow the waitqueue_active to
959 * pass above the actual list_add inside
960 * add_wait_queue critical section. So use a full
961 * memory barrier to serialize the list_add write of
962 * add_wait_queue() with the waitqueue_active read
967 if (waitqueue_active(&ctx
->fault_pending_wqh
))
969 else if (waitqueue_active(&ctx
->event_wqh
))
979 static const struct file_operations userfaultfd_fops
;
981 static int resolve_userfault_fork(struct userfaultfd_ctx
*ctx
,
982 struct userfaultfd_ctx
*new,
983 struct uffd_msg
*msg
)
987 fd
= anon_inode_getfd("[userfaultfd]", &userfaultfd_fops
, new,
988 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
));
992 msg
->arg
.reserved
.reserved1
= 0;
993 msg
->arg
.fork
.ufd
= fd
;
997 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
998 struct uffd_msg
*msg
)
1001 DECLARE_WAITQUEUE(wait
, current
);
1002 struct userfaultfd_wait_queue
*uwq
;
1004 * Handling fork event requires sleeping operations, so
1005 * we drop the event_wqh lock, then do these ops, then
1006 * lock it back and wake up the waiter. While the lock is
1007 * dropped the ewq may go away so we keep track of it
1010 LIST_HEAD(fork_event
);
1011 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1013 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1014 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1015 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1017 set_current_state(TASK_INTERRUPTIBLE
);
1018 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1019 uwq
= find_userfault(ctx
);
1022 * Use a seqcount to repeat the lockless check
1023 * in wake_userfault() to avoid missing
1024 * wakeups because during the refile both
1025 * waitqueue could become empty if this is the
1028 write_seqcount_begin(&ctx
->refile_seq
);
1031 * The fault_pending_wqh.lock prevents the uwq
1032 * to disappear from under us.
1034 * Refile this userfault from
1035 * fault_pending_wqh to fault_wqh, it's not
1036 * pending anymore after we read it.
1038 * Use list_del() by hand (as
1039 * userfaultfd_wake_function also uses
1040 * list_del_init() by hand) to be sure nobody
1041 * changes __remove_wait_queue() to use
1042 * list_del_init() in turn breaking the
1043 * !list_empty_careful() check in
1044 * handle_userfault(). The uwq->wq.head list
1045 * must never be empty at any time during the
1046 * refile, or the waitqueue could disappear
1047 * from under us. The "wait_queue_head_t"
1048 * parameter of __remove_wait_queue() is unused
1051 list_del(&uwq
->wq
.entry
);
1052 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1054 write_seqcount_end(&ctx
->refile_seq
);
1056 /* careful to always initialize msg if ret == 0 */
1058 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1062 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1064 spin_lock(&ctx
->event_wqh
.lock
);
1065 uwq
= find_userfault_evt(ctx
);
1069 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1070 fork_nctx
= (struct userfaultfd_ctx
*)
1072 uwq
->msg
.arg
.reserved
.reserved1
;
1073 list_move(&uwq
->wq
.entry
, &fork_event
);
1075 * fork_nctx can be freed as soon as
1076 * we drop the lock, unless we take a
1079 userfaultfd_ctx_get(fork_nctx
);
1080 spin_unlock(&ctx
->event_wqh
.lock
);
1085 userfaultfd_event_complete(ctx
, uwq
);
1086 spin_unlock(&ctx
->event_wqh
.lock
);
1090 spin_unlock(&ctx
->event_wqh
.lock
);
1092 if (signal_pending(current
)) {
1100 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1102 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1104 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1105 __set_current_state(TASK_RUNNING
);
1106 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1108 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1109 ret
= resolve_userfault_fork(ctx
, fork_nctx
, msg
);
1110 spin_lock_irq(&ctx
->event_wqh
.lock
);
1111 if (!list_empty(&fork_event
)) {
1113 * The fork thread didn't abort, so we can
1114 * drop the temporary refcount.
1116 userfaultfd_ctx_put(fork_nctx
);
1118 uwq
= list_first_entry(&fork_event
,
1122 * If fork_event list wasn't empty and in turn
1123 * the event wasn't already released by fork
1124 * (the event is allocated on fork kernel
1125 * stack), put the event back to its place in
1126 * the event_wq. fork_event head will be freed
1127 * as soon as we return so the event cannot
1128 * stay queued there no matter the current
1131 list_del(&uwq
->wq
.entry
);
1132 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1135 * Leave the event in the waitqueue and report
1136 * error to userland if we failed to resolve
1137 * the userfault fork.
