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 {
84 wait_queue_entry_t wq;
85 struct userfaultfd_ctx *ctx;
89 struct userfaultfd_wake_range {
94 static int userfaultfd_wake_function(wait_queue_entry_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 Program-Order guarantees provided by the scheduler
113 * ensure uwq->waken is visible before the task is woken.
115 ret = wake_up_state(wq->private, mode);
118 * Wake only once, autoremove behavior.
120 * After the effect of list_del_init is visible to the other
121 * CPUs, the waitqueue may disappear from under us, see the
122 * !list_empty_careful() in handle_userfault().
124 * try_to_wake_up() has an implicit smp_mb(), and the
125 * wq->private is read before calling the extern function
126 * "wake_up_state" (which in turns calls try_to_wake_up).
128 list_del_init(&wq->entry);
135 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
137 * @ctx: [in] Pointer to the userfaultfd context.
139 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
141 if (!atomic_inc_not_zero(&ctx->refcount))
146 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
148 * @ctx: [in] Pointer to userfaultfd context.
150 * The userfaultfd context reference must have been previously acquired either
151 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
153 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
155 if (atomic_dec_and_test(&ctx->refcount)) {
156 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
157 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
158 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
159 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
160 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
161 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
162 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
163 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
165 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
169 static inline void msg_init(struct uffd_msg *msg)
171 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
173 * Must use memset to zero out the paddings or kernel data is
174 * leaked to userland.
176 memset(msg, 0, sizeof(struct uffd_msg));
179 static inline struct uffd_msg userfault_msg(unsigned long address,
181 unsigned long reason)
185 msg.event = UFFD_EVENT_PAGEFAULT;
186 msg.arg.pagefault.address = address;
187 if (flags & FAULT_FLAG_WRITE)
189 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
190 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
191 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
192 * was a read fault, otherwise if set it means it's
195 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
196 if (reason & VM_UFFD_WP)
198 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
199 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
200 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
201 * a missing fault, otherwise if set it means it's a
202 * write protect fault.
204 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
208 #ifdef CONFIG_HUGETLB_PAGE
210 * Same functionality as userfaultfd_must_wait below with modifications for
213 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
214 struct vm_area_struct *vma,
215 unsigned long address,
217 unsigned long reason)
219 struct mm_struct *mm = ctx->mm;
223 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
225 pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
232 * Lockless access: we're in a wait_event so it's ok if it
235 if (huge_pte_none(*pte))
237 if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
243 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
244 struct vm_area_struct *vma,
245 unsigned long address,
247 unsigned long reason)
249 return false; /* should never get here */
251 #endif /* CONFIG_HUGETLB_PAGE */
254 * Verify the pagetables are still not ok after having reigstered into
255 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
256 * userfault that has already been resolved, if userfaultfd_read and
257 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
260 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
261 unsigned long address,
263 unsigned long reason)
265 struct mm_struct *mm = ctx->mm;
273 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
275 pgd = pgd_offset(mm, address);
276 if (!pgd_present(*pgd))
278 p4d = p4d_offset(pgd, address);
279 if (!p4d_present(*p4d))
281 pud = pud_offset(p4d, address);
282 if (!pud_present(*pud))
284 pmd = pmd_offset(pud, address);
286 * READ_ONCE must function as a barrier with narrower scope
287 * and it must be equivalent to:
288 * _pmd = *pmd; barrier();
290 * This is to deal with the instability (as in
291 * pmd_trans_unstable) of the pmd.
293 _pmd = READ_ONCE(*pmd);
294 if (!pmd_present(_pmd))
298 if (pmd_trans_huge(_pmd))
302 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
303 * and use the standard pte_offset_map() instead of parsing _pmd.
305 pte = pte_offset_map(pmd, address);
307 * Lockless access: we're in a wait_event so it's ok if it
319 * The locking rules involved in returning VM_FAULT_RETRY depending on
320 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
321 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
322 * recommendation in __lock_page_or_retry is not an understatement.
324 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
325 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
328 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
329 * set, VM_FAULT_RETRY can still be returned if and only if there are
330 * fatal_signal_pending()s, and the mmap_sem must be released before
333 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
335 struct mm_struct *mm = vmf->vma->vm_mm;
336 struct userfaultfd_ctx *ctx;
337 struct userfaultfd_wait_queue uwq;
339 bool must_wait, return_to_userland;
342 ret = VM_FAULT_SIGBUS;
345 * We don't do userfault handling for the final child pid update.
347 * We also don't do userfault handling during
348 * coredumping. hugetlbfs has the special
349 * follow_hugetlb_page() to skip missing pages in the
350 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
351 * the no_page_table() helper in follow_page_mask(), but the
352 * shmem_vm_ops->fault method is invoked even during
353 * coredumping without mmap_sem and it ends up here.
