1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
6 * Copyright (C) 2008-2009 Red Hat, Inc.
7 * Copyright (C) 2015 Red Hat, Inc.
9 * Some part derived from fs/eventfd.c (anon inode setup) and
10 * mm/ksm.c (mm hashing).
13 #include <linux/list.h>
14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/mm.h>
18 #include <linux/mm_inline.h>
19 #include <linux/mmu_notifier.h>
20 #include <linux/poll.h>
21 #include <linux/slab.h>
22 #include <linux/seq_file.h>
23 #include <linux/file.h>
24 #include <linux/bug.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h>
30 #include <linux/security.h>
31 #include <linux/hugetlb.h>
32 #include <linux/swapops.h>
33 #include <linux/miscdevice.h>
34 #include <linux/uio.h>
36 static int sysctl_unprivileged_userfaultfd __read_mostly;
39 static struct ctl_table vm_userfaultfd_table[] = {
41 .procname = "unprivileged_userfaultfd",
42 .data = &sysctl_unprivileged_userfaultfd,
43 .maxlen = sizeof(sysctl_unprivileged_userfaultfd),
45 .proc_handler = proc_dointvec_minmax,
46 .extra1 = SYSCTL_ZERO,
52 static struct kmem_cache *userfaultfd_ctx_cachep __ro_after_init;
54 struct userfaultfd_fork_ctx {
55 struct userfaultfd_ctx *orig;
56 struct userfaultfd_ctx *new;
57 struct list_head list;
60 struct userfaultfd_unmap_ctx {
61 struct userfaultfd_ctx *ctx;
64 struct list_head list;
67 struct userfaultfd_wait_queue {
69 wait_queue_entry_t wq;
70 struct userfaultfd_ctx *ctx;
74 struct userfaultfd_wake_range {
79 /* internal indication that UFFD_API ioctl was successfully executed */
80 #define UFFD_FEATURE_INITIALIZED (1u << 31)
82 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
84 return ctx->features & UFFD_FEATURE_INITIALIZED;
87 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx *ctx)
89 return ctx && (ctx->features & UFFD_FEATURE_WP_ASYNC);
93 * Whether WP_UNPOPULATED is enabled on the uffd context. It is only
94 * meaningful when userfaultfd_wp()==true on the vma and when it's
97 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
99 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
104 return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
107 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
110 const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP;
112 vm_flags_reset(vma, flags);
114 * For shared mappings, we want to enable writenotify while
115 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
116 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
118 if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
119 vma_set_page_prot(vma);
122 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
123 int wake_flags, void *key)
125 struct userfaultfd_wake_range *range = key;
127 struct userfaultfd_wait_queue *uwq;
128 unsigned long start, len;
130 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
132 /* len == 0 means wake all */
133 start = range->start;
135 if (len && (start > uwq->msg.arg.pagefault.address ||
136 start + len <= uwq->msg.arg.pagefault.address))
138 WRITE_ONCE(uwq->waken, true);
140 * The Program-Order guarantees provided by the scheduler
141 * ensure uwq->waken is visible before the task is woken.
143 ret = wake_up_state(wq->private, mode);
146 * Wake only once, autoremove behavior.
148 * After the effect of list_del_init is visible to the other
149 * CPUs, the waitqueue may disappear from under us, see the
150 * !list_empty_careful() in handle_userfault().
152 * try_to_wake_up() has an implicit smp_mb(), and the
153 * wq->private is read before calling the extern function
154 * "wake_up_state" (which in turns calls try_to_wake_up).
156 list_del_init(&wq->entry);
163 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
165 * @ctx: [in] Pointer to the userfaultfd context.
167 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
169 refcount_inc(&ctx->refcount);
173 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
175 * @ctx: [in] Pointer to userfaultfd context.
177 * The userfaultfd context reference must have been previously acquired either
178 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
180 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
182 if (refcount_dec_and_test(&ctx->refcount)) {
183 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
184 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
185 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
186 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
187 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
188 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
189 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
190 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
192 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
196 static inline void msg_init(struct uffd_msg *msg)
198 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
200 * Must use memset to zero out the paddings or kernel data is
201 * leaked to userland.
203 memset(msg, 0, sizeof(struct uffd_msg));
206 static inline struct uffd_msg userfault_msg(unsigned long address,
207 unsigned long real_address,
209 unsigned long reason,
210 unsigned int features)
215 msg.event = UFFD_EVENT_PAGEFAULT;
217 msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
218 real_address : address;
221 * These flags indicate why the userfault occurred:
222 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
223 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
224 * - Neither of these flags being set indicates a MISSING fault.
226 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
227 * fault. Otherwise, it was a read fault.
229 if (flags & FAULT_FLAG_WRITE)
230 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
231 if (reason & VM_UFFD_WP)
232 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
233 if (reason & VM_UFFD_MINOR)
234 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
235 if (features & UFFD_FEATURE_THREAD_ID)
236 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
240 #ifdef CONFIG_HUGETLB_PAGE
242 * Same functionality as userfaultfd_must_wait below with modifications for
245 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
246 struct vm_fault *vmf,
247 unsigned long reason)
249 struct vm_area_struct *vma = vmf->vma;
253 assert_fault_locked(vmf);
255 ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
260 pte = huge_ptep_get(ptep);
263 * Lockless access: we're in a wait_event so it's ok if it
264 * changes under us. PTE markers should be handled the same as none
267 if (huge_pte_none_mostly(pte))
269 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
275 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
276 struct vm_fault *vmf,
277 unsigned long reason)
279 return false; /* should never get here */
281 #endif /* CONFIG_HUGETLB_PAGE */
284 * Verify the pagetables are still not ok after having reigstered into
285 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
286 * userfault that has already been resolved, if userfaultfd_read_iter and
287 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
290 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
291 struct vm_fault *vmf,
292 unsigned long reason)
294 struct mm_struct *mm = ctx->mm;
295 unsigned long address = vmf->address;
304 assert_fault_locked(vmf);
306 pgd = pgd_offset(mm, address);
307 if (!pgd_present(*pgd))
309 p4d = p4d_offset(pgd, address);
310 if (!p4d_present(*p4d))
312 pud = pud_offset(p4d, address);
313 if (!pud_present(*pud))
315 pmd = pmd_offset(pud, address);
317 _pmd = pmdp_get_lockless(pmd);
322 if (!pmd_present(_pmd) || pmd_devmap(_pmd))
325 if (pmd_trans_huge(_pmd)) {
326 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
331 pte = pte_offset_map(pmd, address);
337 * Lockless access: we're in a wait_event so it's ok if it
338 * changes under us. PTE markers should be handled the same as none
341 ptent = ptep_get(pte);
342 if (pte_none_mostly(ptent))
344 if (!pte_write(ptent) && (reason & VM_UFFD_WP))
352 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
354 if (flags & FAULT_FLAG_INTERRUPTIBLE)
355 return TASK_INTERRUPTIBLE;
357 if (flags & FAULT_FLAG_KILLABLE)
358 return TASK_KILLABLE;
360 return TASK_UNINTERRUPTIBLE;
364 * The locking rules involved in returning VM_FAULT_RETRY depending on
365 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
366 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
367 * recommendation in __lock_page_or_retry is not an understatement.
