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 int sysctl_unprivileged_userfaultfd __read_mostly
;
37 static struct kmem_cache
*userfaultfd_ctx_cachep __read_mostly
;
40 * Start with fault_pending_wqh and fault_wqh so they're more likely
41 * to be in the same cacheline.
45 * fault_pending_wqh.lock
49 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
50 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
51 * also taken in IRQ context.
53 struct userfaultfd_ctx
{
54 /* waitqueue head for the pending (i.e. not read) userfaults */
55 wait_queue_head_t fault_pending_wqh
;
56 /* waitqueue head for the userfaults */
57 wait_queue_head_t fault_wqh
;
58 /* waitqueue head for the pseudo fd to wakeup poll/read */
59 wait_queue_head_t fd_wqh
;
60 /* waitqueue head for events */
61 wait_queue_head_t event_wqh
;
62 /* a refile sequence protected by fault_pending_wqh lock */
63 seqcount_spinlock_t refile_seq
;
64 /* pseudo fd refcounting */
66 /* userfaultfd syscall flags */
68 /* features requested from the userspace */
69 unsigned int features
;
72 /* memory mappings are changing because of non-cooperative event */
73 atomic_t mmap_changing
;
74 /* mm with one ore more vmas attached to this userfaultfd_ctx */
78 struct userfaultfd_fork_ctx
{
79 struct userfaultfd_ctx
*orig
;
80 struct userfaultfd_ctx
*new;
81 struct list_head list
;
84 struct userfaultfd_unmap_ctx
{
85 struct userfaultfd_ctx
*ctx
;
88 struct list_head list
;
91 struct userfaultfd_wait_queue
{
93 wait_queue_entry_t wq
;
94 struct userfaultfd_ctx
*ctx
;
98 struct userfaultfd_wake_range
{
103 /* internal indication that UFFD_API ioctl was successfully executed */
104 #define UFFD_FEATURE_INITIALIZED (1u << 31)
106 static bool userfaultfd_is_initialized(struct userfaultfd_ctx
*ctx
)
108 return ctx
->features
& UFFD_FEATURE_INITIALIZED
;
111 static void userfaultfd_set_vm_flags(struct vm_area_struct
*vma
,
114 const bool uffd_wp_changed
= (vma
->vm_flags
^ flags
) & VM_UFFD_WP
;
116 vma
->vm_flags
= flags
;
118 * For shared mappings, we want to enable writenotify while
119 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
120 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
122 if ((vma
->vm_flags
& VM_SHARED
) && uffd_wp_changed
)
123 vma_set_page_prot(vma
);
126 static int userfaultfd_wake_function(wait_queue_entry_t
*wq
, unsigned mode
,
127 int wake_flags
, void *key
)
129 struct userfaultfd_wake_range
*range
= key
;
131 struct userfaultfd_wait_queue
*uwq
;
132 unsigned long start
, len
;
134 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
136 /* len == 0 means wake all */
137 start
= range
->start
;
139 if (len
&& (start
> uwq
->msg
.arg
.pagefault
.address
||
140 start
+ len
<= uwq
->msg
.arg
.pagefault
.address
))
142 WRITE_ONCE(uwq
->waken
, true);
144 * The Program-Order guarantees provided by the scheduler
145 * ensure uwq->waken is visible before the task is woken.
147 ret
= wake_up_state(wq
->private, mode
);
150 * Wake only once, autoremove behavior.
152 * After the effect of list_del_init is visible to the other
153 * CPUs, the waitqueue may disappear from under us, see the
154 * !list_empty_careful() in handle_userfault().
156 * try_to_wake_up() has an implicit smp_mb(), and the
157 * wq->private is read before calling the extern function
158 * "wake_up_state" (which in turns calls try_to_wake_up).
160 list_del_init(&wq
->entry
);
167 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
169 * @ctx: [in] Pointer to the userfaultfd context.
171 static void userfaultfd_ctx_get(struct userfaultfd_ctx
*ctx
)
173 refcount_inc(&ctx
->refcount
);
177 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
179 * @ctx: [in] Pointer to userfaultfd context.
181 * The userfaultfd context reference must have been previously acquired either
182 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
184 static void userfaultfd_ctx_put(struct userfaultfd_ctx
*ctx
)
186 if (refcount_dec_and_test(&ctx
->refcount
)) {
187 VM_BUG_ON(spin_is_locked(&ctx
->fault_pending_wqh
.lock
));
188 VM_BUG_ON(waitqueue_active(&ctx
->fault_pending_wqh
));
189 VM_BUG_ON(spin_is_locked(&ctx
->fault_wqh
.lock
));
190 VM_BUG_ON(waitqueue_active(&ctx
->fault_wqh
));
191 VM_BUG_ON(spin_is_locked(&ctx
->event_wqh
.lock
));
192 VM_BUG_ON(waitqueue_active(&ctx
->event_wqh
));
193 VM_BUG_ON(spin_is_locked(&ctx
->fd_wqh
.lock
));
194 VM_BUG_ON(waitqueue_active(&ctx
->fd_wqh
));
196 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
200 static inline void msg_init(struct uffd_msg
*msg
)
202 BUILD_BUG_ON(sizeof(struct uffd_msg
) != 32);
204 * Must use memset to zero out the paddings or kernel data is
205 * leaked to userland.
207 memset(msg
, 0, sizeof(struct uffd_msg
));
210 static inline struct uffd_msg
userfault_msg(unsigned long address
,
211 unsigned long real_address
,
213 unsigned long reason
,
214 unsigned int features
)
219 msg
.event
= UFFD_EVENT_PAGEFAULT
;
221 msg
.arg
.pagefault
.address
= (features
& UFFD_FEATURE_EXACT_ADDRESS
) ?
222 real_address
: address
;
225 * These flags indicate why the userfault occurred:
226 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
227 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
228 * - Neither of these flags being set indicates a MISSING fault.
230 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
231 * fault. Otherwise, it was a read fault.