355 if (current->flags & (PF_EXITING|PF_DUMPCORE))
359 * Coredumping runs without mmap_sem so we can only check that
360 * the mmap_sem is held, if PF_DUMPCORE was not set.
362 WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
364 ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
368 BUG_ON(ctx->mm != mm);
370 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
371 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
374 * If it's already released don't get it. This avoids to loop
375 * in __get_user_pages if userfaultfd_release waits on the
376 * caller of handle_userfault to release the mmap_sem.
378 if (unlikely(ACCESS_ONCE(ctx->released)))
382 * Check that we can return VM_FAULT_RETRY.
384 * NOTE: it should become possible to return VM_FAULT_RETRY
385 * even if FAULT_FLAG_TRIED is set without leading to gup()
386 * -EBUSY failures, if the userfaultfd is to be extended for
387 * VM_UFFD_WP tracking and we intend to arm the userfault
388 * without first stopping userland access to the memory. For
389 * VM_UFFD_MISSING userfaults this is enough for now.
391 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
393 * Validate the invariant that nowait must allow retry
394 * to be sure not to return SIGBUS erroneously on
395 * nowait invocations.
397 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
398 #ifdef CONFIG_DEBUG_VM
399 if (printk_ratelimit()) {
401 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
410 * Handle nowait, not much to do other than tell it to retry
413 ret = VM_FAULT_RETRY;
414 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
417 /* take the reference before dropping the mmap_sem */
418 userfaultfd_ctx_get(ctx);
420 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
421 uwq.wq.private = current;
422 uwq.msg = userfault_msg(vmf->address, vmf->flags, reason);
427 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
428 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
429 blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
432 spin_lock(&ctx->fault_pending_wqh.lock);
434 * After the __add_wait_queue the uwq is visible to userland
435 * through poll/read().
437 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
439 * The smp_mb() after __set_current_state prevents the reads
440 * following the spin_unlock to happen before the list_add in
443 set_current_state(blocking_state);
444 spin_unlock(&ctx->fault_pending_wqh.lock);
446 if (!is_vm_hugetlb_page(vmf->vma))
447 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
450 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
453 up_read(&mm->mmap_sem);
455 if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
456 (return_to_userland ? !signal_pending(current) :
457 !fatal_signal_pending(current)))) {
458 wake_up_poll(&ctx->fd_wqh, POLLIN);
460 ret |= VM_FAULT_MAJOR;
463 * False wakeups can orginate even from rwsem before
464 * up_read() however userfaults will wait either for a
465 * targeted wakeup on the specific uwq waitqueue from
466 * wake_userfault() or for signals or for uffd
469 while (!READ_ONCE(uwq.waken)) {
471 * This needs the full smp_store_mb()
472 * guarantee as the state write must be
473 * visible to other CPUs before reading
474 * uwq.waken from other CPUs.
476 set_current_state(blocking_state);
477 if (READ_ONCE(uwq.waken) ||
478 READ_ONCE(ctx->released) ||
479 (return_to_userland ? signal_pending(current) :
480 fatal_signal_pending(current)))
486 __set_current_state(TASK_RUNNING);
488 if (return_to_userland) {
489 if (signal_pending(current) &&
490 !fatal_signal_pending(current)) {
492 * If we got a SIGSTOP or SIGCONT and this is
493 * a normal userland page fault, just let
494 * userland return so the signal will be
495 * handled and gdb debugging works. The page
496 * fault code immediately after we return from
497 * this function is going to release the
498 * mmap_sem and it's not depending on it
499 * (unlike gup would if we were not to return
502 * If a fatal signal is pending we still take
503 * the streamlined VM_FAULT_RETRY failure path
504 * and there's no need to retake the mmap_sem
507 down_read(&mm->mmap_sem);
508 ret = VM_FAULT_NOPAGE;
513 * Here we race with the list_del; list_add in
514 * userfaultfd_ctx_read(), however because we don't ever run
515 * list_del_init() to refile across the two lists, the prev
516 * and next pointers will never point to self. list_add also
517 * would never let any of the two pointers to point to
518 * self. So list_empty_careful won't risk to see both pointers
519 * pointing to self at any time during the list refile. The
520 * only case where list_del_init() is called is the full
521 * removal in the wake function and there we don't re-list_add
522 * and it's fine not to block on the spinlock. The uwq on this
523 * kernel stack can be released after the list_del_init.
525 if (!list_empty_careful(&uwq.wq.entry)) {
526 spin_lock(&ctx->fault_pending_wqh.lock);
528 * No need of list_del_init(), the uwq on the stack
529 * will be freed shortly anyway.
