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 vm_flags_reset(vma
, 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
)
258 mmap_assert_locked(ctx
->mm
);
260 ptep
= hugetlb_walk(vma
, address
, vma_mmu_pagesize(vma
));
265 pte
= huge_ptep_get(ptep
);
268 * Lockless access: we're in a wait_event so it's ok if it
269 * changes under us. PTE markers should be handled the same as none
272 if (huge_pte_none_mostly(pte
))
274 if (!huge_pte_write(pte
) && (reason
& VM_UFFD_WP
))
280 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx
*ctx
,
281 struct vm_area_struct
*vma
,
282 unsigned long address
,
284 unsigned long reason
)
286 return false; /* should never get here */
288 #endif /* CONFIG_HUGETLB_PAGE */
291 * Verify the pagetables are still not ok after having reigstered into
292 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
293 * userfault that has already been resolved, if userfaultfd_read and
294 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
297 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx
*ctx
,
298 unsigned long address
,
300 unsigned long reason
)
302 struct mm_struct
*mm
= ctx
->mm
;
310 mmap_assert_locked(mm
);
312 pgd
= pgd_offset(mm
, address
);
313 if (!pgd_present(*pgd
))
315 p4d
= p4d_offset(pgd
, address
);
316 if (!p4d_present(*p4d
))
318 pud
= pud_offset(p4d
, address
);
319 if (!pud_present(*pud
))
321 pmd
= pmd_offset(pud
, address
);
323 * READ_ONCE must function as a barrier with narrower scope
324 * and it must be equivalent to:
325 * _pmd = *pmd; barrier();
327 * This is to deal with the instability (as in
328 * pmd_trans_unstable) of the pmd.
330 _pmd
= READ_ONCE(*pmd
);
335 if (!pmd_present(_pmd
))
338 if (pmd_trans_huge(_pmd
)) {
339 if (!pmd_write(_pmd
) && (reason
& VM_UFFD_WP
))
345 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
346 * and use the standard pte_offset_map() instead of parsing _pmd.
348 pte
= pte_offset_map(pmd
, address
);
350 * Lockless access: we're in a wait_event so it's ok if it
351 * changes under us. PTE markers should be handled the same as none
354 if (pte_none_mostly(*pte
))
356 if (!pte_write(*pte
) && (reason
& VM_UFFD_WP
))
364 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags
)
366 if (flags
& FAULT_FLAG_INTERRUPTIBLE
)
367 return TASK_INTERRUPTIBLE
;
369 if (flags
& FAULT_FLAG_KILLABLE
)
370 return TASK_KILLABLE
;
372 return TASK_UNINTERRUPTIBLE
;
376 * The locking rules involved in returning VM_FAULT_RETRY depending on
377 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
378 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
379 * recommendation in __lock_page_or_retry is not an understatement.
381 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
382 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
385 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
386 * set, VM_FAULT_RETRY can still be returned if and only if there are
387 * fatal_signal_pending()s, and the mmap_lock must be released before
390 vm_fault_t
handle_userfault(struct vm_fault
*vmf
, unsigned long reason
)
392 struct vm_area_struct
*vma
= vmf
->vma
;
393 struct mm_struct
*mm
= vma
->vm_mm
;
394 struct userfaultfd_ctx
*ctx
;
395 struct userfaultfd_wait_queue uwq
;
396 vm_fault_t ret
= VM_FAULT_SIGBUS
;
398 unsigned int blocking_state
;
401 * We don't do userfault handling for the final child pid update.
403 * We also don't do userfault handling during
404 * coredumping. hugetlbfs has the special
405 * follow_hugetlb_page() to skip missing pages in the
406 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
407 * the no_page_table() helper in follow_page_mask(), but the
408 * shmem_vm_ops->fault method is invoked even during
409 * coredumping without mmap_lock and it ends up here.
411 if (current
->flags
& (PF_EXITING
|PF_DUMPCORE
))
415 * Coredumping runs without mmap_lock so we can only check that
416 * the mmap_lock is held, if PF_DUMPCORE was not set.
418 mmap_assert_locked(mm
);
420 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
424 BUG_ON(ctx
->mm
!= mm
);
426 /* Any unrecognized flag is a bug. */
427 VM_BUG_ON(reason
& ~__VM_UFFD_FLAGS
);
428 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
429 VM_BUG_ON(!reason
|| (reason
& (reason
- 1)));
431 if (ctx
->features
& UFFD_FEATURE_SIGBUS
)
433 if (!(vmf
->flags
& FAULT_FLAG_USER
) && (ctx
->flags
& UFFD_USER_MODE_ONLY
))
437 * If it's already released don't get it. This avoids to loop
438 * in __get_user_pages if userfaultfd_release waits on the
439 * caller of handle_userfault to release the mmap_lock.
441 if (unlikely(READ_ONCE(ctx
->released
))) {
443 * Don't return VM_FAULT_SIGBUS in this case, so a non
444 * cooperative manager can close the uffd after the
445 * last UFFDIO_COPY, without risking to trigger an
446 * involuntary SIGBUS if the process was starting the
447 * userfaultfd while the userfaultfd was still armed
448 * (but after the last UFFDIO_COPY). If the uffd
449 * wasn't already closed when the userfault reached
450 * this point, that would normally be solved by
451 * userfaultfd_must_wait returning 'false'.
453 * If we were to return VM_FAULT_SIGBUS here, the non
454 * cooperative manager would be instead forced to
455 * always call UFFDIO_UNREGISTER before it can safely
458 ret
= VM_FAULT_NOPAGE
;
463 * Check that we can return VM_FAULT_RETRY.
465 * NOTE: it should become possible to return VM_FAULT_RETRY
466 * even if FAULT_FLAG_TRIED is set without leading to gup()
467 * -EBUSY failures, if the userfaultfd is to be extended for
468 * VM_UFFD_WP tracking and we intend to arm the userfault
469 * without first stopping userland access to the memory. For
470 * VM_UFFD_MISSING userfaults this is enough for now.
472 if (unlikely(!(vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
))) {
474 * Validate the invariant that nowait must allow retry
475 * to be sure not to return SIGBUS erroneously on
476 * nowait invocations.
478 BUG_ON(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
);
479 #ifdef CONFIG_DEBUG_VM
480 if (printk_ratelimit()) {
482 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
491 * Handle nowait, not much to do other than tell it to retry
494 ret
= VM_FAULT_RETRY
;
495 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
498 /* take the reference before dropping the mmap_lock */
499 userfaultfd_ctx_get(ctx
);
501 init_waitqueue_func_entry(&uwq
.wq
, userfaultfd_wake_function
);
502 uwq
.wq
.private = current
;
503 uwq
.msg
= userfault_msg(vmf
->address
, vmf
->real_address
, vmf
->flags
,
504 reason
, ctx
->features
);
508 blocking_state
= userfaultfd_get_blocking_state(vmf
->flags
);
511 * Take the vma lock now, in order to safely call
512 * userfaultfd_huge_must_wait() later. Since acquiring the
513 * (sleepable) vma lock can modify the current task state, that
514 * must be before explicitly calling set_current_state().