233 if (flags
& FAULT_FLAG_WRITE
)
234 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WRITE
;
235 if (reason
& VM_UFFD_WP
)
236 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_WP
;
237 if (reason
& VM_UFFD_MINOR
)
238 msg
.arg
.pagefault
.flags
|= UFFD_PAGEFAULT_FLAG_MINOR
;
239 if (features
& UFFD_FEATURE_THREAD_ID
)
240 msg
.arg
.pagefault
.feat
.ptid
= task_pid_vnr(current
);
244 #ifdef CONFIG_HUGETLB_PAGE
246 * Same functionality as userfaultfd_must_wait below with modifications for
249 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
250 struct vm_area_struct
*vma
,
251 unsigned long address
,
253 unsigned long reason
)
255 struct mm_struct
*mm
= ctx
->mm
;
259 mmap_assert_locked(mm
);
261 ptep
= huge_pte_offset(mm
, address
, vma_mmu_pagesize(vma
));
267 pte
= huge_ptep_get(ptep
);
270 * Lockless access: we're in a wait_event so it's ok if it
271 * changes under us. PTE markers should be handled the same as none
274 if (huge_pte_none_mostly(pte
))
276 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
282 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
283 struct vm_area_struct
*vma
,
284 unsigned long address
,
286 unsigned long reason
)
288 return false; /* should never get here */
290 #endif /* CONFIG_HUGETLB_PAGE */
293 * Verify the pagetables are still not ok after having reigstered into
294 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
295 * userfault that has already been resolved, if userfaultfd_read and
296 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
299 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
300 unsigned long address
,
302 unsigned long reason
)
304 struct mm_struct
*mm
= ctx
->mm
;
312 mmap_assert_locked(mm
);
314 pgd
= pgd_offset(mm
, address
);
315 if (!pgd_present(*pgd
))
317 p4d
= p4d_offset(pgd
, address
);
318 if (!p4d_present(*p4d
))
320 pud
= pud_offset(p4d
, address
);
321 if (!pud_present(*pud
))
323 pmd
= pmd_offset(pud
, address
);
325 * READ_ONCE must function as a barrier with narrower scope
326 * and it must be equivalent to:
327 * _pmd = *pmd; barrier();
329 * This is to deal with the instability (as in
330 * pmd_trans_unstable) of the pmd.
332 _pmd
= READ_ONCE(*pmd
);
337 if (!pmd_present(_pmd
))
340 if (pmd_trans_huge(_pmd
)) {
341 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
347 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
348 * and use the standard pte_offset_map() instead of parsing _pmd.
350 pte
= pte_offset_map(pmd
, address
);
352 * Lockless access: we're in a wait_event so it's ok if it
353 * changes under us. PTE markers should be handled the same as none
356 if (pte_none_mostly(*pte
))
358 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
366 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
368 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
369 return TASK_INTERRUPTIBLE
;
371 if (flags
& FAULT_FLAG_KILLABLE
)
372 return TASK_KILLABLE
;
374 return TASK_UNINTERRUPTIBLE
;
378 * The locking rules involved in returning VM_FAULT_RETRY depending on
379 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
380 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
381 * recommendation in __lock_page_or_retry is not an understatement.
383 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
384 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
387 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
388 * set, VM_FAULT_RETRY can still be returned if and only if there are
389 * fatal_signal_pending()s, and the mmap_lock must be released before
392 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
394 struct mm_struct
*mm
= vmf
->vma
->vm_mm
;
395 struct userfaultfd_ctx
*ctx
;
396 struct userfaultfd_wait_queue uwq
;
397 vm_fault_t ret
= VM_FAULT_SIGBUS
;
399 unsigned int blocking_state
;
402 * We don't do userfault handling for the final child pid update.
404 * We also don't do userfault handling during
405 * coredumping. hugetlbfs has the special
406 * follow_hugetlb_page() to skip missing pages in the
407 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
408 * the no_page_table() helper in follow_page_mask(), but the
409 * shmem_vm_ops->fault method is invoked even during
410 * coredumping without mmap_lock and it ends up here.
412 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
416 * Coredumping runs without mmap_lock so we can only check that
417 * the mmap_lock is held, if PF_DUMPCORE was not set.
419 mmap_assert_locked(mm
);
421 ctx
= vmf
->vma
->vm_userfaultfd_ctx
.ctx
;
425 BUG_ON(ctx
->mm
!= mm
);
427 /* Any unrecognized flag is a bug. */
428 VM_BUG_ON(reason
& ~__VM_UFFD_FLAGS
);
429 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
430 VM_BUG_ON(!reason
|| (reason
& (reason
- 1)));
432 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
434 if (!(vmf
->flags
& FAULT_FLAG_USER
) && (ctx
->flags
& UFFD_USER_MODE_ONLY
))
438 * If it's already released don't get it. This avoids to loop
439 * in __get_user_pages if userfaultfd_release waits on the
440 * caller of handle_userfault to release the mmap_lock.
442 if (unlikely(READ_ONCE(ctx
->released
))) {
444 * Don't return VM_FAULT_SIGBUS in this case, so a non
445 * cooperative manager can close the uffd after the
446 * last UFFDIO_COPY, without risking to trigger an
447 * involuntary SIGBUS if the process was starting the
448 * userfaultfd while the userfaultfd was still armed
449 * (but after the last UFFDIO_COPY). If the uffd
450 * wasn't already closed when the userfault reached
451 * this point, that would normally be solved by
452 * userfaultfd_must_wait returning 'false'.
454 * If we were to return VM_FAULT_SIGBUS here, the non
455 * cooperative manager would be instead forced to
456 * always call UFFDIO_UNREGISTER before it can safely
459 ret
= VM_FAULT_NOPAGE
;
464 * Check that we can return VM_FAULT_RETRY.
