4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
5 * Copyright (C) 2008-2009 Red Hat, Inc.
6 * Copyright (C) 2015 Red Hat, Inc.
8 * This work is licensed under the terms of the GNU GPL, version 2. See
9 * the COPYING file in the top-level directory.
11 * Some part derived from fs/eventfd.c (anon inode setup) and
12 * mm/ksm.c (mm hashing).
15 #include <linux/list.h>
16 #include <linux/hashtable.h>
17 #include <linux/sched/signal.h>
18 #include <linux/sched/mm.h>
20 #include <linux/poll.h>
21 #include <linux/slab.h>
22 #include <linux/seq_file.h>
23 #include <linux/file.h>
24 #include <linux/bug.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h>
30 #include <linux/security.h>
31 #include <linux/hugetlb.h>
33 static struct kmem_cache
*userfaultfd_ctx_cachep __read_mostly
;
35 enum userfaultfd_state
{
41 * Start with fault_pending_wqh and fault_wqh so they're more likely
42 * to be in the same cacheline.
44 struct userfaultfd_ctx
{
45 /* waitqueue head for the pending (i.e. not read) userfaults */
46 wait_queue_head_t fault_pending_wqh
;
47 /* waitqueue head for the userfaults */
48 wait_queue_head_t fault_wqh
;
49 /* waitqueue head for the pseudo fd to wakeup poll/read */
50 wait_queue_head_t fd_wqh
;
51 /* waitqueue head for events */
52 wait_queue_head_t event_wqh
;
53 /* a refile sequence protected by fault_pending_wqh lock */
54 struct seqcount refile_seq
;
55 /* pseudo fd refcounting */
57 /* userfaultfd syscall flags */
59 /* features requested from the userspace */
60 unsigned int features
;
62 enum userfaultfd_state state
;
65 /* mm with one ore more vmas attached to this userfaultfd_ctx */
69 struct userfaultfd_fork_ctx
{
70 struct userfaultfd_ctx
*orig
;
71 struct userfaultfd_ctx
*new;
72 struct list_head list
;
75 struct userfaultfd_unmap_ctx
{
76 struct userfaultfd_ctx
*ctx
;
79 struct list_head list
;
82 struct userfaultfd_wait_queue
{
84 wait_queue_entry_t wq
;
85 struct userfaultfd_ctx
*ctx
;
89 struct userfaultfd_wake_range
{
94 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
95 int wake_flags
, void *key
)
97 struct userfaultfd_wake_range
*range
= key
;
99 struct userfaultfd_wait_queue
*uwq
;
100 unsigned long start
, len
;
102 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
104 /* len == 0 means wake all */
105 start
= range
->start
;
107 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
108 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
110 WRITE_ONCE(uwq
->waken
, true);
112 * The Program-Order guarantees provided by the scheduler
113 * ensure uwq->waken is visible before the task is woken.
115 ret
= wake_up_state(wq
->private, mode
);
118 * Wake only once, autoremove behavior.
120 * After the effect of list_del_init is visible to the other
121 * CPUs, the waitqueue may disappear from under us, see the
122 * !list_empty_careful() in handle_userfault().
124 * try_to_wake_up() has an implicit smp_mb(), and the
125 * wq->private is read before calling the extern function
126 * "wake_up_state" (which in turns calls try_to_wake_up).
128 list_del_init(&wq
->entry
);
135 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
137 * @ctx: [in] Pointer to the userfaultfd context.
139 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
141 if (!atomic_inc_not_zero(&ctx
->refcount
))
146 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
148 * @ctx: [in] Pointer to userfaultfd context.
150 * The userfaultfd context reference must have been previously acquired either
151 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
153 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
155 if (atomic_dec_and_test(&ctx
->refcount
)) {
156 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
157 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
158 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
159 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
160 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
161 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
162 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
163 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
165 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
169 static inline void msg_init(struct uffd_msg
*msg
)
171 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
173 * Must use memset to zero out the paddings or kernel data is
174 * leaked to userland.
176 memset(msg
, 0, sizeof(struct uffd_msg
));
179 static inline struct uffd_msg
userfault_msg(unsigned long address
,
181 unsigned long reason
,
182 unsigned int features
)
186 msg
.event
= UFFD_EVENT_PAGEFAULT
;
187 msg
.arg
.pagefault
.address
= address
;
188 if (flags
& FAULT_FLAG_WRITE
)
190 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
191 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
192 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
193 * was a read fault, otherwise if set it means it's
196 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
197 if (reason
& VM_UFFD_WP
)
199 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
200 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
201 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
202 * a missing fault, otherwise if set it means it's a
203 * write protect fault.
205 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
206 if (features
& UFFD_FEATURE_THREAD_ID
)
207 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
211 #ifdef CONFIG_HUGETLB_PAGE
213 * Same functionality as userfaultfd_must_wait below with modifications for
216 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
217 struct vm_area_struct
*vma
,
218 unsigned long address
,
220 unsigned long reason
)
222 struct mm_struct
*mm
= ctx
->mm
;
226 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
228 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
234 pte
= huge_ptep_get(ptep
);
237 * Lockless access: we're in a wait_event so it's ok if it
240 if (huge_pte_none(pte
))
242 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
248 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
249 struct vm_area_struct
*vma
,
250 unsigned long address
,
252 unsigned long reason
)
254 return false; /* should never get here */
256 #endif /* CONFIG_HUGETLB_PAGE */
259 * Verify the pagetables are still not ok after having reigstered into
260 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
261 * userfault that has already been resolved, if userfaultfd_read and
262 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
265 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
266 unsigned long address
,
268 unsigned long reason
)
270 struct mm_struct
*mm
= ctx
->mm
;
278 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
280 pgd
= pgd_offset(mm
, address
);
281 if (!pgd_present(*pgd
))
283 p4d
= p4d_offset(pgd
, address
);
284 if (!p4d_present(*p4d
))
286 pud
= pud_offset(p4d
, address
);
287 if (!pud_present(*pud
))
289 pmd
= pmd_offset(pud
, address
);
291 * READ_ONCE must function as a barrier with narrower scope
292 * and it must be equivalent to:
293 * _pmd = *pmd; barrier();
295 * This is to deal with the instability (as in
296 * pmd_trans_unstable) of the pmd.
