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
,
196 unsigned long reason
,
197 unsigned int features
)
201 msg
.event
= UFFD_EVENT_PAGEFAULT
;
203 if (!(features
& UFFD_FEATURE_EXACT_ADDRESS
))
204 address
&= PAGE_MASK
;
205 msg
.arg
.pagefault
.address
= address
;
207 * These flags indicate why the userfault occurred:
208 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
209 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
210 * - Neither of these flags being set indicates a MISSING fault.
212 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
213 * fault. Otherwise, it was a read fault.
215 if (flags
& FAULT_FLAG_WRITE
)
216 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
217 if (reason
& VM_UFFD_WP
)
218 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
219 if (reason
& VM_UFFD_MINOR
)
220 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_MINOR
;
221 if (features
& UFFD_FEATURE_THREAD_ID
)
222 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
226 #ifdef CONFIG_HUGETLB_PAGE
228 * Same functionality as userfaultfd_must_wait below with modifications for
231 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
232 struct vm_area_struct
*vma
,
233 unsigned long address
,
235 unsigned long reason
)
237 struct mm_struct
*mm
= ctx
->mm
;
241 mmap_assert_locked(mm
);
243 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
249 pte
= huge_ptep_get(ptep
);
252 * Lockless access: we're in a wait_event so it's ok if it
253 * changes under us. PTE markers should be handled the same as none
256 if (huge_pte_none_mostly(pte
))
258 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
264 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
265 struct vm_area_struct
*vma
,
266 unsigned long address
,
268 unsigned long reason
)
270 return false; /* should never get here */
272 #endif /* CONFIG_HUGETLB_PAGE */
275 * Verify the pagetables are still not ok after having reigstered into
276 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
277 * userfault that has already been resolved, if userfaultfd_read and
278 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
281 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
282 unsigned long address
,
284 unsigned long reason
)
286 struct mm_struct
*mm
= ctx
->mm
;
294 mmap_assert_locked(mm
);
296 pgd
= pgd_offset(mm
, address
);
297 if (!pgd_present(*pgd
))
299 p4d
= p4d_offset(pgd
, address
);
300 if (!p4d_present(*p4d
))
302 pud
= pud_offset(p4d
, address
);
303 if (!pud_present(*pud
))
305 pmd
= pmd_offset(pud
, address
);
307 * READ_ONCE must function as a barrier with narrower scope
308 * and it must be equivalent to:
309 * _pmd = *pmd; barrier();
311 * This is to deal with the instability (as in
312 * pmd_trans_unstable) of the pmd.
314 _pmd
= READ_ONCE(*pmd
);
319 if (!pmd_present(_pmd
))
322 if (pmd_trans_huge(_pmd
)) {
323 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
329 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
330 * and use the standard pte_offset_map() instead of parsing _pmd.
332 pte
= pte_offset_map(pmd
, address
);
334 * Lockless access: we're in a wait_event so it's ok if it
335 * changes under us. PTE markers should be handled the same as none
338 if (pte_none_mostly(*pte
))
340 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
348 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
350 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
351 return TASK_INTERRUPTIBLE
;
353 if (flags
& FAULT_FLAG_KILLABLE
)
354 return TASK_KILLABLE
;
356 return TASK_UNINTERRUPTIBLE
;
360 * The locking rules involved in returning VM_FAULT_RETRY depending on
361 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
362 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
363 * recommendation in __lock_page_or_retry is not an understatement.
365 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
366 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
369 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
370 * set, VM_FAULT_RETRY can still be returned if and only if there are
371 * fatal_signal_pending()s, and the mmap_lock must be released before
374 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
376 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
377 struct userfaultfd_ctx
*ctx
;
378 struct userfaultfd_wait_queue uwq
;
379 vm_fault_t ret
= VM_FAULT_SIGBUS
;
381 unsigned int blocking_state
;
384 * We don't do userfault handling for the final child pid update.
386 * We also don't do userfault handling during
387 * coredumping. hugetlbfs has the special
388 * follow_hugetlb_page() to skip missing pages in the
389 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
390 * the no_page_table() helper in follow_page_mask(), but the
391 * shmem_vm_ops->fault method is invoked even during
392 * coredumping without mmap_lock and it ends up here.
394 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
398 * Coredumping runs without mmap_lock so we can only check that
399 * the mmap_lock is held, if PF_DUMPCORE was not set.
401 mmap_assert_locked(mm
);
403 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
407 BUG_ON(ctx
->mm
!= mm
);
409 /* Any unrecognized flag is a bug. */
410 VM_BUG_ON(reason
& ~__VM_UFFD_FLAGS
);
411 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
412 VM_BUG_ON(!reason
|| (reason
& (reason
- 1)));
414 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
416 if ((vmf
->flags
& FAULT_FLAG_USER
) == 0 &&
417 ctx
->flags
& UFFD_USER_MODE_ONLY
) {
418 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
419 "sysctl knob to 1 if kernel faults must be handled "
420 "without obtaining CAP_SYS_PTRACE capability\n");
425 * If it's already released don't get it. This avoids to loop
426 * in __get_user_pages if userfaultfd_release waits on the
427 * caller of handle_userfault to release the mmap_lock.