466 * NOTE: it should become possible to return VM_FAULT_RETRY
467 * even if FAULT_FLAG_TRIED is set without leading to gup()
468 * -EBUSY failures, if the userfaultfd is to be extended for
469 * VM_UFFD_WP tracking and we intend to arm the userfault
470 * without first stopping userland access to the memory. For
471 * VM_UFFD_MISSING userfaults this is enough for now.
473 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
475 * Validate the invariant that nowait must allow retry
476 * to be sure not to return SIGBUS erroneously on
477 * nowait invocations.
479 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
480 #ifdef CONFIG_DEBUG_VM
481 if (printk_ratelimit()) {
483 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
492 * Handle nowait, not much to do other than tell it to retry
495 ret
= VM_FAULT_RETRY
;
496 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
499 /* take the reference before dropping the mmap_lock */
500 userfaultfd_ctx_get(ctx
);
502 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
503 uwq
.wq
.private = current
;
504 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->real_address
, vmf
->flags
,
505 reason
, ctx
->features
);
509 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
511 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
513 * After the __add_wait_queue the uwq is visible to userland
514 * through poll/read().
516 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
518 * The smp_mb() after __set_current_state prevents the reads
519 * following the spin_unlock to happen before the list_add in
522 set_current_state(blocking_state
);
523 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
525 if (!is_vm_hugetlb_page(vmf
->vma
))
526 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
529 must_wait
= userfaultfd_huge_must_wait(ctx
, vmf
->vma
,
532 mmap_read_unlock(mm
);
534 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
535 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
539 __set_current_state(TASK_RUNNING
);
542 * Here we race with the list_del; list_add in
543 * userfaultfd_ctx_read(), however because we don't ever run
544 * list_del_init() to refile across the two lists, the prev
545 * and next pointers will never point to self. list_add also
546 * would never let any of the two pointers to point to
547 * self. So list_empty_careful won't risk to see both pointers
548 * pointing to self at any time during the list refile. The
549 * only case where list_del_init() is called is the full
550 * removal in the wake function and there we don't re-list_add
551 * and it's fine not to block on the spinlock. The uwq on this
552 * kernel stack can be released after the list_del_init.
554 if (!list_empty_careful(&uwq
.wq
.entry
)) {
555 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
557 * No need of list_del_init(), the uwq on the stack
558 * will be freed shortly anyway.
560 list_del(&uwq
.wq
.entry
);
561 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
565 * ctx may go away after this if the userfault pseudo fd is
568 userfaultfd_ctx_put(ctx
);
574 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
575 struct userfaultfd_wait_queue
*ewq
)
577 struct userfaultfd_ctx
*release_new_ctx
;
579 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
583 init_waitqueue_entry(&ewq
->wq
, current
);
584 release_new_ctx
= NULL
;
586 spin_lock_irq(&ctx
->event_wqh
.lock
);
588 * After the __add_wait_queue the uwq is visible to userland
589 * through poll/read().
591 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
593 set_current_state(TASK_KILLABLE
);
594 if (ewq
->msg
.event
== 0)
596 if (READ_ONCE(ctx
->released
) ||
597 fatal_signal_pending(current
)) {
599 * &ewq->wq may be queued in fork_event, but
600 * __remove_wait_queue ignores the head
601 * parameter. It would be a problem if it
604 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
605 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
606 struct userfaultfd_ctx
*new;
608 new = (struct userfaultfd_ctx
*)
610 ewq
->msg
.arg
.reserved
.reserved1
;
611 release_new_ctx
= new;
616 spin_unlock_irq(&ctx
->event_wqh
.lock
);
618 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
621 spin_lock_irq(&ctx
->event_wqh
.lock
);
623 __set_current_state(TASK_RUNNING
);
624 spin_unlock_irq(&ctx
->event_wqh
.lock
);
626 if (release_new_ctx
) {
627 struct vm_area_struct
*vma
;
628 struct mm_struct
*mm
= release_new_ctx
->mm
;
629 VMA_ITERATOR(vmi
, mm
, 0);
631 /* the various vma->vm_userfaultfd_ctx still points to it */
633 for_each_vma(vmi
, vma
) {
634 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
635 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
636 userfaultfd_set_vm_flags(vma
,
637 vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
640 mmap_write_unlock(mm
);
642 userfaultfd_ctx_put(release_new_ctx
);
646 * ctx may go away after this if the userfault pseudo fd is
650 atomic_dec(&ctx
->mmap_changing
);
651 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
652 userfaultfd_ctx_put(ctx
);
655 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
656 struct userfaultfd_wait_queue
*ewq
)
659 wake_up_locked(&ctx
->event_wqh
);
660 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
663 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
665 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
666 struct userfaultfd_fork_ctx
*fctx
;
668 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
669 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
670 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
671 userfaultfd_set_vm_flags(vma
, vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
675 list_for_each_entry(fctx
, fcs
, list
)
676 if (fctx
->orig
== octx
) {
682 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
686 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
692 refcount_set(&ctx
->refcount
, 1);
693 ctx
->flags
= octx
->flags
;
694 ctx
->features
= octx
->features
;
695 ctx
->released
= false;
696 atomic_set(&ctx
->mmap_changing
, 0);
697 ctx
->mm
= vma
->vm_mm
;
700 userfaultfd_ctx_get(octx
);
701 atomic_inc(&octx
->mmap_changing
);
704 list_add_tail(&fctx
->list
, fcs
);
707 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
711 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
713 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
714 struct userfaultfd_wait_queue ewq
;
718 ewq
.msg
.event
= UFFD_EVENT_FORK
;
719 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
721 userfaultfd_event_wait_completion(ctx
, &ewq
);
724 void dup_userfaultfd_complete(struct list_head
*fcs
)
726 struct userfaultfd_fork_ctx
*fctx
, *n
;
728 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
730 list_del(&fctx
->list
);
735 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
736 struct vm_userfaultfd_ctx
*vm_ctx
)
738 struct userfaultfd_ctx
*ctx
;
740 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
745 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
747 userfaultfd_ctx_get(ctx
);
748 atomic_inc(&ctx
->mmap_changing
);
750 /* Drop uffd context if remap feature not enabled */
751 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
752 userfaultfd_set_vm_flags(vma
, vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
756 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
757 unsigned long from
, unsigned long to
,
760 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
761 struct userfaultfd_wait_queue ewq
;
766 if (to
& ~PAGE_MASK
) {
767 userfaultfd_ctx_put(ctx
);
773 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
774 ewq
.msg
.arg
.remap
.from
= from
;
775 ewq
.msg
.arg
.remap
.to
= to
;
776 ewq
.msg
.arg
.remap
.len
= len
;
778 userfaultfd_event_wait_completion(ctx
, &ewq
);
781 bool userfaultfd_remove(struct vm_area_struct
*vma
,
782 unsigned long start
, unsigned long end
)
784 struct mm_struct
*mm
= vma
->vm_mm
;
785 struct userfaultfd_ctx
*ctx
;
786 struct userfaultfd_wait_queue ewq
;
788 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
789 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
792 userfaultfd_ctx_get(ctx
);
793 atomic_inc(&ctx
->mmap_changing
);
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 mm_struct
*mm
, unsigned long start
,
821 unsigned long end
, struct list_head
*unmaps
)
823 VMA_ITERATOR(vmi
, mm
, start
);
824 struct vm_area_struct
*vma
;
826 for_each_vma_range(vmi
, vma
, end
) {
827 struct userfaultfd_unmap_ctx
*unmap_ctx
;
828 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
830 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
831 has_unmap_ctx(ctx
, unmaps
, start
, end
))
834 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
838 userfaultfd_ctx_get(ctx
);
839 atomic_inc(&ctx
->mmap_changing
);
840 unmap_ctx
->ctx
= ctx
;
841 unmap_ctx
->start
= start
;
842 unmap_ctx
->end
= end
;
843 list_add_tail(&unmap_ctx
->list
, unmaps
);
849 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
851 struct userfaultfd_unmap_ctx
*ctx
, *n
;
852 struct userfaultfd_wait_queue ewq
;
854 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
857 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
858 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
859 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
861 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
863 list_del(&ctx
->list
);
868 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
870 struct userfaultfd_ctx
*ctx
= file
->private_data
;
871 struct mm_struct
*mm
= ctx
->mm
;
872 struct vm_area_struct
*vma
, *prev
;
873 /* len == 0 means wake all */
874 struct userfaultfd_wake_range range
= { .len
= 0, };
875 unsigned long new_flags
;
876 MA_STATE(mas
, &mm
->mm_mt
, 0, 0);
878 WRITE_ONCE(ctx
->released
, true);
880 if (!mmget_not_zero(mm
))
884 * Flush page faults out of all CPUs. NOTE: all page faults
885 * must be retried without returning VM_FAULT_SIGBUS if
886 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
887 * changes while handle_userfault released the mmap_lock. So
888 * it's critical that released is set to true (above), before
889 * taking the mmap_lock for writing.
893 mas_for_each(&mas
, vma
, ULONG_MAX
) {
895 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
896 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
897 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
901 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
902 prev
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
,
903 new_flags
, vma
->anon_vma
,
904 vma
->vm_file
, vma
->vm_pgoff
,
906 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
914 userfaultfd_set_vm_flags(vma
, new_flags
);
915 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
917 mmap_write_unlock(mm
);
921 * After no new page faults can wait on this fault_*wqh, flush
922 * the last page faults that may have been already waiting on
925 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
926 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
927 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
928 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
930 /* Flush pending events that may still wait on event_wqh */
931 wake_up_all(&ctx
->event_wqh
);
933 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
934 userfaultfd_ctx_put(ctx
);
938 /* fault_pending_wqh.lock must be hold by the caller */
939 static inline struct userfaultfd_wait_queue
*find_userfault_in(
940 wait_queue_head_t
*wqh
)
942 wait_queue_entry_t
*wq
;
943 struct userfaultfd_wait_queue
*uwq
;
945 lockdep_assert_held(&wqh
->lock
);
948 if (!waitqueue_active(wqh
))
950 /* walk in reverse to provide FIFO behavior to read userfaults */
951 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
952 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
957 static inline struct userfaultfd_wait_queue
*find_userfault(
958 struct userfaultfd_ctx
*ctx
)
960 return find_userfault_in(&ctx
->fault_pending_wqh
);
963 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
964 struct userfaultfd_ctx
*ctx
)
966 return find_userfault_in(&ctx
->event_wqh
);
969 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
971 struct userfaultfd_ctx
*ctx
= file
->private_data
;
974 poll_wait(file
, &ctx
->fd_wqh
, wait
);
976 if (!userfaultfd_is_initialized(ctx
))
980 * poll() never guarantees that read won't block.
981 * userfaults can be waken before they're read().
983 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
986 * lockless access to see if there are pending faults
987 * __pollwait last action is the add_wait_queue but
988 * the spin_unlock would allow the waitqueue_active to
989 * pass above the actual list_add inside
990 * add_wait_queue critical section. So use a full
991 * memory barrier to serialize the list_add write of
992 * add_wait_queue() with the waitqueue_active read
997 if (waitqueue_active(&ctx
->fault_pending_wqh
))
999 else if (waitqueue_active(&ctx
->event_wqh
))
1005 static const struct file_operations userfaultfd_fops
;
1007 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
1008 struct inode
*inode
,
1009 struct uffd_msg
*msg
)
1013 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, new,
1014 O_RDONLY
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
1018 msg
->arg
.reserved
.reserved1
= 0;
1019 msg
->arg
.fork
.ufd
= fd
;
1023 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1024 struct uffd_msg
*msg
, struct inode
*inode
)
1027 DECLARE_WAITQUEUE(wait
, current
);
1028 struct userfaultfd_wait_queue
*uwq
;
1030 * Handling fork event requires sleeping operations, so
1031 * we drop the event_wqh lock, then do these ops, then
1032 * lock it back and wake up the waiter. While the lock is
1033 * dropped the ewq may go away so we keep track of it
1036 LIST_HEAD(fork_event
);
1037 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1039 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1040 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1041 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1043 set_current_state(TASK_INTERRUPTIBLE
);
1044 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1045 uwq
= find_userfault(ctx
);
1048 * Use a seqcount to repeat the lockless check
1049 * in wake_userfault() to avoid missing
1050 * wakeups because during the refile both
1051 * waitqueue could become empty if this is the
1054 write_seqcount_begin(&ctx
->refile_seq
);
1057 * The fault_pending_wqh.lock prevents the uwq
1058 * to disappear from under us.
