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/poll.h>
19 #include <linux/slab.h>
20 #include <linux/seq_file.h>
21 #include <linux/file.h>
22 #include <linux/bug.h>
23 #include <linux/anon_inodes.h>
24 #include <linux/syscalls.h>
25 #include <linux/userfaultfd_k.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ioctl.h>
28 #include <linux/security.h>
29 #include <linux/hugetlb.h>
31 int sysctl_unprivileged_userfaultfd __read_mostly = 1;
33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
35 enum userfaultfd_state {
41 * Start with fault_pending_wqh and fault_wqh so they're more likely
42 * to be in the same cacheline.
46 * fault_pending_wqh.lock
50 * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
51 * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
52 * also taken in IRQ context.
54 struct userfaultfd_ctx {
55 /* waitqueue head for the pending (i.e. not read) userfaults */
56 wait_queue_head_t fault_pending_wqh;
57 /* waitqueue head for the userfaults */
58 wait_queue_head_t fault_wqh;
59 /* waitqueue head for the pseudo fd to wakeup poll/read */
60 wait_queue_head_t fd_wqh;
61 /* waitqueue head for events */
62 wait_queue_head_t event_wqh;
63 /* a refile sequence protected by fault_pending_wqh lock */
64 struct seqcount refile_seq;
65 /* pseudo fd refcounting */
67 /* userfaultfd syscall flags */
69 /* features requested from the userspace */
70 unsigned int features;
72 enum userfaultfd_state state;
75 /* memory mappings are changing because of non-cooperative event */
77 /* mm with one ore more vmas attached to this userfaultfd_ctx */
81 struct userfaultfd_fork_ctx {
82 struct userfaultfd_ctx *orig;
83 struct userfaultfd_ctx *new;
84 struct list_head list;
87 struct userfaultfd_unmap_ctx {
88 struct userfaultfd_ctx *ctx;
91 struct list_head list;
94 struct userfaultfd_wait_queue {
96 wait_queue_entry_t wq;
97 struct userfaultfd_ctx *ctx;
101 struct userfaultfd_wake_range {
106 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
107 int wake_flags, void *key)
109 struct userfaultfd_wake_range *range = key;
111 struct userfaultfd_wait_queue *uwq;
112 unsigned long start, len;
114 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
116 /* len == 0 means wake all */
117 start = range->start;
119 if (len && (start > uwq->msg.arg.pagefault.address ||
120 start + len <= uwq->msg.arg.pagefault.address))
122 WRITE_ONCE(uwq->waken, true);
124 * The Program-Order guarantees provided by the scheduler
125 * ensure uwq->waken is visible before the task is woken.
127 ret = wake_up_state(wq->private, mode);
130 * Wake only once, autoremove behavior.
132 * After the effect of list_del_init is visible to the other
133 * CPUs, the waitqueue may disappear from under us, see the
134 * !list_empty_careful() in handle_userfault().
136 * try_to_wake_up() has an implicit smp_mb(), and the
137 * wq->private is read before calling the extern function
138 * "wake_up_state" (which in turns calls try_to_wake_up).
140 list_del_init(&wq->entry);
147 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
149 * @ctx: [in] Pointer to the userfaultfd context.
151 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
153 refcount_inc(&ctx->refcount);
157 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
159 * @ctx: [in] Pointer to userfaultfd context.
161 * The userfaultfd context reference must have been previously acquired either
162 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
164 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
166 if (refcount_dec_and_test(&ctx->refcount)) {
167 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
168 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
169 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
170 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
171 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
172 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
173 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
174 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
176 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
180 static inline void msg_init(struct uffd_msg *msg)
182 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
184 * Must use memset to zero out the paddings or kernel data is
185 * leaked to userland.
187 memset(msg, 0, sizeof(struct uffd_msg));
190 static inline struct uffd_msg userfault_msg(unsigned long address,
192 unsigned long reason,
193 unsigned int features)
197 msg.event = UFFD_EVENT_PAGEFAULT;
198 msg.arg.pagefault.address = address;
199 if (flags & FAULT_FLAG_WRITE)
201 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
202 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
203 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
204 * was a read fault, otherwise if set it means it's
207 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
208 if (reason & VM_UFFD_WP)
210 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
211 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
212 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
213 * a missing fault, otherwise if set it means it's a
214 * write protect fault.