429 if (unlikely(READ_ONCE(ctx
->released
))) {
431 * Don't return VM_FAULT_SIGBUS in this case, so a non
432 * cooperative manager can close the uffd after the
433 * last UFFDIO_COPY, without risking to trigger an
434 * involuntary SIGBUS if the process was starting the
435 * userfaultfd while the userfaultfd was still armed
436 * (but after the last UFFDIO_COPY). If the uffd
437 * wasn't already closed when the userfault reached
438 * this point, that would normally be solved by
439 * userfaultfd_must_wait returning 'false'.
441 * If we were to return VM_FAULT_SIGBUS here, the non
442 * cooperative manager would be instead forced to
443 * always call UFFDIO_UNREGISTER before it can safely
446 ret
= VM_FAULT_NOPAGE
;
451 * Check that we can return VM_FAULT_RETRY.
453 * NOTE: it should become possible to return VM_FAULT_RETRY
454 * even if FAULT_FLAG_TRIED is set without leading to gup()
455 * -EBUSY failures, if the userfaultfd is to be extended for
456 * VM_UFFD_WP tracking and we intend to arm the userfault
457 * without first stopping userland access to the memory. For
458 * VM_UFFD_MISSING userfaults this is enough for now.
460 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
462 * Validate the invariant that nowait must allow retry
463 * to be sure not to return SIGBUS erroneously on
464 * nowait invocations.
466 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
467 #ifdef CONFIG_DEBUG_VM
468 if (printk_ratelimit()) {
470 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
479 * Handle nowait, not much to do other than tell it to retry
482 ret
= VM_FAULT_RETRY
;
483 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
486 /* take the reference before dropping the mmap_lock */
487 userfaultfd_ctx_get(ctx
);
489 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
490 uwq
.wq
.private = current
;
491 uwq
.msg
= userfault_msg(vmf
->real_address
, vmf
->flags
, reason
,
496 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
498 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
500 * After the __add_wait_queue the uwq is visible to userland
501 * through poll/read().
503 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
505 * The smp_mb() after __set_current_state prevents the reads
506 * following the spin_unlock to happen before the list_add in
509 set_current_state(blocking_state
);
510 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
512 if (!is_vm_hugetlb_page(vmf
->vma
))
513 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
516 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
519 mmap_read_unlock(mm
);
521 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
522 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
526 __set_current_state(TASK_RUNNING
);
529 * Here we race with the list_del; list_add in
530 * userfaultfd_ctx_read(), however because we don't ever run
531 * list_del_init() to refile across the two lists, the prev
532 * and next pointers will never point to self. list_add also
533 * would never let any of the two pointers to point to
534 * self. So list_empty_careful won't risk to see both pointers
535 * pointing to self at any time during the list refile. The
536 * only case where list_del_init() is called is the full
537 * removal in the wake function and there we don't re-list_add
538 * and it's fine not to block on the spinlock. The uwq on this
539 * kernel stack can be released after the list_del_init.
541 if (!list_empty_careful(&uwq
.wq
.entry
)) {
542 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
544 * No need of list_del_init(), the uwq on the stack
545 * will be freed shortly anyway.
547 list_del(&uwq
.wq
.entry
);
548 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
552 * ctx may go away after this if the userfault pseudo fd is
555 userfaultfd_ctx_put(ctx
);
561 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
562 struct userfaultfd_wait_queue
*ewq
)
564 struct userfaultfd_ctx
*release_new_ctx
;
566 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
570 init_waitqueue_entry(&ewq
->wq
, current
);
571 release_new_ctx
= NULL
;
573 spin_lock_irq(&ctx
->event_wqh
.lock
);
575 * After the __add_wait_queue the uwq is visible to userland
576 * through poll/read().
