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>
33 #include <linux/miscdevice.h>
35 static int sysctl_unprivileged_userfaultfd __read_mostly
;
38 static struct ctl_table vm_userfaultfd_table
[] = {
40 .procname
= "unprivileged_userfaultfd",
41 .data
= &sysctl_unprivileged_userfaultfd
,
42 .maxlen
= sizeof(sysctl_unprivileged_userfaultfd
),
44 .proc_handler
= proc_dointvec_minmax
,
45 .extra1
= SYSCTL_ZERO
,
51 static struct kmem_cache
*userfaultfd_ctx_cachep __ro_after_init
;
53 struct userfaultfd_fork_ctx
{
54 struct userfaultfd_ctx
*orig
;
55 struct userfaultfd_ctx
*new;
56 struct list_head list
;
59 struct userfaultfd_unmap_ctx
{
60 struct userfaultfd_ctx
*ctx
;
63 struct list_head list
;
66 struct userfaultfd_wait_queue
{
68 wait_queue_entry_t wq
;
69 struct userfaultfd_ctx
*ctx
;
73 struct userfaultfd_wake_range
{
78 /* internal indication that UFFD_API ioctl was successfully executed */
79 #define UFFD_FEATURE_INITIALIZED (1u << 31)
81 static bool userfaultfd_is_initialized(struct userfaultfd_ctx
*ctx
)
83 return ctx
->features
& UFFD_FEATURE_INITIALIZED
;
86 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx
*ctx
)
88 return ctx
&& (ctx
->features
& UFFD_FEATURE_WP_ASYNC
);
92 * Whether WP_UNPOPULATED is enabled on the uffd context. It is only
93 * meaningful when userfaultfd_wp()==true on the vma and when it's
96 bool userfaultfd_wp_unpopulated(struct vm_area_struct
*vma
)
98 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
103 return ctx
->features
& UFFD_FEATURE_WP_UNPOPULATED
;
106 static void userfaultfd_set_vm_flags(struct vm_area_struct
*vma
,
109 const bool uffd_wp_changed
= (vma
->vm_flags
^ flags
) & VM_UFFD_WP
;
111 vm_flags_reset(vma
, flags
);
113 * For shared mappings, we want to enable writenotify while
114 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
115 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
117 if ((vma
->vm_flags
& VM_SHARED
) && uffd_wp_changed
)
118 vma_set_page_prot(vma
);
121 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
122 int wake_flags
, void *key
)
124 struct userfaultfd_wake_range
*range
= key
;
126 struct userfaultfd_wait_queue
*uwq
;
127 unsigned long start
, len
;
129 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
131 /* len == 0 means wake all */
132 start
= range
->start
;
134 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
135 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
137 WRITE_ONCE(uwq
->waken
, true);
139 * The Program-Order guarantees provided by the scheduler
140 * ensure uwq->waken is visible before the task is woken.
142 ret
= wake_up_state(wq
->private, mode
);
145 * Wake only once, autoremove behavior.
147 * After the effect of list_del_init is visible to the other
148 * CPUs, the waitqueue may disappear from under us, see the
149 * !list_empty_careful() in handle_userfault().
151 * try_to_wake_up() has an implicit smp_mb(), and the
152 * wq->private is read before calling the extern function
153 * "wake_up_state" (which in turns calls try_to_wake_up).
155 list_del_init(&wq
->entry
);
162 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
164 * @ctx: [in] Pointer to the userfaultfd context.
166 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
168 refcount_inc(&ctx
->refcount
);
172 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
174 * @ctx: [in] Pointer to userfaultfd context.
176 * The userfaultfd context reference must have been previously acquired either
177 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
179 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
181 if (refcount_dec_and_test(&ctx
->refcount
)) {
182 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
183 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
184 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
185 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
186 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
187 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
188 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
189 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
191 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
195 static inline void msg_init(struct uffd_msg
*msg
)
197 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
199 * Must use memset to zero out the paddings or kernel data is
200 * leaked to userland.
202 memset(msg
, 0, sizeof(struct uffd_msg
));
205 static inline struct uffd_msg
userfault_msg(unsigned long address
,
206 unsigned long real_address
,
208 unsigned long reason
,
209 unsigned int features
)
214 msg
.event
= UFFD_EVENT_PAGEFAULT
;
216 msg
.arg
.pagefault
.address
= (features
& UFFD_FEATURE_EXACT_ADDRESS
) ?
217 real_address
: address
;
220 * These flags indicate why the userfault occurred:
221 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
222 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
223 * - Neither of these flags being set indicates a MISSING fault.
225 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
226 * fault. Otherwise, it was a read fault.
228 if (flags
& FAULT_FLAG_WRITE
)
229 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
230 if (reason
& VM_UFFD_WP
)
231 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
232 if (reason
& VM_UFFD_MINOR
)
233 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_MINOR
;
234 if (features
& UFFD_FEATURE_THREAD_ID
)
235 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
239 #ifdef CONFIG_HUGETLB_PAGE
241 * Same functionality as userfaultfd_must_wait below with modifications for
244 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
245 struct vm_fault
*vmf
,
246 unsigned long reason
)
248 struct vm_area_struct
*vma
= vmf
->vma
;
252 assert_fault_locked(vmf
);
254 ptep
= hugetlb_walk(vma
, vmf
->address
, vma_mmu_pagesize(vma
));
259 pte
= huge_ptep_get(ptep
);
262 * Lockless access: we're in a wait_event so it's ok if it
263 * changes under us. PTE markers should be handled the same as none
266 if (huge_pte_none_mostly(pte
))
268 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
274 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
275 struct vm_fault
*vmf
,
276 unsigned long reason
)
278 return false; /* should never get here */
280 #endif /* CONFIG_HUGETLB_PAGE */
283 * Verify the pagetables are still not ok after having reigstered into
284 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
285 * userfault that has already been resolved, if userfaultfd_read and
286 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
289 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
290 struct vm_fault
*vmf
,
291 unsigned long reason
)
293 struct mm_struct
*mm
= ctx
->mm
;
294 unsigned long address
= vmf
->address
;
303 assert_fault_locked(vmf
);
305 pgd
= pgd_offset(mm
, address
);
306 if (!pgd_present(*pgd
))
308 p4d
= p4d_offset(pgd
, address
);
309 if (!p4d_present(*p4d
))
311 pud
= pud_offset(p4d
, address
);
312 if (!pud_present(*pud
))
314 pmd
= pmd_offset(pud
, address
);
316 _pmd
= pmdp_get_lockless(pmd
);
321 if (!pmd_present(_pmd
) || pmd_devmap(_pmd
))
324 if (pmd_trans_huge(_pmd
)) {
325 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
330 pte
= pte_offset_map(pmd
, address
);
336 * Lockless access: we're in a wait_event so it's ok if it
337 * changes under us. PTE markers should be handled the same as none
340 ptent
= ptep_get(pte
);
341 if (pte_none_mostly(ptent
))
343 if (!pte_write(ptent
) && (reason
& VM_UFFD_WP
))
351 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
353 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
354 return TASK_INTERRUPTIBLE
;
356 if (flags
& FAULT_FLAG_KILLABLE
)
357 return TASK_KILLABLE
;
359 return TASK_UNINTERRUPTIBLE
;
363 * The locking rules involved in returning VM_FAULT_RETRY depending on
364 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
365 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
366 * recommendation in __lock_page_or_retry is not an understatement.