298 _pmd
= READ_ONCE(*pmd
);
299 if (!pmd_present(_pmd
))
303 if (pmd_trans_huge(_pmd
))
307 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
308 * and use the standard pte_offset_map() instead of parsing _pmd.
310 pte
= pte_offset_map(pmd
, address
);
312 * Lockless access: we're in a wait_event so it's ok if it
324 * The locking rules involved in returning VM_FAULT_RETRY depending on
325 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
326 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
327 * recommendation in __lock_page_or_retry is not an understatement.
329 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
330 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
333 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
334 * set, VM_FAULT_RETRY can still be returned if and only if there are
335 * fatal_signal_pending()s, and the mmap_sem must be released before
338 int handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
340 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
341 struct userfaultfd_ctx
*ctx
;
342 struct userfaultfd_wait_queue uwq
;
344 bool must_wait
, return_to_userland
;
347 ret
= VM_FAULT_SIGBUS
;
350 * We don't do userfault handling for the final child pid update.
352 * We also don't do userfault handling during
353 * coredumping. hugetlbfs has the special
354 * follow_hugetlb_page() to skip missing pages in the
355 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
356 * the no_page_table() helper in follow_page_mask(), but the
357 * shmem_vm_ops->fault method is invoked even during
358 * coredumping without mmap_sem and it ends up here.
360 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
364 * Coredumping runs without mmap_sem so we can only check that
365 * the mmap_sem is held, if PF_DUMPCORE was not set.
367 WARN_ON_ONCE(!rwsem_is_locked(&mm
->mmap_sem
));
369 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
373 BUG_ON(ctx
->mm
!= mm
);
375 VM_BUG_ON(reason
& ~(VM_UFFD_MISSING
|VM_UFFD_WP
));
376 VM_BUG_ON(!(reason
& VM_UFFD_MISSING
) ^ !!(reason
& VM_UFFD_WP
));
378 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
382 * If it's already released don't get it. This avoids to loop
383 * in __get_user_pages if userfaultfd_release waits on the
384 * caller of handle_userfault to release the mmap_sem.
386 if (unlikely(READ_ONCE(ctx
->released
))) {
388 * Don't return VM_FAULT_SIGBUS in this case, so a non
389 * cooperative manager can close the uffd after the
390 * last UFFDIO_COPY, without risking to trigger an
391 * involuntary SIGBUS if the process was starting the
392 * userfaultfd while the userfaultfd was still armed
393 * (but after the last UFFDIO_COPY). If the uffd
394 * wasn't already closed when the userfault reached
395 * this point, that would normally be solved by
396 * userfaultfd_must_wait returning 'false'.
398 * If we were to return VM_FAULT_SIGBUS here, the non
399 * cooperative manager would be instead forced to
400 * always call UFFDIO_UNREGISTER before it can safely
403 ret
= VM_FAULT_NOPAGE
;
408 * Check that we can return VM_FAULT_RETRY.
410 * NOTE: it should become possible to return VM_FAULT_RETRY
411 * even if FAULT_FLAG_TRIED is set without leading to gup()
412 * -EBUSY failures, if the userfaultfd is to be extended for
413 * VM_UFFD_WP tracking and we intend to arm the userfault
414 * without first stopping userland access to the memory. For
415 * VM_UFFD_MISSING userfaults this is enough for now.
417 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
419 * Validate the invariant that nowait must allow retry
420 * to be sure not to return SIGBUS erroneously on
421 * nowait invocations.
423 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
424 #ifdef CONFIG_DEBUG_VM
425 if (printk_ratelimit()) {
427 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
436 * Handle nowait, not much to do other than tell it to retry
439 ret
= VM_FAULT_RETRY
;
440 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
443 /* take the reference before dropping the mmap_sem */
444 userfaultfd_ctx_get(ctx
);
446 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
447 uwq
.wq
.private = current
;
448 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
,
454 (vmf
->flags
& (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
)) ==
455 (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
);
456 blocking_state
= return_to_userland
? TASK_INTERRUPTIBLE
:
459 spin_lock(&ctx
->fault_pending_wqh
.lock
);
461 * After the __add_wait_queue the uwq is visible to userland
462 * through poll/read().
464 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
466 * The smp_mb() after __set_current_state prevents the reads
467 * following the spin_unlock to happen before the list_add in
470 set_current_state(blocking_state
);
471 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
473 if (!is_vm_hugetlb_page(vmf
->vma
))
474 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
477 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
480 up_read(&mm
->mmap_sem
);
482 if (likely(must_wait
&& !READ_ONCE(ctx
->released
) &&
483 (return_to_userland
? !signal_pending(current
) :
484 !fatal_signal_pending(current
)))) {
485 wake_up_poll(&ctx
->fd_wqh
, POLLIN
);
487 ret
|= VM_FAULT_MAJOR
;
490 * False wakeups can orginate even from rwsem before
491 * up_read() however userfaults will wait either for a
492 * targeted wakeup on the specific uwq waitqueue from
493 * wake_userfault() or for signals or for uffd
496 while (!READ_ONCE(uwq
.waken
)) {
498 * This needs the full smp_store_mb()
499 * guarantee as the state write must be
500 * visible to other CPUs before reading
501 * uwq.waken from other CPUs.