516 if (is_vm_hugetlb_page(vma
))
517 hugetlb_vma_lock_read(vma
);
519 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
521 * After the __add_wait_queue the uwq is visible to userland
522 * through poll/read().
524 __add_wait_queue(&ctx
->fault_pending_wqh
, &uwq
.wq
);
526 * The smp_mb() after __set_current_state prevents the reads
527 * following the spin_unlock to happen before the list_add in
530 set_current_state(blocking_state
);
531 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
533 if (!is_vm_hugetlb_page(vma
))
534 must_wait
= userfaultfd_must_wait(ctx
, vmf
->address
, vmf
->flags
,
537 must_wait
= userfaultfd_huge_must_wait(ctx
, vma
,
540 if (is_vm_hugetlb_page(vma
))
541 hugetlb_vma_unlock_read(vma
);
542 mmap_read_unlock(mm
);
544 if (likely(must_wait
&& !READ_ONCE(ctx
->released
))) {
545 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
549 __set_current_state(TASK_RUNNING
);
552 * Here we race with the list_del; list_add in
553 * userfaultfd_ctx_read(), however because we don't ever run
554 * list_del_init() to refile across the two lists, the prev
555 * and next pointers will never point to self. list_add also
556 * would never let any of the two pointers to point to
557 * self. So list_empty_careful won't risk to see both pointers
558 * pointing to self at any time during the list refile. The
559 * only case where list_del_init() is called is the full
560 * removal in the wake function and there we don't re-list_add
561 * and it's fine not to block on the spinlock. The uwq on this
562 * kernel stack can be released after the list_del_init.
564 if (!list_empty_careful(&uwq
.wq
.entry
)) {
565 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
567 * No need of list_del_init(), the uwq on the stack
568 * will be freed shortly anyway.
570 list_del(&uwq
.wq
.entry
);
571 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
575 * ctx may go away after this if the userfault pseudo fd is
578 userfaultfd_ctx_put(ctx
);
584 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx
*ctx
,
585 struct userfaultfd_wait_queue
*ewq
)
587 struct userfaultfd_ctx
*release_new_ctx
;
589 if (WARN_ON_ONCE(current
->flags
& PF_EXITING
))
593 init_waitqueue_entry(&ewq
->wq
, current
);
594 release_new_ctx
= NULL
;
596 spin_lock_irq(&ctx
->event_wqh
.lock
);
598 * After the __add_wait_queue the uwq is visible to userland
599 * through poll/read().
601 __add_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
603 set_current_state(TASK_KILLABLE
);
604 if (ewq
->msg
.event
== 0)
606 if (READ_ONCE(ctx
->released
) ||
607 fatal_signal_pending(current
)) {
609 * &ewq->wq may be queued in fork_event, but
610 * __remove_wait_queue ignores the head
611 * parameter. It would be a problem if it
614 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
615 if (ewq
->msg
.event
== UFFD_EVENT_FORK
) {
616 struct userfaultfd_ctx
*new;
618 new = (struct userfaultfd_ctx
*)
620 ewq
->msg
.arg
.reserved
.reserved1
;
621 release_new_ctx
= new;
626 spin_unlock_irq(&ctx
->event_wqh
.lock
);
628 wake_up_poll(&ctx
->fd_wqh
, EPOLLIN
);
631 spin_lock_irq(&ctx
->event_wqh
.lock
);
633 __set_current_state(TASK_RUNNING
);
634 spin_unlock_irq(&ctx
->event_wqh
.lock
);
636 if (release_new_ctx
) {
637 struct vm_area_struct
*vma
;
638 struct mm_struct
*mm
= release_new_ctx
->mm
;
639 VMA_ITERATOR(vmi
, mm
, 0);
641 /* the various vma->vm_userfaultfd_ctx still points to it */
643 for_each_vma(vmi
, vma
) {
644 if (vma
->vm_userfaultfd_ctx
.ctx
== release_new_ctx
) {
645 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
646 userfaultfd_set_vm_flags(vma
,
647 vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
650 mmap_write_unlock(mm
);
652 userfaultfd_ctx_put(release_new_ctx
);
656 * ctx may go away after this if the userfault pseudo fd is
660 atomic_dec(&ctx
->mmap_changing
);
661 VM_BUG_ON(atomic_read(&ctx
->mmap_changing
) < 0);
662 userfaultfd_ctx_put(ctx
);
665 static void userfaultfd_event_complete(struct userfaultfd_ctx
*ctx
,
666 struct userfaultfd_wait_queue
*ewq
)
669 wake_up_locked(&ctx
->event_wqh
);
670 __remove_wait_queue(&ctx
->event_wqh
, &ewq
->wq
);
673 int dup_userfaultfd(struct vm_area_struct
*vma
, struct list_head
*fcs
)
675 struct userfaultfd_ctx
*ctx
= NULL
, *octx
;
676 struct userfaultfd_fork_ctx
*fctx
;
678 octx
= vma
->vm_userfaultfd_ctx
.ctx
;
679 if (!octx
|| !(octx
->features
& UFFD_FEATURE_EVENT_FORK
)) {
680 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
681 userfaultfd_set_vm_flags(vma
, vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
685 list_for_each_entry(fctx
, fcs
, list
)
686 if (fctx
->orig
== octx
) {
692 fctx
= kmalloc(sizeof(*fctx
), GFP_KERNEL
);
696 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
702 refcount_set(&ctx
->refcount
, 1);
703 ctx
->flags
= octx
->flags
;
704 ctx
->features
= octx
->features
;
705 ctx
->released
= false;
706 atomic_set(&ctx
->mmap_changing
, 0);
707 ctx
->mm
= vma
->vm_mm
;
710 userfaultfd_ctx_get(octx
);
711 atomic_inc(&octx
->mmap_changing
);
714 list_add_tail(&fctx
->list
, fcs
);
717 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
721 static void dup_fctx(struct userfaultfd_fork_ctx
*fctx
)
723 struct userfaultfd_ctx
*ctx
= fctx
->orig
;
724 struct userfaultfd_wait_queue ewq
;
728 ewq
.msg
.event
= UFFD_EVENT_FORK
;
729 ewq
.msg
.arg
.reserved
.reserved1
= (unsigned long)fctx
->new;
731 userfaultfd_event_wait_completion(ctx
, &ewq
);
734 void dup_userfaultfd_complete(struct list_head
*fcs
)
736 struct userfaultfd_fork_ctx
*fctx
, *n
;
738 list_for_each_entry_safe(fctx
, n
, fcs
, list
) {
740 list_del(&fctx
->list
);
745 void mremap_userfaultfd_prep(struct vm_area_struct
*vma
,
746 struct vm_userfaultfd_ctx
*vm_ctx
)
748 struct userfaultfd_ctx
*ctx
;
750 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
755 if (ctx
->features
& UFFD_FEATURE_EVENT_REMAP
) {