368 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
369 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
372 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
373 * set, VM_FAULT_RETRY can still be returned if and only if there are
374 * fatal_signal_pending()s, and the mmap_lock must be released before
377 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
379 struct vm_area_struct
*vma
= vmf
->vma
;
380 struct mm_struct
*mm
= vma
->vm_mm
;
381 struct userfaultfd_ctx
*ctx
;
382 struct userfaultfd_wait_queue uwq
;
383 vm_fault_t ret
= VM_FAULT_SIGBUS
;
385 unsigned int blocking_state
;
388 * We don't do userfault handling for the final child pid update.
390 * We also don't do userfault handling during
391 * coredumping. hugetlbfs has the special
392 * hugetlb_follow_page_mask() to skip missing pages in the
393 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
394 * the no_page_table() helper in follow_page_mask(), but the
395 * shmem_vm_ops->fault method is invoked even during
396 * coredumping and it ends up here.
398 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
401 assert_fault_locked(vmf
);
403 ctx
= 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
) && (ctx
->flags
& UFFD_USER_MODE_ONLY
))
420 * If it's already released don't get it. This avoids to loop
421 * in __get_user_pages if userfaultfd_release waits on the
422 * caller of handle_userfault to release the mmap_lock.
424 if (unlikely(READ_ONCE(ctx
->released
))) {
426 * Don't return VM_FAULT_SIGBUS in this case, so a non
427 * cooperative manager can close the uffd after the
428 * last UFFDIO_COPY, without risking to trigger an
429 * involuntary SIGBUS if the process was starting the
430 * userfaultfd while the userfaultfd was still armed
431 * (but after the last UFFDIO_COPY). If the uffd
432 * wasn't already closed when the userfault reached
433 * this point, that would normally be solved by
434 * userfaultfd_must_wait returning 'false'.
436 * If we were to return VM_FAULT_SIGBUS here, the non
437 * cooperative manager would be instead forced to
438 * always call UFFDIO_UNREGISTER before it can safely
441 ret
= VM_FAULT_NOPAGE
;
446 * Check that we can return VM_FAULT_RETRY.
448 * NOTE: it should become possible to return VM_FAULT_RETRY
449 * even if FAULT_FLAG_TRIED is set without leading to gup()
450 * -EBUSY failures, if the userfaultfd is to be extended for
451 * VM_UFFD_WP tracking and we intend to arm the userfault
452 * without first stopping userland access to the memory. For
453 * VM_UFFD_MISSING userfaults this is enough for now.
455 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
457 * Validate the invariant that nowait must allow retry
458 * to be sure not to return SIGBUS erroneously on
459 * nowait invocations.
461 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
462 #ifdef CONFIG_DEBUG_VM
463 if (printk_ratelimit()) {
465 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
474 * Handle nowait, not much to do other than tell it to retry
477 ret
= VM_FAULT_RETRY
;
478 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
481 /* take the reference before dropping the mmap_lock */
482 userfaultfd_ctx_get(ctx
);
484 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
485 uwq
.wq
.private = current
;
486 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->real_address
, vmf
->flags
,
487 reason
, ctx
->features
);
491 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
494 * Take the vma lock now, in order to safely call
495 * userfaultfd_huge_must_wait() later. Since acquiring the
496 * (sleepable) vma lock can modify the current task state, that
497 * must be before explicitly calling set_current_state().
499 if (is_vm_hugetlb_page(vma
))
500 hugetlb_vma_lock_read(vma
);
502 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
504 * After the __add_wait_queue the uwq is visible to userland
505 * through poll/read().
507 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
509 * The smp_mb() after __set_current_state prevents the reads
510 * following the spin_unlock to happen before the list_add in
513 set_current_state(blocking_state
);
514 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
516 if (!is_vm_hugetlb_page(vma
))
517 must_wait
= userfaultfd_must_wait(ctx
, vmf
, reason
);
519 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
, reason
);
520 if (is_vm_hugetlb_page(vma
))
521 hugetlb_vma_unlock_read(vma
);
522 release_fault_lock(vmf
);
524 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
525 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
529 __set_current_state(TASK_RUNNING
);
532 * Here we race with the list_del; list_add in
533 * userfaultfd_ctx_read(), however because we don't ever run
534 * list_del_init() to refile across the two lists, the prev
535 * and next pointers will never point to self. list_add also
536 * would never let any of the two pointers to point to
537 * self. So list_empty_careful won't risk to see both pointers
538 * pointing to self at any time during the list refile. The
539 * only case where list_del_init() is called is the full
540 * removal in the wake function and there we don't re-list_add
541 * and it's fine not to block on the spinlock. The uwq on this
542 * kernel stack can be released after the list_del_init.
544 if (!list_empty_careful(&uwq
.wq
.entry
)) {
545 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
547 * No need of list_del_init(), the uwq on the stack
548 * will be freed shortly anyway.
