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
;
36 * Start with fault_pending_wqh and fault_wqh so they're more likely
37 * to be in the same cacheline.
41 * fault_pending_wqh.lock
45 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
46 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
47 * also taken in IRQ context.
49 struct userfaultfd_ctx
{
50 /* waitqueue head for the pending (i.e. not read) userfaults */
51 wait_queue_head_t fault_pending_wqh
;
52 /* waitqueue head for the userfaults */
53 wait_queue_head_t fault_wqh
;
54 /* waitqueue head for the pseudo fd to wakeup poll/read */
55 wait_queue_head_t fd_wqh
;
56 /* waitqueue head for events */
57 wait_queue_head_t event_wqh
;
58 /* a refile sequence protected by fault_pending_wqh lock */
59 struct seqcount refile_seq
;
60 /* pseudo fd refcounting */
62 /* userfaultfd syscall flags */
64 /* features requested from the userspace */
65 unsigned int features
;
68 /* memory mappings are changing because of non-cooperative event */
70 /* mm with one ore more vmas attached to this userfaultfd_ctx */
74 struct userfaultfd_fork_ctx
{
75 struct userfaultfd_ctx
*orig
;
76 struct userfaultfd_ctx
*new;
77 struct list_head list
;
80 struct userfaultfd_unmap_ctx
{
81 struct userfaultfd_ctx
*ctx
;
84 struct list_head list
;
87 struct userfaultfd_wait_queue
{
89 wait_queue_entry_t wq
;
90 struct userfaultfd_ctx
*ctx
;
94 struct userfaultfd_wake_range
{
99 /* internal indication that UFFD_API ioctl was successfully executed */
100 #define UFFD_FEATURE_INITIALIZED (1u << 31)
102 static bool userfaultfd_is_initialized(struct userfaultfd_ctx
*ctx
)
104 return ctx
->features
& UFFD_FEATURE_INITIALIZED
;
107 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
108 int wake_flags
, void *key
)
110 struct userfaultfd_wake_range
*range
= key
;
112 struct userfaultfd_wait_queue
*uwq
;
113 unsigned long start
, len
;
115 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
117 /* len == 0 means wake all */
118 start
= range
->start
;
120 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
121 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
123 WRITE_ONCE(uwq
->waken
, true);
125 * The Program-Order guarantees provided by the scheduler
126 * ensure uwq->waken is visible before the task is woken.
128 ret
= wake_up_state(wq
->private, mode
);
131 * Wake only once, autoremove behavior.
133 * After the effect of list_del_init is visible to the other
134 * CPUs, the waitqueue may disappear from under us, see the
135 * !list_empty_careful() in handle_userfault().
137 * try_to_wake_up() has an implicit smp_mb(), and the
138 * wq->private is read before calling the extern function
139 * "wake_up_state" (which in turns calls try_to_wake_up).
141 list_del_init(&wq
->entry
);
148 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
150 * @ctx: [in] Pointer to the userfaultfd context.
152 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
154 refcount_inc(&ctx
->refcount
);
158 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
160 * @ctx: [in] Pointer to userfaultfd context.
162 * The userfaultfd context reference must have been previously acquired either
163 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
165 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
167 if (refcount_dec_and_test(&ctx
->refcount
)) {
168 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
169 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
170 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
171 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
172 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
173 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
174 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
175 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
177 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
181 static inline void msg_init(struct uffd_msg
*msg
)
183 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
185 * Must use memset to zero out the paddings or kernel data is
186 * leaked to userland.
188 memset(msg
, 0, sizeof(struct uffd_msg
));
191 static inline struct uffd_msg
userfault_msg(unsigned long address
,
193 unsigned long reason
,
194 unsigned int features
)
198 msg
.event
= UFFD_EVENT_PAGEFAULT
;
199 msg
.arg
.pagefault
.address
= address
;
200 if (flags
& FAULT_FLAG_WRITE
)
202 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
203 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
204 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
205 * was a read fault, otherwise if set it means it's
208 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
209 if (reason
& VM_UFFD_WP
)
211 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
212 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
213 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
214 * a missing fault, otherwise if set it means it's a
215 * write protect fault.
217 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
218 if (features
& UFFD_FEATURE_THREAD_ID
)
219 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
223 #ifdef CONFIG_HUGETLB_PAGE
225 * Same functionality as userfaultfd_must_wait below with modifications for
228 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
229 struct vm_area_struct
*vma
,
230 unsigned long address
,
232 unsigned long reason
)
234 struct mm_struct
*mm
= ctx
->mm
;
238 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
240 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
246 pte
= huge_ptep_get(ptep
);
249 * Lockless access: we're in a wait_event so it's ok if it
252 if (huge_pte_none(pte
))
254 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
260 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
261 struct vm_area_struct
*vma
,
262 unsigned long address
,
264 unsigned long reason
)
266 return false; /* should never get here */
268 #endif /* CONFIG_HUGETLB_PAGE */
271 * Verify the pagetables are still not ok after having reigstered into
272 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
273 * userfault that has already been resolved, if userfaultfd_read and
274 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
277 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
278 unsigned long address
,
280 unsigned long reason
)
282 struct mm_struct
*mm
= ctx
->mm
;
290 VM_BUG_ON(!rwsem_is_locked(&mm
->mmap_sem
));
292 pgd
= pgd_offset(mm
, address
);
293 if (!pgd_present(*pgd
))
295 p4d
= p4d_offset(pgd
, address
);
296 if (!p4d_present(*p4d
))
298 pud
= pud_offset(p4d
, address
);
299 if (!pud_present(*pud
))
301 pmd
= pmd_offset(pud
, address
);
303 * READ_ONCE must function as a barrier with narrower scope
304 * and it must be equivalent to:
305 * _pmd = *pmd; barrier();
307 * This is to deal with the instability (as in
308 * pmd_trans_unstable) of the pmd.
