4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
5 * Copyright (C) 2008-2009 Red Hat, Inc.
6 * Copyright (C) 2015 Red Hat, Inc.
8 * This work is licensed under the terms of the GNU GPL, version 2. See
9 * the COPYING file in the top-level directory.
11 * Some part derived from fs/eventfd.c (anon inode setup) and
12 * mm/ksm.c (mm hashing).
15 #include <linux/list.h>
16 #include <linux/hashtable.h>
17 #include <linux/sched/signal.h>
18 #include <linux/sched/mm.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>
33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
35 enum userfaultfd_state {
41 * Start with fault_pending_wqh and fault_wqh so they're more likely
42 * to be in the same cacheline.
44 struct userfaultfd_ctx {
45 /* waitqueue head for the pending (i.e. not read) userfaults */
46 wait_queue_head_t fault_pending_wqh;
47 /* waitqueue head for the userfaults */
48 wait_queue_head_t fault_wqh;
49 /* waitqueue head for the pseudo fd to wakeup poll/read */
50 wait_queue_head_t fd_wqh;
51 /* waitqueue head for events */
52 wait_queue_head_t event_wqh;
53 /* a refile sequence protected by fault_pending_wqh lock */
54 struct seqcount refile_seq;
55 /* pseudo fd refcounting */
57 /* userfaultfd syscall flags */
59 /* features requested from the userspace */
60 unsigned int features;
62 enum userfaultfd_state state;
65 /* mm with one ore more vmas attached to this userfaultfd_ctx */
69 struct userfaultfd_fork_ctx {
70 struct userfaultfd_ctx *orig;
71 struct userfaultfd_ctx *new;
72 struct list_head list;
75 struct userfaultfd_unmap_ctx {
76 struct userfaultfd_ctx *ctx;
79 struct list_head list;
82 struct userfaultfd_wait_queue {
85 struct userfaultfd_ctx *ctx;
89 struct userfaultfd_wake_range {
94 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
95 int wake_flags, void *key)
97 struct userfaultfd_wake_range *range = key;
99 struct userfaultfd_wait_queue *uwq;
100 unsigned long start, len;
102 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
104 /* len == 0 means wake all */
105 start = range->start;
107 if (len && (start > uwq->msg.arg.pagefault.address ||
108 start + len <= uwq->msg.arg.pagefault.address))
110 WRITE_ONCE(uwq->waken, true);
112 * The implicit smp_mb__before_spinlock in try_to_wake_up()
113 * renders uwq->waken visible to other CPUs before the task is
116 ret = wake_up_state(wq->private, mode);
119 * Wake only once, autoremove behavior.
121 * After the effect of list_del_init is visible to the
122 * other CPUs, the waitqueue may disappear from under
123 * us, see the !list_empty_careful() in
124 * handle_userfault(). try_to_wake_up() has an
125 * implicit smp_mb__before_spinlock, and the
126 * wq->private is read before calling the extern
127 * function "wake_up_state" (which in turns calls
128 * try_to_wake_up). While the spin_lock;spin_unlock;
129 * wouldn't be enough, the smp_mb__before_spinlock is
130 * enough to avoid an explicit smp_mb() here.
132 list_del_init(&wq->task_list);
138 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
140 * @ctx: [in] Pointer to the userfaultfd context.
142 * Returns: In case of success, returns not zero.
144 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
146 if (!atomic_inc_not_zero(&ctx->refcount))
151 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
153 * @ctx: [in] Pointer to userfaultfd context.
155 * The userfaultfd context reference must have been previously acquired either
156 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
158 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
160 if (atomic_dec_and_test(&ctx->refcount)) {
161 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
162 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
163 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
164 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
165 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
166 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
167 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
168 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
170 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
174 static inline void msg_init(struct uffd_msg *msg)
176 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
178 * Must use memset to zero out the paddings or kernel data is
179 * leaked to userland.
181 memset(msg, 0, sizeof(struct uffd_msg));
184 static inline struct uffd_msg userfault_msg(unsigned long address,
186 unsigned long reason)
190 msg.event = UFFD_EVENT_PAGEFAULT;
191 msg.arg.pagefault.address = address;
192 if (flags & FAULT_FLAG_WRITE)
194 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
195 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
196 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
197 * was a read fault, otherwise if set it means it's
200 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
201 if (reason & VM_UFFD_WP)
203 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
204 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
205 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
206 * a missing fault, otherwise if set it means it's a
207 * write protect fault.
