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/mmu_notifier.h>
19 #include <linux/poll.h>
20 #include <linux/slab.h>
21 #include <linux/seq_file.h>
22 #include <linux/file.h>
23 #include <linux/bug.h>
24 #include <linux/anon_inodes.h>
25 #include <linux/syscalls.h>
26 #include <linux/userfaultfd_k.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ioctl.h>
29 #include <linux/security.h>
30 #include <linux/hugetlb.h>
32 int sysctl_unprivileged_userfaultfd __read_mostly
;
34 static struct kmem_cache
*userfaultfd_ctx_cachep __read_mostly
;
37 * Start with fault_pending_wqh and fault_wqh so they're more likely
38 * to be in the same cacheline.
42 * fault_pending_wqh.lock
46 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
47 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
48 * also taken in IRQ context.
50 struct userfaultfd_ctx
{
51 /* waitqueue head for the pending (i.e. not read) userfaults */
52 wait_queue_head_t fault_pending_wqh
;
53 /* waitqueue head for the userfaults */
54 wait_queue_head_t fault_wqh
;
55 /* waitqueue head for the pseudo fd to wakeup poll/read */
56 wait_queue_head_t fd_wqh
;
57 /* waitqueue head for events */
58 wait_queue_head_t event_wqh
;
59 /* a refile sequence protected by fault_pending_wqh lock */
60 seqcount_spinlock_t refile_seq
;
61 /* pseudo fd refcounting */
63 /* userfaultfd syscall flags */
65 /* features requested from the userspace */
66 unsigned int features
;
69 /* memory mappings are changing because of non-cooperative event */
70 atomic_t mmap_changing
;
71 /* mm with one ore more vmas attached to this userfaultfd_ctx */
75 struct userfaultfd_fork_ctx
{
76 struct userfaultfd_ctx
*orig
;
77 struct userfaultfd_ctx
*new;
78 struct list_head list
;
81 struct userfaultfd_unmap_ctx
{
82 struct userfaultfd_ctx
*ctx
;
85 struct list_head list
;
88 struct userfaultfd_wait_queue
{
90 wait_queue_entry_t wq
;
91 struct userfaultfd_ctx
*ctx
;
95 struct userfaultfd_wake_range
{
100 /* internal indication that UFFD_API ioctl was successfully executed */
101 #define UFFD_FEATURE_INITIALIZED (1u << 31)
103 static bool userfaultfd_is_initialized(struct userfaultfd_ctx
*ctx
)
105 return ctx
->features
& UFFD_FEATURE_INITIALIZED
;
108 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
109 int wake_flags
, void *key
)
111 struct userfaultfd_wake_range
*range
= key
;
113 struct userfaultfd_wait_queue
*uwq
;
114 unsigned long start
, len
;
116 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
118 /* len == 0 means wake all */
119 start
= range
->start
;
121 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
122 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
124 WRITE_ONCE(uwq
->waken
, true);
126 * The Program-Order guarantees provided by the scheduler
127 * ensure uwq->waken is visible before the task is woken.
129 ret
= wake_up_state(wq
->private, mode
);
132 * Wake only once, autoremove behavior.
134 * After the effect of list_del_init is visible to the other
135 * CPUs, the waitqueue may disappear from under us, see the
136 * !list_empty_careful() in handle_userfault().
138 * try_to_wake_up() has an implicit smp_mb(), and the
139 * wq->private is read before calling the extern function
140 * "wake_up_state" (which in turns calls try_to_wake_up).
142 list_del_init(&wq
->entry
);
149 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
151 * @ctx: [in] Pointer to the userfaultfd context.
153 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
155 refcount_inc(&ctx
->refcount
);
159 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
161 * @ctx: [in] Pointer to userfaultfd context.
163 * The userfaultfd context reference must have been previously acquired either
164 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
166 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
168 if (refcount_dec_and_test(&ctx
->refcount
)) {
169 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
170 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
171 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
172 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
173 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
174 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
175 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
176 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
178 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
182 static inline void msg_init(struct uffd_msg
*msg
)
184 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
186 * Must use memset to zero out the paddings or kernel data is
187 * leaked to userland.
189 memset(msg
, 0, sizeof(struct uffd_msg
));
192 static inline struct uffd_msg
userfault_msg(unsigned long address
,
194 unsigned long reason
,
195 unsigned int features
)
199 msg
.event
= UFFD_EVENT_PAGEFAULT
;
200 msg
.arg
.pagefault
.address
= address
;
202 * These flags indicate why the userfault occurred:
203 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
204 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
205 * - Neither of these flags being set indicates a MISSING fault.
207 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
208 * fault. Otherwise, it was a read fault.