503 set_current_state(blocking_state
);
504 if (READ_ONCE(uwq
.waken
) ||
505 READ_ONCE(ctx
->released
) ||
506 (return_to_userland
? signal_pending(current
) :
507 fatal_signal_pending(current
)))
513 __set_current_state(TASK_RUNNING
);
515 if (return_to_userland
) {
516 if (signal_pending(current
) &&
517 !fatal_signal_pending(current
)) {
519 * If we got a SIGSTOP or SIGCONT and this is
520 * a normal userland page fault, just let
521 * userland return so the signal will be
522 * handled and gdb debugging works. The page
523 * fault code immediately after we return from
524 * this function is going to release the
525 * mmap_sem and it's not depending on it
526 * (unlike gup would if we were not to return
529 * If a fatal signal is pending we still take
530 * the streamlined VM_FAULT_RETRY failure path
531 * and there's no need to retake the mmap_sem
534 down_read(&mm
->mmap_sem
);
535 ret
= VM_FAULT_NOPAGE
;
540 * Here we race with the list_del; list_add in
541 * userfaultfd_ctx_read(), however because we don't ever run
542 * list_del_init() to refile across the two lists, the prev
543 * and next pointers will never point to self. list_add also
544 * would never let any of the two pointers to point to
545 * self. So list_empty_careful won't risk to see both pointers
546 * pointing to self at any time during the list refile. The
547 * only case where list_del_init() is called is the full
548 * removal in the wake function and there we don't re-list_add
549 * and it's fine not to block on the spinlock. The uwq on this
550 * kernel stack can be released after the list_del_init.
552 if (!list_empty_careful(&uwq
.wq
.entry
)) {
553 spin_lock(&ctx
->fault_pending_wqh
.lock
);
555 * No need of list_del_init(), the uwq on the stack
556 * will be freed shortly anyway.
558 list_del(&uwq
.wq
.entry
);
559 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
563 * ctx may go away after this if the userfault pseudo fd is
566 userfaultfd_ctx_put(ctx
);
572 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
573 struct userfaultfd_wait_queue
*ewq
)
575 struct userfaultfd_ctx
*release_new_ctx
;
577 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
581 init_waitqueue_entry(&ewq
->wq
, current
);
582 release_new_ctx
= NULL
;
584 spin_lock(&ctx
->event_wqh
.lock
);
586 * After the __add_wait_queue the uwq is visible to userland
587 * through poll/read().
589 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
591 set_current_state(TASK_KILLABLE
);
592 if (ewq
->msg
.event
== 0)
594 if (READ_ONCE(ctx
->released
) ||
595 fatal_signal_pending(current
)) {
597 * &ewq->wq may be queued in fork_event, but
598 * __remove_wait_queue ignores the head
599 * parameter. It would be a problem if it
602 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
603 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
604 struct userfaultfd_ctx
*new;
606 new = (struct userfaultfd_ctx
*)
608 ewq
->msg
.arg
.reserved
.reserved1
;
609 release_new_ctx
= new;
614 spin_unlock(&ctx
->event_wqh
.lock
);
616 wake_up_poll(&ctx
->fd_wqh
, POLLIN
);
619 spin_lock(&ctx
->event_wqh
.lock
);
621 __set_current_state(TASK_RUNNING
);
622 spin_unlock(&ctx
->event_wqh
.lock
);
624 if (release_new_ctx
) {
625 struct vm_area_struct
*vma
;
626 struct mm_struct
*mm
= release_new_ctx
->mm
;
628 /* the various vma->vm_userfaultfd_ctx still points to it */
629 down_write(&mm
->mmap_sem
);
630 /* no task can run (and in turn coredump) yet */
631 VM_WARN_ON(!mmget_still_valid(mm
));
632 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
633 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
634 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
635 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
637 up_write(&mm
->mmap_sem
);
639 userfaultfd_ctx_put(release_new_ctx
);
643 * ctx may go away after this if the userfault pseudo fd is
647 userfaultfd_ctx_put(ctx
);
650 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
651 struct userfaultfd_wait_queue
*ewq
)
654 wake_up_locked(&ctx
->event_wqh
);
655 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
658 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
660 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
661 struct userfaultfd_fork_ctx
*fctx
;
663 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
664 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
665 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
666 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
670 list_for_each_entry(fctx
, fcs
, list
)
671 if (fctx
->orig
== octx
) {
677 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
681 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
687 atomic_set(&ctx
->refcount
, 1);
688 ctx
->flags
= octx
->flags
;
689 ctx
->state
= UFFD_STATE_RUNNING
;
690 ctx
->features
= octx
->features
;
691 ctx
->released
= false;
692 ctx
->mm
= vma
->vm_mm
;
695 userfaultfd_ctx_get(octx
);
698 list_add_tail(&fctx
->list
, fcs
);
701 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
705 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
707 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
708 struct userfaultfd_wait_queue ewq
;
712 ewq
.msg
.event
= UFFD_EVENT_FORK
;
713 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
715 userfaultfd_event_wait_completion(ctx
, &ewq
);
718 void dup_userfaultfd_complete(struct list_head
*fcs
)
720 struct userfaultfd_fork_ctx
*fctx
, *n
;
722 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
724 list_del(&fctx
->list
);
729 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
730 struct vm_userfaultfd_ctx
*vm_ctx
)
732 struct userfaultfd_ctx
*ctx
;
734 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
739 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
