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 struct seqcount 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 /* Should pair with userfaultfd_signal_pending() */
343 static inline long userfaultfd_get_blocking_state(unsigned int flags
)
345 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
346 return TASK_INTERRUPTIBLE
;
348 if (flags
& FAULT_FLAG_KILLABLE
)
349 return TASK_KILLABLE
;
351 return TASK_UNINTERRUPTIBLE
;
354 /* Should pair with userfaultfd_get_blocking_state() */
355 static inline bool userfaultfd_signal_pending(unsigned int flags
)
357 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
358 return signal_pending(current
);
360 if (flags
& FAULT_FLAG_KILLABLE
)
361 return fatal_signal_pending(current
);
367 * The locking rules involved in returning VM_FAULT_RETRY depending on
368 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
369 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
370 * recommendation in __lock_page_or_retry is not an understatement.
372 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
373 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
376 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
377 * set, VM_FAULT_RETRY can still be returned if and only if there are
378 * fatal_signal_pending()s, and the mmap_sem must be released before
381 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
383 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
384 struct userfaultfd_ctx
*ctx
;
385 struct userfaultfd_wait_queue uwq
;
386 vm_fault_t ret
= VM_FAULT_SIGBUS
;
391 * We don't do userfault handling for the final child pid update.
393 * We also don't do userfault handling during
394 * coredumping. hugetlbfs has the special
395 * follow_hugetlb_page() to skip missing pages in the
396 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
397 * the no_page_table() helper in follow_page_mask(), but the
398 * shmem_vm_ops->fault method is invoked even during
399 * coredumping without mmap_sem and it ends up here.
401 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
405 * Coredumping runs without mmap_sem so we can only check that
406 * the mmap_sem is held, if PF_DUMPCORE was not set.
408 mmap_assert_locked(mm
);
410 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
414 BUG_ON(ctx
->mm
!= mm
);
416 VM_BUG_ON(reason
& ~(VM_UFFD_MISSING
|VM_UFFD_WP
));
417 VM_BUG_ON(!(reason
& VM_UFFD_MISSING
) ^ !!(reason
& VM_UFFD_WP
));
419 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
423 * If it's already released don't get it. This avoids to loop
424 * in __get_user_pages if userfaultfd_release waits on the
425 * caller of handle_userfault to release the mmap_sem.
427 if (unlikely(READ_ONCE(ctx
->released
))) {
429 * Don't return VM_FAULT_SIGBUS in this case, so a non
430 * cooperative manager can close the uffd after the
431 * last UFFDIO_COPY, without risking to trigger an
432 * involuntary SIGBUS if the process was starting the
433 * userfaultfd while the userfaultfd was still armed
434 * (but after the last UFFDIO_COPY). If the uffd
435 * wasn't already closed when the userfault reached
436 * this point, that would normally be solved by
437 * userfaultfd_must_wait returning 'false'.
439 * If we were to return VM_FAULT_SIGBUS here, the non
440 * cooperative manager would be instead forced to
441 * always call UFFDIO_UNREGISTER before it can safely
444 ret
= VM_FAULT_NOPAGE
;
449 * Check that we can return VM_FAULT_RETRY.
451 * NOTE: it should become possible to return VM_FAULT_RETRY
452 * even if FAULT_FLAG_TRIED is set without leading to gup()
453 * -EBUSY failures, if the userfaultfd is to be extended for
454 * VM_UFFD_WP tracking and we intend to arm the userfault
455 * without first stopping userland access to the memory. For
456 * VM_UFFD_MISSING userfaults this is enough for now.
458 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
460 * Validate the invariant that nowait must allow retry
461 * to be sure not to return SIGBUS erroneously on
462 * nowait invocations.
464 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
465 #ifdef CONFIG_DEBUG_VM
466 if (printk_ratelimit()) {
468 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
477 * Handle nowait, not much to do other than tell it to retry
480 ret
= VM_FAULT_RETRY
;
481 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
484 /* take the reference before dropping the mmap_sem */
485 userfaultfd_ctx_get(ctx
);
487 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
488 uwq
.wq
.private = current
;
489 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
,
494 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
496 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
498 * After the __add_wait_queue the uwq is visible to userland
499 * through poll/read().
501 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
503 * The smp_mb() after __set_current_state prevents the reads
504 * following the spin_unlock to happen before the list_add in
507 set_current_state(blocking_state
);
508 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
510 if (!is_vm_hugetlb_page(vmf
->vma
))
511 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
514 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
517 mmap_read_unlock(mm
);
519 if (likely(must_wait
&& !READ_ONCE(ctx
->released
) &&
520 !userfaultfd_signal_pending(vmf
->flags
))) {
521 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
523 ret
|= VM_FAULT_MAJOR
;
526 * False wakeups can orginate even from rwsem before
527 * up_read() however userfaults will wait either for a
528 * targeted wakeup on the specific uwq waitqueue from
529 * wake_userfault() or for signals or for uffd
532 while (!READ_ONCE(uwq
.waken
)) {
534 * This needs the full smp_store_mb()
535 * guarantee as the state write must be
536 * visible to other CPUs before reading
537 * uwq.waken from other CPUs.