210 if (flags
& FAULT_FLAG_WRITE
)
211 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
212 if (reason
& VM_UFFD_WP
)
213 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
214 if (reason
& VM_UFFD_MINOR
)
215 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_MINOR
;
216 if (features
& UFFD_FEATURE_THREAD_ID
)
217 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
221 #ifdef CONFIG_HUGETLB_PAGE
223 * Same functionality as userfaultfd_must_wait below with modifications for
226 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
227 struct vm_area_struct
*vma
,
228 unsigned long address
,
230 unsigned long reason
)
232 struct mm_struct
*mm
= ctx
->mm
;
236 mmap_assert_locked(mm
);
238 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
244 pte
= huge_ptep_get(ptep
);
247 * Lockless access: we're in a wait_event so it's ok if it
250 if (huge_pte_none(pte
))
252 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
258 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
259 struct vm_area_struct
*vma
,
260 unsigned long address
,
262 unsigned long reason
)
264 return false; /* should never get here */
266 #endif /* CONFIG_HUGETLB_PAGE */
269 * Verify the pagetables are still not ok after having reigstered into
270 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
271 * userfault that has already been resolved, if userfaultfd_read and
272 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
275 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
276 unsigned long address
,
278 unsigned long reason
)
280 struct mm_struct
*mm
= ctx
->mm
;
288 mmap_assert_locked(mm
);
290 pgd
= pgd_offset(mm
, address
);
291 if (!pgd_present(*pgd
))
293 p4d
= p4d_offset(pgd
, address
);
294 if (!p4d_present(*p4d
))
296 pud
= pud_offset(p4d
, address
);
297 if (!pud_present(*pud
))
299 pmd
= pmd_offset(pud
, address
);
301 * READ_ONCE must function as a barrier with narrower scope
302 * and it must be equivalent to:
303 * _pmd = *pmd; barrier();
305 * This is to deal with the instability (as in
306 * pmd_trans_unstable) of the pmd.
308 _pmd
= READ_ONCE(*pmd
);
313 if (!pmd_present(_pmd
))
316 if (pmd_trans_huge(_pmd
)) {
317 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
323 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
324 * and use the standard pte_offset_map() instead of parsing _pmd.
326 pte
= pte_offset_map(pmd
, address
);
328 * Lockless access: we're in a wait_event so it's ok if it
333 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
341 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
343 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
344 return TASK_INTERRUPTIBLE
;
346 if (flags
& FAULT_FLAG_KILLABLE
)
347 return TASK_KILLABLE
;
349 return TASK_UNINTERRUPTIBLE
;
353 * The locking rules involved in returning VM_FAULT_RETRY depending on
354 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
355 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
356 * recommendation in __lock_page_or_retry is not an understatement.
358 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
359 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
362 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
363 * set, VM_FAULT_RETRY can still be returned if and only if there are
364 * fatal_signal_pending()s, and the mmap_lock must be released before
367 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
369 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
370 struct userfaultfd_ctx
*ctx
;
371 struct userfaultfd_wait_queue uwq
;
372 vm_fault_t ret
= VM_FAULT_SIGBUS
;
374 unsigned int blocking_state
;
377 * We don't do userfault handling for the final child pid update.
379 * We also don't do userfault handling during
380 * coredumping. hugetlbfs has the special
381 * follow_hugetlb_page() to skip missing pages in the
382 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
383 * the no_page_table() helper in follow_page_mask(), but the
384 * shmem_vm_ops->fault method is invoked even during
385 * coredumping without mmap_lock and it ends up here.
387 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
391 * Coredumping runs without mmap_lock so we can only check that
392 * the mmap_lock is held, if PF_DUMPCORE was not set.
394 mmap_assert_locked(mm
);
396 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
400 BUG_ON(ctx
->mm
!= mm
);
402 /* Any unrecognized flag is a bug. */
403 VM_BUG_ON(reason
& ~__VM_UFFD_FLAGS
);
404 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
405 VM_BUG_ON(!reason
|| (reason
& (reason
- 1)));
407 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
409 if ((vmf
->flags
& FAULT_FLAG_USER
) == 0 &&
410 ctx
->flags
& UFFD_USER_MODE_ONLY
) {
411 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
412 "sysctl knob to 1 if kernel faults must be handled "
413 "without obtaining CAP_SYS_PTRACE capability\n");
418 * If it's already released don't get it. This avoids to loop
419 * in __get_user_pages if userfaultfd_release waits on the
420 * caller of handle_userfault to release the mmap_lock.
422 if (unlikely(READ_ONCE(ctx
->released
))) {
424 * Don't return VM_FAULT_SIGBUS in this case, so a non
425 * cooperative manager can close the uffd after the
426 * last UFFDIO_COPY, without risking to trigger an
427 * involuntary SIGBUS if the process was starting the
428 * userfaultfd while the userfaultfd was still armed
429 * (but after the last UFFDIO_COPY). If the uffd
430 * wasn't already closed when the userfault reached
431 * this point, that would normally be solved by
432 * userfaultfd_must_wait returning 'false'.
434 * If we were to return VM_FAULT_SIGBUS here, the non
435 * cooperative manager would be instead forced to
436 * always call UFFDIO_UNREGISTER before it can safely
439 ret
= VM_FAULT_NOPAGE
;
444 * Check that we can return VM_FAULT_RETRY.
446 * NOTE: it should become possible to return VM_FAULT_RETRY
447 * even if FAULT_FLAG_TRIED is set without leading to gup()
448 * -EBUSY failures, if the userfaultfd is to be extended for
449 * VM_UFFD_WP tracking and we intend to arm the userfault
450 * without first stopping userland access to the memory. For
451 * VM_UFFD_MISSING userfaults this is enough for now.
453 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
455 * Validate the invariant that nowait must allow retry
456 * to be sure not to return SIGBUS erroneously on
457 * nowait invocations.