741 userfaultfd_ctx_get(ctx
);
743 /* Drop uffd context if remap feature not enabled */
744 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
745 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
749 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
750 unsigned long from
, unsigned long to
,
753 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
754 struct userfaultfd_wait_queue ewq
;
759 if (to
& ~PAGE_MASK
) {
760 userfaultfd_ctx_put(ctx
);
766 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
767 ewq
.msg
.arg
.remap
.from
= from
;
768 ewq
.msg
.arg
.remap
.to
= to
;
769 ewq
.msg
.arg
.remap
.len
= len
;
771 userfaultfd_event_wait_completion(ctx
, &ewq
);
774 bool userfaultfd_remove(struct vm_area_struct
*vma
,
775 unsigned long start
, unsigned long end
)
777 struct mm_struct
*mm
= vma
->vm_mm
;
778 struct userfaultfd_ctx
*ctx
;
779 struct userfaultfd_wait_queue ewq
;
781 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
782 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
785 userfaultfd_ctx_get(ctx
);
786 up_read(&mm
->mmap_sem
);
790 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
791 ewq
.msg
.arg
.remove
.start
= start
;
792 ewq
.msg
.arg
.remove
.end
= end
;
794 userfaultfd_event_wait_completion(ctx
, &ewq
);
799 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
800 unsigned long start
, unsigned long end
)
802 struct userfaultfd_unmap_ctx
*unmap_ctx
;
804 list_for_each_entry(unmap_ctx
, unmaps
, list
)
805 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
806 unmap_ctx
->end
== end
)
812 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
813 unsigned long start
, unsigned long end
,
814 struct list_head
*unmaps
)
816 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
817 struct userfaultfd_unmap_ctx
*unmap_ctx
;
818 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
820 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
821 has_unmap_ctx(ctx
, unmaps
, start
, end
))
824 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
828 userfaultfd_ctx_get(ctx
);
829 unmap_ctx
->ctx
= ctx
;
830 unmap_ctx
->start
= start
;
831 unmap_ctx
->end
= end
;
832 list_add_tail(&unmap_ctx
->list
, unmaps
);
838 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
840 struct userfaultfd_unmap_ctx
*ctx
, *n
;
841 struct userfaultfd_wait_queue ewq
;
843 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
846 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
847 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
848 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
850 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
852 list_del(&ctx
->list
);
857 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
859 struct userfaultfd_ctx
*ctx
= file
->private_data
;
860 struct mm_struct
*mm
= ctx
->mm
;
861 struct vm_area_struct
*vma
, *prev
;
862 /* len == 0 means wake all */
863 struct userfaultfd_wake_range range
= { .len
= 0, };
864 unsigned long new_flags
;
866 WRITE_ONCE(ctx
->released
, true);
868 if (!mmget_not_zero(mm
))
872 * Flush page faults out of all CPUs. NOTE: all page faults
873 * must be retried without returning VM_FAULT_SIGBUS if
874 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
875 * changes while handle_userfault released the mmap_sem. So
876 * it's critical that released is set to true (above), before
877 * taking the mmap_sem for writing.
879 down_write(&mm
->mmap_sem
);
880 if (!mmget_still_valid(mm
))
883 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
885 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
886 !!(vma
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
887 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
891 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
892 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
893 new_flags
, vma
->anon_vma
,
894 vma
->vm_file
, vma
->vm_pgoff
,
901 vma
->vm_flags
= new_flags
;
902 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
905 up_write(&mm
->mmap_sem
);
909 * After no new page faults can wait on this fault_*wqh, flush
910 * the last page faults that may have been already waiting on
913 spin_lock(&ctx
->fault_pending_wqh
.lock
);
914 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
915 __wake_up_locked_key(&ctx
->fault_wqh
, TASK_NORMAL
, &range
);
916 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
918 /* Flush pending events that may still wait on event_wqh */
919 wake_up_all(&ctx
->event_wqh
);
921 wake_up_poll(&ctx
->fd_wqh
, POLLHUP
);
922 userfaultfd_ctx_put(ctx
);
926 /* fault_pending_wqh.lock must be hold by the caller */
927 static inline struct userfaultfd_wait_queue
*find_userfault_in(
928 wait_queue_head_t
*wqh
)
930 wait_queue_entry_t
*wq
;
931 struct userfaultfd_wait_queue
*uwq
;
933 VM_BUG_ON(!spin_is_locked(&wqh
->lock
));
936 if (!waitqueue_active(wqh
))
938 /* walk in reverse to provide FIFO behavior to read userfaults */
939 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
940 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
945 static inline struct userfaultfd_wait_queue
*find_userfault(
946 struct userfaultfd_ctx
*ctx
)
948 return find_userfault_in(&ctx
->fault_pending_wqh
);
951 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
952 struct userfaultfd_ctx
*ctx
)
954 return find_userfault_in(&ctx
->event_wqh
);
957 static unsigned int userfaultfd_poll(struct file
*file
, poll_table
*wait
)
959 struct userfaultfd_ctx
*ctx
= file
->private_data
;
962 poll_wait(file
, &ctx
->fd_wqh
, wait
);
964 switch (ctx
->state
) {
965 case UFFD_STATE_WAIT_API
:
967 case UFFD_STATE_RUNNING
:
969 * poll() never guarantees that read won't block.
970 * userfaults can be waken before they're read().
