4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
5 * Copyright (C) 2008-2009 Red Hat, Inc.
6 * Copyright (C) 2015 Red Hat, Inc.
8 * This work is licensed under the terms of the GNU GPL, version 2. See
9 * the COPYING file in the top-level directory.
11 * Some part derived from fs/eventfd.c (anon inode setup) and
12 * mm/ksm.c (mm hashing).
15 #include <linux/hashtable.h>
16 #include <linux/sched.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>
30 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
32 enum userfaultfd_state {
38 * Start with fault_pending_wqh and fault_wqh so they're more likely
39 * to be in the same cacheline.
41 struct userfaultfd_ctx {
42 /* waitqueue head for the pending (i.e. not read) userfaults */
43 wait_queue_head_t fault_pending_wqh;
44 /* waitqueue head for the userfaults */
45 wait_queue_head_t fault_wqh;
46 /* waitqueue head for the pseudo fd to wakeup poll/read */
47 wait_queue_head_t fd_wqh;
48 /* a refile sequence protected by fault_pending_wqh lock */
49 struct seqcount refile_seq;
50 /* pseudo fd refcounting */
52 /* userfaultfd syscall flags */
55 enum userfaultfd_state state;
58 /* mm with one ore more vmas attached to this userfaultfd_ctx */
62 struct userfaultfd_wait_queue {
65 struct userfaultfd_ctx *ctx;
68 struct userfaultfd_wake_range {
73 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
74 int wake_flags, void *key)
76 struct userfaultfd_wake_range *range = key;
78 struct userfaultfd_wait_queue *uwq;
79 unsigned long start, len;
81 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
83 /* len == 0 means wake all */
86 if (len && (start > uwq->msg.arg.pagefault.address ||
87 start + len <= uwq->msg.arg.pagefault.address))
89 ret = wake_up_state(wq->private, mode);
92 * Wake only once, autoremove behavior.
94 * After the effect of list_del_init is visible to the
95 * other CPUs, the waitqueue may disappear from under
96 * us, see the !list_empty_careful() in
97 * handle_userfault(). try_to_wake_up() has an
98 * implicit smp_mb__before_spinlock, and the
99 * wq->private is read before calling the extern
100 * function "wake_up_state" (which in turns calls
101 * try_to_wake_up). While the spin_lock;spin_unlock;
102 * wouldn't be enough, the smp_mb__before_spinlock is
103 * enough to avoid an explicit smp_mb() here.
105 list_del_init(&wq->task_list);
111 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
113 * @ctx: [in] Pointer to the userfaultfd context.
115 * Returns: In case of success, returns not zero.
117 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
119 if (!atomic_inc_not_zero(&ctx->refcount))
124 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
126 * @ctx: [in] Pointer to userfaultfd context.
128 * The userfaultfd context reference must have been previously acquired either
129 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
131 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
133 if (atomic_dec_and_test(&ctx->refcount)) {
134 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
135 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
136 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
137 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
138 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
139 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
141 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
145 static inline void msg_init(struct uffd_msg *msg)
147 BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
149 * Must use memset to zero out the paddings or kernel data is
150 * leaked to userland.
152 memset(msg, 0, sizeof(struct uffd_msg));
155 static inline struct uffd_msg userfault_msg(unsigned long address,
157 unsigned long reason)
161 msg.event = UFFD_EVENT_PAGEFAULT;
162 msg.arg.pagefault.address = address;
163 if (flags & FAULT_FLAG_WRITE)
165 * If UFFD_FEATURE_PAGEFAULT_FLAG_WRITE was set in the
166 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
167 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
168 * was a read fault, otherwise if set it means it's
171 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
172 if (reason & VM_UFFD_WP)
174 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
175 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
176 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
177 * a missing fault, otherwise if set it means it's a
178 * write protect fault.
180 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
185 * Verify the pagetables are still not ok after having reigstered into
186 * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
187 * userfault that has already been resolved, if userfaultfd_read and
188 * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
191 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
192 unsigned long address,
194 unsigned long reason)
196 struct mm_struct *mm = ctx->mm;
203 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
205 pgd = pgd_offset(mm, address);
206 if (!pgd_present(*pgd))
208 pud = pud_offset(pgd, address);
209 if (!pud_present(*pud))
211 pmd = pmd_offset(pud, address);
213 * READ_ONCE must function as a barrier with narrower scope
214 * and it must be equivalent to:
215 * _pmd = *pmd; barrier();
217 * This is to deal with the instability (as in
218 * pmd_trans_unstable) of the pmd.