757 userfaultfd_ctx_get(ctx
);
758 atomic_inc(&ctx
->mmap_changing
);
760 /* Drop uffd context if remap feature not enabled */
761 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
762 userfaultfd_set_vm_flags(vma
, vma
->vm_flags
& ~__VM_UFFD_FLAGS
);
766 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx
*vm_ctx
,
767 unsigned long from
, unsigned long to
,
770 struct userfaultfd_ctx
*ctx
= vm_ctx
->ctx
;
771 struct userfaultfd_wait_queue ewq
;
776 if (to
& ~PAGE_MASK
) {
777 userfaultfd_ctx_put(ctx
);
783 ewq
.msg
.event
= UFFD_EVENT_REMAP
;
784 ewq
.msg
.arg
.remap
.from
= from
;
785 ewq
.msg
.arg
.remap
.to
= to
;
786 ewq
.msg
.arg
.remap
.len
= len
;
788 userfaultfd_event_wait_completion(ctx
, &ewq
);
791 bool userfaultfd_remove(struct vm_area_struct
*vma
,
792 unsigned long start
, unsigned long end
)
794 struct mm_struct
*mm
= vma
->vm_mm
;
795 struct userfaultfd_ctx
*ctx
;
796 struct userfaultfd_wait_queue ewq
;
798 ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
799 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_REMOVE
))
802 userfaultfd_ctx_get(ctx
);
803 atomic_inc(&ctx
->mmap_changing
);
804 mmap_read_unlock(mm
);
808 ewq
.msg
.event
= UFFD_EVENT_REMOVE
;
809 ewq
.msg
.arg
.remove
.start
= start
;
810 ewq
.msg
.arg
.remove
.end
= end
;
812 userfaultfd_event_wait_completion(ctx
, &ewq
);
817 static bool has_unmap_ctx(struct userfaultfd_ctx
*ctx
, struct list_head
*unmaps
,
818 unsigned long start
, unsigned long end
)
820 struct userfaultfd_unmap_ctx
*unmap_ctx
;
822 list_for_each_entry(unmap_ctx
, unmaps
, list
)
823 if (unmap_ctx
->ctx
== ctx
&& unmap_ctx
->start
== start
&&
824 unmap_ctx
->end
== end
)
830 int userfaultfd_unmap_prep(struct mm_struct
*mm
, unsigned long start
,
831 unsigned long end
, struct list_head
*unmaps
)
833 VMA_ITERATOR(vmi
, mm
, start
);
834 struct vm_area_struct
*vma
;
836 for_each_vma_range(vmi
, vma
, end
) {
837 struct userfaultfd_unmap_ctx
*unmap_ctx
;
838 struct userfaultfd_ctx
*ctx
= vma
->vm_userfaultfd_ctx
.ctx
;
840 if (!ctx
|| !(ctx
->features
& UFFD_FEATURE_EVENT_UNMAP
) ||
841 has_unmap_ctx(ctx
, unmaps
, start
, end
))
844 unmap_ctx
= kzalloc(sizeof(*unmap_ctx
), GFP_KERNEL
);
848 userfaultfd_ctx_get(ctx
);
849 atomic_inc(&ctx
->mmap_changing
);
850 unmap_ctx
->ctx
= ctx
;
851 unmap_ctx
->start
= start
;
852 unmap_ctx
->end
= end
;
853 list_add_tail(&unmap_ctx
->list
, unmaps
);
859 void userfaultfd_unmap_complete(struct mm_struct
*mm
, struct list_head
*uf
)
861 struct userfaultfd_unmap_ctx
*ctx
, *n
;
862 struct userfaultfd_wait_queue ewq
;
864 list_for_each_entry_safe(ctx
, n
, uf
, list
) {
867 ewq
.msg
.event
= UFFD_EVENT_UNMAP
;
868 ewq
.msg
.arg
.remove
.start
= ctx
->start
;
869 ewq
.msg
.arg
.remove
.end
= ctx
->end
;
871 userfaultfd_event_wait_completion(ctx
->ctx
, &ewq
);
873 list_del(&ctx
->list
);
878 static int userfaultfd_release(struct inode
*inode
, struct file
*file
)
880 struct userfaultfd_ctx
*ctx
= file
->private_data
;
881 struct mm_struct
*mm
= ctx
->mm
;
882 struct vm_area_struct
*vma
, *prev
;
883 /* len == 0 means wake all */
884 struct userfaultfd_wake_range range
= { .len
= 0, };
885 unsigned long new_flags
;
886 VMA_ITERATOR(vmi
, mm
, 0);
888 WRITE_ONCE(ctx
->released
, true);
890 if (!mmget_not_zero(mm
))
894 * Flush page faults out of all CPUs. NOTE: all page faults
895 * must be retried without returning VM_FAULT_SIGBUS if
896 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
897 * changes while handle_userfault released the mmap_lock. So
898 * it's critical that released is set to true (above), before
899 * taking the mmap_lock for writing.
903 for_each_vma(vmi
, vma
) {
905 BUG_ON(!!vma
->vm_userfaultfd_ctx
.ctx
^
906 !!(vma
->vm_flags
& __VM_UFFD_FLAGS
));
907 if (vma
->vm_userfaultfd_ctx
.ctx
!= ctx
) {
911 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
912 prev
= vma_merge(&vmi
, mm
, prev
, vma
->vm_start
, vma
->vm_end
,
913 new_flags
, vma
->anon_vma
,
914 vma
->vm_file
, vma
->vm_pgoff
,
916 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
923 userfaultfd_set_vm_flags(vma
, new_flags
);
924 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
926 mmap_write_unlock(mm
);
930 * After no new page faults can wait on this fault_*wqh, flush
931 * the last page faults that may have been already waiting on
934 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
935 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
, &range
);
936 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, &range
);
937 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
939 /* Flush pending events that may still wait on event_wqh */
940 wake_up_all(&ctx
->event_wqh
);
942 wake_up_poll(&ctx
->fd_wqh
, EPOLLHUP
);
943 userfaultfd_ctx_put(ctx
);
947 /* fault_pending_wqh.lock must be hold by the caller */
948 static inline struct userfaultfd_wait_queue
*find_userfault_in(
949 wait_queue_head_t
*wqh
)
951 wait_queue_entry_t
*wq
;
952 struct userfaultfd_wait_queue
*uwq
;
954 lockdep_assert_held(&wqh
->lock
);
957 if (!waitqueue_active(wqh
))
959 /* walk in reverse to provide FIFO behavior to read userfaults */
960 wq
= list_last_entry(&wqh
->head
, typeof(*wq
), entry
);
961 uwq
= container_of(wq
, struct userfaultfd_wait_queue
, wq
);
966 static inline struct userfaultfd_wait_queue
*find_userfault(
967 struct userfaultfd_ctx
*ctx
)
969 return find_userfault_in(&ctx
->fault_pending_wqh
);
972 static inline struct userfaultfd_wait_queue
*find_userfault_evt(
973 struct userfaultfd_ctx
*ctx
)
975 return find_userfault_in(&ctx
->event_wqh
);
978 static __poll_t
userfaultfd_poll(struct file
*file
, poll_table
*wait
)
980 struct userfaultfd_ctx
*ctx
= file
->private_data
;
983 poll_wait(file
, &ctx
->fd_wqh
, wait
);
985 if (!userfaultfd_is_initialized(ctx
))
989 * poll() never guarantees that read won't block.
990 * userfaults can be waken before they're read().
