1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
6 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc.
9 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing).
13 #include <linux/list.h>
14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/mm.h>
18 #include <linux/mmu_notifier.h>
19 #include <linux/poll.h>
20 #include <linux/slab.h>
21 #include <linux/seq_file.h>
22 #include <linux/file.h>
23 #include <linux/bug.h>
24 #include <linux/anon_inodes.h>
25 #include <linux/syscalls.h>
26 #include <linux/userfaultfd_k.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ioctl.h>
29 #include <linux/security.h>
30 #include <linux/hugetlb.h>
32 int sysctl_unprivileged_userfaultfd __read_mostly
;
34 static struct kmem_cache
*userfaultfd_ctx_cachep __read_mostly
;
36 enum userfaultfd_state
{
42 * Start with fault_pending_wqh and fault_wqh so they're more likely
43 * to be in the same cacheline.
47 * fault_pending_wqh.lock
51 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
52 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
53 * also taken in IRQ context.
55 struct userfaultfd_ctx
{
56 /* waitqueue head for the pending (i.e. not read) userfaults */
57 wait_queue_head_t fault_pending_wqh
;
58 /* waitqueue head for the userfaults */
59 wait_queue_head_t fault_wqh
;
60 /* waitqueue head for the pseudo fd to wakeup poll/read */
61 wait_queue_head_t fd_wqh
;
62 /* waitqueue head for events */
63 wait_queue_head_t event_wqh
;
64 /* a refile sequence protected by fault_pending_wqh lock */
65 seqcount_spinlock_t refile_seq
;
66 /* pseudo fd refcounting */
68 /* userfaultfd syscall flags */
70 /* features requested from the userspace */
71 unsigned int features
;
73 enum userfaultfd_state state
;
76 /* memory mappings are changing because of non-cooperative event */
78 /* mm with one ore more vmas attached to this userfaultfd_ctx */
82 struct userfaultfd_fork_ctx
{
83 struct userfaultfd_ctx
*orig
;
84 struct userfaultfd_ctx
*new;
85 struct list_head list
;
88 struct userfaultfd_unmap_ctx
{
89 struct userfaultfd_ctx
*ctx
;
92 struct list_head list
;
95 struct userfaultfd_wait_queue
{
97 wait_queue_entry_t wq
;
98 struct userfaultfd_ctx
*ctx
;
102 struct userfaultfd_wake_range
{
107 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
108 int wake_flags
, void *key
)
110 struct userfaultfd_wake_range
*range
= key
;
112 struct userfaultfd_wait_queue
*uwq
;
113 unsigned long start
, len
;
115 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
117 /* len == 0 means wake all */
118 start
= range
->start
;
120 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
121 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
123 WRITE_ONCE(uwq
->waken
, true);
125 * The Program-Order guarantees provided by the scheduler
126 * ensure uwq->waken is visible before the task is woken.
128 ret
= wake_up_state(wq
->private, mode
);
131 * Wake only once, autoremove behavior.
133 * After the effect of list_del_init is visible to the other
134 * CPUs, the waitqueue may disappear from under us, see the
135 * !list_empty_careful() in handle_userfault().
137 * try_to_wake_up() has an implicit smp_mb(), and the
138 * wq->private is read before calling the extern function
139 * "wake_up_state" (which in turns calls try_to_wake_up).
141 list_del_init(&wq
->entry
);
148 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
150 * @ctx: [in] Pointer to the userfaultfd context.
152 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
154 refcount_inc(&ctx
->refcount
);
158 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
160 * @ctx: [in] Pointer to userfaultfd context.
162 * The userfaultfd context reference must have been previously acquired either
163 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
165 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
167 if (refcount_dec_and_test(&ctx
->refcount
)) {
168 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
169 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
170 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
171 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
172 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
173 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
174 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
175 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
177 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
181 static inline void msg_init(struct uffd_msg
*msg
)
183 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
185 * Must use memset to zero out the paddings or kernel data is
186 * leaked to userland.
188 memset(msg
, 0, sizeof(struct uffd_msg
));
191 static inline struct uffd_msg
userfault_msg(unsigned long address
,
193 unsigned long reason
,
194 unsigned int features
)
198 msg
.event
= UFFD_EVENT_PAGEFAULT
;
199 msg
.arg
.pagefault
.address
= address
;
201 * These flags indicate why the userfault occurred:
202 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
203 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
204 * - Neither of these flags being set indicates a MISSING fault.
206 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
207 * fault. Otherwise, it was a read fault.
209 if (flags
& FAULT_FLAG_WRITE
)
210 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
211 if (reason
& VM_UFFD_WP
)
212 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
213 if (reason
& VM_UFFD_MINOR
)
214 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_MINOR
;
215 if (features
& UFFD_FEATURE_THREAD_ID
)
216 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
220 #ifdef CONFIG_HUGETLB_PAGE
222 * Same functionality as userfaultfd_must_wait below with modifications for
225 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
226 struct vm_area_struct
*vma
,
227 unsigned long address
,
229 unsigned long reason
)
231 struct mm_struct
*mm
= ctx
->mm
;
235 mmap_assert_locked(mm
);
237 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
243 pte
= huge_ptep_get(ptep
);
246 * Lockless access: we're in a wait_event so it's ok if it
249 if (huge_pte_none(pte
))
251 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
257 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
258 struct vm_area_struct
*vma
,
259 unsigned long address
,
261 unsigned long reason
)
263 return false; /* should never get here */
265 #endif /* CONFIG_HUGETLB_PAGE */
268 * Verify the pagetables are still not ok after having reigstered into
269 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
270 * userfault that has already been resolved, if userfaultfd_read and
271 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
274 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
275 unsigned long address
,
277 unsigned long reason
)
279 struct mm_struct
*mm
= ctx
->mm
;
287 mmap_assert_locked(mm
);
289 pgd
= pgd_offset(mm
, address
);
290 if (!pgd_present(*pgd
))
292 p4d
= p4d_offset(pgd
, address
);
293 if (!p4d_present(*p4d
))
295 pud
= pud_offset(p4d
, address
);
296 if (!pud_present(*pud
))
298 pmd
= pmd_offset(pud
, address
);
300 * READ_ONCE must function as a barrier with narrower scope
301 * and it must be equivalent to:
302 * _pmd = *pmd; barrier();
304 * This is to deal with the instability (as in
305 * pmd_trans_unstable) of the pmd.