539 set_current_state(blocking_state
);
540 if (READ_ONCE(uwq
.waken
) ||
541 READ_ONCE(ctx
->released
) ||
542 userfaultfd_signal_pending(vmf
->flags
))
548 __set_current_state(TASK_RUNNING
);
551 * Here we race with the list_del; list_add in
552 * userfaultfd_ctx_read(), however because we don't ever run
553 * list_del_init() to refile across the two lists, the prev
554 * and next pointers will never point to self. list_add also
555 * would never let any of the two pointers to point to
556 * self. So list_empty_careful won't risk to see both pointers
557 * pointing to self at any time during the list refile. The
558 * only case where list_del_init() is called is the full
559 * removal in the wake function and there we don't re-list_add
560 * and it's fine not to block on the spinlock. The uwq on this
561 * kernel stack can be released after the list_del_init.
563 if (!list_empty_careful(&uwq
.wq
.entry
)) {
564 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
566 * No need of list_del_init(), the uwq on the stack
567 * will be freed shortly anyway.
569 list_del(&uwq
.wq
.entry
);
570 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
574 * ctx may go away after this if the userfault pseudo fd is
577 userfaultfd_ctx_put(ctx
);
583 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
584 struct userfaultfd_wait_queue
*ewq
)
586 struct userfaultfd_ctx
*release_new_ctx
;
588 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
592 init_waitqueue_entry(&ewq
->wq
, current
);
593 release_new_ctx
= NULL
;
595 spin_lock_irq(&ctx
->event_wqh
.lock
);
597 * After the __add_wait_queue the uwq is visible to userland
598 * through poll/read().
600 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
602 set_current_state(TASK_KILLABLE
);
603 if (ewq
->msg
.event
== 0)
605 if (READ_ONCE(ctx
->released
) ||
606 fatal_signal_pending(current
)) {
608 * &ewq->wq may be queued in fork_event, but
609 * __remove_wait_queue ignores the head
610 * parameter. It would be a problem if it
613 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
614 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
615 struct userfaultfd_ctx
*new;
617 new = (struct userfaultfd_ctx
*)
619 ewq
->msg
.arg
.reserved
.reserved1
;
620 release_new_ctx
= new;
625 spin_unlock_irq(&ctx
->event_wqh
.lock
);
627 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
630 spin_lock_irq(&ctx
->event_wqh
.lock
);
632 __set_current_state(TASK_RUNNING
);
633 spin_unlock_irq(&ctx
->event_wqh
.lock
);
635 if (release_new_ctx
) {
636 struct vm_area_struct
*vma
;
637 struct mm_struct
*mm
= release_new_ctx
->mm
;
639 /* the various vma->vm_userfaultfd_ctx still points to it */
641 /* no task can run (and in turn coredump) yet */
642 VM_WARN_ON(!mmget_still_valid(mm
));
643 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
644 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
645 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
646 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
648 mmap_write_unlock(mm
);
650 userfaultfd_ctx_put(release_new_ctx
);
654 * ctx may go away after this if the userfault pseudo fd is
658 WRITE_ONCE(ctx
->mmap_changing
, false);
659 userfaultfd_ctx_put(ctx
);
662 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
663 struct userfaultfd_wait_queue
*ewq
)
666 wake_up_locked(&ctx
->event_wqh
);
667 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
670 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
672 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
673 struct userfaultfd_fork_ctx
*fctx
;
675 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
676 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
677 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
678 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
682 list_for_each_entry(fctx
, fcs
, list
)
683 if (fctx
->orig
== octx
) {
689 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
693 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
699 refcount_set(&ctx
->refcount
, 1);
700 ctx
->flags
= octx
->flags
;
701 ctx
->state
= UFFD_STATE_RUNNING
;
702 ctx
->features
= octx
->features
;
703 ctx
->released
= false;
704 ctx
->mmap_changing
= false;
705 ctx
->mm
= vma
->vm_mm
;
708 userfaultfd_ctx_get(octx
);
709 WRITE_ONCE(octx
->mmap_changing
, true);
712 list_add_tail(&fctx
->list
, fcs
);
715 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
719 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
721 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
722 struct userfaultfd_wait_queue ewq
;
726 ewq
.msg
.event
= UFFD_EVENT_FORK
;
727 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
729 userfaultfd_event_wait_completion(ctx
, &ewq
);
732 void dup_userfaultfd_complete(struct list_head
*fcs
)
734 struct userfaultfd_fork_ctx
*fctx
, *n
;
736 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
738 list_del(&fctx
->list
);
743 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
744 struct vm_userfaultfd_ctx
*vm_ctx
)
746 struct userfaultfd_ctx
*ctx
;
748 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
753 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
755 userfaultfd_ctx_get(ctx
);
756 WRITE_ONCE(ctx
->mmap_changing
, true);
758 /* Drop uffd context if remap feature not enabled */
759 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
760 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
764 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