992 if (unlikely(!(file
->f_flags
& O_NONBLOCK
)))
995 * lockless access to see if there are pending faults
996 * __pollwait last action is the add_wait_queue but
997 * the spin_unlock would allow the waitqueue_active to
998 * pass above the actual list_add inside
999 * add_wait_queue critical section. So use a full
1000 * memory barrier to serialize the list_add write of
1001 * add_wait_queue() with the waitqueue_active read
1006 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1008 else if (waitqueue_active(&ctx
->event_wqh
))
1014 static const struct file_operations userfaultfd_fops
;
1016 static int resolve_userfault_fork(struct userfaultfd_ctx
*new,
1017 struct inode
*inode
,
1018 struct uffd_msg
*msg
)
1022 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, new,
1023 O_RDONLY
| (new->flags
& UFFD_SHARED_FCNTL_FLAGS
), inode
);
1027 msg
->arg
.reserved
.reserved1
= 0;
1028 msg
->arg
.fork
.ufd
= fd
;
1032 static ssize_t
userfaultfd_ctx_read(struct userfaultfd_ctx
*ctx
, int no_wait
,
1033 struct uffd_msg
*msg
, struct inode
*inode
)
1036 DECLARE_WAITQUEUE(wait
, current
);
1037 struct userfaultfd_wait_queue
*uwq
;
1039 * Handling fork event requires sleeping operations, so
1040 * we drop the event_wqh lock, then do these ops, then
1041 * lock it back and wake up the waiter. While the lock is
1042 * dropped the ewq may go away so we keep track of it
1045 LIST_HEAD(fork_event
);
1046 struct userfaultfd_ctx
*fork_nctx
= NULL
;
1048 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1049 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1050 __add_wait_queue(&ctx
->fd_wqh
, &wait
);
1052 set_current_state(TASK_INTERRUPTIBLE
);
1053 spin_lock(&ctx
->fault_pending_wqh
.lock
);
1054 uwq
= find_userfault(ctx
);
1057 * Use a seqcount to repeat the lockless check
1058 * in wake_userfault() to avoid missing
1059 * wakeups because during the refile both
1060 * waitqueue could become empty if this is the
1063 write_seqcount_begin(&ctx
->refile_seq
);
1066 * The fault_pending_wqh.lock prevents the uwq
1067 * to disappear from under us.
1069 * Refile this userfault from
1070 * fault_pending_wqh to fault_wqh, it's not
1071 * pending anymore after we read it.
1073 * Use list_del() by hand (as
1074 * userfaultfd_wake_function also uses
1075 * list_del_init() by hand) to be sure nobody
1076 * changes __remove_wait_queue() to use
1077 * list_del_init() in turn breaking the
1078 * !list_empty_careful() check in
1079 * handle_userfault(). The uwq->wq.head list
1080 * must never be empty at any time during the
1081 * refile, or the waitqueue could disappear
1082 * from under us. The "wait_queue_head_t"
1083 * parameter of __remove_wait_queue() is unused
1086 list_del(&uwq
->wq
.entry
);
1087 add_wait_queue(&ctx
->fault_wqh
, &uwq
->wq
);
1089 write_seqcount_end(&ctx
->refile_seq
);
1091 /* careful to always initialize msg if ret == 0 */
1093 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1097 spin_unlock(&ctx
->fault_pending_wqh
.lock
);
1099 spin_lock(&ctx
->event_wqh
.lock
);
1100 uwq
= find_userfault_evt(ctx
);
1104 if (uwq
->msg
.event
== UFFD_EVENT_FORK
) {
1105 fork_nctx
= (struct userfaultfd_ctx
*)
1107 uwq
->msg
.arg
.reserved
.reserved1
;
1108 list_move(&uwq
->wq
.entry
, &fork_event
);
1110 * fork_nctx can be freed as soon as
1111 * we drop the lock, unless we take a
1114 userfaultfd_ctx_get(fork_nctx
);
1115 spin_unlock(&ctx
->event_wqh
.lock
);
1120 userfaultfd_event_complete(ctx
, uwq
);
1121 spin_unlock(&ctx
->event_wqh
.lock
);
1125 spin_unlock(&ctx
->event_wqh
.lock
);
1127 if (signal_pending(current
)) {
1135 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1137 spin_lock_irq(&ctx
->fd_wqh
.lock
);
1139 __remove_wait_queue(&ctx
->fd_wqh
, &wait
);
1140 __set_current_state(TASK_RUNNING
);
1141 spin_unlock_irq(&ctx
->fd_wqh
.lock
);
1143 if (!ret
&& msg
->event
== UFFD_EVENT_FORK
) {
1144 ret
= resolve_userfault_fork(fork_nctx
, inode
, msg
);
1145 spin_lock_irq(&ctx
->event_wqh
.lock
);
1146 if (!list_empty(&fork_event
)) {
1148 * The fork thread didn't abort, so we can
1149 * drop the temporary refcount.
1151 userfaultfd_ctx_put(fork_nctx
);
1153 uwq
= list_first_entry(&fork_event
,
1157 * If fork_event list wasn't empty and in turn
1158 * the event wasn't already released by fork
1159 * (the event is allocated on fork kernel
1160 * stack), put the event back to its place in
1161 * the event_wq. fork_event head will be freed
1162 * as soon as we return so the event cannot
1163 * stay queued there no matter the current
1166 list_del(&uwq
->wq
.entry
);
1167 __add_wait_queue(&ctx
->event_wqh
, &uwq
->wq
);
1170 * Leave the event in the waitqueue and report
1171 * error to userland if we failed to resolve
1172 * the userfault fork.
1175 userfaultfd_event_complete(ctx
, uwq
);
1178 * Here the fork thread aborted and the
1179 * refcount from the fork thread on fork_nctx
1180 * has already been released. We still hold
1181 * the reference we took before releasing the
1182 * lock above. If resolve_userfault_fork
1183 * failed we've to drop it because the
1184 * fork_nctx has to be freed in such case. If
1185 * it succeeded we'll hold it because the new
1186 * uffd references it.
1189 userfaultfd_ctx_put(fork_nctx
);
1191 spin_unlock_irq(&ctx
->event_wqh
.lock
);
1197 static ssize_t
userfaultfd_read(struct file
*file
, char __user
*buf
,
1198 size_t count
, loff_t
*ppos
)
1200 struct userfaultfd_ctx
*ctx
= file
->private_data
;
1201 ssize_t _ret
, ret
= 0;
1202 struct uffd_msg msg
;
1203 int no_wait
= file
->f_flags
& O_NONBLOCK
;
1204 struct inode
*inode
= file_inode(file
);
1206 if (!userfaultfd_is_initialized(ctx
))
1210 if (count
< sizeof(msg
))
1211 return ret
? ret
: -EINVAL
;
1212 _ret
= userfaultfd_ctx_read(ctx
, no_wait
, &msg
, inode
);
1214 return ret
? ret
: _ret
;
1215 if (copy_to_user((__u64 __user
*) buf
, &msg
, sizeof(msg
)))
1216 return ret
? ret
: -EFAULT
;
1219 count
-= sizeof(msg
);
1221 * Allow to read more than one fault at time but only
1222 * block if waiting for the very first one.
