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/mm_inline.h>
19 #include <linux/mmu_notifier.h>
20 #include <linux/poll.h>
21 #include <linux/slab.h>
22 #include <linux/seq_file.h>
23 #include <linux/file.h>
24 #include <linux/bug.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h>
30 #include <linux/security.h>
31 #include <linux/hugetlb.h>
32 #include <linux/swapops.h>
34 int sysctl_unprivileged_userfaultfd __read_mostly
;
36 static struct kmem_cache
*userfaultfd_ctx_cachep __read_mostly
;
39 * Start with fault_pending_wqh and fault_wqh so they're more likely
40 * to be in the same cacheline.
44 * fault_pending_wqh.lock
48 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
49 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
50 * also taken in IRQ context.
52 struct userfaultfd_ctx
{
53 /* waitqueue head for the pending (i.e. not read) userfaults */
54 wait_queue_head_t fault_pending_wqh
;
55 /* waitqueue head for the userfaults */
56 wait_queue_head_t fault_wqh
;
57 /* waitqueue head for the pseudo fd to wakeup poll/read */
58 wait_queue_head_t fd_wqh
;
59 /* waitqueue head for events */
60 wait_queue_head_t event_wqh
;
61 /* a refile sequence protected by fault_pending_wqh lock */
62 seqcount_spinlock_t refile_seq
;
63 /* pseudo fd refcounting */
65 /* userfaultfd syscall flags */
67 /* features requested from the userspace */
68 unsigned int features
;
71 /* memory mappings are changing because of non-cooperative event */
72 atomic_t mmap_changing
;
73 /* mm with one ore more vmas attached to this userfaultfd_ctx */
77 struct userfaultfd_fork_ctx
{
78 struct userfaultfd_ctx
*orig
;
79 struct userfaultfd_ctx
*new;
80 struct list_head list
;
83 struct userfaultfd_unmap_ctx
{
84 struct userfaultfd_ctx
*ctx
;
87 struct list_head list
;
90 struct userfaultfd_wait_queue
{
92 wait_queue_entry_t wq
;
93 struct userfaultfd_ctx
*ctx
;
97 struct userfaultfd_wake_range
{
102 /* internal indication that UFFD_API ioctl was successfully executed */
103 #define UFFD_FEATURE_INITIALIZED (1u << 31)
105 static bool userfaultfd_is_initialized(struct userfaultfd_ctx
*ctx
)
107 return ctx
->features
& UFFD_FEATURE_INITIALIZED
;
110 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
111 int wake_flags
, void *key
)
113 struct userfaultfd_wake_range
*range
= key
;
115 struct userfaultfd_wait_queue
*uwq
;
116 unsigned long start
, len
;
118 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
120 /* len == 0 means wake all */
121 start
= range
->start
;
123 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
124 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
126 WRITE_ONCE(uwq
->waken
, true);
128 * The Program-Order guarantees provided by the scheduler
129 * ensure uwq->waken is visible before the task is woken.
131 ret
= wake_up_state(wq
->private, mode
);
134 * Wake only once, autoremove behavior.
136 * After the effect of list_del_init is visible to the other
137 * CPUs, the waitqueue may disappear from under us, see the
138 * !list_empty_careful() in handle_userfault().
140 * try_to_wake_up() has an implicit smp_mb(), and the
141 * wq->private is read before calling the extern function
142 * "wake_up_state" (which in turns calls try_to_wake_up).
144 list_del_init(&wq
->entry
);
151 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
153 * @ctx: [in] Pointer to the userfaultfd context.
155 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
157 refcount_inc(&ctx
->refcount
);
161 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
163 * @ctx: [in] Pointer to userfaultfd context.
165 * The userfaultfd context reference must have been previously acquired either
166 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
168 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
170 if (refcount_dec_and_test(&ctx
->refcount
)) {
171 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
172 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
173 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
174 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
175 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
176 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
177 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
178 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
180 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
184 static inline void msg_init(struct uffd_msg
*msg
)
186 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
188 * Must use memset to zero out the paddings or kernel data is
189 * leaked to userland.
191 memset(msg
, 0, sizeof(struct uffd_msg
));
194 static inline struct uffd_msg
userfault_msg(unsigned long address
,
195 unsigned long real_address
,
197 unsigned long reason
,
198 unsigned int features
)
203 msg
.event
= UFFD_EVENT_PAGEFAULT
;
205 msg
.arg
.pagefault
.address
= (features
& UFFD_FEATURE_EXACT_ADDRESS
) ?
206 real_address
: address
;
209 * These flags indicate why the userfault occurred:
210 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
211 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
212 * - Neither of these flags being set indicates a MISSING fault.
214 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
215 * fault. Otherwise, it was a read fault.