369 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
370 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
373 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
374 * set, VM_FAULT_RETRY can still be returned if and only if there are
375 * fatal_signal_pending()s, and the mmap_lock must be released before
378 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
380 struct vm_area_struct *vma = vmf->vma;
381 struct mm_struct *mm = vma->vm_mm;
382 struct userfaultfd_ctx *ctx;
383 struct userfaultfd_wait_queue uwq;
384 vm_fault_t ret = VM_FAULT_SIGBUS;
386 unsigned int blocking_state;
389 * We don't do userfault handling for the final child pid update.
391 * We also don't do userfault handling during
392 * coredumping. hugetlbfs has the special
393 * hugetlb_follow_page_mask() to skip missing pages in the
394 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
395 * the no_page_table() helper in follow_page_mask(), but the
396 * shmem_vm_ops->fault method is invoked even during
397 * coredumping and it ends up here.
399 if (current->flags & (PF_EXITING|PF_DUMPCORE))
402 assert_fault_locked(vmf);
404 ctx = vma->vm_userfaultfd_ctx.ctx;
408 BUG_ON(ctx->mm != mm);
410 /* Any unrecognized flag is a bug. */
411 VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
412 /* 0 or > 1 flags set is a bug; we expect exactly 1. */
413 VM_BUG_ON(!reason || (reason & (reason - 1)));
415 if (ctx->features & UFFD_FEATURE_SIGBUS)
417 if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY))
421 * If it's already released don't get it. This avoids to loop
422 * in __get_user_pages if userfaultfd_release waits on the
423 * caller of handle_userfault to release the mmap_lock.
425 if (unlikely(READ_ONCE(ctx->released))) {
427 * Don't return VM_FAULT_SIGBUS in this case, so a non
428 * cooperative manager can close the uffd after the
429 * last UFFDIO_COPY, without risking to trigger an
430 * involuntary SIGBUS if the process was starting the
431 * userfaultfd while the userfaultfd was still armed
432 * (but after the last UFFDIO_COPY). If the uffd
433 * wasn't already closed when the userfault reached
434 * this point, that would normally be solved by
435 * userfaultfd_must_wait returning 'false'.
437 * If we were to return VM_FAULT_SIGBUS here, the non
438 * cooperative manager would be instead forced to
439 * always call UFFDIO_UNREGISTER before it can safely
442 ret = VM_FAULT_NOPAGE;
447 * Check that we can return VM_FAULT_RETRY.
449 * NOTE: it should become possible to return VM_FAULT_RETRY
450 * even if FAULT_FLAG_TRIED is set without leading to gup()
451 * -EBUSY failures, if the userfaultfd is to be extended for
452 * VM_UFFD_WP tracking and we intend to arm the userfault
453 * without first stopping userland access to the memory. For
454 * VM_UFFD_MISSING userfaults this is enough for now.
456 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
458 * Validate the invariant that nowait must allow retry
459 * to be sure not to return SIGBUS erroneously on
460 * nowait invocations.
462 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
463 #ifdef CONFIG_DEBUG_VM
464 if (printk_ratelimit()) {
466 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
475 * Handle nowait, not much to do other than tell it to retry
478 ret = VM_FAULT_RETRY;
479 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
482 /* take the reference before dropping the mmap_lock */
483 userfaultfd_ctx_get(ctx);
485 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
486 uwq.wq.private = current;
487 uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
488 reason, ctx->features);
492 blocking_state = userfaultfd_get_blocking_state(vmf->flags);
495 * Take the vma lock now, in order to safely call
496 * userfaultfd_huge_must_wait() later. Since acquiring the
497 * (sleepable) vma lock can modify the current task state, that
498 * must be before explicitly calling set_current_state().
500 if (is_vm_hugetlb_page(vma))
501 hugetlb_vma_lock_read(vma);
503 spin_lock_irq(&ctx->fault_pending_wqh.lock);
505 * After the __add_wait_queue the uwq is visible to userland
506 * through poll/read().
508 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
510 * The smp_mb() after __set_current_state prevents the reads
511 * following the spin_unlock to happen before the list_add in
514 set_current_state(blocking_state);
515 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
517 if (!is_vm_hugetlb_page(vma))
518 must_wait = userfaultfd_must_wait(ctx, vmf, reason);
520 must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
521 if (is_vm_hugetlb_page(vma))
522 hugetlb_vma_unlock_read(vma);
523 release_fault_lock(vmf);
525 if (likely(must_wait && !READ_ONCE(ctx->released))) {
526 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
530 __set_current_state(TASK_RUNNING);
533 * Here we race with the list_del; list_add in
534 * userfaultfd_ctx_read(), however because we don't ever run
535 * list_del_init() to refile across the two lists, the prev
536 * and next pointers will never point to self. list_add also
537 * would never let any of the two pointers to point to
538 * self. So list_empty_careful won't risk to see both pointers
539 * pointing to self at any time during the list refile. The
540 * only case where list_del_init() is called is the full
541 * removal in the wake function and there we don't re-list_add
542 * and it's fine not to block on the spinlock. The uwq on this
543 * kernel stack can be released after the list_del_init.
545 if (!list_empty_careful(&uwq.wq.entry)) {
546 spin_lock_irq(&ctx->fault_pending_wqh.lock);
548 * No need of list_del_init(), the uwq on the stack
549 * will be freed shortly anyway.
551 list_del(&uwq.wq.entry);
552 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
556 * ctx may go away after this if the userfault pseudo fd is
559 userfaultfd_ctx_put(ctx);
565 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
566 struct userfaultfd_wait_queue *ewq)
568 struct userfaultfd_ctx *release_new_ctx;
570 if (WARN_ON_ONCE(current->flags & PF_EXITING))
574 init_waitqueue_entry(&ewq->wq, current);
575 release_new_ctx = NULL;
577 spin_lock_irq(&ctx->event_wqh.lock);
579 * After the __add_wait_queue the uwq is visible to userland
580 * through poll/read().
