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/mmu_notifier.h>
19 #include <linux/poll.h>
20 #include <linux/slab.h>
21 #include <linux/seq_file.h>
22 #include <linux/file.h>
23 #include <linux/bug.h>
24 #include <linux/anon_inodes.h>
25 #include <linux/syscalls.h>
26 #include <linux/userfaultfd_k.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ioctl.h>
29 #include <linux/security.h>
30 #include <linux/hugetlb.h>
32 int sysctl_unprivileged_userfaultfd __read_mostly;
34 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
36 enum userfaultfd_state {
42 * Start with fault_pending_wqh and fault_wqh so they're more likely
43 * to be in the same cacheline.
47 * fault_pending_wqh.lock
51 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
52 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
53 * also taken in IRQ context.
55 struct userfaultfd_ctx {
56 /* waitqueue head for the pending (i.e. not read) userfaults */
57 wait_queue_head_t fault_pending_wqh;
58 /* waitqueue head for the userfaults */
59 wait_queue_head_t fault_wqh;
60 /* waitqueue head for the pseudo fd to wakeup poll/read */
61 wait_queue_head_t fd_wqh;
62 /* waitqueue head for events */
63 wait_queue_head_t event_wqh;
64 /* a refile sequence protected by fault_pending_wqh lock */
65 seqcount_spinlock_t refile_seq;
66 /* pseudo fd refcounting */
68 /* userfaultfd syscall flags */
70 /* features requested from the userspace */
71 unsigned int features;
73 enum userfaultfd_state state;
76 /* memory mappings are changing because of non-cooperative event */
78 /* mm with one ore more vmas attached to this userfaultfd_ctx */
82 struct userfaultfd_fork_ctx {
83 struct userfaultfd_ctx *orig;
84 struct userfaultfd_ctx *new;
85 struct list_head list;
88 struct userfaultfd_unmap_ctx {
89 struct userfaultfd_ctx *ctx;
92 struct list_head list;
95 struct userfaultfd_wait_queue {
97 wait_queue_entry_t wq;
98 struct userfaultfd_ctx *ctx;
102 struct userfaultfd_wake_range {
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 refcount_inc(&ctx->refcount);
158 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
160 * @ctx: [in] Pointer to userfaultfd context.
162 * The userfaultfd context reference must have been previously acquired either
163 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
165 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
167 if (refcount_dec_and_test(&ctx->refcount)) {
168 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
169 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
170 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
171 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
172 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
173 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
174 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
175 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
177 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
181 static inline void msg_init(struct uffd_msg *msg)
183 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
185 * Must use memset to zero out the paddings or kernel data is
186 * leaked to userland.
188 memset(msg, 0, sizeof(struct uffd_msg));
191 static inline struct uffd_msg userfault_msg(unsigned long address,
193 unsigned long reason,
194 unsigned int features)
198 msg.event = UFFD_EVENT_PAGEFAULT;
199 msg.arg.pagefault.address = address;
200 if (flags & FAULT_FLAG_WRITE)
202 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
203 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
204 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
205 * was a read fault, otherwise if set it means it's
208 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
209 if (reason & VM_UFFD_WP)
211 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
212 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
213 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
214 * a missing fault, otherwise if set it means it's a
215 * write protect fault.
217 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
218 if (features & UFFD_FEATURE_THREAD_ID)
219 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
223 #ifdef CONFIG_HUGETLB_PAGE
225 * Same functionality as userfaultfd_must_wait below with modifications for
228 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
229 struct vm_area_struct *vma,
230 unsigned long address,
232 unsigned long reason)
234 struct mm_struct *mm = ctx->mm;
238 mmap_assert_locked(mm);
240 ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
246 pte = huge_ptep_get(ptep);
249 * Lockless access: we're in a wait_event so it's ok if it
252 if (huge_pte_none(pte))
254 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
260 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
261 struct vm_area_struct *vma,
262 unsigned long address,
264 unsigned long reason)
266 return false; /* should never get here */
268 #endif /* CONFIG_HUGETLB_PAGE */
271 * Verify the pagetables are still not ok after having reigstered into
272 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
273 * userfault that has already been resolved, if userfaultfd_read and
274 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
277 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
278 unsigned long address,
280 unsigned long reason)
282 struct mm_struct *mm = ctx->mm;
290 mmap_assert_locked(mm);
292 pgd = pgd_offset(mm, address);
293 if (!pgd_present(*pgd))
295 p4d = p4d_offset(pgd, address);
296 if (!p4d_present(*p4d))
298 pud = pud_offset(p4d, address);
299 if (!pud_present(*pud))
301 pmd = pmd_offset(pud, address);
303 * READ_ONCE must function as a barrier with narrower scope
304 * and it must be equivalent to:
305 * _pmd = *pmd; barrier();
307 * This is to deal with the instability (as in
308 * pmd_trans_unstable) of the pmd.
310 _pmd = READ_ONCE(*pmd);
315 if (!pmd_present(_pmd))
318 if (pmd_trans_huge(_pmd)) {
319 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
325 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
326 * and use the standard pte_offset_map() instead of parsing _pmd.
