xref: /linux/fs/userfaultfd.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
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 
30 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
31 
32 enum 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  */
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 */
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 
62 struct userfaultfd_wait_queue {
63 	struct uffd_msg msg;
64 	wait_queue_t wq;
65 	struct userfaultfd_ctx *ctx;
66 };
67 
68 struct userfaultfd_wake_range {
69 	unsigned long start;
70 	unsigned long len;
71 };
72 
73 static 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);
106 out:
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  */
117 static 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  */
131 static 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 
145 static 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 
155 static 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  */
191 static 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 
241 out:
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  */
260 int 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 	 * Check that we can return VM_FAULT_RETRY.
291 	 *
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.
298 	 */
299 	if (unlikely(!(flags & FAULT_FLAG_ALLOW_RETRY))) {
300 		/*
301 		 * Validate the invariant that nowait must allow retry
302 		 * to be sure not to return SIGBUS erroneously on
303 		 * nowait invocations.
304 		 */
305 		BUG_ON(flags & FAULT_FLAG_RETRY_NOWAIT);
306 #ifdef CONFIG_DEBUG_VM
307 		if (printk_ratelimit()) {
308 			printk(KERN_WARNING
309 			       "FAULT_FLAG_ALLOW_RETRY missing %x\n", flags);
310 			dump_stack();
311 		}
312 #endif
313 		goto out;
314 	}
315 
316 	/*
317 	 * Handle nowait, not much to do other than tell it to retry
318 	 * and wait.
319 	 */
320 	ret = VM_FAULT_RETRY;
321 	if (flags & FAULT_FLAG_RETRY_NOWAIT)
322 		goto out;
323 
324 	/* take the reference before dropping the mmap_sem */
325 	userfaultfd_ctx_get(ctx);
326 
327 	init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
328 	uwq.wq.private = current;
329 	uwq.msg = userfault_msg(address, flags, reason);
330 	uwq.ctx = ctx;
331 
332 	return_to_userland = (flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
333 		(FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
334 
335 	spin_lock(&ctx->fault_pending_wqh.lock);
336 	/*
337 	 * After the __add_wait_queue the uwq is visible to userland
338 	 * through poll/read().
339 	 */
340 	__add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
341 	/*
342 	 * The smp_mb() after __set_current_state prevents the reads
343 	 * following the spin_unlock to happen before the list_add in
344 	 * __add_wait_queue.
345 	 */
346 	set_current_state(return_to_userland ? TASK_INTERRUPTIBLE :
347 			  TASK_KILLABLE);
348 	spin_unlock(&ctx->fault_pending_wqh.lock);
349 
350 	must_wait = userfaultfd_must_wait(ctx, address, flags, reason);
351 	up_read(&mm->mmap_sem);
352 
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);
357 		schedule();
358 		ret |= VM_FAULT_MAJOR;
359 	}
360 
361 	__set_current_state(TASK_RUNNING);
362 
363 	if (return_to_userland) {
364 		if (signal_pending(current) &&
365 		    !fatal_signal_pending(current)) {
366 			/*
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
375 			 * VM_FAULT_RETRY).
376 			 *
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
380 			 * in such case.
381 			 */
382 			down_read(&mm->mmap_sem);
383 			ret = 0;
384 		}
385 	}
386 
387 	/*
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.
399 	 */
400 	if (!list_empty_careful(&uwq.wq.task_list)) {
401 		spin_lock(&ctx->fault_pending_wqh.lock);
402 		/*
403 		 * No need of list_del_init(), the uwq on the stack
404 		 * will be freed shortly anyway.
405 		 */
406 		list_del(&uwq.wq.task_list);
407 		spin_unlock(&ctx->fault_pending_wqh.lock);
408 	}
409 
410 	/*
411 	 * ctx may go away after this if the userfault pseudo fd is
412 	 * already released.
413 	 */
414 	userfaultfd_ctx_put(ctx);
415 
416 out:
417 	return ret;
418 }
419 
420 static int userfaultfd_release(struct inode *inode, struct file *file)
421 {
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;
428 
429 	ACCESS_ONCE(ctx->released) = true;
430 
431 	/*
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.
438 	 */
439 	down_write(&mm->mmap_sem);
440 	prev = NULL;
441 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
442 		cond_resched();
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) {
446 			prev = vma;
447 			continue;
448 		}
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,
453 				 vma_policy(vma),
454 				 NULL_VM_UFFD_CTX);
455 		if (prev)
456 			vma = prev;
457 		else
458 			prev = vma;
459 		vma->vm_flags = new_flags;
460 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
461 	}
462 	up_write(&mm->mmap_sem);
463 
464 	/*
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
467 	 * the fault_*wqh.