1224 no_wait
= O_NONBLOCK
;
1228 static void __wake_userfault(struct userfaultfd_ctx
*ctx
,
1229 struct userfaultfd_wake_range
*range
)
1231 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
1232 /* wake all in the range and autoremove */
1233 if (waitqueue_active(&ctx
->fault_pending_wqh
))
1234 __wake_up_locked_key(&ctx
->fault_pending_wqh
, TASK_NORMAL
,
1236 if (waitqueue_active(&ctx
->fault_wqh
))
1237 __wake_up(&ctx
->fault_wqh
, TASK_NORMAL
, 1, range
);
1238 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
1241 static __always_inline
void wake_userfault(struct userfaultfd_ctx
*ctx
,
1242 struct userfaultfd_wake_range
*range
)
1248 * To be sure waitqueue_active() is not reordered by the CPU
1249 * before the pagetable update, use an explicit SMP memory
1250 * barrier here. PT lock release or mmap_read_unlock(mm) still
1251 * have release semantics that can allow the
1252 * waitqueue_active() to be reordered before the pte update.
1257 * Use waitqueue_active because it's very frequent to
1258 * change the address space atomically even if there are no
1259 * userfaults yet. So we take the spinlock only when we're
1260 * sure we've userfaults to wake.
1263 seq
= read_seqcount_begin(&ctx
->refile_seq
);
1264 need_wakeup
= waitqueue_active(&ctx
->fault_pending_wqh
) ||
1265 waitqueue_active(&ctx
->fault_wqh
);
1267 } while (read_seqcount_retry(&ctx
->refile_seq
, seq
));
1269 __wake_userfault(ctx
, range
);
1272 static __always_inline
int validate_range(struct mm_struct
*mm
,
1273 __u64 start
, __u64 len
)
1275 __u64 task_size
= mm
->task_size
;
1277 if (start
& ~PAGE_MASK
)
1279 if (len
& ~PAGE_MASK
)
1283 if (start
< mmap_min_addr
)
1285 if (start
>= task_size
)
1287 if (len
> task_size
- start
)
1292 static int userfaultfd_register(struct userfaultfd_ctx
*ctx
,
1295 struct mm_struct
*mm
= ctx
->mm
;
1296 struct vm_area_struct
*vma
, *prev
, *cur
;
1298 struct uffdio_register uffdio_register
;
1299 struct uffdio_register __user
*user_uffdio_register
;
1300 unsigned long vm_flags
, new_flags
;
1303 unsigned long start
, end
, vma_end
;
1304 struct vma_iterator vmi
;
1306 user_uffdio_register
= (struct uffdio_register __user
*) arg
;
1309 if (copy_from_user(&uffdio_register
, user_uffdio_register
,
1310 sizeof(uffdio_register
)-sizeof(__u64
)))
1314 if (!uffdio_register
.mode
)
1316 if (uffdio_register
.mode
& ~UFFD_API_REGISTER_MODES
)
1319 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MISSING
)
1320 vm_flags
|= VM_UFFD_MISSING
;
1321 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
) {
1322 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1325 vm_flags
|= VM_UFFD_WP
;
1327 if (uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
) {
1328 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1331 vm_flags
|= VM_UFFD_MINOR
;
1334 ret
= validate_range(mm
, uffdio_register
.range
.start
,
1335 uffdio_register
.range
.len
);
1339 start
= uffdio_register
.range
.start
;
1340 end
= start
+ uffdio_register
.range
.len
;
1343 if (!mmget_not_zero(mm
))
1347 mmap_write_lock(mm
);
1348 vma_iter_init(&vmi
, mm
, start
);
1349 vma
= vma_find(&vmi
, end
);
1354 * If the first vma contains huge pages, make sure start address
1355 * is aligned to huge page size.
1357 if (is_vm_hugetlb_page(vma
)) {
1358 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1360 if (start
& (vma_hpagesize
- 1))
1365 * Search for not compatible vmas.
1368 basic_ioctls
= false;
1373 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1374 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1376 /* check not compatible vmas */
1378 if (!vma_can_userfault(cur
, vm_flags
))
1382 * UFFDIO_COPY will fill file holes even without
1383 * PROT_WRITE. This check enforces that if this is a
1384 * MAP_SHARED, the process has write permission to the backing
1385 * file. If VM_MAYWRITE is set it also enforces that on a
1386 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1387 * F_WRITE_SEAL can be taken until the vma is destroyed.
1390 if (unlikely(!(cur
->vm_flags
& VM_MAYWRITE
)))
1394 * If this vma contains ending address, and huge pages
1397 if (is_vm_hugetlb_page(cur
) && end
<= cur
->vm_end
&&
1398 end
> cur
->vm_start
) {
1399 unsigned long vma_hpagesize
= vma_kernel_pagesize(cur
);
1403 if (end
& (vma_hpagesize
- 1))
1406 if ((vm_flags
& VM_UFFD_WP
) && !(cur
->vm_flags
& VM_MAYWRITE
))
1410 * Check that this vma isn't already owned by a
1411 * different userfaultfd. We can't allow more than one
1412 * userfaultfd to own a single vma simultaneously or we
1413 * wouldn't know which one to deliver the userfaults to.
1416 if (cur
->vm_userfaultfd_ctx
.ctx
&&
1417 cur
->vm_userfaultfd_ctx
.ctx
!= ctx
)
1421 * Note vmas containing huge pages
1423 if (is_vm_hugetlb_page(cur
))
1424 basic_ioctls
= true;
1427 } for_each_vma_range(vmi
, cur
, end
);
1430 vma_iter_set(&vmi
, start
);
1431 prev
= vma_prev(&vmi
);
1434 for_each_vma_range(vmi
, vma
, end
) {
1437 BUG_ON(!vma_can_userfault(vma
, vm_flags
));
1438 BUG_ON(vma
->vm_userfaultfd_ctx
.ctx
&&
1439 vma
->vm_userfaultfd_ctx
.ctx
!= ctx
);
1440 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1443 * Nothing to do: this vma is already registered into this
1444 * userfaultfd and with the right tracking mode too.
1446 if (vma
->vm_userfaultfd_ctx
.ctx
== ctx
&&
1447 (vma
->vm_flags
& vm_flags
) == vm_flags
)
1450 if (vma
->vm_start
> start
)
1451 start
= vma
->vm_start
;
1452 vma_end
= min(end
, vma
->vm_end
);
1454 new_flags
= (vma
->vm_flags
& ~__VM_UFFD_FLAGS
) | vm_flags
;
1455 prev
= vma_merge(&vmi
, mm
, prev
, start
, vma_end
, new_flags
,
1456 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1458 ((struct vm_userfaultfd_ctx
){ ctx
}),
1459 anon_vma_name(vma
));
1461 /* vma_merge() invalidated the mas */
1465 if (vma
->vm_start
< start
) {
1466 ret
= split_vma(&vmi
, vma
, start
, 1);
1470 if (vma
->vm_end
> end
) {
1471 ret
= split_vma(&vmi
, vma
, end
, 0);
1477 * In the vma_merge() successful mprotect-like case 8:
1478 * the next vma was merged into the current one and
1479 * the current one has not been updated yet.