328 pte = pte_offset_map(pmd, address);
330 * Lockless access: we're in a wait_event so it's ok if it
335 if (!pte_write(*pte) && (reason & VM_UFFD_WP))
343 static inline long userfaultfd_get_blocking_state(unsigned int flags)
345 if (flags & FAULT_FLAG_INTERRUPTIBLE)
346 return TASK_INTERRUPTIBLE;
348 if (flags & FAULT_FLAG_KILLABLE)
349 return TASK_KILLABLE;
351 return TASK_UNINTERRUPTIBLE;
355 * The locking rules involved in returning VM_FAULT_RETRY depending on
356 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
357 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
358 * recommendation in __lock_page_or_retry is not an understatement.
360 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
361 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
364 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
365 * set, VM_FAULT_RETRY can still be returned if and only if there are
366 * fatal_signal_pending()s, and the mmap_lock must be released before
369 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
371 struct mm_struct *mm = vmf->vma->vm_mm;
372 struct userfaultfd_ctx *ctx;
373 struct userfaultfd_wait_queue uwq;
374 vm_fault_t ret = VM_FAULT_SIGBUS;
379 * We don't do userfault handling for the final child pid update.
381 * We also don't do userfault handling during
382 * coredumping. hugetlbfs has the special
383 * follow_hugetlb_page() to skip missing pages in the
384 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
385 * the no_page_table() helper in follow_page_mask(), but the
386 * shmem_vm_ops->fault method is invoked even during
387 * coredumping without mmap_lock and it ends up here.
389 if (current->flags & (PF_EXITING|PF_DUMPCORE))
393 * Coredumping runs without mmap_lock so we can only check that
394 * the mmap_lock is held, if PF_DUMPCORE was not set.
396 mmap_assert_locked(mm);
398 ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
402 BUG_ON(ctx->mm != mm);
404 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
405 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
407 if (ctx->features & UFFD_FEATURE_SIGBUS)
409 if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
410 ctx->flags & UFFD_USER_MODE_ONLY) {
411 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
412 "sysctl knob to 1 if kernel faults must be handled "
413 "without obtaining CAP_SYS_PTRACE capability\n");
418 * If it's already released don't get it. This avoids to loop
419 * in __get_user_pages if userfaultfd_release waits on the
420 * caller of handle_userfault to release the mmap_lock.
422 if (unlikely(READ_ONCE(ctx->released))) {
424 * Don't return VM_FAULT_SIGBUS in this case, so a non
425 * cooperative manager can close the uffd after the
426 * last UFFDIO_COPY, without risking to trigger an
427 * involuntary SIGBUS if the process was starting the
428 * userfaultfd while the userfaultfd was still armed
429 * (but after the last UFFDIO_COPY). If the uffd
430 * wasn't already closed when the userfault reached
431 * this point, that would normally be solved by
432 * userfaultfd_must_wait returning 'false'.
434 * If we were to return VM_FAULT_SIGBUS here, the non
435 * cooperative manager would be instead forced to
436 * always call UFFDIO_UNREGISTER before it can safely
439 ret = VM_FAULT_NOPAGE;
444 * Check that we can return VM_FAULT_RETRY.
446 * NOTE: it should become possible to return VM_FAULT_RETRY
447 * even if FAULT_FLAG_TRIED is set without leading to gup()
448 * -EBUSY failures, if the userfaultfd is to be extended for
449 * VM_UFFD_WP tracking and we intend to arm the userfault
450 * without first stopping userland access to the memory. For
451 * VM_UFFD_MISSING userfaults this is enough for now.
453 if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
455 * Validate the invariant that nowait must allow retry
456 * to be sure not to return SIGBUS erroneously on
457 * nowait invocations.
459 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
460 #ifdef CONFIG_DEBUG_VM
461 if (printk_ratelimit()) {
463 "FAULT_FLAG_ALLOW_RETRY missing %x\n",
472 * Handle nowait, not much to do other than tell it to retry
475 ret = VM_FAULT_RETRY;
476 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
479 /* take the reference before dropping the mmap_lock */
480 userfaultfd_ctx_get(ctx);
482 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
483 uwq.wq.private = current;
484 uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
489 blocking_state = userfaultfd_get_blocking_state(vmf->flags);
491 spin_lock_irq(&ctx->fault_pending_wqh.lock);
493 * After the __add_wait_queue the uwq is visible to userland
494 * through poll/read().
496 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
498 * The smp_mb() after __set_current_state prevents the reads
499 * following the spin_unlock to happen before the list_add in
502 set_current_state(blocking_state);
503 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
505 if (!is_vm_hugetlb_page(vmf->vma))
506 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
509 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
512 mmap_read_unlock(mm);
514 if (likely(must_wait && !READ_ONCE(ctx->released))) {
515 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
519 __set_current_state(TASK_RUNNING);
522 * Here we race with the list_del; list_add in
523 * userfaultfd_ctx_read(), however because we don't ever run
524 * list_del_init() to refile across the two lists, the prev
525 * and next pointers will never point to self. list_add also
526 * would never let any of the two pointers to point to
527 * self. So list_empty_careful won't risk to see both pointers
528 * pointing to self at any time during the list refile. The
529 * only case where list_del_init() is called is the full
530 * removal in the wake function and there we don't re-list_add
531 * and it's fine not to block on the spinlock. The uwq on this
532 * kernel stack can be released after the list_del_init.
534 if (!list_empty_careful(&uwq.wq.entry)) {
535 spin_lock_irq(&ctx->fault_pending_wqh.lock);
537 * No need of list_del_init(), the uwq on the stack
538 * will be freed shortly anyway.