220 _pmd = READ_ONCE(*pmd);
221 if (!pmd_present(_pmd))
225 if (pmd_trans_huge(_pmd))
229 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
230 * and use the standard pte_offset_map() instead of parsing _pmd.
232 pte = pte_offset_map(pmd, address);
234 * Lockless access: we're in a wait_event so it's ok if it
246 * The locking rules involved in returning VM_FAULT_RETRY depending on
247 * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
248 * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
249 * recommendation in __lock_page_or_retry is not an understatement.
251 * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
252 * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
255 * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
256 * set, VM_FAULT_RETRY can still be returned if and only if there are
257 * fatal_signal_pending()s, and the mmap_sem must be released before
260 int handle_userfault(struct vm_area_struct *vma, unsigned long address,
261 unsigned int flags, unsigned long reason)
263 struct mm_struct *mm = vma->vm_mm;
264 struct userfaultfd_ctx *ctx;
265 struct userfaultfd_wait_queue uwq;
267 bool must_wait, return_to_userland;
269 BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
271 ret = VM_FAULT_SIGBUS;
272 ctx = vma->vm_userfaultfd_ctx.ctx;
276 BUG_ON(ctx->mm != mm);
278 VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
279 VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
282 * If it's already released don't get it. This avoids to loop
283 * in __get_user_pages if userfaultfd_release waits on the
284 * caller of handle_userfault to release the mmap_sem.
286 if (unlikely(ACCESS_ONCE(ctx->released)))
290 * We don't do userfault handling for the final child pid update.
292 if (current->flags & PF_EXITING)
296 * Check that we can return VM_FAULT_RETRY.
298 * NOTE: it should become possible to return VM_FAULT_RETRY
299 * even if FAULT_FLAG_TRIED is set without leading to gup()
300 * -EBUSY failures, if the userfaultfd is to be extended for
301 * VM_UFFD_WP tracking and we intend to arm the userfault
302 * without first stopping userland access to the memory. For
303 * VM_UFFD_MISSING userfaults this is enough for now.
305 if (unlikely(!(flags & FAULT_FLAG_ALLOW_RETRY))) {
307 * Validate the invariant that nowait must allow retry
308 * to be sure not to return SIGBUS erroneously on
309 * nowait invocations.
311 BUG_ON(flags & FAULT_FLAG_RETRY_NOWAIT);
312 #ifdef CONFIG_DEBUG_VM
313 if (printk_ratelimit()) {
315 "FAULT_FLAG_ALLOW_RETRY missing %x\n", flags);
323 * Handle nowait, not much to do other than tell it to retry
326 ret = VM_FAULT_RETRY;
327 if (flags & FAULT_FLAG_RETRY_NOWAIT)
330 /* take the reference before dropping the mmap_sem */
331 userfaultfd_ctx_get(ctx);
333 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
334 uwq.wq.private = current;
335 uwq.msg = userfault_msg(address, flags, reason);
338 return_to_userland = (flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
339 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
341 spin_lock(&ctx->fault_pending_wqh.lock);
343 * After the __add_wait_queue the uwq is visible to userland
344 * through poll/read().
