1/*
2 *  fs/userfaultfd.c
3 *
4 *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
5 *  Copyright (C) 2008-2009 Red Hat, Inc.
6 *  Copyright (C) 2015  Red Hat, Inc.
7 *
8 *  This work is licensed under the terms of the GNU GPL, version 2. See
9 *  the COPYING file in the top-level directory.
10 *
11 *  Some part derived from fs/eventfd.c (anon inode setup) and
12 *  mm/ksm.c (mm hashing).
13 */
14
15#include <linux/hashtable.h>
16#include <linux/sched.h>
17#include <linux/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
30static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
31
32enum userfaultfd_state {
33	UFFD_STATE_WAIT_API,
34	UFFD_STATE_RUNNING,
35};
36
37/*
38 * Start with fault_pending_wqh and fault_wqh so they're more likely
39 * to be in the same cacheline.
40 */
41struct 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 */
51	atomic_t refcount;
52	/* userfaultfd syscall flags */
53	unsigned int flags;
54	/* state machine */
55	enum userfaultfd_state state;
56	/* released */
57	bool released;
58	/* mm with one ore more vmas attached to this userfaultfd_ctx */
59	struct mm_struct *mm;
60};
61
62struct userfaultfd_wait_queue {
63	struct uffd_msg msg;
64	wait_queue_t wq;
65	struct userfaultfd_ctx *ctx;
66};
67
68struct userfaultfd_wake_range {
69	unsigned long start;
70	unsigned long len;
71};
72
73static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
74				     int wake_flags, void *key)
75{
76	struct userfaultfd_wake_range *range = key;
77	int ret;
78	struct userfaultfd_wait_queue *uwq;
79	unsigned long start, len;
80
81	uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
82	ret = 0;
83	/* len == 0 means wake all */
84	start = range->start;
85	len = range->len;
86	if (len && (start > uwq->msg.arg.pagefault.address ||
87		    start + len <= uwq->msg.arg.pagefault.address))
88		goto out;
89	ret = wake_up_state(wq->private, mode);
90	if (ret)
91		/*
92		 * Wake only once, autoremove behavior.
93		 *
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.
104		 */
105		list_del_init(&wq->task_list);
106out:
107	return ret;
108}
109
110/**
111 * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
112 * context.
113 * @ctx: [in] Pointer to the userfaultfd context.
114 *
115 * Returns: In case of success, returns not zero.
116 */
117static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
118{
119	if (!atomic_inc_not_zero(&ctx->refcount))
120		BUG();
121}
122
123/**
124 * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
125 * context.
126 * @ctx: [in] Pointer to userfaultfd context.
127 *
128 * The userfaultfd context reference must have been previously acquired either
129 * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
130 */
131static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
132{
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));
140		mmput(ctx->mm);
141		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
142	}
143}
144
145static inline void msg_init(struct uffd_msg *msg)
146{
147	BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
148	/*
149	 * Must use memset to zero out the paddings or kernel data is
150	 * leaked to userland.
151	 */
152	memset(msg, 0, sizeof(struct uffd_msg));
153}
154
155static inline struct uffd_msg userfault_msg(unsigned long address,
156					    unsigned int flags,
157					    unsigned long reason)
158{
159	struct uffd_msg msg;
160	msg_init(&msg);
161	msg.event = UFFD_EVENT_PAGEFAULT;
162	msg.arg.pagefault.address = address;
163	if (flags & FAULT_FLAG_WRITE)
164		/*
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
169		 * a write fault.
170		 */
171		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
172	if (reason & VM_UFFD_WP)
173		/*
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.
179		 */
180		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
181	return msg;
182}
183
184/*
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
189 * threads.
190 */
191static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
192					 unsigned long address,
193					 unsigned long flags,
194					 unsigned long reason)
195{
196	struct mm_struct *mm = ctx->mm;
197	pgd_t *pgd;
198	pud_t *pud;
199	pmd_t *pmd, _pmd;
200	pte_t *pte;
201	bool ret = true;
202
203	VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
204
205	pgd = pgd_offset(mm, address);
206	if (!pgd_present(*pgd))
207		goto out;
208	pud = pud_offset(pgd, address);
209	if (!pud_present(*pud))
210		goto out;
211	pmd = pmd_offset(pud, address);
212	/*
213	 * READ_ONCE must function as a barrier with narrower scope
214	 * and it must be equivalent to:
215	 *	_pmd = *pmd; barrier();
216	 *
217	 * This is to deal with the instability (as in
218	 * pmd_trans_unstable) of the pmd.
