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 * Check that we can return VM_FAULT_RETRY.
292 * NOTE: it should become possible to return VM_FAULT_RETRY
293 * even if FAULT_FLAG_TRIED is set without leading to gup()
294 * -EBUSY failures, if the userfaultfd is to be extended for
295 * VM_UFFD_WP tracking and we intend to arm the userfault
296 * without first stopping userland access to the memory. For
297 * VM_UFFD_MISSING userfaults this is enough for now.
299 if (unlikely(!(flags & FAULT_FLAG_ALLOW_RETRY))) {
301 * Validate the invariant that nowait must allow retry
302 * to be sure not to return SIGBUS erroneously on
303 * nowait invocations.
305 BUG_ON(flags & FAULT_FLAG_RETRY_NOWAIT);
306 #ifdef CONFIG_DEBUG_VM
307 if (printk_ratelimit()) {
309 "FAULT_FLAG_ALLOW_RETRY missing %x\n", flags);
317 * Handle nowait, not much to do other than tell it to retry
320 ret = VM_FAULT_RETRY;
321 if (flags & FAULT_FLAG_RETRY_NOWAIT)
324 /* take the reference before dropping the mmap_sem */
325 userfaultfd_ctx_get(ctx);
327 init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
328 uwq.wq.private = current;
329 uwq.msg = userfault_msg(address, flags, reason);
332 return_to_userland = (flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
333 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
335 spin_lock(&ctx->fault_pending_wqh.lock);
337 * After the __add_wait_queue the uwq is visible to userland
338 * through poll/read().
340 __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
342 * The smp_mb() after __set_current_state prevents the reads
343 * following the spin_unlock to happen before the list_add in
346 set_current_state(return_to_userland ? TASK_INTERRUPTIBLE :
348 spin_unlock(&ctx->fault_pending_wqh.lock);
350 must_wait = userfaultfd_must_wait(ctx, address, flags, reason);
351 up_read(&mm->mmap_sem);
353 if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
354 (return_to_userland ? !signal_pending(current) :
355 !fatal_signal_pending(current)))) {
356 wake_up_poll(&ctx->fd_wqh, POLLIN);
358 ret |= VM_FAULT_MAJOR;
361 __set_current_state(TASK_RUNNING);
363 if (return_to_userland) {
364 if (signal_pending(current) &&
365 !fatal_signal_pending(current)) {
367 * If we got a SIGSTOP or SIGCONT and this is
368 * a normal userland page fault, just let
369 * userland return so the signal will be
370 * handled and gdb debugging works. The page
371 * fault code immediately after we return from
372 * this function is going to release the
373 * mmap_sem and it's not depending on it
374 * (unlike gup would if we were not to return
377 * If a fatal signal is pending we still take
378 * the streamlined VM_FAULT_RETRY failure path
379 * and there's no need to retake the mmap_sem
382 down_read(&mm->mmap_sem);
388 * Here we race with the list_del; list_add in
389 * userfaultfd_ctx_read(), however because we don't ever run
390 * list_del_init() to refile across the two lists, the prev
391 * and next pointers will never point to self. list_add also
392 * would never let any of the two pointers to point to
393 * self. So list_empty_careful won't risk to see both pointers
394 * pointing to self at any time during the list refile. The
395 * only case where list_del_init() is called is the full
396 * removal in the wake function and there we don't re-list_add
397 * and it's fine not to block on the spinlock. The uwq on this
398 * kernel stack can be released after the list_del_init.
400 if (!list_empty_careful(&uwq.wq.task_list)) {
401 spin_lock(&ctx->fault_pending_wqh.lock);
403 * No need of list_del_init(), the uwq on the stack
404 * will be freed shortly anyway.
406 list_del(&uwq.wq.task_list);
407 spin_unlock(&ctx->fault_pending_wqh.lock);
411 * ctx may go away after this if the userfault pseudo fd is
414 userfaultfd_ctx_put(ctx);
420 static int userfaultfd_release(struct inode *inode, struct file *file)
422 struct userfaultfd_ctx *ctx = file->private_data;
423 struct mm_struct *mm = ctx->mm;
424 struct vm_area_struct *vma, *prev;
425 /* len == 0 means wake all */
426 struct userfaultfd_wake_range range = { .len = 0, };
427 unsigned long new_flags;
429 ACCESS_ONCE(ctx->released) = true;
432 * Flush page faults out of all CPUs. NOTE: all page faults
433 * must be retried without returning VM_FAULT_SIGBUS if
434 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
435 * changes while handle_userfault released the mmap_sem. So
436 * it's critical that released is set to true (above), before
437 * taking the mmap_sem for writing.