550 list_del(&uwq
.wq
.entry
);
551 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
555 * ctx may go away after this if the userfault pseudo fd is
558 userfaultfd_ctx_put(ctx
);
564 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
565 struct userfaultfd_wait_queue
*ewq
)
567 struct userfaultfd_ctx
*release_new_ctx
;
569 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
573 init_waitqueue_entry(&ewq
->wq
, current
);
574 release_new_ctx
= NULL
;
576 spin_lock_irq(&ctx
->event_wqh
.lock
);
578 * After the __add_wait_queue the uwq is visible to userland
579 * through poll/read().
581 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
583 set_current_state(TASK_KILLABLE
);
584 if (ewq
->msg
.event
== 0)
586 if (READ_ONCE(ctx
->released
) ||
587 fatal_signal_pending(current
)) {
589 * &ewq->wq may be queued in fork_event, but
590 * __remove_wait_queue ignores the head
591 * parameter. It would be a problem if it
594 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
595 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
596 struct userfaultfd_ctx
*new;
598 new = (struct userfaultfd_ctx
*)
600 ewq
->msg
.arg
.reserved
.reserved1
;
601 release_new_ctx
= new;
606 spin_unlock_irq(&ctx
->event_wqh
.lock
);
608 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
611 spin_lock_irq(&ctx
->event_wqh
.lock
);
613 __set_current_state(TASK_RUNNING
);
614 spin_unlock_irq(&ctx
->event_wqh
.lock
);
616 if (release_new_ctx
) {
617 struct vm_area_struct
*vma
;
618 struct mm_struct
*mm
= release_new_ctx
->mm
;
619 VMA_ITERATOR(vmi
, mm
, 0);
621 /* the various vma->vm_userfaultfd_ctx still points to it */
623 for_each_vma(vmi
, vma
) {
624 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
625 vma_start_write(vma
);
626 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
627 userfaultfd_set_vm_flags(vma
,
628 vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
631 mmap_write_unlock(mm
);
633 userfaultfd_ctx_put(release_new_ctx
);
637 * ctx may go away after this if the userfault pseudo fd is
641 atomic_dec(&ctx
->mmap_changing
);
642 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
643 userfaultfd_ctx_put(ctx
);
646 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
647 struct userfaultfd_wait_queue
*ewq
)
650 wake_up_locked(&ctx
->event_wqh
);
651 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
654 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
656 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
657 struct userfaultfd_fork_ctx
*fctx
;
659 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
660 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
661 vma_start_write(vma
);
662 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
663 userfaultfd_set_vm_flags(vma
, vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
667 list_for_each_entry(fctx
, fcs
, list
)
668 if (fctx
->orig
== octx
) {
674 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
678 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
684 refcount_set(&ctx
->refcount
, 1);
685 ctx
->flags
= octx
->flags
;
686 ctx
->features
= octx
->features
;
687 ctx
->released
= false;
688 init_rwsem(&ctx
->map_changing_lock
);
689 atomic_set(&ctx
->mmap_changing
, 0);
690 ctx
->mm
= vma
->vm_mm
;
693 userfaultfd_ctx_get(octx
);
694 down_write(&octx
->map_changing_lock
);
695 atomic_inc(&octx
->mmap_changing
);
696 up_write(&octx
->map_changing_lock
);
699 list_add_tail(&fctx
->list
, fcs
);
702 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
706 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
708 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
709 struct userfaultfd_wait_queue ewq
;
713 ewq
.msg
.event
= UFFD_EVENT_FORK
;
714 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
716 userfaultfd_event_wait_completion(ctx
, &ewq
);
719 void dup_userfaultfd_complete(struct list_head
*fcs
)
721 struct userfaultfd_fork_ctx
*fctx
, *n
;
723 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
725 list_del(&fctx
->list
);
730 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
731 struct vm_userfaultfd_ctx
*vm_ctx
)
733 struct userfaultfd_ctx
*ctx
;
735 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
740 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
742 userfaultfd_ctx_get(ctx
);
743 down_write(&ctx
->map_changing_lock
);
744 atomic_inc(&ctx
->mmap_changing
);
745 up_write(&ctx
->map_changing_lock
);
747 /* Drop uffd context if remap feature not enabled */
748 vma_start_write(vma
);
749 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
750 userfaultfd_set_vm_flags(vma
, vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
754 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
755 unsigned long from
, unsigned long to
,
758 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
759 struct userfaultfd_wait_queue ewq
;
764 if (to
& ~PAGE_MASK
) {
765 userfaultfd_ctx_put(ctx
);
771 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
772 ewq
.msg
.arg
.remap
.from
= from
;
773 ewq
.msg
.arg
.remap
.to
= to
;
774 ewq
.msg
.arg
.remap
.len
= len
;
776 userfaultfd_event_wait_completion(ctx
, &ewq
);
779 bool userfaultfd_remove(struct vm_area_struct
*vma
,
780 unsigned long start
, unsigned long end
)
782 struct mm_struct
*mm
= vma
->vm_mm
;
783 struct userfaultfd_ctx
*ctx
;
784 struct userfaultfd_wait_queue ewq
;
786 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
787 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
790 userfaultfd_ctx_get(ctx
);
791 down_write(&ctx
->map_changing_lock
);
792 atomic_inc(&ctx
->mmap_changing
);
793 up_write(&ctx
->map_changing_lock
);
794 mmap_read_unlock(mm
);
798 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
799 ewq
.msg
.arg
.remove
.start
= start
;
800 ewq
.msg
.arg
.remove
.end
= end
;
802 userfaultfd_event_wait_completion(ctx
, &ewq
);
807 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
808 unsigned long start
, unsigned long end
)
810 struct userfaultfd_unmap_ctx
*unmap_ctx
;
812 list_for_each_entry(unmap_ctx
, unmaps
, list
)
813 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
814 unmap_ctx
->end
== end
)
820 int userfaultfd_unmap_prep(struct vm_area_struct
*vma
, unsigned long start
,
821 unsigned long end
, struct list_head
*unmaps
)
823 struct userfaultfd_unmap_ctx
*unmap_ctx
;
824 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
826 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
827 has_unmap_ctx(ctx
, unmaps
, start
, end
))
830 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
834 userfaultfd_ctx_get(ctx
);
835 down_write(&ctx
->map_changing_lock
);
836 atomic_inc(&ctx
->mmap_changing
);
837 up_write(&ctx
->map_changing_lock
);
838 unmap_ctx
->ctx
= ctx
;
839 unmap_ctx
->start
= start
;
840 unmap_ctx
->end
= end
;
841 list_add_tail(&unmap_ctx
->list
, unmaps
);
846 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
848 struct userfaultfd_unmap_ctx
*ctx
, *n
;
849 struct userfaultfd_wait_queue ewq
;
851 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
854 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
855 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
856 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
858 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
860 list_del(&ctx
->list
);
865 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
867 struct userfaultfd_ctx
*ctx
= file
->private_data
;
868 struct mm_struct
*mm
= ctx
->mm
;
869 struct vm_area_struct
*vma
, *prev
;
870 /* len == 0 means wake all */
871 struct userfaultfd_wake_range range
= { .len
= 0, };
872 unsigned long new_flags
;
873 VMA_ITERATOR(vmi
, mm
, 0);
875 WRITE_ONCE(ctx
->released
, true);
877 if (!mmget_not_zero(mm
))
881 * Flush page faults out of all CPUs. NOTE: all page faults
882 * must be retried without returning VM_FAULT_SIGBUS if
883 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
884 * changes while handle_userfault released the mmap_lock. So
885 * it's critical that released is set to true (above), before
886 * taking the mmap_lock for writing.