209 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
213 #ifdef CONFIG_HUGETLB_PAGE
215 * Same functionality as userfaultfd_must_wait below with modifications for
218 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
219 unsigned long address,
221 unsigned long reason)
223 struct mm_struct *mm = ctx->mm;
227 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
229 pte = huge_pte_offset(mm, address);
236 * Lockless access: we're in a wait_event so it's ok if it
239 if (huge_pte_none(*pte))
241 if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
247 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
248 unsigned long address,
250 unsigned long reason)
252 return false; /* should never get here */
254 #endif /* CONFIG_HUGETLB_PAGE */
257 * Verify the pagetables are still not ok after having reigstered into
258 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
259 * userfault that has already been resolved, if userfaultfd_read and
260 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
263 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
264 unsigned long address,
266 unsigned long reason)
268 struct mm_struct *mm = ctx->mm;
275 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
277 pgd = pgd_offset(mm, address);
278 if (!pgd_present(*pgd))
280 pud = pud_offset(pgd, address);
281 if (!pud_present(*pud))
283 pmd = pmd_offset(pud, address);
285 * READ_ONCE must function as a barrier with narrower scope
286 * and it must be equivalent to:
287 * _pmd = *pmd; barrier();
289 * This is to deal with the instability (as in
290 * pmd_trans_unstable) of the pmd.
292 _pmd = READ_ONCE(*pmd);
293 if (!pmd_present(_pmd))
297 if (pmd_trans_huge(_pmd))
301 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
302 * and use the standard pte_offset_map() instead of parsing _pmd.
304 pte = pte_offset_map(pmd, address);
306 * Lockless access: we're in a wait_event so it's ok if it
318 * The locking rules involved in returning VM_FAULT_RETRY depending on
319 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
320 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
321 * recommendation in __lock_page_or_retry is not an understatement.
323 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
324 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
327 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
328 * set, VM_FAULT_RETRY can still be returned if and only if there are
329 * fatal_signal_pending()s, and the mmap_sem must be released before
332 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
334 struct mm_struct *mm = vmf->vma->vm_mm;
335 struct userfaultfd_ctx *ctx;
336 struct userfaultfd_wait_queue uwq;
338 bool must_wait, return_to_userland;
341 BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
343 ret = VM_FAULT_SIGBUS;
344 ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
348 BUG_ON(ctx->mm != mm);
350 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
351 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
354 * If it's already released don't get it. This avoids to loop
355 * in __get_user_pages if userfaultfd_release waits on the
356 * caller of handle_userfault to release the mmap_sem.
358 if (unlikely(ACCESS_ONCE(ctx->released)))
362 * We don't do userfault handling for the final child pid update.
364 if (current->flags & PF_EXITING)
368 * Check that we can return VM_FAULT_RETRY.
370 * NOTE: it should become possible to return VM_FAULT_RETRY
371 * even if FAULT_FLAG_TRIED is set without leading to gup()
372 * -EBUSY failures, if the userfaultfd is to be extended for
373 * VM_UFFD_WP tracking and we intend to arm the userfault
374 * without first stopping userland access to the memory. For
375 * VM_UFFD_MISSING userfaults this is enough for now.
377 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
379 * Validate the invariant that nowait must allow retry
380 * to be sure not to return SIGBUS erroneously on
381 * nowait invocations.
383 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
384 #ifdef CONFIG_DEBUG_VM
385 if (printk_ratelimit()) {
387 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
396 * Handle nowait, not much to do other than tell it to retry
399 ret = VM_FAULT_RETRY;
400 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
403 /* take the reference before dropping the mmap_sem */
404 userfaultfd_ctx_get(ctx);
406 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
407 uwq.wq.private = current;
408 uwq.msg = userfault_msg(vmf->address, vmf->flags, reason);
413 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
414 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
415 blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
418 spin_lock(&ctx->fault_pending_wqh.lock);
420 * After the __add_wait_queue the uwq is visible to userland
421 * through poll/read().
423 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
425 * The smp_mb() after __set_current_state prevents the reads
426 * following the spin_unlock to happen before the list_add in
429 set_current_state(blocking_state);
430 spin_unlock(&ctx->fault_pending_wqh.lock);
432 if (!is_vm_hugetlb_page(vmf->vma))
433 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
436 must_wait = userfaultfd_huge_must_wait(ctx, vmf->address,
438 up_read(&mm->mmap_sem);
440 if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
441 (return_to_userland ? !signal_pending(current) :
442 !fatal_signal_pending(current)))) {
443 wake_up_poll(&ctx->fd_wqh, POLLIN);
445 ret |= VM_FAULT_MAJOR;
448 * False wakeups can orginate even from rwsem before
449 * up_read() however userfaults will wait either for a
450 * targeted wakeup on the specific uwq waitqueue from
451 * wake_userfault() or for signals or for uffd
454 while (!READ_ONCE(uwq.waken)) {
456 * This needs the full smp_store_mb()
457 * guarantee as the state write must be
458 * visible to other CPUs before reading
459 * uwq.waken from other CPUs.
