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);
555 spin_unlock(&ctx->event_wqh.lock);
557 wake_up_poll(&ctx->fd_wqh, POLLIN);
560 spin_lock(&ctx->event_wqh.lock);
562 __set_current_state(TASK_RUNNING);
563 spin_unlock(&ctx->event_wqh.lock);
566 * ctx may go away after this if the userfault pseudo fd is
570 userfaultfd_ctx_put(ctx);
573 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
574 struct userfaultfd_wait_queue *ewq)
577 wake_up_locked(&ctx->event_wqh);
578 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
581 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
583 struct userfaultfd_ctx *ctx = NULL, *octx;
584 struct userfaultfd_fork_ctx *fctx;
586 octx = vma->vm_userfaultfd_ctx.ctx;
587 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
588 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
589 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
593 list_for_each_entry(fctx, fcs, list)
594 if (fctx->orig == octx) {
600 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
604 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
610 atomic_set(&ctx->refcount, 1);
611 ctx->flags = octx->flags;
612 ctx->state = UFFD_STATE_RUNNING;
613 ctx->features = octx->features;
614 ctx->released = false;
615 ctx->mm = vma->vm_mm;
616 atomic_inc(&ctx->mm->mm_count);
618 userfaultfd_ctx_get(octx);
621 list_add_tail(&fctx->list, fcs);
624 vma->vm_userfaultfd_ctx.ctx = ctx;
628 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
630 struct userfaultfd_ctx *ctx = fctx->orig;
631 struct userfaultfd_wait_queue ewq;
635 ewq.msg.event = UFFD_EVENT_FORK;
636 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
638 userfaultfd_event_wait_completion(ctx, &ewq);
641 void dup_userfaultfd_complete(struct list_head *fcs)
643 struct userfaultfd_fork_ctx *fctx, *n;
645 list_for_each_entry_safe(fctx, n, fcs, list) {
647 list_del(&fctx->list);
652 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
653 struct vm_userfaultfd_ctx *vm_ctx)
655 struct userfaultfd_ctx *ctx;
657 ctx = vma->vm_userfaultfd_ctx.ctx;
658 if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
660 userfaultfd_ctx_get(ctx);
664 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
665 unsigned long from, unsigned long to,
668 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
669 struct userfaultfd_wait_queue ewq;
674 if (to & ~PAGE_MASK) {
675 userfaultfd_ctx_put(ctx);
681 ewq.msg.event = UFFD_EVENT_REMAP;
682 ewq.msg.arg.remap.from = from;
683 ewq.msg.arg.remap.to = to;
684 ewq.msg.arg.remap.len = len;
686 userfaultfd_event_wait_completion(ctx, &ewq);
689 void userfaultfd_remove(struct vm_area_struct *vma,
690 struct vm_area_struct **prev,
691 unsigned long start, unsigned long end)
693 struct mm_struct *mm = vma->vm_mm;
694 struct userfaultfd_ctx *ctx;
695 struct userfaultfd_wait_queue ewq;
697 ctx = vma->vm_userfaultfd_ctx.ctx;
698 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
701 userfaultfd_ctx_get(ctx);
702 up_read(&mm->mmap_sem);
704 *prev = NULL; /* We wait for ACK w/o the mmap semaphore */
708 ewq.msg.event = UFFD_EVENT_REMOVE;
709 ewq.msg.arg.remove.start = start;
710 ewq.msg.arg.remove.end = end;
712 userfaultfd_event_wait_completion(ctx, &ewq);
714 down_read(&mm->mmap_sem);
717 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
718 unsigned long start, unsigned long end)
720 struct userfaultfd_unmap_ctx *unmap_ctx;
722 list_for_each_entry(unmap_ctx, unmaps, list)
723 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
724 unmap_ctx->end == end)
730 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
731 unsigned long start, unsigned long end,
732 struct list_head *unmaps)