217 if (flags
& FAULT_FLAG_WRITE
)
218 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
219 if (reason
& VM_UFFD_WP
)
220 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
221 if (reason
& VM_UFFD_MINOR
)
222 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_MINOR
;
223 if (features
& UFFD_FEATURE_THREAD_ID
)
224 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
228 #ifdef CONFIG_HUGETLB_PAGE
230 * Same functionality as userfaultfd_must_wait below with modifications for
233 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
234 struct vm_area_struct
*vma
,
235 unsigned long address
,
237 unsigned long reason
)
239 struct mm_struct
*mm
= ctx
->mm
;
243 mmap_assert_locked(mm
);
245 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
251 pte
= huge_ptep_get(ptep
);
254 * Lockless access: we're in a wait_event so it's ok if it
255 * changes under us. PTE markers should be handled the same as none
258 if (huge_pte_none_mostly(pte
))
260 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
266 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
267 struct vm_area_struct
*vma
,
268 unsigned long address
,
270 unsigned long reason
)
272 return false; /* should never get here */
274 #endif /* CONFIG_HUGETLB_PAGE */
277 * Verify the pagetables are still not ok after having reigstered into
278 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
279 * userfault that has already been resolved, if userfaultfd_read and
280 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
283 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
284 unsigned long address
,
286 unsigned long reason
)
288 struct mm_struct
*mm
= ctx
->mm
;
296 mmap_assert_locked(mm
);
298 pgd
= pgd_offset(mm
, address
);
299 if (!pgd_present(*pgd
))
301 p4d
= p4d_offset(pgd
, address
);
302 if (!p4d_present(*p4d
))
304 pud
= pud_offset(p4d
, address
);
305 if (!pud_present(*pud
))
307 pmd
= pmd_offset(pud
, address
);
309 * READ_ONCE must function as a barrier with narrower scope
310 * and it must be equivalent to:
311 * _pmd = *pmd; barrier();
313 * This is to deal with the instability (as in
314 * pmd_trans_unstable) of the pmd.
316 _pmd
= READ_ONCE(*pmd
);
321 if (!pmd_present(_pmd
))
324 if (pmd_trans_huge(_pmd
)) {
325 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
331 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
332 * and use the standard pte_offset_map() instead of parsing _pmd.
334 pte
= pte_offset_map(pmd
, address
);
336 * Lockless access: we're in a wait_event so it's ok if it
337 * changes under us. PTE markers should be handled the same as none
340 if (pte_none_mostly(*pte
))
342 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
350 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
352 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
353 return TASK_INTERRUPTIBLE
;
355 if (flags
& FAULT_FLAG_KILLABLE
)
356 return TASK_KILLABLE
;
358 return TASK_UNINTERRUPTIBLE
;
362 * The locking rules involved in returning VM_FAULT_RETRY depending on
363 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
364 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
365 * recommendation in __lock_page_or_retry is not an understatement.
367 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
368 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
371 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
372 * set, VM_FAULT_RETRY can still be returned if and only if there are
373 * fatal_signal_pending()s, and the mmap_lock must be released before
376 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
378 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
379 struct userfaultfd_ctx
*ctx
;
380 struct userfaultfd_wait_queue uwq
;
381 vm_fault_t ret
= VM_FAULT_SIGBUS
;
383 unsigned int blocking_state
;
386 * We don't do userfault handling for the final child pid update.
388 * We also don't do userfault handling during
389 * coredumping. hugetlbfs has the special
390 * follow_hugetlb_page() to skip missing pages in the
391 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
392 * the no_page_table() helper in follow_page_mask(), but the
393 * shmem_vm_ops->fault method is invoked even during
394 * coredumping without mmap_lock and it ends up here.
396 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
400 * Coredumping runs without mmap_lock so we can only check that
401 * the mmap_lock is held, if PF_DUMPCORE was not set.
403 mmap_assert_locked(mm
);
405 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
409 BUG_ON(ctx
->mm
!= mm
);
411 /* Any unrecognized flag is a bug. */
412 VM_BUG_ON(reason
& ~__VM_UFFD_FLAGS
);
413 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
414 VM_BUG_ON(!reason
|| (reason
& (reason
- 1)));
416 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
418 if ((vmf
->flags
& FAULT_FLAG_USER
) == 0 &&
419 ctx
->flags
& UFFD_USER_MODE_ONLY
) {
420 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
421 "sysctl knob to 1 if kernel faults must be handled "
422 "without obtaining CAP_SYS_PTRACE capability\n");
427 * If it's already released don't get it. This avoids to loop
428 * in __get_user_pages if userfaultfd_release waits on the
429 * caller of handle_userfault to release the mmap_lock.
431 if (unlikely(READ_ONCE(ctx
->released
))) {
433 * Don't return VM_FAULT_SIGBUS in this case, so a non
434 * cooperative manager can close the uffd after the
435 * last UFFDIO_COPY, without risking to trigger an
436 * involuntary SIGBUS if the process was starting the
437 * userfaultfd while the userfaultfd was still armed
438 * (but after the last UFFDIO_COPY). If the uffd
439 * wasn't already closed when the userfault reached
440 * this point, that would normally be solved by
441 * userfaultfd_must_wait returning 'false'.
443 * If we were to return VM_FAULT_SIGBUS here, the non
444 * cooperative manager would be instead forced to
445 * always call UFFDIO_UNREGISTER before it can safely
448 ret
= VM_FAULT_NOPAGE
;
453 * Check that we can return VM_FAULT_RETRY.
455 * NOTE: it should become possible to return VM_FAULT_RETRY
456 * even if FAULT_FLAG_TRIED is set without leading to gup()
457 * -EBUSY failures, if the userfaultfd is to be extended for
458 * VM_UFFD_WP tracking and we intend to arm the userfault
459 * without first stopping userland access to the memory. For
460 * VM_UFFD_MISSING userfaults this is enough for now.
462 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
464 * Validate the invariant that nowait must allow retry
465 * to be sure not to return SIGBUS erroneously on
466 * nowait invocations.