461 set_current_state(blocking_state);
462 if (READ_ONCE(uwq.waken) ||
463 READ_ONCE(ctx->released) ||
464 (return_to_userland ? signal_pending(current) :
465 fatal_signal_pending(current)))
471 __set_current_state(TASK_RUNNING);
473 if (return_to_userland) {
474 if (signal_pending(current) &&
475 !fatal_signal_pending(current)) {
477 * If we got a SIGSTOP or SIGCONT and this is
478 * a normal userland page fault, just let
479 * userland return so the signal will be
480 * handled and gdb debugging works. The page
481 * fault code immediately after we return from
482 * this function is going to release the
483 * mmap_sem and it's not depending on it
484 * (unlike gup would if we were not to return
487 * If a fatal signal is pending we still take
488 * the streamlined VM_FAULT_RETRY failure path
489 * and there's no need to retake the mmap_sem
492 down_read(&mm->mmap_sem);
493 ret = VM_FAULT_NOPAGE;
498 * Here we race with the list_del; list_add in
499 * userfaultfd_ctx_read(), however because we don't ever run
500 * list_del_init() to refile across the two lists, the prev
501 * and next pointers will never point to self. list_add also
502 * would never let any of the two pointers to point to
503 * self. So list_empty_careful won't risk to see both pointers
504 * pointing to self at any time during the list refile. The
505 * only case where list_del_init() is called is the full
506 * removal in the wake function and there we don't re-list_add
507 * and it's fine not to block on the spinlock. The uwq on this
508 * kernel stack can be released after the list_del_init.
510 if (!list_empty_careful(&uwq.wq.task_list)) {
511 spin_lock(&ctx->fault_pending_wqh.lock);
513 * No need of list_del_init(), the uwq on the stack
514 * will be freed shortly anyway.
516 list_del(&uwq.wq.task_list);
517 spin_unlock(&ctx->fault_pending_wqh.lock);
521 * ctx may go away after this if the userfault pseudo fd is
524 userfaultfd_ctx_put(ctx);
530 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
531 struct userfaultfd_wait_queue *ewq)
533 if (WARN_ON_ONCE(current->flags & PF_EXITING))
537 init_waitqueue_entry(&ewq->wq, current);
539 spin_lock(&ctx->event_wqh.lock);
541 * After the __add_wait_queue the uwq is visible to userland
542 * through poll/read().
544 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
546 set_current_state(TASK_KILLABLE);
547 if (ewq->msg.event == 0)
549 if (ACCESS_ONCE(ctx->released) ||
550 fatal_signal_pending(current)) {
551 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
552 if (ewq->msg.event == UFFD_EVENT_FORK) {
553 struct userfaultfd_ctx *new;
555 new = (struct userfaultfd_ctx *)
557 ewq->msg.arg.reserved.reserved1;
559 userfaultfd_ctx_put(new);
564 spin_unlock(&ctx->event_wqh.lock);
566 wake_up_poll(&ctx->fd_wqh, POLLIN);
569 spin_lock(&ctx->event_wqh.lock);
571 __set_current_state(TASK_RUNNING);
572 spin_unlock(&ctx->event_wqh.lock);
575 * ctx may go away after this if the userfault pseudo fd is
579 userfaultfd_ctx_put(ctx);
582 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
583 struct userfaultfd_wait_queue *ewq)
586 wake_up_locked(&ctx->event_wqh);
587 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
590 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
592 struct userfaultfd_ctx *ctx = NULL, *octx;
593 struct userfaultfd_fork_ctx *fctx;
595 octx = vma->vm_userfaultfd_ctx.ctx;
596 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
597 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
598 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
602 list_for_each_entry(fctx, fcs, list)
603 if (fctx->orig == octx) {
609 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
613 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
619 atomic_set(&ctx->refcount, 1);
620 ctx->flags = octx->flags;
621 ctx->state = UFFD_STATE_RUNNING;
622 ctx->features = octx->features;
623 ctx->released = false;
624 ctx->mm = vma->vm_mm;
625 atomic_inc(&ctx->mm->mm_count);
627 userfaultfd_ctx_get(octx);
630 list_add_tail(&fctx->list, fcs);
633 vma->vm_userfaultfd_ctx.ctx = ctx;
637 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
639 struct userfaultfd_ctx *ctx = fctx->orig;
640 struct userfaultfd_wait_queue ewq;
644 ewq.msg.event = UFFD_EVENT_FORK;
645 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
647 userfaultfd_event_wait_completion(ctx, &ewq);
650 void dup_userfaultfd_complete(struct list_head *fcs)
652 struct userfaultfd_fork_ctx *fctx, *n;
654 list_for_each_entry_safe(fctx, n, fcs, list) {
656 list_del(&fctx->list);
661 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
662 struct vm_userfaultfd_ctx *vm_ctx)
664 struct userfaultfd_ctx *ctx;
666 ctx = vma->vm_userfaultfd_ctx.ctx;
667 if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
669 userfaultfd_ctx_get(ctx);
673 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
674 unsigned long from, unsigned long to,
677 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