972 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
975 * lockless access to see if there are pending faults
976 * __pollwait last action is the add_wait_queue but
977 * the spin_unlock would allow the waitqueue_active to
978 * pass above the actual list_add inside
979 * add_wait_queue critical section. So use a full
980 * memory barrier to serialize the list_add write of
981 * add_wait_queue() with the waitqueue_active read
986 if (waitqueue_active(&ctx
->fault_pending_wqh
))
988 else if (waitqueue_active(&ctx
->event_wqh
))
998 static const struct file_operations userfaultfd_fops
;
1000 static int resolve_userfault_fork(struct userfaultfd_ctx
*ctx
,
1001 struct userfaultfd_ctx
*new,
1002 struct uffd_msg
*msg
)
1006 unsigned int flags
= new->flags
& UFFD_SHARED_FCNTL_FLAGS
;
1008 fd
= get_unused_fd_flags(flags
);
1012 file
= anon_inode_getfile("[userfaultfd]", &userfaultfd_fops
, new,
1016 return PTR_ERR(file
);
1019 fd_install(fd
, file
);
1020 msg
->arg
.reserved
.reserved1
= 0;
1021 msg
->arg
.fork
.ufd
= fd
;
1026 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1027 struct uffd_msg
*msg
)
1030 DECLARE_WAITQUEUE(wait
, current
);
1031 struct userfaultfd_wait_queue
*uwq
;
1033 * Handling fork event requires sleeping operations, so
1034 * we drop the event_wqh lock, then do these ops, then
1035 * lock it back and wake up the waiter. While the lock is
1036 * dropped the ewq may go away so we keep track of it
1039 LIST_HEAD(fork_event
);
1040 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1042 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1043 spin_lock(&ctx
->fd_wqh
.lock
);
1044 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1046 set_current_state(TASK_INTERRUPTIBLE
);
1047 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1048 uwq
= find_userfault(ctx
);
1051 * Use a seqcount to repeat the lockless check
1052 * in wake_userfault() to avoid missing
1053 * wakeups because during the refile both
1054 * waitqueue could become empty if this is the
1057 write_seqcount_begin(&ctx
->refile_seq
);
1060 * The fault_pending_wqh.lock prevents the uwq
1061 * to disappear from under us.
1063 * Refile this userfault from
1064 * fault_pending_wqh to fault_wqh, it's not
1065 * pending anymore after we read it.
1067 * Use list_del() by hand (as
1068 * userfaultfd_wake_function also uses
1069 * list_del_init() by hand) to be sure nobody
1070 * changes __remove_wait_queue() to use
1071 * list_del_init() in turn breaking the
1072 * !list_empty_careful() check in
1073 * handle_userfault(). The uwq->wq.head list
1074 * must never be empty at any time during the
1075 * refile, or the waitqueue could disappear
1076 * from under us. The "wait_queue_head_t"
1077 * parameter of __remove_wait_queue() is unused
1080 list_del(&uwq
->wq
.entry
);
1081 __add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1083 write_seqcount_end(&ctx
->refile_seq
);
1085 /* careful to always initialize msg if ret == 0 */
1087 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1091 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1093 spin_lock(&ctx
->event_wqh
.lock
);
1094 uwq
= find_userfault_evt(ctx
);
1098 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1099 fork_nctx
= (struct userfaultfd_ctx
*)
1101 uwq
->msg
.arg
.reserved
.reserved1
;
1102 list_move(&uwq
->wq
.entry
, &fork_event
);
1104 * fork_nctx can be freed as soon as
1105 * we drop the lock, unless we take a
1108 userfaultfd_ctx_get(fork_nctx
);
1109 spin_unlock(&ctx
->event_wqh
.lock
);
1114 userfaultfd_event_complete(ctx
, uwq
);
1115 spin_unlock(&ctx
->event_wqh
.lock
);
1119 spin_unlock(&ctx
->event_wqh
.lock
);
1121 if (signal_pending(current
)) {
1129 spin_unlock(&ctx
->fd_wqh
.lock
);
1131 spin_lock(&ctx
->fd_wqh
.lock
);
1133 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1134 __set_current_state(TASK_RUNNING
);
1135 spin_unlock(&ctx
->fd_wqh
.lock
);
1137 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1138 ret
= resolve_userfault_fork(ctx
, fork_nctx
, msg
);
1139 spin_lock(&ctx
->event_wqh
.lock
);
1140 if (!list_empty(&fork_event
)) {
1142 * The fork thread didn't abort, so we can
1143 * drop the temporary refcount.
1145 userfaultfd_ctx_put(fork_nctx
);
1147 uwq
= list_first_entry(&fork_event
,
1151 * If fork_event list wasn't empty and in turn
1152 * the event wasn't already released by fork
1153 * (the event is allocated on fork kernel
1154 * stack), put the event back to its place in
1155 * the event_wq. fork_event head will be freed
1156 * as soon as we return so the event cannot
1157 * stay queued there no matter the current
1160 list_del(&uwq
->wq
.entry
);
1161 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1164 * Leave the event in the waitqueue and report
1165 * error to userland if we failed to resolve
1166 * the userfault fork.
1169 userfaultfd_event_complete(ctx
, uwq
);
1172 * Here the fork thread aborted and the
1173 * refcount from the fork thread on fork_nctx
1174 * has already been released. We still hold
1175 * the reference we took before releasing the
1176 * lock above. If resolve_userfault_fork
1177 * failed we've to drop it because the
1178 * fork_nctx has to be freed in such case. If
1179 * it succeeded we'll hold it because the new
1180 * uffd references it.