307 _pmd
= READ_ONCE(*pmd
);
312 if (!pmd_present(_pmd
))
315 if (pmd_trans_huge(_pmd
)) {
316 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
322 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
323 * and use the standard pte_offset_map() instead of parsing _pmd.
325 pte
= pte_offset_map(pmd
, address
);
327 * Lockless access: we're in a wait_event so it's ok if it
332 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
340 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
342 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
343 return TASK_INTERRUPTIBLE
;
345 if (flags
& FAULT_FLAG_KILLABLE
)
346 return TASK_KILLABLE
;
348 return TASK_UNINTERRUPTIBLE
;
352 * The locking rules involved in returning VM_FAULT_RETRY depending on
353 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
354 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
355 * recommendation in __lock_page_or_retry is not an understatement.
357 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
358 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
361 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
362 * set, VM_FAULT_RETRY can still be returned if and only if there are
363 * fatal_signal_pending()s, and the mmap_lock must be released before
366 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
368 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
369 struct userfaultfd_ctx
*ctx
;
370 struct userfaultfd_wait_queue uwq
;
371 vm_fault_t ret
= VM_FAULT_SIGBUS
;
373 unsigned int blocking_state
;
376 * We don't do userfault handling for the final child pid update.
378 * We also don't do userfault handling during
379 * coredumping. hugetlbfs has the special
380 * follow_hugetlb_page() to skip missing pages in the
381 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
382 * the no_page_table() helper in follow_page_mask(), but the
383 * shmem_vm_ops->fault method is invoked even during
384 * coredumping without mmap_lock and it ends up here.
386 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
390 * Coredumping runs without mmap_lock so we can only check that
391 * the mmap_lock is held, if PF_DUMPCORE was not set.
393 mmap_assert_locked(mm
);
395 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
399 BUG_ON(ctx
->mm
!= mm
);
401 /* Any unrecognized flag is a bug. */
402 VM_BUG_ON(reason
& ~__VM_UFFD_FLAGS
);
403 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
404 VM_BUG_ON(!reason
|| (reason
& (reason
- 1)));
406 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
408 if ((vmf
->flags
& FAULT_FLAG_USER
) == 0 &&
409 ctx
->flags
& UFFD_USER_MODE_ONLY
) {
410 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
411 "sysctl knob to 1 if kernel faults must be handled "
412 "without obtaining CAP_SYS_PTRACE capability\n");
417 * If it's already released don't get it. This avoids to loop
418 * in __get_user_pages if userfaultfd_release waits on the
419 * caller of handle_userfault to release the mmap_lock.
421 if (unlikely(READ_ONCE(ctx
->released
))) {
423 * Don't return VM_FAULT_SIGBUS in this case, so a non
424 * cooperative manager can close the uffd after the
425 * last UFFDIO_COPY, without risking to trigger an
426 * involuntary SIGBUS if the process was starting the
427 * userfaultfd while the userfaultfd was still armed
428 * (but after the last UFFDIO_COPY). If the uffd
429 * wasn't already closed when the userfault reached
430 * this point, that would normally be solved by
431 * userfaultfd_must_wait returning 'false'.
433 * If we were to return VM_FAULT_SIGBUS here, the non
434 * cooperative manager would be instead forced to
435 * always call UFFDIO_UNREGISTER before it can safely
438 ret
= VM_FAULT_NOPAGE
;
443 * Check that we can return VM_FAULT_RETRY.
445 * NOTE: it should become possible to return VM_FAULT_RETRY
446 * even if FAULT_FLAG_TRIED is set without leading to gup()
447 * -EBUSY failures, if the userfaultfd is to be extended for
448 * VM_UFFD_WP tracking and we intend to arm the userfault
449 * without first stopping userland access to the memory. For
450 * VM_UFFD_MISSING userfaults this is enough for now.
452 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
454 * Validate the invariant that nowait must allow retry
455 * to be sure not to return SIGBUS erroneously on
456 * nowait invocations.
458 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
459 #ifdef CONFIG_DEBUG_VM
460 if (printk_ratelimit()) {
462 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
471 * Handle nowait, not much to do other than tell it to retry
474 ret
= VM_FAULT_RETRY
;
475 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
478 /* take the reference before dropping the mmap_lock */
479 userfaultfd_ctx_get(ctx
);
481 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
482 uwq
.wq
.private = current
;
483 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->flags
, reason
,
488 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
490 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
492 * After the __add_wait_queue the uwq is visible to userland
493 * through poll/read().