468 	 */
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);
473 
474 	wake_up_poll(&ctx->fd_wqh, POLLHUP);
475 	userfaultfd_ctx_put(ctx);
476 	return 0;
477 }
478 
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)
482 {
483 	wait_queue_t *wq;
484 	struct userfaultfd_wait_queue *uwq;
485 
486 	VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
487 
488 	uwq = NULL;
489 	if (!waitqueue_active(&ctx->fault_pending_wqh))
490 		goto out;
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);
495 out:
496 	return uwq;
497 }
498 
499 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
500 {
501 	struct userfaultfd_ctx *ctx = file->private_data;
502 	unsigned int ret;
503 
504 	poll_wait(file, &ctx->fd_wqh, wait);
505 
506 	switch (ctx->state) {
507 	case UFFD_STATE_WAIT_API:
508 		return POLLERR;
509 	case UFFD_STATE_RUNNING:
510 		/*
511 		 * poll() never guarantees that read won't block.
512 		 * userfaults can be waken before they're read().
513 		 */
514 		if (unlikely(!(file->f_flags & O_NONBLOCK)))
515 			return POLLERR;
516 		/*
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
524 		 * below.
525 		 */
526 		ret = 0;
527 		smp_mb();
528 		if (waitqueue_active(&ctx->fault_pending_wqh))
529 			ret = POLLIN;
530 		return ret;
531 	default:
532 		BUG();
533 	}
534 }
535 
536 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
537 				    struct uffd_msg *msg)
538 {
539 	ssize_t ret;
540 	DECLARE_WAITQUEUE(wait, current);
541 	struct userfaultfd_wait_queue *uwq;
542 
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);
546 	for (;;) {
547 		set_current_state(TASK_INTERRUPTIBLE);
548 		spin_lock(&ctx->fault_pending_wqh.lock);
549 		uwq = find_userfault(ctx);
550 		if (uwq) {
551 			/*
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
556 			 * only userfault.
557 			 */
558 			write_seqcount_begin(&ctx->refile_seq);
559 
560 			/*
561 			 * The fault_pending_wqh.lock prevents the uwq
562 			 * to disappear from under us.
563 			 *
564 			 * Refile this userfault from
565 			 * fault_pending_wqh to fault_wqh, it's not
566 			 * pending anymore after we read it.
567 			 *
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
579 			 * anyway.
580 			 */
581 			list_del(&uwq->wq.task_list);
582 			__add_wait_queue(&ctx->fault_wqh, &uwq->wq);
583 
584 			write_seqcount_end(&ctx->refile_seq);
585 
586 			/* careful to always initialize msg if ret == 0 */
587 			*msg = uwq->msg;
588 			spin_unlock(&ctx->fault_pending_wqh.lock);
589 			ret = 0;
590 			break;
591 		}
592 		spin_unlock(&ctx->fault_pending_wqh.lock);
593 		if (signal_pending(current)) {
594 			ret = -ERESTARTSYS;
595 			break;
596 		}
597 		if (no_wait) {
598 			ret = -EAGAIN;
599 			break;
600 		}
601 		spin_unlock(&ctx->fd_wqh.lock);
602 		schedule();
603 		spin_lock(&ctx->fd_wqh.lock);
604 	}
605 	__remove_wait_queue(&ctx->fd_wqh, &wait);
606 	__set_current_state(TASK_RUNNING);
607 	spin_unlock(&ctx->fd_wqh.lock);
608 
609 	return ret;
610 }
611 
612 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
613 				size_t count, loff_t *ppos)
614 {
615 	struct userfaultfd_ctx *ctx = file->private_data;
616 	ssize_t _ret, ret = 0;
617 	struct uffd_msg msg;
618 	int no_wait = file->f_flags & O_NONBLOCK;
619 
620 	if (ctx->state == UFFD_STATE_WAIT_API)
621 		return -EINVAL;
622 
623 	for (;;) {
624 		if (count < sizeof(msg))
625 			return ret ? ret : -EINVAL;
626 		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
627 		if (_ret < 0)
628 			return ret ? ret : _ret;
629 		if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
630 			return ret ? ret : -EFAULT;
631 		ret += sizeof(msg);
632 		buf += sizeof(msg);
633 		count -= sizeof(msg);
634 		/*
635 		 * Allow to read more than one fault at time but only
636 		 * block if waiting for the very first one.