216 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
217 if (features & UFFD_FEATURE_THREAD_ID)
218 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
222 #ifdef CONFIG_HUGETLB_PAGE
224 * Same functionality as userfaultfd_must_wait below with modifications for
227 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
228 struct vm_area_struct *vma,
229 unsigned long address,
231 unsigned long reason)
233 struct mm_struct *mm = ctx->mm;
237 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
239 ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
245 pte = huge_ptep_get(ptep);
248 * Lockless access: we're in a wait_event so it's ok if it
251 if (huge_pte_none(pte))
253 if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
259 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
260 struct vm_area_struct *vma,
261 unsigned long address,
263 unsigned long reason)
265 return false; /* should never get here */
267 #endif /* CONFIG_HUGETLB_PAGE */
270 * Verify the pagetables are still not ok after having reigstered into
271 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
272 * userfault that has already been resolved, if userfaultfd_read and
273 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
276 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
277 unsigned long address,
279 unsigned long reason)
281 struct mm_struct *mm = ctx->mm;
289 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
291 pgd = pgd_offset(mm, address);
292 if (!pgd_present(*pgd))
294 p4d = p4d_offset(pgd, address);
295 if (!p4d_present(*p4d))
297 pud = pud_offset(p4d, address);
298 if (!pud_present(*pud))
300 pmd = pmd_offset(pud, address);
302 * READ_ONCE must function as a barrier with narrower scope
303 * and it must be equivalent to:
304 * _pmd = *pmd; barrier();
306 * This is to deal with the instability (as in
307 * pmd_trans_unstable) of the pmd.
309 _pmd = READ_ONCE(*pmd);
314 if (!pmd_present(_pmd))
317 if (pmd_trans_huge(_pmd))
321 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
322 * and use the standard pte_offset_map() instead of parsing _pmd.
324 pte = pte_offset_map(pmd, address);
326 * Lockless access: we're in a wait_event so it's ok if it
337 /* Should pair with userfaultfd_signal_pending() */
338 static inline long userfaultfd_get_blocking_state(unsigned int flags)
340 if (flags & FAULT_FLAG_INTERRUPTIBLE)
341 return TASK_INTERRUPTIBLE;
343 if (flags & FAULT_FLAG_KILLABLE)
344 return TASK_KILLABLE;
346 return TASK_UNINTERRUPTIBLE;
349 /* Should pair with userfaultfd_get_blocking_state() */
350 static inline bool userfaultfd_signal_pending(unsigned int flags)
352 if (flags & FAULT_FLAG_INTERRUPTIBLE)
353 return signal_pending(current);
355 if (flags & FAULT_FLAG_KILLABLE)
356 return fatal_signal_pending(current);
362 * The locking rules involved in returning VM_FAULT_RETRY depending on
363 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
364 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
365 * recommendation in __lock_page_or_retry is not an understatement.
367 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
368 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
371 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
372 * set, VM_FAULT_RETRY can still be returned if and only if there are
373 * fatal_signal_pending()s, and the mmap_sem must be released before
376 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
378 struct mm_struct *mm = vmf->vma->vm_mm;
379 struct userfaultfd_ctx *ctx;
380 struct userfaultfd_wait_queue uwq;
381 vm_fault_t ret = VM_FAULT_SIGBUS;
386 * We don't do userfault handling for the final child pid update.
388 * We also don't do userfault handling during
389 * coredumping. hugetlbfs has the special
390 * follow_hugetlb_page() to skip missing pages in the
391 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
392 * the no_page_table() helper in follow_page_mask(), but the
393 * shmem_vm_ops->fault method is invoked even during
394 * coredumping without mmap_sem and it ends up here.
396 if (current->flags & (PF_EXITING|PF_DUMPCORE))
400 * Coredumping runs without mmap_sem so we can only check that
401 * the mmap_sem is held, if PF_DUMPCORE was not set.
403 WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
405 ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
409 BUG_ON(ctx->mm != mm);
411 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
412 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
414 if (ctx->features & UFFD_FEATURE_SIGBUS)
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_sem.
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_sem */
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 up_read(&mm->mmap_sem);
514 if (likely(must_wait && !READ_ONCE(ctx->released) &&
515 !userfaultfd_signal_pending(vmf->flags))) {
516 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
518 ret |= VM_FAULT_MAJOR;
521 * False wakeups can orginate even from rwsem before
522 * up_read() however userfaults will wait either for a
523 * targeted wakeup on the specific uwq waitqueue from
524 * wake_userfault() or for signals or for uffd
527 while (!READ_ONCE(uwq.waken)) {
529 * This needs the full smp_store_mb()
530 * guarantee as the state write must be
531 * visible to other CPUs before reading
532 * uwq.waken from other CPUs.
534 set_current_state(blocking_state);
535 if (READ_ONCE(uwq.waken) ||
536 READ_ONCE(ctx->released) ||
537 userfaultfd_signal_pending(vmf->flags))
543 __set_current_state(TASK_RUNNING);
546 * Here we race with the list_del; list_add in
547 * userfaultfd_ctx_read(), however because we don't ever run
548 * list_del_init() to refile across the two lists, the prev
549 * and next pointers will never point to self. list_add also
550 * would never let any of the two pointers to point to
551 * self. So list_empty_careful won't risk to see both pointers
552 * pointing to self at any time during the list refile. The
553 * only case where list_del_init() is called is the full
554 * removal in the wake function and there we don't re-list_add
555 * and it's fine not to block on the spinlock. The uwq on this
556 * kernel stack can be released after the list_del_init.
558 if (!list_empty_careful(&uwq.wq.entry)) {
559 spin_lock_irq(&ctx->fault_pending_wqh.lock);
561 * No need of list_del_init(), the uwq on the stack
562 * will be freed shortly anyway.