1060 * Refile this userfault from
1061 * fault_pending_wqh to fault_wqh, it's not
1062 * pending anymore after we read it.
1064 * Use list_del() by hand (as
1065 * userfaultfd_wake_function also uses
1066 * list_del_init() by hand) to be sure nobody
1067 * changes __remove_wait_queue() to use
1068 * list_del_init() in turn breaking the
1069 * !list_empty_careful() check in
1070 * handle_userfault(). The uwq->wq.head list
1071 * must never be empty at any time during the
1072 * refile, or the waitqueue could disappear
1073 * from under us. The "wait_queue_head_t"
1074 * parameter of __remove_wait_queue() is unused
1077 list_del(&uwq
->wq
.entry
);
1078 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1080 write_seqcount_end(&ctx
->refile_seq
);
1082 /* careful to always initialize msg if ret == 0 */
1084 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1088 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1090 spin_lock(&ctx
->event_wqh
.lock
);
1091 uwq
= find_userfault_evt(ctx
);
1095 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1096 fork_nctx
= (struct userfaultfd_ctx
*)
1098 uwq
->msg
.arg
.reserved
.reserved1
;
1099 list_move(&uwq
->wq
.entry
, &fork_event
);
1101 * fork_nctx can be freed as soon as
1102 * we drop the lock, unless we take a
1105 userfaultfd_ctx_get(fork_nctx
);
1106 spin_unlock(&ctx
->event_wqh
.lock
);
1111 userfaultfd_event_complete(ctx
, uwq
);
1112 spin_unlock(&ctx
->event_wqh
.lock
);
1116 spin_unlock(&ctx
->event_wqh
.lock
);
1118 if (signal_pending(current
)) {
1126 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1128 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1130 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1131 __set_current_state(TASK_RUNNING
);
1132 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1134 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1135 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1136 spin_lock_irq(&ctx
->event_wqh
.lock
);
1137 if (!list_empty(&fork_event
)) {
1139 * The fork thread didn't abort, so we can
1140 * drop the temporary refcount.
1142 userfaultfd_ctx_put(fork_nctx
);
1144 uwq
= list_first_entry(&fork_event
,
1148 * If fork_event list wasn't empty and in turn
1149 * the event wasn't already released by fork
1150 * (the event is allocated on fork kernel
1151 * stack), put the event back to its place in
1152 * the event_wq. fork_event head will be freed
1153 * as soon as we return so the event cannot
1154 * stay queued there no matter the current
1157 list_del(&uwq
->wq
.entry
);
1158 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1161 * Leave the event in the waitqueue and report
1162 * error to userland if we failed to resolve
1163 * the userfault fork.
1166 userfaultfd_event_complete(ctx
, uwq
);
1169 * Here the fork thread aborted and the
1170 * refcount from the fork thread on fork_nctx
1171 * has already been released. We still hold
1172 * the reference we took before releasing the
1173 * lock above. If resolve_userfault_fork
1174 * failed we've to drop it because the
1175 * fork_nctx has to be freed in such case. If
1176 * it succeeded we'll hold it because the new
1177 * uffd references it.
1180 userfaultfd_ctx_put(fork_nctx
);
1182 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1188 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1189 size_t count
, loff_t
*ppos
)
1191 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1192 ssize_t _ret
, ret
= 0;
1193 struct uffd_msg msg
;
1194 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1195 struct inode
*inode
= file_inode(file
);
1197 if (!userfaultfd_is_initialized(ctx
))
1201 if (count
< sizeof(msg
))
1202 return ret
? ret
: -EINVAL
;
1203 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1205 return ret
? ret
: _ret
;
1206 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1207 return ret
? ret
: -EFAULT
;
1210 count
-= sizeof(msg
);
1212 * Allow to read more than one fault at time but only
1213 * block if waiting for the very first one.
1215 no_wait
= O_NONBLOCK
;
1219 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1220 struct userfaultfd_wake_range
*range
)
1222 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1223 /* wake all in the range and autoremove */
1224 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1225 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1227 if (waitqueue_active(&ctx
->fault_wqh
))
1228 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1229 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1232 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1233 struct userfaultfd_wake_range
*range
)
1239 * To be sure waitqueue_active() is not reordered by the CPU
1240 * before the pagetable update, use an explicit SMP memory
1241 * barrier here. PT lock release or mmap_read_unlock(mm) still
1242 * have release semantics that can allow the
1243 * waitqueue_active() to be reordered before the pte update.
1248 * Use waitqueue_active because it's very frequent to
1249 * change the address space atomically even if there are no
1250 * userfaults yet. So we take the spinlock only when we're
1251 * sure we've userfaults to wake.