1183 userfaultfd_ctx_put(fork_nctx
);
1185 spin_unlock(&ctx
->event_wqh
.lock
);
1191 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1192 size_t count
, loff_t
*ppos
)
1194 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1195 ssize_t _ret
, ret
= 0;
1196 struct uffd_msg msg
;
1197 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1199 if (ctx
->state
== UFFD_STATE_WAIT_API
)
1203 if (count
< sizeof(msg
))
1204 return ret
? ret
: -EINVAL
;
1205 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
);
1207 return ret
? ret
: _ret
;
1208 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1209 return ret
? ret
: -EFAULT
;
1212 count
-= sizeof(msg
);
1214 * Allow to read more than one fault at time but only
1215 * block if waiting for the very first one.
1217 no_wait
= O_NONBLOCK
;
1221 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1222 struct userfaultfd_wake_range
*range
)
1224 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1225 /* wake all in the range and autoremove */
1226 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1227 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1229 if (waitqueue_active(&ctx
->fault_wqh
))
1230 __wake_up_locked_key(&ctx
->fault_wqh
, TASK_NORMAL
, range
);
1231 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1234 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1235 struct userfaultfd_wake_range
*range
)
1241 * To be sure waitqueue_active() is not reordered by the CPU
1242 * before the pagetable update, use an explicit SMP memory
1243 * barrier here. PT lock release or up_read(mmap_sem) still
1244 * have release semantics that can allow the
1245 * waitqueue_active() to be reordered before the pte update.
1250 * Use waitqueue_active because it's very frequent to
1251 * change the address space atomically even if there are no
1252 * userfaults yet. So we take the spinlock only when we're
1253 * sure we've userfaults to wake.
1256 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1257 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1258 waitqueue_active(&ctx
->fault_wqh
);
1260 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1262 __wake_userfault(ctx
, range
);
1265 static __always_inline
int validate_range(struct mm_struct
*mm
,
1266 __u64 start
, __u64 len
)
1268 __u64 task_size
= mm
->task_size
;
1270 if (start
& ~PAGE_MASK
)
1272 if (len
& ~PAGE_MASK
)
1276 if (start
< mmap_min_addr
)
1278 if (start
>= task_size
)
1280 if (len
> task_size
- start
)
1285 static inline bool vma_can_userfault(struct vm_area_struct
*vma
)
1287 return vma_is_anonymous(vma
) || is_vm_hugetlb_page(vma
) ||
1291 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1294 struct mm_struct
*mm
= ctx
->mm
;
1295 struct vm_area_struct
*vma
, *prev
, *cur
;
1297 struct uffdio_register uffdio_register
;
1298 struct uffdio_register __user
*user_uffdio_register
;
1299 unsigned long vm_flags
, new_flags
;
1302 unsigned long start
, end
, vma_end
;
1304 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1307 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1308 sizeof(uffdio_register
)-sizeof(__u64
)))
1312 if (!uffdio_register
.mode
)
1314 if (uffdio_register
.mode
& ~(UFFDIO_REGISTER_MODE_MISSING
|
1315 UFFDIO_REGISTER_MODE_WP
))
1318 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1319 vm_flags
|= VM_UFFD_MISSING
;
1320 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1321 vm_flags
|= VM_UFFD_WP
;
1323 * FIXME: remove the below error constraint by
1324 * implementing the wprotect tracking mode.
1330 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1331 uffdio_register
.range
.len
);
1335 start
= uffdio_register
.range
.start
;
1336 end
= start
+ uffdio_register
.range
.len
;
1339 if (!mmget_not_zero(mm
))
1342 down_write(&mm
->mmap_sem
);
1343 if (!mmget_still_valid(mm
))
1345 vma
= find_vma_prev(mm
, start
, &prev
);
1349 /* check that there's at least one vma in the range */
1351 if (vma
->vm_start
>= end
)
1355 * If the first vma contains huge pages, make sure start address
1356 * is aligned to huge page size.
1358 if (is_vm_hugetlb_page(vma
)) {
1359 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1361 if (start
& (vma_hpagesize
- 1))
1366 * Search for not compatible vmas.
1369 basic_ioctls
= false;
1370 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1373 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1374 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1376 /* check not compatible vmas */
1378 if (!vma_can_userfault(cur
))
1382 * UFFDIO_COPY will fill file holes even without
1383 * PROT_WRITE. This check enforces that if this is a
1384 * MAP_SHARED, the process has write permission to the backing
1385 * file. If VM_MAYWRITE is set it also enforces that on a
1386 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1387 * F_WRITE_SEAL can be taken until the vma is destroyed.
1390 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1394 * If this vma contains ending address, and huge pages
1397 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1398 end
> cur
->vm_start
) {
1399 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1403 if (end
& (vma_hpagesize
- 1))
1408 * Check that this vma isn't already owned by a
1409 * different userfaultfd. We can't allow more than one
1410 * userfaultfd to own a single vma simultaneously or we
1411 * wouldn't know which one to deliver the userfaults to.
1414 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1415 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1419 * Note vmas containing huge pages
1421 if (is_vm_hugetlb_page(cur
))
1422 basic_ioctls
= true;
1428 if (vma
->vm_start
< start
)
1435 BUG_ON(!vma_can_userfault(vma
));
1436 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1437 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1438 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1441 * Nothing to do: this vma is already registered into this
1442 * userfaultfd and with the right tracking mode too.
