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;
36 * Start with fault_pending_wqh and fault_wqh so they're more likely
37 * to be in the same cacheline.
41 * fault_pending_wqh.lock
45 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
46 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
47 * also taken in IRQ context.
49 struct userfaultfd_ctx {
50 /* waitqueue head for the pending (i.e. not read) userfaults */
51 wait_queue_head_t fault_pending_wqh;
52 /* waitqueue head for the userfaults */
53 wait_queue_head_t fault_wqh;
54 /* waitqueue head for the pseudo fd to wakeup poll/read */
55 wait_queue_head_t fd_wqh;
56 /* waitqueue head for events */
57 wait_queue_head_t event_wqh;
58 /* a refile sequence protected by fault_pending_wqh lock */
59 struct seqcount refile_seq;
60 /* pseudo fd refcounting */
62 /* userfaultfd syscall flags */
64 /* features requested from the userspace */
65 unsigned int features;
68 /* memory mappings are changing because of non-cooperative event */
70 /* mm with one ore more vmas attached to this userfaultfd_ctx */
74 struct userfaultfd_fork_ctx {
75 struct userfaultfd_ctx *orig;
76 struct userfaultfd_ctx *new;
77 struct list_head list;
80 struct userfaultfd_unmap_ctx {
81 struct userfaultfd_ctx *ctx;
84 struct list_head list;
87 struct userfaultfd_wait_queue {
89 wait_queue_entry_t wq;
90 struct userfaultfd_ctx *ctx;
94 struct userfaultfd_wake_range {
99 /* internal indication that UFFD_API ioctl was successfully executed */
100 #define UFFD_FEATURE_INITIALIZED (1u << 31)
102 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
104 return ctx->features & UFFD_FEATURE_INITIALIZED;
107 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
108 int wake_flags, void *key)
110 struct userfaultfd_wake_range *range = key;
112 struct userfaultfd_wait_queue *uwq;
113 unsigned long start, len;
115 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
117 /* len == 0 means wake all */
118 start = range->start;
120 if (len && (start > uwq->msg.arg.pagefault.address ||
121 start + len <= uwq->msg.arg.pagefault.address))
123 WRITE_ONCE(uwq->waken, true);
125 * The Program-Order guarantees provided by the scheduler
126 * ensure uwq->waken is visible before the task is woken.
128 ret = wake_up_state(wq->private, mode);
131 * Wake only once, autoremove behavior.
133 * After the effect of list_del_init is visible to the other
134 * CPUs, the waitqueue may disappear from under us, see the
135 * !list_empty_careful() in handle_userfault().
137 * try_to_wake_up() has an implicit smp_mb(), and the
138 * wq->private is read before calling the extern function
139 * "wake_up_state" (which in turns calls try_to_wake_up).
141 list_del_init(&wq->entry);
148 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
150 * @ctx: [in] Pointer to the userfaultfd context.
152 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
154 if (!atomic_inc_not_zero(&ctx->refcount))
159 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
161 * @ctx: [in] Pointer to userfaultfd context.
163 * The userfaultfd context reference must have been previously acquired either
164 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
166 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
168 if (atomic_dec_and_test(&ctx->refcount)) {
169 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
170 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
171 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
172 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
173 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
174 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
175 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
176 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
178 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
182 static inline void msg_init(struct uffd_msg *msg)
184 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
186 * Must use memset to zero out the paddings or kernel data is
187 * leaked to userland.
189 memset(msg, 0, sizeof(struct uffd_msg));
192 static inline struct uffd_msg userfault_msg(unsigned long address,
194 unsigned long reason,
195 unsigned int features)
199 msg.event = UFFD_EVENT_PAGEFAULT;
200 msg.arg.pagefault.address = address;
201 if (flags & FAULT_FLAG_WRITE)
203 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
204 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
205 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
206 * was a read fault, otherwise if set it means it's
209 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
210 if (reason & VM_UFFD_WP)
212 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
213 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
214 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
215 * a missing fault, otherwise if set it means it's a
216 * write protect fault.
218 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
219 if (features & UFFD_FEATURE_THREAD_ID)
220 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
224 #ifdef CONFIG_HUGETLB_PAGE
226 * Same functionality as userfaultfd_must_wait below with modifications for
229 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
230 struct vm_area_struct *vma,
231 unsigned long address,
233 unsigned long reason)
235 struct mm_struct *mm = ctx->mm;
239 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
241 ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
247 pte = huge_ptep_get(ptep);
250 * Lockless access: we're in a wait_event so it's ok if it
253 if (huge_pte_none(pte))
255 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
261 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
262 struct vm_area_struct *vma,
263 unsigned long address,
265 unsigned long reason)
267 return false; /* should never get here */
269 #endif /* CONFIG_HUGETLB_PAGE */
272 * Verify the pagetables are still not ok after having reigstered into
273 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
274 * userfault that has already been resolved, if userfaultfd_read and
275 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
278 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
279 unsigned long address,
281 unsigned long reason)
283 struct mm_struct *mm = ctx->mm;
291 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
293 pgd = pgd_offset(mm, address);
294 if (!pgd_present(*pgd))
296 p4d = p4d_offset(pgd, address);
297 if (!p4d_present(*p4d))
299 pud = pud_offset(p4d, address);
300 if (!pud_present(*pud))
302 pmd = pmd_offset(pud, address);
304 * READ_ONCE must function as a barrier with narrower scope
305 * and it must be equivalent to:
306 * _pmd = *pmd; barrier();
308 * This is to deal with the instability (as in
309 * pmd_trans_unstable) of the pmd.