1481 userfaultfd_set_vm_flags(vma
, new_flags
);
1482 vma
->vm_userfaultfd_ctx
.ctx
= ctx
;
1484 if (is_vm_hugetlb_page(vma
) && uffd_disable_huge_pmd_share(vma
))
1485 hugetlb_unshare_all_pmds(vma
);
1489 start
= vma
->vm_end
;
1493 mmap_write_unlock(mm
);
1498 ioctls_out
= basic_ioctls
? UFFD_API_RANGE_IOCTLS_BASIC
:
1499 UFFD_API_RANGE_IOCTLS
;
1502 * Declare the WP ioctl only if the WP mode is
1503 * specified and all checks passed with the range
1505 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_WP
))
1506 ioctls_out
&= ~((__u64
)1 << _UFFDIO_WRITEPROTECT
);
1508 /* CONTINUE ioctl is only supported for MINOR ranges. */
1509 if (!(uffdio_register
.mode
& UFFDIO_REGISTER_MODE_MINOR
))
1510 ioctls_out
&= ~((__u64
)1 << _UFFDIO_CONTINUE
);
1513 * Now that we scanned all vmas we can already tell
1514 * userland which ioctls methods are guaranteed to
1515 * succeed on this range.
1517 if (put_user(ioctls_out
, &user_uffdio_register
->ioctls
))
1524 static int userfaultfd_unregister(struct userfaultfd_ctx
*ctx
,
1527 struct mm_struct
*mm
= ctx
->mm
;
1528 struct vm_area_struct
*vma
, *prev
, *cur
;
1530 struct uffdio_range uffdio_unregister
;
1531 unsigned long new_flags
;
1533 unsigned long start
, end
, vma_end
;
1534 const void __user
*buf
= (void __user
*)arg
;
1535 struct vma_iterator vmi
;
1538 if (copy_from_user(&uffdio_unregister
, buf
, sizeof(uffdio_unregister
)))
1541 ret
= validate_range(mm
, uffdio_unregister
.start
,
1542 uffdio_unregister
.len
);
1546 start
= uffdio_unregister
.start
;
1547 end
= start
+ uffdio_unregister
.len
;
1550 if (!mmget_not_zero(mm
))
1553 mmap_write_lock(mm
);
1555 vma_iter_init(&vmi
, mm
, start
);
1556 vma
= vma_find(&vmi
, end
);
1561 * If the first vma contains huge pages, make sure start address
1562 * is aligned to huge page size.
1564 if (is_vm_hugetlb_page(vma
)) {
1565 unsigned long vma_hpagesize
= vma_kernel_pagesize(vma
);
1567 if (start
& (vma_hpagesize
- 1))
1572 * Search for not compatible vmas.
1579 BUG_ON(!!cur
->vm_userfaultfd_ctx
.ctx
^
1580 !!(cur
->vm_flags
& __VM_UFFD_FLAGS
));
1583 * Check not compatible vmas, not strictly required
1584 * here as not compatible vmas cannot have an
1585 * userfaultfd_ctx registered on them, but this
1586 * provides for more strict behavior to notice
1587 * unregistration errors.
1589 if (!vma_can_userfault(cur
, cur
->vm_flags
))
1593 } for_each_vma_range(vmi
, cur
, end
);
1596 vma_iter_set(&vmi
, start
);
1597 prev
= vma_prev(&vmi
);
1599 for_each_vma_range(vmi
, vma
, end
) {
1602 BUG_ON(!vma_can_userfault(vma
, vma
->vm_flags
));
1605 * Nothing to do: this vma is already registered into this
1606 * userfaultfd and with the right tracking mode too.
1608 if (!vma
->vm_userfaultfd_ctx
.ctx
)
1611 WARN_ON(!(vma
->vm_flags
& VM_MAYWRITE
));
1613 if (vma
->vm_start
> start
)
1614 start
= vma
->vm_start
;
1615 vma_end
= min(end
, vma
->vm_end
);
1617 if (userfaultfd_missing(vma
)) {
1619 * Wake any concurrent pending userfault while
1620 * we unregister, so they will not hang
1621 * permanently and it avoids userland to call
1622 * UFFDIO_WAKE explicitly.
1624 struct userfaultfd_wake_range range
;
1625 range
.start
= start
;
1626 range
.len
= vma_end
- start
;
1627 wake_userfault(vma
->vm_userfaultfd_ctx
.ctx
, &range
);
1630 /* Reset ptes for the whole vma range if wr-protected */
1631 if (userfaultfd_wp(vma
))
1632 uffd_wp_range(mm
, vma
, start
, vma_end
- start
, false);
1634 new_flags
= vma
->vm_flags
& ~__VM_UFFD_FLAGS
;
1635 prev
= vma_merge(&vmi
, mm
, prev
, start
, vma_end
, new_flags
,
1636 vma
->anon_vma
, vma
->vm_file
, vma
->vm_pgoff
,
1638 NULL_VM_UFFD_CTX
, anon_vma_name(vma
));
1643 if (vma
->vm_start
< start
) {
1644 ret
= split_vma(&vmi
, vma
, start
, 1);
1648 if (vma
->vm_end
> end
) {
1649 ret
= split_vma(&vmi
, vma
, end
, 0);
1655 * In the vma_merge() successful mprotect-like case 8:
1656 * the next vma was merged into the current one and
1657 * the current one has not been updated yet.
1659 userfaultfd_set_vm_flags(vma
, new_flags
);
1660 vma
->vm_userfaultfd_ctx
= NULL_VM_UFFD_CTX
;
1664 start
= vma
->vm_end
;
1668 mmap_write_unlock(mm
);
1675 * userfaultfd_wake may be used in combination with the
1676 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1678 static int userfaultfd_wake(struct userfaultfd_ctx
*ctx
,
1682 struct uffdio_range uffdio_wake
;
1683 struct userfaultfd_wake_range range
;
1684 const void __user
*buf
= (void __user
*)arg
;
1687 if (copy_from_user(&uffdio_wake
, buf
, sizeof(uffdio_wake
)))
1690 ret
= validate_range(ctx
->mm
, uffdio_wake
.start
, uffdio_wake
.len
);
1694 range
.start
= uffdio_wake
.start
;
1695 range
.len
= uffdio_wake
.len
;
1698 * len == 0 means wake all and we don't want to wake all here,
1699 * so check it again to be sure.
1701 VM_BUG_ON(!range
.len
);
1703 wake_userfault(ctx
, &range
);
1710 static int userfaultfd_copy(struct userfaultfd_ctx
*ctx
,
1714 struct uffdio_copy uffdio_copy
;
1715 struct uffdio_copy __user
*user_uffdio_copy
;
1716 struct userfaultfd_wake_range range
;
1718 user_uffdio_copy
= (struct uffdio_copy __user
*) arg
;
1721 if (atomic_read(&ctx
->mmap_changing
))
1725 if (copy_from_user(&uffdio_copy
, user_uffdio_copy
,
1726 /* don't copy "copy" last field */
1727 sizeof(uffdio_copy
)-sizeof(__s64
)))
1730 ret
= validate_range(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.len
);
1734 * double check for wraparound just in case. copy_from_user()
1735 * will later check uffdio_copy.src + uffdio_copy.len to fit
1736 * in the userland range.