637 		 */
638 		no_wait = O_NONBLOCK;
639 	}
640 }
641 
642 static void __wake_userfault(struct userfaultfd_ctx *ctx,
643 			     struct userfaultfd_wake_range *range)
644 {
645 	unsigned long start, end;
646 
647 	start = range->start;
648 	end = range->start + range->len;
649 
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,
654 				     range);
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);
658 }
659 
660 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
661 					   struct userfaultfd_wake_range *range)
662 {
663 	unsigned seq;
664 	bool need_wakeup;
665 
666 	/*
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.
672 	 */
673 	smp_mb();
674 
675 	/*
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.
680 	 */
681 	do {
682 		seq = read_seqcount_begin(&ctx->refile_seq);
683 		need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
684 			waitqueue_active(&ctx->fault_wqh);
685 		cond_resched();
686 	} while (read_seqcount_retry(&ctx->refile_seq, seq));
687 	if (need_wakeup)
688 		__wake_userfault(ctx, range);
689 }
690 
691 static __always_inline int validate_range(struct mm_struct *mm,
692 					  __u64 start, __u64 len)
693 {
694 	__u64 task_size = mm->task_size;
695 
696 	if (start & ~PAGE_MASK)
697 		return -EINVAL;
698 	if (len & ~PAGE_MASK)
699 		return -EINVAL;
700 	if (!len)
701 		return -EINVAL;
702 	if (start < mmap_min_addr)
703 		return -EINVAL;
704 	if (start >= task_size)
705 		return -EINVAL;
706 	if (len > task_size - start)
707 		return -EINVAL;
708 	return 0;
709 }
710 
711 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
712 				unsigned long arg)
713 {
714 	struct mm_struct *mm = ctx->mm;
715 	struct vm_area_struct *vma, *prev, *cur;
716 	int ret;
717 	struct uffdio_register uffdio_register;
718 	struct uffdio_register __user *user_uffdio_register;
719 	unsigned long vm_flags, new_flags;
720 	bool found;
721 	unsigned long start, end, vma_end;
722 
723 	user_uffdio_register = (struct uffdio_register __user *) arg;
724 
725 	ret = -EFAULT;
726 	if (copy_from_user(&uffdio_register, user_uffdio_register,
727 			   sizeof(uffdio_register)-sizeof(__u64)))
728 		goto out;
729 
730 	ret = -EINVAL;
731 	if (!uffdio_register.mode)
732 		goto out;
733 	if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
734 				     UFFDIO_REGISTER_MODE_WP))
735 		goto out;
736 	vm_flags = 0;
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;
741 		/*
742 		 * FIXME: remove the below error constraint by
743 		 * implementing the wprotect tracking mode.
744 		 */
745 		ret = -EINVAL;
746 		goto out;
747 	}
748 
749 	ret = validate_range(mm, uffdio_register.range.start,
750 			     uffdio_register.range.len);
751 	if (ret)
752 		goto out;
753 
754 	start = uffdio_register.range.start;
755 	end = start + uffdio_register.range.len;
756 
757 	down_write(&mm->mmap_sem);
758 	vma = find_vma_prev(mm, start, &prev);
759 
760 	ret = -ENOMEM;
761 	if (!vma)
762 		goto out_unlock;
763 
764 	/* check that there's at least one vma in the range */
765 	ret = -EINVAL;
766 	if (vma->vm_start >= end)
767 		goto out_unlock;
768 
769 	/*
770 	 * Search for not compatible vmas.
771 	 *
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).
775 	 */
776 	found = false;
777 	for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
778 		cond_resched();
779 
780 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
781 		       !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
782 
783 		/* check not compatible vmas */
784 		ret = -EINVAL;
785 		if (cur->vm_ops)
786 			goto out_unlock;
787 
788 		/*
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.
793 		 */
794 		ret = -EBUSY;
795 		if (cur->vm_userfaultfd_ctx.ctx &&
796 		    cur->vm_userfaultfd_ctx.ctx != ctx)
797 			goto out_unlock;
798 
799 		found = true;
800 	}
801 	BUG_ON(!found);
802 
803 	if (vma->vm_start < start)
804 		prev = vma;
805 
806 	ret = 0;
807 	do {
808 		cond_resched();
809 
810 		BUG_ON(vma->vm_ops);
811 		BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
812 		       vma->vm_userfaultfd_ctx.ctx != ctx);
813 
814 		/*
815 		 * Nothing to do: this vma is already registered into this
816 		 * userfaultfd and with the right tracking mode too.