310 _pmd
= READ_ONCE(*pmd
);
315 if (!pmd_present(_pmd
))
318 if (pmd_trans_huge(_pmd
))
322 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
323 * and use the standard pte_offset_map() instead of parsing _pmd.
325 pte
= pte_offset_map(pmd
, address
);
327 * Lockless access: we're in a wait_event so it's ok if it
339 * The locking rules involved in returning VM_FAULT_RETRY depending on
340 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
341 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
342 * recommendation in __lock_page_or_retry is not an understatement.
344 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
345 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
348 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
349 * set, VM_FAULT_RETRY can still be returned if and only if there are
350 * fatal_signal_pending()s, and the mmap_sem must be released before
353 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
355 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
356 struct userfaultfd_ctx
*ctx
;
357 struct userfaultfd_wait_queue uwq
;
358 vm_fault_t ret
= VM_FAULT_SIGBUS
;
359 bool must_wait
, return_to_userland
;
363 * We don't do userfault handling for the final child pid update.
365 * We also don't do userfault handling during
366 * coredumping. hugetlbfs has the special
367 * follow_hugetlb_page() to skip missing pages in the
368 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
369 * the no_page_table() helper in follow_page_mask(), but the
370 * shmem_vm_ops->fault method is invoked even during
371 * coredumping without mmap_sem and it ends up here.
373 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
377 * Coredumping runs without mmap_sem so we can only check that
378 * the mmap_sem is held, if PF_DUMPCORE was not set.
380 WARN_ON_ONCE(!rwsem_is_locked(&mm
->mmap_sem
));
382 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
386 BUG_ON(ctx
->mm
!= mm
);
388 VM_BUG_ON(reason
& ~(VM_UFFD_MISSING
|VM_UFFD_WP
));
389 VM_BUG_ON(!(reason
& VM_UFFD_MISSING
) ^ !!(reason
& VM_UFFD_WP
));
391 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
395 * If it's already released don't get it. This avoids to loop
396 * in __get_user_pages if userfaultfd_release waits on the
397 * caller of handle_userfault to release the mmap_sem.
399 if (unlikely(READ_ONCE(ctx
->released
))) {
401 * Don't return VM_FAULT_SIGBUS in this case, so a non
402 * cooperative manager can close the uffd after the
403 * last UFFDIO_COPY, without risking to trigger an
404 * involuntary SIGBUS if the process was starting the
405 * userfaultfd while the userfaultfd was still armed
406 * (but after the last UFFDIO_COPY). If the uffd
407 * wasn't already closed when the userfault reached
408 * this point, that would normally be solved by
409 * userfaultfd_must_wait returning 'false'.
411 * If we were to return VM_FAULT_SIGBUS here, the non
412 * cooperative manager would be instead forced to
413 * always call UFFDIO_UNREGISTER before it can safely
416 ret
= VM_FAULT_NOPAGE
;
421 * Check that we can return VM_FAULT_RETRY.
423 * NOTE: it should become possible to return VM_FAULT_RETRY
424 * even if FAULT_FLAG_TRIED is set without leading to gup()
425 * -EBUSY failures, if the userfaultfd is to be extended for
426 * VM_UFFD_WP tracking and we intend to arm the userfault
427 * without first stopping userland access to the memory. For
428 * VM_UFFD_MISSING userfaults this is enough for now.
430 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
432 * Validate the invariant that nowait must allow retry
433 * to be sure not to return SIGBUS erroneously on
434 * nowait invocations.
436 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
437 #ifdef CONFIG_DEBUG_VM
438 if (printk_ratelimit()) {
440 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
449 * Handle nowait, not much to do other than tell it to retry
452 ret
= VM_FAULT_RETRY
;
453 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
456 /* take the reference before dropping the mmap_sem */
457 userfaultfd_ctx_get(ctx
);
459 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
460 uwq
.wq
.private = current
;
461 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
,
467 (vmf
->flags
& (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
)) ==
468 (FAULT_FLAG_USER
|FAULT_FLAG_KILLABLE
);
469 blocking_state
= return_to_userland
? TASK_INTERRUPTIBLE
:
472 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
474 * After the __add_wait_queue the uwq is visible to userland
475 * through poll/read().