1140 userfaultfd_event_complete(ctx
, uwq
);
1143 * Here the fork thread aborted and the
1144 * refcount from the fork thread on fork_nctx
1145 * has already been released. We still hold
1146 * the reference we took before releasing the
1147 * lock above. If resolve_userfault_fork
1148 * failed we've to drop it because the
1149 * fork_nctx has to be freed in such case. If
1150 * it succeeded we'll hold it because the new
1151 * uffd references it.
1154 userfaultfd_ctx_put(fork_nctx
);
1156 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1162 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1163 size_t count
, loff_t
*ppos
)
1165 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1166 ssize_t _ret
, ret
= 0;
1167 struct uffd_msg msg
;
1168 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1170 if (ctx
->state
== UFFD_STATE_WAIT_API
)
1174 if (count
< sizeof(msg
))
1175 return ret
? ret
: -EINVAL
;
1176 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
);
1178 return ret
? ret
: _ret
;
1179 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1180 return ret
? ret
: -EFAULT
;
1183 count
-= sizeof(msg
);
1185 * Allow to read more than one fault at time but only
1186 * block if waiting for the very first one.
1188 no_wait
= O_NONBLOCK
;
1192 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1193 struct userfaultfd_wake_range
*range
)
1195 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1196 /* wake all in the range and autoremove */
1197 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1198 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1200 if (waitqueue_active(&ctx
->fault_wqh
))
1201 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1202 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1205 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1206 struct userfaultfd_wake_range
*range
)
1212 * To be sure waitqueue_active() is not reordered by the CPU
1213 * before the pagetable update, use an explicit SMP memory
1214 * barrier here. PT lock release or mmap_read_unlock(mm) still
1215 * have release semantics that can allow the
1216 * waitqueue_active() to be reordered before the pte update.
1221 * Use waitqueue_active because it's very frequent to
1222 * change the address space atomically even if there are no
1223 * userfaults yet. So we take the spinlock only when we're
1224 * sure we've userfaults to wake.
1227 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1228 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1229 waitqueue_active(&ctx
->fault_wqh
);
1231 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1233 __wake_userfault(ctx
, range
);
1236 static __always_inline
int validate_range(struct mm_struct
*mm
,
1237 __u64
*start
, __u64 len
)
1239 __u64 task_size
= mm
->task_size
;
1241 *start
= untagged_addr(*start
);
1243 if (*start
& ~PAGE_MASK
)
1245 if (len
& ~PAGE_MASK
)
1249 if (*start
< mmap_min_addr
)
1251 if (*start
>= task_size
)
1253 if (len
> task_size
- *start
)
1258 static inline bool vma_can_userfault(struct vm_area_struct
*vma
,
1259 unsigned long vm_flags
)
1261 /* FIXME: add WP support to hugetlbfs and shmem */
1262 return vma_is_anonymous(vma
) ||
1263 ((is_vm_hugetlb_page(vma
) || vma_is_shmem(vma
)) &&
1264 !(vm_flags
& VM_UFFD_WP
));
1267 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1270 struct mm_struct
*mm
= ctx
->mm
;
1271 struct vm_area_struct
*vma
, *prev
, *cur
;
1273 struct uffdio_register uffdio_register
;
1274 struct uffdio_register __user
*user_uffdio_register
;
1275 unsigned long vm_flags
, new_flags
;
1278 unsigned long start
, end
, vma_end
;
1280 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1283 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1284 sizeof(uffdio_register
)-sizeof(__u64
)))
1288 if (!uffdio_register
.mode
)
1290 if (uffdio_register
.mode
& ~(UFFDIO_REGISTER_MODE_MISSING
|
1291 UFFDIO_REGISTER_MODE_WP
))
1294 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1295 vm_flags
|= VM_UFFD_MISSING
;
1296 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
)
1297 vm_flags
|= VM_UFFD_WP
;
1299 ret
= validate_range(mm
, &uffdio_register
.range
.start
,
1300 uffdio_register
.range
.len
);
1304 start
= uffdio_register
.range
.start
;
1305 end
= start
+ uffdio_register
.range
.len
;
1308 if (!mmget_not_zero(mm
))
1311 mmap_write_lock(mm
);
1312 if (!mmget_still_valid(mm
))
1314 vma
= find_vma_prev(mm
, start
, &prev
);
1318 /* check that there's at least one vma in the range */
1320 if (vma
->vm_start
>= end
)
1324 * If the first vma contains huge pages, make sure start address
1325 * is aligned to huge page size.