311 _pmd = READ_ONCE(*pmd);
316 if (!pmd_present(_pmd))
319 if (pmd_trans_huge(_pmd))
323 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
324 * and use the standard pte_offset_map() instead of parsing _pmd.
326 pte = pte_offset_map(pmd, address);
328 * Lockless access: we're in a wait_event so it's ok if it
340 * The locking rules involved in returning VM_FAULT_RETRY depending on
341 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
342 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
343 * recommendation in __lock_page_or_retry is not an understatement.
345 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
346 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
349 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
350 * set, VM_FAULT_RETRY can still be returned if and only if there are
351 * fatal_signal_pending()s, and the mmap_sem must be released before
354 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
356 struct mm_struct *mm = vmf->vma->vm_mm;
357 struct userfaultfd_ctx *ctx;
358 struct userfaultfd_wait_queue uwq;
359 vm_fault_t ret = VM_FAULT_SIGBUS;
360 bool must_wait, return_to_userland;
364 * We don't do userfault handling for the final child pid update.
366 * We also don't do userfault handling during
367 * coredumping. hugetlbfs has the special
368 * follow_hugetlb_page() to skip missing pages in the
369 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
370 * the no_page_table() helper in follow_page_mask(), but the
371 * shmem_vm_ops->fault method is invoked even during
372 * coredumping without mmap_sem and it ends up here.
374 if (current->flags & (PF_EXITING|PF_DUMPCORE))
378 * Coredumping runs without mmap_sem so we can only check that
379 * the mmap_sem is held, if PF_DUMPCORE was not set.
381 WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
383 ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
387 BUG_ON(ctx->mm != mm);
389 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
390 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
392 if (ctx->features & UFFD_FEATURE_SIGBUS)
396 * If it's already released don't get it. This avoids to loop
397 * in __get_user_pages if userfaultfd_release waits on the
398 * caller of handle_userfault to release the mmap_sem.
400 if (unlikely(READ_ONCE(ctx->released))) {
402 * Don't return VM_FAULT_SIGBUS in this case, so a non
403 * cooperative manager can close the uffd after the
404 * last UFFDIO_COPY, without risking to trigger an
405 * involuntary SIGBUS if the process was starting the
406 * userfaultfd while the userfaultfd was still armed
407 * (but after the last UFFDIO_COPY). If the uffd
408 * wasn't already closed when the userfault reached
409 * this point, that would normally be solved by
410 * userfaultfd_must_wait returning 'false'.
412 * If we were to return VM_FAULT_SIGBUS here, the non
413 * cooperative manager would be instead forced to
414 * always call UFFDIO_UNREGISTER before it can safely
417 ret = VM_FAULT_NOPAGE;
422 * Check that we can return VM_FAULT_RETRY.
424 * NOTE: it should become possible to return VM_FAULT_RETRY
425 * even if FAULT_FLAG_TRIED is set without leading to gup()
426 * -EBUSY failures, if the userfaultfd is to be extended for
427 * VM_UFFD_WP tracking and we intend to arm the userfault
428 * without first stopping userland access to the memory. For
429 * VM_UFFD_MISSING userfaults this is enough for now.
431 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
433 * Validate the invariant that nowait must allow retry
434 * to be sure not to return SIGBUS erroneously on
435 * nowait invocations.
437 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
438 #ifdef CONFIG_DEBUG_VM
439 if (printk_ratelimit()) {
441 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
450 * Handle nowait, not much to do other than tell it to retry
453 ret = VM_FAULT_RETRY;
454 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
457 /* take the reference before dropping the mmap_sem */
458 userfaultfd_ctx_get(ctx);
460 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
461 uwq.wq.private = current;
462 uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
468 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
469 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
470 blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
473 spin_lock_irq(&ctx->fault_pending_wqh.lock);
475 * After the __add_wait_queue the uwq is visible to userland
476 * through poll/read().
478 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
480 * The smp_mb() after __set_current_state prevents the reads
481 * following the spin_unlock to happen before the list_add in
484 set_current_state(blocking_state);
485 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
487 if (!is_vm_hugetlb_page(vmf->vma))
488 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
491 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
494 up_read(&mm->mmap_sem);
496 if (likely(must_wait && !READ_ONCE(ctx->released) &&
497 (return_to_userland ? !signal_pending(current) :
498 !fatal_signal_pending(current)))) {
499 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
501 ret |= VM_FAULT_MAJOR;
504 * False wakeups can orginate even from rwsem before
505 * up_read() however userfaults will wait either for a
506 * targeted wakeup on the specific uwq waitqueue from
507 * wake_userfault() or for signals or for uffd
510 while (!READ_ONCE(uwq.waken)) {
512 * This needs the full smp_store_mb()
513 * guarantee as the state write must be
514 * visible to other CPUs before reading
515 * uwq.waken from other CPUs.