734 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
735 struct userfaultfd_unmap_ctx *unmap_ctx;
736 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
738 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
739 has_unmap_ctx(ctx, unmaps, start, end))
742 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
746 userfaultfd_ctx_get(ctx);
747 unmap_ctx->ctx = ctx;
748 unmap_ctx->start = start;
749 unmap_ctx->end = end;
750 list_add_tail(&unmap_ctx->list, unmaps);
756 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
758 struct userfaultfd_unmap_ctx *ctx, *n;
759 struct userfaultfd_wait_queue ewq;
761 list_for_each_entry_safe(ctx, n, uf, list) {
764 ewq.msg.event = UFFD_EVENT_UNMAP;
765 ewq.msg.arg.remove.start = ctx->start;
766 ewq.msg.arg.remove.end = ctx->end;
768 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
770 list_del(&ctx->list);
775 static int userfaultfd_release(struct inode *inode, struct file *file)
777 struct userfaultfd_ctx *ctx = file->private_data;
778 struct mm_struct *mm = ctx->mm;
779 struct vm_area_struct *vma, *prev;
780 /* len == 0 means wake all */
781 struct userfaultfd_wake_range range = { .len = 0, };
782 unsigned long new_flags;
784 ACCESS_ONCE(ctx->released) = true;
786 if (!mmget_not_zero(mm))
790 * Flush page faults out of all CPUs. NOTE: all page faults
791 * must be retried without returning VM_FAULT_SIGBUS if
792 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
793 * changes while handle_userfault released the mmap_sem. So
794 * it's critical that released is set to true (above), before
795 * taking the mmap_sem for writing.
797 down_write(&mm->mmap_sem);
799 for (vma = mm->mmap; vma; vma = vma->vm_next) {
801 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
802 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
803 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
807 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
808 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
809 new_flags, vma->anon_vma,
810 vma->vm_file, vma->vm_pgoff,
817 vma->vm_flags = new_flags;
818 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
820 up_write(&mm->mmap_sem);
824 * After no new page faults can wait on this fault_*wqh, flush
825 * the last page faults that may have been already waiting on
828 spin_lock(&ctx->fault_pending_wqh.lock);
829 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
830 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
831 spin_unlock(&ctx->fault_pending_wqh.lock);
833 wake_up_poll(&ctx->fd_wqh, POLLHUP);
834 userfaultfd_ctx_put(ctx);
838 /* fault_pending_wqh.lock must be hold by the caller */
839 static inline struct userfaultfd_wait_queue *find_userfault_in(
840 wait_queue_head_t *wqh)
843 struct userfaultfd_wait_queue *uwq;
845 VM_BUG_ON(!spin_is_locked(&wqh->lock));
848 if (!waitqueue_active(wqh))
850 /* walk in reverse to provide FIFO behavior to read userfaults */
851 wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list);
852 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
857 static inline struct userfaultfd_wait_queue *find_userfault(
858 struct userfaultfd_ctx *ctx)
860 return find_userfault_in(&ctx->fault_pending_wqh);
863 static inline struct userfaultfd_wait_queue *find_userfault_evt(
864 struct userfaultfd_ctx *ctx)
866 return find_userfault_in(&ctx->event_wqh);
869 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
871 struct userfaultfd_ctx *ctx = file->private_data;
874 poll_wait(file, &ctx->fd_wqh, wait);
876 switch (ctx->state) {
877 case UFFD_STATE_WAIT_API:
879 case UFFD_STATE_RUNNING:
881 * poll() never guarantees that read won't block.
882 * userfaults can be waken before they're read().