765 unsigned long from
, unsigned long to
,
768 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
769 struct userfaultfd_wait_queue ewq
;
774 if (to
& ~PAGE_MASK
) {
775 userfaultfd_ctx_put(ctx
);
781 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
782 ewq
.msg
.arg
.remap
.from
= from
;
783 ewq
.msg
.arg
.remap
.to
= to
;
784 ewq
.msg
.arg
.remap
.len
= len
;
786 userfaultfd_event_wait_completion(ctx
, &ewq
);
789 bool userfaultfd_remove(struct vm_area_struct
*vma
,
790 unsigned long start
, unsigned long end
)
792 struct mm_struct
*mm
= vma
->vm_mm
;
793 struct userfaultfd_ctx
*ctx
;
794 struct userfaultfd_wait_queue ewq
;
796 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
797 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
800 userfaultfd_ctx_get(ctx
);
801 WRITE_ONCE(ctx
->mmap_changing
, true);
802 mmap_read_unlock(mm
);
806 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
807 ewq
.msg
.arg
.remove
.start
= start
;
808 ewq
.msg
.arg
.remove
.end
= end
;
810 userfaultfd_event_wait_completion(ctx
, &ewq
);
815 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
816 unsigned long start
, unsigned long end
)
818 struct userfaultfd_unmap_ctx
*unmap_ctx
;
820 list_for_each_entry(unmap_ctx
, unmaps
, list
)
821 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
822 unmap_ctx
->end
== end
)
828 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
829 unsigned long start
, unsigned long end
,
830 struct list_head
*unmaps
)
832 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
833 struct userfaultfd_unmap_ctx
*unmap_ctx
;
834 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
836 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
837 has_unmap_ctx(ctx
, unmaps
, start
, end
))
840 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
844 userfaultfd_ctx_get(ctx
);
845 WRITE_ONCE(ctx
->mmap_changing
, true);
846 unmap_ctx
->ctx
= ctx
;
847 unmap_ctx
->start
= start
;
848 unmap_ctx
->end
= end
;
849 list_add_tail(&unmap_ctx
->list
, unmaps
);
855 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
857 struct userfaultfd_unmap_ctx
*ctx
, *n
;
858 struct userfaultfd_wait_queue ewq
;
860 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
863 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
864 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
865 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
867 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
869 list_del(&ctx
->list
);
874 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
876 struct userfaultfd_ctx
*ctx
= file
->private_data
;
877 struct mm_struct
*mm
= ctx
->mm
;
878 struct vm_area_struct
*vma
, *prev
;
879 /* len == 0 means wake all */
880 struct userfaultfd_wake_range range
= { .len
= 0, };
881 unsigned long new_flags
;
884 WRITE_ONCE(ctx
->released
, true);
886 if (!mmget_not_zero(mm
))
890 * Flush page faults out of all CPUs. NOTE: all page faults
891 * must be retried without returning VM_FAULT_SIGBUS if
892 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
893 * changes while handle_userfault released the mmap_sem. So
894 * it's critical that released is set to true (above), before
895 * taking the mmap_sem for writing.
898 still_valid
= mmget_still_valid(mm
);
900 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
902 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
903 !!(vma
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
904 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
908 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
910 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
911 new_flags
, vma
->anon_vma
,
912 vma
->vm_file
, vma
->vm_pgoff
,
920 vma
->vm_flags
= new_flags
;
921 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
923 mmap_write_unlock(mm
);
927 * After no new page faults can wait on this fault_*wqh, flush
928 * the last page faults that may have been already waiting on
931 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
932 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
933 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
934 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
936 /* Flush pending events that may still wait on event_wqh */
937 wake_up_all(&ctx
->event_wqh
);
939 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
940 userfaultfd_ctx_put(ctx
);
944 /* fault_pending_wqh.lock must be hold by the caller */
945 static inline struct userfaultfd_wait_queue
*find_userfault_in(
946 wait_queue_head_t
*wqh
)
948 wait_queue_entry_t
*wq
;
949 struct userfaultfd_wait_queue
*uwq
;
951 lockdep_assert_held(&wqh
->lock
);
954 if (!waitqueue_active(wqh
))
956 /* walk in reverse to provide FIFO behavior to read userfaults */
957 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
958 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
963 static inline struct userfaultfd_wait_queue
*find_userfault(
964 struct userfaultfd_ctx
*ctx
)
966 return find_userfault_in(&ctx
->fault_pending_wqh
);
969 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
970 struct userfaultfd_ctx
*ctx
)
972 return find_userfault_in(&ctx
->event_wqh
);
975 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
977 struct userfaultfd_ctx
*ctx
= file
->private_data
;
980 poll_wait(file
, &ctx
->fd_wqh
, wait
);
982 switch (ctx
->state
) {
983 case UFFD_STATE_WAIT_API
:
985 case UFFD_STATE_RUNNING
:
987 * poll() never guarantees that read won't block.