517 set_current_state(blocking_state);
518 if (READ_ONCE(uwq.waken) ||
519 READ_ONCE(ctx->released) ||
520 (return_to_userland ? signal_pending(current) :
521 fatal_signal_pending(current)))
527 __set_current_state(TASK_RUNNING);
529 if (return_to_userland) {
530 if (signal_pending(current) &&
531 !fatal_signal_pending(current)) {
533 * If we got a SIGSTOP or SIGCONT and this is
534 * a normal userland page fault, just let
535 * userland return so the signal will be
536 * handled and gdb debugging works. The page
537 * fault code immediately after we return from
538 * this function is going to release the
539 * mmap_sem and it's not depending on it
540 * (unlike gup would if we were not to return
543 * If a fatal signal is pending we still take
544 * the streamlined VM_FAULT_RETRY failure path
545 * and there's no need to retake the mmap_sem
548 down_read(&mm->mmap_sem);
549 ret = VM_FAULT_NOPAGE;
554 * Here we race with the list_del; list_add in
555 * userfaultfd_ctx_read(), however because we don't ever run
556 * list_del_init() to refile across the two lists, the prev
557 * and next pointers will never point to self. list_add also
558 * would never let any of the two pointers to point to
559 * self. So list_empty_careful won't risk to see both pointers
560 * pointing to self at any time during the list refile. The
561 * only case where list_del_init() is called is the full
562 * removal in the wake function and there we don't re-list_add
563 * and it's fine not to block on the spinlock. The uwq on this
564 * kernel stack can be released after the list_del_init.
566 if (!list_empty_careful(&uwq.wq.entry)) {
567 spin_lock_irq(&ctx->fault_pending_wqh.lock);
569 * No need of list_del_init(), the uwq on the stack
570 * will be freed shortly anyway.
572 list_del(&uwq.wq.entry);
573 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
577 * ctx may go away after this if the userfault pseudo fd is
580 userfaultfd_ctx_put(ctx);
586 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
587 struct userfaultfd_wait_queue *ewq)
589 struct userfaultfd_ctx *release_new_ctx;
591 if (WARN_ON_ONCE(current->flags & PF_EXITING))
595 init_waitqueue_entry(&ewq->wq, current);
596 release_new_ctx = NULL;
598 spin_lock_irq(&ctx->event_wqh.lock);
600 * After the __add_wait_queue the uwq is visible to userland
601 * through poll/read().
603 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
605 set_current_state(TASK_KILLABLE);
606 if (ewq->msg.event == 0)
608 if (READ_ONCE(ctx->released) ||
609 fatal_signal_pending(current)) {
611 * &ewq->wq may be queued in fork_event, but
612 * __remove_wait_queue ignores the head
613 * parameter. It would be a problem if it
616 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
617 if (ewq->msg.event == UFFD_EVENT_FORK) {
618 struct userfaultfd_ctx *new;
620 new = (struct userfaultfd_ctx *)
622 ewq->msg.arg.reserved.reserved1;
623 release_new_ctx = new;
628 spin_unlock_irq(&ctx->event_wqh.lock);
630 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
633 spin_lock_irq(&ctx->event_wqh.lock);
635 __set_current_state(TASK_RUNNING);
636 spin_unlock_irq(&ctx->event_wqh.lock);
638 if (release_new_ctx) {
639 struct vm_area_struct *vma;
640 struct mm_struct *mm = release_new_ctx->mm;
642 /* the various vma->vm_userfaultfd_ctx still points to it */
643 down_write(&mm->mmap_sem);
644 /* no task can run (and in turn coredump) yet */
645 VM_WARN_ON(!mmget_still_valid(mm));
646 for (vma = mm->mmap; vma; vma = vma->vm_next)
647 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
648 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
649 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
651 up_write(&mm->mmap_sem);
653 userfaultfd_ctx_put(release_new_ctx);
657 * ctx may go away after this if the userfault pseudo fd is
661 WRITE_ONCE(ctx->mmap_changing, false);
662 userfaultfd_ctx_put(ctx);
665 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
666 struct userfaultfd_wait_queue *ewq)
669 wake_up_locked(&ctx->event_wqh);
670 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
673 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
675 struct userfaultfd_ctx *ctx = NULL, *octx;
676 struct userfaultfd_fork_ctx *fctx;
678 octx = vma->vm_userfaultfd_ctx.ctx;
679 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
680 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
681 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
685 list_for_each_entry(fctx, fcs, list)
686 if (fctx->orig == octx) {
692 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
696 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
702 atomic_set(&ctx->refcount, 1);
703 ctx->flags = octx->flags;
704 ctx->features = octx->features;
705 ctx->released = false;
706 ctx->mmap_changing = false;
707 ctx->mm = vma->vm_mm;
710 userfaultfd_ctx_get(octx);
711 WRITE_ONCE(octx->mmap_changing, true);
714 list_add_tail(&fctx->list, fcs);
717 vma->vm_userfaultfd_ctx.ctx = ctx;
721 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
723 struct userfaultfd_ctx *ctx = fctx->orig;
724 struct userfaultfd_wait_queue ewq;
728 ewq.msg.event = UFFD_EVENT_FORK;
729 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
731 userfaultfd_event_wait_completion(ctx, &ewq);
734 void dup_userfaultfd_complete(struct list_head *fcs)
736 struct userfaultfd_fork_ctx *fctx, *n;
738 list_for_each_entry_safe(fctx, n, fcs, list) {
740 list_del(&fctx->list);
745 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
746 struct vm_userfaultfd_ctx *vm_ctx)
748 struct userfaultfd_ctx *ctx;