346 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
348 * The smp_mb() after __set_current_state prevents the reads
349 * following the spin_unlock to happen before the list_add in
352 set_current_state(return_to_userland ? TASK_INTERRUPTIBLE :
354 spin_unlock(&ctx->fault_pending_wqh.lock);
356 must_wait = userfaultfd_must_wait(ctx, address, flags, reason);
357 up_read(&mm->mmap_sem);
359 if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
360 (return_to_userland ? !signal_pending(current) :
361 !fatal_signal_pending(current)))) {
362 wake_up_poll(&ctx->fd_wqh, POLLIN);
364 ret |= VM_FAULT_MAJOR;
367 __set_current_state(TASK_RUNNING);
369 if (return_to_userland) {
370 if (signal_pending(current) &&
371 !fatal_signal_pending(current)) {
373 * If we got a SIGSTOP or SIGCONT and this is
374 * a normal userland page fault, just let
375 * userland return so the signal will be
376 * handled and gdb debugging works. The page
377 * fault code immediately after we return from
378 * this function is going to release the
379 * mmap_sem and it's not depending on it
380 * (unlike gup would if we were not to return
383 * If a fatal signal is pending we still take
384 * the streamlined VM_FAULT_RETRY failure path
385 * and there's no need to retake the mmap_sem
388 down_read(&mm->mmap_sem);
389 ret = VM_FAULT_NOPAGE;
394 * Here we race with the list_del; list_add in
395 * userfaultfd_ctx_read(), however because we don't ever run
396 * list_del_init() to refile across the two lists, the prev
397 * and next pointers will never point to self. list_add also
398 * would never let any of the two pointers to point to
399 * self. So list_empty_careful won't risk to see both pointers
400 * pointing to self at any time during the list refile. The
401 * only case where list_del_init() is called is the full
402 * removal in the wake function and there we don't re-list_add
403 * and it's fine not to block on the spinlock. The uwq on this
404 * kernel stack can be released after the list_del_init.
406 if (!list_empty_careful(&uwq.wq.task_list)) {
407 spin_lock(&ctx->fault_pending_wqh.lock);
409 * No need of list_del_init(), the uwq on the stack
410 * will be freed shortly anyway.
412 list_del(&uwq.wq.task_list);
413 spin_unlock(&ctx->fault_pending_wqh.lock);
417 * ctx may go away after this if the userfault pseudo fd is
420 userfaultfd_ctx_put(ctx);
426 static int userfaultfd_release(struct inode *inode, struct file *file)
428 struct userfaultfd_ctx *ctx = file->private_data;
429 struct mm_struct *mm = ctx->mm;
430 struct vm_area_struct *vma, *prev;
431 /* len == 0 means wake all */
432 struct userfaultfd_wake_range range = { .len = 0, };
433 unsigned long new_flags;
436 ACCESS_ONCE(ctx->released) = true;
438 if (!mmget_not_zero(mm))
442 * Flush page faults out of all CPUs. NOTE: all page faults
443 * must be retried without returning VM_FAULT_SIGBUS if
444 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
445 * changes while handle_userfault released the mmap_sem. So
446 * it's critical that released is set to true (above), before
447 * taking the mmap_sem for writing.
449 down_write(&mm->mmap_sem);
450 still_valid = mmget_still_valid(mm);
452 for (vma = mm->mmap; vma; vma = vma->vm_next) {
454 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
455 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
456 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
460 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
462 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
463 new_flags, vma->anon_vma,
464 vma->vm_file, vma->vm_pgoff,
467 vma_get_anon_name(vma));
473 vma->vm_flags = new_flags;
474 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
476 up_write(&mm->mmap_sem);
480 * After no new page faults can wait on this fault_*wqh, flush
481 * the last page faults that may have been already waiting on
484 spin_lock(&ctx->fault_pending_wqh.lock);
485 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
486 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
487 spin_unlock(&ctx->fault_pending_wqh.lock);
489 wake_up_poll(&ctx->fd_wqh, POLLHUP);
490 userfaultfd_ctx_put(ctx);
494 /* fault_pending_wqh.lock must be hold by the caller */
495 static inline struct userfaultfd_wait_queue *find_userfault(
496 struct userfaultfd_ctx *ctx)
499 struct userfaultfd_wait_queue *uwq;
501 VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
504 if (!waitqueue_active(&ctx->fault_pending_wqh))
506 /* walk in reverse to provide FIFO behavior to read userfaults */
507 wq = list_last_entry(&ctx->fault_pending_wqh.task_list,
508 typeof(*wq), task_list);
509 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
514 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
516 struct userfaultfd_ctx *ctx = file->private_data;
519 poll_wait(file, &ctx->fd_wqh, wait);
521 switch (ctx->state) {
522 case UFFD_STATE_WAIT_API:
524 case UFFD_STATE_RUNNING:
526 * poll() never guarantees that read won't block.
527 * userfaults can be waken before they're read().