988 * userfaults can be waken before they're read().
990 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
993 * lockless access to see if there are pending faults
994 * __pollwait last action is the add_wait_queue but
995 * the spin_unlock would allow the waitqueue_active to
996 * pass above the actual list_add inside
997 * add_wait_queue critical section. So use a full
998 * memory barrier to serialize the list_add write of
999 * add_wait_queue() with the waitqueue_active read
1004 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1006 else if (waitqueue_active(&ctx
->event_wqh
))
1016 static const struct file_operations userfaultfd_fops
;
1018 static int resolve_userfault_fork(struct userfaultfd_ctx
*ctx
,
1019 struct userfaultfd_ctx
*new,
1020 struct uffd_msg
*msg
)
1024 fd
= anon_inode_getfd("[userfaultfd]", &userfaultfd_fops
, new,
1025 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
));
1029 msg
->arg
.reserved
.reserved1
= 0;
1030 msg
->arg
.fork
.ufd
= fd
;
1034 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1035 struct uffd_msg
*msg
)
1038 DECLARE_WAITQUEUE(wait
, current
);
1039 struct userfaultfd_wait_queue
*uwq
;
1041 * Handling fork event requires sleeping operations, so
1042 * we drop the event_wqh lock, then do these ops, then
1043 * lock it back and wake up the waiter. While the lock is
1044 * dropped the ewq may go away so we keep track of it
1047 LIST_HEAD(fork_event
);
1048 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1050 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1051 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1052 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1054 set_current_state(TASK_INTERRUPTIBLE
);
1055 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1056 uwq
= find_userfault(ctx
);
1059 * Use a seqcount to repeat the lockless check
1060 * in wake_userfault() to avoid missing
1061 * wakeups because during the refile both
1062 * waitqueue could become empty if this is the
1065 write_seqcount_begin(&ctx
->refile_seq
);
1068 * The fault_pending_wqh.lock prevents the uwq
1069 * to disappear from under us.
1071 * Refile this userfault from
1072 * fault_pending_wqh to fault_wqh, it's not
1073 * pending anymore after we read it.
1075 * Use list_del() by hand (as
1076 * userfaultfd_wake_function also uses
1077 * list_del_init() by hand) to be sure nobody
1078 * changes __remove_wait_queue() to use
1079 * list_del_init() in turn breaking the
1080 * !list_empty_careful() check in
1081 * handle_userfault(). The uwq->wq.head list
1082 * must never be empty at any time during the
1083 * refile, or the waitqueue could disappear
1084 * from under us. The "wait_queue_head_t"
1085 * parameter of __remove_wait_queue() is unused
1088 list_del(&uwq
->wq
.entry
);
1089 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1091 write_seqcount_end(&ctx
->refile_seq
);
1093 /* careful to always initialize msg if ret == 0 */
1095 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1099 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1101 spin_lock(&ctx
->event_wqh
.lock
);
1102 uwq
= find_userfault_evt(ctx
);
1106 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1107 fork_nctx
= (struct userfaultfd_ctx
*)
1109 uwq
->msg
.arg
.reserved
.reserved1
;
1110 list_move(&uwq
->wq
.entry
, &fork_event
);
1112 * fork_nctx can be freed as soon as
1113 * we drop the lock, unless we take a
1116 userfaultfd_ctx_get(fork_nctx
);
1117 spin_unlock(&ctx
->event_wqh
.lock
);
1122 userfaultfd_event_complete(ctx
, uwq
);
1123 spin_unlock(&ctx
->event_wqh
.lock
);
1127 spin_unlock(&ctx
->event_wqh
.lock
);
1129 if (signal_pending(current
)) {
1137 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1139 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1141 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1142 __set_current_state(TASK_RUNNING
);
1143 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1145 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1146 ret
= resolve_userfault_fork(ctx
, fork_nctx
, msg
);
1147 spin_lock_irq(&ctx
->event_wqh
.lock
);
1148 if (!list_empty(&fork_event
)) {
1150 * The fork thread didn't abort, so we can
1151 * drop the temporary refcount.