459 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
460 #ifdef CONFIG_DEBUG_VM
461 if (printk_ratelimit()) {
463 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
472 * Handle nowait, not much to do other than tell it to retry
475 ret
= VM_FAULT_RETRY
;
476 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
479 /* take the reference before dropping the mmap_lock */
480 userfaultfd_ctx_get(ctx
);
482 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
483 uwq
.wq
.private = current
;
484 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
,
489 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
491 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
493 * After the __add_wait_queue the uwq is visible to userland
494 * through poll/read().
496 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
498 * The smp_mb() after __set_current_state prevents the reads
499 * following the spin_unlock to happen before the list_add in
502 set_current_state(blocking_state
);
503 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
505 if (!is_vm_hugetlb_page(vmf
->vma
))
506 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
509 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
512 mmap_read_unlock(mm
);
514 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
515 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
519 __set_current_state(TASK_RUNNING
);
522 * Here we race with the list_del; list_add in
523 * userfaultfd_ctx_read(), however because we don't ever run
524 * list_del_init() to refile across the two lists, the prev
525 * and next pointers will never point to self. list_add also
526 * would never let any of the two pointers to point to
527 * self. So list_empty_careful won't risk to see both pointers
528 * pointing to self at any time during the list refile. The
529 * only case where list_del_init() is called is the full
530 * removal in the wake function and there we don't re-list_add
531 * and it's fine not to block on the spinlock. The uwq on this
532 * kernel stack can be released after the list_del_init.
534 if (!list_empty_careful(&uwq
.wq
.entry
)) {
535 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
537 * No need of list_del_init(), the uwq on the stack
538 * will be freed shortly anyway.
540 list_del(&uwq
.wq
.entry
);
541 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
545 * ctx may go away after this if the userfault pseudo fd is
548 userfaultfd_ctx_put(ctx
);
554 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
555 struct userfaultfd_wait_queue
*ewq
)
557 struct userfaultfd_ctx
*release_new_ctx
;
559 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
563 init_waitqueue_entry(&ewq
->wq
, current
);
564 release_new_ctx
= NULL
;
566 spin_lock_irq(&ctx
->event_wqh
.lock
);
568 * After the __add_wait_queue the uwq is visible to userland
569 * through poll/read().
571 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
573 set_current_state(TASK_KILLABLE
);
574 if (ewq
->msg
.event
== 0)
576 if (READ_ONCE(ctx
->released
) ||
577 fatal_signal_pending(current
)) {
579 * &ewq->wq may be queued in fork_event, but
580 * __remove_wait_queue ignores the head
581 * parameter. It would be a problem if it
584 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
585 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
586 struct userfaultfd_ctx
*new;
588 new = (struct userfaultfd_ctx
*)
590 ewq
->msg
.arg
.reserved
.reserved1
;
591 release_new_ctx
= new;
596 spin_unlock_irq(&ctx
->event_wqh
.lock
);
598 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
601 spin_lock_irq(&ctx
->event_wqh
.lock
);
603 __set_current_state(TASK_RUNNING
);
604 spin_unlock_irq(&ctx
->event_wqh
.lock
);
606 if (release_new_ctx
) {
607 struct vm_area_struct
*vma
;
608 struct mm_struct
*mm
= release_new_ctx
->mm
;
610 /* the various vma->vm_userfaultfd_ctx still points to it */
612 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
613 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
614 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
615 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
617 mmap_write_unlock(mm
);
619 userfaultfd_ctx_put(release_new_ctx
);
623 * ctx may go away after this if the userfault pseudo fd is
627 atomic_dec(&ctx
->mmap_changing
);
628 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
629 userfaultfd_ctx_put(ctx
);
632 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
633 struct userfaultfd_wait_queue
*ewq
)
636 wake_up_locked(&ctx
->event_wqh
);
637 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
640 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
642 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
643 struct userfaultfd_fork_ctx
*fctx
;
645 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
646 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
647 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
648 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
652 list_for_each_entry(fctx
, fcs
, list
)
653 if (fctx
->orig
== octx
) {
659 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
663 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
669 refcount_set(&ctx
->refcount
, 1);
670 ctx
->flags
= octx
->flags
;
671 ctx
->features
= octx
->features
;
672 ctx
->released
= false;
673 atomic_set(&ctx
->mmap_changing
, 0);
674 ctx
->mm
= vma
->vm_mm
;
677 userfaultfd_ctx_get(octx
);
678 atomic_inc(&octx
->mmap_changing
);
681 list_add_tail(&fctx
->list
, fcs
);
684 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
688 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
690 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
691 struct userfaultfd_wait_queue ewq
;
695 ewq
.msg
.event
= UFFD_EVENT_FORK
;
696 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
698 userfaultfd_event_wait_completion(ctx
, &ewq
);
701 void dup_userfaultfd_complete(struct list_head
*fcs
)
703 struct userfaultfd_fork_ctx
*fctx
, *n
;
705 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