529 if (unlikely(!(file->f_flags & O_NONBLOCK)))
532 * lockless access to see if there are pending faults
533 * __pollwait last action is the add_wait_queue but
534 * the spin_unlock would allow the waitqueue_active to
535 * pass above the actual list_add inside
536 * add_wait_queue critical section. So use a full
537 * memory barrier to serialize the list_add write of
538 * add_wait_queue() with the waitqueue_active read
543 if (waitqueue_active(&ctx->fault_pending_wqh))
551 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
552 struct uffd_msg *msg)
555 DECLARE_WAITQUEUE(wait, current);
556 struct userfaultfd_wait_queue *uwq;
558 /* always take the fd_wqh lock before the fault_pending_wqh lock */
559 spin_lock(&ctx->fd_wqh.lock);
560 __add_wait_queue(&ctx->fd_wqh, &wait);
562 set_current_state(TASK_INTERRUPTIBLE);
563 spin_lock(&ctx->fault_pending_wqh.lock);
564 uwq = find_userfault(ctx);
567 * Use a seqcount to repeat the lockless check
568 * in wake_userfault() to avoid missing
569 * wakeups because during the refile both
570 * waitqueue could become empty if this is the
573 write_seqcount_begin(&ctx->refile_seq);
576 * The fault_pending_wqh.lock prevents the uwq
577 * to disappear from under us.
579 * Refile this userfault from
580 * fault_pending_wqh to fault_wqh, it's not
581 * pending anymore after we read it.
583 * Use list_del() by hand (as
584 * userfaultfd_wake_function also uses
585 * list_del_init() by hand) to be sure nobody
586 * changes __remove_wait_queue() to use
587 * list_del_init() in turn breaking the
588 * !list_empty_careful() check in
589 * handle_userfault(). The uwq->wq.task_list
590 * must never be empty at any time during the
591 * refile, or the waitqueue could disappear
592 * from under us. The "wait_queue_head_t"
593 * parameter of __remove_wait_queue() is unused
596 list_del(&uwq->wq.task_list);
597 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
599 write_seqcount_end(&ctx->refile_seq);
601 /* careful to always initialize msg if ret == 0 */
603 spin_unlock(&ctx->fault_pending_wqh.lock);
607 spin_unlock(&ctx->fault_pending_wqh.lock);
608 if (signal_pending(current)) {
616 spin_unlock(&ctx->fd_wqh.lock);
618 spin_lock(&ctx->fd_wqh.lock);
620 __remove_wait_queue(&ctx->fd_wqh, &wait);
621 __set_current_state(TASK_RUNNING);
622 spin_unlock(&ctx->fd_wqh.lock);
627 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
628 size_t count, loff_t *ppos)
630 struct userfaultfd_ctx *ctx = file->private_data;
631 ssize_t _ret, ret = 0;
633 int no_wait = file->f_flags & O_NONBLOCK;
635 if (ctx->state == UFFD_STATE_WAIT_API)
639 if (count < sizeof(msg))
640 return ret ? ret : -EINVAL;
641 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
643 return ret ? ret : _ret;
644 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
645 return ret ? ret : -EFAULT;
648 count -= sizeof(msg);
650 * Allow to read more than one fault at time but only
651 * block if waiting for the very first one.
653 no_wait = O_NONBLOCK;
657 static void __wake_userfault(struct userfaultfd_ctx *ctx,
658 struct userfaultfd_wake_range *range)
660 unsigned long start, end;
662 start = range->start;
663 end = range->start + range->len;
665 spin_lock(&ctx->fault_pending_wqh.lock);
666 /* wake all in the range and autoremove */
667 if (waitqueue_active(&ctx->fault_pending_wqh))
668 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
670 if (waitqueue_active(&ctx->fault_wqh))
671 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
672 spin_unlock(&ctx->fault_pending_wqh.lock);
675 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
676 struct userfaultfd_wake_range *range)
682 * To be sure waitqueue_active() is not reordered by the CPU
683 * before the pagetable update, use an explicit SMP memory
684 * barrier here. PT lock release or up_read(mmap_sem) still
685 * have release semantics that can allow the
686 * waitqueue_active() to be reordered before the pte update.
691 * Use waitqueue_active because it's very frequent to
692 * change the address space atomically even if there are no
693 * userfaults yet. So we take the spinlock only when we're
694 * sure we've userfaults to wake.