1153 userfaultfd_ctx_put(fork_nctx
);
1155 uwq
= list_first_entry(&fork_event
,
1159 * If fork_event list wasn't empty and in turn
1160 * the event wasn't already released by fork
1161 * (the event is allocated on fork kernel
1162 * stack), put the event back to its place in
1163 * the event_wq. fork_event head will be freed
1164 * as soon as we return so the event cannot
1165 * stay queued there no matter the current
1168 list_del(&uwq
->wq
.entry
);
1169 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1172 * Leave the event in the waitqueue and report
1173 * error to userland if we failed to resolve
1174 * the userfault fork.
1177 userfaultfd_event_complete(ctx
, uwq
);
1180 * Here the fork thread aborted and the
1181 * refcount from the fork thread on fork_nctx
1182 * has already been released. We still hold
1183 * the reference we took before releasing the
1184 * lock above. If resolve_userfault_fork
1185 * failed we've to drop it because the
1186 * fork_nctx has to be freed in such case. If
1187 * it succeeded we'll hold it because the new
1188 * uffd references it.
1191 userfaultfd_ctx_put(fork_nctx
);
1193 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1199 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1200 size_t count
, loff_t
*ppos
)
1202 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1203 ssize_t _ret
, ret
= 0;
1204 struct uffd_msg msg
;
1205 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1207 if (ctx
->state
== UFFD_STATE_WAIT_API
)
1211 if (count
< sizeof(msg
))
1212 return ret
? ret
: -EINVAL
;
1213 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
);
1215 return ret
? ret
: _ret
;
1216 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1217 return ret
? ret
: -EFAULT
;
1220 count
-= sizeof(msg
);
1222 * Allow to read more than one fault at time but only
1223 * block if waiting for the very first one.
1225 no_wait
= O_NONBLOCK
;
1229 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1230 struct userfaultfd_wake_range
*range
)
1232 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1233 /* wake all in the range and autoremove */
1234 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1235 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1237 if (waitqueue_active(&ctx
->fault_wqh
))
1238 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1239 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1242 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1243 struct userfaultfd_wake_range
*range
)
1249 * To be sure waitqueue_active() is not reordered by the CPU
1250 * before the pagetable update, use an explicit SMP memory
1251 * barrier here. PT lock release or mmap_read_unlock(mm) still
1252 * have release semantics that can allow the
1253 * waitqueue_active() to be reordered before the pte update.
1258 * Use waitqueue_active because it's very frequent to
1259 * change the address space atomically even if there are no
1260 * userfaults yet. So we take the spinlock only when we're
1261 * sure we've userfaults to wake.
1264 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1265 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1266 waitqueue_active(&ctx
->fault_wqh
);
1268 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1270 __wake_userfault(ctx
, range
);
1273 static __always_inline
int validate_range(struct mm_struct
*mm
,
1274 __u64
*start
, __u64 len
)
1276 __u64 task_size
= mm
->task_size
;
1278 *start
= untagged_addr(*start
);
1280 if (*start
& ~PAGE_MASK
)
1282 if (len
& ~PAGE_MASK
)
1286 if (*start
< mmap_min_addr
)
1288 if (*start
>= task_size
)
1290 if (len
> task_size
- *start
)
1295 static inline bool vma_can_userfault(struct vm_area_struct
*vma
,
1296 unsigned long vm_flags
)
1298 /* FIXME: add WP support to hugetlbfs and shmem */
1299 return vma_is_anonymous(vma
) ||
1300 ((is_vm_hugetlb_page(vma
) || vma_is_shmem(vma
)) &&
1301 !(vm_flags
& VM_UFFD_WP
));
1304 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1307 struct mm_struct
*mm
= ctx
->mm
;
1308 struct vm_area_struct
*vma
, *prev
, *cur
;
1310 struct uffdio_register uffdio_register
;
1311 struct uffdio_register __user
*user_uffdio_register
;
1312 unsigned long vm_flags
, new_flags
;
1315 unsigned long start
, end
, vma_end
;
1317 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1320 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1321 sizeof(uffdio_register
)-sizeof(__u64
)))
1325 if (!uffdio_register
.mode
)
1327 if (uffdio_register
.mode
& ~(UFFDIO_REGISTER_MODE_MISSING
|
1328 UFFDIO_REGISTER_MODE_WP
))
1331 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1332 vm_flags
|= VM_UFFD_MISSING
;
1333 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
)
1334 vm_flags
|= VM_UFFD_WP
;
1336 ret
= validate_range(mm
, &uffdio_register
.range
.start
,
1337 uffdio_register
.range
.len
);
1341 start
= uffdio_register
.range
.start
;
1342 end
= start
+ uffdio_register
.range
.len
;
1345 if (!mmget_not_zero(mm
))
1348 mmap_write_lock(mm
);
1349 if (!mmget_still_valid(mm
))
1351 vma
= find_vma_prev(mm
, start
, &prev
);
1355 /* check that there's at least one vma in the range */
1357 if (vma
->vm_start
>= end
)
1361 * If the first vma contains huge pages, make sure start address
1362 * is aligned to huge page size.