578 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
580 set_current_state(TASK_KILLABLE
);
581 if (ewq
->msg
.event
== 0)
583 if (READ_ONCE(ctx
->released
) ||
584 fatal_signal_pending(current
)) {
586 * &ewq->wq may be queued in fork_event, but
587 * __remove_wait_queue ignores the head
588 * parameter. It would be a problem if it
591 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
592 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
593 struct userfaultfd_ctx
*new;
595 new = (struct userfaultfd_ctx
*)
597 ewq
->msg
.arg
.reserved
.reserved1
;
598 release_new_ctx
= new;
603 spin_unlock_irq(&ctx
->event_wqh
.lock
);
605 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
608 spin_lock_irq(&ctx
->event_wqh
.lock
);
610 __set_current_state(TASK_RUNNING
);
611 spin_unlock_irq(&ctx
->event_wqh
.lock
);
613 if (release_new_ctx
) {
614 struct vm_area_struct
*vma
;
615 struct mm_struct
*mm
= release_new_ctx
->mm
;
617 /* the various vma->vm_userfaultfd_ctx still points to it */
619 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
620 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
621 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
622 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
624 mmap_write_unlock(mm
);
626 userfaultfd_ctx_put(release_new_ctx
);
630 * ctx may go away after this if the userfault pseudo fd is
634 atomic_dec(&ctx
->mmap_changing
);
635 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
636 userfaultfd_ctx_put(ctx
);
639 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
640 struct userfaultfd_wait_queue
*ewq
)
643 wake_up_locked(&ctx
->event_wqh
);
644 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
647 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
649 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
650 struct userfaultfd_fork_ctx
*fctx
;
652 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
653 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
654 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
655 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
659 list_for_each_entry(fctx
, fcs
, list
)
660 if (fctx
->orig
== octx
) {
666 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
670 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
676 refcount_set(&ctx
->refcount
, 1);
677 ctx
->flags
= octx
->flags
;
678 ctx
->features
= octx
->features
;
679 ctx
->released
= false;
680 atomic_set(&ctx
->mmap_changing
, 0);
681 ctx
->mm
= vma
->vm_mm
;
684 userfaultfd_ctx_get(octx
);
685 atomic_inc(&octx
->mmap_changing
);
688 list_add_tail(&fctx
->list
, fcs
);
691 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
695 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
697 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
698 struct userfaultfd_wait_queue ewq
;
702 ewq
.msg
.event
= UFFD_EVENT_FORK
;
703 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
705 userfaultfd_event_wait_completion(ctx
, &ewq
);
708 void dup_userfaultfd_complete(struct list_head
*fcs
)
710 struct userfaultfd_fork_ctx
*fctx
, *n
;
712 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
714 list_del(&fctx
->list
);
719 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
720 struct vm_userfaultfd_ctx
*vm_ctx
)
722 struct userfaultfd_ctx
*ctx
;
724 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
729 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
731 userfaultfd_ctx_get(ctx
);
732 atomic_inc(&ctx
->mmap_changing
);
734 /* Drop uffd context if remap feature not enabled */
735 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
736 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
740 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
741 unsigned long from
, unsigned long to
,
744 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
745 struct userfaultfd_wait_queue ewq
;
750 if (to
& ~PAGE_MASK
) {
751 userfaultfd_ctx_put(ctx
);
757 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
758 ewq
.msg
.arg
.remap
.from
= from
;
759 ewq
.msg
.arg
.remap
.to
= to
;
760 ewq
.msg
.arg
.remap
.len
= len
;
762 userfaultfd_event_wait_completion(ctx
, &ewq
);
765 bool userfaultfd_remove(struct vm_area_struct
*vma
,
766 unsigned long start
, unsigned long end
)
768 struct mm_struct
*mm
= vma
->vm_mm
;
769 struct userfaultfd_ctx
*ctx
;
770 struct userfaultfd_wait_queue ewq
;
772 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
773 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
776 userfaultfd_ctx_get(ctx
);
777 atomic_inc(&ctx
->mmap_changing
);
778 mmap_read_unlock(mm
);
782 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
783 ewq
.msg
.arg
.remove
.start
= start
;
784 ewq
.msg
.arg
.remove
.end
= end
;
786 userfaultfd_event_wait_completion(ctx
, &ewq
);
791 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
792 unsigned long start
, unsigned long end
)
794 struct userfaultfd_unmap_ctx
*unmap_ctx
;
796 list_for_each_entry(unmap_ctx
, unmaps
, list
)
797 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
798 unmap_ctx
->end
== end
)
804 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
805 unsigned long start
, unsigned long end
,
806 struct list_head
*unmaps
)
808 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
809 struct userfaultfd_unmap_ctx
*unmap_ctx
;
810 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
812 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
813 has_unmap_ctx(ctx
, unmaps
, start
, end
))
816 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
820 userfaultfd_ctx_get(ctx
);
821 atomic_inc(&ctx
->mmap_changing
);
822 unmap_ctx
->ctx
= ctx
;
823 unmap_ctx
->start
= start
;
824 unmap_ctx
->end
= end
;
825 list_add_tail(&unmap_ctx
->list
, unmaps
);
831 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
833 struct userfaultfd_unmap_ctx
*ctx
, *n
;
834 struct userfaultfd_wait_queue ewq
;
836 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
839 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
840 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
841 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
843 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
845 list_del(&ctx
->list
);
850 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
852 struct userfaultfd_ctx
*ctx
= file
->private_data
;
853 struct mm_struct
*mm
= ctx
->mm
;
854 struct vm_area_struct
*vma
, *prev
;
855 /* len == 0 means wake all */
856 struct userfaultfd_wake_range range
= { .len
= 0, };
857 unsigned long new_flags
;
859 WRITE_ONCE(ctx
->released
, true);
861 if (!mmget_not_zero(mm
))
865 * Flush page faults out of all CPUs. NOTE: all page faults
866 * must be retried without returning VM_FAULT_SIGBUS if
867 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
868 * changes while handle_userfault released the mmap_lock. So
869 * it's critical that released is set to true (above), before
870 * taking the mmap_lock for writing.