219	 */
220	_pmd = READ_ONCE(*pmd);
221	if (!pmd_present(_pmd))
222		goto out;
223
224	ret = false;
225	if (pmd_trans_huge(_pmd))
226		goto out;
227
228	/*
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.
231	 */
232	pte = pte_offset_map(pmd, address);
233	/*
234	 * Lockless access: we're in a wait_event so it's ok if it
235	 * changes under us.
236	 */
237	if (pte_none(*pte))
238		ret = true;
239	pte_unmap(pte);
240
241out:
242	return ret;
243}
244
245/*
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.
250 *
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
253 * not set.
254 *
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
258 * returning it.
259 */
260int handle_userfault(struct vm_area_struct *vma, unsigned long address,
261		     unsigned int flags, unsigned long reason)
262{
263	struct mm_struct *mm = vma->vm_mm;
264	struct userfaultfd_ctx *ctx;
265	struct userfaultfd_wait_queue uwq;
266	int ret;
267	bool must_wait, return_to_userland;
268
269	BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
270
271	ret = VM_FAULT_SIGBUS;
272	ctx = vma->vm_userfaultfd_ctx.ctx;
273	if (!ctx)
274		goto out;
275
276	BUG_ON(ctx->mm != mm);
277
278	VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
279	VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
280
281	/*
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.
285	 */
286	if (unlikely(ACCESS_ONCE(ctx->released)))
287		goto out;
288
289	/*
290	 * We don't do userfault handling for the final child pid update.
291	 */
292	if (current->flags & PF_EXITING)
293		goto out;
294
295	/*
296	 * Check that we can return VM_FAULT_RETRY.
297	 *
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.
304	 */
305	if (unlikely(!(flags & FAULT_FLAG_ALLOW_RETRY))) {
306		/*
307		 * Validate the invariant that nowait must allow retry
308		 * to be sure not to return SIGBUS erroneously on
309		 * nowait invocations.
310		 */
311		BUG_ON(flags & FAULT_FLAG_RETRY_NOWAIT);
312#ifdef CONFIG_DEBUG_VM
313		if (printk_ratelimit()) {
314			printk(KERN_WARNING
315			       "FAULT_FLAG_ALLOW_RETRY missing %x\n", flags);
316			dump_stack();
317		}
318#endif
319		goto out;
320	}
321
322	/*
323	 * Handle nowait, not much to do other than tell it to retry
324	 * and wait.
325	 */
326	ret = VM_FAULT_RETRY;
327	if (flags & FAULT_FLAG_RETRY_NOWAIT)
328		goto out;
329
330	/* take the reference before dropping the mmap_sem */
331	userfaultfd_ctx_get(ctx);
332
333	init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
334	uwq.wq.private = current;
335	uwq.msg = userfault_msg(address, flags, reason);
336	uwq.ctx = ctx;
337
338	return_to_userland = (flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
339		(FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
340
341	spin_lock(&ctx->fault_pending_wqh.lock);
342	/*
343	 * After the __add_wait_queue the uwq is visible to userland
344	 * through poll/read().
345	 */
346	__add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
347	/*
348	 * The smp_mb() after __set_current_state prevents the reads
349	 * following the spin_unlock to happen before the list_add in
350	 * __add_wait_queue.
351	 */
352	set_current_state(return_to_userland ? TASK_INTERRUPTIBLE :
353			  TASK_KILLABLE);
354	spin_unlock(&ctx->fault_pending_wqh.lock);
355
356	must_wait = userfaultfd_must_wait(ctx, address, flags, reason);
357	up_read(&mm->mmap_sem);
358
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);
363		schedule();
364		ret |= VM_FAULT_MAJOR;
365	}
366
367	__set_current_state(TASK_RUNNING);
368
369	if (return_to_userland) {
370		if (signal_pending(current) &&
371		    !fatal_signal_pending(current)) {
372			/*
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
381			 * VM_FAULT_RETRY).
382			 *
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
386			 * in such case.
387			 */
388			down_read(&mm->mmap_sem);
389			ret = 0;
390		}
391	}
392
393	/*
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.
405	 */
406	if (!list_empty_careful(&uwq.wq.task_list)) {
407		spin_lock(&ctx->fault_pending_wqh.lock);
408		/*
409		 * No need of list_del_init(), the uwq on the stack
410		 * will be freed shortly anyway.
411		 */
412		list_del(&uwq.wq.task_list);
413		spin_unlock(&ctx->fault_pending_wqh.lock);
414	}
415
416	/*
417	 * ctx may go away after this if the userfault pseudo fd is
418	 * already released.