582 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
584 set_current_state(TASK_KILLABLE);
585 if (ewq->msg.event == 0)
587 if (READ_ONCE(ctx->released) ||
588 fatal_signal_pending(current)) {
590 * &ewq->wq may be queued in fork_event, but
591 * __remove_wait_queue ignores the head
592 * parameter. It would be a problem if it
595 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
596 if (ewq->msg.event == UFFD_EVENT_FORK) {
597 struct userfaultfd_ctx *new;
599 new = (struct userfaultfd_ctx *)
601 ewq->msg.arg.reserved.reserved1;
602 release_new_ctx = new;
607 spin_unlock_irq(&ctx->event_wqh.lock);
609 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
612 spin_lock_irq(&ctx->event_wqh.lock);
614 __set_current_state(TASK_RUNNING);
615 spin_unlock_irq(&ctx->event_wqh.lock);
617 if (release_new_ctx) {
618 struct vm_area_struct *vma;
619 struct mm_struct *mm = release_new_ctx->mm;
620 VMA_ITERATOR(vmi, mm, 0);
622 /* the various vma->vm_userfaultfd_ctx still points to it */
624 for_each_vma(vmi, vma) {
625 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
626 vma_start_write(vma);
627 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
628 userfaultfd_set_vm_flags(vma,
629 vma->vm_flags & ~__VM_UFFD_FLAGS);
632 mmap_write_unlock(mm);
634 userfaultfd_ctx_put(release_new_ctx);
638 * ctx may go away after this if the userfault pseudo fd is
642 atomic_dec(&ctx->mmap_changing);
643 VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
644 userfaultfd_ctx_put(ctx);
647 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
648 struct userfaultfd_wait_queue *ewq)
651 wake_up_locked(&ctx->event_wqh);
652 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
655 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
657 struct userfaultfd_ctx *ctx = NULL, *octx;
658 struct userfaultfd_fork_ctx *fctx;
660 octx = vma->vm_userfaultfd_ctx.ctx;
661 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
662 vma_start_write(vma);
663 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
664 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
668 list_for_each_entry(fctx, fcs, list)
669 if (fctx->orig == octx) {
675 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
679 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
685 refcount_set(&ctx->refcount, 1);
686 ctx->flags = octx->flags;
687 ctx->features = octx->features;
688 ctx->released = false;
689 init_rwsem(&ctx->map_changing_lock);
690 atomic_set(&ctx->mmap_changing, 0);
691 ctx->mm = vma->vm_mm;
694 userfaultfd_ctx_get(octx);
695 down_write(&octx->map_changing_lock);
696 atomic_inc(&octx->mmap_changing);
697 up_write(&octx->map_changing_lock);
700 list_add_tail(&fctx->list, fcs);
703 vma->vm_userfaultfd_ctx.ctx = ctx;
707 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
709 struct userfaultfd_ctx *ctx = fctx->orig;
710 struct userfaultfd_wait_queue ewq;
714 ewq.msg.event = UFFD_EVENT_FORK;
715 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
717 userfaultfd_event_wait_completion(ctx, &ewq);
720 void dup_userfaultfd_complete(struct list_head *fcs)
722 struct userfaultfd_fork_ctx *fctx, *n;
724 list_for_each_entry_safe(fctx, n, fcs, list) {
726 list_del(&fctx->list);
731 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
732 struct vm_userfaultfd_ctx *vm_ctx)
734 struct userfaultfd_ctx *ctx;
736 ctx = vma->vm_userfaultfd_ctx.ctx;
741 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
743 userfaultfd_ctx_get(ctx);
744 down_write(&ctx->map_changing_lock);
745 atomic_inc(&ctx->mmap_changing);
746 up_write(&ctx->map_changing_lock);
748 /* Drop uffd context if remap feature not enabled */
749 vma_start_write(vma);
750 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
751 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
755 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
756 unsigned long from, unsigned long to,
759 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
760 struct userfaultfd_wait_queue ewq;
765 if (to & ~PAGE_MASK) {
766 userfaultfd_ctx_put(ctx);
772 ewq.msg.event = UFFD_EVENT_REMAP;
773 ewq.msg.arg.remap.from = from;
774 ewq.msg.arg.remap.to = to;
775 ewq.msg.arg.remap.len = len;
777 userfaultfd_event_wait_completion(ctx, &ewq);
780 bool userfaultfd_remove(struct vm_area_struct *vma,
781 unsigned long start, unsigned long end)
783 struct mm_struct *mm = vma->vm_mm;
784 struct userfaultfd_ctx *ctx;
785 struct userfaultfd_wait_queue ewq;
787 ctx = vma->vm_userfaultfd_ctx.ctx;
788 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
791 userfaultfd_ctx_get(ctx);
792 down_write(&ctx->map_changing_lock);
793 atomic_inc(&ctx->mmap_changing);
794 up_write(&ctx->map_changing_lock);
795 mmap_read_unlock(mm);
799 ewq.msg.event = UFFD_EVENT_REMOVE;
800 ewq.msg.arg.remove.start = start;
801 ewq.msg.arg.remove.end = end;
803 userfaultfd_event_wait_completion(ctx, &ewq);
808 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
809 unsigned long start, unsigned long end)
811 struct userfaultfd_unmap_ctx *unmap_ctx;
813 list_for_each_entry(unmap_ctx, unmaps, list)
814 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
815 unmap_ctx->end == end)
821 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
822 unsigned long end, struct list_head *unmaps)
824 struct userfaultfd_unmap_ctx *unmap_ctx;
825 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
827 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
828 has_unmap_ctx(ctx, unmaps, start, end))
831 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
835 userfaultfd_ctx_get(ctx);
836 down_write(&ctx->map_changing_lock);
837 atomic_inc(&ctx->mmap_changing);
838 up_write(&ctx->map_changing_lock);
839 unmap_ctx->ctx = ctx;
840 unmap_ctx->start = start;
841 unmap_ctx->end = end;
842 list_add_tail(&unmap_ctx->list, unmaps);
847 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
849 struct userfaultfd_unmap_ctx *ctx, *n;
850 struct userfaultfd_wait_queue ewq;
852 list_for_each_entry_safe(ctx, n, uf, list) {
855 ewq.msg.event = UFFD_EVENT_UNMAP;
856 ewq.msg.arg.remove.start = ctx->start;
857 ewq.msg.arg.remove.end = ctx->end;
859 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
861 list_del(&ctx->list);
866 static int userfaultfd_release(struct inode *inode, struct file *file)
868 struct userfaultfd_ctx *ctx = file->private_data;
869 struct mm_struct *mm = ctx->mm;
870 struct vm_area_struct *vma, *prev;
871 /* len == 0 means wake all */
872 struct userfaultfd_wake_range range = { .len = 0, };
873 unsigned long new_flags;
874 VMA_ITERATOR(vmi, mm, 0);
876 WRITE_ONCE(ctx->released, true);
878 if (!mmget_not_zero(mm))
882 * Flush page faults out of all CPUs. NOTE: all page faults
883 * must be retried without returning VM_FAULT_SIGBUS if
884 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
885 * changes while handle_userfault released the mmap_lock. So
886 * it's critical that released is set to true (above), before
887 * taking the mmap_lock for writing.