678 struct userfaultfd_wait_queue ewq;
683 if (to & ~PAGE_MASK) {
684 userfaultfd_ctx_put(ctx);
690 ewq.msg.event = UFFD_EVENT_REMAP;
691 ewq.msg.arg.remap.from = from;
692 ewq.msg.arg.remap.to = to;
693 ewq.msg.arg.remap.len = len;
695 userfaultfd_event_wait_completion(ctx, &ewq);
698 bool userfaultfd_remove(struct vm_area_struct *vma,
699 unsigned long start, unsigned long end)
701 struct mm_struct *mm = vma->vm_mm;
702 struct userfaultfd_ctx *ctx;
703 struct userfaultfd_wait_queue ewq;
705 ctx = vma->vm_userfaultfd_ctx.ctx;
706 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
709 userfaultfd_ctx_get(ctx);
710 up_read(&mm->mmap_sem);
714 ewq.msg.event = UFFD_EVENT_REMOVE;
715 ewq.msg.arg.remove.start = start;
716 ewq.msg.arg.remove.end = end;
718 userfaultfd_event_wait_completion(ctx, &ewq);
723 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
724 unsigned long start, unsigned long end)
726 struct userfaultfd_unmap_ctx *unmap_ctx;
728 list_for_each_entry(unmap_ctx, unmaps, list)
729 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
730 unmap_ctx->end == end)
736 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
737 unsigned long start, unsigned long end,
738 struct list_head *unmaps)
740 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
741 struct userfaultfd_unmap_ctx *unmap_ctx;
742 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
744 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
745 has_unmap_ctx(ctx, unmaps, start, end))
748 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
752 userfaultfd_ctx_get(ctx);
753 unmap_ctx->ctx = ctx;
754 unmap_ctx->start = start;
755 unmap_ctx->end = end;
756 list_add_tail(&unmap_ctx->list, unmaps);
762 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
764 struct userfaultfd_unmap_ctx *ctx, *n;
765 struct userfaultfd_wait_queue ewq;
767 list_for_each_entry_safe(ctx, n, uf, list) {
770 ewq.msg.event = UFFD_EVENT_UNMAP;
771 ewq.msg.arg.remove.start = ctx->start;
772 ewq.msg.arg.remove.end = ctx->end;
774 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
776 list_del(&ctx->list);
781 static int userfaultfd_release(struct inode *inode, struct file *file)
783 struct userfaultfd_ctx *ctx = file->private_data;
784 struct mm_struct *mm = ctx->mm;
785 struct vm_area_struct *vma, *prev;
786 /* len == 0 means wake all */
787 struct userfaultfd_wake_range range = { .len = 0, };
788 unsigned long new_flags;
790 ACCESS_ONCE(ctx->released) = true;
792 if (!mmget_not_zero(mm))
796 * Flush page faults out of all CPUs. NOTE: all page faults
797 * must be retried without returning VM_FAULT_SIGBUS if
798 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
799 * changes while handle_userfault released the mmap_sem. So
800 * it's critical that released is set to true (above), before
801 * taking the mmap_sem for writing.
803 down_write(&mm->mmap_sem);
805 for (vma = mm->mmap; vma; vma = vma->vm_next) {
807 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
808 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
809 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
813 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
814 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
815 new_flags, vma->anon_vma,
816 vma->vm_file, vma->vm_pgoff,
823 vma->vm_flags = new_flags;
824 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
826 up_write(&mm->mmap_sem);
830 * After no new page faults can wait on this fault_*wqh, flush
831 * the last page faults that may have been already waiting on
834 spin_lock(&ctx->fault_pending_wqh.lock);
835 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
836 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
837 spin_unlock(&ctx->fault_pending_wqh.lock);
839 wake_up_poll(&ctx->fd_wqh, POLLHUP);
840 userfaultfd_ctx_put(ctx);
844 /* fault_pending_wqh.lock must be hold by the caller */
845 static inline struct userfaultfd_wait_queue *find_userfault_in(
846 wait_queue_head_t *wqh)
849 struct userfaultfd_wait_queue *uwq;
851 VM_BUG_ON(!spin_is_locked(&wqh->lock));
854 if (!waitqueue_active(wqh))
856 /* walk in reverse to provide FIFO behavior to read userfaults */
857 wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list);
858 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
863 static inline struct userfaultfd_wait_queue *find_userfault(
864 struct userfaultfd_ctx *ctx)
866 return find_userfault_in(&ctx->fault_pending_wqh);
869 static inline struct userfaultfd_wait_queue *find_userfault_evt(
870 struct userfaultfd_ctx *ctx)
872 return find_userfault_in(&ctx->event_wqh);
875 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
877 struct userfaultfd_ctx *ctx = file->private_data;
880 poll_wait(file, &ctx->fd_wqh, wait);
882 switch (ctx->state) {
883 case UFFD_STATE_WAIT_API:
885 case UFFD_STATE_RUNNING:
887 * poll() never guarantees that read won't block.
888 * userfaults can be waken before they're read().