750 ctx = vma->vm_userfaultfd_ctx.ctx;
755 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
757 userfaultfd_ctx_get(ctx);
758 WRITE_ONCE(ctx->mmap_changing, true);
760 /* Drop uffd context if remap feature not enabled */
761 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
762 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
766 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
767 unsigned long from, unsigned long to,
770 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
771 struct userfaultfd_wait_queue ewq;
776 if (to & ~PAGE_MASK) {
777 userfaultfd_ctx_put(ctx);
783 ewq.msg.event = UFFD_EVENT_REMAP;
784 ewq.msg.arg.remap.from = from;
785 ewq.msg.arg.remap.to = to;
786 ewq.msg.arg.remap.len = len;
788 userfaultfd_event_wait_completion(ctx, &ewq);
791 bool userfaultfd_remove(struct vm_area_struct *vma,
792 unsigned long start, unsigned long end)
794 struct mm_struct *mm = vma->vm_mm;
795 struct userfaultfd_ctx *ctx;
796 struct userfaultfd_wait_queue ewq;
798 ctx = vma->vm_userfaultfd_ctx.ctx;
799 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
802 userfaultfd_ctx_get(ctx);
803 WRITE_ONCE(ctx->mmap_changing, true);
804 up_read(&mm->mmap_sem);
808 ewq.msg.event = UFFD_EVENT_REMOVE;
809 ewq.msg.arg.remove.start = start;
810 ewq.msg.arg.remove.end = end;
812 userfaultfd_event_wait_completion(ctx, &ewq);
817 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
818 unsigned long start, unsigned long end)
820 struct userfaultfd_unmap_ctx *unmap_ctx;
822 list_for_each_entry(unmap_ctx, unmaps, list)
823 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
824 unmap_ctx->end == end)
830 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
831 unsigned long start, unsigned long end,
832 struct list_head *unmaps)
834 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
835 struct userfaultfd_unmap_ctx *unmap_ctx;
836 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
838 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
839 has_unmap_ctx(ctx, unmaps, start, end))
842 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
846 userfaultfd_ctx_get(ctx);
847 WRITE_ONCE(ctx->mmap_changing, true);
848 unmap_ctx->ctx = ctx;
849 unmap_ctx->start = start;
850 unmap_ctx->end = end;
851 list_add_tail(&unmap_ctx->list, unmaps);
857 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
859 struct userfaultfd_unmap_ctx *ctx, *n;
860 struct userfaultfd_wait_queue ewq;
862 list_for_each_entry_safe(ctx, n, uf, list) {
865 ewq.msg.event = UFFD_EVENT_UNMAP;
866 ewq.msg.arg.remove.start = ctx->start;
867 ewq.msg.arg.remove.end = ctx->end;
869 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
871 list_del(&ctx->list);
876 static int userfaultfd_release(struct inode *inode, struct file *file)
878 struct userfaultfd_ctx *ctx = file->private_data;
879 struct mm_struct *mm = ctx->mm;
880 struct vm_area_struct *vma, *prev;
881 /* len == 0 means wake all */
882 struct userfaultfd_wake_range range = { .len = 0, };
883 unsigned long new_flags;
886 WRITE_ONCE(ctx->released, true);
888 if (!mmget_not_zero(mm))
892 * Flush page faults out of all CPUs. NOTE: all page faults
893 * must be retried without returning VM_FAULT_SIGBUS if
894 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
895 * changes while handle_userfault released the mmap_sem. So
896 * it's critical that released is set to true (above), before
897 * taking the mmap_sem for writing.
899 down_write(&mm->mmap_sem);
900 still_valid = mmget_still_valid(mm);
902 for (vma = mm->mmap; vma; vma = vma->vm_next) {
904 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
905 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
906 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
910 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
912 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
913 new_flags, vma->anon_vma,
914 vma->vm_file, vma->vm_pgoff,
922 vma->vm_flags = new_flags;
923 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
925 up_write(&mm->mmap_sem);
929 * After no new page faults can wait on this fault_*wqh, flush
930 * the last page faults that may have been already waiting on
933 spin_lock_irq(&ctx->fault_pending_wqh.lock);
934 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
935 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
936 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
938 /* Flush pending events that may still wait on event_wqh */
939 wake_up_all(&ctx->event_wqh);
941 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
942 userfaultfd_ctx_put(ctx);
946 /* fault_pending_wqh.lock must be hold by the caller */
947 static inline struct userfaultfd_wait_queue *find_userfault_in(
948 wait_queue_head_t *wqh)
950 wait_queue_entry_t *wq;
951 struct userfaultfd_wait_queue *uwq;
953 VM_BUG_ON(!spin_is_locked(&wqh->lock));
956 if (!waitqueue_active(wqh))
958 /* walk in reverse to provide FIFO behavior to read userfaults */
959 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
960 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
965 static inline struct userfaultfd_wait_queue *find_userfault(
966 struct userfaultfd_ctx *ctx)
968 return find_userfault_in(&ctx->fault_pending_wqh);
971 static inline struct userfaultfd_wait_queue *find_userfault_evt(
972 struct userfaultfd_ctx *ctx)
974 return find_userfault_in(&ctx->event_wqh);
977 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
979 struct userfaultfd_ctx *ctx = file->private_data;
982 poll_wait(file, &ctx->fd_wqh, wait);
984 if (!userfaultfd_is_initialized(ctx))
988 * poll() never guarantees that read won't block.