891 for_each_vma(vmi, vma) {
893 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
894 !!(vma->vm_flags & __VM_UFFD_FLAGS));
895 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
899 /* Reset ptes for the whole vma range if wr-protected */
900 if (userfaultfd_wp(vma))
901 uffd_wp_range(vma, vma->vm_start,
902 vma->vm_end - vma->vm_start, false);
903 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
904 vma = vma_modify_flags_uffd(&vmi, prev, vma, vma->vm_start,
905 vma->vm_end, new_flags,
908 vma_start_write(vma);
909 userfaultfd_set_vm_flags(vma, new_flags);
910 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
914 mmap_write_unlock(mm);
918 * After no new page faults can wait on this fault_*wqh, flush
919 * the last page faults that may have been already waiting on
922 spin_lock_irq(&ctx->fault_pending_wqh.lock);
923 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
924 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
925 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
927 /* Flush pending events that may still wait on event_wqh */
928 wake_up_all(&ctx->event_wqh);
930 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
931 userfaultfd_ctx_put(ctx);
935 /* fault_pending_wqh.lock must be hold by the caller */
936 static inline struct userfaultfd_wait_queue *find_userfault_in(
937 wait_queue_head_t *wqh)
939 wait_queue_entry_t *wq;
940 struct userfaultfd_wait_queue *uwq;
942 lockdep_assert_held(&wqh->lock);
945 if (!waitqueue_active(wqh))
947 /* walk in reverse to provide FIFO behavior to read userfaults */
948 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
949 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
954 static inline struct userfaultfd_wait_queue *find_userfault(
955 struct userfaultfd_ctx *ctx)
957 return find_userfault_in(&ctx->fault_pending_wqh);
960 static inline struct userfaultfd_wait_queue *find_userfault_evt(
961 struct userfaultfd_ctx *ctx)
963 return find_userfault_in(&ctx->event_wqh);
966 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
968 struct userfaultfd_ctx *ctx = file->private_data;
971 poll_wait(file, &ctx->fd_wqh, wait);
973 if (!userfaultfd_is_initialized(ctx))
977 * poll() never guarantees that read won't block.
978 * userfaults can be waken before they're read().
980 if (unlikely(!(file->f_flags & O_NONBLOCK)))
983 * lockless access to see if there are pending faults
984 * __pollwait last action is the add_wait_queue but
985 * the spin_unlock would allow the waitqueue_active to
986 * pass above the actual list_add inside
987 * add_wait_queue critical section. So use a full
988 * memory barrier to serialize the list_add write of
989 * add_wait_queue() with the waitqueue_active read
994 if (waitqueue_active(&ctx->fault_pending_wqh))
996 else if (waitqueue_active(&ctx->event_wqh))
1002 static const struct file_operations userfaultfd_fops;
1004 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
1005 struct inode *inode,
1006 struct uffd_msg *msg)
1010 fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, new,
1011 O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1015 msg->arg.reserved.reserved1 = 0;
1016 msg->arg.fork.ufd = fd;
1020 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1021 struct uffd_msg *msg, struct inode *inode)
1024 DECLARE_WAITQUEUE(wait, current);
1025 struct userfaultfd_wait_queue *uwq;
1027 * Handling fork event requires sleeping operations, so
1028 * we drop the event_wqh lock, then do these ops, then
1029 * lock it back and wake up the waiter. While the lock is
1030 * dropped the ewq may go away so we keep track of it
1033 LIST_HEAD(fork_event);
1034 struct userfaultfd_ctx *fork_nctx = NULL;
1036 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1037 spin_lock_irq(&ctx->fd_wqh.lock);
1038 __add_wait_queue(&ctx->fd_wqh, &wait);
1040 set_current_state(TASK_INTERRUPTIBLE);
1041 spin_lock(&ctx->fault_pending_wqh.lock);
1042 uwq = find_userfault(ctx);
1045 * Use a seqcount to repeat the lockless check
1046 * in wake_userfault() to avoid missing
1047 * wakeups because during the refile both
1048 * waitqueue could become empty if this is the
1051 write_seqcount_begin(&ctx->refile_seq);
1054 * The fault_pending_wqh.lock prevents the uwq
1055 * to disappear from under us.
1057 * Refile this userfault from
1058 * fault_pending_wqh to fault_wqh, it's not
1059 * pending anymore after we read it.