890 if (unlikely(!(file->f_flags & O_NONBLOCK)))
893 * lockless access to see if there are pending faults
894 * __pollwait last action is the add_wait_queue but
895 * the spin_unlock would allow the waitqueue_active to
896 * pass above the actual list_add inside
897 * add_wait_queue critical section. So use a full
898 * memory barrier to serialize the list_add write of
899 * add_wait_queue() with the waitqueue_active read
904 if (waitqueue_active(&ctx->fault_pending_wqh))
906 else if (waitqueue_active(&ctx->event_wqh))
916 static const struct file_operations userfaultfd_fops;
918 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
919 struct userfaultfd_ctx *new,
920 struct uffd_msg *msg)
924 unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
926 fd = get_unused_fd_flags(flags);
930 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
934 return PTR_ERR(file);
937 fd_install(fd, file);
938 msg->arg.reserved.reserved1 = 0;
939 msg->arg.fork.ufd = fd;
944 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
945 struct uffd_msg *msg)
948 DECLARE_WAITQUEUE(wait, current);
949 struct userfaultfd_wait_queue *uwq;
951 * Handling fork event requires sleeping operations, so
952 * we drop the event_wqh lock, then do these ops, then
953 * lock it back and wake up the waiter. While the lock is
954 * dropped the ewq may go away so we keep track of it
957 LIST_HEAD(fork_event);
958 struct userfaultfd_ctx *fork_nctx = NULL;
960 /* always take the fd_wqh lock before the fault_pending_wqh lock */
961 spin_lock(&ctx->fd_wqh.lock);
962 __add_wait_queue(&ctx->fd_wqh, &wait);
964 set_current_state(TASK_INTERRUPTIBLE);
965 spin_lock(&ctx->fault_pending_wqh.lock);
966 uwq = find_userfault(ctx);
969 * Use a seqcount to repeat the lockless check
970 * in wake_userfault() to avoid missing
971 * wakeups because during the refile both
972 * waitqueue could become empty if this is the
975 write_seqcount_begin(&ctx->refile_seq);
978 * The fault_pending_wqh.lock prevents the uwq
979 * to disappear from under us.
981 * Refile this userfault from
982 * fault_pending_wqh to fault_wqh, it's not
983 * pending anymore after we read it.
985 * Use list_del() by hand (as
986 * userfaultfd_wake_function also uses
987 * list_del_init() by hand) to be sure nobody
988 * changes __remove_wait_queue() to use
989 * list_del_init() in turn breaking the
990 * !list_empty_careful() check in
991 * handle_userfault(). The uwq->wq.task_list
992 * must never be empty at any time during the
993 * refile, or the waitqueue could disappear
994 * from under us. The "wait_queue_head_t"
995 * parameter of __remove_wait_queue() is unused
998 list_del(&uwq->wq.task_list);
999 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1001 write_seqcount_end(&ctx->refile_seq);
1003 /* careful to always initialize msg if ret == 0 */
1005 spin_unlock(&ctx->fault_pending_wqh.lock);
1009 spin_unlock(&ctx->fault_pending_wqh.lock);
1011 spin_lock(&ctx->event_wqh.lock);
1012 uwq = find_userfault_evt(ctx);
1016 if (uwq->msg.event == UFFD_EVENT_FORK) {
1017 fork_nctx = (struct userfaultfd_ctx *)
1019 uwq->msg.arg.reserved.reserved1;
1020 list_move(&uwq->wq.task_list, &fork_event);
1021 spin_unlock(&ctx->event_wqh.lock);
1026 userfaultfd_event_complete(ctx, uwq);
1027 spin_unlock(&ctx->event_wqh.lock);
1031 spin_unlock(&ctx->event_wqh.lock);
1033 if (signal_pending(current)) {
1041 spin_unlock(&ctx->fd_wqh.lock);
1043 spin_lock(&ctx->fd_wqh.lock);
1045 __remove_wait_queue(&ctx->fd_wqh, &wait);
1046 __set_current_state(TASK_RUNNING);
1047 spin_unlock(&ctx->fd_wqh.lock);
1049 if (!ret && msg->event == UFFD_EVENT_FORK) {
1050 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1053 spin_lock(&ctx->event_wqh.lock);
1054 if (!list_empty(&fork_event)) {
1055 uwq = list_first_entry(&fork_event,
1058 list_del(&uwq->wq.task_list);
1059 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1060 userfaultfd_event_complete(ctx, uwq);
1062 spin_unlock(&ctx->event_wqh.lock);
1069 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1070 size_t count, loff_t *ppos)
1072 struct userfaultfd_ctx *ctx = file->private_data;
1073 ssize_t _ret, ret = 0;
1074 struct uffd_msg msg;
1075 int no_wait = file->f_flags & O_NONBLOCK;
1077 if (ctx->state == UFFD_STATE_WAIT_API)
1081 if (count < sizeof(msg))
1082 return ret ? ret : -EINVAL;
1083 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1085 return ret ? ret : _ret;
1086 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1087 return ret ? ret : -EFAULT;
1090 count -= sizeof(msg);
1092 * Allow to read more than one fault at time but only
1093 * block if waiting for the very first one.