439 down_write(&mm->mmap_sem);
441 for (vma = mm->mmap; vma; vma = vma->vm_next) {
443 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
444 !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
445 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
449 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
450 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
451 new_flags, vma->anon_vma,
452 vma->vm_file, vma->vm_pgoff,
459 vma->vm_flags = new_flags;
460 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
462 up_write(&mm->mmap_sem);
465 * After no new page faults can wait on this fault_*wqh, flush
466 * the last page faults that may have been already waiting on
469 spin_lock(&ctx->fault_pending_wqh.lock);
470 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
471 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
472 spin_unlock(&ctx->fault_pending_wqh.lock);
474 wake_up_poll(&ctx->fd_wqh, POLLHUP);
475 userfaultfd_ctx_put(ctx);
479 /* fault_pending_wqh.lock must be hold by the caller */
480 static inline struct userfaultfd_wait_queue *find_userfault(
481 struct userfaultfd_ctx *ctx)
484 struct userfaultfd_wait_queue *uwq;
486 VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
489 if (!waitqueue_active(&ctx->fault_pending_wqh))
491 /* walk in reverse to provide FIFO behavior to read userfaults */
492 wq = list_last_entry(&ctx->fault_pending_wqh.task_list,
493 typeof(*wq), task_list);
494 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
499 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
501 struct userfaultfd_ctx *ctx = file->private_data;
504 poll_wait(file, &ctx->fd_wqh, wait);
506 switch (ctx->state) {
507 case UFFD_STATE_WAIT_API:
509 case UFFD_STATE_RUNNING:
511 * poll() never guarantees that read won't block.
512 * userfaults can be waken before they're read().
514 if (unlikely(!(file->f_flags & O_NONBLOCK)))
517 * lockless access to see if there are pending faults
518 * __pollwait last action is the add_wait_queue but
519 * the spin_unlock would allow the waitqueue_active to
520 * pass above the actual list_add inside
521 * add_wait_queue critical section. So use a full
522 * memory barrier to serialize the list_add write of
523 * add_wait_queue() with the waitqueue_active read
528 if (waitqueue_active(&ctx->fault_pending_wqh))
536 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
537 struct uffd_msg *msg)
540 DECLARE_WAITQUEUE(wait, current);
541 struct userfaultfd_wait_queue *uwq;
543 /* always take the fd_wqh lock before the fault_pending_wqh lock */
544 spin_lock(&ctx->fd_wqh.lock);
545 __add_wait_queue(&ctx->fd_wqh, &wait);
547 set_current_state(TASK_INTERRUPTIBLE);
548 spin_lock(&ctx->fault_pending_wqh.lock);
549 uwq = find_userfault(ctx);
552 * Use a seqcount to repeat the lockless check
553 * in wake_userfault() to avoid missing
554 * wakeups because during the refile both
555 * waitqueue could become empty if this is the
558 write_seqcount_begin(&ctx->refile_seq);
561 * The fault_pending_wqh.lock prevents the uwq
562 * to disappear from under us.
564 * Refile this userfault from
565 * fault_pending_wqh to fault_wqh, it's not
566 * pending anymore after we read it.
568 * Use list_del() by hand (as
569 * userfaultfd_wake_function also uses
570 * list_del_init() by hand) to be sure nobody
571 * changes __remove_wait_queue() to use
572 * list_del_init() in turn breaking the
573 * !list_empty_careful() check in
574 * handle_userfault(). The uwq->wq.task_list
575 * must never be empty at any time during the
576 * refile, or the waitqueue could disappear
577 * from under us. The "wait_queue_head_t"
578 * parameter of __remove_wait_queue() is unused
581 list_del(&uwq->wq.task_list);
582 __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
584 write_seqcount_end(&ctx->refile_seq);
586 /* careful to always initialize msg if ret == 0 */
588 spin_unlock(&ctx->fault_pending_wqh.lock);
592 spin_unlock(&ctx->fault_pending_wqh.lock);
593 if (signal_pending(current)) {
601 spin_unlock(&ctx->fd_wqh.lock);
603 spin_lock(&ctx->fd_wqh.lock);
605 __remove_wait_queue(&ctx->fd_wqh, &wait);
606 __set_current_state(TASK_RUNNING);
607 spin_unlock(&ctx->fd_wqh.lock);
612 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
613 size_t count, loff_t *ppos)
615 struct userfaultfd_ctx *ctx = file->private_data;
616 ssize_t _ret, ret = 0;
618 int no_wait = file->f_flags & O_NONBLOCK;
620 if (ctx->state == UFFD_STATE_WAIT_API)
624 if (count < sizeof(msg))
625 return ret ? ret : -EINVAL;
626 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
628 return ret ? ret : _ret;
629 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
630 return ret ? ret : -EFAULT;
633 count -= sizeof(msg);
635 * Allow to read more than one fault at time but only
636 * block if waiting for the very first one.