989 * userfaults can be waken before they're read().
991 if (unlikely(!(file->f_flags & O_NONBLOCK)))
994 * lockless access to see if there are pending faults
995 * __pollwait last action is the add_wait_queue but
996 * the spin_unlock would allow the waitqueue_active to
997 * pass above the actual list_add inside
998 * add_wait_queue critical section. So use a full
999 * memory barrier to serialize the list_add write of
1000 * add_wait_queue() with the waitqueue_active read
1005 if (waitqueue_active(&ctx->fault_pending_wqh))
1007 else if (waitqueue_active(&ctx->event_wqh))
1013 static const struct file_operations userfaultfd_fops;
1015 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
1016 struct userfaultfd_ctx *new,
1017 struct uffd_msg *msg)
1021 fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
1022 O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
1026 msg->arg.reserved.reserved1 = 0;
1027 msg->arg.fork.ufd = fd;
1031 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1032 struct uffd_msg *msg)
1035 DECLARE_WAITQUEUE(wait, current);
1036 struct userfaultfd_wait_queue *uwq;
1038 * Handling fork event requires sleeping operations, so
1039 * we drop the event_wqh lock, then do these ops, then
1040 * lock it back and wake up the waiter. While the lock is
1041 * dropped the ewq may go away so we keep track of it
1044 LIST_HEAD(fork_event);
1045 struct userfaultfd_ctx *fork_nctx = NULL;
1047 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1048 spin_lock_irq(&ctx->fd_wqh.lock);
1049 __add_wait_queue(&ctx->fd_wqh, &wait);
1051 set_current_state(TASK_INTERRUPTIBLE);
1052 spin_lock(&ctx->fault_pending_wqh.lock);
1053 uwq = find_userfault(ctx);
1056 * Use a seqcount to repeat the lockless check
1057 * in wake_userfault() to avoid missing
1058 * wakeups because during the refile both
1059 * waitqueue could become empty if this is the
1062 write_seqcount_begin(&ctx->refile_seq);
1065 * The fault_pending_wqh.lock prevents the uwq
1066 * to disappear from under us.
1068 * Refile this userfault from
1069 * fault_pending_wqh to fault_wqh, it's not
1070 * pending anymore after we read it.
1072 * Use list_del() by hand (as
1073 * userfaultfd_wake_function also uses
1074 * list_del_init() by hand) to be sure nobody
1075 * changes __remove_wait_queue() to use
1076 * list_del_init() in turn breaking the
1077 * !list_empty_careful() check in
1078 * handle_userfault(). The uwq->wq.head list
1079 * must never be empty at any time during the
1080 * refile, or the waitqueue could disappear
1081 * from under us. The "wait_queue_head_t"
1082 * parameter of __remove_wait_queue() is unused
1085 list_del(&uwq->wq.entry);
1086 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1088 write_seqcount_end(&ctx->refile_seq);
1090 /* careful to always initialize msg if ret == 0 */
1092 spin_unlock(&ctx->fault_pending_wqh.lock);
1096 spin_unlock(&ctx->fault_pending_wqh.lock);
1098 spin_lock(&ctx->event_wqh.lock);
1099 uwq = find_userfault_evt(ctx);
1103 if (uwq->msg.event == UFFD_EVENT_FORK) {
1104 fork_nctx = (struct userfaultfd_ctx *)
1106 uwq->msg.arg.reserved.reserved1;
1107 list_move(&uwq->wq.entry, &fork_event);
1109 * fork_nctx can be freed as soon as
1110 * we drop the lock, unless we take a
1113 userfaultfd_ctx_get(fork_nctx);
1114 spin_unlock(&ctx->event_wqh.lock);
1119 userfaultfd_event_complete(ctx, uwq);
1120 spin_unlock(&ctx->event_wqh.lock);
1124 spin_unlock(&ctx->event_wqh.lock);
1126 if (signal_pending(current)) {
1134 spin_unlock_irq(&ctx->fd_wqh.lock);
1136 spin_lock_irq(&ctx->fd_wqh.lock);
1138 __remove_wait_queue(&ctx->fd_wqh, &wait);
1139 __set_current_state(TASK_RUNNING);
1140 spin_unlock_irq(&ctx->fd_wqh.lock);
1142 if (!ret && msg->event == UFFD_EVENT_FORK) {
1143 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1144 spin_lock_irq(&ctx->event_wqh.lock);
1145 if (!list_empty(&fork_event)) {
1147 * The fork thread didn't abort, so we can
1148 * drop the temporary refcount.