531 list_del(&uwq.wq.entry);
532 spin_unlock(&ctx->fault_pending_wqh.lock);
536 * ctx may go away after this if the userfault pseudo fd is
539 userfaultfd_ctx_put(ctx);
545 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
546 struct userfaultfd_wait_queue *ewq)
548 if (WARN_ON_ONCE(current->flags & PF_EXITING))
552 init_waitqueue_entry(&ewq->wq, current);
554 spin_lock(&ctx->event_wqh.lock);
556 * After the __add_wait_queue the uwq is visible to userland
557 * through poll/read().
559 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
561 set_current_state(TASK_KILLABLE);
562 if (ewq->msg.event == 0)
564 if (ACCESS_ONCE(ctx->released) ||
565 fatal_signal_pending(current)) {
566 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
567 if (ewq->msg.event == UFFD_EVENT_FORK) {
568 struct userfaultfd_ctx *new;
570 new = (struct userfaultfd_ctx *)
572 ewq->msg.arg.reserved.reserved1;
574 userfaultfd_ctx_put(new);
579 spin_unlock(&ctx->event_wqh.lock);
581 wake_up_poll(&ctx->fd_wqh, POLLIN);
584 spin_lock(&ctx->event_wqh.lock);
586 __set_current_state(TASK_RUNNING);
587 spin_unlock(&ctx->event_wqh.lock);
590 * ctx may go away after this if the userfault pseudo fd is
594 userfaultfd_ctx_put(ctx);
597 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
598 struct userfaultfd_wait_queue *ewq)
601 wake_up_locked(&ctx->event_wqh);
602 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
605 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
607 struct userfaultfd_ctx *ctx = NULL, *octx;
608 struct userfaultfd_fork_ctx *fctx;
610 octx = vma->vm_userfaultfd_ctx.ctx;
611 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
612 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
613 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
617 list_for_each_entry(fctx, fcs, list)
618 if (fctx->orig == octx) {
624 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
628 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
634 atomic_set(&ctx->refcount, 1);
635 ctx->flags = octx->flags;
636 ctx->state = UFFD_STATE_RUNNING;
637 ctx->features = octx->features;
638 ctx->released = false;
639 ctx->mm = vma->vm_mm;
640 atomic_inc(&ctx->mm->mm_count);
642 userfaultfd_ctx_get(octx);
645 list_add_tail(&fctx->list, fcs);
648 vma->vm_userfaultfd_ctx.ctx = ctx;
652 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
654 struct userfaultfd_ctx *ctx = fctx->orig;
655 struct userfaultfd_wait_queue ewq;
659 ewq.msg.event = UFFD_EVENT_FORK;
660 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
662 userfaultfd_event_wait_completion(ctx, &ewq);
665 void dup_userfaultfd_complete(struct list_head *fcs)
667 struct userfaultfd_fork_ctx *fctx, *n;
669 list_for_each_entry_safe(fctx, n, fcs, list) {
671 list_del(&fctx->list);
676 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
677 struct vm_userfaultfd_ctx *vm_ctx)
679 struct userfaultfd_ctx *ctx;
681 ctx = vma->vm_userfaultfd_ctx.ctx;
682 if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
684 userfaultfd_ctx_get(ctx);
688 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
689 unsigned long from, unsigned long to,
692 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
693 struct userfaultfd_wait_queue ewq;
698 if (to & ~PAGE_MASK) {
699 userfaultfd_ctx_put(ctx);
705 ewq.msg.event = UFFD_EVENT_REMAP;
706 ewq.msg.arg.remap.from = from;
707 ewq.msg.arg.remap.to = to;
708 ewq.msg.arg.remap.len = len;
710 userfaultfd_event_wait_completion(ctx, &ewq);
713 bool userfaultfd_remove(struct vm_area_struct *vma,
714 unsigned long start, unsigned long end)
716 struct mm_struct *mm = vma->vm_mm;
717 struct userfaultfd_ctx *ctx;
718 struct userfaultfd_wait_queue ewq;
720 ctx = vma->vm_userfaultfd_ctx.ctx;
721 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
724 userfaultfd_ctx_get(ctx);
725 up_read(&mm->mmap_sem);
729 ewq.msg.event = UFFD_EVENT_REMOVE;
730 ewq.msg.arg.remove.start = start;
731 ewq.msg.arg.remove.end = end;
733 userfaultfd_event_wait_completion(ctx, &ewq);
738 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
739 unsigned long start, unsigned long end)
741 struct userfaultfd_unmap_ctx *unmap_ctx;
743 list_for_each_entry(unmap_ctx, unmaps, list)
744 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
745 unmap_ctx->end == end)
751 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
752 unsigned long start, unsigned long end,
753 struct list_head *unmaps)
755 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
756 struct userfaultfd_unmap_ctx *unmap_ctx;
757 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
759 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
760 has_unmap_ctx(ctx, unmaps, start, end))
763 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
767 userfaultfd_ctx_get(ctx);
768 unmap_ctx->ctx = ctx;
769 unmap_ctx->start = start;
770 unmap_ctx->end = end;
771 list_add_tail(&unmap_ctx->list, unmaps);
777 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
779 struct userfaultfd_unmap_ctx *ctx, *n;
780 struct userfaultfd_wait_queue ewq;
782 list_for_each_entry_safe(ctx, n, uf, list) {
785 ewq.msg.event = UFFD_EVENT_UNMAP;