884 if (unlikely(!(file->f_flags & O_NONBLOCK)))
887 * lockless access to see if there are pending faults
888 * __pollwait last action is the add_wait_queue but
889 * the spin_unlock would allow the waitqueue_active to
890 * pass above the actual list_add inside
891 * add_wait_queue critical section. So use a full
892 * memory barrier to serialize the list_add write of
893 * add_wait_queue() with the waitqueue_active read
898 if (waitqueue_active(&ctx->fault_pending_wqh))
900 else if (waitqueue_active(&ctx->event_wqh))
910 static const struct file_operations userfaultfd_fops;
912 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
913 struct userfaultfd_ctx *new,
914 struct uffd_msg *msg)
918 unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
920 fd = get_unused_fd_flags(flags);
924 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
928 return PTR_ERR(file);
931 fd_install(fd, file);
932 msg->arg.reserved.reserved1 = 0;
933 msg->arg.fork.ufd = fd;
938 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
939 struct uffd_msg *msg)
942 DECLARE_WAITQUEUE(wait, current);
943 struct userfaultfd_wait_queue *uwq;
945 * Handling fork event requires sleeping operations, so
946 * we drop the event_wqh lock, then do these ops, then
947 * lock it back and wake up the waiter. While the lock is
948 * dropped the ewq may go away so we keep track of it
951 LIST_HEAD(fork_event);
952 struct userfaultfd_ctx *fork_nctx = NULL;
954 /* always take the fd_wqh lock before the fault_pending_wqh lock */
955 spin_lock(&ctx->fd_wqh.lock);
956 __add_wait_queue(&ctx->fd_wqh, &wait);
958 set_current_state(TASK_INTERRUPTIBLE);
959 spin_lock(&ctx->fault_pending_wqh.lock);
960 uwq = find_userfault(ctx);
963 * Use a seqcount to repeat the lockless check
964 * in wake_userfault() to avoid missing
965 * wakeups because during the refile both
966 * waitqueue could become empty if this is the
969 write_seqcount_begin(&ctx->refile_seq);
972 * The fault_pending_wqh.lock prevents the uwq
973 * to disappear from under us.
975 * Refile this userfault from
976 * fault_pending_wqh to fault_wqh, it's not
977 * pending anymore after we read it.
979 * Use list_del() by hand (as
980 * userfaultfd_wake_function also uses
981 * list_del_init() by hand) to be sure nobody
982 * changes __remove_wait_queue() to use
983 * list_del_init() in turn breaking the
984 * !list_empty_careful() check in
985 * handle_userfault(). The uwq->wq.task_list
986 * must never be empty at any time during the
987 * refile, or the waitqueue could disappear
988 * from under us. The "wait_queue_head_t"
989 * parameter of __remove_wait_queue() is unused
992 list_del(&uwq->wq.task_list);
993 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
995 write_seqcount_end(&ctx->refile_seq);
997 /* careful to always initialize msg if ret == 0 */
999 spin_unlock(&ctx->fault_pending_wqh.lock);
1003 spin_unlock(&ctx->fault_pending_wqh.lock);
1005 spin_lock(&ctx->event_wqh.lock);
1006 uwq = find_userfault_evt(ctx);
1010 if (uwq->msg.event == UFFD_EVENT_FORK) {
1011 fork_nctx = (struct userfaultfd_ctx *)
1013 uwq->msg.arg.reserved.reserved1;
1014 list_move(&uwq->wq.task_list, &fork_event);
1015 spin_unlock(&ctx->event_wqh.lock);
1020 userfaultfd_event_complete(ctx, uwq);
1021 spin_unlock(&ctx->event_wqh.lock);
1025 spin_unlock(&ctx->event_wqh.lock);
1027 if (signal_pending(current)) {
1035 spin_unlock(&ctx->fd_wqh.lock);
1037 spin_lock(&ctx->fd_wqh.lock);
1039 __remove_wait_queue(&ctx->fd_wqh, &wait);
1040 __set_current_state(TASK_RUNNING);
1041 spin_unlock(&ctx->fd_wqh.lock);
1043 if (!ret && msg->event == UFFD_EVENT_FORK) {
1044 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1047 spin_lock(&ctx->event_wqh.lock);
1048 if (!list_empty(&fork_event)) {
1049 uwq = list_first_entry(&fork_event,
1052 list_del(&uwq->wq.task_list);
1053 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1054 userfaultfd_event_complete(ctx, uwq);
1056 spin_unlock(&ctx->event_wqh.lock);
1063 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1064 size_t count, loff_t *ppos)
1066 struct userfaultfd_ctx *ctx = file->private_data;
1067 ssize_t _ret, ret = 0;
1068 struct uffd_msg msg;
1069 int no_wait = file->f_flags & O_NONBLOCK;
1071 if (ctx->state == UFFD_STATE_WAIT_API)
1075 if (count < sizeof(msg))
1076 return ret ? ret : -EINVAL;
1077 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1079 return ret ? ret : _ret;
1080 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1081 return ret ? ret : -EFAULT;
1084 count -= sizeof(msg);
1086 * Allow to read more than one fault at time but only
1087 * block if waiting for the very first one.