1739 if (uffdio_copy
.src
+ uffdio_copy
.len
<= uffdio_copy
.src
)
1741 if (uffdio_copy
.mode
& ~(UFFDIO_COPY_MODE_DONTWAKE
|UFFDIO_COPY_MODE_WP
))
1743 if (mmget_not_zero(ctx
->mm
)) {
1744 ret
= mcopy_atomic(ctx
->mm
, uffdio_copy
.dst
, uffdio_copy
.src
,
1745 uffdio_copy
.len
, &ctx
->mmap_changing
,
1751 if (unlikely(put_user(ret
, &user_uffdio_copy
->copy
)))
1756 /* len == 0 would wake all */
1758 if (!(uffdio_copy
.mode
& UFFDIO_COPY_MODE_DONTWAKE
)) {
1759 range
.start
= uffdio_copy
.dst
;
1760 wake_userfault(ctx
, &range
);
1762 ret
= range
.len
== uffdio_copy
.len
? 0 : -EAGAIN
;
1767 static int userfaultfd_zeropage(struct userfaultfd_ctx
*ctx
,
1771 struct uffdio_zeropage uffdio_zeropage
;
1772 struct uffdio_zeropage __user
*user_uffdio_zeropage
;
1773 struct userfaultfd_wake_range range
;
1775 user_uffdio_zeropage
= (struct uffdio_zeropage __user
*) arg
;
1778 if (atomic_read(&ctx
->mmap_changing
))
1782 if (copy_from_user(&uffdio_zeropage
, user_uffdio_zeropage
,
1783 /* don't copy "zeropage" last field */
1784 sizeof(uffdio_zeropage
)-sizeof(__s64
)))
1787 ret
= validate_range(ctx
->mm
, uffdio_zeropage
.range
.start
,
1788 uffdio_zeropage
.range
.len
);
1792 if (uffdio_zeropage
.mode
& ~UFFDIO_ZEROPAGE_MODE_DONTWAKE
)
1795 if (mmget_not_zero(ctx
->mm
)) {
1796 ret
= mfill_zeropage(ctx
->mm
, uffdio_zeropage
.range
.start
,
1797 uffdio_zeropage
.range
.len
,
1798 &ctx
->mmap_changing
);
1803 if (unlikely(put_user(ret
, &user_uffdio_zeropage
->zeropage
)))
1807 /* len == 0 would wake all */
1810 if (!(uffdio_zeropage
.mode
& UFFDIO_ZEROPAGE_MODE_DONTWAKE
)) {
1811 range
.start
= uffdio_zeropage
.range
.start
;
1812 wake_userfault(ctx
, &range
);
1814 ret
= range
.len
== uffdio_zeropage
.range
.len
? 0 : -EAGAIN
;
1819 static int userfaultfd_writeprotect(struct userfaultfd_ctx
*ctx
,
1823 struct uffdio_writeprotect uffdio_wp
;
1824 struct uffdio_writeprotect __user
*user_uffdio_wp
;
1825 struct userfaultfd_wake_range range
;
1826 bool mode_wp
, mode_dontwake
;
1828 if (atomic_read(&ctx
->mmap_changing
))
1831 user_uffdio_wp
= (struct uffdio_writeprotect __user
*) arg
;
1833 if (copy_from_user(&uffdio_wp
, user_uffdio_wp
,
1834 sizeof(struct uffdio_writeprotect
)))
1837 ret
= validate_range(ctx
->mm
, uffdio_wp
.range
.start
,
1838 uffdio_wp
.range
.len
);
1842 if (uffdio_wp
.mode
& ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE
|
1843 UFFDIO_WRITEPROTECT_MODE_WP
))
1846 mode_wp
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_WP
;
1847 mode_dontwake
= uffdio_wp
.mode
& UFFDIO_WRITEPROTECT_MODE_DONTWAKE
;
1849 if (mode_wp
&& mode_dontwake
)
1852 if (mmget_not_zero(ctx
->mm
)) {
1853 ret
= mwriteprotect_range(ctx
->mm
, uffdio_wp
.range
.start
,
1854 uffdio_wp
.range
.len
, mode_wp
,
1855 &ctx
->mmap_changing
);
1864 if (!mode_wp
&& !mode_dontwake
) {
1865 range
.start
= uffdio_wp
.range
.start
;
1866 range
.len
= uffdio_wp
.range
.len
;
1867 wake_userfault(ctx
, &range
);
1872 static int userfaultfd_continue(struct userfaultfd_ctx
*ctx
, unsigned long arg
)
1875 struct uffdio_continue uffdio_continue
;
1876 struct uffdio_continue __user
*user_uffdio_continue
;
1877 struct userfaultfd_wake_range range
;
1879 user_uffdio_continue
= (struct uffdio_continue __user
*)arg
;
1882 if (atomic_read(&ctx
->mmap_changing
))
1886 if (copy_from_user(&uffdio_continue
, user_uffdio_continue
,
1887 /* don't copy the output fields */
1888 sizeof(uffdio_continue
) - (sizeof(__s64
))))
1891 ret
= validate_range(ctx
->mm
, uffdio_continue
.range
.start
,
1892 uffdio_continue
.range
.len
);
1897 /* double check for wraparound just in case. */
1898 if (uffdio_continue
.range
.start
+ uffdio_continue
.range
.len
<=
1899 uffdio_continue
.range
.start
) {
1902 if (uffdio_continue
.mode
& ~UFFDIO_CONTINUE_MODE_DONTWAKE
)
1905 if (mmget_not_zero(ctx
->mm
)) {
1906 ret
= mcopy_continue(ctx
->mm
, uffdio_continue
.range
.start
,
1907 uffdio_continue
.range
.len
,
1908 &ctx
->mmap_changing
);
1914 if (unlikely(put_user(ret
, &user_uffdio_continue
->mapped
)))
1919 /* len == 0 would wake all */
1922 if (!(uffdio_continue
.mode
& UFFDIO_CONTINUE_MODE_DONTWAKE
)) {
1923 range
.start
= uffdio_continue
.range
.start
;
1924 wake_userfault(ctx
, &range
);
1926 ret
= range
.len
== uffdio_continue
.range
.len
? 0 : -EAGAIN
;
1932 static inline unsigned int uffd_ctx_features(__u64 user_features
)
1935 * For the current set of features the bits just coincide. Set
1936 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1938 return (unsigned int)user_features
| UFFD_FEATURE_INITIALIZED
;
1942 * userland asks for a certain API version and we return which bits
1943 * and ioctl commands are implemented in this kernel for such API
1944 * version or -EINVAL if unknown.