495 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
497 * The smp_mb() after __set_current_state prevents the reads
498 * following the spin_unlock to happen before the list_add in
501 set_current_state(blocking_state
);
502 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
504 if (!is_vm_hugetlb_page(vmf
->vma
))
505 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
508 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
511 mmap_read_unlock(mm
);
513 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
514 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
518 __set_current_state(TASK_RUNNING
);
521 * Here we race with the list_del; list_add in
522 * userfaultfd_ctx_read(), however because we don't ever run
523 * list_del_init() to refile across the two lists, the prev
524 * and next pointers will never point to self. list_add also
525 * would never let any of the two pointers to point to
526 * self. So list_empty_careful won't risk to see both pointers
527 * pointing to self at any time during the list refile. The
528 * only case where list_del_init() is called is the full
529 * removal in the wake function and there we don't re-list_add
530 * and it's fine not to block on the spinlock. The uwq on this
531 * kernel stack can be released after the list_del_init.
533 if (!list_empty_careful(&uwq
.wq
.entry
)) {
534 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
536 * No need of list_del_init(), the uwq on the stack
537 * will be freed shortly anyway.
539 list_del(&uwq
.wq
.entry
);
540 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
544 * ctx may go away after this if the userfault pseudo fd is
547 userfaultfd_ctx_put(ctx
);
553 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
554 struct userfaultfd_wait_queue
*ewq
)
556 struct userfaultfd_ctx
*release_new_ctx
;
558 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
562 init_waitqueue_entry(&ewq
->wq
, current
);
563 release_new_ctx
= NULL
;
565 spin_lock_irq(&ctx
->event_wqh
.lock
);
567 * After the __add_wait_queue the uwq is visible to userland
568 * through poll/read().
570 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
572 set_current_state(TASK_KILLABLE
);
573 if (ewq
->msg
.event
== 0)
575 if (READ_ONCE(ctx
->released
) ||
576 fatal_signal_pending(current
)) {
578 * &ewq->wq may be queued in fork_event, but
579 * __remove_wait_queue ignores the head
580 * parameter. It would be a problem if it
583 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
584 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
585 struct userfaultfd_ctx
*new;
587 new = (struct userfaultfd_ctx
*)
589 ewq
->msg
.arg
.reserved
.reserved1
;
590 release_new_ctx
= new;
595 spin_unlock_irq(&ctx
->event_wqh
.lock
);
597 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
600 spin_lock_irq(&ctx
->event_wqh
.lock
);
602 __set_current_state(TASK_RUNNING
);
603 spin_unlock_irq(&ctx
->event_wqh
.lock
);
605 if (release_new_ctx
) {
606 struct vm_area_struct
*vma
;
607 struct mm_struct
*mm
= release_new_ctx
->mm
;
609 /* the various vma->vm_userfaultfd_ctx still points to it */
611 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
612 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
613 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
614 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
616 mmap_write_unlock(mm
);
618 userfaultfd_ctx_put(release_new_ctx
);
622 * ctx may go away after this if the userfault pseudo fd is
626 WRITE_ONCE(ctx
->mmap_changing
, false);
627 userfaultfd_ctx_put(ctx
);
630 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
631 struct userfaultfd_wait_queue
*ewq
)
634 wake_up_locked(&ctx
->event_wqh
);
635 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
638 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
640 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
641 struct userfaultfd_fork_ctx
*fctx
;
643 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
644 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
645 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
646 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
650 list_for_each_entry(fctx
, fcs
, list
)
651 if (fctx
->orig
== octx
) {
657 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
661 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
667 refcount_set(&ctx
->refcount
, 1);
668 ctx
->flags
= octx
->flags
;
669 ctx
->state
= UFFD_STATE_RUNNING
;
670 ctx
->features
= octx
->features
;
671 ctx
->released
= false;
672 ctx
->mmap_changing
= false;
673 ctx
->mm
= vma
->vm_mm
;
676 userfaultfd_ctx_get(octx
);
677 WRITE_ONCE(octx
->mmap_changing
, true);
680 list_add_tail(&fctx
->list
, fcs
);
683 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
687 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
689 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
690 struct userfaultfd_wait_queue ewq
;
694 ewq
.msg
.event
= UFFD_EVENT_FORK
;
695 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
697 userfaultfd_event_wait_completion(ctx
, &ewq
);
700 void dup_userfaultfd_complete(struct list_head
*fcs
)
702 struct userfaultfd_fork_ctx
*fctx
, *n
;
704 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
706 list_del(&fctx
->list
);
711 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
712 struct vm_userfaultfd_ctx
*vm_ctx
)
714 struct userfaultfd_ctx
*ctx
;
716 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
721 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
723 userfaultfd_ctx_get(ctx
);
724 WRITE_ONCE(ctx
->mmap_changing
, true);
726 /* Drop uffd context if remap feature not enabled */
727 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
728 vma
->vm_flags
&= ~__VM_UFFD_FLAGS
;
732 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
733 unsigned long from
, unsigned long to
,
736 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
737 struct userfaultfd_wait_queue ewq
;
742 if (to
& ~PAGE_MASK
) {
743 userfaultfd_ctx_put(ctx
);
749 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
750 ewq
.msg
.arg
.remap
.from
= from
;
751 ewq
.msg
.arg
.remap
.to
= to
;
752 ewq
.msg
.arg
.remap
.len
= len
;
754 userfaultfd_event_wait_completion(ctx
, &ewq
);
757 bool userfaultfd_remove(struct vm_area_struct
*vma
,
758 unsigned long start
, unsigned long end
)
760 struct mm_struct
*mm
= vma
->vm_mm
;
761 struct userfaultfd_ctx
*ctx
;
762 struct userfaultfd_wait_queue ewq
;
764 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
765 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
768 userfaultfd_ctx_get(ctx
);
769 WRITE_ONCE(ctx
->mmap_changing
, true);
770 mmap_read_unlock(mm