1327 if (is_vm_hugetlb_page(vma
)) {
1328 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1330 if (start
& (vma_hpagesize
- 1))
1335 * Search for not compatible vmas.
1338 basic_ioctls
= false;
1339 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1342 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1343 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1345 /* check not compatible vmas */
1347 if (!vma_can_userfault(cur
, vm_flags
))
1351 * UFFDIO_COPY will fill file holes even without
1352 * PROT_WRITE. This check enforces that if this is a
1353 * MAP_SHARED, the process has write permission to the backing
1354 * file. If VM_MAYWRITE is set it also enforces that on a
1355 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1356 * F_WRITE_SEAL can be taken until the vma is destroyed.
1359 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1363 * If this vma contains ending address, and huge pages
1366 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1367 end
> cur
->vm_start
) {
1368 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1372 if (end
& (vma_hpagesize
- 1))
1375 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1379 * Check that this vma isn't already owned by a
1380 * different userfaultfd. We can't allow more than one
1381 * userfaultfd to own a single vma simultaneously or we
1382 * wouldn't know which one to deliver the userfaults to.
1385 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1386 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1390 * Note vmas containing huge pages
1392 if (is_vm_hugetlb_page(cur
))
1393 basic_ioctls
= true;
1399 if (vma
->vm_start
< start
)
1406 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1407 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1408 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1409 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1412 * Nothing to do: this vma is already registered into this
1413 * userfaultfd and with the right tracking mode too.
1415 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1416 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1419 if (vma
->vm_start
> start
)
1420 start
= vma
->vm_start
;
1421 vma_end
= min(end
, vma
->vm_end
);
1423 new_flags
= (vma
->vm_flags
&
1424 ~(VM_UFFD_MISSING
|VM_UFFD_WP
)) | vm_flags
;
1425 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1426 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1428 ((struct vm_userfaultfd_ctx
){ ctx
}));
1433 if (vma
->vm_start
< start
) {
1434 ret
= split_vma(mm
, vma
, start
, 1);
1438 if (vma
->vm_end
> end
) {
1439 ret
= split_vma(mm
, vma
, end
, 0);
1445 * In the vma_merge() successful mprotect-like case 8:
1446 * the next vma was merged into the current one and
1447 * the current one has not been updated yet.
1449 vma
->vm_flags
= new_flags
;
1450 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1454 start
= vma
->vm_end
;
1456 } while (vma
&& vma
->vm_start
< end
);
1458 mmap_write_unlock(mm
);
1463 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1464 UFFD_API_RANGE_IOCTLS
;
1467 * Declare the WP ioctl only if the WP mode is
1468 * specified and all checks passed with the range
1470 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1471 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1474 * Now that we scanned all vmas we can already tell
1475 * userland which ioctls methods are guaranteed to
1476 * succeed on this range.
1478 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1485 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1488 struct mm_struct
*mm
= ctx
->mm
;
1489 struct vm_area_struct
*vma
, *prev
, *cur
;
1491 struct uffdio_range uffdio_unregister
;
1492 unsigned long new_flags
;
1494 unsigned long start
, end
, vma_end
;
1495 const void __user
*buf
= (void __user
*)arg
;
1498 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1501 ret
= validate_range(mm
, &uffdio_unregister
.start
,
1502 uffdio_unregister
.len
);
1506 start
= uffdio_unregister
.start
;
1507 end
= start
+ uffdio_unregister
.len
;
1510 if (!mmget_not_zero(mm
))
1513 mmap_write_lock(mm
);
1514 if (!mmget_still_valid(mm
))
1516 vma
= find_vma_prev(mm
, start
, &prev
);
1520 /* check that there's at least one vma in the range */
1522 if (vma
->vm_start
>= end
)
1526 * If the first vma contains huge pages, make sure start address
1527 * is aligned to huge page size.
1529 if (is_vm_hugetlb_page(vma
)) {
1530 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1532 if (start
& (vma_hpagesize
- 1))
1537 * Search for not compatible vmas.
1541 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1544 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1545 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1548 * Check not compatible vmas, not strictly required
1549 * here as not compatible vmas cannot have an
1550 * userfaultfd_ctx registered on them, but this
1551 * provides for more strict behavior to notice
1552 * unregistration errors.
1554 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1561 if (vma
->vm_start
< start
)
1568 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1571 * Nothing to do: this vma is already registered into this
1572 * userfaultfd and with the right tracking mode too.