1089 no_wait = O_NONBLOCK;
1093 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1094 struct userfaultfd_wake_range *range)
1096 unsigned long start, end;
1098 start = range->start;
1099 end = range->start + range->len;
1101 spin_lock(&ctx->fault_pending_wqh.lock);
1102 /* wake all in the range and autoremove */
1103 if (waitqueue_active(&ctx->fault_pending_wqh))
1104 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1106 if (waitqueue_active(&ctx->fault_wqh))
1107 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1108 spin_unlock(&ctx->fault_pending_wqh.lock);
1111 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1112 struct userfaultfd_wake_range *range)
1118 * To be sure waitqueue_active() is not reordered by the CPU
1119 * before the pagetable update, use an explicit SMP memory
1120 * barrier here. PT lock release or up_read(mmap_sem) still
1121 * have release semantics that can allow the
1122 * waitqueue_active() to be reordered before the pte update.
1127 * Use waitqueue_active because it's very frequent to
1128 * change the address space atomically even if there are no
1129 * userfaults yet. So we take the spinlock only when we're
1130 * sure we've userfaults to wake.
1133 seq = read_seqcount_begin(&ctx->refile_seq);
1134 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1135 waitqueue_active(&ctx->fault_wqh);
1137 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1139 __wake_userfault(ctx, range);
1142 static __always_inline int validate_range(struct mm_struct *mm,
1143 __u64 start, __u64 len)
1145 __u64 task_size = mm->task_size;
1147 if (start & ~PAGE_MASK)
1149 if (len & ~PAGE_MASK)
1153 if (start < mmap_min_addr)
1155 if (start >= task_size)
1157 if (len > task_size - start)
1162 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1164 return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1168 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1171 struct mm_struct *mm = ctx->mm;
1172 struct vm_area_struct *vma, *prev, *cur;
1174 struct uffdio_register uffdio_register;
1175 struct uffdio_register __user *user_uffdio_register;
1176 unsigned long vm_flags, new_flags;
1178 bool non_anon_pages;
1179 unsigned long start, end, vma_end;
1181 user_uffdio_register = (struct uffdio_register __user *) arg;
1184 if (copy_from_user(&uffdio_register, user_uffdio_register,
1185 sizeof(uffdio_register)-sizeof(__u64)))
1189 if (!uffdio_register.mode)
1191 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1192 UFFDIO_REGISTER_MODE_WP))
1195 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1196 vm_flags |= VM_UFFD_MISSING;
1197 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1198 vm_flags |= VM_UFFD_WP;
1200 * FIXME: remove the below error constraint by
1201 * implementing the wprotect tracking mode.
1207 ret = validate_range(mm, uffdio_register.range.start,
1208 uffdio_register.range.len);
1212 start = uffdio_register.range.start;
1213 end = start + uffdio_register.range.len;
1216 if (!mmget_not_zero(mm))
1219 down_write(&mm->mmap_sem);
1220 vma = find_vma_prev(mm, start, &prev);
1224 /* check that there's at least one vma in the range */
1226 if (vma->vm_start >= end)
1230 * If the first vma contains huge pages, make sure start address
1231 * is aligned to huge page size.