817 		 */
818 		if (vma->vm_userfaultfd_ctx.ctx == ctx &&
819 		    (vma->vm_flags & vm_flags) == vm_flags)
820 			goto skip;
821 
822 		if (vma->vm_start > start)
823 			start = vma->vm_start;
824 		vma_end = min(end, vma->vm_end);
825 
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,
829 				 vma_policy(vma),
830 				 ((struct vm_userfaultfd_ctx){ ctx }));
831 		if (prev) {
832 			vma = prev;
833 			goto next;
834 		}
835 		if (vma->vm_start < start) {
836 			ret = split_vma(mm, vma, start, 1);
837 			if (ret)
838 				break;
839 		}
840 		if (vma->vm_end > end) {
841 			ret = split_vma(mm, vma, end, 0);
842 			if (ret)
843 				break;
844 		}
845 	next:
846 		/*
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.
850 		 */
851 		vma->vm_flags = new_flags;
852 		vma->vm_userfaultfd_ctx.ctx = ctx;
853 
854 	skip:
855 		prev = vma;
856 		start = vma->vm_end;
857 		vma = vma->vm_next;
858 	} while (vma && vma->vm_start < end);
859 out_unlock:
860 	up_write(&mm->mmap_sem);
861 	if (!ret) {
862 		/*
863 		 * Now that we scanned all vmas we can already tell
864 		 * userland which ioctls methods are guaranteed to
865 		 * succeed on this range.
866 		 */
867 		if (put_user(UFFD_API_RANGE_IOCTLS,
868 			     &user_uffdio_register->ioctls))
869 			ret = -EFAULT;
870 	}
871 out:
872 	return ret;
873 }
874 
875 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
876 				  unsigned long arg)
877 {
878 	struct mm_struct *mm = ctx->mm;
879 	struct vm_area_struct *vma, *prev, *cur;
880 	int ret;
881 	struct uffdio_range uffdio_unregister;
882 	unsigned long new_flags;
883 	bool found;
884 	unsigned long start, end, vma_end;
885 	const void __user *buf = (void __user *)arg;
886 
887 	ret = -EFAULT;
888 	if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
889 		goto out;
890 
891 	ret = validate_range(mm, uffdio_unregister.start,
892 			     uffdio_unregister.len);
893 	if (ret)
894 		goto out;
895 
896 	start = uffdio_unregister.start;
897 	end = start + uffdio_unregister.len;
898 
899 	down_write(&mm->mmap_sem);
900 	vma = find_vma_prev(mm, start, &prev);
901 
902 	ret = -ENOMEM;
903 	if (!vma)
904 		goto out_unlock;
905 
906 	/* check that there's at least one vma in the range */
907 	ret = -EINVAL;
908 	if (vma->vm_start >= end)
909 		goto out_unlock;
910 
911 	/*
912 	 * Search for not compatible vmas.
913 	 *
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).
917 	 */
918 	found = false;
919 	ret = -EINVAL;
920 	for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
921 		cond_resched();
922 
923 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
924 		       !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
925 
926 		/*
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.
932 		 */
933 		if (cur->vm_ops)
934 			goto out_unlock;
935 
936 		found = true;
937 	}
938 	BUG_ON(!found);
939 
940 	if (vma->vm_start < start)
941 		prev = vma;
942 
943 	ret = 0;
944 	do {
945 		cond_resched();
946 
947 		BUG_ON(vma->vm_ops);
948 
949 		/*
950 		 * Nothing to do: this vma is already registered into this
951 		 * userfaultfd and with the right tracking mode too.
952 		 */
953 		if (!vma->vm_userfaultfd_ctx.ctx)
954 			goto skip;
955 
956 		if (vma->vm_start > start)
957 			start = vma->vm_start;
958 		vma_end = min(end, vma->vm_end);
959 
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,
963 				 vma_policy(vma),
964 				 NULL_VM_UFFD_CTX);
965 		if (prev) {
966 			vma = prev;
967 			goto next;
968 		}
969 		if (vma->vm_start < start) {
970 			ret = split_vma(mm, vma, start, 1);
971 			if (ret)
972 				break;
973 		}
974 		if (vma->vm_end > end) {
975 			ret = split_vma(mm, vma, end, 0);
976 			if (ret)
977 				break;
978 		}
979 	next:
980 		/*
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.