1574 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1577 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1579 if (vma
->vm_start
> start
)
1580 start
= vma
->vm_start
;
1581 vma_end
= min(end
, vma
->vm_end
);
1583 if (userfaultfd_missing(vma
)) {
1585 * Wake any concurrent pending userfault while
1586 * we unregister, so they will not hang
1587 * permanently and it avoids userland to call
1588 * UFFDIO_WAKE explicitly.
1590 struct userfaultfd_wake_range range
;
1591 range
.start
= start
;
1592 range
.len
= vma_end
- start
;
1593 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1596 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
1597 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1598 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1605 if (vma
->vm_start
< start
) {
1606 ret
= split_vma(mm
, vma
, start
, 1);
1610 if (vma
->vm_end
> end
) {
1611 ret
= split_vma(mm
, vma
, end
, 0);
1617 * In the vma_merge() successful mprotect-like case 8:
1618 * the next vma was merged into the current one and
1619 * the current one has not been updated yet.
1621 vma
->vm_flags
= new_flags
;
1622 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1626 start
= vma
->vm_end
;
1628 } while (vma
&& vma
->vm_start
< end
);
1630 mmap_write_unlock(mm
);
1637 * userfaultfd_wake may be used in combination with the
1638 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1640 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1644 struct uffdio_range uffdio_wake
;
1645 struct userfaultfd_wake_range range
;
1646 const void __user
*buf
= (void __user
*)arg
;
1649 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1652 ret
= validate_range(ctx
->mm
, &uffdio_wake
.start
, uffdio_wake
.len
);
1656 range
.start
= uffdio_wake
.start
;
1657 range
.len
= uffdio_wake
.len
;
1660 * len == 0 means wake all and we don't want to wake all here,
1661 * so check it again to be sure.
1663 VM_BUG_ON(!range
.len
);
1665 wake_userfault(ctx
, &range
);
1672 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1676 struct uffdio_copy uffdio_copy
;
1677 struct uffdio_copy __user
*user_uffdio_copy
;
1678 struct userfaultfd_wake_range range
;
1680 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1683 if (READ_ONCE(ctx
->mmap_changing
))
1687 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1688 /* don't copy "copy" last field */
1689 sizeof(uffdio_copy
)-sizeof(__s64
)))
1692 ret
= validate_range(ctx
->mm
, &uffdio_copy
.dst
, uffdio_copy
.len
);
1696 * double check for wraparound just in case. copy_from_user()
1697 * will later check uffdio_copy.src + uffdio_copy.len to fit
1698 * in the userland range.
1701 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1703 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1705 if (mmget_not_zero(ctx
->mm
)) {
1706 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1707 uffdio_copy
.len
, &ctx
->mmap_changing
,
1713 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1718 /* len == 0 would wake all */
1720 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1721 range
.start
= uffdio_copy
.dst
;
1722 wake_userfault(ctx
, &range
);
1724 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1729 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1733 struct uffdio_zeropage uffdio_zeropage
;
1734 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1735 struct userfaultfd_wake_range range
;
1737 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1740 if (READ_ONCE(ctx
->mmap_changing
))
1744 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1745 /* don't copy "zeropage" last field */
1746 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1749 ret
= validate_range(ctx
->mm
, &uffdio_zeropage
.range
.start
,
1750 uffdio_zeropage
.range
.len
);
1754 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1757 if (mmget_not_zero(ctx
->mm
)) {
1758 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1759 uffdio_zeropage
.range
.len
,
1760 &ctx
->mmap_changing
);
1765 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1769 /* len == 0 would wake all */
1772 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1773 range
.start
= uffdio_zeropage
.range
.start
;
1774 wake_userfault(ctx
, &range
);
1776 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1781 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1785 struct uffdio_writeprotect uffdio_wp
;
1786 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1787 struct userfaultfd_wake_range range
;
1788 bool mode_wp
, mode_dontwake
;
1790 if (READ_ONCE(ctx
->mmap_changing
))
1793 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1795 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1796 sizeof(struct uffdio_writeprotect
)))
1799 ret
= validate_range(ctx
->mm
, &uffdio_wp
.range
.start
,
1800 uffdio_wp
.range
.len
);
1804 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1805 UFFDIO_WRITEPROTECT_MODE_WP
))
1808 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1809 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1811 if (mode_wp
&& mode_dontwake
)
1814 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1815 uffdio_wp
.range
.len
, mode_wp
,
1816 &ctx
->mmap_changing
);
1820 if (!mode_wp
&& !mode_dontwake
) {
1821 range
.start
= uffdio_wp
.range
.start
;
1822 range
.len
= uffdio_wp
.range
.len
;
1823 wake_userfault(ctx
, &range
);
1828 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1831 * For the current set of features the bits just coincide
1833 return (unsigned int)user_features
;
1837 * userland asks for a certain API version and we return which bits
1838 * and ioctl commands are implemented in this kernel for such API
1839 * version or -EINVAL if unknown.