1233 if (is_vm_hugetlb_page(vma)) {
1234 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1236 if (start & (vma_hpagesize - 1))
1241 * Search for not compatible vmas.
1244 non_anon_pages = false;
1245 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1248 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1249 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1251 /* check not compatible vmas */
1253 if (!vma_can_userfault(cur))
1256 * If this vma contains ending address, and huge pages
1259 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1260 end > cur->vm_start) {
1261 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1265 if (end & (vma_hpagesize - 1))
1270 * Check that this vma isn't already owned by a
1271 * different userfaultfd. We can't allow more than one
1272 * userfaultfd to own a single vma simultaneously or we
1273 * wouldn't know which one to deliver the userfaults to.
1276 if (cur->vm_userfaultfd_ctx.ctx &&
1277 cur->vm_userfaultfd_ctx.ctx != ctx)
1281 * Note vmas containing huge pages
1283 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1284 non_anon_pages = true;
1290 if (vma->vm_start < start)
1297 BUG_ON(!vma_can_userfault(vma));
1298 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1299 vma->vm_userfaultfd_ctx.ctx != ctx);
1302 * Nothing to do: this vma is already registered into this
1303 * userfaultfd and with the right tracking mode too.
1305 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1306 (vma->vm_flags & vm_flags) == vm_flags)
1309 if (vma->vm_start > start)
1310 start = vma->vm_start;
1311 vma_end = min(end, vma->vm_end);
1313 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1314 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1315 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1317 ((struct vm_userfaultfd_ctx){ ctx }));
1322 if (vma->vm_start < start) {
1323 ret = split_vma(mm, vma, start, 1);
1327 if (vma->vm_end > end) {
1328 ret = split_vma(mm, vma, end, 0);
1334 * In the vma_merge() successful mprotect-like case 8:
1335 * the next vma was merged into the current one and
1336 * the current one has not been updated yet.
1338 vma->vm_flags = new_flags;
1339 vma->vm_userfaultfd_ctx.ctx = ctx;
1343 start = vma->vm_end;
1345 } while (vma && vma->vm_start < end);
1347 up_write(&mm->mmap_sem);
1351 * Now that we scanned all vmas we can already tell
1352 * userland which ioctls methods are guaranteed to
1353 * succeed on this range.
1355 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1356 UFFD_API_RANGE_IOCTLS,
1357 &user_uffdio_register->ioctls))
1364 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1367 struct mm_struct *mm = ctx->mm;
1368 struct vm_area_struct *vma, *prev, *cur;
1370 struct uffdio_range uffdio_unregister;
1371 unsigned long new_flags;
1373 unsigned long start, end, vma_end;
1374 const void __user *buf = (void __user *)arg;
1377 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1380 ret = validate_range(mm, uffdio_unregister.start,
1381 uffdio_unregister.len);
1385 start = uffdio_unregister.start;
1386 end = start + uffdio_unregister.len;
1389 if (!mmget_not_zero(mm))
1392 down_write(&mm->mmap_sem);
1393 vma = find_vma_prev(mm, start, &prev);
1397 /* check that there's at least one vma in the range */
1399 if (vma->vm_start >= end)
1403 * If the first vma contains huge pages, make sure start address
1404 * is aligned to huge page size.
1406 if (is_vm_hugetlb_page(vma)) {
1407 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1409 if (start & (vma_hpagesize - 1))
1414 * Search for not compatible vmas.
1418 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1421 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1422 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1425 * Check not compatible vmas, not strictly required
1426 * here as not compatible vmas cannot have an
1427 * userfaultfd_ctx registered on them, but this
1428 * provides for more strict behavior to notice
1429 * unregistration errors.
1431 if (!vma_can_userfault(cur))
1438 if (vma->vm_start < start)
1445 BUG_ON(!vma_can_userfault(vma));
1448 * Nothing to do: this vma is already registered into this
1449 * userfaultfd and with the right tracking mode too.