564 list_del(&uwq.wq.entry);
565 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
569 * ctx may go away after this if the userfault pseudo fd is
572 userfaultfd_ctx_put(ctx);
578 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
579 struct userfaultfd_wait_queue *ewq)
581 struct userfaultfd_ctx *release_new_ctx;
583 if (WARN_ON_ONCE(current->flags & PF_EXITING))
587 init_waitqueue_entry(&ewq->wq, current);
588 release_new_ctx = NULL;
590 spin_lock_irq(&ctx->event_wqh.lock);
592 * After the __add_wait_queue the uwq is visible to userland
593 * through poll/read().
595 __add_wait_queue(&ctx->event_wqh, &ewq->wq);
597 set_current_state(TASK_KILLABLE);
598 if (ewq->msg.event == 0)
600 if (READ_ONCE(ctx->released) ||
601 fatal_signal_pending(current)) {
603 * &ewq->wq may be queued in fork_event, but
604 * __remove_wait_queue ignores the head
605 * parameter. It would be a problem if it
608 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
609 if (ewq->msg.event == UFFD_EVENT_FORK) {
610 struct userfaultfd_ctx *new;
612 new = (struct userfaultfd_ctx *)
614 ewq->msg.arg.reserved.reserved1;
615 release_new_ctx = new;
620 spin_unlock_irq(&ctx->event_wqh.lock);
622 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
625 spin_lock_irq(&ctx->event_wqh.lock);
627 __set_current_state(TASK_RUNNING);
628 spin_unlock_irq(&ctx->event_wqh.lock);
630 if (release_new_ctx) {
631 struct vm_area_struct *vma;
632 struct mm_struct *mm = release_new_ctx->mm;
634 /* the various vma->vm_userfaultfd_ctx still points to it */
635 down_write(&mm->mmap_sem);
636 /* no task can run (and in turn coredump) yet */
637 VM_WARN_ON(!mmget_still_valid(mm));
638 for (vma = mm->mmap; vma; vma = vma->vm_next)
639 if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
640 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
641 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
643 up_write(&mm->mmap_sem);
645 userfaultfd_ctx_put(release_new_ctx);
649 * ctx may go away after this if the userfault pseudo fd is
653 WRITE_ONCE(ctx->mmap_changing, false);
654 userfaultfd_ctx_put(ctx);
657 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
658 struct userfaultfd_wait_queue *ewq)
661 wake_up_locked(&ctx->event_wqh);
662 __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
665 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
667 struct userfaultfd_ctx *ctx = NULL, *octx;
668 struct userfaultfd_fork_ctx *fctx;
670 octx = vma->vm_userfaultfd_ctx.ctx;
671 if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
672 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
673 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
677 list_for_each_entry(fctx, fcs, list)
678 if (fctx->orig == octx) {
684 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
688 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
694 refcount_set(&ctx->refcount, 1);
695 ctx->flags = octx->flags;
696 ctx->state = UFFD_STATE_RUNNING;
697 ctx->features = octx->features;
698 ctx->released = false;
699 ctx->mmap_changing = false;
700 ctx->mm = vma->vm_mm;
703 userfaultfd_ctx_get(octx);
704 WRITE_ONCE(octx->mmap_changing, true);
707 list_add_tail(&fctx->list, fcs);
710 vma->vm_userfaultfd_ctx.ctx = ctx;
714 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
716 struct userfaultfd_ctx *ctx = fctx->orig;
717 struct userfaultfd_wait_queue ewq;
721 ewq.msg.event = UFFD_EVENT_FORK;
722 ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
724 userfaultfd_event_wait_completion(ctx, &ewq);
727 void dup_userfaultfd_complete(struct list_head *fcs)
729 struct userfaultfd_fork_ctx *fctx, *n;
731 list_for_each_entry_safe(fctx, n, fcs, list) {
733 list_del(&fctx->list);
738 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
739 struct vm_userfaultfd_ctx *vm_ctx)
741 struct userfaultfd_ctx *ctx;
743 ctx = vma->vm_userfaultfd_ctx.ctx;
748 if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
750 userfaultfd_ctx_get(ctx);
751 WRITE_ONCE(ctx->mmap_changing, true);
753 /* Drop uffd context if remap feature not enabled */
754 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
755 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
759 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
760 unsigned long from, unsigned long to,
763 struct userfaultfd_ctx *ctx = vm_ctx->ctx;
764 struct userfaultfd_wait_queue ewq;
769 if (to & ~PAGE_MASK) {
770 userfaultfd_ctx_put(ctx);
776 ewq.msg.event = UFFD_EVENT_REMAP;
777 ewq.msg.arg.remap.from = from;
778 ewq.msg.arg.remap.to = to;
779 ewq.msg.arg.remap.len = len;
781 userfaultfd_event_wait_completion(ctx, &ewq);
784 bool userfaultfd_remove(struct vm_area_struct *vma,
785 unsigned long start, unsigned long end)
787 struct mm_struct *mm = vma->vm_mm;
788 struct userfaultfd_ctx *ctx;