477 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
479 * The smp_mb() after __set_current_state prevents the reads
480 * following the spin_unlock to happen before the list_add in
483 set_current_state(blocking_state
);
484 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
486 if (!is_vm_hugetlb_page(vmf
->vma
))
487 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
490 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
493 up_read(&mm
->mmap_sem
);
495 if (likely(must_wait
&& !READ_ONCE(ctx
->released
) &&
496 (return_to_userland
? !signal_pending(current
) :
497 !fatal_signal_pending(current
)))) {
498 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
500 ret
|= VM_FAULT_MAJOR
;
503 * False wakeups can orginate even from rwsem before
504 * up_read() however userfaults will wait either for a
505 * targeted wakeup on the specific uwq waitqueue from
506 * wake_userfault() or for signals or for uffd
509 while (!READ_ONCE(uwq
.waken
)) {
511 * This needs the full smp_store_mb()
512 * guarantee as the state write must be
513 * visible to other CPUs before reading
514 * uwq.waken from other CPUs.
516 set_current_state(blocking_state
);
517 if (READ_ONCE(uwq
.waken
) ||
518 READ_ONCE(ctx
->released
) ||
519 (return_to_userland
? signal_pending(current
) :
520 fatal_signal_pending(current
)))
526 __set_current_state(TASK_RUNNING
);
528 if (return_to_userland
) {
529 if (signal_pending(current
) &&
530 !fatal_signal_pending(current
)) {
532 * If we got a SIGSTOP or SIGCONT and this is
533 * a normal userland page fault, just let
534 * userland return so the signal will be
535 * handled and gdb debugging works. The page
536 * fault code immediately after we return from
537 * this function is going to release the
538 * mmap_sem and it's not depending on it
539 * (unlike gup would if we were not to return
542 * If a fatal signal is pending we still take
543 * the streamlined VM_FAULT_RETRY failure path
544 * and there's no need to retake the mmap_sem
547 down_read(&mm
->mmap_sem
);
548 ret
= VM_FAULT_NOPAGE
;
553 * Here we race with the list_del; list_add in
554 * userfaultfd_ctx_read(), however because we don't ever run
555 * list_del_init() to refile across the two lists, the prev
556 * and next pointers will never point to self. list_add also
557 * would never let any of the two pointers to point to
558 * self. So list_empty_careful won't risk to see both pointers
559 * pointing to self at any time during the list refile. The
560 * only case where list_del_init() is called is the full
561 * removal in the wake function and there we don't re-list_add
562 * and it's fine not to block on the spinlock. The uwq on this
563 * kernel stack can be released after the list_del_init.
565 if (!list_empty_careful(&uwq
.wq
.entry
)) {
566 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
568 * No need of list_del_init(), the uwq on the stack
569 * will be freed shortly anyway.
571 list_del(&uwq
.wq
.entry
);
572 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
576 * ctx may go away after this if the userfault pseudo fd is
579 userfaultfd_ctx_put(ctx
);
585 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
586 struct userfaultfd_wait_queue
*ewq
)
588 struct userfaultfd_ctx
*release_new_ctx
;
590 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
594 init_waitqueue_entry(&ewq
->wq
, current
);
595 release_new_ctx
= NULL
;
597 spin_lock_irq(&ctx
->event_wqh
.lock
);
599 * After the __add_wait_queue the uwq is visible to userland
600 * through poll/read().
602 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
604 set_current_state(TASK_KILLABLE
);
605 if (ewq
->msg
.event
== 0)
607 if (READ_ONCE(ctx
->released
) ||
608 fatal_signal_pending(current
)) {
610 * &ewq->wq may be queued in fork_event, but
611 * __remove_wait_queue ignores the head
612 * parameter. It would be a problem if it
615 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
616 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
617 struct userfaultfd_ctx
*new;
619 new = (struct userfaultfd_ctx
*)
621 ewq
->msg
.arg
.reserved
.reserved1
;
622 release_new_ctx
= new;
627 spin_unlock_irq(&ctx
->event_wqh
.lock
);
629 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
632 spin_lock_irq(&ctx
->event_wqh
.lock
);
634 __set_current_state(TASK_RUNNING
);
635 spin_unlock_irq(&ctx
->event_wqh
.lock
);
637 if (release_new_ctx
) {
638 struct vm_area_struct
*vma
;
639 struct mm_struct
*mm
= release_new_ctx
->mm
;
641 /* the various vma->vm_userfaultfd_ctx still points to it */
642 down_write(&mm
->mmap_sem
);
643 /* no task can run (and in turn coredump) yet */
644 VM_WARN_ON(!mmget_still_valid(mm
));
645 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
646 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
647 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
648 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
650 up_write(&mm
->mmap_sem
);
652 userfaultfd_ctx_put(release_new_ctx
);
656 * ctx may go away after this if the userfault pseudo fd is
660 WRITE_ONCE(ctx
->mmap_changing
, false);
661 userfaultfd_ctx_put(ctx
);
664 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
665 struct userfaultfd_wait_queue
*ewq
)
668 wake_up_locked(&ctx
->event_wqh
);
