xref: /openbmc/linux/fs/userfaultfd.c (revision 6726d552)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  fs/userfaultfd.c
4  *
5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
6  *  Copyright (C) 2008-2009 Red Hat, Inc.
7  *  Copyright (C) 2015  Red Hat, Inc.
8  *
9  *  Some part derived from fs/eventfd.c (anon inode setup) and
10  *  mm/ksm.c (mm hashing).
11  */
12 
13 #include <linux/list.h>
14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/mm.h>
17 #include <linux/mm.h>
18 #include <linux/mm_inline.h>
19 #include <linux/mmu_notifier.h>
20 #include <linux/poll.h>
21 #include <linux/slab.h>
22 #include <linux/seq_file.h>
23 #include <linux/file.h>
24 #include <linux/bug.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/syscalls.h>
27 #include <linux/userfaultfd_k.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ioctl.h>
30 #include <linux/security.h>
31 #include <linux/hugetlb.h>
32 #include <linux/swapops.h>
33 
34 int sysctl_unprivileged_userfaultfd __read_mostly;
35 
36 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
37 
38 /*
39  * Start with fault_pending_wqh and fault_wqh so they're more likely
40  * to be in the same cacheline.
41  *
42  * Locking order:
43  *	fd_wqh.lock
44  *		fault_pending_wqh.lock
45  *			fault_wqh.lock
46  *		event_wqh.lock
47  *
48  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
49  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
50  * also taken in IRQ context.
51  */
52 struct userfaultfd_ctx {
53 	/* waitqueue head for the pending (i.e. not read) userfaults */
54 	wait_queue_head_t fault_pending_wqh;
55 	/* waitqueue head for the userfaults */
56 	wait_queue_head_t fault_wqh;
57 	/* waitqueue head for the pseudo fd to wakeup poll/read */
58 	wait_queue_head_t fd_wqh;
59 	/* waitqueue head for events */
60 	wait_queue_head_t event_wqh;
61 	/* a refile sequence protected by fault_pending_wqh lock */
62 	seqcount_spinlock_t refile_seq;
63 	/* pseudo fd refcounting */
64 	refcount_t refcount;
65 	/* userfaultfd syscall flags */
66 	unsigned int flags;
67 	/* features requested from the userspace */
68 	unsigned int features;
69 	/* released */
70 	bool released;
71 	/* memory mappings are changing because of non-cooperative event */
72 	atomic_t mmap_changing;
73 	/* mm with one ore more vmas attached to this userfaultfd_ctx */
74 	struct mm_struct *mm;
75 };
76 
77 struct userfaultfd_fork_ctx {
78 	struct userfaultfd_ctx *orig;
79 	struct userfaultfd_ctx *new;
80 	struct list_head list;
81 };
82 
83 struct userfaultfd_unmap_ctx {
84 	struct userfaultfd_ctx *ctx;
85 	unsigned long start;
86 	unsigned long end;
87 	struct list_head list;
88 };
89 
90 struct userfaultfd_wait_queue {
91 	struct uffd_msg msg;
92 	wait_queue_entry_t wq;
93 	struct userfaultfd_ctx *ctx;
94 	bool waken;
95 };
96 
97 struct userfaultfd_wake_range {
98 	unsigned long start;
99 	unsigned long len;
100 };
101 
102 /* internal indication that UFFD_API ioctl was successfully executed */
103 #define UFFD_FEATURE_INITIALIZED		(1u << 31)
104 
105 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
106 {
107 	return ctx->features & UFFD_FEATURE_INITIALIZED;
108 }
109 
110 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
111 				     int wake_flags, void *key)
112 {
113 	struct userfaultfd_wake_range *range = key;
114 	int ret;
115 	struct userfaultfd_wait_queue *uwq;
116 	unsigned long start, len;
117 
118 	uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
119 	ret = 0;
120 	/* len == 0 means wake all */
121 	start = range->start;
122 	len = range->len;
123 	if (len && (start > uwq->msg.arg.pagefault.address ||
124 		    start + len <= uwq->msg.arg.pagefault.address))
125 		goto out;
126 	WRITE_ONCE(uwq->waken, true);
127 	/*
128 	 * The Program-Order guarantees provided by the scheduler
129 	 * ensure uwq->waken is visible before the task is woken.
130 	 */
131 	ret = wake_up_state(wq->private, mode);
132 	if (ret) {
133 		/*
134 		 * Wake only once, autoremove behavior.
135 		 *
136 		 * After the effect of list_del_init is visible to the other
137 		 * CPUs, the waitqueue may disappear from under us, see the
138 		 * !list_empty_careful() in handle_userfault().
139 		 *
140 		 * try_to_wake_up() has an implicit smp_mb(), and the
141 		 * wq->private is read before calling the extern function
142 		 * "wake_up_state" (which in turns calls try_to_wake_up).
143 		 */
144 		list_del_init(&wq->entry);
145 	}
146 out:
147 	return ret;
148 }
149 
150 /**
151  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
152  * context.
153  * @ctx: [in] Pointer to the userfaultfd context.