468 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
469 #ifdef CONFIG_DEBUG_VM
470 if (printk_ratelimit()) {
472 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
481 * Handle nowait, not much to do other than tell it to retry
484 ret
= VM_FAULT_RETRY
;
485 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
488 /* take the reference before dropping the mmap_lock */
489 userfaultfd_ctx_get(ctx
);
491 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
492 uwq
.wq
.private = current
;
493 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->real_address
, vmf
->flags
,
494 reason
, ctx
->features
);
498 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
500 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
502 * After the __add_wait_queue the uwq is visible to userland
503 * through poll/read().
505 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
507 * The smp_mb() after __set_current_state prevents the reads
508 * following the spin_unlock to happen before the list_add in
511 set_current_state(blocking_state
);
512 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
514 if (!is_vm_hugetlb_page(vmf
->vma
))
515 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
518 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
521 mmap_read_unlock(mm
);
523 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
524 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
528 __set_current_state(TASK_RUNNING
);
531 * Here we race with the list_del; list_add in
532 * userfaultfd_ctx_read(), however because we don't ever run
533 * list_del_init() to refile across the two lists, the prev
534 * and next pointers will never point to self. list_add also
535 * would never let any of the two pointers to point to
536 * self. So list_empty_careful won't risk to see both pointers
537 * pointing to self at any time during the list refile. The
538 * only case where list_del_init() is called is the full
539 * removal in the wake function and there we don't re-list_add
540 * and it's fine not to block on the spinlock. The uwq on this
541 * kernel stack can be released after the list_del_init.
543 if (!list_empty_careful(&uwq
.wq
.entry
)) {
544 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
546 * No need of list_del_init(), the uwq on the stack
547 * will be freed shortly anyway.
549 list_del(&uwq
.wq
.entry
);
550 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
554 * ctx may go away after this if the userfault pseudo fd is
557 userfaultfd_ctx_put(ctx
);
563 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
564 struct userfaultfd_wait_queue
*ewq
)
566 struct userfaultfd_ctx
*release_new_ctx
;
568 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
572 init_waitqueue_entry(&ewq
->wq
, current
);
573 release_new_ctx
= NULL
;
575 spin_lock_irq(&ctx
->event_wqh
.lock
);
577 * After the __add_wait_queue the uwq is visible to userland
578 * through poll/read().
580 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
582 set_current_state(TASK_KILLABLE
);
583 if (ewq
->msg
.event
== 0)
585 if (READ_ONCE(ctx
->released
) ||
586 fatal_signal_pending(current
)) {
588 * &ewq->wq may be queued in fork_event, but
589 * __remove_wait_queue ignores the head
590 * parameter. It would be a problem if it
593 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
594 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
595 struct userfaultfd_ctx
*new;
597 new = (struct userfaultfd_ctx
*)
599 ewq
->msg
.arg
.reserved
.reserved1
;
600 release_new_ctx
= new;
605 spin_unlock_irq(&ctx
->event_wqh
.lock
);
607 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
610 spin_lock_irq(&ctx
->event_wqh
.lock
);
612 __set_current_state(TASK_RUNNING
);
613 spin_unlock_irq(&ctx
->event_wqh
.lock
);
615 if (release_new_ctx
) {
616 struct vm_area_struct
*vma
;
617 struct mm_struct
*mm
= release_new_ctx
->mm
;
619 /* the various vma->vm_userfaultfd_ctx still points to it */
621 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
622 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
623 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
624 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
626 mmap_write_unlock(mm
);
628 userfaultfd_ctx_put(release_new_ctx
);
632 * ctx may go away after this if the userfault pseudo fd is
636 atomic_dec(&ctx
->mmap_changing
);
637 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
638 userfaultfd_ctx_put(ctx
);
641 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
642 struct userfaultfd_wait_queue
*ewq
)
645 wake_up_locked(&ctx
->event_wqh
);
646 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
649 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
651 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
652 struct userfaultfd_fork_ctx
*fctx
;
654 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
655 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
656 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
657 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
661 list_for_each_entry(fctx
, fcs
, list
)
662 if (fctx
->orig
== octx
) {
668 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
672 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
678 refcount_set(&ctx
->refcount
, 1);
679 ctx
->flags
= octx
->flags
;
680 ctx
->features
= octx
->features
;
681 ctx
->released
= false;
682 atomic_set(&ctx
->mmap_changing
, 0);
683 ctx
->mm
= vma
->vm_mm
;
686 userfaultfd_ctx_get(octx
);
687 atomic_inc(&octx
->mmap_changing
);
690 list_add_tail(&fctx
->list
, fcs
);
693 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
697 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
699 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
700 struct userfaultfd_wait_queue ewq
;
704 ewq
.msg
.event
= UFFD_EVENT_FORK
;
705 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
707 userfaultfd_event_wait_completion(ctx
, &ewq
);
710 void dup_userfaultfd_complete(struct list_head
*fcs
)
712 struct userfaultfd_fork_ctx
*fctx
, *n
;
714 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
716 list_del(&fctx
->list
);
721 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
722 struct vm_userfaultfd_ctx
*vm_ctx
)
724 struct userfaultfd_ctx
*ctx
;
726 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
731 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
733 userfaultfd_ctx_get(ctx
);
734 atomic_inc(&ctx
->mmap_changing
);
736 /* Drop uffd context if remap feature not enabled */
737 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
738 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
742 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
743 unsigned long from
, unsigned long to
,
746 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
747 struct userfaultfd_wait_queue ewq
;
752 if (to
& ~PAGE_MASK
) {
753 userfaultfd_ctx_put(ctx
);
759 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
760 ewq
.msg
.arg
.remap
.from
= from
;
761 ewq
.msg
.arg
.remap
.to
= to
;
762 ewq
.msg
.arg
.remap
.len
= len
;
764 userfaultfd_event_wait_completion(ctx