707 list_del(&fctx
->list
);
712 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
713 struct vm_userfaultfd_ctx
*vm_ctx
)
715 struct userfaultfd_ctx
*ctx
;
717 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
722 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
724 userfaultfd_ctx_get(ctx
);
725 atomic_inc(&ctx
->mmap_changing
);
727 /* Drop uffd context if remap feature not enabled */
728 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
729 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
733 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
734 unsigned long from
, unsigned long to
,
737 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
738 struct userfaultfd_wait_queue ewq
;
743 if (to
& ~PAGE_MASK
) {
744 userfaultfd_ctx_put(ctx
);
750 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
751 ewq
.msg
.arg
.remap
.from
= from
;
752 ewq
.msg
.arg
.remap
.to
= to
;
753 ewq
.msg
.arg
.remap
.len
= len
;
755 userfaultfd_event_wait_completion(ctx
, &ewq
);
758 bool userfaultfd_remove(struct vm_area_struct
*vma
,
759 unsigned long start
, unsigned long end
)
761 struct mm_struct
*mm
= vma
->vm_mm
;
762 struct userfaultfd_ctx
*ctx
;
763 struct userfaultfd_wait_queue ewq
;
765 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
766 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
769 userfaultfd_ctx_get(ctx
);
770 atomic_inc(&ctx
->mmap_changing
);
771 mmap_read_unlock(mm
);
775 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
776 ewq
.msg
.arg
.remove
.start
= start
;
777 ewq
.msg
.arg
.remove
.end
= end
;
779 userfaultfd_event_wait_completion(ctx
, &ewq
);
784 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
785 unsigned long start
, unsigned long end
)
787 struct userfaultfd_unmap_ctx
*unmap_ctx
;
789 list_for_each_entry(unmap_ctx
, unmaps
, list
)
790 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
791 unmap_ctx
->end
== end
)
797 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
798 unsigned long start
, unsigned long end
,
799 struct list_head
*unmaps
)
801 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
802 struct userfaultfd_unmap_ctx
*unmap_ctx
;
803 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
805 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
806 has_unmap_ctx(ctx
, unmaps
, start
, end
))
809 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
813 userfaultfd_ctx_get(ctx
);
814 atomic_inc(&ctx
->mmap_changing
);
815 unmap_ctx
->ctx
= ctx
;
816 unmap_ctx
->start
= start
;
817 unmap_ctx
->end
= end
;
818 list_add_tail(&unmap_ctx
->list
, unmaps
);
824 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
826 struct userfaultfd_unmap_ctx
*ctx
, *n
;
827 struct userfaultfd_wait_queue ewq
;
829 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
832 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
833 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
834 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
836 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
838 list_del(&ctx
->list
);
843 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
845 struct userfaultfd_ctx
*ctx
= file
->private_data
;
846 struct mm_struct
*mm
= ctx
->mm
;
847 struct vm_area_struct
*vma
, *prev
;
848 /* len == 0 means wake all */
849 struct userfaultfd_wake_range range
= { .len
= 0, };
850 unsigned long new_flags
;
852 WRITE_ONCE(ctx
->released
, true);
854 if (!mmget_not_zero(mm
))
858 * Flush page faults out of all CPUs. NOTE: all page faults
859 * must be retried without returning VM_FAULT_SIGBUS if
860 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
861 * changes while handle_userfault released the mmap_lock. So
862 * it's critical that released is set to true (above), before
863 * taking the mmap_lock for writing.
867 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
869 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
870 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
871 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
875 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
876 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
877 new_flags
, vma
->anon_vma
,
878 vma
->vm_file
, vma
->vm_pgoff
,
885 vma
->vm_flags
= new_flags
;
886 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
888 mmap_write_unlock(mm
);
892 * After no new page faults can wait on this fault_*wqh, flush
893 * the last page faults that may have been already waiting on
896 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
897 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
898 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
899 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
901 /* Flush pending events that may still wait on event_wqh */
902 wake_up_all(&ctx
->event_wqh
);
904 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
905 userfaultfd_ctx_put(ctx
);
909 /* fault_pending_wqh.lock must be hold by the caller */
910 static inline struct userfaultfd_wait_queue
*find_userfault_in(
911 wait_queue_head_t
*wqh
)
913 wait_queue_entry_t
*wq
;
914 struct userfaultfd_wait_queue
*uwq
;
916 lockdep_assert_held(&wqh
->lock
);
919 if (!waitqueue_active(wqh
))
921 /* walk in reverse to provide FIFO behavior to read userfaults */
922 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
923 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
928 static inline struct userfaultfd_wait_queue
*find_userfault(
929 struct userfaultfd_ctx
*ctx
)
931 return find_userfault_in(&ctx
->fault_pending_wqh
);
934 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
935 struct userfaultfd_ctx
*ctx
)
937 return find_userfault_in(&ctx
->event_wqh
);
940 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
942 struct userfaultfd_ctx
*ctx
= file
->private_data
;
945 poll_wait(file
, &ctx
->fd_wqh
, wait
);
947 if (!userfaultfd_is_initialized(ctx
))
951 * poll() never guarantees that read won't block.
952 * userfaults can be waken before they're read().