419	 */
420	userfaultfd_ctx_put(ctx);
421
422out:
423	return ret;
424}
425
426static int userfaultfd_release(struct inode *inode, struct file *file)
427{
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;
434
435	ACCESS_ONCE(ctx->released) = true;
436
437	/*
438	 * Flush page faults out of all CPUs. NOTE: all page faults
439	 * must be retried without returning VM_FAULT_SIGBUS if
440	 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
441	 * changes while handle_userfault released the mmap_sem. So
442	 * it's critical that released is set to true (above), before
443	 * taking the mmap_sem for writing.
444	 */
445	down_write(&mm->mmap_sem);
446	prev = NULL;
447	for (vma = mm->mmap; vma; vma = vma->vm_next) {
448		cond_resched();
449		BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
450		       !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
451		if (vma->vm_userfaultfd_ctx.ctx != ctx) {
452			prev = vma;
453			continue;
454		}
455		new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
456		prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
457				 new_flags, vma->anon_vma,
458				 vma->vm_file, vma->vm_pgoff,
459				 vma_policy(vma),
460				 NULL_VM_UFFD_CTX);
461		if (prev)
462			vma = prev;
463		else
464			prev = vma;
465		vma->vm_flags = new_flags;
466		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
467	}
468	up_write(&mm->mmap_sem);
469
470	/*
471	 * After no new page faults can wait on this fault_*wqh, flush
472	 * the last page faults that may have been already waiting on
473	 * the fault_*wqh.
474	 */
475	spin_lock(&ctx->fault_pending_wqh.lock);
476	__wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
477	__wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
478	spin_unlock(&ctx->fault_pending_wqh.lock);
479
480	wake_up_poll(&ctx->fd_wqh, POLLHUP);
481	userfaultfd_ctx_put(ctx);
482	return 0;
483}
484
485/* fault_pending_wqh.lock must be hold by the caller */
486static inline struct userfaultfd_wait_queue *find_userfault(
487	struct userfaultfd_ctx *ctx)
488{
489	wait_queue_t *wq;
490	struct userfaultfd_wait_queue *uwq;
491
492	VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
493
494	uwq = NULL;
495	if (!waitqueue_active(&ctx->fault_pending_wqh))
496		goto out;
497	/* walk in reverse to provide FIFO behavior to read userfaults */
498	wq = list_last_entry(&ctx->fault_pending_wqh.task_list,
499			     typeof(*wq), task_list);
500	uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
501out:
502	return uwq;
503}
504
505static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
506{
507	struct userfaultfd_ctx *ctx = file->private_data;
508	unsigned int ret;
509
510	poll_wait(file, &ctx->fd_wqh, wait);
511
512	switch (ctx->state) {
513	case UFFD_STATE_WAIT_API:
514		return POLLERR;
515	case UFFD_STATE_RUNNING:
516		/*
517		 * poll() never guarantees that read won't block.
518		 * userfaults can be waken before they're read().
519		 */
520		if (unlikely(!(file->f_flags & O_NONBLOCK)))
521			return POLLERR;
522		/*
523		 * lockless access to see if there are pending faults
524		 * __pollwait last action is the add_wait_queue but
525		 * the spin_unlock would allow the waitqueue_active to
526		 * pass above the actual list_add inside
527		 * add_wait_queue critical section. So use a full
528		 * memory barrier to serialize the list_add write of
529		 * add_wait_queue() with the waitqueue_active read
530		 * below.
531		 */
532		ret = 0;
533		smp_mb();
534		if (waitqueue_active(&ctx->fault_pending_wqh))
535			ret = POLLIN;
536		return ret;
537	default:
538		BUG();
539	}
540}
541
542static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
543				    struct uffd_msg *msg)
544{
545	ssize_t ret;
546	DECLARE_WAITQUEUE(wait, current);
547	struct userfaultfd_wait_queue *uwq;
548
549	/* always take the fd_wqh lock before the fault_pending_wqh lock */
550	spin_lock(&ctx->fd_wqh.lock);
551	__add_wait_queue(&ctx->fd_wqh, &wait);
552	for (;;) {
553		set_current_state(TASK_INTERRUPTIBLE);
554		spin_lock(&ctx->fault_pending_wqh.lock);
555		uwq = find_userfault(ctx);
556		if (uwq) {
557			/*
558			 * Use a seqcount to repeat the lockless check
559			 * in wake_userfault() to avoid missing
560			 * wakeups because during the refile both
561			 * waitqueue could become empty if this is the
562			 * only userfault.
563			 */
564			write_seqcount_begin(&ctx->refile_seq);
565
566			/*
567			 * The fault_pending_wqh.lock prevents the uwq
568			 * to disappear from under us.