789 struct userfaultfd_wait_queue ewq;
791 ctx = vma->vm_userfaultfd_ctx.ctx;
792 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
795 userfaultfd_ctx_get(ctx);
796 WRITE_ONCE(ctx->mmap_changing, true);
797 up_read(&mm->mmap_sem);
801 ewq.msg.event = UFFD_EVENT_REMOVE;
802 ewq.msg.arg.remove.start = start;
803 ewq.msg.arg.remove.end = end;
805 userfaultfd_event_wait_completion(ctx, &ewq);
810 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
811 unsigned long start, unsigned long end)
813 struct userfaultfd_unmap_ctx *unmap_ctx;
815 list_for_each_entry(unmap_ctx, unmaps, list)
816 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
817 unmap_ctx->end == end)
823 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
824 unsigned long start, unsigned long end,
825 struct list_head *unmaps)
827 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
828 struct userfaultfd_unmap_ctx *unmap_ctx;
829 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
831 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
832 has_unmap_ctx(ctx, unmaps, start, end))
835 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
839 userfaultfd_ctx_get(ctx);
840 WRITE_ONCE(ctx->mmap_changing, true);
841 unmap_ctx->ctx = ctx;
842 unmap_ctx->start = start;
843 unmap_ctx->end = end;
844 list_add_tail(&unmap_ctx->list, unmaps);
850 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
852 struct userfaultfd_unmap_ctx *ctx, *n;
853 struct userfaultfd_wait_queue ewq;
855 list_for_each_entry_safe(ctx, n, uf, list) {
858 ewq.msg.event = UFFD_EVENT_UNMAP;
859 ewq.msg.arg.remove.start = ctx->start;
860 ewq.msg.arg.remove.end = ctx->end;
862 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
864 list_del(&ctx->list);
869 static int userfaultfd_release(struct inode *inode, struct file *file)
871 struct userfaultfd_ctx *ctx = file->private_data;
872 struct mm_struct *mm = ctx->mm;
873 struct vm_area_struct *vma, *prev;
874 /* len == 0 means wake all */
875 struct userfaultfd_wake_range range = { .len = 0, };
876 unsigned long new_flags;
879 WRITE_ONCE(ctx->released, true);
881 if (!mmget_not_zero(mm))
885 * Flush page faults out of all CPUs. NOTE: all page faults
886 * must be retried without returning VM_FAULT_SIGBUS if
887 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
888 * changes while handle_userfault released the mmap_sem. So
889 * it's critical that released is set to true (above), before
890 * taking the mmap_sem for writing.
892 down_write(&mm->mmap_sem);
893 still_valid = mmget_still_valid(mm);
895 for (vma = mm->mmap; vma; vma = vma->vm_next) {
897 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
898 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
899 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
903 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
905 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
906 new_flags, vma->anon_vma,
907 vma->vm_file, vma->vm_pgoff,
915 vma->vm_flags = new_flags;
916 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
918 up_write(&mm->mmap_sem);
922 * After no new page faults can wait on this fault_*wqh, flush
923 * the last page faults that may have been already waiting on
926 spin_lock_irq(&ctx->fault_pending_wqh.lock);
927 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
928 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
929 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
931 /* Flush pending events that may still wait on event_wqh */
932 wake_up_all(&ctx->event_wqh);
934 wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
935 userfaultfd_ctx_put(ctx);
939 /* fault_pending_wqh.lock must be hold by the caller */
940 static inline struct userfaultfd_wait_queue *find_userfault_in(
941 wait_queue_head_t *wqh)
943 wait_queue_entry_t *wq;
944 struct userfaultfd_wait_queue *uwq;
946 lockdep_assert_held(&wqh->lock);
949 if (!waitqueue_active(wqh))
951 /* walk in reverse to provide FIFO behavior to read userfaults */
952 wq = list_last_entry(&wqh->head, typeof(*wq), entry);
953 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
958 static inline struct userfaultfd_wait_queue *find_userfault(
959 struct userfaultfd_ctx *ctx)
961 return find_userfault_in(&ctx->fault_pending_wqh);
964 static inline struct userfaultfd_wait_queue *find_userfault_evt(
965 struct userfaultfd_ctx *ctx)
967 return find_userfault_in(&ctx->event_wqh);
970 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
972 struct userfaultfd_ctx *ctx = file->private_data;
975 poll_wait(file, &ctx->fd_wqh, wait);
977 switch (ctx->state) {
978 case UFFD_STATE_WAIT_API:
980 case UFFD_STATE_RUNNING:
982 * poll() never guarantees that read won't block.
983 * userfaults can be waken before they're read().