);
774 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
775 ewq
.msg
.arg
.remove
.start
= start
;
776 ewq
.msg
.arg
.remove
.end
= end
;
778 userfaultfd_event_wait_completion(ctx
, &ewq
);
783 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
784 unsigned long start
, unsigned long end
)
786 struct userfaultfd_unmap_ctx
*unmap_ctx
;
788 list_for_each_entry(unmap_ctx
, unmaps
, list
)
789 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
790 unmap_ctx
->end
== end
)
796 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
,
797 unsigned long start
, unsigned long end
,
798 struct list_head
*unmaps
)
800 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
801 struct userfaultfd_unmap_ctx
*unmap_ctx
;
802 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
804 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
805 has_unmap_ctx(ctx
, unmaps
, start
, end
))
808 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
812 userfaultfd_ctx_get(ctx
);
813 WRITE_ONCE(ctx
->mmap_changing
, true);
814 unmap_ctx
->ctx
= ctx
;
815 unmap_ctx
->start
= start
;
816 unmap_ctx
->end
= end
;
817 list_add_tail(&unmap_ctx
->list
, unmaps
);
823 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
825 struct userfaultfd_unmap_ctx
*ctx
, *n
;
826 struct userfaultfd_wait_queue ewq
;
828 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
831 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
832 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
833 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
835 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
837 list_del(&ctx
->list
);
842 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
844 struct userfaultfd_ctx
*ctx
= file
->private_data
;
845 struct mm_struct
*mm
= ctx
->mm
;
846 struct vm_area_struct
*vma
, *prev
;
847 /* len == 0 means wake all */
848 struct userfaultfd_wake_range range
= { .len
= 0, };
849 unsigned long new_flags
;
851 WRITE_ONCE(ctx
->released
, true);
853 if (!mmget_not_zero(mm
))
857 * Flush page faults out of all CPUs. NOTE: all page faults
858 * must be retried without returning VM_FAULT_SIGBUS if
859 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
860 * changes while handle_userfault released the mmap_lock. So
861 * it's critical that released is set to true (above), before
862 * taking the mmap_lock for writing.
866 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
868 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
869 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
870 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
874 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
875 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
876 new_flags
, vma
->anon_vma
,
877 vma
->vm_file
, vma
->vm_pgoff
,
884 vma
->vm_flags
= new_flags
;
885 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
887 mmap_write_unlock(mm
);
891 * After no new page faults can wait on this fault_*wqh, flush
892 * the last page faults that may have been already waiting on
895 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
896 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
897 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
898 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
900 /* Flush pending events that may still wait on event_wqh */
901 wake_up_all(&ctx
->event_wqh
);
903 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
904 userfaultfd_ctx_put(ctx
);
908 /* fault_pending_wqh.lock must be hold by the caller */
909 static inline struct userfaultfd_wait_queue
*find_userfault_in(
910 wait_queue_head_t
*wqh
)
912 wait_queue_entry_t
*wq
;
913 struct userfaultfd_wait_queue
*uwq
;
915 lockdep_assert_held(&wqh
->lock
);
918 if (!waitqueue_active(wqh
))
920 /* walk in reverse to provide FIFO behavior to read userfaults */
921 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
922 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
927 static inline struct userfaultfd_wait_queue
*find_userfault(
928 struct userfaultfd_ctx
*ctx
)
930 return find_userfault_in(&ctx
->fault_pending_wqh
);
933 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
934 struct userfaultfd_ctx
*ctx
)
936 return find_userfault_in(&ctx
->event_wqh
);
939 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
941 struct userfaultfd_ctx
*ctx
= file
->private_data
;
944 poll_wait(file
, &ctx
->fd_wqh
, wait
);
946 switch (ctx
->state
) {
947 case UFFD_STATE_WAIT_API
:
949 case UFFD_STATE_RUNNING
:
951 * poll() never guarantees that read won't block.
952 * userfaults can be waken before they're read().
954 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
957 * lockless access to see if there are pending faults
958 * __pollwait last action is the add_wait_queue but
959 * the spin_unlock would allow the waitqueue_active to
960 * pass above the actual list_add inside
961 * add_wait_queue critical section. So use a full
962 * memory barrier to serialize the list_add write of
963 * add_wait_queue() with the waitqueue_active read
968 if (waitqueue_active(&ctx
->fault_pending_wqh
))
970 else if (waitqueue_active(&ctx
->event_wqh
))
980 static const struct file_operations userfaultfd_fops
;
982 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
984 struct uffd_msg
*msg
)
988 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, new,
989 O_RDWR
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
993 msg
->arg
.reserved
.reserved1
= 0;
994 msg
->arg
.fork
.ufd
= fd
;
998 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
999 struct uffd_msg
*msg
, struct inode
*inode
)
1002 DECLARE_WAITQUEUE(wait
, current
);
1003 struct userfaultfd_wait_queue
*uwq
;
1005 * Handling fork event requires sleeping operations, so
1006 * we drop the event_wqh lock, then do these ops, then
1007 * lock it back and wake up the waiter. While the lock is
1008 * dropped the ewq may go away so we keep track of it
1011 LIST_HEAD(fork_event
);
1012 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1014 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1015 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1016 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1018 set_current_state(TASK_INTERRUPTIBLE
);
1019 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1020 uwq
= find_userfault(ctx
);
1023 * Use a seqcount to repeat the lockless check
1024 * in wake_userfault() to avoid missing
1025 * wakeups because during the refile both
1026 * waitqueue could become empty if this is the
1029 write_seqcount_begin(&ctx
->refile_seq
);
1032 * The fault_pending_wqh.lock prevents the uwq
1033 * to disappear from under us.