569			 *
570			 * Refile this userfault from
571			 * fault_pending_wqh to fault_wqh, it's not
572			 * pending anymore after we read it.
573			 *
574			 * Use list_del() by hand (as
575			 * userfaultfd_wake_function also uses
576			 * list_del_init() by hand) to be sure nobody
577			 * changes __remove_wait_queue() to use
578			 * list_del_init() in turn breaking the
579			 * !list_empty_careful() check in
580			 * handle_userfault(). The uwq->wq.task_list
581			 * must never be empty at any time during the
582			 * refile, or the waitqueue could disappear
583			 * from under us. The "wait_queue_head_t"
584			 * parameter of __remove_wait_queue() is unused
585			 * anyway.
586			 */
587			list_del(&uwq->wq.task_list);
588			__add_wait_queue(&ctx->fault_wqh, &uwq->wq);
589
590			write_seqcount_end(&ctx->refile_seq);
591
592			/* careful to always initialize msg if ret == 0 */
593			*msg = uwq->msg;
594			spin_unlock(&ctx->fault_pending_wqh.lock);
595			ret = 0;
596			break;
597		}
598		spin_unlock(&ctx->fault_pending_wqh.lock);
599		if (signal_pending(current)) {
600			ret = -ERESTARTSYS;
601			break;
602		}
603		if (no_wait) {
604			ret = -EAGAIN;
605			break;
606		}
607		spin_unlock(&ctx->fd_wqh.lock);
608		schedule();
609		spin_lock(&ctx->fd_wqh.lock);
610	}
611	__remove_wait_queue(&ctx->fd_wqh, &wait);
612	__set_current_state(TASK_RUNNING);
613	spin_unlock(&ctx->fd_wqh.lock);
614
615	return ret;
616}
617
618static ssize_t userfaultfd_read(struct file *file, char __user *buf,
619				size_t count, loff_t *ppos)
620{
621	struct userfaultfd_ctx *ctx = file->private_data;
622	ssize_t _ret, ret = 0;
623	struct uffd_msg msg;
624	int no_wait = file->f_flags & O_NONBLOCK;
625
626	if (ctx->state == UFFD_STATE_WAIT_API)
627		return -EINVAL;
628
629	for (;;) {
630		if (count < sizeof(msg))
631			return ret ? ret : -EINVAL;
632		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
633		if (_ret < 0)
634			return ret ? ret : _ret;
635		if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
636			return ret ? ret : -EFAULT;
637		ret += sizeof(msg);
638		buf += sizeof(msg);
639		count -= sizeof(msg);
640		/*
641		 * Allow to read more than one fault at time but only
642		 * block if waiting for the very first one.
643		 */
644		no_wait = O_NONBLOCK;
645	}
646}
647
648static void __wake_userfault(struct userfaultfd_ctx *ctx,
649			     struct userfaultfd_wake_range *range)
650{
651	unsigned long start, end;
652
653	start = range->start;
654	end = range->start + range->len;
655
656	spin_lock(&ctx->fault_pending_wqh.lock);
657	/* wake all in the range and autoremove */
658	if (waitqueue_active(&ctx->fault_pending_wqh))
659		__wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
660				     range);
661	if (waitqueue_active(&ctx->fault_wqh))
662		__wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
663	spin_unlock(&ctx->fault_pending_wqh.lock);
664}
665
666static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
667					   struct userfaultfd_wake_range *range)
668{
669	unsigned seq;
670	bool need_wakeup;
671
672	/*
673	 * To be sure waitqueue_active() is not reordered by the CPU
674	 * before the pagetable update, use an explicit SMP memory
675	 * barrier here. PT lock release or up_read(mmap_sem) still
676	 * have release semantics that can allow the
677	 * waitqueue_active() to be reordered before the pte update.
678	 */
679	smp_mb();
680
681	/*
682	 * Use waitqueue_active because it's very frequent to
683	 * change the address space atomically even if there are no
684	 * userfaults yet. So we take the spinlock only when we're
685	 * sure we've userfaults to wake.