985 if (unlikely(!(file->f_flags & O_NONBLOCK)))
988 * lockless access to see if there are pending faults
989 * __pollwait last action is the add_wait_queue but
990 * the spin_unlock would allow the waitqueue_active to
991 * pass above the actual list_add inside
992 * add_wait_queue critical section. So use a full
993 * memory barrier to serialize the list_add write of
994 * add_wait_queue() with the waitqueue_active read
999 if (waitqueue_active(&ctx->fault_pending_wqh))
1001 else if (waitqueue_active(&ctx->event_wqh))
1011 static const struct file_operations userfaultfd_fops;
1013 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
1014 struct userfaultfd_ctx *new,
1015 struct uffd_msg *msg)
1019 fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
1020 O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
1024 msg->arg.reserved.reserved1 = 0;
1025 msg->arg.fork.ufd = fd;
1029 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1030 struct uffd_msg *msg)
1033 DECLARE_WAITQUEUE(wait, current);
1034 struct userfaultfd_wait_queue *uwq;
1036 * Handling fork event requires sleeping operations, so
1037 * we drop the event_wqh lock, then do these ops, then
1038 * lock it back and wake up the waiter. While the lock is
1039 * dropped the ewq may go away so we keep track of it
1042 LIST_HEAD(fork_event);
1043 struct userfaultfd_ctx *fork_nctx = NULL;
1045 /* always take the fd_wqh lock before the fault_pending_wqh lock */
1046 spin_lock_irq(&ctx->fd_wqh.lock);
1047 __add_wait_queue(&ctx->fd_wqh, &wait);
1049 set_current_state(TASK_INTERRUPTIBLE);
1050 spin_lock(&ctx->fault_pending_wqh.lock);
1051 uwq = find_userfault(ctx);
1054 * Use a seqcount to repeat the lockless check
1055 * in wake_userfault() to avoid missing
1056 * wakeups because during the refile both
1057 * waitqueue could become empty if this is the
1060 write_seqcount_begin(&ctx->refile_seq);
1063 * The fault_pending_wqh.lock prevents the uwq
1064 * to disappear from under us.
1066 * Refile this userfault from
1067 * fault_pending_wqh to fault_wqh, it's not
1068 * pending anymore after we read it.
1070 * Use list_del() by hand (as
1071 * userfaultfd_wake_function also uses
1072 * list_del_init() by hand) to be sure nobody
1073 * changes __remove_wait_queue() to use
1074 * list_del_init() in turn breaking the
1075 * !list_empty_careful() check in
1076 * handle_userfault(). The uwq->wq.head list
1077 * must never be empty at any time during the
1078 * refile, or the waitqueue could disappear
1079 * from under us. The "wait_queue_head_t"
1080 * parameter of __remove_wait_queue() is unused
1083 list_del(&uwq->wq.entry);
1084 add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1086 write_seqcount_end(&ctx->refile_seq);
1088 /* careful to always initialize msg if ret == 0 */
1090 spin_unlock(&ctx->fault_pending_wqh.lock);
1094 spin_unlock(&ctx->fault_pending_wqh.lock);
1096 spin_lock(&ctx->event_wqh.lock);
1097 uwq = find_userfault_evt(ctx);
1101 if (uwq->msg.event == UFFD_EVENT_FORK) {
1102 fork_nctx = (struct userfaultfd_ctx *)
1104 uwq->msg.arg.reserved.reserved1;
1105 list_move(&uwq->wq.entry, &fork_event);
1107 * fork_nctx can be freed as soon as
1108 * we drop the lock, unless we take a
1111 userfaultfd_ctx_get(fork_nctx);
1112 spin_unlock(&ctx->event_wqh.lock);
1117 userfaultfd_event_complete(ctx, uwq);
1118 spin_unlock(&ctx->event_wqh.lock);
1122 spin_unlock(&ctx->event_wqh.lock);
1124 if (signal_pending(current)) {
1132 spin_unlock_irq(&ctx->fd_wqh.lock);
1134 spin_lock_irq(&ctx->fd_wqh.lock);
1136 __remove_wait_queue(&ctx->fd_wqh, &wait);
1137 __set_current_state(TASK_RUNNING);
1138 spin_unlock_irq(&ctx->fd_wqh.lock);
1140 if (!ret && msg->event == UFFD_EVENT_FORK) {
1141 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1142 spin_lock_irq(&ctx->event_wqh.lock);
1143 if (!list_empty(&fork_event)) {
1145 * The fork thread didn't abort, so we can
1146 * drop the temporary refcount.
1148 userfaultfd_ctx_put(fork_nctx);
1150 uwq = list_first_entry(&fork_event,
1154 * If fork_event list wasn't empty and in turn
1155 * the event wasn't already released by fork
1156 * (the event is allocated on fork kernel
1157 * stack), put the event back to its place in
1158 * the event_wq. fork_event head will be freed
1159 * as soon as we return so the event cannot
1160 * stay queued there no matter the current
1163 list_del(&uwq->wq.entry);
1164 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1167 * Leave the event in the waitqueue and report
1168 * error to userland if we failed to resolve
1169 * the userfault fork.
1172 userfaultfd_event_complete(ctx, uwq);
1175 * Here the fork thread aborted and the
1176 * refcount from the fork thread on fork_nctx
1177 * has already been released. We still hold
1178 * the reference we took before releasing the
1179 * lock above. If resolve_userfault_fork
1180 * failed we've to drop it because the
1181 * fork_nctx has to be freed in such case. If
1182 * it succeeded we'll hold it because the new
1183 * uffd references it.