669 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
672 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
674 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
675 struct userfaultfd_fork_ctx
*fctx
;
677 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
678 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
679 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
680 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
684 list_for_each_entry(fctx
, fcs
, list
)
685 if (fctx
->orig
== octx
) {
691 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
695 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
701 refcount_set(&ctx
->refcount
, 1);
702 ctx
->flags
= octx
->flags
;
703 ctx
->features
= octx
->features
;
704 ctx
->released
= false;
705 ctx
->mmap_changing
= false;
706 ctx
->mm
= vma
->vm_mm
;
709 userfaultfd_ctx_get(octx
);
710 WRITE_ONCE(octx
->mmap_changing
, true);
713 list_add_tail(&fctx
->list
, fcs
);
716 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
720 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
722 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
723 struct userfaultfd_wait_queue ewq
;
727 ewq
.msg
.event
= UFFD_EVENT_FORK
;
728 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
730 userfaultfd_event_wait_completion(ctx
, &ewq
);
733 void dup_userfaultfd_complete(struct list_head
*fcs
)
735 struct userfaultfd_fork_ctx
*fctx
, *n
;
737 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
739 list_del(&fctx
->list
);
744 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
745 struct vm_userfaultfd_ctx
*vm_ctx
)
747 struct userfaultfd_ctx
*ctx
;
749 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
754 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
756 userfaultfd_ctx_get(ctx
);
757 WRITE_ONCE(ctx
->mmap_changing
, true);
759 /* Drop uffd context if remap feature not enabled */
760 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
761 vma
->vm_flags
&= ~(VM_UFFD_WP
| VM_UFFD_MISSING
);
765 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
766 unsigned long from
, unsigned long to
,
769 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
770 struct userfaultfd_wait_queue ewq
;
775 if (to
& ~PAGE_MASK
) {
776 userfaultfd_ctx_put(ctx
);
782 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
783 ewq
.msg
.arg
.remap
.from
= from
;
784 ewq
.msg
.arg
.remap
.to
= to
;
785 ewq
.msg
.arg
.remap
.len
= len
;
787 userfaultfd_event_wait_completion(ctx
, &ewq
);
790 bool userfaultfd_remove(struct vm_area_struct
*vma
,
791 unsigned long start
, unsigned long end
)
793 struct mm_struct
*mm
= vma
->vm_mm
;
794 struct userfaultfd_ctx
*ctx
;
795 struct userfaultfd_wait_queue ewq
;
797 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
798 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
801 userfaultfd_ctx_get(ctx
);
802 WRITE_ONCE(ctx
->mmap_changing
, true);
803 up_read(&mm
->mmap_sem
);
807 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
808 ewq
.msg
.arg
.remove
.start
= start
;
809 ewq
.msg
.arg
.remove
.end
= end
;
811 userfaultfd_event_wait_completion(ctx
, &ewq
);
816 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
817 unsigned long start
, unsigned long end
)
819 struct userfaultfd_unmap_ctx
*unmap_ctx
;
821 list_for_each_entry(unmap_ctx
, unmaps
, list
)
822 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
823 unmap_ctx
->end
== end
)
829 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
830 unsigned long start
, unsigned long end
,
831 struct list_head
*unmaps
)
833 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
834 struct userfaultfd_unmap_ctx
*unmap_ctx
;
835 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
837 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
838 has_unmap_ctx(ctx
, unmaps
, start
, end
))
841 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
845 userfaultfd_ctx_get(ctx
);
846 WRITE_ONCE(ctx
->mmap_changing
, true);
847 unmap_ctx
->ctx
= ctx
;
848 unmap_ctx
->start
= start
;
849 unmap_ctx
->end
= end
;
850 list_add_tail(&unmap_ctx
->list
, unmaps
);
856 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
858 struct userfaultfd_unmap_ctx
*ctx
, *n
;
859 struct userfaultfd_wait_queue ewq
;
861 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
864 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
865 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
866 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
868 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
870 list_del(&ctx
->list
);
875 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
877 struct userfaultfd_ctx
*ctx
= file
->private_data
;
878 struct mm_struct
*mm
= ctx
->mm
;
879 struct vm_area_struct
*vma
, *prev
;
880 /* len == 0 means wake all */
881 struct userfaultfd_wake_range range
= { .len
= 0, };
882 unsigned long new_flags
;
885 WRITE_ONCE(ctx
->released
, true);
887 if (!mmget_not_zero(mm
))
891 * Flush page faults out of all CPUs. NOTE: all page faults
892 * must be retried without returning VM_FAULT_SIGBUS if
893 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
894 * changes while handle_userfault released the mmap_sem. So
895 * it's critical that released is set to true (above), before
896 * taking the mmap_sem for writing.