954 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
957 * lockless access to see if there are pending faults
958 * __pollwait last action is the add_wait_queue but
959 * the spin_unlock would allow the waitqueue_active to
960 * pass above the actual list_add inside
961 * add_wait_queue critical section. So use a full
962 * memory barrier to serialize the list_add write of
963 * add_wait_queue() with the waitqueue_active read
968 if (waitqueue_active(&ctx
->fault_pending_wqh
))
970 else if (waitqueue_active(&ctx
->event_wqh
))
976 static const struct file_operations userfaultfd_fops
;
978 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
980 struct uffd_msg
*msg
)
984 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, new,
985 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
989 msg
->arg
.reserved
.reserved1
= 0;
990 msg
->arg
.fork
.ufd
= fd
;
994 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
995 struct uffd_msg
*msg
, struct inode
*inode
)
998 DECLARE_WAITQUEUE(wait
, current
);
999 struct userfaultfd_wait_queue
*uwq
;
1001 * Handling fork event requires sleeping operations, so
1002 * we drop the event_wqh lock, then do these ops, then
1003 * lock it back and wake up the waiter. While the lock is
1004 * dropped the ewq may go away so we keep track of it
1007 LIST_HEAD(fork_event
);
1008 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1010 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1011 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1012 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1014 set_current_state(TASK_INTERRUPTIBLE
);
1015 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1016 uwq
= find_userfault(ctx
);
1019 * Use a seqcount to repeat the lockless check
1020 * in wake_userfault() to avoid missing
1021 * wakeups because during the refile both
1022 * waitqueue could become empty if this is the
1025 write_seqcount_begin(&ctx
->refile_seq
);
1028 * The fault_pending_wqh.lock prevents the uwq
1029 * to disappear from under us.
1031 * Refile this userfault from
1032 * fault_pending_wqh to fault_wqh, it's not
1033 * pending anymore after we read it.
1035 * Use list_del() by hand (as
1036 * userfaultfd_wake_function also uses
1037 * list_del_init() by hand) to be sure nobody
1038 * changes __remove_wait_queue() to use
1039 * list_del_init() in turn breaking the
1040 * !list_empty_careful() check in
1041 * handle_userfault(). The uwq->wq.head list
1042 * must never be empty at any time during the
1043 * refile, or the waitqueue could disappear
1044 * from under us. The "wait_queue_head_t"
1045 * parameter of __remove_wait_queue() is unused
1048 list_del(&uwq
->wq
.entry
);
1049 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1051 write_seqcount_end(&ctx
->refile_seq
);
1053 /* careful to always initialize msg if ret == 0 */
1055 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1059 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1061 spin_lock(&ctx
->event_wqh
.lock
);
1062 uwq
= find_userfault_evt(ctx
);
1066 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1067 fork_nctx
= (struct userfaultfd_ctx
*)
1069 uwq
->msg
.arg
.reserved
.reserved1
;
1070 list_move(&uwq
->wq
.entry
, &fork_event
);
1072 * fork_nctx can be freed as soon as
1073 * we drop the lock, unless we take a
1076 userfaultfd_ctx_get(fork_nctx
);
1077 spin_unlock(&ctx
->event_wqh
.lock
);
1082 userfaultfd_event_complete(ctx
, uwq
);
1083 spin_unlock(&ctx
->event_wqh
.lock
);
1087 spin_unlock(&ctx
->event_wqh
.lock
);
1089 if (signal_pending(current
)) {
1097 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1099 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1101 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1102 __set_current_state(TASK_RUNNING
);
1103 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1105 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1106 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1107 spin_lock_irq(&ctx
->event_wqh
.lock
);
1108 if (!list_empty(&fork_event
)) {
1110 * The fork thread didn't abort, so we can
1111 * drop the temporary refcount.
1113 userfaultfd_ctx_put(fork_nctx
);
1115 uwq
= list_first_entry(&fork_event
,
1119 * If fork_event list wasn't empty and in turn
1120 * the event wasn't already released by fork
1121 * (the event is allocated on fork kernel
1122 * stack), put the event back to its place in
1123 * the event_wq. fork_event head will be freed
1124 * as soon as we return so the event cannot
1125 * stay queued there no matter the current
1128 list_del(&uwq
->wq
.entry
);
1129 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1132 * Leave the event in the waitqueue and report
1133 * error to userland if we failed to resolve
1134 * the userfault fork.
1137 userfaultfd_event_complete(ctx
, uwq
);
1140 * Here the fork thread aborted and the
1141 * refcount from the fork thread on fork_nctx
1142 * has already been released. We still hold
1143 * the reference we took before releasing the
1144 * lock above. If resolve_userfault_fork
1145 * failed we've to drop it because the
1146 * fork_nctx has to be freed in such case. If
1147 * it succeeded we'll hold it because the new
1148 * uffd references it.
1151 userfaultfd_ctx_put(fork_nctx
);
1153 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1159 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1160 size_t count
, loff_t
*ppos
)
1162 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1163 ssize_t _ret
, ret
= 0;
1164 struct uffd_msg msg
;
1165 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1166 struct inode
*inode
= file_inode(file
);
1168 if (!userfaultfd_is_initialized(ctx
))
1172 if (count
< sizeof(msg
))
1173 return ret
? ret
: -EINVAL
;
1174 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1176 return ret
? ret
: _ret
;
1177 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1178 return ret
? ret
: -EFAULT
;
1181 count
-= sizeof(msg
);
1183 * Allow to read more than one fault at time but only
1184 * block if waiting for the very first one.
1186 no_wait
= O_NONBLOCK
;
1190 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1191 struct userfaultfd_wake_range
*range
)
1193 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1194 /* wake all in the range and autoremove */
1195 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1196 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1198 if (waitqueue_active(&ctx
->fault_wqh
))
1199 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1200 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1203 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1204 struct userfaultfd_wake_range
*range
)
1210 * To be sure waitqueue_active() is not reordered by the CPU
1211 * before the pagetable update, use an explicit SMP memory
1212 * barrier here. PT lock release or mmap_read_unlock(mm) still
1213 * have release semantics that can allow the
1214 * waitqueue_active() to be reordered before the pte update.