697 seq = read_seqcount_begin(&ctx->refile_seq);
698 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
699 waitqueue_active(&ctx->fault_wqh);
701 } while (read_seqcount_retry(&ctx->refile_seq, seq));
703 __wake_userfault(ctx, range);
706 static __always_inline int validate_range(struct mm_struct *mm,
707 __u64 start, __u64 len)
709 __u64 task_size = mm->task_size;
711 if (start & ~PAGE_MASK)
713 if (len & ~PAGE_MASK)
717 if (start < mmap_min_addr)
719 if (start >= task_size)
721 if (len > task_size - start)
726 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
729 struct mm_struct *mm = ctx->mm;
730 struct vm_area_struct *vma, *prev, *cur;
732 struct uffdio_register uffdio_register;
733 struct uffdio_register __user *user_uffdio_register;
734 unsigned long vm_flags, new_flags;
736 unsigned long start, end, vma_end;
738 user_uffdio_register = (struct uffdio_register __user *) arg;
741 if (copy_from_user(&uffdio_register, user_uffdio_register,
742 sizeof(uffdio_register)-sizeof(__u64)))
746 if (!uffdio_register.mode)
748 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
749 UFFDIO_REGISTER_MODE_WP))
752 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
753 vm_flags |= VM_UFFD_MISSING;
754 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
755 vm_flags |= VM_UFFD_WP;
757 * FIXME: remove the below error constraint by
758 * implementing the wprotect tracking mode.
764 ret = validate_range(mm, uffdio_register.range.start,
765 uffdio_register.range.len);
769 start = uffdio_register.range.start;
770 end = start + uffdio_register.range.len;
773 if (!mmget_not_zero(mm))
776 down_write(&mm->mmap_sem);
777 if (!mmget_still_valid(mm))
779 vma = find_vma_prev(mm, start, &prev);
783 /* check that there's at least one vma in the range */
785 if (vma->vm_start >= end)
789 * Search for not compatible vmas.
791 * FIXME: this shall be relaxed later so that it doesn't fail
792 * on tmpfs backed vmas (in addition to the current allowance
793 * on anonymous vmas).
796 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
799 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
800 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
802 /* check not compatible vmas */
808 * UFFDIO_COPY will fill file holes even without
809 * PROT_WRITE. This check enforces that if this is a
810 * MAP_SHARED, the process has write permission to the backing
811 * file. If VM_MAYWRITE is set it also enforces that on a
812 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
813 * F_WRITE_SEAL can be taken until the vma is destroyed.
816 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
820 * Check that this vma isn't already owned by a
821 * different userfaultfd. We can't allow more than one
822 * userfaultfd to own a single vma simultaneously or we
823 * wouldn't know which one to deliver the userfaults to.
826 if (cur->vm_userfaultfd_ctx.ctx &&
827 cur->vm_userfaultfd_ctx.ctx != ctx)
834 if (vma->vm_start < start)
842 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
843 vma->vm_userfaultfd_ctx.ctx != ctx);
844 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
847 * Nothing to do: this vma is already registered into this
848 * userfaultfd and with the right tracking mode too.
850 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
851 (vma->vm_flags & vm_flags) == vm_flags)
854 if (vma->vm_start > start)
855 start = vma->vm_start;
856 vma_end = min(end, vma->vm_end);
858 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
859 prev = vma_merge(mm, prev, start, vma_end, new_flags,
860 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
862 ((struct vm_userfaultfd_ctx){ ctx }),
863 vma_get_anon_name(vma));
868 if (vma->vm_start < start) {
869 ret = split_vma(mm, vma, start, 1);
873 if (vma->vm_end > end) {
874 ret = split_vma(mm, vma, end, 0);
880 * In the vma_merge() successful mprotect-like case 8:
881 * the next vma was merged into the current one and
882 * the current one has not been updated yet.
884 vma->vm_flags = new_flags;
885 vma->vm_userfaultfd_ctx.ctx = ctx;
891 } while (vma && vma->vm_start < end);
893 up_write(&mm->mmap_sem);
897 * Now that we scanned all vmas we can already tell
898 * userland which ioctls methods are guaranteed to
899 * succeed on this range.