1061 * Use list_del() by hand (as
1062 * userfaultfd_wake_function also uses
1063 * list_del_init() by hand) to be sure nobody
1064 * changes __remove_wait_queue() to use
1065 * list_del_init() in turn breaking the
1066 * !list_empty_careful() check in
1067 * handle_userfault(). The uwq->wq.head list
1068 * must never be empty at any time during the
1069 * refile, or the waitqueue could disappear
1070 * from under us. The "wait_queue_head_t"
1071 * parameter of __remove_wait_queue() is unused
1074 list_del(&uwq->wq.entry);
1075 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1077 write_seqcount_end(&ctx->refile_seq);
1079 /* careful to always initialize msg if ret == 0 */
1081 spin_unlock(&ctx->fault_pending_wqh.lock);
1085 spin_unlock(&ctx->fault_pending_wqh.lock);
1087 spin_lock(&ctx->event_wqh.lock);
1088 uwq = find_userfault_evt(ctx);
1092 if (uwq->msg.event == UFFD_EVENT_FORK) {
1093 fork_nctx = (struct userfaultfd_ctx *)
1095 uwq->msg.arg.reserved.reserved1;
1096 list_move(&uwq->wq.entry, &fork_event);
1098 * fork_nctx can be freed as soon as
1099 * we drop the lock, unless we take a
1102 userfaultfd_ctx_get(fork_nctx);
1103 spin_unlock(&ctx->event_wqh.lock);
1108 userfaultfd_event_complete(ctx, uwq);
1109 spin_unlock(&ctx->event_wqh.lock);
1113 spin_unlock(&ctx->event_wqh.lock);
1115 if (signal_pending(current)) {
1123 spin_unlock_irq(&ctx->fd_wqh.lock);
1125 spin_lock_irq(&ctx->fd_wqh.lock);
1127 __remove_wait_queue(&ctx->fd_wqh, &wait);
1128 __set_current_state(TASK_RUNNING);
1129 spin_unlock_irq(&ctx->fd_wqh.lock);
1131 if (!ret && msg->event == UFFD_EVENT_FORK) {
1132 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1133 spin_lock_irq(&ctx->event_wqh.lock);
1134 if (!list_empty(&fork_event)) {
1136 * The fork thread didn't abort, so we can
1137 * drop the temporary refcount.
1139 userfaultfd_ctx_put(fork_nctx);
1141 uwq = list_first_entry(&fork_event,
1145 * If fork_event list wasn't empty and in turn
1146 * the event wasn't already released by fork
1147 * (the event is allocated on fork kernel
1148 * stack), put the event back to its place in
1149 * the event_wq. fork_event head will be freed
1150 * as soon as we return so the event cannot
1151 * stay queued there no matter the current
1154 list_del(&uwq->wq.entry);
1155 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1158 * Leave the event in the waitqueue and report
1159 * error to userland if we failed to resolve
1160 * the userfault fork.
1163 userfaultfd_event_complete(ctx, uwq);
1166 * Here the fork thread aborted and the
1167 * refcount from the fork thread on fork_nctx
1168 * has already been released. We still hold
1169 * the reference we took before releasing the
1170 * lock above. If resolve_userfault_fork
1171 * failed we've to drop it because the
1172 * fork_nctx has to be freed in such case. If
1173 * it succeeded we'll hold it because the new
1174 * uffd references it.
1177 userfaultfd_ctx_put(fork_nctx);
1179 spin_unlock_irq(&ctx->event_wqh.lock);
1185 static ssize_t userfaultfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
1187 struct file *file = iocb->ki_filp;
1188 struct userfaultfd_ctx *ctx = file->private_data;
1189 ssize_t _ret, ret = 0;
1190 struct uffd_msg msg;
1191 struct inode *inode = file_inode(file);
1194 if (!userfaultfd_is_initialized(ctx))
1197 no_wait = file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT;
1199 if (iov_iter_count(to) < sizeof(msg))
1200 return ret ? ret : -EINVAL;
1201 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1203 return ret ? ret : _ret;
1204 _ret = !copy_to_iter_full(&msg, sizeof(msg), to);
1206 return ret ? ret : -EFAULT;
1209 * Allow to read more than one fault at time but only
1210 * block if waiting for the very first one.
1216 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1217 struct userfaultfd_wake_range *range)
1219 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1220 /* wake all in the range and autoremove */
1221 if (waitqueue_active(&ctx->fault_pending_wqh))
1222 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1224 if (waitqueue_active(&ctx->fault_wqh))
1225 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1226 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1229 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1230 struct userfaultfd_wake_range *range)
1236 * To be sure waitqueue_active() is not reordered by the CPU
1237 * before the pagetable update, use an explicit SMP memory
1238 * barrier here. PT lock release or mmap_read_unlock(mm) still
1239 * have release semantics that can allow the
1240 * waitqueue_active() to be reordered before the pte update.
1245 * Use waitqueue_active because it's very frequent to
1246 * change the address space atomically even if there are no
1247 * userfaults yet. So we take the spinlock only when we're
1248 * sure we've userfaults to wake.
1251 seq = read_seqcount_begin(&ctx->refile_seq);
1252 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1253 waitqueue_active(&ctx->fault_wqh);
1255 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1257 __wake_userfault(ctx, range);
1260 static __always_inline int validate_unaligned_range(
1261 struct mm_struct *mm, __u64 start, __u64 len)
1263 __u64 task_size = mm->task_size;
1265 if (len & ~PAGE_MASK)
1269 if (start < mmap_min_addr)
1271 if (start >= task_size)
1273 if (len > task_size - start)
1275 if (start + len <= start)
1280 static __always_inline int validate_range(struct mm_struct *mm,
1281 __u64 start, __u64 len)
1283 if (start & ~PAGE_MASK)
1286 return validate_unaligned_range(mm, start, len);
1289 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1292 struct mm_struct *mm = ctx->mm;
1293 struct vm_area_struct *vma, *prev, *cur;
1295 struct uffdio_register uffdio_register;
1296 struct uffdio_register __user *user_uffdio_register;
1297 unsigned long vm_flags, new_flags;
1300 unsigned long start, end, vma_end;
1301 struct vma_iterator vmi;
1302 bool wp_async = userfaultfd_wp_async_ctx(ctx);
1304 user_uffdio_register = (struct uffdio_register __user *) arg;
1307 if (copy_from_user(&uffdio_register, user_uffdio_register,
1308 sizeof(uffdio_register)-sizeof(__u64)))
1312 if (!uffdio_register.mode)
1314 if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1317 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1318 vm_flags |= VM_UFFD_MISSING;
1319 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1320 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1323 vm_flags |= VM_UFFD_WP;
1325 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1326 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1329 vm_flags |= VM_UFFD_MINOR;
1332 ret = validate_range(mm, uffdio_register.range.start,
1333 uffdio_register.range.len);
1337 start = uffdio_register.range.start;
1338 end = start + uffdio_register.range.len;
1341 if (!mmget_not_zero(mm))
1345 mmap_write_lock(mm);
1346 vma_iter_init(&vmi, mm, start);
1347 vma = vma_find(&vmi, end);
1352 * If the first vma contains huge pages, make sure start address
1353 * is aligned to huge page size.
1355 if (is_vm_hugetlb_page(vma)) {
1356 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1358 if (start & (vma_hpagesize - 1))
1363 * Search for not compatible vmas.