1254 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1255 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1256 waitqueue_active(&ctx
->fault_wqh
);
1258 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1260 __wake_userfault(ctx
, range
);
1263 static __always_inline
int validate_range(struct mm_struct
*mm
,
1264 __u64 start
, __u64 len
)
1266 __u64 task_size
= mm
->task_size
;
1268 if (start
& ~PAGE_MASK
)
1270 if (len
& ~PAGE_MASK
)
1274 if (start
< mmap_min_addr
)
1276 if (start
>= task_size
)
1278 if (len
> task_size
- start
)
1283 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1286 struct mm_struct
*mm
= ctx
->mm
;
1287 struct vm_area_struct
*vma
, *prev
, *cur
;
1289 struct uffdio_register uffdio_register
;
1290 struct uffdio_register __user
*user_uffdio_register
;
1291 unsigned long vm_flags
, new_flags
;
1294 unsigned long start
, end
, vma_end
;
1295 MA_STATE(mas
, &mm
->mm_mt
, 0, 0);
1297 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1300 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1301 sizeof(uffdio_register
)-sizeof(__u64
)))
1305 if (!uffdio_register
.mode
)
1307 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1310 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1311 vm_flags
|= VM_UFFD_MISSING
;
1312 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1313 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1316 vm_flags
|= VM_UFFD_WP
;
1318 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1319 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1322 vm_flags
|= VM_UFFD_MINOR
;
1325 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1326 uffdio_register
.range
.len
);
1330 start
= uffdio_register
.range
.start
;
1331 end
= start
+ uffdio_register
.range
.len
;
1334 if (!mmget_not_zero(mm
))
1337 mmap_write_lock(mm
);
1338 mas_set(&mas
, start
);
1339 vma
= mas_find(&mas
, ULONG_MAX
);
1343 /* check that there's at least one vma in the range */
1345 if (vma
->vm_start
>= end
)
1349 * If the first vma contains huge pages, make sure start address
1350 * is aligned to huge page size.
1352 if (is_vm_hugetlb_page(vma
)) {
1353 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1355 if (start
& (vma_hpagesize
- 1))
1360 * Search for not compatible vmas.
1363 basic_ioctls
= false;
1364 for (cur
= vma
; cur
; cur
= mas_next(&mas
, end
- 1)) {
1367 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1368 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1370 /* check not compatible vmas */
1372 if (!vma_can_userfault(cur
, vm_flags
))
1376 * UFFDIO_COPY will fill file holes even without
1377 * PROT_WRITE. This check enforces that if this is a
1378 * MAP_SHARED, the process has write permission to the backing
1379 * file. If VM_MAYWRITE is set it also enforces that on a
1380 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1381 * F_WRITE_SEAL can be taken until the vma is destroyed.
1384 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1388 * If this vma contains ending address, and huge pages
1391 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1392 end
> cur
->vm_start
) {
1393 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1397 if (end
& (vma_hpagesize
- 1))
1400 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1404 * Check that this vma isn't already owned by a
1405 * different userfaultfd. We can't allow more than one
1406 * userfaultfd to own a single vma simultaneously or we
1407 * wouldn't know which one to deliver the userfaults to.
1410 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1411 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1415 * Note vmas containing huge pages
1417 if (is_vm_hugetlb_page(cur
))
1418 basic_ioctls
= true;
1424 mas_set(&mas
, start
);
1425 prev
= mas_prev(&mas
, 0);
1427 mas_next(&mas
, ULONG_MAX
);
1433 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1434 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1435 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1436 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1439 * Nothing to do: this vma is already registered into this
1440 * userfaultfd and with the right tracking mode too.
1442 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1443 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1446 if (vma
->vm_start
> start
)
1447 start
= vma
->vm_start
;
1448 vma_end
= min(end
, vma
->vm_end
);
1450 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1451 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1452 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1454 ((struct vm_userfaultfd_ctx
){ ctx
}),
1455 anon_vma_name(vma
));
1457 /* vma_merge() invalidated the mas */
1462 if (vma
->vm_start
< start
) {
1463 ret
= split_vma(mm
, vma
, start
, 1);
1466 /* split_vma() invalidated the mas */
1469 if (vma
->vm_end
> end
) {
1470 ret
= split_vma(mm
, vma
, end
, 0);
1473 /* split_vma() invalidated the mas */
1478 * In the vma_merge() successful mprotect-like case 8:
1479 * the next vma was merged into the current one and
1480 * the current one has not been updated yet.
1482 userfaultfd_set_vm_flags(vma
, new_flags
);
1483 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1485 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1486 hugetlb_unshare_all_pmds(vma
);
1490 start
= vma
->vm_end
;
1491 vma
= mas_next(&mas
, end
- 1);
1494 mmap_write_unlock(mm
);
1499 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1500 UFFD_API_RANGE_IOCTLS
;
1503 * Declare the WP ioctl only if the WP mode is
1504 * specified and all checks passed with the range
1506 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1507 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1509 /* CONTINUE ioctl is only supported for MINOR ranges. */
1510 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
))
1511 ioctls_out
&= ~((__u64
)1 << _UFFDIO_CONTINUE
);
1514 * Now that we scanned all vmas we can already tell
1515 * userland which ioctls methods are guaranteed to
1516 * succeed on this range.
1518 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1525 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1528 struct mm_struct
*mm
= ctx
->mm
;
1529 struct vm_area_struct
*vma
, *prev
, *cur
;
1531 struct uffdio_range uffdio_unregister
;
1532 unsigned long new_flags
;
1534 unsigned long start
, end
, vma_end
;
1535 const void __user
*buf
= (void __user
*)arg
;
1536 MA_STATE(mas
, &mm
->mm_mt
, 0, 0);
1539 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1542 ret
= validate_range(mm
, uffdio_unregister
.start
,
1543 uffdio_unregister
.len
);
1547 start
= uffdio_unregister
.start
;
1548 end
= start
+ uffdio_unregister
.len
;
1551 if (!mmget_not_zero(mm
))
1554 mmap_write_lock(mm
);
1555 mas_set(&mas
, start
);
1556 vma
= mas_find(&mas
, ULONG_MAX
);
1560 /* check that there's at least one vma in the range */
1562 if (vma
->vm_start
>= end
)
1566 * If the first vma contains huge pages, make sure start address
1567 * is aligned to huge page size.
1569 if (is_vm_hugetlb_page(vma
)) {
1570 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1572 if (start
& (vma_hpagesize
- 1))
1577 * Search for not compatible vmas.