686	 */
687	do {
688		seq = read_seqcount_begin(&ctx->refile_seq);
689		need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
690			waitqueue_active(&ctx->fault_wqh);
691		cond_resched();
692	} while (read_seqcount_retry(&ctx->refile_seq, seq));
693	if (need_wakeup)
694		__wake_userfault(ctx, range);
695}
696
697static __always_inline int validate_range(struct mm_struct *mm,
698					  __u64 start, __u64 len)
699{
700	__u64 task_size = mm->task_size;
701
702	if (start & ~PAGE_MASK)
703		return -EINVAL;
704	if (len & ~PAGE_MASK)
705		return -EINVAL;
706	if (!len)
707		return -EINVAL;
708	if (start < mmap_min_addr)
709		return -EINVAL;
710	if (start >= task_size)
711		return -EINVAL;
712	if (len > task_size - start)
713		return -EINVAL;
714	return 0;
715}
716
717static int userfaultfd_register(struct userfaultfd_ctx *ctx,
718				unsigned long arg)
719{
720	struct mm_struct *mm = ctx->mm;
721	struct vm_area_struct *vma, *prev, *cur;
722	int ret;
723	struct uffdio_register uffdio_register;
724	struct uffdio_register __user *user_uffdio_register;
725	unsigned long vm_flags, new_flags;
726	bool found;
727	unsigned long start, end, vma_end;
728
729	user_uffdio_register = (struct uffdio_register __user *) arg;
730
731	ret = -EFAULT;
732	if (copy_from_user(&uffdio_register, user_uffdio_register,
733			   sizeof(uffdio_register)-sizeof(__u64)))
734		goto out;
735
736	ret = -EINVAL;
737	if (!uffdio_register.mode)
738		goto out;
739	if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
740				     UFFDIO_REGISTER_MODE_WP))
741		goto out;
742	vm_flags = 0;
743	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
744		vm_flags |= VM_UFFD_MISSING;
745	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
746		vm_flags |= VM_UFFD_WP;
747		/*
748		 * FIXME: remove the below error constraint by
749		 * implementing the wprotect tracking mode.
750		 */
751		ret = -EINVAL;
752		goto out;
753	}
754
755	ret = validate_range(mm, uffdio_register.range.start,
756			     uffdio_register.range.len);
757	if (ret)
758		goto out;
759
760	start = uffdio_register.range.start;
761	end = start + uffdio_register.range.len;
762
763	down_write(&mm->mmap_sem);
764	vma = find_vma_prev(mm, start, &prev);
765
766	ret = -ENOMEM;
767	if (!vma)
768		goto out_unlock;
769
770	/* check that there's at least one vma in the range */
771	ret = -EINVAL;
772	if (vma->vm_start >= end)
773		goto out_unlock;
774
775	/*
776	 * Search for not compatible vmas.
777	 *
778	 * FIXME: this shall be relaxed later so that it doesn't fail
779	 * on tmpfs backed vmas (in addition to the current allowance
780	 * on anonymous vmas).
781	 */
782	found = false;
783	for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
784		cond_resched();
785
786		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
787		       !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
788
789		/* check not compatible vmas */
790		ret = -EINVAL;
791		if (cur->vm_ops)
792			goto out_unlock;
793
794		/*
795		 * Check that this vma isn't already owned by a
796		 * different userfaultfd. We can't allow more than one
797		 * userfaultfd to own a single vma simultaneously or we
798		 * wouldn't know which one to deliver the userfaults to.
799		 */
800		ret = -EBUSY;
801		if (cur->vm_userfaultfd_ctx.ctx &&
802		    cur->vm_userfaultfd_ctx.ctx != ctx)
803			goto out_unlock;
804
805		found = true;
806	}
807	BUG_ON(!found);
808
809	if (vma->vm_start < start)
810		prev = vma;
811
812	ret = 0;
813	do {
814		cond_resched();
815
816		BUG_ON(vma->vm_ops);
817		BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
818		       vma->vm_userfaultfd_ctx.ctx != ctx);
819
820		/*
821		 * Nothing to do: this vma is already registered into this
822		 * userfaultfd and with the right tracking mode too.
823		 */
824		if (vma->vm_userfaultfd_ctx.ctx == ctx &&
825		    (vma->vm_flags & vm_flags) == vm_flags)
826			goto skip;
827
828		if (vma->vm_start > start)
829			start = vma->vm_start;
830		vma_end = min(end, vma->vm_end);
831
832		new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
833		prev = vma_merge(mm, prev, start, vma_end, new_flags,
834				 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
835				 vma_policy(vma),
836				 ((struct vm_userfaultfd_ctx){ ctx }));
837		if (prev) {
838			vma = prev;
839			goto next;
840		}
841		if (vma->vm_start < start) {
842			ret = split_vma(mm, vma, start, 1);
843			if (ret)
844				break;
845		}
846		if (vma->vm_end > end) {
847			ret = split_vma(mm, vma, end, 0);
848			if (ret)
849				break;
850		}
851	next:
852		/*
853		 * In the vma_merge() successful mprotect-like case 8:
854		 * the next vma was merged into the current one and
855		 * the current one has not been updated yet.