1035 * Refile this userfault from
1036 * fault_pending_wqh to fault_wqh, it's not
1037 * pending anymore after we read it.
1039 * Use list_del() by hand (as
1040 * userfaultfd_wake_function also uses
1041 * list_del_init() by hand) to be sure nobody
1042 * changes __remove_wait_queue() to use
1043 * list_del_init() in turn breaking the
1044 * !list_empty_careful() check in
1045 * handle_userfault(). The uwq->wq.head list
1046 * must never be empty at any time during the
1047 * refile, or the waitqueue could disappear
1048 * from under us. The "wait_queue_head_t"
1049 * parameter of __remove_wait_queue() is unused
1052 list_del(&uwq
->wq
.entry
);
1053 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1055 write_seqcount_end(&ctx
->refile_seq
);
1057 /* careful to always initialize msg if ret == 0 */
1059 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1063 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1065 spin_lock(&ctx
->event_wqh
.lock
);
1066 uwq
= find_userfault_evt(ctx
);
1070 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1071 fork_nctx
= (struct userfaultfd_ctx
*)
1073 uwq
->msg
.arg
.reserved
.reserved1
;
1074 list_move(&uwq
->wq
.entry
, &fork_event
);
1076 * fork_nctx can be freed as soon as
1077 * we drop the lock, unless we take a
1080 userfaultfd_ctx_get(fork_nctx
);
1081 spin_unlock(&ctx
->event_wqh
.lock
);
1086 userfaultfd_event_complete(ctx
, uwq
);
1087 spin_unlock(&ctx
->event_wqh
.lock
);
1091 spin_unlock(&ctx
->event_wqh
.lock
);
1093 if (signal_pending(current
)) {
1101 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1103 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1105 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1106 __set_current_state(TASK_RUNNING
);
1107 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1109 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1110 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1111 spin_lock_irq(&ctx
->event_wqh
.lock
);
1112 if (!list_empty(&fork_event
)) {
1114 * The fork thread didn't abort, so we can
1115 * drop the temporary refcount.
1117 userfaultfd_ctx_put(fork_nctx
);
1119 uwq
= list_first_entry(&fork_event
,
1123 * If fork_event list wasn't empty and in turn
1124 * the event wasn't already released by fork
1125 * (the event is allocated on fork kernel
1126 * stack), put the event back to its place in
1127 * the event_wq. fork_event head will be freed
1128 * as soon as we return so the event cannot
1129 * stay queued there no matter the current
1132 list_del(&uwq
->wq
.entry
);
1133 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1136 * Leave the event in the waitqueue and report
1137 * error to userland if we failed to resolve
1138 * the userfault fork.
1141 userfaultfd_event_complete(ctx
, uwq
);
1144 * Here the fork thread aborted and the
1145 * refcount from the fork thread on fork_nctx
1146 * has already been released. We still hold
1147 * the reference we took before releasing the
1148 * lock above. If resolve_userfault_fork
1149 * failed we've to drop it because the
1150 * fork_nctx has to be freed in such case. If
1151 * it succeeded we'll hold it because the new
1152 * uffd references it.
1155 userfaultfd_ctx_put(fork_nctx
);
1157 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1163 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1164 size_t count
, loff_t
*ppos
)
1166 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1167 ssize_t _ret
, ret
= 0;
1168 struct uffd_msg msg
;
1169 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1170 struct inode
*inode
= file_inode(file
);
1172 if (ctx
->state
== UFFD_STATE_WAIT_API
)
1176 if (count
< sizeof(msg
))
1177 return ret
? ret
: -EINVAL
;
1178 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1180 return ret
? ret
: _ret
;
1181 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1182 return ret
? ret
: -EFAULT
;
1185 count
-= sizeof(msg
);
1187 * Allow to read more than one fault at time but only
1188 * block if waiting for the very first one.
1190 no_wait
= O_NONBLOCK
;
1194 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1195 struct userfaultfd_wake_range
*range
)
1197 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1198 /* wake all in the range and autoremove */
1199 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1200 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1202 if (waitqueue_active(&ctx
->fault_wqh
))
1203 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1204 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1207 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1208 struct userfaultfd_wake_range
*range
)
1214 * To be sure waitqueue_active() is not reordered by the CPU
1215 * before the pagetable update, use an explicit SMP memory
1216 * barrier here. PT lock release or mmap_read_unlock(mm) still
1217 * have release semantics that can allow the
1218 * waitqueue_active() to be reordered before the pte update.
1223 * Use waitqueue_active because it's very frequent to
1224 * change the address space atomically even if there are no
1225 * userfaults yet. So we take the spinlock only when we're
1226 * sure we've userfaults to wake.