540 list_del(&uwq.wq.entry);
541 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
545 * ctx may go away after this if the userfault pseudo fd is
548 userfaultfd_ctx_put(ctx);
554 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
555 struct userfaultfd_wait_queue *ewq)
557 struct userfaultfd_ctx *release_new_ctx;
559 if (WARN_ON_ONCE(current->flags & PF_EXITING))
563 init_waitqueue_entry(&ewq->wq, current);
564 release_new_ctx = NULL;
566 spin_lock_irq(&ctx->event_wqh.lock);
568 * After the __add_wait_queue the uwq is visible to userland
569 * through poll/read().
571 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
573 set_current_state(TASK_KILLABLE);
574 if (ewq->msg.event == 0)
576 if (READ_ONCE(ctx->released) ||
577 fatal_signal_pending(current)) {
579 * &ewq->wq may be queued in fork_event, but
580 * __remove_wait_queue ignores the head
581 * parameter. It would be a problem if it
584 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
585 if (ewq->msg.event == UFFD_EVENT_FORK) {
586 struct userfaultfd_ctx *new;
588 new = (struct userfaultfd_ctx *)
590 ewq->msg.arg.reserved.reserved1;
591 release_new_ctx = new;
596 spin_unlock_irq(&ctx->event_wqh.lock);
598 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
601 spin_lock_irq(&ctx->event_wqh.lock);
603 __set_current_state(TASK_RUNNING);
604 spin_unlock_irq(&ctx->event_wqh.lock);
606 if (release_new_ctx) {
607 struct vm_area_struct *vma;
608 struct mm_struct *mm = release_new_ctx->mm;
610 /* the various vma->vm_userfaultfd_ctx still points to it */
612 for (vma = mm->mmap; vma; vma = vma->vm_next)
613 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
614 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
615 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
617 mmap_write_unlock(mm);
619 userfaultfd_ctx_put(release_new_ctx);
623 * ctx may go away after this if the userfault pseudo fd is
627 WRITE_ONCE(ctx->mmap_changing, false);
628 userfaultfd_ctx_put(ctx);
631 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
632 struct userfaultfd_wait_queue *ewq)
635 wake_up_locked(&ctx->event_wqh);
636 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
639 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
641 struct userfaultfd_ctx *ctx = NULL, *octx;
642 struct userfaultfd_fork_ctx *fctx;
644 octx = vma->vm_userfaultfd_ctx.ctx;
645 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
646 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
647 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
651 list_for_each_entry(fctx, fcs, list)
652 if (fctx->orig == octx) {
658 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
662 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
668 refcount_set(&ctx->refcount, 1);
669 ctx->flags = octx->flags;
670 ctx->state = UFFD_STATE_RUNNING;
671 ctx->features = octx->features;
672 ctx->released = false;
673 ctx->mmap_changing = false;
674 ctx->mm = vma->vm_mm;
677 userfaultfd_ctx_get(octx);
678 WRITE_ONCE(octx->mmap_changing, true);
681 list_add_tail(&fctx->list, fcs);
684 vma->vm_userfaultfd_ctx.ctx = ctx;
688 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
690 struct userfaultfd_ctx *ctx = fctx->orig;
691 struct userfaultfd_wait_queue ewq;
695 ewq.msg.event = UFFD_EVENT_FORK;
696 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
698 userfaultfd_event_wait_completion(ctx, &ewq);
701 void dup_userfaultfd_complete(struct list_head *fcs)
703 struct userfaultfd_fork_ctx *fctx, *n;
705 list_for_each_entry_safe(fctx, n, fcs, list) {
707 list_del(&fctx->list);
712 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
713 struct vm_userfaultfd_ctx *vm_ctx)
715 struct userfaultfd_ctx *ctx;
717 ctx = vma->vm_userfaultfd_ctx.ctx;
722 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
724 userfaultfd_ctx_get(ctx);
725 WRITE_ONCE(ctx->mmap_changing, true);
727 /* Drop uffd context if remap feature not enabled */
728 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
729 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
733 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
734 unsigned long from, unsigned long to,
737 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
738 struct userfaultfd_wait_queue ewq;
743 if (to & ~PAGE_MASK) {
744 userfaultfd_ctx_put(ctx);
750 ewq.msg.event = UFFD_EVENT_REMAP;
751 ewq.msg.arg.remap.from = from;
752 ewq.msg.arg.remap.to = to;
753 ewq.msg.arg.remap.len = len;
755 userfaultfd_event_wait_completion(ctx, &ewq);
758 bool userfaultfd_remove(struct vm_area_struct *vma,
759 unsigned long start, unsigned long end)
761 struct mm_struct *mm = vma->vm_mm;
762 struct userfaultfd_ctx *ctx;
763 struct userfaultfd_wait_queue ewq;
765 ctx = vma->vm_userfaultfd_ctx.ctx;
766 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
769 userfaultfd_ctx_get(ctx);
770 WRITE_ONCE(ctx->mmap_changing, true);
771 mmap_read_unlock(mm);
775 ewq.msg.event = UFFD_EVENT_REMOVE;
776 ewq.msg.arg.remove.start = start;
777 ewq.msg.arg.remove.end = end;
779 userfaultfd_event_wait_completion(ctx, &ewq);
784 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
785 unsigned long start, unsigned long end)
787 struct userfaultfd_unmap_ctx *unmap_ctx;
789 list_for_each_entry(unmap_ctx, unmaps, list)
790 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
791 unmap_ctx->end == end)
797 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
798 unsigned long start, unsigned long end,
799 struct list_head *unmaps)
801 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
802 struct userfaultfd_unmap_ctx *unmap_ctx;
803 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