1150 userfaultfd_ctx_put(fork_nctx);
1152 uwq = list_first_entry(&fork_event,
1156 * If fork_event list wasn't empty and in turn
1157 * the event wasn't already released by fork
1158 * (the event is allocated on fork kernel
1159 * stack), put the event back to its place in
1160 * the event_wq. fork_event head will be freed
1161 * as soon as we return so the event cannot
1162 * stay queued there no matter the current
1165 list_del(&uwq->wq.entry);
1166 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1169 * Leave the event in the waitqueue and report
1170 * error to userland if we failed to resolve
1171 * the userfault fork.
1174 userfaultfd_event_complete(ctx, uwq);
1177 * Here the fork thread aborted and the
1178 * refcount from the fork thread on fork_nctx
1179 * has already been released. We still hold
1180 * the reference we took before releasing the
1181 * lock above. If resolve_userfault_fork
1182 * failed we've to drop it because the
1183 * fork_nctx has to be freed in such case. If
1184 * it succeeded we'll hold it because the new
1185 * uffd references it.
1188 userfaultfd_ctx_put(fork_nctx);
1190 spin_unlock_irq(&ctx->event_wqh.lock);
1196 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1197 size_t count, loff_t *ppos)
1199 struct userfaultfd_ctx *ctx = file->private_data;
1200 ssize_t _ret, ret = 0;
1201 struct uffd_msg msg;
1202 int no_wait = file->f_flags & O_NONBLOCK;
1204 if (!userfaultfd_is_initialized(ctx))
1208 if (count < sizeof(msg))
1209 return ret ? ret : -EINVAL;
1210 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1212 return ret ? ret : _ret;
1213 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1214 return ret ? ret : -EFAULT;
1217 count -= sizeof(msg);
1219 * Allow to read more than one fault at time but only
1220 * block if waiting for the very first one.
1222 no_wait = O_NONBLOCK;
1226 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1227 struct userfaultfd_wake_range *range)
1229 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1230 /* wake all in the range and autoremove */
1231 if (waitqueue_active(&ctx->fault_pending_wqh))
1232 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1234 if (waitqueue_active(&ctx->fault_wqh))
1235 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1236 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1239 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1240 struct userfaultfd_wake_range *range)
1246 * To be sure waitqueue_active() is not reordered by the CPU
1247 * before the pagetable update, use an explicit SMP memory
1248 * barrier here. PT lock release or up_read(mmap_sem) still
1249 * have release semantics that can allow the
1250 * waitqueue_active() to be reordered before the pte update.
1255 * Use waitqueue_active because it's very frequent to
1256 * change the address space atomically even if there are no
1257 * userfaults yet. So we take the spinlock only when we're
1258 * sure we've userfaults to wake.
1261 seq = read_seqcount_begin(&ctx->refile_seq);
1262 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1263 waitqueue_active(&ctx->fault_wqh);
1265 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1267 __wake_userfault(ctx, range);
1270 static __always_inline int validate_range(struct mm_struct *mm,
1271 __u64 start, __u64 len)
1273 __u64 task_size = mm->task_size;
1275 if (start & ~PAGE_MASK)
1277 if (len & ~PAGE_MASK)
1281 if (start < mmap_min_addr)
1283 if (start >= task_size)
1285 if (len > task_size - start)
1290 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1292 return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1296 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1299 struct mm_struct *mm = ctx->mm;
1300 struct vm_area_struct *vma, *prev, *cur;
1302 struct uffdio_register uffdio_register;
1303 struct uffdio_register __user *user_uffdio_register;
1304 unsigned long vm_flags, new_flags;
1307 unsigned long start, end, vma_end;
1309 user_uffdio_register = (struct uffdio_register __user *) arg;
1312 if (copy_from_user(&uffdio_register, user_uffdio_register,
1313 sizeof(uffdio_register)-sizeof(__u64)))
1317 if (!uffdio_register.mode)
1319 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1320 UFFDIO_REGISTER_MODE_WP))
1323 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1324 vm_flags |= VM_UFFD_MISSING;
1325 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1326 vm_flags |= VM_UFFD_WP;
1328 * FIXME: remove the below error constraint by
1329 * implementing the wprotect tracking mode.
1335 ret = validate_range(mm, uffdio_register.range.start,
1336 uffdio_register.range.len);
1340 start = uffdio_register.range.start;
1341 end = start + uffdio_register.range.len;
1344 if (!mmget_not_zero(mm))
1347 down_write(&mm->mmap_sem);
1348 if (!mmget_still_valid(mm))
1350 vma = find_vma_prev(mm, start, &prev);
1354 /* check that there's at least one vma in the range */
1356 if (vma->vm_start >= end)
1360 * If the first vma contains huge pages, make sure start address
1361 * is aligned to huge page size.
1363 if (is_vm_hugetlb_page(vma)) {
1364 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1366 if (start & (vma_hpagesize - 1))
1371 * Search for not compatible vmas.
1374 basic_ioctls = false;
1375 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1378 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1379 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1381 /* check not compatible vmas */
1383 if (!vma_can_userfault(cur))
1387 * UFFDIO_COPY will fill file holes even without
1388 * PROT_WRITE. This check enforces that if this is a
1389 * MAP_SHARED, the process has write permission to the backing
1390 * file. If VM_MAYWRITE is set it also enforces that on a
1391 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1392 * F_WRITE_SEAL can be taken until the vma is destroyed.