1219 * Use waitqueue_active because it's very frequent to
1220 * change the address space atomically even if there are no
1221 * userfaults yet. So we take the spinlock only when we're
1222 * sure we've userfaults to wake.
1225 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1226 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1227 waitqueue_active(&ctx
->fault_wqh
);
1229 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1231 __wake_userfault(ctx
, range
);
1234 static __always_inline
int validate_range(struct mm_struct
*mm
,
1235 __u64 start
, __u64 len
)
1237 __u64 task_size
= mm
->task_size
;
1239 if (start
& ~PAGE_MASK
)
1241 if (len
& ~PAGE_MASK
)
1245 if (start
< mmap_min_addr
)
1247 if (start
>= task_size
)
1249 if (len
> task_size
- start
)
1254 static inline bool vma_can_userfault(struct vm_area_struct
*vma
,
1255 unsigned long vm_flags
)
1257 /* FIXME: add WP support to hugetlbfs and shmem */
1258 if (vm_flags
& VM_UFFD_WP
) {
1259 if (is_vm_hugetlb_page(vma
) || vma_is_shmem(vma
))
1263 if (vm_flags
& VM_UFFD_MINOR
) {
1264 if (!(is_vm_hugetlb_page(vma
) || vma_is_shmem(vma
)))
1268 return vma_is_anonymous(vma
) || is_vm_hugetlb_page(vma
) ||
1272 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1275 struct mm_struct
*mm
= ctx
->mm
;
1276 struct vm_area_struct
*vma
, *prev
, *cur
;
1278 struct uffdio_register uffdio_register
;
1279 struct uffdio_register __user
*user_uffdio_register
;
1280 unsigned long vm_flags
, new_flags
;
1283 unsigned long start
, end
, vma_end
;
1285 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1288 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1289 sizeof(uffdio_register
)-sizeof(__u64
)))
1293 if (!uffdio_register
.mode
)
1295 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1298 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1299 vm_flags
|= VM_UFFD_MISSING
;
1300 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1301 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1304 vm_flags
|= VM_UFFD_WP
;
1306 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1307 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1310 vm_flags
|= VM_UFFD_MINOR
;
1313 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1314 uffdio_register
.range
.len
);
1318 start
= uffdio_register
.range
.start
;
1319 end
= start
+ uffdio_register
.range
.len
;
1322 if (!mmget_not_zero(mm
))
1325 mmap_write_lock(mm
);
1326 vma
= find_vma_prev(mm
, start
, &prev
);
1330 /* check that there's at least one vma in the range */
1332 if (vma
->vm_start
>= end
)
1336 * If the first vma contains huge pages, make sure start address
1337 * is aligned to huge page size.
1339 if (is_vm_hugetlb_page(vma
)) {
1340 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1342 if (start
& (vma_hpagesize
- 1))
1347 * Search for not compatible vmas.
1350 basic_ioctls
= false;
1351 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1354 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1355 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1357 /* check not compatible vmas */
1359 if (!vma_can_userfault(cur
, vm_flags
))
1363 * UFFDIO_COPY will fill file holes even without
1364 * PROT_WRITE. This check enforces that if this is a
1365 * MAP_SHARED, the process has write permission to the backing
1366 * file. If VM_MAYWRITE is set it also enforces that on a
1367 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1368 * F_WRITE_SEAL can be taken until the vma is destroyed.
1371 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1375 * If this vma contains ending address, and huge pages
1378 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1379 end
> cur
->vm_start
) {
1380 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1384 if (end
& (vma_hpagesize
- 1))
1387 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1391 * Check that this vma isn't already owned by a
1392 * different userfaultfd. We can't allow more than one
1393 * userfaultfd to own a single vma simultaneously or we
1394 * wouldn't know which one to deliver the userfaults to.
1397 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1398 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1402 * Note vmas containing huge pages
1404 if (is_vm_hugetlb_page(cur
))
1405 basic_ioctls
= true;
1411 if (vma
->vm_start
< start
)
1418 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1419 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1420 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1421 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1424 * Nothing to do: this vma is already registered into this
1425 * userfaultfd and with the right tracking mode too.
1427 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1428 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1431 if (vma
->vm_start
> start
)
1432 start
= vma
->vm_start
;
1433 vma_end
= min(end
, vma
->vm_end
);
1435 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1436 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1437 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1439 ((struct vm_userfaultfd_ctx
){ ctx
}));
1444 if (vma
->vm_start
< start
) {
1445 ret
= split_vma(mm
, vma
, start
, 1);
1449 if (vma
->vm_end
> end
) {
1450 ret
= split_vma(mm
, vma
, end
, 0);
1456 * In the vma_merge() successful mprotect-like case 8:
1457 * the next vma was merged into the current one and
1458 * the current one has not been updated yet.
1460 vma
->vm_flags
= new_flags
;
1461 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1463 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1464 hugetlb_unshare_all_pmds(vma
);
1468 start
= vma
->vm_end
;
1470 } while (vma
&& vma
->vm_start
< end
);
1472 mmap_write_unlock(mm
);
1477 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1478 UFFD_API_RANGE_IOCTLS
;
1481 * Declare the WP ioctl only if the WP mode is
1482 * specified and all checks passed with the range
1484 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1485 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1487 /* CONTINUE ioctl is only supported for MINOR ranges. */
1488 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
))
1489 ioctls_out
&= ~((__u64
)1 << _UFFDIO_CONTINUE
);
1492 * Now that we scanned all vmas we can already tell
1493 * userland which ioctls methods are guaranteed to
1494 * succeed on this range.