890 for_each_vma(vmi
, vma
) {
892 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
893 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
894 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
898 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
899 vma
= vma_modify_flags_uffd(&vmi
, prev
, vma
, vma
->vm_start
,
900 vma
->vm_end
, new_flags
,
903 vma_start_write(vma
);
904 userfaultfd_set_vm_flags(vma
, new_flags
);
905 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
909 mmap_write_unlock(mm
);
913 * After no new page faults can wait on this fault_*wqh, flush
914 * the last page faults that may have been already waiting on
917 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
918 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
919 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
920 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
922 /* Flush pending events that may still wait on event_wqh */
923 wake_up_all(&ctx
->event_wqh
);
925 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
926 userfaultfd_ctx_put(ctx
);
930 /* fault_pending_wqh.lock must be hold by the caller */
931 static inline struct userfaultfd_wait_queue
*find_userfault_in(
932 wait_queue_head_t
*wqh
)
934 wait_queue_entry_t
*wq
;
935 struct userfaultfd_wait_queue
*uwq
;
937 lockdep_assert_held(&wqh
->lock
);
940 if (!waitqueue_active(wqh
))
942 /* walk in reverse to provide FIFO behavior to read userfaults */
943 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
944 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
949 static inline struct userfaultfd_wait_queue
*find_userfault(
950 struct userfaultfd_ctx
*ctx
)
952 return find_userfault_in(&ctx
->fault_pending_wqh
);
955 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
956 struct userfaultfd_ctx
*ctx
)
958 return find_userfault_in(&ctx
->event_wqh
);
961 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
963 struct userfaultfd_ctx
*ctx
= file
->private_data
;
966 poll_wait(file
, &ctx
->fd_wqh
, wait
);
968 if (!userfaultfd_is_initialized(ctx
))
972 * poll() never guarantees that read won't block.
973 * userfaults can be waken before they're read().
975 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
978 * lockless access to see if there are pending faults
979 * __pollwait last action is the add_wait_queue but
980 * the spin_unlock would allow the waitqueue_active to
981 * pass above the actual list_add inside
982 * add_wait_queue critical section. So use a full
983 * memory barrier to serialize the list_add write of
984 * add_wait_queue() with the waitqueue_active read
989 if (waitqueue_active(&ctx
->fault_pending_wqh
))
991 else if (waitqueue_active(&ctx
->event_wqh
))
997 static const struct file_operations userfaultfd_fops
;
999 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
1000 struct inode
*inode
,
1001 struct uffd_msg
*msg
)
1005 fd
= anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops
, new,
1006 O_RDONLY
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
1010 msg
->arg
.reserved
.reserved1
= 0;
1011 msg
->arg
.fork
.ufd
= fd
;
1015 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1016 struct uffd_msg
*msg
, struct inode
*inode
)
1019 DECLARE_WAITQUEUE(wait
, current
);
1020 struct userfaultfd_wait_queue
*uwq
;
1022 * Handling fork event requires sleeping operations, so
1023 * we drop the event_wqh lock, then do these ops, then
1024 * lock it back and wake up the waiter. While the lock is
1025 * dropped the ewq may go away so we keep track of it
1028 LIST_HEAD(fork_event
);
1029 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1031 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1032 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1033 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1035 set_current_state(TASK_INTERRUPTIBLE
);
1036 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1037 uwq
= find_userfault(ctx
);
1040 * Use a seqcount to repeat the lockless check
1041 * in wake_userfault() to avoid missing
1042 * wakeups because during the refile both
1043 * waitqueue could become empty if this is the
1046 write_seqcount_begin(&ctx
->refile_seq
);
1049 * The fault_pending_wqh.lock prevents the uwq
1050 * to disappear from under us.
1052 * Refile this userfault from
1053 * fault_pending_wqh to fault_wqh, it's not
1054 * pending anymore after we read it.
1056 * Use list_del() by hand (as
1057 * userfaultfd_wake_function also uses
1058 * list_del_init() by hand) to be sure nobody
1059 * changes __remove_wait_queue() to use
1060 * list_del_init() in turn breaking the
1061 * !list_empty_careful() check in
1062 * handle_userfault(). The uwq->wq.head list
1063 * must never be empty at any time during the
1064 * refile, or the waitqueue could disappear
1065 * from under us. The "wait_queue_head_t"
1066 * parameter of __remove_wait_queue() is unused
1069 list_del(&uwq
->wq
.entry
);
1070 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1072 write_seqcount_end(&ctx
->refile_seq
);
1074 /* careful to always initialize msg if ret == 0 */
1076 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1080 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1082 spin_lock(&ctx
->event_wqh
.lock
);
1083 uwq
= find_userfault_evt(ctx
);
1087 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1088 fork_nctx
= (struct userfaultfd_ctx
*)
1090 uwq
->msg
.arg
.reserved
.reserved1
;
1091 list_move(&uwq
->wq
.entry
, &fork_event
);
1093 * fork_nctx can be freed as soon as
1094 * we drop the lock, unless we take a
1097 userfaultfd_ctx_get(fork_nctx
);
1098 spin_unlock(&ctx
->event_wqh
.lock
);
1103 userfaultfd_event_complete(ctx
, uwq
);
1104 spin_unlock(&ctx
->event_wqh
.lock
);
1108 spin_unlock(&ctx
->event_wqh
.lock
);
1110 if (signal_pending(current
)) {
1118 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1120 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1122 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1123 __set_current_state(TASK_RUNNING
);
1124 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1126 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1127 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1128 spin_lock_irq(&ctx
->event_wqh
.lock
);
1129 if (!list_empty(&fork_event
)) {
1131 * The fork thread didn't abort, so we can
1132 * drop the temporary refcount.