638 no_wait = O_NONBLOCK;
642 static void __wake_userfault(struct userfaultfd_ctx *ctx,
643 struct userfaultfd_wake_range *range)
645 unsigned long start, end;
647 start = range->start;
648 end = range->start + range->len;
650 spin_lock(&ctx->fault_pending_wqh.lock);
651 /* wake all in the range and autoremove */
652 if (waitqueue_active(&ctx->fault_pending_wqh))
653 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
655 if (waitqueue_active(&ctx->fault_wqh))
656 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
657 spin_unlock(&ctx->fault_pending_wqh.lock);
660 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
661 struct userfaultfd_wake_range *range)
667 * To be sure waitqueue_active() is not reordered by the CPU
668 * before the pagetable update, use an explicit SMP memory
669 * barrier here. PT lock release or up_read(mmap_sem) still
670 * have release semantics that can allow the
671 * waitqueue_active() to be reordered before the pte update.
676 * Use waitqueue_active because it's very frequent to
677 * change the address space atomically even if there are no
678 * userfaults yet. So we take the spinlock only when we're
679 * sure we've userfaults to wake.
682 seq = read_seqcount_begin(&ctx->refile_seq);
683 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
684 waitqueue_active(&ctx->fault_wqh);
686 } while (read_seqcount_retry(&ctx->refile_seq, seq));
688 __wake_userfault(ctx, range);
691 static __always_inline int validate_range(struct mm_struct *mm,
692 __u64 start, __u64 len)
694 __u64 task_size = mm->task_size;
696 if (start & ~PAGE_MASK)
698 if (len & ~PAGE_MASK)
702 if (start < mmap_min_addr)
704 if (start >= task_size)
706 if (len > task_size - start)
711 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
714 struct mm_struct *mm = ctx->mm;
715 struct vm_area_struct *vma, *prev, *cur;
717 struct uffdio_register uffdio_register;
718 struct uffdio_register __user *user_uffdio_register;
719 unsigned long vm_flags, new_flags;
721 unsigned long start, end, vma_end;
723 user_uffdio_register = (struct uffdio_register __user *) arg;
726 if (copy_from_user(&uffdio_register, user_uffdio_register,
727 sizeof(uffdio_register)-sizeof(__u64)))
731 if (!uffdio_register.mode)
733 if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
734 UFFDIO_REGISTER_MODE_WP))
737 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
738 vm_flags |= VM_UFFD_MISSING;
739 if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
740 vm_flags |= VM_UFFD_WP;
742 * FIXME: remove the below error constraint by
743 * implementing the wprotect tracking mode.
749 ret = validate_range(mm, uffdio_register.range.start,
750 uffdio_register.range.len);
754 start = uffdio_register.range.start;
755 end = start + uffdio_register.range.len;
757 down_write(&mm->mmap_sem);
758 vma = find_vma_prev(mm, start, &prev);
764 /* check that there's at least one vma in the range */
766 if (vma->vm_start >= end)
770 * Search for not compatible vmas.
772 * FIXME: this shall be relaxed later so that it doesn't fail
773 * on tmpfs backed vmas (in addition to the current allowance
774 * on anonymous vmas).
777 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
780 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
781 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
783 /* check not compatible vmas */
789 * Check that this vma isn't already owned by a
790 * different userfaultfd. We can't allow more than one
791 * userfaultfd to own a single vma simultaneously or we
792 * wouldn't know which one to deliver the userfaults to.
795 if (cur->vm_userfaultfd_ctx.ctx &&
796 cur->vm_userfaultfd_ctx.ctx != ctx)
803 if (vma->vm_start < start)
811 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
812 vma->vm_userfaultfd_ctx.ctx != ctx);
815 * Nothing to do: this vma is already registered into this
816 * userfaultfd and with the right tracking mode too.
818 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
819 (vma->vm_flags & vm_flags) == vm_flags)
822 if (vma->vm_start > start)
823 start = vma->vm_start;
824 vma_end = min(end, vma->vm_end);
826 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
827 prev = vma_merge(mm, prev, start, vma_end, new_flags,
828 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
830 ((struct vm_userfaultfd_ctx){ ctx }));
835 if (vma->vm_start < start) {
836 ret = split_vma(mm, vma, start, 1);
840 if (vma->vm_end > end) {
841 ret = split_vma(mm, vma, end, 0);
847 * In the vma_merge() successful mprotect-like case 8:
848 * the next vma was merged into the current one and
849 * the current one has not been updated yet.