1395 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1399 * If this vma contains ending address, and huge pages
1402 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1403 end > cur->vm_start) {
1404 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1408 if (end & (vma_hpagesize - 1))
1413 * Check that this vma isn't already owned by a
1414 * different userfaultfd. We can't allow more than one
1415 * userfaultfd to own a single vma simultaneously or we
1416 * wouldn't know which one to deliver the userfaults to.
1419 if (cur->vm_userfaultfd_ctx.ctx &&
1420 cur->vm_userfaultfd_ctx.ctx != ctx)
1424 * Note vmas containing huge pages
1426 if (is_vm_hugetlb_page(cur))
1427 basic_ioctls = true;
1433 if (vma->vm_start < start)
1440 BUG_ON(!vma_can_userfault(vma));
1441 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1442 vma->vm_userfaultfd_ctx.ctx != ctx);
1443 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1446 * Nothing to do: this vma is already registered into this
1447 * userfaultfd and with the right tracking mode too.
1449 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1450 (vma->vm_flags & vm_flags) == vm_flags)
1453 if (vma->vm_start > start)
1454 start = vma->vm_start;
1455 vma_end = min(end, vma->vm_end);
1457 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1458 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1459 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1461 ((struct vm_userfaultfd_ctx){ ctx }));
1466 if (vma->vm_start < start) {
1467 ret = split_vma(mm, vma, start, 1);
1471 if (vma->vm_end > end) {
1472 ret = split_vma(mm, vma, end, 0);
1478 * In the vma_merge() successful mprotect-like case 8:
1479 * the next vma was merged into the current one and
1480 * the current one has not been updated yet.
1482 vma->vm_flags = new_flags;
1483 vma->vm_userfaultfd_ctx.ctx = ctx;
1487 start = vma->vm_end;
1489 } while (vma && vma->vm_start < end);
1491 up_write(&mm->mmap_sem);
1495 * Now that we scanned all vmas we can already tell
1496 * userland which ioctls methods are guaranteed to
1497 * succeed on this range.
1499 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1500 UFFD_API_RANGE_IOCTLS,
1501 &user_uffdio_register->ioctls))
1508 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1511 struct mm_struct *mm = ctx->mm;
1512 struct vm_area_struct *vma, *prev, *cur;
1514 struct uffdio_range uffdio_unregister;
1515 unsigned long new_flags;
1517 unsigned long start, end, vma_end;
1518 const void __user *buf = (void __user *)arg;
1521 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1524 ret = validate_range(mm, uffdio_unregister.start,
1525 uffdio_unregister.len);
1529 start = uffdio_unregister.start;
1530 end = start + uffdio_unregister.len;
1533 if (!mmget_not_zero(mm))
1536 down_write(&mm->mmap_sem);
1537 if (!mmget_still_valid(mm))
1539 vma = find_vma_prev(mm, start, &prev);
1543 /* check that there's at least one vma in the range */
1545 if (vma->vm_start >= end)
1549 * If the first vma contains huge pages, make sure start address
1550 * is aligned to huge page size.
1552 if (is_vm_hugetlb_page(vma)) {
1553 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1555 if (start & (vma_hpagesize - 1))
1560 * Search for not compatible vmas.
1564 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1567 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1568 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1571 * Check not compatible vmas, not strictly required
1572 * here as not compatible vmas cannot have an
1573 * userfaultfd_ctx registered on them, but this
1574 * provides for more strict behavior to notice
1575 * unregistration errors.
1577 if (!vma_can_userfault(cur))
1584 if (vma->vm_start < start)
1591 BUG_ON(!vma_can_userfault(vma));
1594 * Nothing to do: this vma is already registered into this
1595 * userfaultfd and with the right tracking mode too.
1597 if (!vma->vm_userfaultfd_ctx.ctx)
1600 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1602 if (vma->vm_start > start)
1603 start = vma->vm_start;
1604 vma_end = min(end, vma->vm_end);
1606 if (userfaultfd_missing(vma)) {
1608 * Wake any concurrent pending userfault while
1609 * we unregister, so they will not hang
1610 * permanently and it avoids userland to call
1611 * UFFDIO_WAKE explicitly.
1613 struct userfaultfd_wake_range range;
1614 range.start = start;
1615 range.len = vma_end - start;
1616 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1619 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1620 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1621 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1628 if (vma->vm_start < start) {
1629 ret = split_vma(mm, vma, start, 1);
1633 if (vma->vm_end > end) {
1634 ret = split_vma(mm, vma, end, 0);
1640 * In the vma_merge() successful mprotect-like case 8:
1641 * the next vma was merged into the current one and
1642 * the current one has not been updated yet.
1644 vma->vm_flags = new_flags;
1645 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1649 start = vma->vm_end;
1651 } while (vma && vma->vm_start < end);
1653 up_write(&mm->mmap_sem);
1660 * userfaultfd_wake may be used in combination with the
1661 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1663 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1667 struct uffdio_range uffdio_wake;
1668 struct userfaultfd_wake_range range;
1669 const void __user *buf = (void __user *)arg;
1672 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1675 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1679 range.start = uffdio_wake.start;
1680 range.len = uffdio_wake.len;
1683 * len == 0 means wake all and we don't want to wake all here,
1684 * so check it again to be sure.