1134 userfaultfd_ctx_put(fork_nctx
);
1136 uwq
= list_first_entry(&fork_event
,
1140 * If fork_event list wasn't empty and in turn
1141 * the event wasn't already released by fork
1142 * (the event is allocated on fork kernel
1143 * stack), put the event back to its place in
1144 * the event_wq. fork_event head will be freed
1145 * as soon as we return so the event cannot
1146 * stay queued there no matter the current
1149 list_del(&uwq
->wq
.entry
);
1150 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1153 * Leave the event in the waitqueue and report
1154 * error to userland if we failed to resolve
1155 * the userfault fork.
1158 userfaultfd_event_complete(ctx
, uwq
);
1161 * Here the fork thread aborted and the
1162 * refcount from the fork thread on fork_nctx
1163 * has already been released. We still hold
1164 * the reference we took before releasing the
1165 * lock above. If resolve_userfault_fork
1166 * failed we've to drop it because the
1167 * fork_nctx has to be freed in such case. If
1168 * it succeeded we'll hold it because the new
1169 * uffd references it.
1172 userfaultfd_ctx_put(fork_nctx
);
1174 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1180 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1181 size_t count
, loff_t
*ppos
)
1183 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1184 ssize_t _ret
, ret
= 0;
1185 struct uffd_msg msg
;
1186 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1187 struct inode
*inode
= file_inode(file
);
1189 if (!userfaultfd_is_initialized(ctx
))
1193 if (count
< sizeof(msg
))
1194 return ret
? ret
: -EINVAL
;
1195 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1197 return ret
? ret
: _ret
;
1198 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1199 return ret
? ret
: -EFAULT
;
1202 count
-= sizeof(msg
);
1204 * Allow to read more than one fault at time but only
1205 * block if waiting for the very first one.
1207 no_wait
= O_NONBLOCK
;
1211 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1212 struct userfaultfd_wake_range
*range
)
1214 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1215 /* wake all in the range and autoremove */
1216 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1217 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1219 if (waitqueue_active(&ctx
->fault_wqh
))
1220 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1221 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1224 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1225 struct userfaultfd_wake_range
*range
)
1231 * To be sure waitqueue_active() is not reordered by the CPU
1232 * before the pagetable update, use an explicit SMP memory
1233 * barrier here. PT lock release or mmap_read_unlock(mm) still
1234 * have release semantics that can allow the
1235 * waitqueue_active() to be reordered before the pte update.
1240 * Use waitqueue_active because it's very frequent to
1241 * change the address space atomically even if there are no
1242 * userfaults yet. So we take the spinlock only when we're
1243 * sure we've userfaults to wake.
1246 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1247 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1248 waitqueue_active(&ctx
->fault_wqh
);
1250 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1252 __wake_userfault(ctx
, range
);
1255 static __always_inline
int validate_unaligned_range(
1256 struct mm_struct
*mm
, __u64 start
, __u64 len
)
1258 __u64 task_size
= mm
->task_size
;
1260 if (len
& ~PAGE_MASK
)
1264 if (start
< mmap_min_addr
)
1266 if (start
>= task_size
)
1268 if (len
> task_size
- start
)
1270 if (start
+ len
<= start
)
1275 static __always_inline
int validate_range(struct mm_struct
*mm
,
1276 __u64 start
, __u64 len
)
1278 if (start
& ~PAGE_MASK
)
1281 return validate_unaligned_range(mm
, start
, len
);
1284 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1287 struct mm_struct
*mm
= ctx
->mm
;
1288 struct vm_area_struct
*vma
, *prev
, *cur
;
1290 struct uffdio_register uffdio_register
;
1291 struct uffdio_register __user
*user_uffdio_register
;
1292 unsigned long vm_flags
, new_flags
;
1295 unsigned long start
, end
, vma_end
;
1296 struct vma_iterator vmi
;
1297 bool wp_async
= userfaultfd_wp_async_ctx(ctx
);
1299 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1302 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1303 sizeof(uffdio_register
)-sizeof(__u64
)))
1307 if (!uffdio_register
.mode
)
1309 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1312 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1313 vm_flags
|= VM_UFFD_MISSING
;
1314 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1315 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1318 vm_flags
|= VM_UFFD_WP
;
1320 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1321 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1324 vm_flags
|= VM_UFFD_MINOR
;
1327 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1328 uffdio_register
.range
.len
);
1332 start
= uffdio_register
.range
.start
;
1333 end
= start
+ uffdio_register
.range
.len
;
1336 if (!mmget_not_zero(mm
))
1340 mmap_write_lock(mm
);
1341 vma_iter_init(&vmi
, mm
, start
);
1342 vma
= vma_find(&vmi
, end
);
1347 * If the first vma contains huge pages, make sure start address
1348 * is aligned to huge page size.
1350 if (is_vm_hugetlb_page(vma
)) {
1351 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1353 if (start
& (vma_hpagesize
- 1))
1358 * Search for not compatible vmas.
1361 basic_ioctls
= false;
1366 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1367 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1369 /* check not compatible vmas */
1371 if (!vma_can_userfault(cur
, vm_flags
, wp_async
))
1375 * UFFDIO_COPY will fill file holes even without
1376 * PROT_WRITE. This check enforces that if this is a
1377 * MAP_SHARED, the process has write permission to the backing
1378 * file. If VM_MAYWRITE is set it also enforces that on a
1379 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1380 * F_WRITE_SEAL can be taken until the vma is destroyed.