1366 basic_ioctls = false;
1371 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1372 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1374 /* check not compatible vmas */
1376 if (!vma_can_userfault(cur, vm_flags, wp_async))
1380 * UFFDIO_COPY will fill file holes even without
1381 * PROT_WRITE. This check enforces that if this is a
1382 * MAP_SHARED, the process has write permission to the backing
1383 * file. If VM_MAYWRITE is set it also enforces that on a
1384 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1385 * F_WRITE_SEAL can be taken until the vma is destroyed.
1388 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1392 * If this vma contains ending address, and huge pages
1395 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1396 end > cur->vm_start) {
1397 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1401 if (end & (vma_hpagesize - 1))
1404 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1408 * Check that this vma isn't already owned by a
1409 * different userfaultfd. We can't allow more than one
1410 * userfaultfd to own a single vma simultaneously or we
1411 * wouldn't know which one to deliver the userfaults to.
1414 if (cur->vm_userfaultfd_ctx.ctx &&
1415 cur->vm_userfaultfd_ctx.ctx != ctx)
1419 * Note vmas containing huge pages
1421 if (is_vm_hugetlb_page(cur))
1422 basic_ioctls = true;
1425 } for_each_vma_range(vmi, cur, end);
1428 vma_iter_set(&vmi, start);
1429 prev = vma_prev(&vmi);
1430 if (vma->vm_start < start)
1434 for_each_vma_range(vmi, vma, end) {
1437 BUG_ON(!vma_can_userfault(vma, vm_flags, wp_async));
1438 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1439 vma->vm_userfaultfd_ctx.ctx != ctx);
1440 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1443 * Nothing to do: this vma is already registered into this
1444 * userfaultfd and with the right tracking mode too.
1446 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1447 (vma->vm_flags & vm_flags) == vm_flags)
1450 if (vma->vm_start > start)
1451 start = vma->vm_start;
1452 vma_end = min(end, vma->vm_end);
1454 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1455 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1457 (struct vm_userfaultfd_ctx){ctx});
1464 * In the vma_merge() successful mprotect-like case 8:
1465 * the next vma was merged into the current one and
1466 * the current one has not been updated yet.
1468 vma_start_write(vma);
1469 userfaultfd_set_vm_flags(vma, new_flags);
1470 vma->vm_userfaultfd_ctx.ctx = ctx;
1472 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1473 hugetlb_unshare_all_pmds(vma);
1477 start = vma->vm_end;
1481 mmap_write_unlock(mm);
1486 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1487 UFFD_API_RANGE_IOCTLS;
1490 * Declare the WP ioctl only if the WP mode is
1491 * specified and all checks passed with the range
1493 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1494 ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1496 /* CONTINUE ioctl is only supported for MINOR ranges. */
1497 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1498 ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1501 * Now that we scanned all vmas we can already tell
1502 * userland which ioctls methods are guaranteed to
1503 * succeed on this range.
1505 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1512 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1515 struct mm_struct *mm = ctx->mm;
1516 struct vm_area_struct *vma, *prev, *cur;
1518 struct uffdio_range uffdio_unregister;
1519 unsigned long new_flags;
1521 unsigned long start, end, vma_end;
1522 const void __user *buf = (void __user *)arg;
1523 struct vma_iterator vmi;
1524 bool wp_async = userfaultfd_wp_async_ctx(ctx);
1527 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1530 ret = validate_range(mm, uffdio_unregister.start,
1531 uffdio_unregister.len);
1535 start = uffdio_unregister.start;
1536 end = start + uffdio_unregister.len;
1539 if (!mmget_not_zero(mm))
1542 mmap_write_lock(mm);
1544 vma_iter_init(&vmi, mm, start);
1545 vma = vma_find(&vmi, end);
1550 * If the first vma contains huge pages, make sure start address
1551 * is aligned to huge page size.
1553 if (is_vm_hugetlb_page(vma)) {
1554 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1556 if (start & (vma_hpagesize - 1))
1561 * Search for not compatible vmas.
1568 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1569 !!(cur->vm_flags & __VM_UFFD_FLAGS));
1572 * Check not compatible vmas, not strictly required
1573 * here as not compatible vmas cannot have an
1574 * userfaultfd_ctx registered on them, but this
1575 * provides for more strict behavior to notice
1576 * unregistration errors.
1578 if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
1582 } for_each_vma_range(vmi, cur, end);
1585 vma_iter_set(&vmi, start);
1586 prev = vma_prev(&vmi);
1587 if (vma->vm_start < start)
1591 for_each_vma_range(vmi, vma, end) {
1594 BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
1597 * Nothing to do: this vma is already registered into this
1598 * userfaultfd and with the right tracking mode too.
1600 if (!vma->vm_userfaultfd_ctx.ctx)
1603 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1605 if (vma->vm_start > start)
1606 start = vma->vm_start;
1607 vma_end = min(end, vma->vm_end);
1609 if (userfaultfd_missing(vma)) {
1611 * Wake any concurrent pending userfault while
1612 * we unregister, so they will not hang
1613 * permanently and it avoids userland to call
1614 * UFFDIO_WAKE explicitly.
1616 struct userfaultfd_wake_range range;
1617 range.start = start;
1618 range.len = vma_end - start;
1619 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1622 /* Reset ptes for the whole vma range if wr-protected */
1623 if (userfaultfd_wp(vma))
1624 uffd_wp_range(vma, start, vma_end - start, false);
1626 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1627 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1628 new_flags, NULL_VM_UFFD_CTX);
1635 * In the vma_merge() successful mprotect-like case 8:
1636 * the next vma was merged into the current one and
1637 * the current one has not been updated yet.
1639 vma_start_write(vma);
1640 userfaultfd_set_vm_flags(vma, new_flags);
1641 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1645 start = vma->vm_end;
1649 mmap_write_unlock(mm);
1656 * userfaultfd_wake may be used in combination with the
1657 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1659 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1663 struct uffdio_range uffdio_wake;
1664 struct userfaultfd_wake_range range;
1665 const void __user *buf = (void __user *)arg;
1668 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1671 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1675 range.start = uffdio_wake.start;
1676 range.len = uffdio_wake.len;
1679 * len == 0 means wake all and we don't want to wake all here,
1680 * so check it again to be sure.