1496 if (put_user(ioctls_out
, &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 mmap_write_lock(mm
);
1532 vma
= find_vma_prev(mm
, start
, &prev
);
1536 /* check that there's at least one vma in the range */
1538 if (vma
->vm_start
>= end
)
1542 * If the first vma contains huge pages, make sure start address
1543 * is aligned to huge page size.
1545 if (is_vm_hugetlb_page(vma
)) {
1546 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1548 if (start
& (vma_hpagesize
- 1))
1553 * Search for not compatible vmas.
1557 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1560 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1561 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1564 * Check not compatible vmas, not strictly required
1565 * here as not compatible vmas cannot have an
1566 * userfaultfd_ctx registered on them, but this
1567 * provides for more strict behavior to notice
1568 * unregistration errors.
1570 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1577 if (vma
->vm_start
< start
)
1584 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1587 * Nothing to do: this vma is already registered into this
1588 * userfaultfd and with the right tracking mode too.
1590 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1593 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1595 if (vma
->vm_start
> start
)
1596 start
= vma
->vm_start
;
1597 vma_end
= min(end
, vma
->vm_end
);
1599 if (userfaultfd_missing(vma
)) {
1601 * Wake any concurrent pending userfault while
1602 * we unregister, so they will not hang
1603 * permanently and it avoids userland to call
1604 * UFFDIO_WAKE explicitly.
1606 struct userfaultfd_wake_range range
;
1607 range
.start
= start
;
1608 range
.len
= vma_end
- start
;
1609 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1612 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1613 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1614 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1621 if (vma
->vm_start
< start
) {
1622 ret
= split_vma(mm
, vma
, start
, 1);
1626 if (vma
->vm_end
> end
) {
1627 ret
= split_vma(mm
, vma
, end
, 0);
1633 * In the vma_merge() successful mprotect-like case 8:
1634 * the next vma was merged into the current one and
1635 * the current one has not been updated yet.
1637 vma
->vm_flags
= new_flags
;
1638 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1642 start
= vma
->vm_end
;
1644 } while (vma
&& vma
->vm_start
< end
);
1646 mmap_write_unlock(mm
);
1653 * userfaultfd_wake may be used in combination with the
1654 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1656 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1660 struct uffdio_range uffdio_wake
;
1661 struct userfaultfd_wake_range range
;
1662 const void __user
*buf
= (void __user
*)arg
;
1665 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1668 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1672 range
.start
= uffdio_wake
.start
;
1673 range
.len
= uffdio_wake
.len
;
1676 * len == 0 means wake all and we don't want to wake all here,
1677 * so check it again to be sure.
1679 VM_BUG_ON(!range
.len
);
1681 wake_userfault(ctx
, &range
);
1688 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1692 struct uffdio_copy uffdio_copy
;
1693 struct uffdio_copy __user
*user_uffdio_copy
;
1694 struct userfaultfd_wake_range range
;
1696 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1699 if (atomic_read(&ctx
->mmap_changing
))
1703 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1704 /* don't copy "copy" last field */
1705 sizeof(uffdio_copy
)-sizeof(__s64
)))
1708 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1712 * double check for wraparound just in case. copy_from_user()
1713 * will later check uffdio_copy.src + uffdio_copy.len to fit
1714 * in the userland range.
1717 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1719 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1721 if (mmget_not_zero(ctx
->mm
)) {
1722 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1723 uffdio_copy
.len
, &ctx
->mmap_changing
,
1729 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1734 /* len == 0 would wake all */
1736 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1737 range
.start
= uffdio_copy
.dst
;
1738 wake_userfault(ctx
, &range
);
1740 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1745 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1749 struct uffdio_zeropage uffdio_zeropage
;
1750 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1751 struct userfaultfd_wake_range range
;
1753 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1756 if (atomic_read(&ctx
->mmap_changing
))
1760 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1761 /* don't copy "zeropage" last field */
1762 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1765 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1766 uffdio_zeropage
.range
.len
);
1770 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1773 if (mmget_not_zero(ctx
->mm
)) {
1774 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1775 uffdio_zeropage
.range
.len
,
1776 &ctx
->mmap_changing
);
1781 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1785 /* len == 0 would wake all */
1788 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1789 range
.start
= uffdio_zeropage
.range
.start
;
1790 wake_userfault(ctx
, &range
);
1792 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1797 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1801 struct uffdio_writeprotect uffdio_wp
;
1802 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1803 struct userfaultfd_wake_range range
;
1804 bool mode_wp
, mode_dontwake
;
1806 if (atomic_read(&ctx
->mmap_changing
))
1809 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1811 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1812 sizeof(struct uffdio_writeprotect
)))
1815 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1816 uffdio_wp
.range
.len
);
1820 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1821 UFFDIO_WRITEPROTECT_MODE_WP
))
1824 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1825 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1827 if (mode_wp
&& mode_dontwake
)
1830 if (mmget_not_zero(ctx
->mm
)) {
1831 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1832 uffdio_wp
.range
.len
, mode_wp
,
1833 &ctx
->mmap_changing
);
1842 if (!mode_wp
&& !mode_dontwake
) {
1843 range
.start
= uffdio_wp
.range
.start
;
1844 range
.len
= uffdio_wp
.range
.len
;
1845 wake_userfault(ctx
, &range
);
1850 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1853 struct uffdio_continue uffdio_continue
;
1854 struct uffdio_continue __user
*user_uffdio_continue
;
1855 struct userfaultfd_wake_range range
;
1857 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1860 if (atomic_read(&ctx
->mmap_changing
))
1864 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1865 /* don't copy the output fields */
1866 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1869 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1870 uffdio_continue
.range
.len
);
1875 /* double check for wraparound just in case. */
1876 if (uffdio_continue
.range
.start
+ uffdio_continue
.range
.len
<=
1877 uffdio_continue
.range
.start
) {
1880 if (uffdio_continue
.mode
& ~UFFDIO_CONTINUE_MODE_DONTWAKE
)
1883 if (mmget_not_zero(ctx
->mm
)) {
1884 ret
= mcopy_continue(ctx
->mm
, uffdio_continue
.range
.start
,
1885 uffdio_continue
.range
.len
,
1886 &ctx
->mmap_changing
);
1892 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1897 /* len == 0 would wake all */
1900 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1901 range
.start
= uffdio_continue
.range
.start
;
1902 wake_userfault(ctx
, &range
);
1904 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1910 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1913 * For the current set of features the bits just coincide. Set
1914 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1916 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1920 * userland asks for a certain API version and we return which bits
1921 * and ioctl commands are implemented in this kernel for such API
1922 * version or -EINVAL if unknown.