901 if (put_user(UFFD_API_RANGE_IOCTLS,
902 &user_uffdio_register->ioctls))
909 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
912 struct mm_struct *mm = ctx->mm;
913 struct vm_area_struct *vma, *prev, *cur;
915 struct uffdio_range uffdio_unregister;
916 unsigned long new_flags;
918 unsigned long start, end, vma_end;
919 const void __user *buf = (void __user *)arg;
922 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
925 ret = validate_range(mm, uffdio_unregister.start,
926 uffdio_unregister.len);
930 start = uffdio_unregister.start;
931 end = start + uffdio_unregister.len;
934 if (!mmget_not_zero(mm))
937 down_write(&mm->mmap_sem);
938 if (!mmget_still_valid(mm))
940 vma = find_vma_prev(mm, start, &prev);
944 /* check that there's at least one vma in the range */
946 if (vma->vm_start >= end)
950 * Search for not compatible vmas.
952 * FIXME: this shall be relaxed later so that it doesn't fail
953 * on tmpfs backed vmas (in addition to the current allowance
954 * on anonymous vmas).
958 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
961 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
962 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
965 * Check not compatible vmas, not strictly required
966 * here as not compatible vmas cannot have an
967 * userfaultfd_ctx registered on them, but this
968 * provides for more strict behavior to notice
969 * unregistration errors.
978 if (vma->vm_start < start)
986 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
989 * Nothing to do: this vma is already registered into this
990 * userfaultfd and with the right tracking mode too.
992 if (!vma->vm_userfaultfd_ctx.ctx)
995 if (vma->vm_start > start)
996 start = vma->vm_start;
997 vma_end = min(end, vma->vm_end);
999 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1000 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1001 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1004 vma_get_anon_name(vma));
1009 if (vma->vm_start < start) {
1010 ret = split_vma(mm, vma, start, 1);
1014 if (vma->vm_end > end) {
1015 ret = split_vma(mm, vma, end, 0);
1021 * In the vma_merge() successful mprotect-like case 8:
1022 * the next vma was merged into the current one and
1023 * the current one has not been updated yet.
1025 vma->vm_flags = new_flags;
1026 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1030 start = vma->vm_end;
1032 } while (vma && vma->vm_start < end);
1034 up_write(&mm->mmap_sem);
1041 * userfaultfd_wake may be used in combination with the
1042 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1044 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1048 struct uffdio_range uffdio_wake;
1049 struct userfaultfd_wake_range range;
1050 const void __user *buf = (void __user *)arg;
1053 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1056 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1060 range.start = uffdio_wake.start;
1061 range.len = uffdio_wake.len;
1064 * len == 0 means wake all and we don't want to wake all here,
1065 * so check it again to be sure.
1067 VM_BUG_ON(!range.len);
1069 wake_userfault(ctx, &range);
1076 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1080 struct uffdio_copy uffdio_copy;
1081 struct uffdio_copy __user *user_uffdio_copy;
1082 struct userfaultfd_wake_range range;
1084 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1087 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1088 /* don't copy "copy" last field */
1089 sizeof(uffdio_copy)-sizeof(__s64)))
1092 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1096 * double check for wraparound just in case. copy_from_user()
1097 * will later check uffdio_copy.src + uffdio_copy.len to fit
1098 * in the userland range.
1101 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1103 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1105 if (mmget_not_zero(ctx->mm)) {
1106 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1110 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1115 /* len == 0 would wake all */
1117 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1118 range.start = uffdio_copy.dst;
1119 wake_userfault(ctx, &range);
1121 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1126 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1130 struct uffdio_zeropage uffdio_zeropage;
1131 struct uffdio_zeropage __user *user_uffdio_zeropage;
1132 struct userfaultfd_wake_range range;
1134 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1137 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1138 /* don't copy "zeropage" last field */
1139 sizeof(uffdio_zeropage)-sizeof(__s64)))
1142 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1143 uffdio_zeropage.range.len);
1147 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1150 if (mmget_not_zero(ctx->mm)) {
1151 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1152 uffdio_zeropage.range.len);
1155 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1159 /* len == 0 would wake all */
1162 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1163 range.start = uffdio_zeropage.range.start;
1164 wake_userfault(ctx, &range);
1166 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1172 * userland asks for a certain API version and we return which bits
1173 * and ioctl commands are implemented in this kernel for such API
1174 * version or -EINVAL if unknown.