856		 */
857		vma->vm_flags = new_flags;
858		vma->vm_userfaultfd_ctx.ctx = ctx;
859
860	skip:
861		prev = vma;
862		start = vma->vm_end;
863		vma = vma->vm_next;
864	} while (vma && vma->vm_start < end);
865out_unlock:
866	up_write(&mm->mmap_sem);
867	if (!ret) {
868		/*
869		 * Now that we scanned all vmas we can already tell
870		 * userland which ioctls methods are guaranteed to
871		 * succeed on this range.
872		 */
873		if (put_user(UFFD_API_RANGE_IOCTLS,
874			     &user_uffdio_register->ioctls))
875			ret = -EFAULT;
876	}
877out:
878	return ret;
879}
880
881static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
882				  unsigned long arg)
883{
884	struct mm_struct *mm = ctx->mm;
885	struct vm_area_struct *vma, *prev, *cur;
886	int ret;
887	struct uffdio_range uffdio_unregister;
888	unsigned long new_flags;
889	bool found;
890	unsigned long start, end, vma_end;
891	const void __user *buf = (void __user *)arg;
892
893	ret = -EFAULT;
894	if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
895		goto out;
896
897	ret = validate_range(mm, uffdio_unregister.start,
898			     uffdio_unregister.len);
899	if (ret)
900		goto out;
901
902	start = uffdio_unregister.start;
903	end = start + uffdio_unregister.len;
904
905	down_write(&mm->mmap_sem);
906	vma = find_vma_prev(mm, start, &prev);
907
908	ret = -ENOMEM;
909	if (!vma)
910		goto out_unlock;
911
912	/* check that there's at least one vma in the range */
913	ret = -EINVAL;
914	if (vma->vm_start >= end)
915		goto out_unlock;
916
917	/*
918	 * Search for not compatible vmas.
919	 *
920	 * FIXME: this shall be relaxed later so that it doesn't fail
921	 * on tmpfs backed vmas (in addition to the current allowance
922	 * on anonymous vmas).
923	 */
924	found = false;
925	ret = -EINVAL;
926	for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
927		cond_resched();
928
929		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
930		       !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
931
932		/*
933		 * Check not compatible vmas, not strictly required
934		 * here as not compatible vmas cannot have an
935		 * userfaultfd_ctx registered on them, but this
936		 * provides for more strict behavior to notice
937		 * unregistration errors.
938		 */
939		if (cur->vm_ops)
940			goto out_unlock;
941
942		found = true;
943	}
944	BUG_ON(!found);
945
946	if (vma->vm_start < start)
947		prev = vma;
948
949	ret = 0;
950	do {
951		cond_resched();
952
953		BUG_ON(vma->vm_ops);
954
955		/*
956		 * Nothing to do: this vma is already registered into this
957		 * userfaultfd and with the right tracking mode too.
958		 */
959		if (!vma->vm_userfaultfd_ctx.ctx)
960			goto skip;
961
962		if (vma->vm_start > start)
963			start = vma->vm_start;
964		vma_end = min(end, vma->vm_end);
965
966		new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
967		prev = vma_merge(mm, prev, start, vma_end, new_flags,
968				 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
969				 vma_policy(vma),
970				 NULL_VM_UFFD_CTX);
971		if (prev) {
972			vma = prev;
973			goto next;
974		}
975		if (vma->vm_start < start) {
976			ret = split_vma(mm, vma, start, 1);
977			if (ret)
978				break;
979		}
980		if (vma->vm_end > end) {
981			ret = split_vma(mm, vma, end, 0);
982			if (ret)
983				break;
984		}
985	next:
986		/*
987		 * In the vma_merge() successful mprotect-like case 8:
988		 * the next vma was merged into the current one and
989		 * the current one has not been updated yet.
990		 */
991		vma->vm_flags = new_flags;
992		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
993
994	skip:
995		prev = vma;
996		start = vma->vm_end;
997		vma = vma->vm_next;
998	} while (vma && vma->vm_start < end);
999out_unlock:
1000	up_write(&mm->mmap_sem);
1001out:
1002	return ret;
1003}
1004
1005/*
1006 * userfaultfd_wake may be used in combination with the
1007 * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1008 */
1009static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1010			    unsigned long arg)
1011{
1012	int ret;
1013	struct uffdio_range uffdio_wake;
1014	struct userfaultfd_wake_range range;
1015	const void __user *buf = (void __user *)arg;
1016
1017	ret = -EFAULT;
1018	if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1019		goto out;
1020
1021	ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1022	if (ret)
1023		goto out;
1024
1025	range.start = uffdio_wake.start;
1026	range.len = uffdio_wake.len;
1027
1028	/*
1029	 * len == 0 means wake all and we don't want to wake all here,
1030	 * so check it again to be sure.