1229 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1230 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1231 waitqueue_active(&ctx
->fault_wqh
);
1233 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1235 __wake_userfault(ctx
, range
);
1238 static __always_inline
int validate_range(struct mm_struct
*mm
,
1239 __u64 start
, __u64 len
)
1241 __u64 task_size
= mm
->task_size
;
1243 if (start
& ~PAGE_MASK
)
1245 if (len
& ~PAGE_MASK
)
1249 if (start
< mmap_min_addr
)
1251 if (start
>= task_size
)
1253 if (len
> task_size
- start
)
1258 static inline bool vma_can_userfault(struct vm_area_struct
*vma
,
1259 unsigned long vm_flags
)
1261 /* FIXME: add WP support to hugetlbfs and shmem */
1262 if (vm_flags
& VM_UFFD_WP
) {
1263 if (is_vm_hugetlb_page(vma
) || vma_is_shmem(vma
))
1267 if (vm_flags
& VM_UFFD_MINOR
) {
1268 if (!(is_vm_hugetlb_page(vma
) || vma_is_shmem(vma
)))
1272 return vma_is_anonymous(vma
) || is_vm_hugetlb_page(vma
) ||
1276 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1279 struct mm_struct
*mm
= ctx
->mm
;
1280 struct vm_area_struct
*vma
, *prev
, *cur
;
1282 struct uffdio_register uffdio_register
;
1283 struct uffdio_register __user
*user_uffdio_register
;
1284 unsigned long vm_flags
, new_flags
;
1287 unsigned long start
, end
, vma_end
;
1289 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1292 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1293 sizeof(uffdio_register
)-sizeof(__u64
)))
1297 if (!uffdio_register
.mode
)
1299 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1302 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1303 vm_flags
|= VM_UFFD_MISSING
;
1304 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1305 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1308 vm_flags
|= VM_UFFD_WP
;
1310 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1311 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1314 vm_flags
|= VM_UFFD_MINOR
;
1317 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1318 uffdio_register
.range
.len
);
1322 start
= uffdio_register
.range
.start
;
1323 end
= start
+ uffdio_register
.range
.len
;
1326 if (!mmget_not_zero(mm
))
1329 mmap_write_lock(mm
);
1330 vma
= find_vma_prev(mm
, start
, &prev
);
1334 /* check that there's at least one vma in the range */
1336 if (vma
->vm_start
>= end
)
1340 * If the first vma contains huge pages, make sure start address
1341 * is aligned to huge page size.
1343 if (is_vm_hugetlb_page(vma
)) {
1344 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1346 if (start
& (vma_hpagesize
- 1))
1351 * Search for not compatible vmas.
1354 basic_ioctls
= false;
1355 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1358 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1359 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1361 /* check not compatible vmas */
1363 if (!vma_can_userfault(cur
, vm_flags
))
1367 * UFFDIO_COPY will fill file holes even without
1368 * PROT_WRITE. This check enforces that if this is a
1369 * MAP_SHARED, the process has write permission to the backing
1370 * file. If VM_MAYWRITE is set it also enforces that on a
1371 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1372 * F_WRITE_SEAL can be taken until the vma is destroyed.
1375 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1379 * If this vma contains ending address, and huge pages
1382 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1383 end
> cur
->vm_start
) {
1384 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1388 if (end
& (vma_hpagesize
- 1))
1391 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1395 * Check that this vma isn't already owned by a
1396 * different userfaultfd. We can't allow more than one
1397 * userfaultfd to own a single vma simultaneously or we
1398 * wouldn't know which one to deliver the userfaults to.
1401 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1402 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1406 * Note vmas containing huge pages
1408 if (is_vm_hugetlb_page(cur
))
1409 basic_ioctls
= true;
1415 if (vma
->vm_start
< start
)
1422 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1423 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1424 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1425 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1428 * Nothing to do: this vma is already registered into this
1429 * userfaultfd and with the right tracking mode too.
1431 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1432 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1435 if (vma
->vm_start
> start
)
1436 start
= vma
->vm_start
;
1437 vma_end
= min(end
, vma
->vm_end
);
1439 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1440 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1441 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1443 ((struct vm_userfaultfd_ctx
){ ctx
}));
1448 if (vma
->vm_start
< start
) {
1449 ret
= split_vma(mm
, vma
, start
, 1);
1453 if (vma
->vm_end
> end
) {
1454 ret
= split_vma(mm
, vma
, end
, 0);
1460 * In the vma_merge() successful mprotect-like case 8:
1461 * the next vma was merged into the current one and
1462 * the current one has not been updated yet.
1464 vma
->vm_flags
= new_flags
;
1465 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1467 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1468 hugetlb_unshare_all_pmds(vma
);
1472 start
= vma
->vm_end
;
1474 } while (vma
&& vma
->vm_start
< end
);
1476 mmap_write_unlock(mm
);
1481 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1482 UFFD_API_RANGE_IOCTLS
;
1485 * Declare the WP ioctl only if the WP mode is
1486 * specified and all checks passed with the range
1488 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1489 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1491 /* CONTINUE ioctl is only supported for MINOR ranges. */
1492 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
))
1493 ioctls_out
&= ~((__u64
)1 << _UFFDIO_CONTINUE
);
1496 * Now that we scanned all vmas we can already tell
1497 * userland which ioctls methods are guaranteed to
1498 * succeed on this range.
1500 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1507 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1510 struct mm_struct
*mm
= ctx
->mm
;
1511 struct vm_area_struct
*vma
, *prev
, *cur
;
1513 struct uffdio_range uffdio_unregister
;
1514 unsigned long new_flags
;
1516 unsigned long start
, end
, vma_end
;
1517 const void __user
*buf
= (void __user
*)arg
;
1520 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1523 ret
= validate_range(mm
, uffdio_unregister
.start
,
1524 uffdio_unregister
.len
);
1528 start
= uffdio_unregister
.start
;
1529 end
= start
+ uffdio_unregister
.len
;
1532 if (!mmget_not_zero(mm
))
1535 mmap_write_lock(mm
);
1536 vma
= find_vma_prev(mm
, start
, &prev
);
1540 /* check that there's at least one vma in the range */
1542 if (vma
->vm_start
>= end
)
1546 * If the first vma contains huge pages, make sure start address
1547 * is aligned to huge page size.