898 down_write(&mm
->mmap_sem
);
899 still_valid
= mmget_still_valid(mm
);
901 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
903 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
904 !!(vma
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
905 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
909 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
911 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
912 new_flags
, vma
->anon_vma
,
913 vma
->vm_file
, vma
->vm_pgoff
,
921 vma
->vm_flags
= new_flags
;
922 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
924 up_write(&mm
->mmap_sem
);
928 * After no new page faults can wait on this fault_*wqh, flush
929 * the last page faults that may have been already waiting on
932 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
933 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
934 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
935 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
937 /* Flush pending events that may still wait on event_wqh */
938 wake_up_all(&ctx
->event_wqh
);
940 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
941 userfaultfd_ctx_put(ctx
);
945 /* fault_pending_wqh.lock must be hold by the caller */
946 static inline struct userfaultfd_wait_queue
*find_userfault_in(
947 wait_queue_head_t
*wqh
)
949 wait_queue_entry_t
*wq
;
950 struct userfaultfd_wait_queue
*uwq
;
952 lockdep_assert_held(&wqh
->lock
);
955 if (!waitqueue_active(wqh
))
957 /* walk in reverse to provide FIFO behavior to read userfaults */
958 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
959 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
964 static inline struct userfaultfd_wait_queue
*find_userfault(
965 struct userfaultfd_ctx
*ctx
)
967 return find_userfault_in(&ctx
->fault_pending_wqh
);
970 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
971 struct userfaultfd_ctx
*ctx
)
973 return find_userfault_in(&ctx
->event_wqh
);
976 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
978 struct userfaultfd_ctx
*ctx
= file
->private_data
;
981 poll_wait(file
, &ctx
->fd_wqh
, wait
);
983 if (!userfaultfd_is_initialized(ctx
))
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
))
1012 static const struct file_operations userfaultfd_fops
;
1014 static int resolve_userfault_fork(struct userfaultfd_ctx
*ctx
,
1015 struct userfaultfd_ctx
*new,
1016 struct uffd_msg
*msg
)
1020 fd
= anon_inode_getfd("[userfaultfd]", &userfaultfd_fops
, new,
1021 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
));
1025 msg
->arg
.reserved
.reserved1
= 0;
1026 msg
->arg
.fork
.ufd
= fd
;
1030 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1031 struct uffd_msg
*msg
)
1034 DECLARE_WAITQUEUE(wait
, current
);
1035 struct userfaultfd_wait_queue
*uwq
;
1037 * Handling fork event requires sleeping operations, so
1038 * we drop the event_wqh lock, then do these ops, then
1039 * lock it back and wake up the waiter. While the lock is
1040 * dropped the ewq may go away so we keep track of it
1043 LIST_HEAD(fork_event
);
1044 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1046 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1047 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1048 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1050 set_current_state(TASK_INTERRUPTIBLE
);
1051 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1052 uwq
= find_userfault(ctx
);
1055 * Use a seqcount to repeat the lockless check
1056 * in wake_userfault() to avoid missing
1057 * wakeups because during the refile both
1058 * waitqueue could become empty if this is the
1061 write_seqcount_begin(&ctx
->refile_seq
);
1064 * The fault_pending_wqh.lock prevents the uwq
1065 * to disappear from under us.
1067 * Refile this userfault from
1068 * fault_pending_wqh to fault_wqh, it's not
1069 * pending anymore after we read it.
1071 * Use list_del() by hand (as
1072 * userfaultfd_wake_function also uses
1073 * list_del_init() by hand) to be sure nobody
1074 * changes __remove_wait_queue() to use
1075 * list_del_init() in turn breaking the
1076 * !list_empty_careful() check in
1077 * handle_userfault(). The uwq->wq.head list
1078 * must never be empty at any time during the
1079 * refile, or the waitqueue could disappear
1080 * from under us. The "wait_queue_head_t"
1081 * parameter of __remove_wait_queue() is unused
1084 list_del(&uwq
->wq
.entry
);
1085 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1087 write_seqcount_end(&ctx
->refile_seq
);
1089 /* careful to always initialize msg if ret == 0 */
1091 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1095 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1097 spin_lock(&ctx
->event_wqh
.lock
);
1098 uwq
= find_userfault_evt(ctx
);
1102 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1103 fork_nctx
= (struct userfaultfd_ctx
*)
1105 uwq
->msg
.arg
.reserved
.reserved1
;
1106 list_move(&uwq
->wq
.entry
, &fork_event
);
1108 * fork_nctx can be freed as soon as
1109 * we drop the lock, unless we take a
1112 userfaultfd_ctx_get(fork_nctx
);
1113 spin_unlock(&ctx
->event_wqh
.lock
);
1118 userfaultfd_event_complete(ctx
, uwq
);
1119 spin_unlock(&ctx
->event_wqh
.lock
);
1123 spin_unlock(&ctx
->event_wqh
.lock
);
1125 if (signal_pending(current
)) {
1133 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1135 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1137 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1138 __set_current_state(TASK_RUNNING
);
1139 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1141 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1142 ret
= resolve_userfault_fork(ctx
, fork_nctx
, msg
);
1143 spin_lock_irq(&ctx
->event_wqh
.lock
);
1144 if (!list_empty(&fork_event
)) {
1146 * The fork thread didn't abort, so we can
1147 * drop the temporary refcount.
1149 userfaultfd_ctx_put(fork_nctx
);
1151 uwq
= list_first_entry(&fork_event
,
1155 * If fork_event list wasn't empty and in turn
1156 * the event wasn't already released by fork
1157 * (the event is allocated on fork kernel
1158 * stack), put the event back to its place in
1159 * the event_wq. fork_event head will be freed
1160 * as soon as we return so the event cannot
1161 * stay queued there no matter the current
1164 list_del(&uwq
->wq
.entry
);
1165 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1168 * Leave the event in the waitqueue and report
1169 * error to userland if we failed to resolve
1170 * the userfault fork.
1173 userfaultfd_event_complete(ctx
, uwq
);
1176 * Here the fork thread aborted and the
1177 * refcount from the fork thread on fork_nctx
1178 * has already been released. We still hold
1179 * the reference we took before releasing the
1180 * lock above. If resolve_userfault_fork
1181 * failed we've to drop it because the
1182 * fork_nctx has to be freed in such case. If
1183 * it succeeded we'll hold it because the new
1184 * uffd references it.