1444 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1445 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1448 if (vma
->vm_start
> start
)
1449 start
= vma
->vm_start
;
1450 vma_end
= min(end
, vma
->vm_end
);
1452 new_flags
= (vma
->vm_flags
& ~vm_flags
) | vm_flags
;
1453 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1454 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1456 ((struct vm_userfaultfd_ctx
){ ctx
}));
1461 if (vma
->vm_start
< start
) {
1462 ret
= split_vma(mm
, vma
, start
, 1);
1466 if (vma
->vm_end
> end
) {
1467 ret
= split_vma(mm
, vma
, end
, 0);
1473 * In the vma_merge() successful mprotect-like case 8:
1474 * the next vma was merged into the current one and
1475 * the current one has not been updated yet.
1477 vma
->vm_flags
= new_flags
;
1478 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1482 start
= vma
->vm_end
;
1484 } while (vma
&& vma
->vm_start
< end
);
1486 up_write(&mm
->mmap_sem
);
1490 * Now that we scanned all vmas we can already tell
1491 * userland which ioctls methods are guaranteed to
1492 * succeed on this range.
1494 if (put_user(basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1495 UFFD_API_RANGE_IOCTLS
,
1496 &user_uffdio_register
->ioctls
))
1503 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1506 struct mm_struct
*mm
= ctx
->mm
;
1507 struct vm_area_struct
*vma
, *prev
, *cur
;
1509 struct uffdio_range uffdio_unregister
;
1510 unsigned long new_flags
;
1512 unsigned long start
, end
, vma_end
;
1513 const void __user
*buf
= (void __user
*)arg
;
1516 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1519 ret
= validate_range(mm
, uffdio_unregister
.start
,
1520 uffdio_unregister
.len
);
1524 start
= uffdio_unregister
.start
;
1525 end
= start
+ uffdio_unregister
.len
;
1528 if (!mmget_not_zero(mm
))
1531 down_write(&mm
->mmap_sem
);
1532 if (!mmget_still_valid(mm
))
1534 vma
= find_vma_prev(mm
, start
, &prev
);
1538 /* check that there's at least one vma in the range */
1540 if (vma
->vm_start
>= end
)
1544 * If the first vma contains huge pages, make sure start address
1545 * is aligned to huge page size.
1547 if (is_vm_hugetlb_page(vma
)) {
1548 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1550 if (start
& (vma_hpagesize
- 1))
1555 * Search for not compatible vmas.
1559 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1562 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1563 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1566 * Check not compatible vmas, not strictly required
1567 * here as not compatible vmas cannot have an
1568 * userfaultfd_ctx registered on them, but this
1569 * provides for more strict behavior to notice
1570 * unregistration errors.
1572 if (!vma_can_userfault(cur
))
1579 if (vma
->vm_start
< start
)
1586 BUG_ON(!vma_can_userfault(vma
));
1589 * Nothing to do: this vma is already registered into this
1590 * userfaultfd and with the right tracking mode too.
1592 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1595 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1597 if (vma
->vm_start
> start
)
1598 start
= vma
->vm_start
;
1599 vma_end
= min(end
, vma
->vm_end
);
1601 if (userfaultfd_missing(vma
)) {
1603 * Wake any concurrent pending userfault while
1604 * we unregister, so they will not hang
1605 * permanently and it avoids userland to call
1606 * UFFDIO_WAKE explicitly.
1608 struct userfaultfd_wake_range range
;
1609 range
.start
= start
;
1610 range
.len
= vma_end
- start
;
1611 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1614 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
1615 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1616 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1623 if (vma
->vm_start
< start
) {
1624 ret
= split_vma(mm
, vma
, start
, 1);
1628 if (vma
->vm_end
> end
) {
1629 ret
= split_vma(mm
, vma
, end
, 0);
1635 * In the vma_merge() successful mprotect-like case 8:
1636 * the next vma was merged into the current one and
1637 * the current one has not been updated yet.
1639 vma
->vm_flags
= new_flags
;
1640 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1644 start
= vma
->vm_end
;
1646 } while (vma
&& vma
->vm_start
< end
);
1648 up_write(&mm
->mmap_sem
);
1655 * userfaultfd_wake may be used in combination with the
1656 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1658 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1662 struct uffdio_range uffdio_wake
;
1663 struct userfaultfd_wake_range range
;
1664 const void __user
*buf
= (void __user
*)arg
;
1667 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1670 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1674 range
.start
= uffdio_wake
.start
;
1675 range
.len
= uffdio_wake
.len
;
1678 * len == 0 means wake all and we don't want to wake all here,
1679 * so check it again to be sure.
1681 VM_BUG_ON(!range
.len
);
1683 wake_userfault(ctx
, &range
);
1690 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1694 struct uffdio_copy uffdio_copy
;
1695 struct uffdio_copy __user
*user_uffdio_copy
;
1696 struct userfaultfd_wake_range range
;
1698 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1701 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1702 /* don't copy "copy" last field */
1703 sizeof(uffdio_copy
)-sizeof(__s64
)))
1706 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1710 * double check for wraparound just in case. copy_from_user()
1711 * will later check uffdio_copy.src + uffdio_copy.len to fit
1712 * in the userland range.