1095 no_wait = O_NONBLOCK;
1099 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1100 struct userfaultfd_wake_range *range)
1102 unsigned long start, end;
1104 start = range->start;
1105 end = range->start + range->len;
1107 spin_lock(&ctx->fault_pending_wqh.lock);
1108 /* wake all in the range and autoremove */
1109 if (waitqueue_active(&ctx->fault_pending_wqh))
1110 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1112 if (waitqueue_active(&ctx->fault_wqh))
1113 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1114 spin_unlock(&ctx->fault_pending_wqh.lock);
1117 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1118 struct userfaultfd_wake_range *range)
1124 * To be sure waitqueue_active() is not reordered by the CPU
1125 * before the pagetable update, use an explicit SMP memory
1126 * barrier here. PT lock release or up_read(mmap_sem) still
1127 * have release semantics that can allow the
1128 * waitqueue_active() to be reordered before the pte update.
1133 * Use waitqueue_active because it's very frequent to
1134 * change the address space atomically even if there are no
1135 * userfaults yet. So we take the spinlock only when we're
1136 * sure we've userfaults to wake.
1139 seq = read_seqcount_begin(&ctx->refile_seq);
1140 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1141 waitqueue_active(&ctx->fault_wqh);
1143 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1145 __wake_userfault(ctx, range);
1148 static __always_inline int validate_range(struct mm_struct *mm,
1149 __u64 start, __u64 len)
1151 __u64 task_size = mm->task_size;
1153 if (start & ~PAGE_MASK)
1155 if (len & ~PAGE_MASK)
1159 if (start < mmap_min_addr)
1161 if (start >= task_size)
1163 if (len > task_size - start)
1168 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1170 return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1174 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1177 struct mm_struct *mm = ctx->mm;
1178 struct vm_area_struct *vma, *prev, *cur;
1180 struct uffdio_register uffdio_register;
1181 struct uffdio_register __user *user_uffdio_register;
1182 unsigned long vm_flags, new_flags;
1184 bool non_anon_pages;
1185 unsigned long start, end, vma_end;
1187 user_uffdio_register = (struct uffdio_register __user *) arg;
1190 if (copy_from_user(&uffdio_register, user_uffdio_register,
1191 sizeof(uffdio_register)-sizeof(__u64)))
1195 if (!uffdio_register.mode)
1197 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1198 UFFDIO_REGISTER_MODE_WP))
1201 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1202 vm_flags |= VM_UFFD_MISSING;
1203 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1204 vm_flags |= VM_UFFD_WP;
1206 * FIXME: remove the below error constraint by
1207 * implementing the wprotect tracking mode.
1213 ret = validate_range(mm, uffdio_register.range.start,
1214 uffdio_register.range.len);
1218 start = uffdio_register.range.start;
1219 end = start + uffdio_register.range.len;
1222 if (!mmget_not_zero(mm))
1225 down_write(&mm->mmap_sem);
1226 vma = find_vma_prev(mm, start, &prev);
1230 /* check that there's at least one vma in the range */
1232 if (vma->vm_start >= end)
1236 * If the first vma contains huge pages, make sure start address
1237 * is aligned to huge page size.
1239 if (is_vm_hugetlb_page(vma)) {
1240 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1242 if (start & (vma_hpagesize - 1))
1247 * Search for not compatible vmas.
1250 non_anon_pages = false;
1251 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1254 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1255 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1257 /* check not compatible vmas */
1259 if (!vma_can_userfault(cur))
1262 * If this vma contains ending address, and huge pages
1265 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1266 end > cur->vm_start) {
1267 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1271 if (end & (vma_hpagesize - 1))
1276 * Check that this vma isn't already owned by a
1277 * different userfaultfd. We can't allow more than one
1278 * userfaultfd to own a single vma simultaneously or we
1279 * wouldn't know which one to deliver the userfaults to.
1282 if (cur->vm_userfaultfd_ctx.ctx &&
1283 cur->vm_userfaultfd_ctx.ctx != ctx)
1287 * Note vmas containing huge pages
1289 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1290 non_anon_pages = true;
1296 if (vma->vm_start < start)
1303 BUG_ON(!vma_can_userfault(vma));
1304 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1305 vma->vm_userfaultfd_ctx.ctx != ctx);
1308 * Nothing to do: this vma is already registered into this
1309 * userfaultfd and with the right tracking mode too.
1311 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1312 (vma->vm_flags & vm_flags) == vm_flags)
1315 if (vma->vm_start > start)
1316 start = vma->vm_start;
1317 vma_end = min(end, vma->vm_end);
1319 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1320 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1321 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1323 ((struct vm_userfaultfd_ctx){ ctx }));
1328 if (vma->vm_start < start) {
1329 ret = split_vma(mm, vma, start, 1);
1333 if (vma->vm_end > end) {
1334 ret = split_vma(mm, vma, end, 0);
1340 * In the vma_merge() successful mprotect-like case 8:
1341 * the next vma was merged into the current one and
1342 * the current one has not been updated yet.