786 ewq.msg.arg.remove.start = ctx->start;
787 ewq.msg.arg.remove.end = ctx->end;
789 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
791 list_del(&ctx->list);
796 static int userfaultfd_release(struct inode *inode, struct file *file)
798 struct userfaultfd_ctx *ctx = file->private_data;
799 struct mm_struct *mm = ctx->mm;
800 struct vm_area_struct *vma, *prev;
801 /* len == 0 means wake all */
802 struct userfaultfd_wake_range range = { .len = 0, };
803 unsigned long new_flags;
805 ACCESS_ONCE(ctx->released) = true;
807 if (!mmget_not_zero(mm))
811 * Flush page faults out of all CPUs. NOTE: all page faults
812 * must be retried without returning VM_FAULT_SIGBUS if
813 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
814 * changes while handle_userfault released the mmap_sem. So
815 * it's critical that released is set to true (above), before
816 * taking the mmap_sem for writing.
818 down_write(&mm->mmap_sem);
820 for (vma = mm->mmap; vma; vma = vma->vm_next) {
822 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
823 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
824 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
828 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
829 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
830 new_flags, vma->anon_vma,
831 vma->vm_file, vma->vm_pgoff,
838 vma->vm_flags = new_flags;
839 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
841 up_write(&mm->mmap_sem);
845 * After no new page faults can wait on this fault_*wqh, flush
846 * the last page faults that may have been already waiting on
849 spin_lock(&ctx->fault_pending_wqh.lock);
850 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
851 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
852 spin_unlock(&ctx->fault_pending_wqh.lock);
854 /* Flush pending events that may still wait on event_wqh */
855 wake_up_all(&ctx->event_wqh);
857 wake_up_poll(&ctx->fd_wqh, POLLHUP);
858 userfaultfd_ctx_put(ctx);
862 /* fault_pending_wqh.lock must be hold by the caller */
863 static inline struct userfaultfd_wait_queue *find_userfault_in(
864 wait_queue_head_t *wqh)
866 wait_queue_entry_t *wq;
867 struct userfaultfd_wait_queue *uwq;
869 VM_BUG_ON(!spin_is_locked(&wqh->lock));
872 if (!waitqueue_active(wqh))
874 /* walk in reverse to provide FIFO behavior to read userfaults */
875 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
876 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
881 static inline struct userfaultfd_wait_queue *find_userfault(
882 struct userfaultfd_ctx *ctx)
884 return find_userfault_in(&ctx->fault_pending_wqh);
887 static inline struct userfaultfd_wait_queue *find_userfault_evt(
888 struct userfaultfd_ctx *ctx)
890 return find_userfault_in(&ctx->event_wqh);
893 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
895 struct userfaultfd_ctx *ctx = file->private_data;
898 poll_wait(file, &ctx->fd_wqh, wait);
900 switch (ctx->state) {
901 case UFFD_STATE_WAIT_API:
903 case UFFD_STATE_RUNNING:
905 * poll() never guarantees that read won't block.
906 * userfaults can be waken before they're read().
908 if (unlikely(!(file->f_flags & O_NONBLOCK)))
911 * lockless access to see if there are pending faults
912 * __pollwait last action is the add_wait_queue but
913 * the spin_unlock would allow the waitqueue_active to
914 * pass above the actual list_add inside
915 * add_wait_queue critical section. So use a full
916 * memory barrier to serialize the list_add write of
917 * add_wait_queue() with the waitqueue_active read
922 if (waitqueue_active(&ctx->fault_pending_wqh))
924 else if (waitqueue_active(&ctx->event_wqh))
934 static const struct file_operations userfaultfd_fops;
936 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
937 struct userfaultfd_ctx *new,
938 struct uffd_msg *msg)
942 unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
944 fd = get_unused_fd_flags(flags);
948 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
952 return PTR_ERR(file);
955 fd_install(fd, file);
956 msg->arg.reserved.reserved1 = 0;
957 msg->arg.fork.ufd = fd;
962 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
963 struct uffd_msg *msg)
966 DECLARE_WAITQUEUE(wait, current);
967 struct userfaultfd_wait_queue *uwq;
969 * Handling fork event requires sleeping operations, so
970 * we drop the event_wqh lock, then do these ops, then
971 * lock it back and wake up the waiter. While the lock is
972 * dropped the ewq may go away so we keep track of it
975 LIST_HEAD(fork_event);
976 struct userfaultfd_ctx *fork_nctx = NULL;
978 /* always take the fd_wqh lock before the fault_pending_wqh lock */
979 spin_lock(&ctx->fd_wqh.lock);
980 __add_wait_queue(&ctx->fd_wqh, &wait);
982 set_current_state(TASK_INTERRUPTIBLE);
983 spin_lock(&ctx->fault_pending_wqh.lock);
984 uwq = find_userfault(ctx);
987 * Use a seqcount to repeat the lockless check
988 * in wake_userfault() to avoid missing
989 * wakeups because during the refile both
990 * waitqueue could become empty if this is the
993 write_seqcount_begin(&ctx->refile_seq);
996 * The fault_pending_wqh.lock prevents the uwq
997 * to disappear from under us.