1682 VM_BUG_ON(!range.len);
1684 wake_userfault(ctx, &range);
1691 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1695 struct uffdio_copy uffdio_copy;
1696 struct uffdio_copy __user *user_uffdio_copy;
1697 struct userfaultfd_wake_range range;
1698 uffd_flags_t flags = 0;
1700 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1703 if (atomic_read(&ctx->mmap_changing))
1707 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1708 /* don't copy "copy" last field */
1709 sizeof(uffdio_copy)-sizeof(__s64)))
1712 ret = validate_unaligned_range(ctx->mm, uffdio_copy.src,
1716 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1721 if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1723 if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1724 flags |= MFILL_ATOMIC_WP;
1725 if (mmget_not_zero(ctx->mm)) {
1726 ret = mfill_atomic_copy(ctx, uffdio_copy.dst, uffdio_copy.src,
1727 uffdio_copy.len, flags);
1732 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1737 /* len == 0 would wake all */
1739 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1740 range.start = uffdio_copy.dst;
1741 wake_userfault(ctx, &range);
1743 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1748 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1752 struct uffdio_zeropage uffdio_zeropage;
1753 struct uffdio_zeropage __user *user_uffdio_zeropage;
1754 struct userfaultfd_wake_range range;
1756 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1759 if (atomic_read(&ctx->mmap_changing))
1763 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1764 /* don't copy "zeropage" last field */
1765 sizeof(uffdio_zeropage)-sizeof(__s64)))
1768 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1769 uffdio_zeropage.range.len);
1773 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1776 if (mmget_not_zero(ctx->mm)) {
1777 ret = mfill_atomic_zeropage(ctx, uffdio_zeropage.range.start,
1778 uffdio_zeropage.range.len);
1783 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1787 /* len == 0 would wake all */
1790 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1791 range.start = uffdio_zeropage.range.start;
1792 wake_userfault(ctx, &range);
1794 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1799 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1803 struct uffdio_writeprotect uffdio_wp;
1804 struct uffdio_writeprotect __user *user_uffdio_wp;
1805 struct userfaultfd_wake_range range;
1806 bool mode_wp, mode_dontwake;
1808 if (atomic_read(&ctx->mmap_changing))
1811 user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1813 if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1814 sizeof(struct uffdio_writeprotect)))
1817 ret = validate_range(ctx->mm, uffdio_wp.range.start,
1818 uffdio_wp.range.len);
1822 if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1823 UFFDIO_WRITEPROTECT_MODE_WP))
1826 mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1827 mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1829 if (mode_wp && mode_dontwake)
1832 if (mmget_not_zero(ctx->mm)) {
1833 ret = mwriteprotect_range(ctx, uffdio_wp.range.start,
1834 uffdio_wp.range.len, mode_wp);
1843 if (!mode_wp && !mode_dontwake) {
1844 range.start = uffdio_wp.range.start;
1845 range.len = uffdio_wp.range.len;
1846 wake_userfault(ctx, &range);
1851 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1854 struct uffdio_continue uffdio_continue;
1855 struct uffdio_continue __user *user_uffdio_continue;
1856 struct userfaultfd_wake_range range;
1857 uffd_flags_t flags = 0;
1859 user_uffdio_continue = (struct uffdio_continue __user *)arg;
1862 if (atomic_read(&ctx->mmap_changing))
1866 if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1867 /* don't copy the output fields */
1868 sizeof(uffdio_continue) - (sizeof(__s64))))
1871 ret = validate_range(ctx->mm, uffdio_continue.range.start,
1872 uffdio_continue.range.len);
1877 if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1878 UFFDIO_CONTINUE_MODE_WP))
1880 if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1881 flags |= MFILL_ATOMIC_WP;
1883 if (mmget_not_zero(ctx->mm)) {
1884 ret = mfill_atomic_continue(ctx, uffdio_continue.range.start,
1885 uffdio_continue.range.len, flags);
1891 if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1896 /* len == 0 would wake all */
1899 if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1900 range.start = uffdio_continue.range.start;
1901 wake_userfault(ctx, &range);
1903 ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1909 static inline int userfaultfd_poison(struct userfaultfd_ctx *ctx, unsigned long arg)
1912 struct uffdio_poison uffdio_poison;
1913 struct uffdio_poison __user *user_uffdio_poison;
1914 struct userfaultfd_wake_range range;
1916 user_uffdio_poison = (struct uffdio_poison __user *)arg;
1919 if (atomic_read(&ctx->mmap_changing))
1923 if (copy_from_user(&uffdio_poison, user_uffdio_poison,
1924 /* don't copy the output fields */
1925 sizeof(uffdio_poison) - (sizeof(__s64))))
1928 ret = validate_range(ctx->mm, uffdio_poison.range.start,
1929 uffdio_poison.range.len);
1934 if (uffdio_poison.mode & ~UFFDIO_POISON_MODE_DONTWAKE)
1937 if (mmget_not_zero(ctx->mm)) {
1938 ret = mfill_atomic_poison(ctx, uffdio_poison.range.start,
1939 uffdio_poison.range.len, 0);
1945 if (unlikely(put_user(ret, &user_uffdio_poison->updated)))
1950 /* len == 0 would wake all */
1953 if (!(uffdio_poison.mode & UFFDIO_POISON_MODE_DONTWAKE)) {
1954 range.start = uffdio_poison.range.start;
1955 wake_userfault(ctx, &range);
1957 ret = range.len == uffdio_poison.range.len ? 0 : -EAGAIN;
1963 bool userfaultfd_wp_async(struct vm_area_struct *vma)
1965 return userfaultfd_wp_async_ctx(vma->vm_userfaultfd_ctx.ctx);
1968 static inline unsigned int uffd_ctx_features(__u64 user_features)
1971 * For the current set of features the bits just coincide. Set
1972 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1974 return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1977 static int userfaultfd_move(struct userfaultfd_ctx *ctx,
1981 struct uffdio_move uffdio_move;
1982 struct uffdio_move __user *user_uffdio_move;
1983 struct userfaultfd_wake_range range;
1984 struct mm_struct *mm = ctx->mm;
1986 user_uffdio_move = (struct uffdio_move __user *) arg;
1988 if (atomic_read(&ctx->mmap_changing))
1991 if (copy_from_user(&uffdio_move, user_uffdio_move,
1992 /* don't copy "move" last field */
1993 sizeof(uffdio_move)-sizeof(__s64)))
1996 /* Do not allow cross-mm moves. */
1997 if (mm != current->mm)
2000 ret = validate_range(mm, uffdio_move.dst, uffdio_move.len);
2004 ret = validate_range(mm, uffdio_move.src, uffdio_move.len);
2008 if (uffdio_move.mode & ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES|
2009 UFFDIO_MOVE_MODE_DONTWAKE))
2012 if (mmget_not_zero(mm)) {
2013 ret = move_pages(ctx, uffdio_move.dst, uffdio_move.src,
2014 uffdio_move.len, uffdio_move.mode);
2020 if (unlikely(put_user(ret, &user_uffdio_move->move)))
2025 /* len == 0 would wake all */
2028 if (!(uffdio_move.mode & UFFDIO_MOVE_MODE_DONTWAKE)) {
2029 range.start = uffdio_move.dst;
2030 wake_userfault(ctx, &range);
2032 ret = range.len == uffdio_move.len ? 0 : -EAGAIN;
2039 * userland asks for a certain API version and we return which bits
2040 * and ioctl commands are implemented in this kernel for such API
2041 * version or -EINVAL if unknown.