874 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
876 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
877 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
878 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
882 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
883 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
884 new_flags
, vma
->anon_vma
,
885 vma
->vm_file
, vma
->vm_pgoff
,
887 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
892 vma
->vm_flags
= new_flags
;
893 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
895 mmap_write_unlock(mm
);
899 * After no new page faults can wait on this fault_*wqh, flush
900 * the last page faults that may have been already waiting on
903 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
904 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
905 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
906 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
908 /* Flush pending events that may still wait on event_wqh */
909 wake_up_all(&ctx
->event_wqh
);
911 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
912 userfaultfd_ctx_put(ctx
);
916 /* fault_pending_wqh.lock must be hold by the caller */
917 static inline struct userfaultfd_wait_queue
*find_userfault_in(
918 wait_queue_head_t
*wqh
)
920 wait_queue_entry_t
*wq
;
921 struct userfaultfd_wait_queue
*uwq
;
923 lockdep_assert_held(&wqh
->lock
);
926 if (!waitqueue_active(wqh
))
928 /* walk in reverse to provide FIFO behavior to read userfaults */
929 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
930 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
935 static inline struct userfaultfd_wait_queue
*find_userfault(
936 struct userfaultfd_ctx
*ctx
)
938 return find_userfault_in(&ctx
->fault_pending_wqh
);
941 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
942 struct userfaultfd_ctx
*ctx
)
944 return find_userfault_in(&ctx
->event_wqh
);
947 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
949 struct userfaultfd_ctx
*ctx
= file
->private_data
;
952 poll_wait(file
, &ctx
->fd_wqh
, wait
);
954 if (!userfaultfd_is_initialized(ctx
))
958 * poll() never guarantees that read won't block.
959 * userfaults can be waken before they're read().
961 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
964 * lockless access to see if there are pending faults
965 * __pollwait last action is the add_wait_queue but
966 * the spin_unlock would allow the waitqueue_active to
967 * pass above the actual list_add inside
968 * add_wait_queue critical section. So use a full
969 * memory barrier to serialize the list_add write of
970 * add_wait_queue() with the waitqueue_active read
975 if (waitqueue_active(&ctx
->fault_pending_wqh
))
977 else if (waitqueue_active(&ctx
->event_wqh
))
983 static const struct file_operations userfaultfd_fops
;
985 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
987 struct uffd_msg
*msg
)
991 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, new,
992 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
996 msg
->arg
.reserved
.reserved1
= 0;
997 msg
->arg
.fork
.ufd
= fd
;
1001 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1002 struct uffd_msg
*msg
, struct inode
*inode
)
1005 DECLARE_WAITQUEUE(wait
, current
);
1006 struct userfaultfd_wait_queue
*uwq
;
1008 * Handling fork event requires sleeping operations, so
1009 * we drop the event_wqh lock, then do these ops, then
1010 * lock it back and wake up the waiter. While the lock is
1011 * dropped the ewq may go away so we keep track of it
1014 LIST_HEAD(fork_event
);
1015 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1017 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1018 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1019 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1021 set_current_state(TASK_INTERRUPTIBLE
);
1022 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1023 uwq
= find_userfault(ctx
);
1026 * Use a seqcount to repeat the lockless check
1027 * in wake_userfault() to avoid missing
1028 * wakeups because during the refile both
1029 * waitqueue could become empty if this is the
1032 write_seqcount_begin(&ctx
->refile_seq
);
1035 * The fault_pending_wqh.lock prevents the uwq
1036 * to disappear from under us.
1038 * Refile this userfault from
1039 * fault_pending_wqh to fault_wqh, it's not
1040 * pending anymore after we read it.
1042 * Use list_del() by hand (as
1043 * userfaultfd_wake_function also uses
1044 * list_del_init() by hand) to be sure nobody
1045 * changes __remove_wait_queue() to use
1046 * list_del_init() in turn breaking the
1047 * !list_empty_careful() check in
1048 * handle_userfault(). The uwq->wq.head list
1049 * must never be empty at any time during the
1050 * refile, or the waitqueue could disappear
1051 * from under us. The "wait_queue_head_t"
1052 * parameter of __remove_wait_queue() is unused
1055 list_del(&uwq
->wq
.entry
);
1056 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1058 write_seqcount_end(&ctx
->refile_seq
);
1060 /* careful to always initialize msg if ret == 0 */
1062 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1066 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1068 spin_lock(&ctx
->event_wqh
.lock
);
1069 uwq
= find_userfault_evt(ctx
);
1073 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1074 fork_nctx
= (struct userfaultfd_ctx
*)
1076 uwq
->msg
.arg
.reserved
.reserved1
;
1077 list_move(&uwq
->wq
.entry
, &fork_event
);
1079 * fork_nctx can be freed as soon as
1080 * we drop the lock, unless we take a
1083 userfaultfd_ctx_get(fork_nctx
);
1084 spin_unlock(&ctx
->event_wqh
.lock
);
1089 userfaultfd_event_complete(ctx
, uwq
);
1090 spin_unlock(&ctx
->event_wqh
.lock
);
1094 spin_unlock(&ctx
->event_wqh
.lock
);
1096 if (signal_pending(current
)) {
1104 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1106 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1108 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1109 __set_current_state(TASK_RUNNING
);
1110 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1112 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1113 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1114 spin_lock_irq(&ctx
->event_wqh
.lock
);
1115 if (!list_empty(&fork_event
)) {
1117 * The fork thread didn't abort, so we can
1118 * drop the temporary refcount.