999 * Refile this userfault from
1000 * fault_pending_wqh to fault_wqh, it's not
1001 * pending anymore after we read it.
1003 * Use list_del() by hand (as
1004 * userfaultfd_wake_function also uses
1005 * list_del_init() by hand) to be sure nobody
1006 * changes __remove_wait_queue() to use
1007 * list_del_init() in turn breaking the
1008 * !list_empty_careful() check in
1009 * handle_userfault(). The uwq->wq.head list
1010 * must never be empty at any time during the
1011 * refile, or the waitqueue could disappear
1012 * from under us. The "wait_queue_head_t"
1013 * parameter of __remove_wait_queue() is unused
1016 list_del(&uwq->wq.entry);
1017 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1019 write_seqcount_end(&ctx->refile_seq);
1021 /* careful to always initialize msg if ret == 0 */
1023 spin_unlock(&ctx->fault_pending_wqh.lock);
1027 spin_unlock(&ctx->fault_pending_wqh.lock);
1029 spin_lock(&ctx->event_wqh.lock);
1030 uwq = find_userfault_evt(ctx);
1034 if (uwq->msg.event == UFFD_EVENT_FORK) {
1035 fork_nctx = (struct userfaultfd_ctx *)
1037 uwq->msg.arg.reserved.reserved1;
1038 list_move(&uwq->wq.entry, &fork_event);
1039 spin_unlock(&ctx->event_wqh.lock);
1044 userfaultfd_event_complete(ctx, uwq);
1045 spin_unlock(&ctx->event_wqh.lock);
1049 spin_unlock(&ctx->event_wqh.lock);
1051 if (signal_pending(current)) {
1059 spin_unlock(&ctx->fd_wqh.lock);
1061 spin_lock(&ctx->fd_wqh.lock);
1063 __remove_wait_queue(&ctx->fd_wqh, &wait);
1064 __set_current_state(TASK_RUNNING);
1065 spin_unlock(&ctx->fd_wqh.lock);
1067 if (!ret && msg->event == UFFD_EVENT_FORK) {
1068 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1071 spin_lock(&ctx->event_wqh.lock);
1072 if (!list_empty(&fork_event)) {
1073 uwq = list_first_entry(&fork_event,
1076 list_del(&uwq->wq.entry);
1077 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1078 userfaultfd_event_complete(ctx, uwq);
1080 spin_unlock(&ctx->event_wqh.lock);
1087 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1088 size_t count, loff_t *ppos)
1090 struct userfaultfd_ctx *ctx = file->private_data;
1091 ssize_t _ret, ret = 0;
1092 struct uffd_msg msg;
1093 int no_wait = file->f_flags & O_NONBLOCK;
1095 if (ctx->state == UFFD_STATE_WAIT_API)
1099 if (count < sizeof(msg))
1100 return ret ? ret : -EINVAL;
1101 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1103 return ret ? ret : _ret;
1104 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1105 return ret ? ret : -EFAULT;
1108 count -= sizeof(msg);
1110 * Allow to read more than one fault at time but only
1111 * block if waiting for the very first one.
1113 no_wait = O_NONBLOCK;
1117 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1118 struct userfaultfd_wake_range *range)
1120 spin_lock(&ctx->fault_pending_wqh.lock);
1121 /* wake all in the range and autoremove */
1122 if (waitqueue_active(&ctx->fault_pending_wqh))
1123 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1125 if (waitqueue_active(&ctx->fault_wqh))
1126 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1127 spin_unlock(&ctx->fault_pending_wqh.lock);
1130 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1131 struct userfaultfd_wake_range *range)
1137 * To be sure waitqueue_active() is not reordered by the CPU
1138 * before the pagetable update, use an explicit SMP memory
1139 * barrier here. PT lock release or up_read(mmap_sem) still
1140 * have release semantics that can allow the
1141 * waitqueue_active() to be reordered before the pte update.