, &ewq
);
767 bool userfaultfd_remove(struct vm_area_struct
*vma
,
768 unsigned long start
, unsigned long end
)
770 struct mm_struct
*mm
= vma
->vm_mm
;
771 struct userfaultfd_ctx
*ctx
;
772 struct userfaultfd_wait_queue ewq
;
774 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
775 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
778 userfaultfd_ctx_get(ctx
);
779 atomic_inc(&ctx
->mmap_changing
);
780 mmap_read_unlock(mm
);
784 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
785 ewq
.msg
.arg
.remove
.start
= start
;
786 ewq
.msg
.arg
.remove
.end
= end
;
788 userfaultfd_event_wait_completion(ctx
, &ewq
);
793 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
794 unsigned long start
, unsigned long end
)
796 struct userfaultfd_unmap_ctx
*unmap_ctx
;
798 list_for_each_entry(unmap_ctx
, unmaps
, list
)
799 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
800 unmap_ctx
->end
== end
)
806 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
807 unsigned long start
, unsigned long end
,
808 struct list_head
*unmaps
)
810 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
811 struct userfaultfd_unmap_ctx
*unmap_ctx
;
812 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
814 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
815 has_unmap_ctx(ctx
, unmaps
, start
, end
))
818 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
822 userfaultfd_ctx_get(ctx
);
823 atomic_inc(&ctx
->mmap_changing
);
824 unmap_ctx
->ctx
= ctx
;
825 unmap_ctx
->start
= start
;
826 unmap_ctx
->end
= end
;
827 list_add_tail(&unmap_ctx
->list
, unmaps
);
833 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
835 struct userfaultfd_unmap_ctx
*ctx
, *n
;
836 struct userfaultfd_wait_queue ewq
;
838 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
841 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
842 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
843 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
845 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
847 list_del(&ctx
->list
);
852 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
854 struct userfaultfd_ctx
*ctx
= file
->private_data
;
855 struct mm_struct
*mm
= ctx
->mm
;
856 struct vm_area_struct
*vma
, *prev
;
857 /* len == 0 means wake all */
858 struct userfaultfd_wake_range range
= { .len
= 0, };
859 unsigned long new_flags
;
861 WRITE_ONCE(ctx
->released
, true);
863 if (!mmget_not_zero(mm
))
867 * Flush page faults out of all CPUs. NOTE: all page faults
868 * must be retried without returning VM_FAULT_SIGBUS if
869 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
870 * changes while handle_userfault released the mmap_lock. So
871 * it's critical that released is set to true (above), before
872 * taking the mmap_lock for writing.
876 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
878 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
879 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
880 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
884 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
885 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
886 new_flags
, vma
->anon_vma
,
887 vma
->vm_file
, vma
->vm_pgoff
,
889 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
894 vma
->vm_flags
= new_flags
;
895 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
897 mmap_write_unlock(mm
);
901 * After no new page faults can wait on this fault_*wqh, flush
902 * the last page faults that may have been already waiting on
905 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
906 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
907 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
908 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
910 /* Flush pending events that may still wait on event_wqh */
911 wake_up_all(&ctx
->event_wqh
);
913 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
914 userfaultfd_ctx_put(ctx
);
918 /* fault_pending_wqh.lock must be hold by the caller */
919 static inline struct userfaultfd_wait_queue
*find_userfault_in(
920 wait_queue_head_t
*wqh
)
922 wait_queue_entry_t
*wq
;
923 struct userfaultfd_wait_queue
*uwq
;
925 lockdep_assert_held(&wqh
->lock
);
928 if (!waitqueue_active(wqh
))
930 /* walk in reverse to provide FIFO behavior to read userfaults */
931 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
932 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
937 static inline struct userfaultfd_wait_queue
*find_userfault(
938 struct userfaultfd_ctx
*ctx
)
940 return find_userfault_in(&ctx
->fault_pending_wqh
);
943 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
944 struct userfaultfd_ctx
*ctx
)
946 return find_userfault_in(&ctx
->event_wqh
);
949 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
951 struct userfaultfd_ctx
*ctx
= file
->private_data
;
954 poll_wait(file
, &ctx
->fd_wqh
, wait
);
956 if (!userfaultfd_is_initialized(ctx
))
960 * poll() never guarantees that read won't block.
961 * userfaults can be waken before they're read().
963 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
966 * lockless access to see if there are pending faults
967 * __pollwait last action is the add_wait_queue but
968 * the spin_unlock would allow the waitqueue_active to
969 * pass above the actual list_add inside
970 * add_wait_queue critical section. So use a full
971 * memory barrier to serialize the list_add write of
972 * add_wait_queue() with the waitqueue_active read
977 if (waitqueue_active(&ctx
->fault_pending_wqh
))
979 else if (waitqueue_active(&ctx
->event_wqh
))
985 static const struct file_operations userfaultfd_fops
;
987 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
989 struct uffd_msg
*msg
)
993 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, new,
994 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
998 msg
->arg
.reserved
.reserved1
= 0;
999 msg
->arg
.fork
.ufd
= fd
;
1003 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1004 struct uffd_msg
*msg
, struct inode
*inode
)
1007 DECLARE_WAITQUEUE(wait
, current
);
1008 struct userfaultfd_wait_queue
*uwq
;
1010 * Handling fork event requires sleeping operations, so
1011 * we drop the event_wqh lock, then do these ops, then
1012 * lock it back and wake up the waiter. While the lock is
1013 * dropped the ewq may go away so we keep track of it
1016 LIST_HEAD(fork_event
);
1017 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1019 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1020 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1021 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1023 set_current_state(TASK_INTERRUPTIBLE
);
1024 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1025 uwq
= find_userfault(ctx
);
1028 * Use a seqcount to repeat the lockless check
1029 * in wake_userfault() to avoid missing
1030 * wakeups because during the refile both
1031 * waitqueue could become empty if this is the
1034 write_seqcount_begin(&ctx
->refile_seq
);
1037 * The fault_pending_wqh.lock prevents the uwq
1038 * to disappear from under us.