1549 if (is_vm_hugetlb_page(vma
)) {
1550 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1552 if (start
& (vma_hpagesize
- 1))
1557 * Search for not compatible vmas.
1561 for (cur
= vma
; cur
&& cur
->vm_start
< end
; cur
= cur
->vm_next
) {
1564 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1565 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1568 * Check not compatible vmas, not strictly required
1569 * here as not compatible vmas cannot have an
1570 * userfaultfd_ctx registered on them, but this
1571 * provides for more strict behavior to notice
1572 * unregistration errors.
1574 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1581 if (vma
->vm_start
< start
)
1588 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1591 * Nothing to do: this vma is already registered into this
1592 * userfaultfd and with the right tracking mode too.
1594 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1597 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1599 if (vma
->vm_start
> start
)
1600 start
= vma
->vm_start
;
1601 vma_end
= min(end
, vma
->vm_end
);
1603 if (userfaultfd_missing(vma
)) {
1605 * Wake any concurrent pending userfault while
1606 * we unregister, so they will not hang
1607 * permanently and it avoids userland to call
1608 * UFFDIO_WAKE explicitly.
1610 struct userfaultfd_wake_range range
;
1611 range
.start
= start
;
1612 range
.len
= vma_end
- start
;
1613 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1616 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1617 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1618 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1625 if (vma
->vm_start
< start
) {
1626 ret
= split_vma(mm
, vma
, start
, 1);
1630 if (vma
->vm_end
> end
) {
1631 ret
= split_vma(mm
, vma
, end
, 0);
1637 * In the vma_merge() successful mprotect-like case 8:
1638 * the next vma was merged into the current one and
1639 * the current one has not been updated yet.
1641 vma
->vm_flags
= new_flags
;
1642 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1646 start
= vma
->vm_end
;
1648 } while (vma
&& vma
->vm_start
< end
);
1650 mmap_write_unlock(mm
);
1657 * userfaultfd_wake may be used in combination with the
1658 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1660 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1664 struct uffdio_range uffdio_wake
;
1665 struct userfaultfd_wake_range range
;
1666 const void __user
*buf
= (void __user
*)arg
;
1669 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1672 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1676 range
.start
= uffdio_wake
.start
;
1677 range
.len
= uffdio_wake
.len
;
1680 * len == 0 means wake all and we don't want to wake all here,
1681 * so check it again to be sure.
1683 VM_BUG_ON(!range
.len
);
1685 wake_userfault(ctx
, &range
);
1692 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1696 struct uffdio_copy uffdio_copy
;
1697 struct uffdio_copy __user
*user_uffdio_copy
;
1698 struct userfaultfd_wake_range range
;
1700 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1703 if (READ_ONCE(ctx
->mmap_changing
))
1707 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1708 /* don't copy "copy" last field */
1709 sizeof(uffdio_copy
)-sizeof(__s64
)))
1712 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1716 * double check for wraparound just in case. copy_from_user()
1717 * will later check uffdio_copy.src + uffdio_copy.len to fit
1718 * in the userland range.
1721 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1723 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1725 if (mmget_not_zero(ctx
->mm
)) {
1726 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1727 uffdio_copy
.len
, &ctx
->mmap_changing
,
1733 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1738 /* len == 0 would wake all */
1740 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1741 range
.start
= uffdio_copy
.dst
;
1742 wake_userfault(ctx
, &range
);
1744 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1749 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1753 struct uffdio_zeropage uffdio_zeropage
;
1754 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1755 struct userfaultfd_wake_range range
;
1757 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1760 if (READ_ONCE(ctx
->mmap_changing
))
1764 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1765 /* don't copy "zeropage" last field */
1766 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1769 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1770 uffdio_zeropage
.range
.len
);
1774 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1777 if (mmget_not_zero(ctx
->mm
)) {
1778 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1779 uffdio_zeropage
.range
.len
,
1780 &ctx
->mmap_changing
);
1785 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1789 /* len == 0 would wake all */
1792 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1793 range
.start
= uffdio_zeropage
.range
.start
;
1794 wake_userfault(ctx
, &range
);
1796 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1801 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1805 struct uffdio_writeprotect uffdio_wp
;
1806 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1807 struct userfaultfd_wake_range range
;
1808 bool mode_wp
, mode_dontwake
;
1810 if (READ_ONCE(ctx
->mmap_changing
))
1813 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1815 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1816 sizeof(struct uffdio_writeprotect
)))
1819 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1820 uffdio_wp
.range
.len
);
1824 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1825 UFFDIO_WRITEPROTECT_MODE_WP
))
1828 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1829 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1831 if (mode_wp
&& mode_dontwake
)
1834 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1835 uffdio_wp
.range
.len
, mode_wp
,
1836 &ctx
->mmap_changing
);
1840 if (!mode_wp
&& !mode_dontwake
) {
1841 range
.start
= uffdio_wp
.range
.start
;
1842 range
.len
= uffdio_wp
.range
.len
;
1843 wake_userfault(ctx
, &range
);
1848 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1851 struct uffdio_continue uffdio_continue
;
1852 struct uffdio_continue __user
*user_uffdio_continue
;
1853 struct userfaultfd_wake_range range
;
1855 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1858 if (READ_ONCE(ctx
->mmap_changing
))
1862 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1863 /* don't copy the output fields */
1864 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1867 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1868 uffdio_continue
.range
.len
);
1873 /* double check for wraparound just in case. */
1874 if (uffdio_continue
.range
.start
+ uffdio_continue
.range
.len
<=
1875 uffdio_continue
.range
.start
) {
1878 if (uffdio_continue
.mode
& ~UFFDIO_CONTINUE_MODE_DONTWAKE
)
1881 if (mmget_not_zero(ctx
->mm
)) {
1882 ret
= mcopy_continue(ctx
->mm
, uffdio_continue
.range
.start
,
1883 uffdio_continue
.range
.len
,
1884 &ctx
->mmap_changing
);
1890 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1895 /* len == 0 would wake all */
1898 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1899 range
.start
= uffdio_continue
.range
.start
;
1900 wake_userfault(ctx
, &range
);
1902 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1908 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1911 * For the current set of features the bits just coincide
1913 return (unsigned int)user_features
;
1917 * userland asks for a certain API version and we return which bits
1918 * and ioctl commands are implemented in this kernel for such API
1919 * version or -EINVAL if unknown.