805 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
806 has_unmap_ctx(ctx, unmaps, start, end))
809 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
813 userfaultfd_ctx_get(ctx);
814 WRITE_ONCE(ctx->mmap_changing, true);
815 unmap_ctx->ctx = ctx;
816 unmap_ctx->start = start;
817 unmap_ctx->end = end;
818 list_add_tail(&unmap_ctx->list, unmaps);
824 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
826 struct userfaultfd_unmap_ctx *ctx, *n;
827 struct userfaultfd_wait_queue ewq;
829 list_for_each_entry_safe(ctx, n, uf, list) {
832 ewq.msg.event = UFFD_EVENT_UNMAP;
833 ewq.msg.arg.remove.start = ctx->start;
834 ewq.msg.arg.remove.end = ctx->end;
836 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
838 list_del(&ctx->list);
843 static int userfaultfd_release(struct inode *inode, struct file *file)
845 struct userfaultfd_ctx *ctx = file->private_data;
846 struct mm_struct *mm = ctx->mm;
847 struct vm_area_struct *vma, *prev;
848 /* len == 0 means wake all */
849 struct userfaultfd_wake_range range = { .len = 0, };
850 unsigned long new_flags;
852 WRITE_ONCE(ctx->released, true);
854 if (!mmget_not_zero(mm))
858 * Flush page faults out of all CPUs. NOTE: all page faults
859 * must be retried without returning VM_FAULT_SIGBUS if
860 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
861 * changes while handle_userfault released the mmap_lock. So
862 * it's critical that released is set to true (above), before
863 * taking the mmap_lock for writing.
867 for (vma = mm->mmap; vma; vma = vma->vm_next) {
869 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
870 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
871 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
875 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
876 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
877 new_flags, vma->anon_vma,
878 vma->vm_file, vma->vm_pgoff,
885 vma->vm_flags = new_flags;
886 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
888 mmap_write_unlock(mm);
892 * After no new page faults can wait on this fault_*wqh, flush
893 * the last page faults that may have been already waiting on
896 spin_lock_irq(&ctx->fault_pending_wqh.lock);
897 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
898 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
899 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
901 /* Flush pending events that may still wait on event_wqh */
902 wake_up_all(&ctx->event_wqh);
904 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
905 userfaultfd_ctx_put(ctx);
909 /* fault_pending_wqh.lock must be hold by the caller */
910 static inline struct userfaultfd_wait_queue *find_userfault_in(
911 wait_queue_head_t *wqh)
913 wait_queue_entry_t *wq;
914 struct userfaultfd_wait_queue *uwq;
916 lockdep_assert_held(&wqh->lock);
919 if (!waitqueue_active(wqh))
921 /* walk in reverse to provide FIFO behavior to read userfaults */
922 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
923 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
928 static inline struct userfaultfd_wait_queue *find_userfault(
929 struct userfaultfd_ctx *ctx)
931 return find_userfault_in(&ctx->fault_pending_wqh);
934 static inline struct userfaultfd_wait_queue *find_userfault_evt(
935 struct userfaultfd_ctx *ctx)
937 return find_userfault_in(&ctx->event_wqh);
940 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
942 struct userfaultfd_ctx *ctx = file->private_data;
945 poll_wait(file, &ctx->fd_wqh, wait);
947 switch (ctx->state) {
948 case UFFD_STATE_WAIT_API:
950 case UFFD_STATE_RUNNING:
952 * poll() never guarantees that read won't block.
953 * userfaults can be waken before they're read().
955 if (unlikely(!(file->f_flags & O_NONBLOCK)))
958 * lockless access to see if there are pending faults
959 * __pollwait last action is the add_wait_queue but
960 * the spin_unlock would allow the waitqueue_active to
961 * pass above the actual list_add inside
962 * add_wait_queue critical section. So use a full
963 * memory barrier to serialize the list_add write of
964 * add_wait_queue() with the waitqueue_active read
969 if (waitqueue_active(&ctx->fault_pending_wqh))
971 else if (waitqueue_active(&ctx->event_wqh))
981 static const struct file_operations userfaultfd_fops;
983 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
985 struct uffd_msg *msg)
989 fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
990 O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
994 msg->arg.reserved.reserved1 = 0;
995 msg->arg.fork.ufd = fd;
999 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1000 struct uffd_msg *msg, struct inode *inode)
1003 DECLARE_WAITQUEUE(wait, current);
1004 struct userfaultfd_wait_queue *uwq;
1006 * Handling fork event requires sleeping operations, so
1007 * we drop the event_wqh lock, then do these ops, then
1008 * lock it back and wake up the waiter. While the lock is
1009 * dropped the ewq may go away so we keep track of it
1012 LIST_HEAD(fork_event);
1013 struct userfaultfd_ctx *fork_nctx = NULL;
1015 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1016 spin_lock_irq(&ctx->fd_wqh.lock);
1017 __add_wait_queue(&ctx->fd_wqh, &wait);
1019 set_current_state(TASK_INTERRUPTIBLE);
1020 spin_lock(&ctx->fault_pending_wqh.lock);
1021 uwq = find_userfault(ctx);
1024 * Use a seqcount to repeat the lockless check
1025 * in wake_userfault() to avoid missing
1026 * wakeups because during the refile both
1027 * waitqueue could become empty if this is the
1030 write_seqcount_begin(&ctx->refile_seq);
1033 * The fault_pending_wqh.lock prevents the uwq
1034 * to disappear from under us.