1946 static int userfaultfd_api(struct userfaultfd_ctx
*ctx
,
1949 struct uffdio_api uffdio_api
;
1950 void __user
*buf
= (void __user
*)arg
;
1951 unsigned int ctx_features
;
1956 if (copy_from_user(&uffdio_api
, buf
, sizeof(uffdio_api
)))
1958 /* Ignore unsupported features (userspace built against newer kernel) */
1959 features
= uffdio_api
.features
& UFFD_API_FEATURES
;
1961 if ((features
& UFFD_FEATURE_EVENT_FORK
) && !capable(CAP_SYS_PTRACE
))
1963 /* report all available features and ioctls to userland */
1964 uffdio_api
.features
= UFFD_API_FEATURES
;
1965 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1966 uffdio_api
.features
&=
1967 ~(UFFD_FEATURE_MINOR_HUGETLBFS
| UFFD_FEATURE_MINOR_SHMEM
);
1969 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1970 uffdio_api
.features
&= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP
;
1972 #ifndef CONFIG_PTE_MARKER_UFFD_WP
1973 uffdio_api
.features
&= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM
;
1975 uffdio_api
.ioctls
= UFFD_API_IOCTLS
;
1977 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1980 /* only enable the requested features for this uffd context */
1981 ctx_features
= uffd_ctx_features(features
);
1983 if (cmpxchg(&ctx
->features
, 0, ctx_features
) != 0)
1990 memset(&uffdio_api
, 0, sizeof(uffdio_api
));
1991 if (copy_to_user(buf
, &uffdio_api
, sizeof(uffdio_api
)))
1996 static long userfaultfd_ioctl(struct file
*file
, unsigned cmd
,
2000 struct userfaultfd_ctx
*ctx
= file
->private_data
;
2002 if (cmd
!= UFFDIO_API
&& !userfaultfd_is_initialized(ctx
))
2007 ret
= userfaultfd_api(ctx
, arg
);
2009 case UFFDIO_REGISTER
:
2010 ret
= userfaultfd_register(ctx
, arg
);
2012 case UFFDIO_UNREGISTER
:
2013 ret
= userfaultfd_unregister(ctx
, arg
);
2016 ret
= userfaultfd_wake(ctx
, arg
);
2019 ret
= userfaultfd_copy(ctx
, arg
);
2021 case UFFDIO_ZEROPAGE
:
2022 ret
= userfaultfd_zeropage(ctx
, arg
);
2024 case UFFDIO_WRITEPROTECT
:
2025 ret
= userfaultfd_writeprotect(ctx
, arg
);
2027 case UFFDIO_CONTINUE
:
2028 ret
= userfaultfd_continue(ctx
, arg
);
2034 #ifdef CONFIG_PROC_FS
2035 static void userfaultfd_show_fdinfo(struct seq_file
*m
, struct file
*f
)
2037 struct userfaultfd_ctx
*ctx
= f
->private_data
;
2038 wait_queue_entry_t
*wq
;
2039 unsigned long pending
= 0, total
= 0;
2041 spin_lock_irq(&ctx
->fault_pending_wqh
.lock
);
2042 list_for_each_entry(wq
, &ctx
->fault_pending_wqh
.head
, entry
) {
2046 list_for_each_entry(wq
, &ctx
->fault_wqh
.head
, entry
) {
2049 spin_unlock_irq(&ctx
->fault_pending_wqh
.lock
);
2052 * If more protocols will be added, there will be all shown
2053 * separated by a space. Like this:
2054 * protocols: aa:... bb:...
2056 seq_printf(m
, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2057 pending
, total
, UFFD_API
, ctx
->features
,
2058 UFFD_API_IOCTLS
|UFFD_API_RANGE_IOCTLS
);
2062 static const struct file_operations userfaultfd_fops
= {
2063 #ifdef CONFIG_PROC_FS
2064 .show_fdinfo
= userfaultfd_show_fdinfo
,
2066 .release
= userfaultfd_release
,
2067 .poll
= userfaultfd_poll
,
2068 .read
= userfaultfd_read
,
2069 .unlocked_ioctl
= userfaultfd_ioctl
,
2070 .compat_ioctl
= compat_ptr_ioctl
,
2071 .llseek
= noop_llseek
,
2074 static void init_once_userfaultfd_ctx(void *mem
)
2076 struct userfaultfd_ctx
*ctx
= (struct userfaultfd_ctx
*) mem
;
2078 init_waitqueue_head(&ctx
->fault_pending_wqh
);
2079 init_waitqueue_head(&ctx
->fault_wqh
);
2080 init_waitqueue_head(&ctx
->event_wqh
);
2081 init_waitqueue_head(&ctx
->fd_wqh
);
2082 seqcount_spinlock_init(&ctx
->refile_seq
, &ctx
->fault_pending_wqh
.lock
);
2085 static int new_userfaultfd(int flags
)
2087 struct userfaultfd_ctx
*ctx
;
2090 BUG_ON(!current
->mm
);
2092 /* Check the UFFD_* constants for consistency. */
2093 BUILD_BUG_ON(UFFD_USER_MODE_ONLY
& UFFD_SHARED_FCNTL_FLAGS
);
2094 BUILD_BUG_ON(UFFD_CLOEXEC
!= O_CLOEXEC
);
2095 BUILD_BUG_ON(UFFD_NONBLOCK
!= O_NONBLOCK
);
2097 if (flags
& ~(UFFD_SHARED_FCNTL_FLAGS
| UFFD_USER_MODE_ONLY
))
2100 ctx
= kmem_cache_alloc(userfaultfd_ctx_cachep
, GFP_KERNEL
);
2104 refcount_set(&ctx
->refcount
, 1);
2107 ctx
->released
= false;
2108 atomic_set(&ctx
->mmap_changing
, 0);
2109 ctx
->mm
= current
->mm
;
2110 /* prevent the mm struct to be freed */
2113 fd
= anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops
, ctx
,
2114 O_RDONLY
| (flags
& UFFD_SHARED_FCNTL_FLAGS
), NULL
);
2117 kmem_cache_free(userfaultfd_ctx_cachep
, ctx
);
2122 static inline bool userfaultfd_syscall_allowed(int flags
)
2124 /* Userspace-only page faults are always allowed */
2125 if (flags
& UFFD_USER_MODE_ONLY
)
2129 * The user is requesting a userfaultfd which can handle kernel faults.
2130 * Privileged users are always allowed to do this.
2132 if (capable(CAP_SYS_PTRACE
))
2135 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2136 return sysctl_unprivileged_userfaultfd
;
2139 SYSCALL_DEFINE1(userfaultfd
, int, flags
)
2141 if (!userfaultfd_syscall_allowed(flags
))
2144 return new_userfaultfd(flags
);
2147 static long userfaultfd_dev_ioctl(struct file
*file
, unsigned int cmd
, unsigned long flags
)
2149 if (cmd
!= USERFAULTFD_IOC_NEW
)
2152 return new_userfaultfd(flags
);
2155 static const struct file_operations userfaultfd_dev_fops
= {
2156 .unlocked_ioctl
= userfaultfd_dev_ioctl
,
2157 .compat_ioctl
= userfaultfd_dev_ioctl
,
2158 .owner
= THIS_MODULE
,
2159 .llseek
= noop_llseek
,
2162 static struct miscdevice userfaultfd_misc
= {
2163 .minor
= MISC_DYNAMIC_MINOR
,
2164 .name
= "userfaultfd",
2165 .fops
= &userfaultfd_dev_fops
2168 static int __init
userfaultfd_init(void)
2172 ret
= misc_register(&userfaultfd_misc
);
2176 userfaultfd_ctx_cachep
= kmem_cache_create("userfaultfd_ctx_cache",
2177 sizeof(struct userfaultfd_ctx
),
2179 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
,
2180 init_once_userfaultfd_ctx
);
2183 __initcall(userfaultfd_init
);