1921 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1924 struct uffdio_api uffdio_api
;
1925 void __user
*buf
= (void __user
*)arg
;
1930 if (ctx
->state
!= UFFD_STATE_WAIT_API
)
1933 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1935 features
= uffdio_api
.features
;
1937 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
1940 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1942 /* report all available features and ioctls to userland */
1943 uffdio_api
.features
= UFFD_API_FEATURES
;
1944 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1945 uffdio_api
.features
&=
1946 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
1948 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1949 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
1951 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1953 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1955 ctx
->state
= UFFD_STATE_RUNNING
;
1956 /* only enable the requested features for this uffd context */
1957 ctx
->features
= uffd_ctx_features(features
);
1962 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1963 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1968 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
1972 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1974 if (cmd
!= UFFDIO_API
&& ctx
->state
== UFFD_STATE_WAIT_API
)
1979 ret
= userfaultfd_api(ctx
, arg
);
1981 case UFFDIO_REGISTER
:
1982 ret
= userfaultfd_register(ctx
, arg
);
1984 case UFFDIO_UNREGISTER
:
1985 ret
= userfaultfd_unregister(ctx
, arg
);
1988 ret
= userfaultfd_wake(ctx
, arg
);
1991 ret
= userfaultfd_copy(ctx
, arg
);
1993 case UFFDIO_ZEROPAGE
:
1994 ret
= userfaultfd_zeropage(ctx
, arg
);
1996 case UFFDIO_WRITEPROTECT
:
1997 ret
= userfaultfd_writeprotect(ctx
, arg
);
1999 case UFFDIO_CONTINUE
:
2000 ret
= userfaultfd_continue(ctx
, arg
);
2006 #ifdef CONFIG_PROC_FS
2007 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2009 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2010 wait_queue_entry_t
*wq
;
2011 unsigned long pending
= 0, total
= 0;
2013 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2014 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2018 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2021 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2024 * If more protocols will be added, there will be all shown
2025 * separated by a space. Like this:
2026 * protocols: aa:... bb:...
2028 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2029 pending
, total
, UFFD_API
, ctx
->features
,
2030 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2034 static const struct file_operations userfaultfd_fops
= {
2035 #ifdef CONFIG_PROC_FS
2036 .show_fdinfo
= userfaultfd_show_fdinfo
,
2038 .release
= userfaultfd_release
,
2039 .poll
= userfaultfd_poll
,
2040 .read
= userfaultfd_read
,
2041 .unlocked_ioctl
= userfaultfd_ioctl
,
2042 .compat_ioctl
= compat_ptr_ioctl
,
2043 .llseek
= noop_llseek
,
2046 static void init_once_userfaultfd_ctx(void *mem
)
2048 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2050 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2051 init_waitqueue_head(&ctx
->fault_wqh
);
2052 init_waitqueue_head(&ctx
->event_wqh
);
2053 init_waitqueue_head(&ctx
->fd_wqh
);
2054 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2057 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2059 struct userfaultfd_ctx
*ctx
;
2062 if (!sysctl_unprivileged_userfaultfd
&&
2063 (flags
& UFFD_USER_MODE_ONLY
) == 0 &&
2064 !capable(CAP_SYS_PTRACE
)) {
2065 printk_once(KERN_WARNING
"uffd: Set unprivileged_userfaultfd "
2066 "sysctl knob to 1 if kernel faults must be handled "
2067 "without obtaining CAP_SYS_PTRACE capability\n");
2071 BUG_ON(!current
->mm
);
2073 /* Check the UFFD_* constants for consistency. */
2074 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2075 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2076 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2078 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2081 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2085 refcount_set(&ctx
->refcount
, 1);
2088 ctx
->state
= UFFD_STATE_WAIT_API
;
2089 ctx
->released
= false;
2090 ctx
->mmap_changing
= false;
2091 ctx
->mm
= current
->mm
;
2092 /* prevent the mm struct to be freed */
2095 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, ctx
,
2096 O_RDWR
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2099 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2104 static int __init
userfaultfd_init(void)
2106 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2107 sizeof(struct userfaultfd_ctx
),
2109 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2110 init_once_userfaultfd_ctx
);
2113 __initcall(userfaultfd_init
);