1383 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1387 * If this vma contains ending address, and huge pages
1390 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1391 end
> cur
->vm_start
) {
1392 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1396 if (end
& (vma_hpagesize
- 1))
1399 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1403 * Check that this vma isn't already owned by a
1404 * different userfaultfd. We can't allow more than one
1405 * userfaultfd to own a single vma simultaneously or we
1406 * wouldn't know which one to deliver the userfaults to.
1409 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1410 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1414 * Note vmas containing huge pages
1416 if (is_vm_hugetlb_page(cur
))
1417 basic_ioctls
= true;
1420 } for_each_vma_range(vmi
, cur
, end
);
1423 vma_iter_set(&vmi
, start
);
1424 prev
= vma_prev(&vmi
);
1425 if (vma
->vm_start
< start
)
1429 for_each_vma_range(vmi
, vma
, end
) {
1432 BUG_ON(!vma_can_userfault(vma
, vm_flags
, wp_async
));
1433 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1434 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1435 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1438 * Nothing to do: this vma is already registered into this
1439 * userfaultfd and with the right tracking mode too.
1441 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1442 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1445 if (vma
->vm_start
> start
)
1446 start
= vma
->vm_start
;
1447 vma_end
= min(end
, vma
->vm_end
);
1449 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1450 vma
= vma_modify_flags_uffd(&vmi
, prev
, vma
, start
, vma_end
,
1452 (struct vm_userfaultfd_ctx
){ctx
});
1459 * In the vma_merge() successful mprotect-like case 8:
1460 * the next vma was merged into the current one and
1461 * the current one has not been updated yet.
1463 vma_start_write(vma
);
1464 userfaultfd_set_vm_flags(vma
, 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
;
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
;
1518 struct vma_iterator vmi
;
1519 bool wp_async
= userfaultfd_wp_async_ctx(ctx
);
1522 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1525 ret
= validate_range(mm
, uffdio_unregister
.start
,
1526 uffdio_unregister
.len
);
1530 start
= uffdio_unregister
.start
;
1531 end
= start
+ uffdio_unregister
.len
;
1534 if (!mmget_not_zero(mm
))
1537 mmap_write_lock(mm
);
1539 vma_iter_init(&vmi
, mm
, start
);
1540 vma
= vma_find(&vmi
, end
);
1545 * If the first vma contains huge pages, make sure start address
1546 * is aligned to huge page size.
1548 if (is_vm_hugetlb_page(vma
)) {
1549 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1551 if (start
& (vma_hpagesize
- 1))
1556 * Search for not compatible vmas.
1563 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1564 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1567 * Check not compatible vmas, not strictly required
1568 * here as not compatible vmas cannot have an
1569 * userfaultfd_ctx registered on them, but this
1570 * provides for more strict behavior to notice
1571 * unregistration errors.
1573 if (!vma_can_userfault(cur
, cur
->vm_flags
, wp_async
))
1577 } for_each_vma_range(vmi
, cur
, end
);
1580 vma_iter_set(&vmi
, start
);
1581 prev
= vma_prev(&vmi
);
1582 if (vma
->vm_start
< start
)
1586 for_each_vma_range(vmi
, vma
, end
) {
1589 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
, wp_async
));
1592 * Nothing to do: this vma is already registered into this
1593 * userfaultfd and with the right tracking mode too.
1595 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1598 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1600 if (vma
->vm_start
> start
)
1601 start
= vma
->vm_start
;
1602 vma_end
= min(end
, vma
->vm_end
);
1604 if (userfaultfd_missing(vma
)) {
1606 * Wake any concurrent pending userfault while
1607 * we unregister, so they will not hang
1608 * permanently and it avoids userland to call
1609 * UFFDIO_WAKE explicitly.
1611 struct userfaultfd_wake_range range
;
1612 range
.start
= start
;
1613 range
.len
= vma_end
- start
;
1614 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1617 /* Reset ptes for the whole vma range if wr-protected */
1618 if (userfaultfd_wp(vma
))
1619 uffd_wp_range(vma
, start
, vma_end
- start
, false);
1621 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1622 vma
= vma_modify_flags_uffd(&vmi
, prev
, vma
, start
, vma_end
,
1623 new_flags
, NULL_VM_UFFD_CTX
);
1630 * In the vma_merge() successful mprotect-like case 8:
1631 * the next vma was merged into the current one and
1632 * the current one has not been updated yet.
1634 vma_start_write(vma
);
1635 userfaultfd_set_vm_flags(vma
, new_flags
);
1636 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1640 start
= vma
->vm_end
;
1644 mmap_write_unlock(mm
);
1651 * userfaultfd_wake may be used in combination with the
1652 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1654 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1658 struct uffdio_range uffdio_wake
;
1659 struct userfaultfd_wake_range range
;
1660 const void __user
*buf
= (void __user
*)arg
;
1663 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1666 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1670 range
.start
= uffdio_wake
.start
;
1671 range
.len
= uffdio_wake
.len
;
1674 * len == 0 means wake all and we don't want to wake all here,
1675 * so check it again to be sure.