1581 for (cur
= vma
; cur
; cur
= mas_next(&mas
, end
- 1)) {
1584 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1585 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1588 * Check not compatible vmas, not strictly required
1589 * here as not compatible vmas cannot have an
1590 * userfaultfd_ctx registered on them, but this
1591 * provides for more strict behavior to notice
1592 * unregistration errors.
1594 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1601 mas_set(&mas
, start
);
1602 prev
= mas_prev(&mas
, 0);
1604 mas_next(&mas
, ULONG_MAX
);
1610 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1613 * Nothing to do: this vma is already registered into this
1614 * userfaultfd and with the right tracking mode too.
1616 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1619 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1621 if (vma
->vm_start
> start
)
1622 start
= vma
->vm_start
;
1623 vma_end
= min(end
, vma
->vm_end
);
1625 if (userfaultfd_missing(vma
)) {
1627 * Wake any concurrent pending userfault while
1628 * we unregister, so they will not hang
1629 * permanently and it avoids userland to call
1630 * UFFDIO_WAKE explicitly.
1632 struct userfaultfd_wake_range range
;
1633 range
.start
= start
;
1634 range
.len
= vma_end
- start
;
1635 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1638 /* Reset ptes for the whole vma range if wr-protected */
1639 if (userfaultfd_wp(vma
))
1640 uffd_wp_range(mm
, vma
, start
, vma_end
- start
, false);
1642 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1643 prev
= vma_merge(mm
, prev
, start
, vma_end
, new_flags
,
1644 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1646 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
1652 if (vma
->vm_start
< start
) {
1653 ret
= split_vma(mm
, vma
, start
, 1);
1658 if (vma
->vm_end
> end
) {
1659 ret
= split_vma(mm
, vma
, end
, 0);
1666 * In the vma_merge() successful mprotect-like case 8:
1667 * the next vma was merged into the current one and
1668 * the current one has not been updated yet.
1670 userfaultfd_set_vm_flags(vma
, new_flags
);
1671 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1675 start
= vma
->vm_end
;
1676 vma
= mas_next(&mas
, end
- 1);
1679 mmap_write_unlock(mm
);
1686 * userfaultfd_wake may be used in combination with the
1687 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1689 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1693 struct uffdio_range uffdio_wake
;
1694 struct userfaultfd_wake_range range
;
1695 const void __user
*buf
= (void __user
*)arg
;
1698 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1701 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1705 range
.start
= uffdio_wake
.start
;
1706 range
.len
= uffdio_wake
.len
;
1709 * len == 0 means wake all and we don't want to wake all here,
1710 * so check it again to be sure.
1712 VM_BUG_ON(!range
.len
);
1714 wake_userfault(ctx
, &range
);
1721 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1725 struct uffdio_copy uffdio_copy
;
1726 struct uffdio_copy __user
*user_uffdio_copy
;
1727 struct userfaultfd_wake_range range
;
1729 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1732 if (atomic_read(&ctx
->mmap_changing
))
1736 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1737 /* don't copy "copy" last field */
1738 sizeof(uffdio_copy
)-sizeof(__s64
)))
1741 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1745 * double check for wraparound just in case. copy_from_user()
1746 * will later check uffdio_copy.src + uffdio_copy.len to fit
1747 * in the userland range.
1750 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1752 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1754 if (mmget_not_zero(ctx
->mm
)) {
1755 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1756 uffdio_copy
.len
, &ctx
->mmap_changing
,
1762 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1767 /* len == 0 would wake all */
1769 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1770 range
.start
= uffdio_copy
.dst
;
1771 wake_userfault(ctx
, &range
);
1773 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1778 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1782 struct uffdio_zeropage uffdio_zeropage
;
1783 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1784 struct userfaultfd_wake_range range
;
1786 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1789 if (atomic_read(&ctx
->mmap_changing
))
1793 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1794 /* don't copy "zeropage" last field */
1795 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1798 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1799 uffdio_zeropage
.range
.len
);
1803 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1806 if (mmget_not_zero(ctx
->mm
)) {
1807 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1808 uffdio_zeropage
.range
.len
,
1809 &ctx
->mmap_changing
);
1814 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1818 /* len == 0 would wake all */
1821 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1822 range
.start
= uffdio_zeropage
.range
.start
;
1823 wake_userfault(ctx
, &range
);
1825 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1830 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1834 struct uffdio_writeprotect uffdio_wp
;
1835 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1836 struct userfaultfd_wake_range range
;
1837 bool mode_wp
, mode_dontwake
;
1839 if (atomic_read(&ctx
->mmap_changing
))
1842 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1844 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1845 sizeof(struct uffdio_writeprotect
)))
1848 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1849 uffdio_wp
.range
.len
);
1853 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1854 UFFDIO_WRITEPROTECT_MODE_WP
))
1857 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1858 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1860 if (mode_wp
&& mode_dontwake
)
1863 if (mmget_not_zero(ctx
->mm
)) {
1864 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1865 uffdio_wp
.range
.len
, mode_wp
,
1866 &ctx
->mmap_changing
);
1875 if (!mode_wp
&& !mode_dontwake
) {
1876 range
.start
= uffdio_wp
.range
.start
;
1877 range
.len
= uffdio_wp
.range
.len
;
1878 wake_userfault(ctx
, &range
);
1883 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1886 struct uffdio_continue uffdio_continue
;
1887 struct uffdio_continue __user
*user_uffdio_continue
;
1888 struct userfaultfd_wake_range range
;
1890 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1893 if (atomic_read(&ctx
->mmap_changing
))
1897 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1898 /* don't copy the output fields */
1899 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1902 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1903 uffdio_continue
.range
.len
);
1908 /* double check for wraparound just in case. */
1909 if (uffdio_continue
.range
.start
+ uffdio_continue
.range
.len
<=
1910 uffdio_continue
.range
.start
) {
1913 if (uffdio_continue
.mode
& ~UFFDIO_CONTINUE_MODE_DONTWAKE
)
1916 if (mmget_not_zero(ctx
->mm
)) {
1917 ret
= mcopy_continue(ctx
->mm
, uffdio_continue
.range
.start
,
1918 uffdio_continue
.range
.len
,
1919 &ctx
->mmap_changing
);
1925 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1930 /* len == 0 would wake all */
1933 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1934 range
.start
= uffdio_continue
.range
.start
;
1935 wake_userfault(ctx
, &range
);
1937 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1943 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1946 * For the current set of features the bits just coincide. Set
1947 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1949 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1953 * userland asks for a certain API version and we return which bits
1954 * and ioctl commands are implemented in this kernel for such API
1955 * version or -EINVAL if unknown.