1120 userfaultfd_ctx_put(fork_nctx
);
1122 uwq
= list_first_entry(&fork_event
,
1126 * If fork_event list wasn't empty and in turn
1127 * the event wasn't already released by fork
1128 * (the event is allocated on fork kernel
1129 * stack), put the event back to its place in
1130 * the event_wq. fork_event head will be freed
1131 * as soon as we return so the event cannot
1132 * stay queued there no matter the current
1135 list_del(&uwq
->wq
.entry
);
1136 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1139 * Leave the event in the waitqueue and report
1140 * error to userland if we failed to resolve
1141 * the userfault fork.
1144 userfaultfd_event_complete(ctx
, uwq
);
1147 * Here the fork thread aborted and the
1148 * refcount from the fork thread on fork_nctx
1149 * has already been released. We still hold
1150 * the reference we took before releasing the
1151 * lock above. If resolve_userfault_fork
1152 * failed we've to drop it because the
1153 * fork_nctx has to be freed in such case. If
1154 * it succeeded we'll hold it because the new
1155 * uffd references it.
1158 userfaultfd_ctx_put(fork_nctx
);
1160 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1166 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1167 size_t count
, loff_t
*ppos
)
1169 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1170 ssize_t _ret
, ret
= 0;
1171 struct uffd_msg msg
;
1172 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1173 struct inode
*inode
= file_inode(file
);
1175 if (!userfaultfd_is_initialized(ctx
))
1179 if (count
< sizeof(msg
))
1180 return ret
? ret
: -EINVAL
;
1181 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1183 return ret
? ret
: _ret
;
1184 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1185 return ret
? ret
: -EFAULT
;
1188 count
-= sizeof(msg
);
1190 * Allow to read more than one fault at time but only
1191 * block if waiting for the very first one.
1193 no_wait
= O_NONBLOCK
;
1197 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1198 struct userfaultfd_wake_range
*range
)
1200 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1201 /* wake all in the range and autoremove */
1202 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1203 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1205 if (waitqueue_active(&ctx
->fault_wqh
))
1206 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1207 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1210 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1211 struct userfaultfd_wake_range
*range
)
1217 * To be sure waitqueue_active() is not reordered by the CPU
1218 * before the pagetable update, use an explicit SMP memory
1219 * barrier here. PT lock release or mmap_read_unlock(mm) still
1220 * have release semantics that can allow the
1221 * waitqueue_active() to be reordered before the pte update.
1226 * Use waitqueue_active because it's very frequent to
1227 * change the address space atomically even if there are no
1228 * userfaults yet. So we take the spinlock only when we're
1229 * sure we've userfaults to wake.
1232 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1233 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1234 waitqueue_active(&ctx
->fault_wqh
);
1236 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1238 __wake_userfault(ctx
, range
);
1241 static __always_inline
int validate_range(struct mm_struct
*mm
,
1242 __u64 start
, __u64 len
)
1244 __u64 task_size
= mm
->task_size
;
1246 if (start
& ~PAGE_MASK
)
1248 if (len
& ~PAGE_MASK
)
1252 if (start
< mmap_min_addr
)
1254 if (start
>= task_size
)
1256 if (len
> task_size
- start
)
1261 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1264 struct mm_struct
*mm
= ctx
->mm
;
1265 struct vm_area_struct
*vma
, *prev
, *cur
;
1267 struct uffdio_register uffdio_register
;
1268 struct uffdio_register __user
*user_uffdio_register
;
1269 unsigned long vm_flags
, new_flags
;
1272 unsigned long start
, end
, vma_end
;
1274 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1277 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1278 sizeof(uffdio_register
)-sizeof(__u64
)))
1282 if (!uffdio_register
.mode
)
1284 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1287 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1288 vm_flags
|= VM_UFFD_MISSING
;
1289 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1290 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1293 vm_flags
|= VM_UFFD_WP
;
1295 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1296 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1299 vm_flags
|= VM_UFFD_MINOR
;
1302 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1303 uffdio_register
.range
.len
);
1307 start
= uffdio_register
.range
.start
;
1308 end
= start
+ uffdio_register
.range
.len
;
1311 if (!mmget_not_zero(mm
))
1314 mmap_write_lock(mm
);
1315 vma
= find_vma_prev(mm
, start
, &prev
);
1319 /* check that there's at least one vma in the range */
1321 if (vma
->vm_start
>= end
)
1325 * If the first vma contains huge pages, make sure start address
1326 * is aligned to huge page size.