1036 * Refile this userfault from
1037 * fault_pending_wqh to fault_wqh, it's not
1038 * pending anymore after we read it.
1040 * Use list_del() by hand (as
1041 * userfaultfd_wake_function also uses
1042 * list_del_init() by hand) to be sure nobody
1043 * changes __remove_wait_queue() to use
1044 * list_del_init() in turn breaking the
1045 * !list_empty_careful() check in
1046 * handle_userfault(). The uwq->wq.head list
1047 * must never be empty at any time during the
1048 * refile, or the waitqueue could disappear
1049 * from under us. The "wait_queue_head_t"
1050 * parameter of __remove_wait_queue() is unused
1053 list_del(&uwq->wq.entry);
1054 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1056 write_seqcount_end(&ctx->refile_seq);
1058 /* careful to always initialize msg if ret == 0 */
1060 spin_unlock(&ctx->fault_pending_wqh.lock);
1064 spin_unlock(&ctx->fault_pending_wqh.lock);
1066 spin_lock(&ctx->event_wqh.lock);
1067 uwq = find_userfault_evt(ctx);
1071 if (uwq->msg.event == UFFD_EVENT_FORK) {
1072 fork_nctx = (struct userfaultfd_ctx *)
1074 uwq->msg.arg.reserved.reserved1;
1075 list_move(&uwq->wq.entry, &fork_event);
1077 * fork_nctx can be freed as soon as
1078 * we drop the lock, unless we take a
1081 userfaultfd_ctx_get(fork_nctx);
1082 spin_unlock(&ctx->event_wqh.lock);
1087 userfaultfd_event_complete(ctx, uwq);
1088 spin_unlock(&ctx->event_wqh.lock);
1092 spin_unlock(&ctx->event_wqh.lock);
1094 if (signal_pending(current)) {
1102 spin_unlock_irq(&ctx->fd_wqh.lock);
1104 spin_lock_irq(&ctx->fd_wqh.lock);
1106 __remove_wait_queue(&ctx->fd_wqh, &wait);
1107 __set_current_state(TASK_RUNNING);
1108 spin_unlock_irq(&ctx->fd_wqh.lock);
1110 if (!ret && msg->event == UFFD_EVENT_FORK) {
1111 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1112 spin_lock_irq(&ctx->event_wqh.lock);
1113 if (!list_empty(&fork_event)) {
1115 * The fork thread didn't abort, so we can
1116 * drop the temporary refcount.
1118 userfaultfd_ctx_put(fork_nctx);
1120 uwq = list_first_entry(&fork_event,
1124 * If fork_event list wasn't empty and in turn
1125 * the event wasn't already released by fork
1126 * (the event is allocated on fork kernel
1127 * stack), put the event back to its place in
1128 * the event_wq. fork_event head will be freed
1129 * as soon as we return so the event cannot
1130 * stay queued there no matter the current
1133 list_del(&uwq->wq.entry);
1134 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1137 * Leave the event in the waitqueue and report
1138 * error to userland if we failed to resolve
1139 * the userfault fork.
1142 userfaultfd_event_complete(ctx, uwq);
1145 * Here the fork thread aborted and the
1146 * refcount from the fork thread on fork_nctx
1147 * has already been released. We still hold
1148 * the reference we took before releasing the
1149 * lock above. If resolve_userfault_fork
1150 * failed we've to drop it because the
1151 * fork_nctx has to be freed in such case. If
1152 * it succeeded we'll hold it because the new
1153 * uffd references it.
1156 userfaultfd_ctx_put(fork_nctx);
1158 spin_unlock_irq(&ctx->event_wqh.lock);
1164 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1165 size_t count, loff_t *ppos)
1167 struct userfaultfd_ctx *ctx = file->private_data;
1168 ssize_t _ret, ret = 0;
1169 struct uffd_msg msg;
1170 int no_wait = file->f_flags & O_NONBLOCK;
1171 struct inode *inode = file_inode(file);
1173 if (ctx->state == UFFD_STATE_WAIT_API)
1177 if (count < sizeof(msg))
1178 return ret ? ret : -EINVAL;
1179 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1181 return ret ? ret : _ret;
1182 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1183 return ret ? ret : -EFAULT;
1186 count -= sizeof(msg);
1188 * Allow to read more than one fault at time but only
1189 * block if waiting for the very first one.
1191 no_wait = O_NONBLOCK;
1195 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1196 struct userfaultfd_wake_range *range)
1198 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1199 /* wake all in the range and autoremove */
1200 if (waitqueue_active(&ctx->fault_pending_wqh))
1201 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1203 if (waitqueue_active(&ctx->fault_wqh))
1204 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1205 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1208 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1209 struct userfaultfd_wake_range *range)
1215 * To be sure waitqueue_active() is not reordered by the CPU
1216 * before the pagetable update, use an explicit SMP memory
1217 * barrier here. PT lock release or mmap_read_unlock(mm) still
1218 * have release semantics that can allow the
1219 * waitqueue_active() to be reordered before the pte update.