1040 * Refile this userfault from
1041 * fault_pending_wqh to fault_wqh, it's not
1042 * pending anymore after we read it.
1044 * Use list_del() by hand (as
1045 * userfaultfd_wake_function also uses
1046 * list_del_init() by hand) to be sure nobody
1047 * changes __remove_wait_queue() to use
1048 * list_del_init() in turn breaking the
1049 * !list_empty_careful() check in
1050 * handle_userfault(). The uwq->wq.head list
1051 * must never be empty at any time during the
1052 * refile, or the waitqueue could disappear
1053 * from under us. The "wait_queue_head_t"
1054 * parameter of __remove_wait_queue() is unused
1057 list_del(&uwq
->wq
.entry
);
1058 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1060 write_seqcount_end(&ctx
->refile_seq
);
1062 /* careful to always initialize msg if ret == 0 */
1064 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1068 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1070 spin_lock(&ctx
->event_wqh
.lock
);
1071 uwq
= find_userfault_evt(ctx
);
1075 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1076 fork_nctx
= (struct userfaultfd_ctx
*)
1078 uwq
->msg
.arg
.reserved
.reserved1
;
1079 list_move(&uwq
->wq
.entry
, &fork_event
);
1081 * fork_nctx can be freed as soon as
1082 * we drop the lock, unless we take a
1085 userfaultfd_ctx_get(fork_nctx
);
1086 spin_unlock(&ctx
->event_wqh
.lock
);
1091 userfaultfd_event_complete(ctx
, uwq
);
1092 spin_unlock(&ctx
->event_wqh
.lock
);
1096 spin_unlock(&ctx
->event_wqh
.lock
);
1098 if (signal_pending(current
)) {
1106 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1108 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1110 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1111 __set_current_state(TASK_RUNNING
);
1112 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1114 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1115 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1116 spin_lock_irq(&ctx
->event_wqh
.lock
);
1117 if (!list_empty(&fork_event
)) {
1119 * The fork thread didn't abort, so we can
1120 * drop the temporary refcount.
1122 userfaultfd_ctx_put(fork_nctx
);
1124 uwq
= list_first_entry(&fork_event
,
1128 * If fork_event list wasn't empty and in turn
1129 * the event wasn't already released by fork
1130 * (the event is allocated on fork kernel
1131 * stack), put the event back to its place in
1132 * the event_wq. fork_event head will be freed
1133 * as soon as we return so the event cannot
1134 * stay queued there no matter the current
1137 list_del(&uwq
->wq
.entry
);
1138 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1141 * Leave the event in the waitqueue and report
1142 * error to userland if we failed to resolve
1143 * the userfault fork.
1146 userfaultfd_event_complete(ctx
, uwq
);
1149 * Here the fork thread aborted and the
1150 * refcount from the fork thread on fork_nctx
1151 * has already been released. We still hold
1152 * the reference we took before releasing the
1153 * lock above. If resolve_userfault_fork
1154 * failed we've to drop it because the
1155 * fork_nctx has to be freed in such case. If
1156 * it succeeded we'll hold it because the new
1157 * uffd references it.
1160 userfaultfd_ctx_put(fork_nctx
);
1162 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1168 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1169 size_t count
, loff_t
*ppos
)
1171 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1172 ssize_t _ret
, ret
= 0;
1173 struct uffd_msg msg
;
1174 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1175 struct inode
*inode
= file_inode(file
);
1177 if (!userfaultfd_is_initialized(ctx
))
1181 if (count
< sizeof(msg
))
1182 return ret
? ret
: -EINVAL
;
1183 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1185 return ret
? ret
: _ret
;
1186 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1187 return ret
? ret
: -EFAULT
;
1190 count
-= sizeof(msg
);
1192 * Allow to read more than one fault at time but only
1193 * block if waiting for the very first one.
1195 no_wait
= O_NONBLOCK
;
1199 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1200 struct userfaultfd_wake_range
*range
)
1202 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1203 /* wake all in the range and autoremove */
1204 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1205 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1207 if (waitqueue_active(&ctx
->fault_wqh
))
1208 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1209 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1212 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1213 struct userfaultfd_wake_range
*range
)
1219 * To be sure waitqueue_active() is not reordered by the CPU
1220 * before the pagetable update, use an explicit SMP memory
1221 * barrier here. PT lock release or mmap_read_unlock(mm) still
1222 * have release semantics that can allow the
1223 * waitqueue_active() to be reordered before the pte update.
1228 * Use waitqueue_active because it's very frequent to
1229 * change the address space atomically even if there are no
1230 * userfaults yet. So we take the spinlock only when we're
1231 * sure we've userfaults to wake.