1187 userfaultfd_ctx_put(fork_nctx
);
1189 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1195 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1196 size_t count
, loff_t
*ppos
)
1198 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1199 ssize_t _ret
, ret
= 0;
1200 struct uffd_msg msg
;
1201 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1203 if (!userfaultfd_is_initialized(ctx
))
1207 if (count
< sizeof(msg
))
1208 return ret
? ret
: -EINVAL
;
1209 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
);
1211 return ret
? ret
: _ret
;
1212 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1213 return ret
? ret
: -EFAULT
;
1216 count
-= sizeof(msg
);
1218 * Allow to read more than one fault at time but only
1219 * block if waiting for the very first one.
1221 no_wait
= O_NONBLOCK
;
1225 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1226 struct userfaultfd_wake_range
*range
)
1228 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1229 /* wake all in the range and autoremove */
1230 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1231 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1233 if (waitqueue_active(&ctx
->fault_wqh
))
1234 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1235 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1238 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1239 struct userfaultfd_wake_range
*range
)
1245 * To be sure waitqueue_active() is not reordered by the CPU
1246 * before the pagetable update, use an explicit SMP memory
1247 * barrier here. PT lock release or up_read(mmap_sem) still
1248 * have release semantics that can allow the
1249 * waitqueue_active() to be reordered before the pte update.
1254 * Use waitqueue_active because it's very frequent to
1255 * change the address space atomically even if there are no
1256 * userfaults yet. So we take the spinlock only when we're
1257 * sure we've userfaults to wake.
1260 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1261 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1262 waitqueue_active(&ctx
->fault_wqh
);
1264 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1266 __wake_userfault(ctx
, range
);
1269 static __always_inline
int validate_range(struct mm_struct
*mm
,
1270 __u64 start
, __u64 len
)
1272 __u64 task_size
= mm
->task_size
;
1274 if (start
& ~PAGE_MASK
)
1276 if (len
& ~PAGE_MASK
)
1280 if (start
< mmap_min_addr
)
1282 if (start
>= task_size
)
1284 if (len
> task_size
- start
)
1289 static inline bool vma_can_userfault(struct vm_area_struct
*vma
)
1291 return vma_is_anonymous(vma
) || is_vm_hugetlb_page(vma
) ||
1295 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1298 struct mm_struct
*mm
= ctx
->mm
;
1299 struct vm_area_struct
*vma
, *prev
, *cur
;
1301 struct uffdio_register uffdio_register
;
1302 struct uffdio_register __user
*user_uffdio_register
;
1303 unsigned long vm_flags
, new_flags
;
1306 unsigned long start
, end
, vma_end
;
1308 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1311 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1312 sizeof(uffdio_register
)-sizeof(__u64
)))
1316 if (!uffdio_register
.mode
)
1318 if (uffdio_register
.mode
& ~(UFFDIO_REGISTER_MODE_MISSING
|
1319 UFFDIO_REGISTER_MODE_WP
))
1322 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1323 vm_flags
|= VM_UFFD_MISSING
;
1324 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1325 vm_flags
|= VM_UFFD_WP
;
1327 * FIXME: remove the below error constraint by
1328 * implementing the wprotect tracking mode.
1334 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1335 uffdio_register
.range
.len
);
1339 start
= uffdio_register
.range
.start
;
1340 end
= start
+ uffdio_register
.range
.len
;
1343 if (!mmget_not_zero(mm
))
1346 down_write(&mm
->mmap_sem
);
1347 if (!mmget_still_valid(mm
))
1349 vma
= find_vma_prev(mm
, start
, &prev
);
1353 /* check that there's at least one vma in the range */
1355 if (vma
->vm_start
>= end
)
1359 * If the first vma contains huge pages, make sure start address
1360 * is aligned to huge page size.
1362 if (is_vm_hugetlb_page(vma
)) {
1363 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1365 if (start
& (vma_hpagesize
- 1))
1370 * Search for not compatible vmas.
1373 basic_ioctls
= false;
1374 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1377 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1378 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1380 /* check not compatible vmas */
1382 if (!vma_can_userfault(cur
))
1386 * UFFDIO_COPY will fill file holes even without
1387 * PROT_WRITE. This check enforces that if this is a
1388 * MAP_SHARED, the process has write permission to the backing
1389 * file. If VM_MAYWRITE is set it also enforces that on a
1390 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1391 * F_WRITE_SEAL can be taken until the vma is destroyed.
1394 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1398 * If this vma contains ending address, and huge pages
1401 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1402 end
> cur
->vm_start
) {
1403 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1407 if (end
& (vma_hpagesize
- 1))
1412 * Check that this vma isn't already owned by a
1413 * different userfaultfd. We can't allow more than one
1414 * userfaultfd to own a single vma simultaneously or we
1415 * wouldn't know which one to deliver the userfaults to.
1418 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1419 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1423 * Note vmas containing huge pages
1425 if (is_vm_hugetlb_page(cur
))
1426 basic_ioctls
= true;
1432 if (vma
->vm_start
< start
)
1439 BUG_ON(!vma_can_userfault(vma
));
1440 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1441 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1442 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1445 * Nothing to do: this vma is already registered into this
1446 * userfaultfd and with the right tracking mode too.