1224 * Use waitqueue_active because it's very frequent to
1225 * change the address space atomically even if there are no
1226 * userfaults yet. So we take the spinlock only when we're
1227 * sure we've userfaults to wake.
1230 seq = read_seqcount_begin(&ctx->refile_seq);
1231 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1232 waitqueue_active(&ctx->fault_wqh);
1234 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1236 __wake_userfault(ctx, range);
1239 static __always_inline int validate_range(struct mm_struct *mm,
1240 __u64 *start, __u64 len)
1242 __u64 task_size = mm->task_size;
1244 *start = untagged_addr(*start);
1246 if (*start & ~PAGE_MASK)
1248 if (len & ~PAGE_MASK)
1252 if (*start < mmap_min_addr)
1254 if (*start >= task_size)
1256 if (len > task_size - *start)
1261 static inline bool vma_can_userfault(struct vm_area_struct *vma,
1262 unsigned long vm_flags)
1264 /* FIXME: add WP support to hugetlbfs and shmem */
1265 return vma_is_anonymous(vma) ||
1266 ((is_vm_hugetlb_page(vma) || vma_is_shmem(vma)) &&
1267 !(vm_flags & VM_UFFD_WP));
1270 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1273 struct mm_struct *mm = ctx->mm;
1274 struct vm_area_struct *vma, *prev, *cur;
1276 struct uffdio_register uffdio_register;
1277 struct uffdio_register __user *user_uffdio_register;
1278 unsigned long vm_flags, new_flags;
1281 unsigned long start, end, vma_end;
1283 user_uffdio_register = (struct uffdio_register __user *) arg;
1286 if (copy_from_user(&uffdio_register, user_uffdio_register,
1287 sizeof(uffdio_register)-sizeof(__u64)))
1291 if (!uffdio_register.mode)
1293 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1294 UFFDIO_REGISTER_MODE_WP))
1297 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1298 vm_flags |= VM_UFFD_MISSING;
1299 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)
1300 vm_flags |= VM_UFFD_WP;
1302 ret = validate_range(mm, &uffdio_register.range.start,
1303 uffdio_register.range.len);
1307 start = uffdio_register.range.start;
1308 end = start + uffdio_register.range.len;
1311 if (!mmget_not_zero(mm))
1314 mmap_write_lock(mm);
1315 vma = find_vma_prev(mm, start, &prev);
1319 /* check that there's at least one vma in the range */
1321 if (vma->vm_start >= end)
1325 * If the first vma contains huge pages, make sure start address
1326 * is aligned to huge page size.
1328 if (is_vm_hugetlb_page(vma)) {
1329 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1331 if (start & (vma_hpagesize - 1))
1336 * Search for not compatible vmas.
1339 basic_ioctls = false;
1340 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1343 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1344 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1346 /* check not compatible vmas */
1348 if (!vma_can_userfault(cur, vm_flags))
1352 * UFFDIO_COPY will fill file holes even without
1353 * PROT_WRITE. This check enforces that if this is a
1354 * MAP_SHARED, the process has write permission to the backing
1355 * file. If VM_MAYWRITE is set it also enforces that on a
1356 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1357 * F_WRITE_SEAL can be taken until the vma is destroyed.
1360 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1364 * If this vma contains ending address, and huge pages
1367 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1368 end > cur->vm_start) {
1369 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1373 if (end & (vma_hpagesize - 1))
1376 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1380 * Check that this vma isn't already owned by a
1381 * different userfaultfd. We can't allow more than one
1382 * userfaultfd to own a single vma simultaneously or we
1383 * wouldn't know which one to deliver the userfaults to.
1386 if (cur->vm_userfaultfd_ctx.ctx &&
1387 cur->vm_userfaultfd_ctx.ctx != ctx)
1391 * Note vmas containing huge pages
1393 if (is_vm_hugetlb_page(cur))
1394 basic_ioctls = true;
1400 if (vma->vm_start < start)
1407 BUG_ON(!vma_can_userfault(vma, vm_flags));
1408 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1409 vma->vm_userfaultfd_ctx.ctx != ctx);
1410 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1413 * Nothing to do: this vma is already registered into this
1414 * userfaultfd and with the right tracking mode too.
1416 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1417 (vma->vm_flags & vm_flags) == vm_flags)
1420 if (vma->vm_start > start)
1421 start = vma->vm_start;
1422 vma_end = min(end, vma->vm_end);
1424 new_flags = (vma->vm_flags &
1425 ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags;
1426 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1427 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1429 ((struct vm_userfaultfd_ctx){ ctx }));
1434 if (vma->vm_start < start) {
1435 ret = split_vma(mm, vma, start, 1);
1439 if (vma->vm_end > end) {
1440 ret = split_vma(mm, vma, end, 0);
1446 * In the vma_merge() successful mprotect-like case 8:
1447 * the next vma was merged into the current one and
1448 * the current one has not been updated yet.