1364 if (is_vm_hugetlb_page(vma
)) {
1365 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1367 if (start
& (vma_hpagesize
- 1))
1372 * Search for not compatible vmas.
1375 basic_ioctls
= false;
1376 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1379 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1380 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1382 /* check not compatible vmas */
1384 if (!vma_can_userfault(cur
, vm_flags
))
1388 * UFFDIO_COPY will fill file holes even without
1389 * PROT_WRITE. This check enforces that if this is a
1390 * MAP_SHARED, the process has write permission to the backing
1391 * file. If VM_MAYWRITE is set it also enforces that on a
1392 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1393 * F_WRITE_SEAL can be taken until the vma is destroyed.
1396 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1400 * If this vma contains ending address, and huge pages
1403 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1404 end
> cur
->vm_start
) {
1405 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1409 if (end
& (vma_hpagesize
- 1))
1412 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1416 * Check that this vma isn't already owned by a
1417 * different userfaultfd. We can't allow more than one
1418 * userfaultfd to own a single vma simultaneously or we
1419 * wouldn't know which one to deliver the userfaults to.
1422 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1423 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1427 * Note vmas containing huge pages
1429 if (is_vm_hugetlb_page(cur
))
1430 basic_ioctls
= true;
1436 if (vma
->vm_start
< start
)
1443 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1444 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1445 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1446 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1449 * Nothing to do: this vma is already registered into this
1450 * userfaultfd and with the right tracking mode too.
1452 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1453 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1456 if (vma
->vm_start
> start
)
1457 start
= vma
->vm_start
;
1458 vma_end
= min(end
, vma
->vm_end
);
1460 new_flags
= (vma
->vm_flags
&
1461 ~(VM_UFFD_MISSING
|VM_UFFD_WP
)) | vm_flags
;
1462 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1463 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1465 ((struct vm_userfaultfd_ctx
){ ctx
}));
1470 if (vma
->vm_start
< start
) {
1471 ret
= split_vma(mm
, vma
, start
, 1);
1475 if (vma
->vm_end
> end
) {
1476 ret
= split_vma(mm
, vma
, end
, 0);
1482 * In the vma_merge() successful mprotect-like case 8:
1483 * the next vma was merged into the current one and
1484 * the current one has not been updated yet.
1486 vma
->vm_flags
= new_flags
;
1487 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1491 start
= vma
->vm_end
;
1493 } while (vma
&& vma
->vm_start
< end
);
1495 mmap_write_unlock(mm
);
1500 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1501 UFFD_API_RANGE_IOCTLS
;
1504 * Declare the WP ioctl only if the WP mode is
1505 * specified and all checks passed with the range
1507 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1508 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1511 * Now that we scanned all vmas we can already tell
1512 * userland which ioctls methods are guaranteed to
1513 * succeed on this range.
1515 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1522 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1525 struct mm_struct
*mm
= ctx
->mm
;
1526 struct vm_area_struct
*vma
, *prev
, *cur
;
1528 struct uffdio_range uffdio_unregister
;
1529 unsigned long new_flags
;
1531 unsigned long start
, end
, vma_end
;
1532 const void __user
*buf
= (void __user
*)arg
;
1535 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1538 ret
= validate_range(mm
, &uffdio_unregister
.start
,
1539 uffdio_unregister
.len
);
1543 start
= uffdio_unregister
.start
;
1544 end
= start
+ uffdio_unregister
.len
;
1547 if (!mmget_not_zero(mm
))
1550 mmap_write_lock(mm
);
1551 if (!mmget_still_valid(mm
))
1553 vma
= find_vma_prev(mm
, start
, &prev
);
1557 /* check that there's at least one vma in the range */
1559 if (vma
->vm_start
>= end
)
1563 * If the first vma contains huge pages, make sure start address
1564 * is aligned to huge page size.
1566 if (is_vm_hugetlb_page(vma
)) {
1567 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1569 if (start
& (vma_hpagesize
- 1))
1574 * Search for not compatible vmas.