1448 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1449 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1452 if (vma
->vm_start
> start
)
1453 start
= vma
->vm_start
;
1454 vma_end
= min(end
, vma
->vm_end
);
1456 new_flags
= (vma
->vm_flags
& ~vm_flags
) | vm_flags
;
1457 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1458 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1460 ((struct vm_userfaultfd_ctx
){ ctx
}));
1465 if (vma
->vm_start
< start
) {
1466 ret
= split_vma(mm
, vma
, start
, 1);
1470 if (vma
->vm_end
> end
) {
1471 ret
= split_vma(mm
, vma
, end
, 0);
1477 * In the vma_merge() successful mprotect-like case 8:
1478 * the next vma was merged into the current one and
1479 * the current one has not been updated yet.
1481 vma
->vm_flags
= new_flags
;
1482 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1486 start
= vma
->vm_end
;
1488 } while (vma
&& vma
->vm_start
< end
);
1490 up_write(&mm
->mmap_sem
);
1494 * Now that we scanned all vmas we can already tell
1495 * userland which ioctls methods are guaranteed to
1496 * succeed on this range.
1498 if (put_user(basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1499 UFFD_API_RANGE_IOCTLS
,
1500 &user_uffdio_register
->ioctls
))
1507 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1510 struct mm_struct
*mm
= ctx
->mm
;
1511 struct vm_area_struct
*vma
, *prev
, *cur
;
1513 struct uffdio_range uffdio_unregister
;
1514 unsigned long new_flags
;
1516 unsigned long start
, end
, vma_end
;
1517 const void __user
*buf
= (void __user
*)arg
;
1520 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1523 ret
= validate_range(mm
, uffdio_unregister
.start
,
1524 uffdio_unregister
.len
);
1528 start
= uffdio_unregister
.start
;
1529 end
= start
+ uffdio_unregister
.len
;
1532 if (!mmget_not_zero(mm
))
1535 down_write(&mm
->mmap_sem
);
1536 if (!mmget_still_valid(mm
))
1538 vma
= find_vma_prev(mm
, start
, &prev
);
1542 /* check that there's at least one vma in the range */
1544 if (vma
->vm_start
>= end
)
1548 * If the first vma contains huge pages, make sure start address
1549 * is aligned to huge page size.
1551 if (is_vm_hugetlb_page(vma
)) {
1552 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1554 if (start
& (vma_hpagesize
- 1))
1559 * Search for not compatible vmas.
1563 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1566 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1567 !!(cur
->vm_flags
& (VM_UFFD_MISSING
| VM_UFFD_WP
)));
1570 * Check not compatible vmas, not strictly required
1571 * here as not compatible vmas cannot have an
1572 * userfaultfd_ctx registered on them, but this
1573 * provides for more strict behavior to notice
1574 * unregistration errors.
1576 if (!vma_can_userfault(cur
))
1583 if (vma
->vm_start
< start
)
1590 BUG_ON(!vma_can_userfault(vma
));
1593 * Nothing to do: this vma is already registered into this
1594 * userfaultfd and with the right tracking mode too.
1596 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1599 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1601 if (vma
->vm_start
> start
)
1602 start
= vma
->vm_start
;
1603 vma_end
= min(end
, vma
->vm_end
);
1605 if (userfaultfd_missing(vma
)) {
1607 * Wake any concurrent pending userfault while
1608 * we unregister, so they will not hang
1609 * permanently and it avoids userland to call
1610 * UFFDIO_WAKE explicitly.
1612 struct userfaultfd_wake_range range
;
1613 range
.start
= start
;
1614 range
.len
= vma_end
- start
;
1615 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1618 new_flags
= vma
->vm_flags
& ~(VM_UFFD_MISSING
| VM_UFFD_WP
);
1619 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1620 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1627 if (vma
->vm_start
< start
) {
1628 ret
= split_vma(mm
, vma
, start
, 1);
1632 if (vma
->vm_end
> end
) {
1633 ret
= split_vma(mm
, vma
, end
, 0);
1639 * In the vma_merge() successful mprotect-like case 8:
1640 * the next vma was merged into the current one and
1641 * the current one has not been updated yet.
1643 vma
->vm_flags
= new_flags
;
1644 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1648 start
= vma
->vm_end
;
1650 } while (vma
&& vma
->vm_start
< end
);
1652 up_write(&mm
->mmap_sem
);
1659 * userfaultfd_wake may be used in combination with the
1660 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1662 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1666 struct uffdio_range uffdio_wake
;
1667 struct userfaultfd_wake_range range
;
1668 const void __user
*buf
= (void __user
*)arg
;
1671 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1674 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1678 range
.start
= uffdio_wake
.start
;
1679 range
.len
= uffdio_wake
.len
;
1682 * len == 0 means wake all and we don't want to wake all here,
1683 * so check it again to be sure.
1685 VM_BUG_ON(!range
.len
);
1687 wake_userfault(ctx
, &range
);
1694 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1698 struct uffdio_copy uffdio_copy
;
1699 struct uffdio_copy __user
*user_uffdio_copy
;
1700 struct userfaultfd_wake_range range
;
1702 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1705 if (READ_ONCE(ctx
->mmap_changing
))
1709 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1710 /* don't copy "copy" last field */
1711 sizeof(uffdio_copy
)-sizeof(__s64
)))
1714 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1718 * double check for wraparound just in case. copy_from_user()
1719 * will later check uffdio_copy.src + uffdio_copy.len to fit
1720 * in the userland range.