1186 userfaultfd_ctx_put(fork_nctx);
1188 spin_unlock_irq(&ctx->event_wqh.lock);
1194 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1195 size_t count, loff_t *ppos)
1197 struct userfaultfd_ctx *ctx = file->private_data;
1198 ssize_t _ret, ret = 0;
1199 struct uffd_msg msg;
1200 int no_wait = file->f_flags & O_NONBLOCK;
1202 if (ctx->state == UFFD_STATE_WAIT_API)
1206 if (count < sizeof(msg))
1207 return ret ? ret : -EINVAL;
1208 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1210 return ret ? ret : _ret;
1211 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1212 return ret ? ret : -EFAULT;
1215 count -= sizeof(msg);
1217 * Allow to read more than one fault at time but only
1218 * block if waiting for the very first one.
1220 no_wait = O_NONBLOCK;
1224 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1225 struct userfaultfd_wake_range *range)
1227 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1228 /* wake all in the range and autoremove */
1229 if (waitqueue_active(&ctx->fault_pending_wqh))
1230 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1232 if (waitqueue_active(&ctx->fault_wqh))
1233 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1234 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1237 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1238 struct userfaultfd_wake_range *range)
1244 * To be sure waitqueue_active() is not reordered by the CPU
1245 * before the pagetable update, use an explicit SMP memory
1246 * barrier here. PT lock release or up_read(mmap_sem) still
1247 * have release semantics that can allow the
1248 * waitqueue_active() to be reordered before the pte update.
1253 * Use waitqueue_active because it's very frequent to
1254 * change the address space atomically even if there are no
1255 * userfaults yet. So we take the spinlock only when we're
1256 * sure we've userfaults to wake.
1259 seq = read_seqcount_begin(&ctx->refile_seq);
1260 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1261 waitqueue_active(&ctx->fault_wqh);
1263 } while (read_seqcount_retry(&ctx->refile_seq, seq));
1265 __wake_userfault(ctx, range);
1268 static __always_inline int validate_range(struct mm_struct *mm,
1269 __u64 *start, __u64 len)
1271 __u64 task_size = mm->task_size;
1273 *start = untagged_addr(*start);
1275 if (*start & ~PAGE_MASK)
1277 if (len & ~PAGE_MASK)
1281 if (*start < mmap_min_addr)
1283 if (*start >= task_size)
1285 if (len > task_size - *start)
1290 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1292 return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1296 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1299 struct mm_struct *mm = ctx->mm;
1300 struct vm_area_struct *vma, *prev, *cur;
1302 struct uffdio_register uffdio_register;
1303 struct uffdio_register __user *user_uffdio_register;
1304 unsigned long vm_flags, new_flags;
1307 unsigned long start, end, vma_end;
1309 user_uffdio_register = (struct uffdio_register __user *) arg;
1312 if (copy_from_user(&uffdio_register, user_uffdio_register,
1313 sizeof(uffdio_register)-sizeof(__u64)))
1317 if (!uffdio_register.mode)
1319 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1320 UFFDIO_REGISTER_MODE_WP))
1323 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1324 vm_flags |= VM_UFFD_MISSING;
1325 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1326 vm_flags |= VM_UFFD_WP;
1328 * FIXME: remove the below error constraint by
1329 * implementing the wprotect tracking mode.
1335 ret = validate_range(mm, &uffdio_register.range.start,
1336 uffdio_register.range.len);
1340 start = uffdio_register.range.start;
1341 end = start + uffdio_register.range.len;
1344 if (!mmget_not_zero(mm))
1347 down_write(&mm->mmap_sem);
1348 if (!mmget_still_valid(mm))
1350 vma = find_vma_prev(mm, start, &prev);
1354 /* check that there's at least one vma in the range */
1356 if (vma->vm_start >= end)
1360 * If the first vma contains huge pages, make sure start address
1361 * is aligned to huge page size.
1363 if (is_vm_hugetlb_page(vma)) {
1364 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1366 if (start & (vma_hpagesize - 1))
1371 * Search for not compatible vmas.
1374 basic_ioctls = false;
1375 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1378 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1379 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1381 /* check not compatible vmas */
1383 if (!vma_can_userfault(cur))
1387 * UFFDIO_COPY will fill file holes even without
1388 * PROT_WRITE. This check enforces that if this is a
1389 * MAP_SHARED, the process has write permission to the backing
1390 * file. If VM_MAYWRITE is set it also enforces that on a
1391 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1392 * F_WRITE_SEAL can be taken until the vma is destroyed.
1395 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1399 * If this vma contains ending address, and huge pages
1402 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1403 end > cur->vm_start) {
1404 unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1408 if (end & (vma_hpagesize - 1))
1413 * Check that this vma isn't already owned by a
1414 * different userfaultfd. We can't allow more than one
1415 * userfaultfd to own a single vma simultaneously or we
1416 * wouldn't know which one to deliver the userfaults to.