1677 VM_BUG_ON(!range
.len
);
1679 wake_userfault(ctx
, &range
);
1686 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1690 struct uffdio_copy uffdio_copy
;
1691 struct uffdio_copy __user
*user_uffdio_copy
;
1692 struct userfaultfd_wake_range range
;
1693 uffd_flags_t flags
= 0;
1695 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1698 if (atomic_read(&ctx
->mmap_changing
))
1702 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1703 /* don't copy "copy" last field */
1704 sizeof(uffdio_copy
)-sizeof(__s64
)))
1707 ret
= validate_unaligned_range(ctx
->mm
, uffdio_copy
.src
,
1711 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1716 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1718 if (uffdio_copy
.mode
& UFFDIO_COPY_MODE_WP
)
1719 flags
|= MFILL_ATOMIC_WP
;
1720 if (mmget_not_zero(ctx
->mm
)) {
1721 ret
= mfill_atomic_copy(ctx
, uffdio_copy
.dst
, uffdio_copy
.src
,
1722 uffdio_copy
.len
, flags
);
1727 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1732 /* len == 0 would wake all */
1734 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1735 range
.start
= uffdio_copy
.dst
;
1736 wake_userfault(ctx
, &range
);
1738 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1743 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1747 struct uffdio_zeropage uffdio_zeropage
;
1748 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1749 struct userfaultfd_wake_range range
;
1751 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1754 if (atomic_read(&ctx
->mmap_changing
))
1758 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1759 /* don't copy "zeropage" last field */
1760 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1763 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1764 uffdio_zeropage
.range
.len
);
1768 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1771 if (mmget_not_zero(ctx
->mm
)) {
1772 ret
= mfill_atomic_zeropage(ctx
, uffdio_zeropage
.range
.start
,
1773 uffdio_zeropage
.range
.len
);
1778 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1782 /* len == 0 would wake all */
1785 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1786 range
.start
= uffdio_zeropage
.range
.start
;
1787 wake_userfault(ctx
, &range
);
1789 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1794 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1798 struct uffdio_writeprotect uffdio_wp
;
1799 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1800 struct userfaultfd_wake_range range
;
1801 bool mode_wp
, mode_dontwake
;
1803 if (atomic_read(&ctx
->mmap_changing
))
1806 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1808 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1809 sizeof(struct uffdio_writeprotect
)))
1812 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1813 uffdio_wp
.range
.len
);
1817 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1818 UFFDIO_WRITEPROTECT_MODE_WP
))
1821 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1822 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1824 if (mode_wp
&& mode_dontwake
)
1827 if (mmget_not_zero(ctx
->mm
)) {
1828 ret
= mwriteprotect_range(ctx
, uffdio_wp
.range
.start
,
1829 uffdio_wp
.range
.len
, mode_wp
);
1838 if (!mode_wp
&& !mode_dontwake
) {
1839 range
.start
= uffdio_wp
.range
.start
;
1840 range
.len
= uffdio_wp
.range
.len
;
1841 wake_userfault(ctx
, &range
);
1846 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1849 struct uffdio_continue uffdio_continue
;
1850 struct uffdio_continue __user
*user_uffdio_continue
;
1851 struct userfaultfd_wake_range range
;
1852 uffd_flags_t flags
= 0;
1854 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1857 if (atomic_read(&ctx
->mmap_changing
))
1861 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1862 /* don't copy the output fields */
1863 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1866 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1867 uffdio_continue
.range
.len
);
1872 if (uffdio_continue
.mode
& ~(UFFDIO_CONTINUE_MODE_DONTWAKE
|
1873 UFFDIO_CONTINUE_MODE_WP
))
1875 if (uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_WP
)
1876 flags
|= MFILL_ATOMIC_WP
;
1878 if (mmget_not_zero(ctx
->mm
)) {
1879 ret
= mfill_atomic_continue(ctx
, uffdio_continue
.range
.start
,
1880 uffdio_continue
.range
.len
, flags
);
1886 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1891 /* len == 0 would wake all */
1894 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1895 range
.start
= uffdio_continue
.range
.start
;
1896 wake_userfault(ctx
, &range
);
1898 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1904 static inline int userfaultfd_poison(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1907 struct uffdio_poison uffdio_poison
;
1908 struct uffdio_poison __user
*user_uffdio_poison
;
1909 struct userfaultfd_wake_range range
;
1911 user_uffdio_poison
= (struct uffdio_poison __user
*)arg
;
1914 if (atomic_read(&ctx
->mmap_changing
))
1918 if (copy_from_user(&uffdio_poison
, user_uffdio_poison
,
1919 /* don't copy the output fields */
1920 sizeof(uffdio_poison
) - (sizeof(__s64
))))
1923 ret
= validate_range(ctx
->mm
, uffdio_poison
.range
.start
,
1924 uffdio_poison
.range
.len
);
1929 if (uffdio_poison
.mode
& ~UFFDIO_POISON_MODE_DONTWAKE
)
1932 if (mmget_not_zero(ctx
->mm
)) {
1933 ret
= mfill_atomic_poison(ctx
, uffdio_poison
.range
.start
,
1934 uffdio_poison
.range
.len
, 0);
1940 if (unlikely(put_user(ret
, &user_uffdio_poison
->updated
)))
1945 /* len == 0 would wake all */
1948 if (!(uffdio_poison
.mode
& UFFDIO_POISON_MODE_DONTWAKE
)) {
1949 range
.start
= uffdio_poison
.range
.start
;
1950 wake_userfault(ctx
, &range
);
1952 ret
= range
.len
== uffdio_poison
.range
.len
? 0 : -EAGAIN
;
1958 bool userfaultfd_wp_async(struct vm_area_struct
*vma
)
1960 return userfaultfd_wp_async_ctx(vma
->vm_userfaultfd_ctx
.ctx
);
1963 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1966 * For the current set of features the bits just coincide. Set
1967 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1969 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1972 static int userfaultfd_move(struct userfaultfd_ctx
*ctx
,
1976 struct uffdio_move uffdio_move
;
1977 struct uffdio_move __user
*user_uffdio_move
;
1978 struct userfaultfd_wake_range range
;
1979 struct mm_struct
*mm
= ctx
->mm
;
1981 user_uffdio_move
= (struct uffdio_move __user
*) arg
;
1983 if (atomic_read(&ctx
->mmap_changing
))
1986 if (copy_from_user(&uffdio_move
, user_uffdio_move
,
1987 /* don't copy "move" last field */
1988 sizeof(uffdio_move
)-sizeof(__s64
)))
1991 /* Do not allow cross-mm moves. */
1992 if (mm
!= current
->mm
)
1995 ret
= validate_range(mm
, uffdio_move
.dst
, uffdio_move
.len
);
1999 ret
= validate_range(mm
, uffdio_move
.src
, uffdio_move
.len
);
2003 if (uffdio_move
.mode
& ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES
|
2004 UFFDIO_MOVE_MODE_DONTWAKE
))
2007 if (mmget_not_zero(mm
)) {
2008 ret
= move_pages(ctx
, uffdio_move
.dst
, uffdio_move
.src
,
2009 uffdio_move
.len
, uffdio_move
.mode
);
2015 if (unlikely(put_user(ret
, &user_uffdio_move
->move
)))
2020 /* len == 0 would wake all */
2023 if (!(uffdio_move
.mode
& UFFDIO_MOVE_MODE_DONTWAKE
)) {
2024 range
.start
= uffdio_move
.dst
;
2025 wake_userfault(ctx
, &range
);
2027 ret
= range
.len
== uffdio_move
.len
? 0 : -EAGAIN
;
2034 * userland asks for a certain API version and we return which bits
2035 * and ioctl commands are implemented in this kernel for such API
2036 * version or -EINVAL if unknown.