1234 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1235 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1236 waitqueue_active(&ctx
->fault_wqh
);
1238 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1240 __wake_userfault(ctx
, range
);
1243 static __always_inline
int validate_range(struct mm_struct
*mm
,
1244 __u64 start
, __u64 len
)
1246 __u64 task_size
= mm
->task_size
;
1248 if (start
& ~PAGE_MASK
)
1250 if (len
& ~PAGE_MASK
)
1254 if (start
< mmap_min_addr
)
1256 if (start
>= task_size
)
1258 if (len
> task_size
- start
)
1263 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1266 struct mm_struct
*mm
= ctx
->mm
;
1267 struct vm_area_struct
*vma
, *prev
, *cur
;
1269 struct uffdio_register uffdio_register
;
1270 struct uffdio_register __user
*user_uffdio_register
;
1271 unsigned long vm_flags
, new_flags
;
1274 unsigned long start
, end
, vma_end
;
1276 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1279 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1280 sizeof(uffdio_register
)-sizeof(__u64
)))
1284 if (!uffdio_register
.mode
)
1286 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1289 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1290 vm_flags
|= VM_UFFD_MISSING
;
1291 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1292 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1295 vm_flags
|= VM_UFFD_WP
;
1297 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1298 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1301 vm_flags
|= VM_UFFD_MINOR
;
1304 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1305 uffdio_register
.range
.len
);
1309 start
= uffdio_register
.range
.start
;
1310 end
= start
+ uffdio_register
.range
.len
;
1313 if (!mmget_not_zero(mm
))
1316 mmap_write_lock(mm
);
1317 vma
= find_vma_prev(mm
, start
, &prev
);
1321 /* check that there's at least one vma in the range */
1323 if (vma
->vm_start
>= end
)
1327 * If the first vma contains huge pages, make sure start address
1328 * is aligned to huge page size.
1330 if (is_vm_hugetlb_page(vma
)) {
1331 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1333 if (start
& (vma_hpagesize
- 1))
1338 * Search for not compatible vmas.
1341 basic_ioctls
= false;
1342 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1345 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1346 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1348 /* check not compatible vmas */
1350 if (!vma_can_userfault(cur
, vm_flags
))
1354 * UFFDIO_COPY will fill file holes even without
1355 * PROT_WRITE. This check enforces that if this is a
1356 * MAP_SHARED, the process has write permission to the backing
1357 * file. If VM_MAYWRITE is set it also enforces that on a
1358 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1359 * F_WRITE_SEAL can be taken until the vma is destroyed.
1362 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1366 * If this vma contains ending address, and huge pages
1369 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1370 end
> cur
->vm_start
) {
1371 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1375 if (end
& (vma_hpagesize
- 1))
1378 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1382 * Check that this vma isn't already owned by a
1383 * different userfaultfd. We can't allow more than one
1384 * userfaultfd to own a single vma simultaneously or we
1385 * wouldn't know which one to deliver the userfaults to.
1388 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1389 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1393 * Note vmas containing huge pages
1395 if (is_vm_hugetlb_page(cur
))
1396 basic_ioctls
= true;
1402 if (vma
->vm_start
< start
)
1409 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1410 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1411 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1412 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1415 * Nothing to do: this vma is already registered into this
1416 * userfaultfd and with the right tracking mode too.
1418 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1419 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1422 if (vma
->vm_start
> start
)
1423 start
= vma
->vm_start
;
1424 vma_end
= min(end
, vma
->vm_end
);
1426 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1427 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1428 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1430 ((struct vm_userfaultfd_ctx
){ ctx
}),
1431 anon_vma_name(vma
));
1436 if (vma
->vm_start
< start
) {
1437 ret
= split_vma(mm
, vma
, start
, 1);
1441 if (vma
->vm_end
> end
) {
1442 ret
= split_vma(mm
, vma
, end
, 0);
1448 * In the vma_merge() successful mprotect-like case 8:
1449 * the next vma was merged into the current one and
1450 * the current one has not been updated yet.
1452 vma
->vm_flags
= new_flags
;
1453 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1455 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1456 hugetlb_unshare_all_pmds(vma
);
1460 start
= vma
->vm_end
;
1462 } while (vma
&& vma
->vm_start
< end
);
1464 mmap_write_unlock(mm
);
1469 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1470 UFFD_API_RANGE_IOCTLS
;
1473 * Declare the WP ioctl only if the WP mode is
1474 * specified and all checks passed with the range
1476 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1477 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1479 /* CONTINUE ioctl is only supported for MINOR ranges. */
1480 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
))
1481 ioctls_out
&= ~((__u64
)1 << _UFFDIO_CONTINUE
);
1484 * Now that we scanned all vmas we can already tell
1485 * userland which ioctls methods are guaranteed to
1486 * succeed on this range.
1488 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1495 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1498 struct mm_struct
*mm
= ctx
->mm
;
1499 struct vm_area_struct
*vma
, *prev
, *cur
;
1501 struct uffdio_range uffdio_unregister
;
1502 unsigned long new_flags
;
1504 unsigned long start
, end
, vma_end
;
1505 const void __user
*buf
= (void __user
*)arg
;
1508 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1511 ret
= validate_range(mm
, uffdio_unregister
.start
,
1512 uffdio_unregister
.len
);
1516 start
= uffdio_unregister
.start
;
1517 end
= start
+ uffdio_unregister
.len
;
1520 if (!mmget_not_zero(mm
))
1523 mmap_write_lock(mm
);
1524 vma
= find_vma_prev(mm
, start
, &prev
);
1528 /* check that there's at least one vma in the range */
1530 if (vma
->vm_start
>= end
)
1534 * If the first vma contains huge pages, make sure start address
1535 * is aligned to huge page size.