1419 if (cur->vm_userfaultfd_ctx.ctx &&
1420 cur->vm_userfaultfd_ctx.ctx != ctx)
1424 * Note vmas containing huge pages
1426 if (is_vm_hugetlb_page(cur))
1427 basic_ioctls = true;
1433 if (vma->vm_start < start)
1440 BUG_ON(!vma_can_userfault(vma));
1441 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1442 vma->vm_userfaultfd_ctx.ctx != ctx);
1443 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1446 * Nothing to do: this vma is already registered into this
1447 * userfaultfd and with the right tracking mode too.
1449 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1450 (vma->vm_flags & vm_flags) == vm_flags)
1453 if (vma->vm_start > start)
1454 start = vma->vm_start;
1455 vma_end = min(end, vma->vm_end);
1457 new_flags = (vma->vm_flags &
1458 ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags;
1459 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1460 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1462 ((struct vm_userfaultfd_ctx){ ctx }));
1467 if (vma->vm_start < start) {
1468 ret = split_vma(mm, vma, start, 1);
1472 if (vma->vm_end > end) {
1473 ret = split_vma(mm, vma, end, 0);
1479 * In the vma_merge() successful mprotect-like case 8:
1480 * the next vma was merged into the current one and
1481 * the current one has not been updated yet.
1483 vma->vm_flags = new_flags;
1484 vma->vm_userfaultfd_ctx.ctx = ctx;
1488 start = vma->vm_end;
1490 } while (vma && vma->vm_start < end);
1492 up_write(&mm->mmap_sem);
1496 * Now that we scanned all vmas we can already tell
1497 * userland which ioctls methods are guaranteed to
1498 * succeed on this range.
1500 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1501 UFFD_API_RANGE_IOCTLS,
1502 &user_uffdio_register->ioctls))
1509 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1512 struct mm_struct *mm = ctx->mm;
1513 struct vm_area_struct *vma, *prev, *cur;
1515 struct uffdio_range uffdio_unregister;
1516 unsigned long new_flags;
1518 unsigned long start, end, vma_end;
1519 const void __user *buf = (void __user *)arg;
1522 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1525 ret = validate_range(mm, &uffdio_unregister.start,
1526 uffdio_unregister.len);
1530 start = uffdio_unregister.start;
1531 end = start + uffdio_unregister.len;
1534 if (!mmget_not_zero(mm))
1537 down_write(&mm->mmap_sem);
1538 if (!mmget_still_valid(mm))
1540 vma = find_vma_prev(mm, start, &prev);
1544 /* check that there's at least one vma in the range */
1546 if (vma->vm_start >= 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.
1565 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1568 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1569 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
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))
1585 if (vma->vm_start < start)
1592 BUG_ON(!vma_can_userfault(vma));
1595 * Nothing to do: this vma is already registered into this
1596 * userfaultfd and with the right tracking mode too.
1598 if (!vma->vm_userfaultfd_ctx.ctx)
1601 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1603 if (vma->vm_start > start)
1604 start = vma->vm_start;
1605 vma_end = min(end, vma->vm_end);
1607 if (userfaultfd_missing(vma)) {
1609 * Wake any concurrent pending userfault while
1610 * we unregister, so they will not hang
1611 * permanently and it avoids userland to call
1612 * UFFDIO_WAKE explicitly.
1614 struct userfaultfd_wake_range range;
1615 range.start = start;
1616 range.len = vma_end - start;
1617 wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1620 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1621 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1622 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1629 if (vma->vm_start < start) {
1630 ret = split_vma(mm, vma, start, 1);
1634 if (vma->vm_end > end) {
1635 ret = split_vma(mm, vma, end, 0);
1641 * In the vma_merge() successful mprotect-like case 8:
1642 * the next vma was merged into the current one and
1643 * the current one has not been updated yet.
1645 vma->vm_flags = new_flags;
1646 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1650 start = vma->vm_end;
1652 } while (vma && vma->vm_start < end);
1654 up_write(&mm->mmap_sem);
1661 * userfaultfd_wake may be used in combination with the
1662 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1664 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1668 struct uffdio_range uffdio_wake;
1669 struct userfaultfd_wake_range range;
1670 const void __user *buf = (void __user *)arg;
1673 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1676 ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len);
1680 range.start = uffdio_wake.start;
1681 range.len = uffdio_wake.len;
1684 * len == 0 means wake all and we don't want to wake all here,
1685 * so check it again to be sure.
1687 VM_BUG_ON(!range.len);
1689 wake_userfault(ctx, &range);
1696 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1700 struct uffdio_copy uffdio_copy;
1701 struct uffdio_copy __user *user_uffdio_copy;
1702 struct userfaultfd_wake_range range;
1704 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1707 if (READ_ONCE(ctx->mmap_changing))
1711 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1712 /* don't copy "copy" last field */
1713 sizeof(uffdio_copy)-sizeof(__s64)))
1716 ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len);
1720 * double check for wraparound just in case. copy_from_user()
1721 * will later check uffdio_copy.src + uffdio_copy.len to fit
1722 * in the userland range.