984 		 */
985 		vma->vm_flags = new_flags;
986 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
987 
988 	skip:
989 		prev = vma;
990 		start = vma->vm_end;
991 		vma = vma->vm_next;
992 	} while (vma && vma->vm_start < end);
993 out_unlock:
994 	up_write(&mm->mmap_sem);
995 out:
996 	return ret;
997 }
998 
999 /*
1000  * userfaultfd_wake may be used in combination with the
1001  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1002  */
1003 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1004 			    unsigned long arg)
1005 {
1006 	int ret;
1007 	struct uffdio_range uffdio_wake;
1008 	struct userfaultfd_wake_range range;
1009 	const void __user *buf = (void __user *)arg;
1010 
1011 	ret = -EFAULT;
1012 	if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1013 		goto out;
1014 
1015 	ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1016 	if (ret)
1017 		goto out;
1018 
1019 	range.start = uffdio_wake.start;
1020 	range.len = uffdio_wake.len;
1021 
1022 	/*
1023 	 * len == 0 means wake all and we don't want to wake all here,
1024 	 * so check it again to be sure.
1025 	 */
1026 	VM_BUG_ON(!range.len);
1027 
1028 	wake_userfault(ctx, &range);
1029 	ret = 0;
1030 
1031 out:
1032 	return ret;
1033 }
1034 
1035 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1036 			    unsigned long arg)
1037 {
1038 	__s64 ret;
1039 	struct uffdio_copy uffdio_copy;
1040 	struct uffdio_copy __user *user_uffdio_copy;
1041 	struct userfaultfd_wake_range range;
1042 
1043 	user_uffdio_copy = (struct uffdio_copy __user *) arg;
1044 
1045 	ret = -EFAULT;
1046 	if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1047 			   /* don't copy "copy" last field */
1048 			   sizeof(uffdio_copy)-sizeof(__s64)))
1049 		goto out;
1050 
1051 	ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1052 	if (ret)
1053 		goto out;
1054 	/*
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.
1058 	 */
1059 	ret = -EINVAL;
1060 	if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1061 		goto out;
1062 	if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1063 		goto out;
1064 
1065 	ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1066 			   uffdio_copy.len);
1067 	if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1068 		return -EFAULT;
1069 	if (ret < 0)
1070 		goto out;
1071 	BUG_ON(!ret);
1072 	/* len == 0 would wake all */
1073 	range.len = ret;
1074 	if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1075 		range.start = uffdio_copy.dst;
1076 		wake_userfault(ctx, &range);
1077 	}
1078 	ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1079 out:
1080 	return ret;
1081 }
1082 
1083 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1084 				unsigned long arg)
1085 {
1086 	__s64 ret;
1087 	struct uffdio_zeropage uffdio_zeropage;
1088 	struct uffdio_zeropage __user *user_uffdio_zeropage;
1089 	struct userfaultfd_wake_range range;
1090 
1091 	user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1092 
1093 	ret = -EFAULT;
1094 	if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1095 			   /* don't copy "zeropage" last field */
1096 			   sizeof(uffdio_zeropage)-sizeof(__s64)))
1097 		goto out;
1098 
1099 	ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1100 			     uffdio_zeropage.range.len);
1101 	if (ret)
1102 		goto out;
1103 	ret = -EINVAL;
1104 	if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1105 		goto out;
1106 
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)))
1110 		return -EFAULT;
1111 	if (ret < 0)
1112 		goto out;
1113 	/* len == 0 would wake all */
1114 	BUG_ON(!ret);
1115 	range.len = ret;
1116 	if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1117 		range.start = uffdio_zeropage.range.start;
1118 		wake_userfault(ctx, &range);
1119 	}
1120 	ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1121 out:
1122 	return ret;
1123 }
1124 
1125 /*
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.