1344 vma->vm_flags = new_flags;
1345 vma->vm_userfaultfd_ctx.ctx = ctx;
1349 start = vma->vm_end;
1351 } while (vma && vma->vm_start < end);
1353 up_write(&mm->mmap_sem);
1357 * Now that we scanned all vmas we can already tell
1358 * userland which ioctls methods are guaranteed to
1359 * succeed on this range.
1361 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1362 UFFD_API_RANGE_IOCTLS,
1363 &user_uffdio_register->ioctls))
1370 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1373 struct mm_struct *mm = ctx->mm;
1374 struct vm_area_struct *vma, *prev, *cur;
1376 struct uffdio_range uffdio_unregister;
1377 unsigned long new_flags;
1379 unsigned long start, end, vma_end;
1380 const void __user *buf = (void __user *)arg;
1383 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1386 ret = validate_range(mm, uffdio_unregister.start,
1387 uffdio_unregister.len);
1391 start = uffdio_unregister.start;
1392 end = start + uffdio_unregister.len;
1395 if (!mmget_not_zero(mm))
1398 down_write(&mm->mmap_sem);
1399 vma = find_vma_prev(mm, start, &prev);
1403 /* check that there's at least one vma in the range */
1405 if (vma->vm_start >= end)
1409 * If the first vma contains huge pages, make sure start address
1410 * is aligned to huge page size.
1412 if (is_vm_hugetlb_page(vma)) {
1413 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1415 if (start & (vma_hpagesize - 1))
1420 * Search for not compatible vmas.
1424 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1427 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1428 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1431 * Check not compatible vmas, not strictly required
1432 * here as not compatible vmas cannot have an
1433 * userfaultfd_ctx registered on them, but this
1434 * provides for more strict behavior to notice
1435 * unregistration errors.
1437 if (!vma_can_userfault(cur))
1444 if (vma->vm_start < start)
1451 BUG_ON(!vma_can_userfault(vma));
1454 * Nothing to do: this vma is already registered into this
1455 * userfaultfd and with the right tracking mode too.
1457 if (!vma->vm_userfaultfd_ctx.ctx)
1460 if (vma->vm_start > start)
1461 start = vma->vm_start;
1462 vma_end = min(end, vma->vm_end);
1464 if (userfaultfd_missing(vma)) {
1466 * Wake any concurrent pending userfault while
1467 * we unregister, so they will not hang
1468 * permanently and it avoids userland to call
1469 * UFFDIO_WAKE explicitly.
1471 struct userfaultfd_wake_range range;
1472 range.start = start;
1473 range.len = vma_end - start;
1474 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1477 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1478 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1479 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1486 if (vma->vm_start < start) {
1487 ret = split_vma(mm, vma, start, 1);
1491 if (vma->vm_end > end) {
1492 ret = split_vma(mm, vma, end, 0);
1498 * In the vma_merge() successful mprotect-like case 8:
1499 * the next vma was merged into the current one and
1500 * the current one has not been updated yet.
1502 vma->vm_flags = new_flags;
1503 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1507 start = vma->vm_end;
1509 } while (vma && vma->vm_start < end);
1511 up_write(&mm->mmap_sem);
1518 * userfaultfd_wake may be used in combination with the
1519 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1521 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1525 struct uffdio_range uffdio_wake;
1526 struct userfaultfd_wake_range range;
1527 const void __user *buf = (void __user *)arg;
1530 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1533 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1537 range.start = uffdio_wake.start;
1538 range.len = uffdio_wake.len;
1541 * len == 0 means wake all and we don't want to wake all here,
1542 * so check it again to be sure.
1544 VM_BUG_ON(!range.len);
1546 wake_userfault(ctx, &range);
1553 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1557 struct uffdio_copy uffdio_copy;
1558 struct uffdio_copy __user *user_uffdio_copy;
1559 struct userfaultfd_wake_range range;
1561 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1564 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1565 /* don't copy "copy" last field */
1566 sizeof(uffdio_copy)-sizeof(__s64)))
1569 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1573 * double check for wraparound just in case. copy_from_user()
1574 * will later check uffdio_copy.src + uffdio_copy.len to fit
1575 * in the userland range.
1578 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1580 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1582 if (mmget_not_zero(ctx->mm)) {
1583 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1589 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1594 /* len == 0 would wake all */
1596 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1597 range.start = uffdio_copy.dst;
1598 wake_userfault(ctx, &range);
1600 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1605 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1609 struct uffdio_zeropage uffdio_zeropage;
1610 struct uffdio_zeropage __user *user_uffdio_zeropage;
1611 struct userfaultfd_wake_range range;
1613 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1616 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1617 /* don't copy "zeropage" last field */
1618 sizeof(uffdio_zeropage)-sizeof(__s64)))
1621 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1622 uffdio_zeropage.range.len);
1626 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1629 if (mmget_not_zero(ctx->mm)) {
1630 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1631 uffdio_zeropage.range.len);
1634 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1638 /* len == 0 would wake all */
1641 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1642 range.start = uffdio_zeropage.range.start;
1643 wake_userfault(ctx, &range);
1645 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1650 static inline unsigned int uffd_ctx_features(__u64 user_features)
1653 * For the current set of features the bits just coincide
1655 return (unsigned int)user_features;
1659 * userland asks for a certain API version and we return which bits
1660 * and ioctl commands are implemented in this kernel for such API
1661 * version or -EINVAL if unknown.