1328 if (is_vm_hugetlb_page(vma
)) {
1329 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1331 if (start
& (vma_hpagesize
- 1))
1336 * Search for not compatible vmas.
1339 basic_ioctls
= false;
1340 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1343 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1344 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1346 /* check not compatible vmas */
1348 if (!vma_can_userfault(cur
, vm_flags
))
1352 * UFFDIO_COPY will fill file holes even without
1353 * PROT_WRITE. This check enforces that if this is a
1354 * MAP_SHARED, the process has write permission to the backing
1355 * file. If VM_MAYWRITE is set it also enforces that on a
1356 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1357 * F_WRITE_SEAL can be taken until the vma is destroyed.
1360 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1364 * If this vma contains ending address, and huge pages
1367 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1368 end
> cur
->vm_start
) {
1369 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1373 if (end
& (vma_hpagesize
- 1))
1376 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1380 * Check that this vma isn't already owned by a
1381 * different userfaultfd. We can't allow more than one
1382 * userfaultfd to own a single vma simultaneously or we
1383 * wouldn't know which one to deliver the userfaults to.
1386 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1387 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1391 * Note vmas containing huge pages
1393 if (is_vm_hugetlb_page(cur
))
1394 basic_ioctls
= true;
1400 if (vma
->vm_start
< start
)
1407 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1408 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1409 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1410 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1413 * Nothing to do: this vma is already registered into this
1414 * userfaultfd and with the right tracking mode too.
1416 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1417 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1420 if (vma
->vm_start
> start
)
1421 start
= vma
->vm_start
;
1422 vma_end
= min(end
, vma
->vm_end
);
1424 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1425 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1426 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1428 ((struct vm_userfaultfd_ctx
){ ctx
}),
1429 anon_vma_name(vma
));
1434 if (vma
->vm_start
< start
) {
1435 ret
= split_vma(mm
, vma
, start
, 1);
1439 if (vma
->vm_end
> end
) {
1440 ret
= split_vma(mm
, vma
, end
, 0);
1446 * In the vma_merge() successful mprotect-like case 8:
1447 * the next vma was merged into the current one and
1448 * the current one has not been updated yet.
1450 vma
->vm_flags
= new_flags
;
1451 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1453 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1454 hugetlb_unshare_all_pmds(vma
);
1458 start
= vma
->vm_end
;
1460 } while (vma
&& vma
->vm_start
< end
);
1462 mmap_write_unlock(mm
);
1467 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1468 UFFD_API_RANGE_IOCTLS
;
1471 * Declare the WP ioctl only if the WP mode is
1472 * specified and all checks passed with the range
1474 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1475 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1477 /* CONTINUE ioctl is only supported for MINOR ranges. */
1478 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
))
1479 ioctls_out
&= ~((__u64
)1 << _UFFDIO_CONTINUE
);
1482 * Now that we scanned all vmas we can already tell
1483 * userland which ioctls methods are guaranteed to
1484 * succeed on this range.
1486 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1493 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1496 struct mm_struct
*mm
= ctx
->mm
;
1497 struct vm_area_struct
*vma
, *prev
, *cur
;
1499 struct uffdio_range uffdio_unregister
;
1500 unsigned long new_flags
;
1502 unsigned long start
, end
, vma_end
;
1503 const void __user
*buf
= (void __user
*)arg
;
1506 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1509 ret
= validate_range(mm
, uffdio_unregister
.start
,
1510 uffdio_unregister
.len
);
1514 start
= uffdio_unregister
.start
;
1515 end
= start
+ uffdio_unregister
.len
;
1518 if (!mmget_not_zero(mm
))
1521 mmap_write_lock(mm
);
1522 vma
= find_vma_prev(mm
, start
, &prev
);
1526 /* check that there's at least one vma in the range */
1528 if (vma
->vm_start
>= end
)
1532 * If the first vma contains huge pages, make sure start address
1533 * is aligned to huge page size.
1535 if (is_vm_hugetlb_page(vma
)) {
1536 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1538 if (start
& (vma_hpagesize
- 1))
1543 * Search for not compatible vmas.
1547 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1550 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1551 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1554 * Check not compatible vmas, not strictly required
1555 * here as not compatible vmas cannot have an
1556 * userfaultfd_ctx registered on them, but this
1557 * provides for more strict behavior to notice
1558 * unregistration errors.