851 vma->vm_flags = new_flags;
852 vma->vm_userfaultfd_ctx.ctx = ctx;
858 } while (vma && vma->vm_start < end);
860 up_write(&mm->mmap_sem);
863 * Now that we scanned all vmas we can already tell
864 * userland which ioctls methods are guaranteed to
865 * succeed on this range.
867 if (put_user(UFFD_API_RANGE_IOCTLS,
868 &user_uffdio_register->ioctls))
875 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
878 struct mm_struct *mm = ctx->mm;
879 struct vm_area_struct *vma, *prev, *cur;
881 struct uffdio_range uffdio_unregister;
882 unsigned long new_flags;
884 unsigned long start, end, vma_end;
885 const void __user *buf = (void __user *)arg;
888 if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
891 ret = validate_range(mm, uffdio_unregister.start,
892 uffdio_unregister.len);
896 start = uffdio_unregister.start;
897 end = start + uffdio_unregister.len;
899 down_write(&mm->mmap_sem);
900 vma = find_vma_prev(mm, start, &prev);
906 /* check that there's at least one vma in the range */
908 if (vma->vm_start >= end)
912 * Search for not compatible vmas.
914 * FIXME: this shall be relaxed later so that it doesn't fail
915 * on tmpfs backed vmas (in addition to the current allowance
916 * on anonymous vmas).
920 for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
923 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
924 !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
927 * Check not compatible vmas, not strictly required
928 * here as not compatible vmas cannot have an
929 * userfaultfd_ctx registered on them, but this
930 * provides for more strict behavior to notice
931 * unregistration errors.
940 if (vma->vm_start < start)
950 * Nothing to do: this vma is already registered into this
951 * userfaultfd and with the right tracking mode too.
953 if (!vma->vm_userfaultfd_ctx.ctx)
956 if (vma->vm_start > start)
957 start = vma->vm_start;
958 vma_end = min(end, vma->vm_end);
960 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
961 prev = vma_merge(mm, prev, start, vma_end, new_flags,
962 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
969 if (vma->vm_start < start) {
970 ret = split_vma(mm, vma, start, 1);
974 if (vma->vm_end > end) {
975 ret = split_vma(mm, vma, end, 0);
981 * In the vma_merge() successful mprotect-like case 8:
982 * the next vma was merged into the current one and
983 * the current one has not been updated yet.
985 vma->vm_flags = new_flags;
986 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
992 } while (vma && vma->vm_start < end);
994 up_write(&mm->mmap_sem);
1000 * userfaultfd_wake may be used in combination with the
1001 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1003 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1007 struct uffdio_range uffdio_wake;
1008 struct userfaultfd_wake_range range;
1009 const void __user *buf = (void __user *)arg;
1012 if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1015 ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1019 range.start = uffdio_wake.start;
1020 range.len = uffdio_wake.len;
1023 * len == 0 means wake all and we don't want to wake all here,
1024 * so check it again to be sure.
1026 VM_BUG_ON(!range.len);
1028 wake_userfault(ctx, &range);
1035 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1039 struct uffdio_copy uffdio_copy;
1040 struct uffdio_copy __user *user_uffdio_copy;
1041 struct userfaultfd_wake_range range;
1043 user_uffdio_copy = (struct uffdio_copy __user *) arg;
1046 if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1047 /* don't copy "copy" last field */
1048 sizeof(uffdio_copy)-sizeof(__s64)))
1051 ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1055 * double check for wraparound just in case. copy_from_user()
1056 * will later check uffdio_copy.src + uffdio_copy.len to fit
1057 * in the userland range.
1060 if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1062 if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1065 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1067 if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1072 /* len == 0 would wake all */
1074 if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1075 range.start = uffdio_copy.dst;
1076 wake_userfault(ctx, &range);
1078 ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1083 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1087 struct uffdio_zeropage uffdio_zeropage;
1088 struct uffdio_zeropage __user *user_uffdio_zeropage;
1089 struct userfaultfd_wake_range range;
1091 user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1094 if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1095 /* don't copy "zeropage" last field */
1096 sizeof(uffdio_zeropage)-sizeof(__s64)))
1099 ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1100 uffdio_zeropage.range.len);
1104 if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1107 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1108 uffdio_zeropage.range.len);
1109 if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1113 /* len == 0 would wake all */
1116 if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1117 range.start = uffdio_zeropage.range.start;
1118 wake_userfault(ctx, &range);
1120 ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1126 * userland asks for a certain API version and we return which bits
1127 * and ioctl commands are implemented in this kernel for such API
1128 * version or -EINVAL if unknown.