1663 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1666 struct uffdio_api uffdio_api;
1667 void __user *buf = (void __user *)arg;
1672 if (ctx->state != UFFD_STATE_WAIT_API)
1675 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1677 features = uffdio_api.features;
1678 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1679 memset(&uffdio_api, 0, sizeof(uffdio_api));
1680 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1685 /* report all available features and ioctls to userland */
1686 uffdio_api.features = UFFD_API_FEATURES;
1687 uffdio_api.ioctls = UFFD_API_IOCTLS;
1689 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1691 ctx->state = UFFD_STATE_RUNNING;
1692 /* only enable the requested features for this uffd context */
1693 ctx->features = uffd_ctx_features(features);
1699 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1703 struct userfaultfd_ctx *ctx = file->private_data;
1705 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1710 ret = userfaultfd_api(ctx, arg);
1712 case UFFDIO_REGISTER:
1713 ret = userfaultfd_register(ctx, arg);
1715 case UFFDIO_UNREGISTER:
1716 ret = userfaultfd_unregister(ctx, arg);
1719 ret = userfaultfd_wake(ctx, arg);
1722 ret = userfaultfd_copy(ctx, arg);
1724 case UFFDIO_ZEROPAGE:
1725 ret = userfaultfd_zeropage(ctx, arg);
1731 #ifdef CONFIG_PROC_FS
1732 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1734 struct userfaultfd_ctx *ctx = f->private_data;
1736 struct userfaultfd_wait_queue *uwq;
1737 unsigned long pending = 0, total = 0;
1739 spin_lock(&ctx->fault_pending_wqh.lock);
1740 list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1741 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1745 list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1746 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1749 spin_unlock(&ctx->fault_pending_wqh.lock);
1752 * If more protocols will be added, there will be all shown
1753 * separated by a space. Like this:
1754 * protocols: aa:... bb:...
1756 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1757 pending, total, UFFD_API, UFFD_API_FEATURES,
1758 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1762 static const struct file_operations userfaultfd_fops = {
1763 #ifdef CONFIG_PROC_FS
1764 .show_fdinfo = userfaultfd_show_fdinfo,
1766 .release = userfaultfd_release,
1767 .poll = userfaultfd_poll,
1768 .read = userfaultfd_read,
1769 .unlocked_ioctl = userfaultfd_ioctl,
1770 .compat_ioctl = userfaultfd_ioctl,
1771 .llseek = noop_llseek,
1774 static void init_once_userfaultfd_ctx(void *mem)
1776 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1778 init_waitqueue_head(&ctx->fault_pending_wqh);
1779 init_waitqueue_head(&ctx->fault_wqh);
1780 init_waitqueue_head(&ctx->event_wqh);
1781 init_waitqueue_head(&ctx->fd_wqh);
1782 seqcount_init(&ctx->refile_seq);
1786 * userfaultfd_file_create - Creates a userfaultfd file pointer.
1787 * @flags: Flags for the userfaultfd file.
1789 * This function creates a userfaultfd file pointer, w/out installing
1790 * it into the fd table. This is useful when the userfaultfd file is
1791 * used during the initialization of data structures that require
1792 * extra setup after the userfaultfd creation. So the userfaultfd
1793 * creation is split into the file pointer creation phase, and the
1794 * file descriptor installation phase. In this way races with
1795 * userspace closing the newly installed file descriptor can be
1796 * avoided. Returns a userfaultfd file pointer, or a proper error
1799 static struct file *userfaultfd_file_create(int flags)
1802 struct userfaultfd_ctx *ctx;
1804 BUG_ON(!current->mm);
1806 /* Check the UFFD_* constants for consistency. */
1807 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1808 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1810 file = ERR_PTR(-EINVAL);
1811 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1814 file = ERR_PTR(-ENOMEM);
1815 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1819 atomic_set(&ctx->refcount, 1);
1822 ctx->state = UFFD_STATE_WAIT_API;
1823 ctx->released = false;
1824 ctx->mm = current->mm;
1825 /* prevent the mm struct to be freed */
1828 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1829 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1832 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1838 SYSCALL_DEFINE1(userfaultfd, int, flags)
1843 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1848 file = userfaultfd_file_create(flags);
1850 error = PTR_ERR(file);
1851 goto err_put_unused_fd;
1853 fd_install(fd, file);
1863 static int __init userfaultfd_init(void)
1865 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1866 sizeof(struct userfaultfd_ctx),
1868 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1869 init_once_userfaultfd_ctx);
1872 __initcall(userfaultfd_init);