1578 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1581 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1582 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1585 * Check not compatible vmas, not strictly required
1586 * here as not compatible vmas cannot have an
1587 * userfaultfd_ctx registered on them, but this
1588 * provides for more strict behavior to notice
1589 * unregistration errors.
1591 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1598 if (vma
->vm_start
< start
)
1605 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1608 * Nothing to do: this vma is already registered into this
1609 * userfaultfd and with the right tracking mode too.
1611 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1614 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1616 if (vma
->vm_start
> start
)
1617 start
= vma
->vm_start
;
1618 vma_end
= min(end
, vma
->vm_end
);
1620 if (userfaultfd_missing(vma
)) {
1622 * Wake any concurrent pending userfault while
1623 * we unregister, so they will not hang
1624 * permanently and it avoids userland to call
1625 * UFFDIO_WAKE explicitly.
1627 struct userfaultfd_wake_range range
;
1628 range
.start
= start
;
1629 range
.len
= vma_end
- start
;
1630 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1633 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
1634 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1635 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1642 if (vma
->vm_start
< start
) {
1643 ret
= split_vma(mm
, vma
, start
, 1);
1647 if (vma
->vm_end
> end
) {
1648 ret
= split_vma(mm
, vma
, end
, 0);
1654 * In the vma_merge() successful mprotect-like case 8:
1655 * the next vma was merged into the current one and
1656 * the current one has not been updated yet.
1658 vma
->vm_flags
= new_flags
;
1659 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1663 start
= vma
->vm_end
;
1665 } while (vma
&& vma
->vm_start
< end
);
1667 mmap_write_unlock(mm
);
1674 * userfaultfd_wake may be used in combination with the
1675 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1677 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1681 struct uffdio_range uffdio_wake
;
1682 struct userfaultfd_wake_range range
;
1683 const void __user
*buf
= (void __user
*)arg
;
1686 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1689 ret
= validate_range(ctx
->mm
, &uffdio_wake
.start
, uffdio_wake
.len
);
1693 range
.start
= uffdio_wake
.start
;
1694 range
.len
= uffdio_wake
.len
;
1697 * len == 0 means wake all and we don't want to wake all here,
1698 * so check it again to be sure.
1700 VM_BUG_ON(!range
.len
);
1702 wake_userfault(ctx
, &range
);
1709 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1713 struct uffdio_copy uffdio_copy
;
1714 struct uffdio_copy __user
*user_uffdio_copy
;
1715 struct userfaultfd_wake_range range
;
1717 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1720 if (READ_ONCE(ctx
->mmap_changing
))
1724 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1725 /* don't copy "copy" last field */
1726 sizeof(uffdio_copy
)-sizeof(__s64
)))
1729 ret
= validate_range(ctx
->mm
, &uffdio_copy
.dst
, uffdio_copy
.len
);
1733 * double check for wraparound just in case. copy_from_user()
1734 * will later check uffdio_copy.src + uffdio_copy.len to fit
1735 * in the userland range.
1738 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1740 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1742 if (mmget_not_zero(ctx
->mm
)) {
1743 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1744 uffdio_copy
.len
, &ctx
->mmap_changing
,
1750 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1755 /* len == 0 would wake all */
1757 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1758 range
.start
= uffdio_copy
.dst
;
1759 wake_userfault(ctx
, &range
);
1761 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1766 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1770 struct uffdio_zeropage uffdio_zeropage
;
1771 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1772 struct userfaultfd_wake_range range
;
1774 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1777 if (READ_ONCE(ctx
->mmap_changing
))
1781 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1782 /* don't copy "zeropage" last field */
1783 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1786 ret
= validate_range(ctx
->mm
, &uffdio_zeropage
.range
.start
,
1787 uffdio_zeropage
.range
.len
);
1791 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1794 if (mmget_not_zero(ctx
->mm
)) {
1795 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1796 uffdio_zeropage
.range
.len
,
1797 &ctx
->mmap_changing
);
1802 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1806 /* len == 0 would wake all */
1809 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1810 range
.start
= uffdio_zeropage
.range
.start
;
1811 wake_userfault(ctx
, &range
);
1813 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1818 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1822 struct uffdio_writeprotect uffdio_wp
;
1823 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1824 struct userfaultfd_wake_range range
;
1825 bool mode_wp
, mode_dontwake
;
1827 if (READ_ONCE(ctx
->mmap_changing
))
1830 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1832 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1833 sizeof(struct uffdio_writeprotect
)))
1836 ret
= validate_range(ctx
->mm
, &uffdio_wp
.range
.start
,
1837 uffdio_wp
.range
.len
);
1841 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1842 UFFDIO_WRITEPROTECT_MODE_WP
))
1845 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1846 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1848 if (mode_wp
&& mode_dontwake
)
1851 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1852 uffdio_wp
.range
.len
, mode_wp
,
1853 &ctx
->mmap_changing
);
1857 if (!mode_wp
&& !mode_dontwake
) {
1858 range
.start
= uffdio_wp
.range
.start
;
1859 range
.len
= uffdio_wp
.range
.len
;
1860 wake_userfault(ctx
, &range
);
1865 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1868 * For the current set of features the bits just coincide
1870 return (unsigned int)user_features
;
1874 * userland asks for a certain API version and we return which bits
1875 * and ioctl commands are implemented in this kernel for such API
1876 * version or -EINVAL if unknown.