1537 if (is_vm_hugetlb_page(vma
)) {
1538 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1540 if (start
& (vma_hpagesize
- 1))
1545 * Search for not compatible vmas.
1549 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1552 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1553 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1556 * Check not compatible vmas, not strictly required
1557 * here as not compatible vmas cannot have an
1558 * userfaultfd_ctx registered on them, but this
1559 * provides for more strict behavior to notice
1560 * unregistration errors.
1562 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1569 if (vma
->vm_start
< start
)
1576 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1579 * Nothing to do: this vma is already registered into this
1580 * userfaultfd and with the right tracking mode too.
1582 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1585 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1587 if (vma
->vm_start
> start
)
1588 start
= vma
->vm_start
;
1589 vma_end
= min(end
, vma
->vm_end
);
1591 if (userfaultfd_missing(vma
)) {
1593 * Wake any concurrent pending userfault while
1594 * we unregister, so they will not hang
1595 * permanently and it avoids userland to call
1596 * UFFDIO_WAKE explicitly.
1598 struct userfaultfd_wake_range range
;
1599 range
.start
= start
;
1600 range
.len
= vma_end
- start
;
1601 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1604 /* Reset ptes for the whole vma range if wr-protected */
1605 if (userfaultfd_wp(vma
))
1606 uffd_wp_range(mm
, vma
, start
, vma_end
- start
, false);
1608 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1609 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1610 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1612 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
1617 if (vma
->vm_start
< start
) {
1618 ret
= split_vma(mm
, vma
, start
, 1);
1622 if (vma
->vm_end
> end
) {
1623 ret
= split_vma(mm
, vma
, end
, 0);
1629 * In the vma_merge() successful mprotect-like case 8:
1630 * the next vma was merged into the current one and
1631 * the current one has not been updated yet.
1633 vma
->vm_flags
= new_flags
;
1634 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1638 start
= vma
->vm_end
;
1640 } while (vma
&& vma
->vm_start
< end
);
1642 mmap_write_unlock(mm
);
1649 * userfaultfd_wake may be used in combination with the
1650 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1652 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1656 struct uffdio_range uffdio_wake
;
1657 struct userfaultfd_wake_range range
;
1658 const void __user
*buf
= (void __user
*)arg
;
1661 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1664 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1668 range
.start
= uffdio_wake
.start
;
1669 range
.len
= uffdio_wake
.len
;
1672 * len == 0 means wake all and we don't want to wake all here,
1673 * so check it again to be sure.
1675 VM_BUG_ON(!range
.len
);
1677 wake_userfault(ctx
, &range
);
1684 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1688 struct uffdio_copy uffdio_copy
;
1689 struct uffdio_copy __user
*user_uffdio_copy
;
1690 struct userfaultfd_wake_range range
;
1692 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1695 if (atomic_read(&ctx
->mmap_changing
))
1699 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1700 /* don't copy "copy" last field */
1701 sizeof(uffdio_copy
)-sizeof(__s64
)))
1704 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1708 * double check for wraparound just in case. copy_from_user()
1709 * will later check uffdio_copy.src + uffdio_copy.len to fit
1710 * in the userland range.
1713 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1715 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1717 if (mmget_not_zero(ctx
->mm
)) {
1718 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1719 uffdio_copy
.len
, &ctx
->mmap_changing
,
1725 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1730 /* len == 0 would wake all */
1732 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1733 range
.start
= uffdio_copy
.dst
;
1734 wake_userfault(ctx
, &range
);
1736 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1741 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1745 struct uffdio_zeropage uffdio_zeropage
;
1746 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1747 struct userfaultfd_wake_range range
;
1749 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1752 if (atomic_read(&ctx
->mmap_changing
))
1756 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1757 /* don't copy "zeropage" last field */
1758 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1761 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1762 uffdio_zeropage
.range
.len
);
1766 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1769 if (mmget_not_zero(ctx
->mm
)) {
1770 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1771 uffdio_zeropage
.range
.len
,
1772 &ctx
->mmap_changing
);
1777 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1781 /* len == 0 would wake all */
1784 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1785 range
.start
= uffdio_zeropage
.range
.start
;
1786 wake_userfault(ctx
, &range
);
1788 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1793 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1797 struct uffdio_writeprotect uffdio_wp
;
1798 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1799 struct userfaultfd_wake_range range
;
1800 bool mode_wp
, mode_dontwake
;
1802 if (atomic_read(&ctx
->mmap_changing
))
1805 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1807 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1808 sizeof(struct uffdio_writeprotect
)))
1811 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1812 uffdio_wp
.range
.len
);
1816 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1817 UFFDIO_WRITEPROTECT_MODE_WP
))
1820 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1821 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1823 if (mode_wp
&& mode_dontwake
)
1826 if (mmget_not_zero(ctx
->mm
)) {
1827 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1828 uffdio_wp
.range
.len
, mode_wp
,
1829 &ctx
->mmap_changing
);
1838 if (!mode_wp
&& !mode_dontwake
) {
1839 range
.start
= uffdio_wp
.range
.start
;
1840 range
.len
= uffdio_wp
.range
.len
;
1841 wake_userfault(ctx
, &range
);
1846 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1849 struct uffdio_continue uffdio_continue
;
1850 struct uffdio_continue __user
*user_uffdio_continue
;
1851 struct userfaultfd_wake_range range
;
1853 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1856 if (atomic_read(&ctx
->mmap_changing
))
1860 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1861 /* don't copy the output fields */
1862 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1865 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1866 uffdio_continue
.range
.len
);
1871 /* double check for wraparound just in case. */
1872 if (uffdio_continue
.range
.start
+ uffdio_continue
.range
.len
<=
1873 uffdio_continue
.range
.start
) {
1876 if (uffdio_continue
.mode
& ~UFFDIO_CONTINUE_MODE_DONTWAKE
)
1879 if (mmget_not_zero(ctx
->mm
)) {
1880 ret
= mcopy_continue(ctx
->mm
, uffdio_continue
.range
.start
,
1881 uffdio_continue
.range
.len
,
1882 &ctx
->mmap_changing
);
1888 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1893 /* len == 0 would wake all */
1896 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1897 range
.start
= uffdio_continue
.range
.start
;
1898 wake_userfault(ctx
, &range
);
1900 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1906 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1909 * For the current set of features the bits just coincide. Set
1910 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1912 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1916 * userland asks for a certain API version and we return which bits
1917 * and ioctl commands are implemented in this kernel for such API
1918 * version or -EINVAL if unknown.