1450 vma->vm_flags = new_flags;
1451 vma->vm_userfaultfd_ctx.ctx = ctx;
1453 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1454 hugetlb_unshare_all_pmds(vma);
1458 start = vma->vm_end;
1460 } while (vma && vma->vm_start < end);
1462 mmap_write_unlock(mm);
1467 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1468 UFFD_API_RANGE_IOCTLS;
1471 * Declare the WP ioctl only if the WP mode is
1472 * specified and all checks passed with the range
1474 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1475 ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1478 * Now that we scanned all vmas we can already tell
1479 * userland which ioctls methods are guaranteed to
1480 * succeed on this range.
1482 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1489 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1492 struct mm_struct *mm = ctx->mm;
1493 struct vm_area_struct *vma, *prev, *cur;
1495 struct uffdio_range uffdio_unregister;
1496 unsigned long new_flags;
1498 unsigned long start, end, vma_end;
1499 const void __user *buf = (void __user *)arg;
1502 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1505 ret = validate_range(mm, &uffdio_unregister.start,
1506 uffdio_unregister.len);
1510 start = uffdio_unregister.start;
1511 end = start + uffdio_unregister.len;
1514 if (!mmget_not_zero(mm))
1517 mmap_write_lock(mm);
1518 vma = find_vma_prev(mm, start, &prev);
1522 /* check that there's at least one vma in the range */
1524 if (vma->vm_start >= end)
1528 * If the first vma contains huge pages, make sure start address
1529 * is aligned to huge page size.
1531 if (is_vm_hugetlb_page(vma)) {
1532 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1534 if (start & (vma_hpagesize - 1))
1539 * Search for not compatible vmas.
1543 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1546 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1547 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1550 * Check not compatible vmas, not strictly required
1551 * here as not compatible vmas cannot have an
1552 * userfaultfd_ctx registered on them, but this
1553 * provides for more strict behavior to notice
1554 * unregistration errors.
1556 if (!vma_can_userfault(cur, cur->vm_flags))
1563 if (vma->vm_start < start)
1570 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1573 * Nothing to do: this vma is already registered into this
1574 * userfaultfd and with the right tracking mode too.
1576 if (!vma->vm_userfaultfd_ctx.ctx)
1579 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1581 if (vma->vm_start > start)
1582 start = vma->vm_start;
1583 vma_end = min(end, vma->vm_end);
1585 if (userfaultfd_missing(vma)) {
1587 * Wake any concurrent pending userfault while
1588 * we unregister, so they will not hang
1589 * permanently and it avoids userland to call
1590 * UFFDIO_WAKE explicitly.
1592 struct userfaultfd_wake_range range;
1593 range.start = start;
1594 range.len = vma_end - start;
1595 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1598 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1599 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1600 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1607 if (vma->vm_start < start) {
1608 ret = split_vma(mm, vma, start, 1);
1612 if (vma->vm_end > end) {
1613 ret = split_vma(mm, vma, end, 0);
1619 * In the vma_merge() successful mprotect-like case 8:
1620 * the next vma was merged into the current one and
1621 * the current one has not been updated yet.
1623 vma->vm_flags = new_flags;
1624 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1628 start = vma->vm_end;
1630 } while (vma && vma->vm_start < end);
1632 mmap_write_unlock(mm);
1639 * userfaultfd_wake may be used in combination with the
1640 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1642 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1646 struct uffdio_range uffdio_wake;
1647 struct userfaultfd_wake_range range;
1648 const void __user *buf = (void __user *)arg;
1651 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1654 ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len);
1658 range.start = uffdio_wake.start;
1659 range.len = uffdio_wake.len;
1662 * len == 0 means wake all and we don't want to wake all here,
1663 * so check it again to be sure.
1665 VM_BUG_ON(!range.len);
1667 wake_userfault(ctx, &range);
1674 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1678 struct uffdio_copy uffdio_copy;
1679 struct uffdio_copy __user *user_uffdio_copy;
1680 struct userfaultfd_wake_range range;
1682 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1685 if (READ_ONCE(ctx->mmap_changing))
1689 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1690 /* don't copy "copy" last field */
1691 sizeof(uffdio_copy)-sizeof(__s64)))
1694 ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len);
1698 * double check for wraparound just in case. copy_from_user()
1699 * will later check uffdio_copy.src + uffdio_copy.len to fit
1700 * in the userland range.