1715 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1717 if (uffdio_copy
.mode
& ~UFFDIO_COPY_MODE_DONTWAKE
)
1719 if (mmget_not_zero(ctx
->mm
)) {
1720 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1726 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1731 /* len == 0 would wake all */
1733 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1734 range
.start
= uffdio_copy
.dst
;
1735 wake_userfault(ctx
, &range
);
1737 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1742 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1746 struct uffdio_zeropage uffdio_zeropage
;
1747 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1748 struct userfaultfd_wake_range range
;
1750 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1753 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1754 /* don't copy "zeropage" last field */
1755 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1758 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1759 uffdio_zeropage
.range
.len
);
1763 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1766 if (mmget_not_zero(ctx
->mm
)) {
1767 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1768 uffdio_zeropage
.range
.len
);
1773 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1777 /* len == 0 would wake all */
1780 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1781 range
.start
= uffdio_zeropage
.range
.start
;
1782 wake_userfault(ctx
, &range
);
1784 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1789 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1792 * For the current set of features the bits just coincide
1794 return (unsigned int)user_features
;
1798 * userland asks for a certain API version and we return which bits
1799 * and ioctl commands are implemented in this kernel for such API
1800 * version or -EINVAL if unknown.
1802 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1805 struct uffdio_api uffdio_api
;
1806 void __user
*buf
= (void __user
*)arg
;
1811 if (ctx
->state
!= UFFD_STATE_WAIT_API
)
1814 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1816 features
= uffdio_api
.features
;
1817 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
)) {
1818 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1819 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1824 /* report all available features and ioctls to userland */
1825 uffdio_api
.features
= UFFD_API_FEATURES
;
1826 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1828 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1830 ctx
->state
= UFFD_STATE_RUNNING
;
1831 /* only enable the requested features for this uffd context */
1832 ctx
->features
= uffd_ctx_features(features
);
1838 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1842 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1844 if (cmd
!= UFFDIO_API
&& ctx
->state
== UFFD_STATE_WAIT_API
)
1849 ret
= userfaultfd_api(ctx
, arg
);
1851 case UFFDIO_REGISTER
:
1852 ret
= userfaultfd_register(ctx
, arg
);
1854 case UFFDIO_UNREGISTER
:
1855 ret
= userfaultfd_unregister(ctx
, arg
);
1858 ret
= userfaultfd_wake(ctx
, arg
);
1861 ret
= userfaultfd_copy(ctx
, arg
);
1863 case UFFDIO_ZEROPAGE
:
1864 ret
= userfaultfd_zeropage(ctx
, arg
);
1870 #ifdef CONFIG_PROC_FS
1871 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
1873 struct userfaultfd_ctx
*ctx
= f
->private_data
;
1874 wait_queue_entry_t
*wq
;
1875 struct userfaultfd_wait_queue
*uwq
;
1876 unsigned long pending
= 0, total
= 0;
1878 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1879 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
1880 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
1884 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
1885 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
1888 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1891 * If more protocols will be added, there will be all shown
1892 * separated by a space. Like this:
1893 * protocols: aa:... bb:...
1895 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1896 pending
, total
, UFFD_API
, ctx
->features
,
1897 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
1901 static const struct file_operations userfaultfd_fops
= {
1902 #ifdef CONFIG_PROC_FS
1903 .show_fdinfo
= userfaultfd_show_fdinfo
,
1905 .release
= userfaultfd_release
,
1906 .poll
= userfaultfd_poll
,
1907 .read
= userfaultfd_read
,
1908 .unlocked_ioctl
= userfaultfd_ioctl
,
1909 .compat_ioctl
= userfaultfd_ioctl
,
1910 .llseek
= noop_llseek
,
1913 static void init_once_userfaultfd_ctx(void *mem
)
1915 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
1917 init_waitqueue_head(&ctx
->fault_pending_wqh
);
1918 init_waitqueue_head(&ctx
->fault_wqh
);
1919 init_waitqueue_head(&ctx
->event_wqh
);
1920 init_waitqueue_head(&ctx
->fd_wqh
);
1921 seqcount_init(&ctx
->refile_seq
);
1925 * userfaultfd_file_create - Creates a userfaultfd file pointer.
1926 * @flags: Flags for the userfaultfd file.
1928 * This function creates a userfaultfd file pointer, w/out installing
1929 * it into the fd table. This is useful when the userfaultfd file is
1930 * used during the initialization of data structures that require
1931 * extra setup after the userfaultfd creation. So the userfaultfd
1932 * creation is split into the file pointer creation phase, and the
1933 * file descriptor installation phase. In this way races with
1934 * userspace closing the newly installed file descriptor can be
1935 * avoided. Returns a userfaultfd file pointer, or a proper error
1938 static struct file
*userfaultfd_file_create(int flags
)
1941 struct userfaultfd_ctx
*ctx
;
1943 BUG_ON(!current
->mm
);
1945 /* Check the UFFD_* constants for consistency. */
1946 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
1947 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
1949 file
= ERR_PTR(-EINVAL
);
1950 if (flags
& ~UFFD_SHARED_FCNTL_FLAGS
)
1953 file
= ERR_PTR(-ENOMEM
);
1954 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
1958 atomic_set(&ctx
->refcount
, 1);
1961 ctx
->state
= UFFD_STATE_WAIT_API
;
1962 ctx
->released
= false;
1963 ctx
->mm
= current
->mm
;
1964 /* prevent the mm struct to be freed */
1967 file
= anon_inode_getfile("[userfaultfd]", &userfaultfd_fops
, ctx
,
1968 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
));
1971 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
1977 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
1982 error
= get_unused_fd_flags(flags
& UFFD_SHARED_FCNTL_FLAGS
);
1987 file
= userfaultfd_file_create(flags
);
1989 error
= PTR_ERR(file
);
1990 goto err_put_unused_fd
;
1992 fd_install(fd
, file
);
2002 static int __init
userfaultfd_init(void)
2004 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2005 sizeof(struct userfaultfd_ctx
),
2007 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2008 init_once_userfaultfd_ctx
);
2011 __initcall(userfaultfd_init
);