1129  */
1130 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1131 			   unsigned long arg)
1132 {
1133 	struct uffdio_api uffdio_api;
1134 	void __user *buf = (void __user *)arg;
1135 	int ret;
1136 
1137 	ret = -EINVAL;
1138 	if (ctx->state != UFFD_STATE_WAIT_API)
1139 		goto out;
1140 	ret = -EFAULT;
1141 	if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1142 		goto out;
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)))
1146 			goto out;
1147 		ret = -EINVAL;
1148 		goto out;
1149 	}
1150 	uffdio_api.features = UFFD_API_FEATURES;
1151 	uffdio_api.ioctls = UFFD_API_IOCTLS;
1152 	ret = -EFAULT;
1153 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1154 		goto out;
1155 	ctx->state = UFFD_STATE_RUNNING;
1156 	ret = 0;
1157 out:
1158 	return ret;
1159 }
1160 
1161 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1162 			      unsigned long arg)
1163 {
1164 	int ret = -EINVAL;
1165 	struct userfaultfd_ctx *ctx = file->private_data;
1166 
1167 	if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1168 		return -EINVAL;
1169 
1170 	switch(cmd) {
1171 	case UFFDIO_API:
1172 		ret = userfaultfd_api(ctx, arg);
1173 		break;
1174 	case UFFDIO_REGISTER:
1175 		ret = userfaultfd_register(ctx, arg);
1176 		break;
1177 	case UFFDIO_UNREGISTER:
1178 		ret = userfaultfd_unregister(ctx, arg);
1179 		break;
1180 	case UFFDIO_WAKE:
1181 		ret = userfaultfd_wake(ctx, arg);
1182 		break;
1183 	case UFFDIO_COPY:
1184 		ret = userfaultfd_copy(ctx, arg);
1185 		break;
1186 	case UFFDIO_ZEROPAGE:
1187 		ret = userfaultfd_zeropage(ctx, arg);
1188 		break;
1189 	}
1190 	return ret;
1191 }
1192 
1193 #ifdef CONFIG_PROC_FS
1194 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1195 {
1196 	struct userfaultfd_ctx *ctx = f->private_data;
1197 	wait_queue_t *wq;
1198 	struct userfaultfd_wait_queue *uwq;
1199 	unsigned long pending = 0, total = 0;
1200 
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);
1204 		pending++;
1205 		total++;
1206 	}
1207 	list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1208 		uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1209 		total++;
1210 	}
1211 	spin_unlock(&ctx->fault_pending_wqh.lock);
1212 
1213 	/*
1214 	 * If more protocols will be added, there will be all shown
1215 	 * separated by a space. Like this:
1216 	 *	protocols: aa:... bb:...
1217 	 */
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);
1221 }
1222 #endif
1223 
1224 static const struct file_operations userfaultfd_fops = {
1225 #ifdef CONFIG_PROC_FS
1226 	.show_fdinfo	= userfaultfd_show_fdinfo,
1227 #endif
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,
1234 };
1235 
1236 static void init_once_userfaultfd_ctx(void *mem)
1237 {
1238 	struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1239 
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);
1244 }
1245 
1246 /**
1247  * userfaultfd_file_create - Creates an userfaultfd file pointer.
1248  * @flags: Flags for the userfaultfd file.
1249  *
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
1258  * pointer.
1259  */
1260 static struct file *userfaultfd_file_create(int flags)
1261 {
1262 	struct file *file;
1263 	struct userfaultfd_ctx *ctx;
1264 
1265 	BUG_ON(!current->mm);
1266 
1267 	/* Check the UFFD_* constants for consistency.  */
1268 	BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1269 	BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1270 
1271 	file = ERR_PTR(-EINVAL);
1272 	if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1273 		goto out;
1274 
1275 	file = ERR_PTR(-ENOMEM);
1276 	ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1277 	if (!ctx)
1278 		goto out;
1279 
1280 	atomic_set(&ctx->refcount, 1);
1281 	ctx->flags = flags;
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);
1287 
1288 	file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1289 				  O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1290 	if (IS_ERR(file)) {
1291 		mmput(ctx->mm);
1292 		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1293 	}
1294 out:
1295 	return file;
1296 }
1297 
1298 SYSCALL_DEFINE1(userfaultfd, int, flags)
1299 {
1300 	int fd, error;
1301 	struct file *file;
1302 
1303 	error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1304 	if (error < 0)
1305 		return error;
1306 	fd = error;
1307 
1308 	file = userfaultfd_file_create(flags);
1309 	if (IS_ERR(file)) {
1310 		error = PTR_ERR(file);
1311 		goto err_put_unused_fd;
1312 	}
1313 	fd_install(fd, file);
1314 
1315 	return fd;
1316 
1317 err_put_unused_fd:
1318 	put_unused_fd(fd);
1319 
1320 	return error;
1321 }
1322 
1323 static int __init userfaultfd_init(void)
1324 {
1325 	userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1326 						sizeof(struct userfaultfd_ctx),
1327 						0,
1328 						SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1329 						init_once_userfaultfd_ctx);
1330 	return 0;
1331 }
1332 __initcall(userfaultfd_init);
1333