1451 if (!vma->vm_userfaultfd_ctx.ctx)
1454 if (vma->vm_start > start)
1455 start = vma->vm_start;
1456 vma_end = min(end, vma->vm_end);
1458 if (userfaultfd_missing(vma)) {
1460 * Wake any concurrent pending userfault while
1461 * we unregister, so they will not hang
1462 * permanently and it avoids userland to call
1463 * UFFDIO_WAKE explicitly.
1465 struct userfaultfd_wake_range range;
1466 range.start = start;
1467 range.len = vma_end - start;
1468 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1471 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1472 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1473 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1480 if (vma->vm_start < start) {
1481 ret = split_vma(mm, vma, start, 1);
1485 if (vma->vm_end > end) {
1486 ret = split_vma(mm, vma, end, 0);
1492 * In the vma_merge() successful mprotect-like case 8:
1493 * the next vma was merged into the current one and
1494 * the current one has not been updated yet.
1496 vma->vm_flags = new_flags;
1497 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1501 start = vma->vm_end;
1503 } while (vma && vma->vm_start < end);
1505 up_write(&mm->mmap_sem);
1512 * userfaultfd_wake may be used in combination with the
1513 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1515 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1519 struct uffdio_range uffdio_wake;
1520 struct userfaultfd_wake_range range;
1521 const void __user *buf = (void __user *)arg;
1524 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1527 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1531 range.start = uffdio_wake.start;
1532 range.len = uffdio_wake.len;
1535 * len == 0 means wake all and we don't want to wake all here,
1536 * so check it again to be sure.
1538 VM_BUG_ON(!range.len);
1540 wake_userfault(ctx, &range);
1547 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1551 struct uffdio_copy uffdio_copy;
1552 struct uffdio_copy __user *user_uffdio_copy;
1553 struct userfaultfd_wake_range range;
1555 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1558 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1559 /* don't copy "copy" last field */
1560 sizeof(uffdio_copy)-sizeof(__s64)))
1563 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1567 * double check for wraparound just in case. copy_from_user()
1568 * will later check uffdio_copy.src + uffdio_copy.len to fit
1569 * in the userland range.
1572 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1574 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1576 if (mmget_not_zero(ctx->mm)) {
1577 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1583 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1588 /* len == 0 would wake all */
1590 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1591 range.start = uffdio_copy.dst;
1592 wake_userfault(ctx, &range);
1594 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1599 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1603 struct uffdio_zeropage uffdio_zeropage;
1604 struct uffdio_zeropage __user *user_uffdio_zeropage;
1605 struct userfaultfd_wake_range range;
1607 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1610 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1611 /* don't copy "zeropage" last field */
1612 sizeof(uffdio_zeropage)-sizeof(__s64)))
1615 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1616 uffdio_zeropage.range.len);
1620 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1623 if (mmget_not_zero(ctx->mm)) {
1624 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1625 uffdio_zeropage.range.len);
1628 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1632 /* len == 0 would wake all */
1635 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1636 range.start = uffdio_zeropage.range.start;
1637 wake_userfault(ctx, &range);
1639 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1644 static inline unsigned int uffd_ctx_features(__u64 user_features)
1647 * For the current set of features the bits just coincide
1649 return (unsigned int)user_features;
1653 * userland asks for a certain API version and we return which bits
1654 * and ioctl commands are implemented in this kernel for such API
1655 * version or -EINVAL if unknown.