1841 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1844 struct uffdio_api uffdio_api
;
1845 void __user
*buf
= (void __user
*)arg
;
1850 if (ctx
->state
!= UFFD_STATE_WAIT_API
)
1853 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1855 features
= uffdio_api
.features
;
1857 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
1860 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1862 /* report all available features and ioctls to userland */
1863 uffdio_api
.features
= UFFD_API_FEATURES
;
1864 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1866 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1868 ctx
->state
= UFFD_STATE_RUNNING
;
1869 /* only enable the requested features for this uffd context */
1870 ctx
->features
= uffd_ctx_features(features
);
1875 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1876 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1881 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1885 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1887 if (cmd
!= UFFDIO_API
&& ctx
->state
== UFFD_STATE_WAIT_API
)
1892 ret
= userfaultfd_api(ctx
, arg
);
1894 case UFFDIO_REGISTER
:
1895 ret
= userfaultfd_register(ctx
, arg
);
1897 case UFFDIO_UNREGISTER
:
1898 ret
= userfaultfd_unregister(ctx
, arg
);
1901 ret
= userfaultfd_wake(ctx
, arg
);
1904 ret
= userfaultfd_copy(ctx
, arg
);
1906 case UFFDIO_ZEROPAGE
:
1907 ret
= userfaultfd_zeropage(ctx
, arg
);
1909 case UFFDIO_WRITEPROTECT
:
1910 ret
= userfaultfd_writeprotect(ctx
, arg
);
1916 #ifdef CONFIG_PROC_FS
1917 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
1919 struct userfaultfd_ctx
*ctx
= f
->private_data
;
1920 wait_queue_entry_t
*wq
;
1921 unsigned long pending
= 0, total
= 0;
1923 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1924 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
1928 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
1931 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1934 * If more protocols will be added, there will be all shown
1935 * separated by a space. Like this:
1936 * protocols: aa:... bb:...
1938 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1939 pending
, total
, UFFD_API
, ctx
->features
,
1940 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
1944 static const struct file_operations userfaultfd_fops
= {
1945 #ifdef CONFIG_PROC_FS
1946 .show_fdinfo
= userfaultfd_show_fdinfo
,
1948 .release
= userfaultfd_release
,
1949 .poll
= userfaultfd_poll
,
1950 .read
= userfaultfd_read
,
1951 .unlocked_ioctl
= userfaultfd_ioctl
,
1952 .compat_ioctl
= compat_ptr_ioctl
,
1953 .llseek
= noop_llseek
,
1956 static void init_once_userfaultfd_ctx(void *mem
)
1958 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
1960 init_waitqueue_head(&ctx
->fault_pending_wqh
);
1961 init_waitqueue_head(&ctx
->fault_wqh
);
1962 init_waitqueue_head(&ctx
->event_wqh
);
1963 init_waitqueue_head(&ctx
->fd_wqh
);
1964 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
1967 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
1969 struct userfaultfd_ctx
*ctx
;
1972 if (!sysctl_unprivileged_userfaultfd
&& !capable(CAP_SYS_PTRACE
))
1975 BUG_ON(!current
->mm
);
1977 /* Check the UFFD_* constants for consistency. */
1978 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
1979 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
1981 if (flags
& ~UFFD_SHARED_FCNTL_FLAGS
)
1984 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
1988 refcount_set(&ctx
->refcount
, 1);
1991 ctx
->state
= UFFD_STATE_WAIT_API
;
1992 ctx
->released
= false;
1993 ctx
->mmap_changing
= false;
1994 ctx
->mm
= current
->mm
;
1995 /* prevent the mm struct to be freed */
1998 fd
= anon_inode_getfd("[userfaultfd]", &userfaultfd_fops
, ctx
,
1999 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
));
2002 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2007 static int __init
userfaultfd_init(void)
2009 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2010 sizeof(struct userfaultfd_ctx
),
2012 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2013 init_once_userfaultfd_ctx
);
2016 __initcall(userfaultfd_init
);