1130 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1133 struct uffdio_api uffdio_api;
1134 void __user *buf = (void __user *)arg;
1138 if (ctx->state != UFFD_STATE_WAIT_API)
1141 if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1143 if (uffdio_api.api != UFFD_API || uffdio_api.features) {
1144 memset(&uffdio_api, 0, sizeof(uffdio_api));
1145 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1150 uffdio_api.features = UFFD_API_FEATURES;
1151 uffdio_api.ioctls = UFFD_API_IOCTLS;
1153 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1155 ctx->state = UFFD_STATE_RUNNING;
1161 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1165 struct userfaultfd_ctx *ctx = file->private_data;
1167 if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1172 ret = userfaultfd_api(ctx, arg);
1174 case UFFDIO_REGISTER:
1175 ret = userfaultfd_register(ctx, arg);
1177 case UFFDIO_UNREGISTER:
1178 ret = userfaultfd_unregister(ctx, arg);
1181 ret = userfaultfd_wake(ctx, arg);
1184 ret = userfaultfd_copy(ctx, arg);
1186 case UFFDIO_ZEROPAGE:
1187 ret = userfaultfd_zeropage(ctx, arg);
1193 #ifdef CONFIG_PROC_FS
1194 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1196 struct userfaultfd_ctx *ctx = f->private_data;
1198 struct userfaultfd_wait_queue *uwq;
1199 unsigned long pending = 0, total = 0;
1201 spin_lock(&ctx->fault_pending_wqh.lock);
1202 list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1203 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1207 list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1208 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1211 spin_unlock(&ctx->fault_pending_wqh.lock);
1214 * If more protocols will be added, there will be all shown
1215 * separated by a space. Like this:
1216 * protocols: aa:... bb:...
1218 seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1219 pending, total, UFFD_API, UFFD_API_FEATURES,
1220 UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1224 static const struct file_operations userfaultfd_fops = {
1225 #ifdef CONFIG_PROC_FS
1226 .show_fdinfo = userfaultfd_show_fdinfo,
1228 .release = userfaultfd_release,
1229 .poll = userfaultfd_poll,
1230 .read = userfaultfd_read,
1231 .unlocked_ioctl = userfaultfd_ioctl,
1232 .compat_ioctl = userfaultfd_ioctl,
1233 .llseek = noop_llseek,
1236 static void init_once_userfaultfd_ctx(void *mem)
1238 struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1240 init_waitqueue_head(&ctx->fault_pending_wqh);
1241 init_waitqueue_head(&ctx->fault_wqh);
1242 init_waitqueue_head(&ctx->fd_wqh);
1243 seqcount_init(&ctx->refile_seq);
1247 * userfaultfd_file_create - Creates an userfaultfd file pointer.
1248 * @flags: Flags for the userfaultfd file.
1250 * This function creates an userfaultfd file pointer, w/out installing
1251 * it into the fd table. This is useful when the userfaultfd file is
1252 * used during the initialization of data structures that require
1253 * extra setup after the userfaultfd creation. So the userfaultfd
1254 * creation is split into the file pointer creation phase, and the
1255 * file descriptor installation phase. In this way races with
1256 * userspace closing the newly installed file descriptor can be
1257 * avoided. Returns an userfaultfd file pointer, or a proper error
1260 static struct file *userfaultfd_file_create(int flags)
1263 struct userfaultfd_ctx *ctx;
1265 BUG_ON(!current->mm);
1267 /* Check the UFFD_* constants for consistency. */
1268 BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1269 BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1271 file = ERR_PTR(-EINVAL);
1272 if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1275 file = ERR_PTR(-ENOMEM);
1276 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1280 atomic_set(&ctx->refcount, 1);
1282 ctx->state = UFFD_STATE_WAIT_API;
1283 ctx->released = false;
1284 ctx->mm = current->mm;
1285 /* prevent the mm struct to be freed */
1286 atomic_inc(&ctx->mm->mm_users);
1288 file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1289 O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1292 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1298 SYSCALL_DEFINE1(userfaultfd, int, flags)
1303 error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1308 file = userfaultfd_file_create(flags);
1310 error = PTR_ERR(file);
1311 goto err_put_unused_fd;
1313 fd_install(fd, file);
1323 static int __init userfaultfd_init(void)
1325 userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1326 sizeof(struct userfaultfd_ctx),
1328 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1329 init_once_userfaultfd_ctx);
1332 __initcall(userfaultfd_init);