1146 * Use waitqueue_active because it's very frequent to
1147 * change the address space atomically even if there are no
1148 * userfaults yet. So we take the spinlock only when we're
1149 * sure we've userfaults to wake.
1152 seq = read_seqcount_begin(&ctx->refile_seq);
1153 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1154 waitqueue_active(&ctx->fault_wqh);
1156 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1158 __wake_userfault(ctx, range);
1161 static __always_inline int validate_range(struct mm_struct *mm,
1162 __u64 start, __u64 len)
1164 __u64 task_size = mm->task_size;
1166 if (start & ~PAGE_MASK)
1168 if (len & ~PAGE_MASK)
1172 if (start < mmap_min_addr)
1174 if (start >= task_size)
1176 if (len > task_size - start)
1181 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1183 return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1187 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1190 struct mm_struct *mm = ctx->mm;
1191 struct vm_area_struct *vma, *prev, *cur;
1193 struct uffdio_register uffdio_register;
1194 struct uffdio_register __user *user_uffdio_register;
1195 unsigned long vm_flags, new_flags;
1197 bool non_anon_pages;
1198 unsigned long start, end, vma_end;
1200 user_uffdio_register = (struct uffdio_register __user *) arg;
1203 if (copy_from_user(&uffdio_register, user_uffdio_register,
1204 sizeof(uffdio_register)-sizeof(__u64)))
1208 if (!uffdio_register.mode)
1210 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1211 UFFDIO_REGISTER_MODE_WP))
1214 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1215 vm_flags |= VM_UFFD_MISSING;
1216 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1217 vm_flags |= VM_UFFD_WP;
1219 * FIXME: remove the below error constraint by
1220 * implementing the wprotect tracking mode.
1226 ret = validate_range(mm, uffdio_register.range.start,
1227 uffdio_register.range.len);
1231 start = uffdio_register.range.start;
1232 end = start + uffdio_register.range.len;
1235 if (!mmget_not_zero(mm))
1238 down_write(&mm->mmap_sem);
1239 vma = find_vma_prev(mm, start, &prev);
1243 /* check that there's at least one vma in the range */
1245 if (vma->vm_start >= end)
1249 * If the first vma contains huge pages, make sure start address
1250 * is aligned to huge page size.
1252 if (is_vm_hugetlb_page(vma)) {
1253 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1255 if (start & (vma_hpagesize - 1))
1260 * Search for not compatible vmas.
1263 non_anon_pages = false;
1264 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1267 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1268 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1270 /* check not compatible vmas */
1272 if (!vma_can_userfault(cur))
1275 * If this vma contains ending address, and huge pages
1278 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1279 end > cur->vm_start) {
1280 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1284 if (end & (vma_hpagesize - 1))
1289 * Check that this vma isn't already owned by a
1290 * different userfaultfd. We can't allow more than one
1291 * userfaultfd to own a single vma simultaneously or we
1292 * wouldn't know which one to deliver the userfaults to.
1295 if (cur->vm_userfaultfd_ctx.ctx &&
1296 cur->vm_userfaultfd_ctx.ctx != ctx)
1300 * Note vmas containing huge pages
1302 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1303 non_anon_pages = true;
1309 if (vma->vm_start < start)
1316 BUG_ON(!vma_can_userfault(vma));
1317 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1318 vma->vm_userfaultfd_ctx.ctx != ctx);
1321 * Nothing to do: this vma is already registered into this
1322 * userfaultfd and with the right tracking mode too.
1324 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1325 (vma->vm_flags & vm_flags) == vm_flags)
1328 if (vma->vm_start > start)
1329 start = vma->vm_start;
1330 vma_end = min(end, vma->vm_end);
1332 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1333 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1334 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1336 ((struct vm_userfaultfd_ctx){ ctx }));
1341 if (vma->vm_start < start) {
1342 ret = split_vma(mm, vma, start, 1);
1346 if (vma->vm_end > end) {
1347 ret = split_vma(mm, vma, end, 0);
1353 * In the vma_merge() successful mprotect-like case 8:
1354 * the next vma was merged into the current one and
1355 * the current one has not been updated yet.
1357 vma->vm_flags = new_flags;
1358 vma->vm_userfaultfd_ctx.ctx = ctx;
1362 start = vma->vm_end;
1364 } while (vma && vma->vm_start < end);
1366 up_write(&mm->mmap_sem);
1370 * Now that we scanned all vmas we can already tell
1371 * userland which ioctls methods are guaranteed to
1372 * succeed on this range.