1920 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1923 struct uffdio_api uffdio_api
;
1924 void __user
*buf
= (void __user
*)arg
;
1925 unsigned int ctx_features
;
1930 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1932 /* Ignore unsupported features (userspace built against newer kernel) */
1933 features
= uffdio_api
.features
& UFFD_API_FEATURES
;
1935 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1937 /* report all available features and ioctls to userland */
1938 uffdio_api
.features
= UFFD_API_FEATURES
;
1939 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1940 uffdio_api
.features
&=
1941 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
1943 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1944 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
1946 #ifndef CONFIG_PTE_MARKER_UFFD_WP
1947 uffdio_api
.features
&= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM
;
1949 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1951 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1954 /* only enable the requested features for this uffd context */
1955 ctx_features
= uffd_ctx_features(features
);
1957 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
1964 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1965 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1970 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1974 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1976 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
1981 ret
= userfaultfd_api(ctx
, arg
);
1983 case UFFDIO_REGISTER
:
1984 ret
= userfaultfd_register(ctx
, arg
);
1986 case UFFDIO_UNREGISTER
:
1987 ret
= userfaultfd_unregister(ctx
, arg
);
1990 ret
= userfaultfd_wake(ctx
, arg
);
1993 ret
= userfaultfd_copy(ctx
, arg
);
1995 case UFFDIO_ZEROPAGE
:
1996 ret
= userfaultfd_zeropage(ctx
, arg
);
1998 case UFFDIO_WRITEPROTECT
:
1999 ret
= userfaultfd_writeprotect(ctx
, arg
);
2001 case UFFDIO_CONTINUE
:
2002 ret
= userfaultfd_continue(ctx
, arg
);
2008 #ifdef CONFIG_PROC_FS
2009 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2011 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2012 wait_queue_entry_t
*wq
;
2013 unsigned long pending
= 0, total
= 0;
2015 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2016 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2020 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2023 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2026 * If more protocols will be added, there will be all shown
2027 * separated by a space. Like this:
2028 * protocols: aa:... bb:...
2030 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2031 pending
, total
, UFFD_API
, ctx
->features
,
2032 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2036 static const struct file_operations userfaultfd_fops
= {
2037 #ifdef CONFIG_PROC_FS
2038 .show_fdinfo
= userfaultfd_show_fdinfo
,
2040 .release
= userfaultfd_release
,
2041 .poll
= userfaultfd_poll
,
2042 .read
= userfaultfd_read
,
2043 .unlocked_ioctl
= userfaultfd_ioctl
,
2044 .compat_ioctl
= compat_ptr_ioctl
,
2045 .llseek
= noop_llseek
,
2048 static void init_once_userfaultfd_ctx(void *mem
)
2050 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2052 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2053 init_waitqueue_head(&ctx
->fault_wqh
);
2054 init_waitqueue_head(&ctx
->event_wqh
);
2055 init_waitqueue_head(&ctx
->fd_wqh
);
2056 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2059 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2061 struct userfaultfd_ctx
*ctx
;
2064 if (!sysctl_unprivileged_userfaultfd
&&
2065 (flags
& UFFD_USER_MODE_ONLY
) == 0 &&
2066 !capable(CAP_SYS_PTRACE
)) {
2067 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
2068 "sysctl knob to 1 if kernel faults must be handled "
2069 "without obtaining CAP_SYS_PTRACE capability\n");
2073 BUG_ON(!current
->mm
);
2075 /* Check the UFFD_* constants for consistency. */
2076 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2077 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2078 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2080 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2083 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2087 refcount_set(&ctx
->refcount
, 1);
2090 ctx
->released
= false;
2091 atomic_set(&ctx
->mmap_changing
, 0);
2092 ctx
->mm
= current
->mm
;
2093 /* prevent the mm struct to be freed */
2096 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, ctx
,
2097 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2100 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2105 static int __init
userfaultfd_init(void)
2107 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2108 sizeof(struct userfaultfd_ctx
),
2110 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2111 init_once_userfaultfd_ctx
);
2114 __initcall(userfaultfd_init
);