1878 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1881 struct uffdio_api uffdio_api
;
1882 void __user
*buf
= (void __user
*)arg
;
1887 if (ctx
->state
!= UFFD_STATE_WAIT_API
)
1890 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1892 features
= uffdio_api
.features
;
1894 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
1897 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1899 /* report all available features and ioctls to userland */
1900 uffdio_api
.features
= UFFD_API_FEATURES
;
1901 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1903 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1905 ctx
->state
= UFFD_STATE_RUNNING
;
1906 /* only enable the requested features for this uffd context */
1907 ctx
->features
= uffd_ctx_features(features
);
1912 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1913 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1918 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1922 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1924 if (cmd
!= UFFDIO_API
&& ctx
->state
== UFFD_STATE_WAIT_API
)
1929 ret
= userfaultfd_api(ctx
, arg
);
1931 case UFFDIO_REGISTER
:
1932 ret
= userfaultfd_register(ctx
, arg
);
1934 case UFFDIO_UNREGISTER
:
1935 ret
= userfaultfd_unregister(ctx
, arg
);
1938 ret
= userfaultfd_wake(ctx
, arg
);
1941 ret
= userfaultfd_copy(ctx
, arg
);
1943 case UFFDIO_ZEROPAGE
:
1944 ret
= userfaultfd_zeropage(ctx
, arg
);
1946 case UFFDIO_WRITEPROTECT
:
1947 ret
= userfaultfd_writeprotect(ctx
, arg
);
1953 #ifdef CONFIG_PROC_FS
1954 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
1956 struct userfaultfd_ctx
*ctx
= f
->private_data
;
1957 wait_queue_entry_t
*wq
;
1958 unsigned long pending
= 0, total
= 0;
1960 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1961 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
1965 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
1968 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1971 * If more protocols will be added, there will be all shown
1972 * separated by a space. Like this:
1973 * protocols: aa:... bb:...
1975 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1976 pending
, total
, UFFD_API
, ctx
->features
,
1977 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
1981 static const struct file_operations userfaultfd_fops
= {
1982 #ifdef CONFIG_PROC_FS
1983 .show_fdinfo
= userfaultfd_show_fdinfo
,
1985 .release
= userfaultfd_release
,
1986 .poll
= userfaultfd_poll
,
1987 .read
= userfaultfd_read
,
1988 .unlocked_ioctl
= userfaultfd_ioctl
,
1989 .compat_ioctl
= compat_ptr_ioctl
,
1990 .llseek
= noop_llseek
,
1993 static void init_once_userfaultfd_ctx(void *mem
)
1995 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
1997 init_waitqueue_head(&ctx
->fault_pending_wqh
);
1998 init_waitqueue_head(&ctx
->fault_wqh
);
1999 init_waitqueue_head(&ctx
->event_wqh
);
2000 init_waitqueue_head(&ctx
->fd_wqh
);
2001 seqcount_init(&ctx
->refile_seq
);
2004 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2006 struct userfaultfd_ctx
*ctx
;
2009 if (!sysctl_unprivileged_userfaultfd
&& !capable(CAP_SYS_PTRACE
))
2012 BUG_ON(!current
->mm
);
2014 /* Check the UFFD_* constants for consistency. */
2015 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2016 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2018 if (flags
& ~UFFD_SHARED_FCNTL_FLAGS
)
2021 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2025 refcount_set(&ctx
->refcount
, 1);
2028 ctx
->state
= UFFD_STATE_WAIT_API
;
2029 ctx
->released
= false;
2030 ctx
->mmap_changing
= false;
2031 ctx
->mm
= current
->mm
;
2032 /* prevent the mm struct to be freed */
2035 fd
= anon_inode_getfd("[userfaultfd]", &userfaultfd_fops
, ctx
,
2036 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
));
2039 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2044 static int __init
userfaultfd_init(void)
2046 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2047 sizeof(struct userfaultfd_ctx
),
2049 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2050 init_once_userfaultfd_ctx
);
2053 __initcall(userfaultfd_init
);