1703 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1705 if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1707 if (mmget_not_zero(ctx->mm)) {
1708 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1709 uffdio_copy.len, &ctx->mmap_changing,
1715 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1720 /* len == 0 would wake all */
1722 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1723 range.start = uffdio_copy.dst;
1724 wake_userfault(ctx, &range);
1726 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1731 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1735 struct uffdio_zeropage uffdio_zeropage;
1736 struct uffdio_zeropage __user *user_uffdio_zeropage;
1737 struct userfaultfd_wake_range range;
1739 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1742 if (READ_ONCE(ctx->mmap_changing))
1746 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1747 /* don't copy "zeropage" last field */
1748 sizeof(uffdio_zeropage)-sizeof(__s64)))
1751 ret = validate_range(ctx->mm, &uffdio_zeropage.range.start,
1752 uffdio_zeropage.range.len);
1756 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1759 if (mmget_not_zero(ctx->mm)) {
1760 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1761 uffdio_zeropage.range.len,
1762 &ctx->mmap_changing);
1767 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1771 /* len == 0 would wake all */
1774 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1775 range.start = uffdio_zeropage.range.start;
1776 wake_userfault(ctx, &range);
1778 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1783 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1787 struct uffdio_writeprotect uffdio_wp;
1788 struct uffdio_writeprotect __user *user_uffdio_wp;
1789 struct userfaultfd_wake_range range;
1790 bool mode_wp, mode_dontwake;
1792 if (READ_ONCE(ctx->mmap_changing))
1795 user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1797 if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1798 sizeof(struct uffdio_writeprotect)))
1801 ret = validate_range(ctx->mm, &uffdio_wp.range.start,
1802 uffdio_wp.range.len);
1806 if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1807 UFFDIO_WRITEPROTECT_MODE_WP))
1810 mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1811 mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1813 if (mode_wp && mode_dontwake)
1816 ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1817 uffdio_wp.range.len, mode_wp,
1818 &ctx->mmap_changing);
1822 if (!mode_wp && !mode_dontwake) {
1823 range.start = uffdio_wp.range.start;
1824 range.len = uffdio_wp.range.len;
1825 wake_userfault(ctx, &range);
1830 static inline unsigned int uffd_ctx_features(__u64 user_features)
1833 * For the current set of features the bits just coincide
1835 return (unsigned int)user_features;
1839 * userland asks for a certain API version and we return which bits
1840 * and ioctl commands are implemented in this kernel for such API
1841 * version or -EINVAL if unknown.
1843 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1846 struct uffdio_api uffdio_api;
1847 void __user *buf = (void __user *)arg;
1852 if (ctx->state != UFFD_STATE_WAIT_API)
1855 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1857 features = uffdio_api.features;
1859 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1862 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1864 /* report all available features and ioctls to userland */
1865 uffdio_api.features = UFFD_API_FEATURES;
1866 uffdio_api.ioctls = UFFD_API_IOCTLS;
1868 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1870 ctx->state = UFFD_STATE_RUNNING;
1871 /* only enable the requested features for this uffd context */
1872 ctx->features = uffd_ctx_features(features);
1877 memset(&uffdio_api, 0, sizeof(uffdio_api));
1878 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1883 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1887 struct userfaultfd_ctx *ctx = file->private_data;
1889 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1894 ret = userfaultfd_api(ctx, arg);
1896 case UFFDIO_REGISTER:
1897 ret = userfaultfd_register(ctx, arg);
1899 case UFFDIO_UNREGISTER:
1900 ret = userfaultfd_unregister(ctx, arg);
1903 ret = userfaultfd_wake(ctx, arg);
1906 ret = userfaultfd_copy(ctx, arg);
1908 case UFFDIO_ZEROPAGE:
1909 ret = userfaultfd_zeropage(ctx, arg);
1911 case UFFDIO_WRITEPROTECT:
1912 ret = userfaultfd_writeprotect(ctx, arg);
1918 #ifdef CONFIG_PROC_FS
1919 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1921 struct userfaultfd_ctx *ctx = f->private_data;
1922 wait_queue_entry_t *wq;
1923 unsigned long pending = 0, total = 0;
1925 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1926 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1930 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1933 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1936 * If more protocols will be added, there will be all shown
1937 * separated by a space. Like this:
1938 * protocols: aa:... bb:...
1940 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1941 pending, total, UFFD_API, ctx->features,
1942 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1946 static const struct file_operations userfaultfd_fops = {
1947 #ifdef CONFIG_PROC_FS
1948 .show_fdinfo = userfaultfd_show_fdinfo,
1950 .release = userfaultfd_release,
1951 .poll = userfaultfd_poll,
1952 .read = userfaultfd_read,
1953 .unlocked_ioctl = userfaultfd_ioctl,
1954 .compat_ioctl = compat_ptr_ioctl,
1955 .llseek = noop_llseek,
1958 static void init_once_userfaultfd_ctx(void *mem)
1960 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1962 init_waitqueue_head(&ctx->fault_pending_wqh);
1963 init_waitqueue_head(&ctx->fault_wqh);
1964 init_waitqueue_head(&ctx->event_wqh);
1965 init_waitqueue_head(&ctx->fd_wqh);
1966 seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
1969 SYSCALL_DEFINE1(userfaultfd, int, flags)
1971 struct userfaultfd_ctx *ctx;
1974 if (!sysctl_unprivileged_userfaultfd &&
1975 (flags & UFFD_USER_MODE_ONLY) == 0 &&
1976 !capable(CAP_SYS_PTRACE)) {
1977 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
1978 "sysctl knob to 1 if kernel faults must be handled "
1979 "without obtaining CAP_SYS_PTRACE capability\n");
1983 BUG_ON(!current->mm);
1985 /* Check the UFFD_* constants for consistency. */
1986 BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
1987 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1988 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1990 if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
1993 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1997 refcount_set(&ctx->refcount, 1);
2000 ctx->state = UFFD_STATE_WAIT_API;
2001 ctx->released = false;
2002 ctx->mmap_changing = false;
2003 ctx->mm = current->mm;
2004 /* prevent the mm struct to be freed */
2007 fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2008 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2011 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2016 static int __init userfaultfd_init(void)
2018 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2019 sizeof(struct userfaultfd_ctx),
2021 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2022 init_once_userfaultfd_ctx);
2025 __initcall(userfaultfd_init);