1957 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1960 struct uffdio_api uffdio_api
;
1961 void __user
*buf
= (void __user
*)arg
;
1962 unsigned int ctx_features
;
1967 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1969 /* Ignore unsupported features (userspace built against newer kernel) */
1970 features
= uffdio_api
.features
& UFFD_API_FEATURES
;
1972 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1974 /* report all available features and ioctls to userland */
1975 uffdio_api
.features
= UFFD_API_FEATURES
;
1976 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1977 uffdio_api
.features
&=
1978 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
1980 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1981 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
1983 #ifndef CONFIG_PTE_MARKER_UFFD_WP
1984 uffdio_api
.features
&= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM
;
1986 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1988 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1991 /* only enable the requested features for this uffd context */
1992 ctx_features
= uffd_ctx_features(features
);
1994 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
2001 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
2002 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
2007 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
2011 struct userfaultfd_ctx
*ctx
= file
->private_data
;
2013 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
2018 ret
= userfaultfd_api(ctx
, arg
);
2020 case UFFDIO_REGISTER
:
2021 ret
= userfaultfd_register(ctx
, arg
);
2023 case UFFDIO_UNREGISTER
:
2024 ret
= userfaultfd_unregister(ctx
, arg
);
2027 ret
= userfaultfd_wake(ctx
, arg
);
2030 ret
= userfaultfd_copy(ctx
, arg
);
2032 case UFFDIO_ZEROPAGE
:
2033 ret
= userfaultfd_zeropage(ctx
, arg
);
2035 case UFFDIO_WRITEPROTECT
:
2036 ret
= userfaultfd_writeprotect(ctx
, arg
);
2038 case UFFDIO_CONTINUE
:
2039 ret
= userfaultfd_continue(ctx
, arg
);
2045 #ifdef CONFIG_PROC_FS
2046 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2048 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2049 wait_queue_entry_t
*wq
;
2050 unsigned long pending
= 0, total
= 0;
2052 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2053 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2057 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2060 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2063 * If more protocols will be added, there will be all shown
2064 * separated by a space. Like this:
2065 * protocols: aa:... bb:...
2067 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2068 pending
, total
, UFFD_API
, ctx
->features
,
2069 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2073 static const struct file_operations userfaultfd_fops
= {
2074 #ifdef CONFIG_PROC_FS
2075 .show_fdinfo
= userfaultfd_show_fdinfo
,
2077 .release
= userfaultfd_release
,
2078 .poll
= userfaultfd_poll
,
2079 .read
= userfaultfd_read
,
2080 .unlocked_ioctl
= userfaultfd_ioctl
,
2081 .compat_ioctl
= compat_ptr_ioctl
,
2082 .llseek
= noop_llseek
,
2085 static void init_once_userfaultfd_ctx(void *mem
)
2087 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2089 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2090 init_waitqueue_head(&ctx
->fault_wqh
);
2091 init_waitqueue_head(&ctx
->event_wqh
);
2092 init_waitqueue_head(&ctx
->fd_wqh
);
2093 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2096 static int new_userfaultfd(int flags
)
2098 struct userfaultfd_ctx
*ctx
;
2101 BUG_ON(!current
->mm
);
2103 /* Check the UFFD_* constants for consistency. */
2104 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2105 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2106 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2108 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2111 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2115 refcount_set(&ctx
->refcount
, 1);
2118 ctx
->released
= false;
2119 atomic_set(&ctx
->mmap_changing
, 0);
2120 ctx
->mm
= current
->mm
;
2121 /* prevent the mm struct to be freed */
2124 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, ctx
,
2125 O_RDONLY
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2128 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2133 static inline bool userfaultfd_syscall_allowed(int flags
)
2135 /* Userspace-only page faults are always allowed */
2136 if (flags
& UFFD_USER_MODE_ONLY
)
2140 * The user is requesting a userfaultfd which can handle kernel faults.
2141 * Privileged users are always allowed to do this.
2143 if (capable(CAP_SYS_PTRACE
))
2146 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2147 return sysctl_unprivileged_userfaultfd
;
2150 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2152 if (!userfaultfd_syscall_allowed(flags
))
2155 return new_userfaultfd(flags
);
2158 static long userfaultfd_dev_ioctl(struct file
*file
, unsigned int cmd
, unsigned long flags
)
2160 if (cmd
!= USERFAULTFD_IOC_NEW
)
2163 return new_userfaultfd(flags
);
2166 static const struct file_operations userfaultfd_dev_fops
= {
2167 .unlocked_ioctl
= userfaultfd_dev_ioctl
,
2168 .compat_ioctl
= userfaultfd_dev_ioctl
,
2169 .owner
= THIS_MODULE
,
2170 .llseek
= noop_llseek
,
2173 static struct miscdevice userfaultfd_misc
= {
2174 .minor
= MISC_DYNAMIC_MINOR
,
2175 .name
= "userfaultfd",
2176 .fops
= &userfaultfd_dev_fops
2179 static int __init
userfaultfd_init(void)
2183 ret
= misc_register(&userfaultfd_misc
);
2187 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2188 sizeof(struct userfaultfd_ctx
),
2190 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2191 init_once_userfaultfd_ctx
);
2194 __initcall(userfaultfd_init
);