1031	 */
1032	VM_BUG_ON(!range.len);
1033
1034	wake_userfault(ctx, &range);
1035	ret = 0;
1036
1037out:
1038	return ret;
1039}
1040
1041static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1042			    unsigned long arg)
1043{
1044	__s64 ret;
1045	struct uffdio_copy uffdio_copy;
1046	struct uffdio_copy __user *user_uffdio_copy;
1047	struct userfaultfd_wake_range range;
1048
1049	user_uffdio_copy = (struct uffdio_copy __user *) arg;
1050
1051	ret = -EFAULT;
1052	if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1053			   /* don't copy "copy" last field */
1054			   sizeof(uffdio_copy)-sizeof(__s64)))
1055		goto out;
1056
1057	ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1058	if (ret)
1059		goto out;
1060	/*
1061	 * double check for wraparound just in case. copy_from_user()
1062	 * will later check uffdio_copy.src + uffdio_copy.len to fit
1063	 * in the userland range.
1064	 */
1065	ret = -EINVAL;
1066	if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1067		goto out;
1068	if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1069		goto out;
1070
1071	ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1072			   uffdio_copy.len);
1073	if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1074		return -EFAULT;
1075	if (ret < 0)
1076		goto out;
1077	BUG_ON(!ret);
1078	/* len == 0 would wake all */
1079	range.len = ret;
1080	if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1081		range.start = uffdio_copy.dst;
1082		wake_userfault(ctx, &range);
1083	}
1084	ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1085out:
1086	return ret;
1087}
1088
1089static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1090				unsigned long arg)
1091{
1092	__s64 ret;
1093	struct uffdio_zeropage uffdio_zeropage;
1094	struct uffdio_zeropage __user *user_uffdio_zeropage;
1095	struct userfaultfd_wake_range range;
1096
1097	user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1098
1099	ret = -EFAULT;
1100	if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1101			   /* don't copy "zeropage" last field */
1102			   sizeof(uffdio_zeropage)-sizeof(__s64)))
1103		goto out;
1104
1105	ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1106			     uffdio_zeropage.range.len);
1107	if (ret)
1108		goto out;
1109	ret = -EINVAL;
1110	if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1111		goto out;
1112
1113	ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1114			     uffdio_zeropage.range.len);
1115	if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1116		return -EFAULT;
1117	if (ret < 0)
1118		goto out;
1119	/* len == 0 would wake all */
1120	BUG_ON(!ret);
1121	range.len = ret;
1122	if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1123		range.start = uffdio_zeropage.range.start;
1124		wake_userfault(ctx, &range);
1125	}
1126	ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1127out:
1128	return ret;
1129}
1130
1131/*
1132 * userland asks for a certain API version and we return which bits
1133 * and ioctl commands are implemented in this kernel for such API
1134 * version or -EINVAL if unknown.
1135 */
1136static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1137			   unsigned long arg)
1138{
1139	struct uffdio_api uffdio_api;
1140	void __user *buf = (void __user *)arg;
1141	int ret;
1142
1143	ret = -EINVAL;
1144	if (ctx->state != UFFD_STATE_WAIT_API)
1145		goto out;
1146	ret = -EFAULT;
1147	if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1148		goto out;
1149	if (uffdio_api.api != UFFD_API || uffdio_api.features) {
1150		memset(&uffdio_api, 0, sizeof(uffdio_api));
1151		if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1152			goto out;
1153		ret = -EINVAL;
1154		goto out;
1155	}
1156	uffdio_api.features = UFFD_API_FEATURES;
1157	uffdio_api.ioctls = UFFD_API_IOCTLS;
1158	ret = -EFAULT;
1159	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1160		goto out;
1161	ctx->state = UFFD_STATE_RUNNING;
1162	ret = 0;
1163out:
1164	return ret;
1165}
1166
1167static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1168			      unsigned long arg)
1169{
1170	int ret = -EINVAL;
1171	struct userfaultfd_ctx *ctx = file->private_data;
1172
1173	if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1174		return -EINVAL;
1175
1176	switch(cmd) {
1177	case UFFDIO_API:
1178		ret = userfaultfd_api(ctx, arg);
1179		break;
1180	case UFFDIO_REGISTER:
1181		ret = userfaultfd_register(ctx, arg);
1182		break;
1183	case UFFDIO_UNREGISTER:
1184		ret = userfaultfd_unregister(ctx, arg);
1185		break;
1186	case UFFDIO_WAKE:
1187		ret = userfaultfd_wake(ctx, arg);
1188		break;
1189	case UFFDIO_COPY:
1190		ret = userfaultfd_copy(ctx, arg);
1191		break;
1192	case UFFDIO_ZEROPAGE:
1193		ret = userfaultfd_zeropage(ctx, arg);
1194		break;
1195	}
1196	return ret;
1197}
1198
1199#ifdef CONFIG_PROC_FS
1200static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1201{
1202	struct userfaultfd_ctx *ctx = f->private_data;
1203	wait_queue_t *wq;
1204	struct userfaultfd_wait_queue *uwq;
1205	unsigned long pending = 0, total = 0;
1206
1207	spin_lock(&ctx->fault_pending_wqh.lock);
1208	list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1209		uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1210		pending++;
1211		total++;
1212	}
1213	list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1214		uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1215		total++;
1216	}
1217	spin_unlock(&ctx->fault_pending_wqh.lock);
1218
1219	/*
1220	 * If more protocols will be added, there will be all shown
1221	 * separated by a space. Like this:
1222	 *	protocols: aa:... bb:...