1560 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1567 if (vma
->vm_start
< start
)
1574 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1577 * Nothing to do: this vma is already registered into this
1578 * userfaultfd and with the right tracking mode too.
1580 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1583 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1585 if (vma
->vm_start
> start
)
1586 start
= vma
->vm_start
;
1587 vma_end
= min(end
, vma
->vm_end
);
1589 if (userfaultfd_missing(vma
)) {
1591 * Wake any concurrent pending userfault while
1592 * we unregister, so they will not hang
1593 * permanently and it avoids userland to call
1594 * UFFDIO_WAKE explicitly.
1596 struct userfaultfd_wake_range range
;
1597 range
.start
= start
;
1598 range
.len
= vma_end
- start
;
1599 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1602 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1603 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1604 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1606 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
1611 if (vma
->vm_start
< start
) {
1612 ret
= split_vma(mm
, vma
, start
, 1);
1616 if (vma
->vm_end
> end
) {
1617 ret
= split_vma(mm
, vma
, end
, 0);
1623 * In the vma_merge() successful mprotect-like case 8:
1624 * the next vma was merged into the current one and
1625 * the current one has not been updated yet.
1627 vma
->vm_flags
= new_flags
;
1628 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1632 start
= vma
->vm_end
;
1634 } while (vma
&& vma
->vm_start
< end
);
1636 mmap_write_unlock(mm
);
1643 * userfaultfd_wake may be used in combination with the
1644 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1646 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1650 struct uffdio_range uffdio_wake
;
1651 struct userfaultfd_wake_range range
;
1652 const void __user
*buf
= (void __user
*)arg
;
1655 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1658 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1662 range
.start
= uffdio_wake
.start
;
1663 range
.len
= uffdio_wake
.len
;
1666 * len == 0 means wake all and we don't want to wake all here,
1667 * so check it again to be sure.
1669 VM_BUG_ON(!range
.len
);
1671 wake_userfault(ctx
, &range
);
1678 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1682 struct uffdio_copy uffdio_copy
;
1683 struct uffdio_copy __user
*user_uffdio_copy
;
1684 struct userfaultfd_wake_range range
;
1686 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1689 if (atomic_read(&ctx
->mmap_changing
))
1693 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1694 /* don't copy "copy" last field */
1695 sizeof(uffdio_copy
)-sizeof(__s64
)))
1698 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1702 * double check for wraparound just in case. copy_from_user()
1703 * will later check uffdio_copy.src + uffdio_copy.len to fit
1704 * in the userland range.
1707 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1709 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1711 if (mmget_not_zero(ctx
->mm
)) {
1712 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1713 uffdio_copy
.len
, &ctx
->mmap_changing
,
1719 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1724 /* len == 0 would wake all */
1726 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1727 range
.start
= uffdio_copy
.dst
;
1728 wake_userfault(ctx
, &range
);
1730 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1735 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1739 struct uffdio_zeropage uffdio_zeropage
;
1740 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1741 struct userfaultfd_wake_range range
;
1743 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1746 if (atomic_read(&ctx
->mmap_changing
))
1750 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1751 /* don't copy "zeropage" last field */
1752 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1755 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1756 uffdio_zeropage
.range
.len
);
1760 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1763 if (mmget_not_zero(ctx
->mm
)) {
1764 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1765 uffdio_zeropage
.range
.len
,
1766 &ctx
->mmap_changing
);
1771 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1775 /* len == 0 would wake all */
1778 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1779 range
.start
= uffdio_zeropage
.range
.start
;
1780 wake_userfault(ctx
, &range
);
1782 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1787 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1791 struct uffdio_writeprotect uffdio_wp
;
1792 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1793 struct userfaultfd_wake_range range
;
1794 bool mode_wp
, mode_dontwake
;
1796 if (atomic_read(&ctx
->mmap_changing
))
1799 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1801 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1802 sizeof(struct uffdio_writeprotect
)))
1805 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1806 uffdio_wp
.range
.len
);
1810 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1811 UFFDIO_WRITEPROTECT_MODE_WP
))
1814 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1815 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1817 if (mode_wp
&& mode_dontwake
)
1820 if (mmget_not_zero(ctx
->mm
)) {
1821 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1822 uffdio_wp
.range
.len
, mode_wp
,
1823 &ctx
->mmap_changing
);
1832 if (!mode_wp
&& !mode_dontwake
) {
1833 range
.start
= uffdio_wp
.range
.start
;
1834 range
.len
= uffdio_wp
.range
.len
;
1835 wake_userfault(ctx
, &range
);
1840 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1843 struct uffdio_continue uffdio_continue
;
1844 struct uffdio_continue __user
*user_uffdio_continue
;
1845 struct userfaultfd_wake_range range
;
1847 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1850 if (atomic_read(&ctx
->mmap_changing
))
1854 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1855 /* don't copy the output fields */
1856 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1859 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1860 uffdio_continue
.range
.len
);
1865 /* double check for wraparound just in case. */
1866 if (uffdio_continue
.range
.start
+ uffdio_continue
.range
.len
<=
1867 uffdio_continue
.range
.start
) {
1870 if (uffdio_continue
.mode
& ~UFFDIO_CONTINUE_MODE_DONTWAKE
)
1873 if (mmget_not_zero(ctx
->mm
)) {
1874 ret
= mcopy_continue(ctx
->mm
, uffdio_continue
.range
.start
,
1875 uffdio_continue
.range
.len
,
1876 &ctx
->mmap_changing
);
1882 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1887 /* len == 0 would wake all */
1890 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1891 range
.start
= uffdio_continue
.range
.start
;
1892 wake_userfault(ctx
, &range
);
1894 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1900 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1903 * For the current set of features the bits just coincide. Set
1904 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1906 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1910 * userland asks for a certain API version and we return which bits
1911 * and ioctl commands are implemented in this kernel for such API
1912 * version or -EINVAL if unknown.