1686 VM_BUG_ON(!range.len);
1688 wake_userfault(ctx, &range);
1695 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1699 struct uffdio_copy uffdio_copy;
1700 struct uffdio_copy __user *user_uffdio_copy;
1701 struct userfaultfd_wake_range range;
1703 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1706 if (READ_ONCE(ctx->mmap_changing))
1710 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1711 /* don't copy "copy" last field */
1712 sizeof(uffdio_copy)-sizeof(__s64)))
1715 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1719 * double check for wraparound just in case. copy_from_user()
1720 * will later check uffdio_copy.src + uffdio_copy.len to fit
1721 * in the userland range.
1724 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1726 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1728 if (mmget_not_zero(ctx->mm)) {
1729 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1730 uffdio_copy.len, &ctx->mmap_changing);
1735 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1740 /* len == 0 would wake all */
1742 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1743 range.start = uffdio_copy.dst;
1744 wake_userfault(ctx, &range);
1746 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1751 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1755 struct uffdio_zeropage uffdio_zeropage;
1756 struct uffdio_zeropage __user *user_uffdio_zeropage;
1757 struct userfaultfd_wake_range range;
1759 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1762 if (READ_ONCE(ctx->mmap_changing))
1766 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1767 /* don't copy "zeropage" last field */
1768 sizeof(uffdio_zeropage)-sizeof(__s64)))
1771 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1772 uffdio_zeropage.range.len);
1776 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1779 if (mmget_not_zero(ctx->mm)) {
1780 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1781 uffdio_zeropage.range.len,
1782 &ctx->mmap_changing);
1787 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1791 /* len == 0 would wake all */
1794 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1795 range.start = uffdio_zeropage.range.start;
1796 wake_userfault(ctx, &range);
1798 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1803 static inline unsigned int uffd_ctx_features(__u64 user_features)
1806 * For the current set of features the bits just coincide. Set
1807 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1809 return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1813 * userland asks for a certain API version and we return which bits
1814 * and ioctl commands are implemented in this kernel for such API
1815 * version or -EINVAL if unknown.
1817 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1820 struct uffdio_api uffdio_api;
1821 void __user *buf = (void __user *)arg;
1822 unsigned int ctx_features;
1827 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1829 features = uffdio_api.features;
1831 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1834 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1836 /* report all available features and ioctls to userland */
1837 uffdio_api.features = UFFD_API_FEATURES;
1838 uffdio_api.ioctls = UFFD_API_IOCTLS;
1840 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1843 /* only enable the requested features for this uffd context */
1844 ctx_features = uffd_ctx_features(features);
1846 if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
1853 memset(&uffdio_api, 0, sizeof(uffdio_api));
1854 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1859 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1863 struct userfaultfd_ctx *ctx = file->private_data;
1865 if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
1870 ret = userfaultfd_api(ctx, arg);
1872 case UFFDIO_REGISTER:
1873 ret = userfaultfd_register(ctx, arg);
1875 case UFFDIO_UNREGISTER:
1876 ret = userfaultfd_unregister(ctx, arg);
1879 ret = userfaultfd_wake(ctx, arg);
1882 ret = userfaultfd_copy(ctx, arg);
1884 case UFFDIO_ZEROPAGE:
1885 ret = userfaultfd_zeropage(ctx, arg);
1891 #ifdef CONFIG_PROC_FS
1892 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1894 struct userfaultfd_ctx *ctx = f->private_data;
1895 wait_queue_entry_t *wq;
1896 unsigned long pending = 0, total = 0;
1898 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1899 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1903 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1906 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1909 * If more protocols will be added, there will be all shown
1910 * separated by a space. Like this:
1911 * protocols: aa:... bb:...
1913 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1914 pending, total, UFFD_API, ctx->features,
1915 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1919 static const struct file_operations userfaultfd_fops = {
1920 #ifdef CONFIG_PROC_FS
1921 .show_fdinfo = userfaultfd_show_fdinfo,
1923 .release = userfaultfd_release,
1924 .poll = userfaultfd_poll,
1925 .read = userfaultfd_read,
1926 .unlocked_ioctl = userfaultfd_ioctl,
1927 .compat_ioctl = userfaultfd_ioctl,
1928 .llseek = noop_llseek,
1931 static void init_once_userfaultfd_ctx(void *mem)
1933 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1935 init_waitqueue_head(&ctx->fault_pending_wqh);
1936 init_waitqueue_head(&ctx->fault_wqh);
1937 init_waitqueue_head(&ctx->event_wqh);
1938 init_waitqueue_head(&ctx->fd_wqh);
1939 seqcount_init(&ctx->refile_seq);
1942 SYSCALL_DEFINE1(userfaultfd, int, flags)
1944 struct userfaultfd_ctx *ctx;
1947 BUG_ON(!current->mm);
1949 /* Check the UFFD_* constants for consistency. */
1950 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1951 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1953 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1956 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1960 atomic_set(&ctx->refcount, 1);
1963 ctx->released = false;
1964 ctx->mmap_changing = false;
1965 ctx->mm = current->mm;
1966 /* prevent the mm struct to be freed */
1969 fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
1970 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1973 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1978 static int __init userfaultfd_init(void)
1980 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1981 sizeof(struct userfaultfd_ctx),
1983 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1984 init_once_userfaultfd_ctx);
1987 __initcall(userfaultfd_init);