1723 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1725 if (uffdio_copy
.mode
& ~UFFDIO_COPY_MODE_DONTWAKE
)
1727 if (mmget_not_zero(ctx
->mm
)) {
1728 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1729 uffdio_copy
.len
, &ctx
->mmap_changing
);
1734 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1739 /* len == 0 would wake all */
1741 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1742 range
.start
= uffdio_copy
.dst
;
1743 wake_userfault(ctx
, &range
);
1745 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1750 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1754 struct uffdio_zeropage uffdio_zeropage
;
1755 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1756 struct userfaultfd_wake_range range
;
1758 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1761 if (READ_ONCE(ctx
->mmap_changing
))
1765 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1766 /* don't copy "zeropage" last field */
1767 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1770 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1771 uffdio_zeropage
.range
.len
);
1775 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1778 if (mmget_not_zero(ctx
->mm
)) {
1779 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1780 uffdio_zeropage
.range
.len
,
1781 &ctx
->mmap_changing
);
1786 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1790 /* len == 0 would wake all */
1793 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1794 range
.start
= uffdio_zeropage
.range
.start
;
1795 wake_userfault(ctx
, &range
);
1797 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1802 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1805 * For the current set of features the bits just coincide. Set
1806 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1808 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1812 * userland asks for a certain API version and we return which bits
1813 * and ioctl commands are implemented in this kernel for such API
1814 * version or -EINVAL if unknown.
1816 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1819 struct uffdio_api uffdio_api
;
1820 void __user
*buf
= (void __user
*)arg
;
1821 unsigned int ctx_features
;
1826 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1828 features
= uffdio_api
.features
;
1830 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
1833 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1835 /* report all available features and ioctls to userland */
1836 uffdio_api
.features
= UFFD_API_FEATURES
;
1837 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1839 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1842 /* only enable the requested features for this uffd context */
1843 ctx_features
= uffd_ctx_features(features
);
1845 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
1852 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1853 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1858 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1862 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1864 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
1869 ret
= userfaultfd_api(ctx
, arg
);
1871 case UFFDIO_REGISTER
:
1872 ret
= userfaultfd_register(ctx
, arg
);
1874 case UFFDIO_UNREGISTER
:
1875 ret
= userfaultfd_unregister(ctx
, arg
);
1878 ret
= userfaultfd_wake(ctx
, arg
);
1881 ret
= userfaultfd_copy(ctx
, arg
);
1883 case UFFDIO_ZEROPAGE
:
1884 ret
= userfaultfd_zeropage(ctx
, arg
);
1890 #ifdef CONFIG_PROC_FS
1891 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
1893 struct userfaultfd_ctx
*ctx
= f
->private_data
;
1894 wait_queue_entry_t
*wq
;
1895 unsigned long pending
= 0, total
= 0;
1897 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1898 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
1902 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
1905 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1908 * If more protocols will be added, there will be all shown
1909 * separated by a space. Like this:
1910 * protocols: aa:... bb:...
1912 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1913 pending
, total
, UFFD_API
, ctx
->features
,
1914 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
1918 static const struct file_operations userfaultfd_fops
= {
1919 #ifdef CONFIG_PROC_FS
1920 .show_fdinfo
= userfaultfd_show_fdinfo
,
1922 .release
= userfaultfd_release
,
1923 .poll
= userfaultfd_poll
,
1924 .read
= userfaultfd_read
,
1925 .unlocked_ioctl
= userfaultfd_ioctl
,
1926 .compat_ioctl
= userfaultfd_ioctl
,
1927 .llseek
= noop_llseek
,
1930 static void init_once_userfaultfd_ctx(void *mem
)
1932 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
1934 init_waitqueue_head(&ctx
->fault_pending_wqh
);
1935 init_waitqueue_head(&ctx
->fault_wqh
);
1936 init_waitqueue_head(&ctx
->event_wqh
);
1937 init_waitqueue_head(&ctx
->fd_wqh
);
1938 seqcount_init(&ctx
->refile_seq
);
1941 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
1943 struct userfaultfd_ctx
*ctx
;
1946 if (!sysctl_unprivileged_userfaultfd
&& !capable(CAP_SYS_PTRACE
))
1949 BUG_ON(!current
->mm
);
1951 /* Check the UFFD_* constants for consistency. */
1952 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
1953 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
1955 if (flags
& ~UFFD_SHARED_FCNTL_FLAGS
)
1958 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
1962 refcount_set(&ctx
->refcount
, 1);
1965 ctx
->released
= false;
1966 ctx
->mmap_changing
= false;
1967 ctx
->mm
= current
->mm
;
1968 /* prevent the mm struct to be freed */
1971 fd
= anon_inode_getfd("[userfaultfd]", &userfaultfd_fops
, ctx
,
1972 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
));
1975 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
1980 static int __init
userfaultfd_init(void)
1982 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
1983 sizeof(struct userfaultfd_ctx
),
1985 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
1986 init_once_userfaultfd_ctx
);
1989 __initcall(userfaultfd_init
);