1924 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1927 struct uffdio_api uffdio_api
;
1928 void __user
*buf
= (void __user
*)arg
;
1929 unsigned int ctx_features
;
1934 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1936 features
= uffdio_api
.features
;
1938 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
1941 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1943 /* report all available features and ioctls to userland */
1944 uffdio_api
.features
= UFFD_API_FEATURES
;
1945 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1946 uffdio_api
.features
&=
1947 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
1949 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1950 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
1952 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1954 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1957 /* only enable the requested features for this uffd context */
1958 ctx_features
= uffd_ctx_features(features
);
1960 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
1967 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1968 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1973 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1977 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1979 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
1984 ret
= userfaultfd_api(ctx
, arg
);
1986 case UFFDIO_REGISTER
:
1987 ret
= userfaultfd_register(ctx
, arg
);
1989 case UFFDIO_UNREGISTER
:
1990 ret
= userfaultfd_unregister(ctx
, arg
);
1993 ret
= userfaultfd_wake(ctx
, arg
);
1996 ret
= userfaultfd_copy(ctx
, arg
);
1998 case UFFDIO_ZEROPAGE
:
1999 ret
= userfaultfd_zeropage(ctx
, arg
);
2001 case UFFDIO_WRITEPROTECT
:
2002 ret
= userfaultfd_writeprotect(ctx
, arg
);
2004 case UFFDIO_CONTINUE
:
2005 ret
= userfaultfd_continue(ctx
, arg
);
2011 #ifdef CONFIG_PROC_FS
2012 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2014 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2015 wait_queue_entry_t
*wq
;
2016 unsigned long pending
= 0, total
= 0;
2018 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2019 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2023 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2026 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2029 * If more protocols will be added, there will be all shown
2030 * separated by a space. Like this:
2031 * protocols: aa:... bb:...
2033 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2034 pending
, total
, UFFD_API
, ctx
->features
,
2035 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2039 static const struct file_operations userfaultfd_fops
= {
2040 #ifdef CONFIG_PROC_FS
2041 .show_fdinfo
= userfaultfd_show_fdinfo
,
2043 .release
= userfaultfd_release
,
2044 .poll
= userfaultfd_poll
,
2045 .read
= userfaultfd_read
,
2046 .unlocked_ioctl
= userfaultfd_ioctl
,
2047 .compat_ioctl
= compat_ptr_ioctl
,
2048 .llseek
= noop_llseek
,
2051 static void init_once_userfaultfd_ctx(void *mem
)
2053 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2055 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2056 init_waitqueue_head(&ctx
->fault_wqh
);
2057 init_waitqueue_head(&ctx
->event_wqh
);
2058 init_waitqueue_head(&ctx
->fd_wqh
);
2059 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2062 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2064 struct userfaultfd_ctx
*ctx
;
2067 if (!sysctl_unprivileged_userfaultfd
&&
2068 (flags
& UFFD_USER_MODE_ONLY
) == 0 &&
2069 !capable(CAP_SYS_PTRACE
)) {
2070 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
2071 "sysctl knob to 1 if kernel faults must be handled "
2072 "without obtaining CAP_SYS_PTRACE capability\n");
2076 BUG_ON(!current
->mm
);
2078 /* Check the UFFD_* constants for consistency. */
2079 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2080 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2081 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2083 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2086 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2090 refcount_set(&ctx
->refcount
, 1);
2093 ctx
->released
= false;
2094 atomic_set(&ctx
->mmap_changing
, 0);
2095 ctx
->mm
= current
->mm
;
2096 /* prevent the mm struct to be freed */
2099 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, ctx
,
2100 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2103 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2108 static int __init
userfaultfd_init(void)
2110 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2111 sizeof(struct userfaultfd_ctx
),
2113 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2114 init_once_userfaultfd_ctx
);
2117 __initcall(userfaultfd_init
);