1657 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1660 struct uffdio_api uffdio_api;
1661 void __user *buf = (void __user *)arg;
1666 if (ctx->state != UFFD_STATE_WAIT_API)
1669 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1671 features = uffdio_api.features;
1672 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1673 memset(&uffdio_api, 0, sizeof(uffdio_api));
1674 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1679 /* report all available features and ioctls to userland */
1680 uffdio_api.features = UFFD_API_FEATURES;
1681 uffdio_api.ioctls = UFFD_API_IOCTLS;
1683 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1685 ctx->state = UFFD_STATE_RUNNING;
1686 /* only enable the requested features for this uffd context */
1687 ctx->features = uffd_ctx_features(features);
1693 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1697 struct userfaultfd_ctx *ctx = file->private_data;
1699 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1704 ret = userfaultfd_api(ctx, arg);
1706 case UFFDIO_REGISTER:
1707 ret = userfaultfd_register(ctx, arg);
1709 case UFFDIO_UNREGISTER:
1710 ret = userfaultfd_unregister(ctx, arg);
1713 ret = userfaultfd_wake(ctx, arg);
1716 ret = userfaultfd_copy(ctx, arg);
1718 case UFFDIO_ZEROPAGE:
1719 ret = userfaultfd_zeropage(ctx, arg);
1725 #ifdef CONFIG_PROC_FS
1726 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1728 struct userfaultfd_ctx *ctx = f->private_data;
1730 struct userfaultfd_wait_queue *uwq;
1731 unsigned long pending = 0, total = 0;
1733 spin_lock(&ctx->fault_pending_wqh.lock);
1734 list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1735 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1739 list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1740 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1743 spin_unlock(&ctx->fault_pending_wqh.lock);
1746 * If more protocols will be added, there will be all shown
1747 * separated by a space. Like this:
1748 * protocols: aa:... bb:...
1750 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1751 pending, total, UFFD_API, UFFD_API_FEATURES,
1752 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1756 static const struct file_operations userfaultfd_fops = {
1757 #ifdef CONFIG_PROC_FS
1758 .show_fdinfo = userfaultfd_show_fdinfo,
1760 .release = userfaultfd_release,
1761 .poll = userfaultfd_poll,
1762 .read = userfaultfd_read,
1763 .unlocked_ioctl = userfaultfd_ioctl,
1764 .compat_ioctl = userfaultfd_ioctl,
1765 .llseek = noop_llseek,
1768 static void init_once_userfaultfd_ctx(void *mem)
1770 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1772 init_waitqueue_head(&ctx->fault_pending_wqh);
1773 init_waitqueue_head(&ctx->fault_wqh);
1774 init_waitqueue_head(&ctx->event_wqh);
1775 init_waitqueue_head(&ctx->fd_wqh);
1776 seqcount_init(&ctx->refile_seq);
1780 * userfaultfd_file_create - Creates a userfaultfd file pointer.
1781 * @flags: Flags for the userfaultfd file.
1783 * This function creates a userfaultfd file pointer, w/out installing
1784 * it into the fd table. This is useful when the userfaultfd file is
1785 * used during the initialization of data structures that require
1786 * extra setup after the userfaultfd creation. So the userfaultfd
1787 * creation is split into the file pointer creation phase, and the
1788 * file descriptor installation phase. In this way races with
1789 * userspace closing the newly installed file descriptor can be
1790 * avoided. Returns a userfaultfd file pointer, or a proper error
1793 static struct file *userfaultfd_file_create(int flags)
1796 struct userfaultfd_ctx *ctx;
1798 BUG_ON(!current->mm);
1800 /* Check the UFFD_* constants for consistency. */
1801 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1802 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1804 file = ERR_PTR(-EINVAL);
1805 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1808 file = ERR_PTR(-ENOMEM);
1809 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1813 atomic_set(&ctx->refcount, 1);
1816 ctx->state = UFFD_STATE_WAIT_API;
1817 ctx->released = false;
1818 ctx->mm = current->mm;
1819 /* prevent the mm struct to be freed */
1822 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1823 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1826 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1832 SYSCALL_DEFINE1(userfaultfd, int, flags)
1837 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1842 file = userfaultfd_file_create(flags);
1844 error = PTR_ERR(file);
1845 goto err_put_unused_fd;
1847 fd_install(fd, file);
1857 static int __init userfaultfd_init(void)
1859 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1860 sizeof(struct userfaultfd_ctx),
1862 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1863 init_once_userfaultfd_ctx);
1866 __initcall(userfaultfd_init);