1374 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1375 UFFD_API_RANGE_IOCTLS,
1376 &user_uffdio_register->ioctls))
1383 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1386 struct mm_struct *mm = ctx->mm;
1387 struct vm_area_struct *vma, *prev, *cur;
1389 struct uffdio_range uffdio_unregister;
1390 unsigned long new_flags;
1392 unsigned long start, end, vma_end;
1393 const void __user *buf = (void __user *)arg;
1396 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1399 ret = validate_range(mm, uffdio_unregister.start,
1400 uffdio_unregister.len);
1404 start = uffdio_unregister.start;
1405 end = start + uffdio_unregister.len;
1408 if (!mmget_not_zero(mm))
1411 down_write(&mm->mmap_sem);
1412 vma = find_vma_prev(mm, start, &prev);
1416 /* check that there's at least one vma in the range */
1418 if (vma->vm_start >= end)
1422 * If the first vma contains huge pages, make sure start address
1423 * is aligned to huge page size.
1425 if (is_vm_hugetlb_page(vma)) {
1426 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1428 if (start & (vma_hpagesize - 1))
1433 * Search for not compatible vmas.
1437 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1440 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1441 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1444 * Check not compatible vmas, not strictly required
1445 * here as not compatible vmas cannot have an
1446 * userfaultfd_ctx registered on them, but this
1447 * provides for more strict behavior to notice
1448 * unregistration errors.
1450 if (!vma_can_userfault(cur))
1457 if (vma->vm_start < start)
1464 BUG_ON(!vma_can_userfault(vma));
1467 * Nothing to do: this vma is already registered into this
1468 * userfaultfd and with the right tracking mode too.
1470 if (!vma->vm_userfaultfd_ctx.ctx)
1473 if (vma->vm_start > start)
1474 start = vma->vm_start;
1475 vma_end = min(end, vma->vm_end);
1477 if (userfaultfd_missing(vma)) {
1479 * Wake any concurrent pending userfault while
1480 * we unregister, so they will not hang
1481 * permanently and it avoids userland to call
1482 * UFFDIO_WAKE explicitly.
1484 struct userfaultfd_wake_range range;
1485 range.start = start;
1486 range.len = vma_end - start;
1487 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1490 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1491 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1492 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1499 if (vma->vm_start < start) {
1500 ret = split_vma(mm, vma, start, 1);
1504 if (vma->vm_end > end) {
1505 ret = split_vma(mm, vma, end, 0);
1511 * In the vma_merge() successful mprotect-like case 8:
1512 * the next vma was merged into the current one and
1513 * the current one has not been updated yet.
1515 vma->vm_flags = new_flags;
1516 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1520 start = vma->vm_end;
1522 } while (vma && vma->vm_start < end);
1524 up_write(&mm->mmap_sem);
1531 * userfaultfd_wake may be used in combination with the
1532 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1534 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1538 struct uffdio_range uffdio_wake;
1539 struct userfaultfd_wake_range range;
1540 const void __user *buf = (void __user *)arg;
1543 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1546 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1550 range.start = uffdio_wake.start;
1551 range.len = uffdio_wake.len;
1554 * len == 0 means wake all and we don't want to wake all here,
1555 * so check it again to be sure.
1557 VM_BUG_ON(!range.len);
1559 wake_userfault(ctx, &range);
1566 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1570 struct uffdio_copy uffdio_copy;
1571 struct uffdio_copy __user *user_uffdio_copy;
1572 struct userfaultfd_wake_range range;
1574 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1577 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1578 /* don't copy "copy" last field */
1579 sizeof(uffdio_copy)-sizeof(__s64)))
1582 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1586 * double check for wraparound just in case. copy_from_user()
1587 * will later check uffdio_copy.src + uffdio_copy.len to fit
1588 * in the userland range.
1591 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1593 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1595 if (mmget_not_zero(ctx->mm)) {
1596 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1602 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1607 /* len == 0 would wake all */
1609 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1610 range.start = uffdio_copy.dst;
1611 wake_userfault(ctx, &range);
1613 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1618 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1622 struct uffdio_zeropage uffdio_zeropage;
1623 struct uffdio_zeropage __user *user_uffdio_zeropage;
1624 struct userfaultfd_wake_range range;
1626 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1629 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1630 /* don't copy "zeropage" last field */
1631 sizeof(uffdio_zeropage)-sizeof(__s64)))
1634 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1635 uffdio_zeropage.range.len);
1639 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1642 if (mmget_not_zero(ctx->mm)) {
1643 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1644 uffdio_zeropage.range.len);
1649 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1653 /* len == 0 would wake all */
1656 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1657 range.start = uffdio_zeropage.range.start;
1658 wake_userfault(ctx, &range);
1660 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1665 static inline unsigned int uffd_ctx_features(__u64 user_features)
1668 * For the current set of features the bits just coincide
1670 return (unsigned int)user_features;
1674 * userland asks for a certain API version and we return which bits
1675 * and ioctl commands are implemented in this kernel for such API
1676 * version or -EINVAL if unknown.