1223	 */
1224	seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1225		   pending, total, UFFD_API, UFFD_API_FEATURES,
1226		   UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1227}
1228#endif
1229
1230static const struct file_operations userfaultfd_fops = {
1231#ifdef CONFIG_PROC_FS
1232	.show_fdinfo	= userfaultfd_show_fdinfo,
1233#endif
1234	.release	= userfaultfd_release,
1235	.poll		= userfaultfd_poll,
1236	.read		= userfaultfd_read,
1237	.unlocked_ioctl = userfaultfd_ioctl,
1238	.compat_ioctl	= userfaultfd_ioctl,
1239	.llseek		= noop_llseek,
1240};
1241
1242static void init_once_userfaultfd_ctx(void *mem)
1243{
1244	struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1245
1246	init_waitqueue_head(&ctx->fault_pending_wqh);
1247	init_waitqueue_head(&ctx->fault_wqh);
1248	init_waitqueue_head(&ctx->fd_wqh);
1249	seqcount_init(&ctx->refile_seq);
1250}
1251
1252/**
1253 * userfaultfd_file_create - Creates an userfaultfd file pointer.
1254 * @flags: Flags for the userfaultfd file.
1255 *
1256 * This function creates an userfaultfd file pointer, w/out installing
1257 * it into the fd table. This is useful when the userfaultfd file is
1258 * used during the initialization of data structures that require
1259 * extra setup after the userfaultfd creation. So the userfaultfd
1260 * creation is split into the file pointer creation phase, and the
1261 * file descriptor installation phase.  In this way races with
1262 * userspace closing the newly installed file descriptor can be
1263 * avoided.  Returns an userfaultfd file pointer, or a proper error
1264 * pointer.
1265 */
1266static struct file *userfaultfd_file_create(int flags)
1267{
1268	struct file *file;
1269	struct userfaultfd_ctx *ctx;
1270
1271	BUG_ON(!current->mm);
1272
1273	/* Check the UFFD_* constants for consistency.  */
1274	BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1275	BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1276
1277	file = ERR_PTR(-EINVAL);
1278	if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1279		goto out;
1280
1281	file = ERR_PTR(-ENOMEM);
1282	ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1283	if (!ctx)
1284		goto out;
1285
1286	atomic_set(&ctx->refcount, 1);
1287	ctx->flags = flags;
1288	ctx->state = UFFD_STATE_WAIT_API;
1289	ctx->released = false;
1290	ctx->mm = current->mm;
1291	/* prevent the mm struct to be freed */
1292	atomic_inc(&ctx->mm->mm_users);
1293
1294	file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1295				  O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1296	if (IS_ERR(file)) {
1297		mmput(ctx->mm);
1298		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1299	}
1300out:
1301	return file;
1302}
1303
1304SYSCALL_DEFINE1(userfaultfd, int, flags)
1305{
1306	int fd, error;
1307	struct file *file;
1308
1309	error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1310	if (error < 0)
1311		return error;
1312	fd = error;
1313
1314	file = userfaultfd_file_create(flags);
1315	if (IS_ERR(file)) {
1316		error = PTR_ERR(file);
1317		goto err_put_unused_fd;
1318	}
1319	fd_install(fd, file);
1320
1321	return fd;
1322
1323err_put_unused_fd:
1324	put_unused_fd(fd);
1325
1326	return error;
1327}
1328
1329static int __init userfaultfd_init(void)
1330{
1331	userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1332						sizeof(struct userfaultfd_ctx),
1333						0,
1334						SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1335						init_once_userfaultfd_ctx);
1336	return 0;
1337}
1338__initcall(userfaultfd_init);
1339