1914 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1917 struct uffdio_api uffdio_api
;
1918 void __user
*buf
= (void __user
*)arg
;
1919 unsigned int ctx_features
;
1924 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1926 features
= uffdio_api
.features
;
1928 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
1931 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1933 /* report all available features and ioctls to userland */
1934 uffdio_api
.features
= UFFD_API_FEATURES
;
1935 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1936 uffdio_api
.features
&=
1937 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
1939 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1940 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
1942 #ifndef CONFIG_PTE_MARKER_UFFD_WP
1943 uffdio_api
.features
&= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM
;
1945 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1947 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1950 /* only enable the requested features for this uffd context */
1951 ctx_features
= uffd_ctx_features(features
);
1953 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
1960 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1961 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1966 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1970 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1972 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
1977 ret
= userfaultfd_api(ctx
, arg
);
1979 case UFFDIO_REGISTER
:
1980 ret
= userfaultfd_register(ctx
, arg
);
1982 case UFFDIO_UNREGISTER
:
1983 ret
= userfaultfd_unregister(ctx
, arg
);
1986 ret
= userfaultfd_wake(ctx
, arg
);
1989 ret
= userfaultfd_copy(ctx
, arg
);
1991 case UFFDIO_ZEROPAGE
:
1992 ret
= userfaultfd_zeropage(ctx
, arg
);
1994 case UFFDIO_WRITEPROTECT
:
1995 ret
= userfaultfd_writeprotect(ctx
, arg
);
1997 case UFFDIO_CONTINUE
:
1998 ret
= userfaultfd_continue(ctx
, arg
);
2004 #ifdef CONFIG_PROC_FS
2005 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2007 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2008 wait_queue_entry_t
*wq
;
2009 unsigned long pending
= 0, total
= 0;
2011 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2012 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2016 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2019 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2022 * If more protocols will be added, there will be all shown
2023 * separated by a space. Like this:
2024 * protocols: aa:... bb:...
2026 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2027 pending
, total
, UFFD_API
, ctx
->features
,
2028 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2032 static const struct file_operations userfaultfd_fops
= {
2033 #ifdef CONFIG_PROC_FS
2034 .show_fdinfo
= userfaultfd_show_fdinfo
,
2036 .release
= userfaultfd_release
,
2037 .poll
= userfaultfd_poll
,
2038 .read
= userfaultfd_read
,
2039 .unlocked_ioctl
= userfaultfd_ioctl
,
2040 .compat_ioctl
= compat_ptr_ioctl
,
2041 .llseek
= noop_llseek
,
2044 static void init_once_userfaultfd_ctx(void *mem
)
2046 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2048 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2049 init_waitqueue_head(&ctx
->fault_wqh
);
2050 init_waitqueue_head(&ctx
->event_wqh
);
2051 init_waitqueue_head(&ctx
->fd_wqh
);
2052 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2055 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2057 struct userfaultfd_ctx
*ctx
;
2060 if (!sysctl_unprivileged_userfaultfd
&&
2061 (flags
& UFFD_USER_MODE_ONLY
) == 0 &&
2062 !capable(CAP_SYS_PTRACE
)) {
2063 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
2064 "sysctl knob to 1 if kernel faults must be handled "
2065 "without obtaining CAP_SYS_PTRACE capability\n");
2069 BUG_ON(!current
->mm
);
2071 /* Check the UFFD_* constants for consistency. */
2072 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2073 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2074 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2076 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2079 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2083 refcount_set(&ctx
->refcount
, 1);
2086 ctx
->released
= false;
2087 atomic_set(&ctx
->mmap_changing
, 0);
2088 ctx
->mm
= current
->mm
;
2089 /* prevent the mm struct to be freed */
2092 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, ctx
,
2093 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2096 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2101 static int __init
userfaultfd_init(void)
2103 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2104 sizeof(struct userfaultfd_ctx
),
2106 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2107 init_once_userfaultfd_ctx
);
2110 __initcall(userfaultfd_init
);