2043 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
2046 struct uffdio_api uffdio_api;
2047 void __user *buf = (void __user *)arg;
2048 unsigned int ctx_features;
2053 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
2055 features = uffdio_api.features;
2057 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
2060 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
2063 /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2064 if (features & UFFD_FEATURE_WP_ASYNC)
2065 features |= UFFD_FEATURE_WP_UNPOPULATED;
2067 /* report all available features and ioctls to userland */
2068 uffdio_api.features = UFFD_API_FEATURES;
2069 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2070 uffdio_api.features &=
2071 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
2073 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2074 uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
2076 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2077 uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
2078 uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
2079 uffdio_api.features &= ~UFFD_FEATURE_WP_ASYNC;
2081 uffdio_api.ioctls = UFFD_API_IOCTLS;
2083 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2086 /* only enable the requested features for this uffd context */
2087 ctx_features = uffd_ctx_features(features);
2089 if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
2096 memset(&uffdio_api, 0, sizeof(uffdio_api));
2097 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2102 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2106 struct userfaultfd_ctx *ctx = file->private_data;
2108 if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2113 ret = userfaultfd_api(ctx, arg);
2115 case UFFDIO_REGISTER:
2116 ret = userfaultfd_register(ctx, arg);
2118 case UFFDIO_UNREGISTER:
2119 ret = userfaultfd_unregister(ctx, arg);
2122 ret = userfaultfd_wake(ctx, arg);
2125 ret = userfaultfd_copy(ctx, arg);
2127 case UFFDIO_ZEROPAGE:
2128 ret = userfaultfd_zeropage(ctx, arg);
2131 ret = userfaultfd_move(ctx, arg);
2133 case UFFDIO_WRITEPROTECT:
2134 ret = userfaultfd_writeprotect(ctx, arg);
2136 case UFFDIO_CONTINUE:
2137 ret = userfaultfd_continue(ctx, arg);
2140 ret = userfaultfd_poison(ctx, arg);
2146 #ifdef CONFIG_PROC_FS
2147 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2149 struct userfaultfd_ctx *ctx = f->private_data;
2150 wait_queue_entry_t *wq;
2151 unsigned long pending = 0, total = 0;
2153 spin_lock_irq(&ctx->fault_pending_wqh.lock);
2154 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2158 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2161 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2164 * If more protocols will be added, there will be all shown
2165 * separated by a space. Like this:
2166 * protocols: aa:... bb:...
2168 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2169 pending, total, UFFD_API, ctx->features,
2170 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2174 static const struct file_operations userfaultfd_fops = {
2175 #ifdef CONFIG_PROC_FS
2176 .show_fdinfo = userfaultfd_show_fdinfo,
2178 .release = userfaultfd_release,
2179 .poll = userfaultfd_poll,
2180 .read_iter = userfaultfd_read_iter,
2181 .unlocked_ioctl = userfaultfd_ioctl,
2182 .compat_ioctl = compat_ptr_ioctl,
2183 .llseek = noop_llseek,
2186 static void init_once_userfaultfd_ctx(void *mem)
2188 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2190 init_waitqueue_head(&ctx->fault_pending_wqh);
2191 init_waitqueue_head(&ctx->fault_wqh);
2192 init_waitqueue_head(&ctx->event_wqh);
2193 init_waitqueue_head(&ctx->fd_wqh);
2194 seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2197 static int new_userfaultfd(int flags)
2199 struct userfaultfd_ctx *ctx;
2203 BUG_ON(!current->mm);
2205 /* Check the UFFD_* constants for consistency. */
2206 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2207 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2208 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2210 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2213 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2217 refcount_set(&ctx->refcount, 1);
2220 ctx->released = false;
2221 init_rwsem(&ctx->map_changing_lock);
2222 atomic_set(&ctx->mmap_changing, 0);
2223 ctx->mm = current->mm;
2225 fd = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
2229 /* Create a new inode so that the LSM can block the creation. */
2230 file = anon_inode_create_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
2231 O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2237 /* prevent the mm struct to be freed */
2239 file->f_mode |= FMODE_NOWAIT;
2240 fd_install(fd, file);
2243 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2247 static inline bool userfaultfd_syscall_allowed(int flags)
2249 /* Userspace-only page faults are always allowed */
2250 if (flags & UFFD_USER_MODE_ONLY)
2254 * The user is requesting a userfaultfd which can handle kernel faults.
2255 * Privileged users are always allowed to do this.
2257 if (capable(CAP_SYS_PTRACE))
2260 /* Otherwise, access to kernel fault handling is sysctl controlled. */
2261 return sysctl_unprivileged_userfaultfd;
2264 SYSCALL_DEFINE1(userfaultfd, int, flags)
2266 if (!userfaultfd_syscall_allowed(flags))
2269 return new_userfaultfd(flags);
2272 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2274 if (cmd != USERFAULTFD_IOC_NEW)
2277 return new_userfaultfd(flags);
2280 static const struct file_operations userfaultfd_dev_fops = {
2281 .unlocked_ioctl = userfaultfd_dev_ioctl,
2282 .compat_ioctl = userfaultfd_dev_ioctl,
2283 .owner = THIS_MODULE,
2284 .llseek = noop_llseek,
2287 static struct miscdevice userfaultfd_misc = {
2288 .minor = MISC_DYNAMIC_MINOR,
2289 .name = "userfaultfd",
2290 .fops = &userfaultfd_dev_fops
2293 static int __init userfaultfd_init(void)
2297 ret = misc_register(&userfaultfd_misc);
2301 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2302 sizeof(struct userfaultfd_ctx),
2304 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2305 init_once_userfaultfd_ctx);
2306 #ifdef CONFIG_SYSCTL
2307 register_sysctl_init("vm", vm_userfaultfd_table);
2311 __initcall(userfaultfd_init);