1176 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1179 struct uffdio_api uffdio_api;
1180 void __user *buf = (void __user *)arg;
1184 if (ctx->state != UFFD_STATE_WAIT_API)
1187 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1189 if (uffdio_api.api != UFFD_API || uffdio_api.features) {
1190 memset(&uffdio_api, 0, sizeof(uffdio_api));
1191 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1196 uffdio_api.features = UFFD_API_FEATURES;
1197 uffdio_api.ioctls = UFFD_API_IOCTLS;
1199 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1201 ctx->state = UFFD_STATE_RUNNING;
1207 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1211 struct userfaultfd_ctx *ctx = file->private_data;
1213 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1218 ret = userfaultfd_api(ctx, arg);
1220 case UFFDIO_REGISTER:
1221 ret = userfaultfd_register(ctx, arg);
1223 case UFFDIO_UNREGISTER:
1224 ret = userfaultfd_unregister(ctx, arg);
1227 ret = userfaultfd_wake(ctx, arg);
1230 ret = userfaultfd_copy(ctx, arg);
1232 case UFFDIO_ZEROPAGE:
1233 ret = userfaultfd_zeropage(ctx, arg);
1239 #ifdef CONFIG_PROC_FS
1240 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1242 struct userfaultfd_ctx *ctx = f->private_data;
1244 struct userfaultfd_wait_queue *uwq;
1245 unsigned long pending = 0, total = 0;
1247 spin_lock(&ctx->fault_pending_wqh.lock);
1248 list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1249 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1253 list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1254 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1257 spin_unlock(&ctx->fault_pending_wqh.lock);
1260 * If more protocols will be added, there will be all shown
1261 * separated by a space. Like this:
1262 * protocols: aa:... bb:...
1264 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1265 pending, total, UFFD_API, UFFD_API_FEATURES,
1266 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1270 static const struct file_operations userfaultfd_fops = {
1271 #ifdef CONFIG_PROC_FS
1272 .show_fdinfo = userfaultfd_show_fdinfo,
1274 .release = userfaultfd_release,
1275 .poll = userfaultfd_poll,
1276 .read = userfaultfd_read,
1277 .unlocked_ioctl = userfaultfd_ioctl,
1278 .compat_ioctl = userfaultfd_ioctl,
1279 .llseek = noop_llseek,
1282 static void init_once_userfaultfd_ctx(void *mem)
1284 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1286 init_waitqueue_head(&ctx->fault_pending_wqh);
1287 init_waitqueue_head(&ctx->fault_wqh);
1288 init_waitqueue_head(&ctx->fd_wqh);
1289 seqcount_init(&ctx->refile_seq);
1293 * userfaultfd_file_create - Creates an userfaultfd file pointer.
1294 * @flags: Flags for the userfaultfd file.
1296 * This function creates an userfaultfd file pointer, w/out installing
1297 * it into the fd table. This is useful when the userfaultfd file is
1298 * used during the initialization of data structures that require
1299 * extra setup after the userfaultfd creation. So the userfaultfd
1300 * creation is split into the file pointer creation phase, and the
1301 * file descriptor installation phase. In this way races with
1302 * userspace closing the newly installed file descriptor can be
1303 * avoided. Returns an userfaultfd file pointer, or a proper error
1306 static struct file *userfaultfd_file_create(int flags)
1309 struct userfaultfd_ctx *ctx;
1311 BUG_ON(!current->mm);
1313 /* Check the UFFD_* constants for consistency. */
1314 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1315 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1317 file = ERR_PTR(-EINVAL);
1318 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1321 file = ERR_PTR(-ENOMEM);
1322 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1326 atomic_set(&ctx->refcount, 1);
1328 ctx->state = UFFD_STATE_WAIT_API;
1329 ctx->released = false;
1330 ctx->mm = current->mm;
1331 /* prevent the mm struct to be freed */
1332 atomic_inc(&ctx->mm->mm_count);
1334 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1335 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1338 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1344 SYSCALL_DEFINE1(userfaultfd, int, flags)
1349 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1354 file = userfaultfd_file_create(flags);
1356 error = PTR_ERR(file);
1357 goto err_put_unused_fd;
1359 fd_install(fd, file);
1369 static int __init userfaultfd_init(void)
1371 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1372 sizeof(struct userfaultfd_ctx),
1374 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1375 init_once_userfaultfd_ctx);
1378 __initcall(userfaultfd_init);