2038 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
2041 struct uffdio_api uffdio_api
;
2042 void __user
*buf
= (void __user
*)arg
;
2043 unsigned int ctx_features
;
2048 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
2050 features
= uffdio_api
.features
;
2052 if (uffdio_api
.api
!= UFFD_API
|| (features
& ~UFFD_API_FEATURES
))
2055 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
2058 /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2059 if (features
& UFFD_FEATURE_WP_ASYNC
)
2060 features
|= UFFD_FEATURE_WP_UNPOPULATED
;
2062 /* report all available features and ioctls to userland */
2063 uffdio_api
.features
= UFFD_API_FEATURES
;
2064 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2065 uffdio_api
.features
&=
2066 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
2068 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2069 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
2071 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2072 uffdio_api
.features
&= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM
;
2073 uffdio_api
.features
&= ~UFFD_FEATURE_WP_UNPOPULATED
;
2074 uffdio_api
.features
&= ~UFFD_FEATURE_WP_ASYNC
;
2076 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
2078 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
2081 /* only enable the requested features for this uffd context */
2082 ctx_features
= uffd_ctx_features(features
);
2084 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
2091 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
2092 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
2097 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
2101 struct userfaultfd_ctx
*ctx
= file
->private_data
;
2103 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
2108 ret
= userfaultfd_api(ctx
, arg
);
2110 case UFFDIO_REGISTER
:
2111 ret
= userfaultfd_register(ctx
, arg
);
2113 case UFFDIO_UNREGISTER
:
2114 ret
= userfaultfd_unregister(ctx
, arg
);
2117 ret
= userfaultfd_wake(ctx
, arg
);
2120 ret
= userfaultfd_copy(ctx
, arg
);
2122 case UFFDIO_ZEROPAGE
:
2123 ret
= userfaultfd_zeropage(ctx
, arg
);
2126 ret
= userfaultfd_move(ctx
, arg
);
2128 case UFFDIO_WRITEPROTECT
:
2129 ret
= userfaultfd_writeprotect(ctx
, arg
);
2131 case UFFDIO_CONTINUE
:
2132 ret
= userfaultfd_continue(ctx
, arg
);
2135 ret
= userfaultfd_poison(ctx
, arg
);
2141 #ifdef CONFIG_PROC_FS
2142 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2144 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2145 wait_queue_entry_t
*wq
;
2146 unsigned long pending
= 0, total
= 0;
2148 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2149 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2153 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2156 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2159 * If more protocols will be added, there will be all shown
2160 * separated by a space. Like this:
2161 * protocols: aa:... bb:...
2163 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2164 pending
, total
, UFFD_API
, ctx
->features
,
2165 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2169 static const struct file_operations userfaultfd_fops
= {
2170 #ifdef CONFIG_PROC_FS
2171 .show_fdinfo
= userfaultfd_show_fdinfo
,
2173 .release
= userfaultfd_release
,
2174 .poll
= userfaultfd_poll
,
2175 .read
= userfaultfd_read
,
2176 .unlocked_ioctl
= userfaultfd_ioctl
,
2177 .compat_ioctl
= compat_ptr_ioctl
,
2178 .llseek
= noop_llseek
,
2181 static void init_once_userfaultfd_ctx(void *mem
)
2183 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2185 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2186 init_waitqueue_head(&ctx
->fault_wqh
);
2187 init_waitqueue_head(&ctx
->event_wqh
);
2188 init_waitqueue_head(&ctx
->fd_wqh
);
2189 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2192 static int new_userfaultfd(int flags
)
2194 struct userfaultfd_ctx
*ctx
;
2197 BUG_ON(!current
->mm
);
2199 /* Check the UFFD_* constants for consistency. */
2200 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2201 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2202 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2204 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2207 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2211 refcount_set(&ctx
->refcount
, 1);
2214 ctx
->released
= false;
2215 init_rwsem(&ctx
->map_changing_lock
);
2216 atomic_set(&ctx
->mmap_changing
, 0);
2217 ctx
->mm
= current
->mm
;
2218 /* prevent the mm struct to be freed */
2221 /* Create a new inode so that the LSM can block the creation. */
2222 fd
= anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops
, ctx
,
2223 O_RDONLY
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2226 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2231 static inline bool userfaultfd_syscall_allowed(int flags
)
2233 /* Userspace-only page faults are always allowed */
2234 if (flags
& UFFD_USER_MODE_ONLY
)
2238 * The user is requesting a userfaultfd which can handle kernel faults.
2239 * Privileged users are always allowed to do this.
2241 if (capable(CAP_SYS_PTRACE
))
2244 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2245 return sysctl_unprivileged_userfaultfd
;
2248 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2250 if (!userfaultfd_syscall_allowed(flags
))
2253 return new_userfaultfd(flags
);
2256 static long userfaultfd_dev_ioctl(struct file
*file
, unsigned int cmd
, unsigned long flags
)
2258 if (cmd
!= USERFAULTFD_IOC_NEW
)
2261 return new_userfaultfd(flags
);
2264 static const struct file_operations userfaultfd_dev_fops
= {
2265 .unlocked_ioctl
= userfaultfd_dev_ioctl
,
2266 .compat_ioctl
= userfaultfd_dev_ioctl
,
2267 .owner
= THIS_MODULE
,
2268 .llseek
= noop_llseek
,
2271 static struct miscdevice userfaultfd_misc
= {
2272 .minor
= MISC_DYNAMIC_MINOR
,
2273 .name
= "userfaultfd",
2274 .fops
= &userfaultfd_dev_fops
2277 static int __init
userfaultfd_init(void)
2281 ret
= misc_register(&userfaultfd_misc
);
2285 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2286 sizeof(struct userfaultfd_ctx
),
2288 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2289 init_once_userfaultfd_ctx
);
2290 #ifdef CONFIG_SYSCTL
2291 register_sysctl_init("vm", vm_userfaultfd_table
);
2295 __initcall(userfaultfd_init
);