1725 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1727 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1729 if (mmget_not_zero(ctx->mm)) {
1730 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1731 uffdio_copy.len, &ctx->mmap_changing);
1736 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1741 /* len == 0 would wake all */
1743 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1744 range.start = uffdio_copy.dst;
1745 wake_userfault(ctx, &range);
1747 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1752 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1756 struct uffdio_zeropage uffdio_zeropage;
1757 struct uffdio_zeropage __user *user_uffdio_zeropage;
1758 struct userfaultfd_wake_range range;
1760 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1763 if (READ_ONCE(ctx->mmap_changing))
1767 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1768 /* don't copy "zeropage" last field */
1769 sizeof(uffdio_zeropage)-sizeof(__s64)))
1772 ret = validate_range(ctx->mm, &uffdio_zeropage.range.start,
1773 uffdio_zeropage.range.len);
1777 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1780 if (mmget_not_zero(ctx->mm)) {
1781 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1782 uffdio_zeropage.range.len,
1783 &ctx->mmap_changing);
1788 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1792 /* len == 0 would wake all */
1795 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1796 range.start = uffdio_zeropage.range.start;
1797 wake_userfault(ctx, &range);
1799 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1804 static inline unsigned int uffd_ctx_features(__u64 user_features)
1807 * For the current set of features the bits just coincide
1809 return (unsigned int)user_features;
1813 * userland asks for a certain API version and we return which bits
1814 * and ioctl commands are implemented in this kernel for such API
1815 * version or -EINVAL if unknown.
1817 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1820 struct uffdio_api uffdio_api;
1821 void __user *buf = (void __user *)arg;
1826 if (ctx->state != UFFD_STATE_WAIT_API)
1829 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1831 features = uffdio_api.features;
1833 if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1836 if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1838 /* report all available features and ioctls to userland */
1839 uffdio_api.features = UFFD_API_FEATURES;
1840 uffdio_api.ioctls = UFFD_API_IOCTLS;
1842 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1844 ctx->state = UFFD_STATE_RUNNING;
1845 /* only enable the requested features for this uffd context */
1846 ctx->features = uffd_ctx_features(features);
1851 memset(&uffdio_api, 0, sizeof(uffdio_api));
1852 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1857 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1861 struct userfaultfd_ctx *ctx = file->private_data;
1863 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1868 ret = userfaultfd_api(ctx, arg);
1870 case UFFDIO_REGISTER:
1871 ret = userfaultfd_register(ctx, arg);
1873 case UFFDIO_UNREGISTER:
1874 ret = userfaultfd_unregister(ctx, arg);
1877 ret = userfaultfd_wake(ctx, arg);
1880 ret = userfaultfd_copy(ctx, arg);
1882 case UFFDIO_ZEROPAGE:
1883 ret = userfaultfd_zeropage(ctx, arg);
1889 #ifdef CONFIG_PROC_FS
1890 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1892 struct userfaultfd_ctx *ctx = f->private_data;
1893 wait_queue_entry_t *wq;
1894 unsigned long pending = 0, total = 0;
1896 spin_lock_irq(&ctx->fault_pending_wqh.lock);
1897 list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1901 list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1904 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1907 * If more protocols will be added, there will be all shown
1908 * separated by a space. Like this:
1909 * protocols: aa:... bb:...
1911 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1912 pending, total, UFFD_API, ctx->features,
1913 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1917 static const struct file_operations userfaultfd_fops = {
1918 #ifdef CONFIG_PROC_FS
1919 .show_fdinfo = userfaultfd_show_fdinfo,
1921 .release = userfaultfd_release,
1922 .poll = userfaultfd_poll,
1923 .read = userfaultfd_read,
1924 .unlocked_ioctl = userfaultfd_ioctl,
1925 .compat_ioctl = compat_ptr_ioctl,
1926 .llseek = noop_llseek,
1929 static void init_once_userfaultfd_ctx(void *mem)
1931 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1933 init_waitqueue_head(&ctx->fault_pending_wqh);
1934 init_waitqueue_head(&ctx->fault_wqh);
1935 init_waitqueue_head(&ctx->event_wqh);
1936 init_waitqueue_head(&ctx->fd_wqh);
1937 seqcount_init(&ctx->refile_seq);
1940 SYSCALL_DEFINE1(userfaultfd, int, flags)
1942 struct userfaultfd_ctx *ctx;
1945 if (!sysctl_unprivileged_userfaultfd && !capable(CAP_SYS_PTRACE))
1948 BUG_ON(!current->mm);
1950 /* Check the UFFD_* constants for consistency. */
1951 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1952 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1954 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1957 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1961 refcount_set(&ctx->refcount, 1);
1964 ctx->state = UFFD_STATE_WAIT_API;
1965 ctx->released = false;
1966 ctx->mmap_changing = false;
1967 ctx->mm = current->mm;
1968 /* prevent the mm struct to be freed */
1971 fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
1972 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1975 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1980 static int __init userfaultfd_init(void)
1982 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1983 sizeof(struct userfaultfd_ctx),
1985 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1986 init_once_userfaultfd_ctx);
1989 __initcall(userfaultfd_init);