1678 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1681 struct uffdio_api uffdio_api;
1682 void __user *buf = (void __user *)arg;
1687 if (ctx->state != UFFD_STATE_WAIT_API)
1690 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1692 features = uffdio_api.features;
1693 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1694 memset(&uffdio_api, 0, sizeof(uffdio_api));
1695 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1700 /* report all available features and ioctls to userland */
1701 uffdio_api.features = UFFD_API_FEATURES;
1702 uffdio_api.ioctls = UFFD_API_IOCTLS;
1704 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1706 ctx->state = UFFD_STATE_RUNNING;
1707 /* only enable the requested features for this uffd context */
1708 ctx->features = uffd_ctx_features(features);
1714 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1718 struct userfaultfd_ctx *ctx = file->private_data;
1720 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1725 ret = userfaultfd_api(ctx, arg);
1727 case UFFDIO_REGISTER:
1728 ret = userfaultfd_register(ctx, arg);
1730 case UFFDIO_UNREGISTER:
1731 ret = userfaultfd_unregister(ctx, arg);
1734 ret = userfaultfd_wake(ctx, arg);
1737 ret = userfaultfd_copy(ctx, arg);
1739 case UFFDIO_ZEROPAGE:
1740 ret = userfaultfd_zeropage(ctx, arg);
1746 #ifdef CONFIG_PROC_FS
1747 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1749 struct userfaultfd_ctx *ctx = f->private_data;
1750 wait_queue_entry_t *wq;
1751 struct userfaultfd_wait_queue *uwq;
1752 unsigned long pending = 0, total = 0;
1754 spin_lock(&ctx->fault_pending_wqh.lock);
1755 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1756 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1760 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1761 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1764 spin_unlock(&ctx->fault_pending_wqh.lock);
1767 * If more protocols will be added, there will be all shown
1768 * separated by a space. Like this:
1769 * protocols: aa:... bb:...
1771 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1772 pending, total, UFFD_API, ctx->features,
1773 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1777 static const struct file_operations userfaultfd_fops = {
1778 #ifdef CONFIG_PROC_FS
1779 .show_fdinfo = userfaultfd_show_fdinfo,
1781 .release = userfaultfd_release,
1782 .poll = userfaultfd_poll,
1783 .read = userfaultfd_read,
1784 .unlocked_ioctl = userfaultfd_ioctl,
1785 .compat_ioctl = userfaultfd_ioctl,
1786 .llseek = noop_llseek,
1789 static void init_once_userfaultfd_ctx(void *mem)
1791 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1793 init_waitqueue_head(&ctx->fault_pending_wqh);
1794 init_waitqueue_head(&ctx->fault_wqh);
1795 init_waitqueue_head(&ctx->event_wqh);
1796 init_waitqueue_head(&ctx->fd_wqh);
1797 seqcount_init(&ctx->refile_seq);
1801 * userfaultfd_file_create - Creates a userfaultfd file pointer.
1802 * @flags: Flags for the userfaultfd file.
1804 * This function creates a userfaultfd file pointer, w/out installing
1805 * it into the fd table. This is useful when the userfaultfd file is
1806 * used during the initialization of data structures that require
1807 * extra setup after the userfaultfd creation. So the userfaultfd
1808 * creation is split into the file pointer creation phase, and the
1809 * file descriptor installation phase. In this way races with
1810 * userspace closing the newly installed file descriptor can be
1811 * avoided. Returns a userfaultfd file pointer, or a proper error
1814 static struct file *userfaultfd_file_create(int flags)
1817 struct userfaultfd_ctx *ctx;
1819 BUG_ON(!current->mm);
1821 /* Check the UFFD_* constants for consistency. */
1822 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1823 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1825 file = ERR_PTR(-EINVAL);
1826 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1829 file = ERR_PTR(-ENOMEM);
1830 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1834 atomic_set(&ctx->refcount, 1);
1837 ctx->state = UFFD_STATE_WAIT_API;
1838 ctx->released = false;
1839 ctx->mm = current->mm;
1840 /* prevent the mm struct to be freed */
1843 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1844 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1847 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1853 SYSCALL_DEFINE1(userfaultfd, int, flags)
1858 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1863 file = userfaultfd_file_create(flags);
1865 error = PTR_ERR(file);
1866 goto err_put_unused_fd;
1868 fd_install(fd, file);
1878 static int __init userfaultfd_init(void)
1880 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1881 sizeof(struct userfaultfd_ctx),
1883 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1884 init_once_userfaultfd_ctx);
1887 __initcall(userfaultfd_init);