154  */
155 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
156 {
157 	refcount_inc(&ctx->refcount);
158 }
159 
160 /**
161  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
162  * context.
163  * @ctx: [in] Pointer to userfaultfd context.
164  *
165  * The userfaultfd context reference must have been previously acquired either
166  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
167  */
168 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
169 {
170 	if (refcount_dec_and_test(&ctx->refcount)) {
171 		VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
172 		VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
173 		VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
174 		VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
175 		VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
176 		VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
177 		VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
178 		VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
179 		mmdrop(ctx->mm);
180 		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
181 	}
182 }
183 
184 static inline void msg_init(struct uffd_msg *msg)
185 {
186 	BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
187 	/*
188 	 * Must use memset to zero out the paddings or kernel data is
189 	 * leaked to userland.
190 	 */
191 	memset(msg, 0, sizeof(struct uffd_msg));
192 }
193 
194 static inline struct uffd_msg userfault_msg(unsigned long address,
195 					    unsigned long real_address,
196 					    unsigned int flags,
197 					    unsigned long reason,
198 					    unsigned int features)
199 {
200 	struct uffd_msg msg;
201 
202 	msg_init(&msg);
203 	msg.event = UFFD_EVENT_PAGEFAULT;
204 
205 	msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
206 				    real_address : address;
207 
208 	/*
209 	 * These flags indicate why the userfault occurred:
210 	 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
211 	 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
212 	 * - Neither of these flags being set indicates a MISSING fault.
213 	 *
214 	 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
215 	 * fault. Otherwise, it was a read fault.
216 	 */
217 	if (flags & FAULT_FLAG_WRITE)
218 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
219 	if (reason & VM_UFFD_WP)
220 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
221 	if (reason & VM_UFFD_MINOR)
222 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
223 	if (features & UFFD_FEATURE_THREAD_ID)
224 		msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
225 	return msg;
226 }
227 
228 #ifdef CONFIG_HUGETLB_PAGE
229 /*
230  * Same functionality as userfaultfd_must_wait below with modifications for
231  * hugepmd ranges.
232  */
233 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
234 					 struct vm_area_struct *vma,
235 					 unsigned long address,
236 					 unsigned long flags,
237 					 unsigned long reason)
238 {
239 	struct mm_struct *mm = ctx->mm;
240 	pte_t *ptep, pte;
241 	bool ret = true;
242 
243 	mmap_assert_locked(mm);
244 
245 	ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
246 
247 	if (!ptep)
248 		goto out;
249 
250 	ret = false;
251 	pte = huge_ptep_get(ptep);
252 
253 	/*
254 	 * Lockless access: we're in a wait_event so it's ok if it
255 	 * changes under us.  PTE markers should be handled the same as none
256 	 * ptes here.
257 	 */
258 	if (huge_pte_none_mostly(pte))
259 		ret = true;
260 	if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
261 		ret = true;
262 out:
263 	return ret;
264 }
265 #else
266 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
267 					 struct vm_area_struct *vma,
268 					 unsigned long address,
269 					 unsigned long flags,
270 					 unsigned long reason)
271 {
272 	return false;	/* should never get here */
273 }
274 #endif /* CONFIG_HUGETLB_PAGE */
275 
276 /*
277  * Verify the pagetables are still not ok after having reigstered into
278  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
279  * userfault that has already been resolved, if userfaultfd_read and
280  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
281  * threads.
282  */
283 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
284 					 unsigned long address,
285 					 unsigned long flags,
286 					 unsigned long reason)
287 {
288 	struct mm_struct *mm = ctx->mm;
289 	pgd_t *pgd;
290 	p4d_t *p4d;
291 	pud_t *pud;
292 	pmd_t *pmd, _pmd;
293 	pte_t *pte;
294 	bool ret = true;
295 
296 	mmap_assert_locked(mm);
297 
298 	pgd = pgd_offset(mm, address);
299 	if (!pgd_present(*pgd))
300 		goto out;
301 	p4d = p4d_offset(pgd, address);
302 	if (!p4d_present(*p4d))
303 		goto out;
304 	pud = pud_offset(p4d, address);
305 	if (!pud_present(*pud))
306 		goto out;
307 	pmd = pmd_offset(pud, address);
308 	/*
309 	 * READ_ONCE must function as a barrier with narrower scope
310 	 * and it must be equivalent to:
311 	 *	_pmd = *pmd; barrier();
312 	 *
313 	 * This is to deal with the instability (as in
314 	 * pmd_trans_unstable) of the pmd.
315 	 */
316 	_pmd = READ_ONCE(*pmd);
317 	if (pmd_none(_pmd))
318 		goto out;
319 
320 	ret = false;
321 	if (!pmd_present(_pmd))
322 		goto out;
323 
324 	if (pmd_trans_huge(_pmd)) {
325 		if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
326 			ret = true;
327 		goto out;
328 	}
329 
330 	/*
331 	 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
332 	 * and use the standard pte_offset_map() instead of parsing _pmd.
333 	 */
334 	pte = pte_offset_map(pmd, address);
335 	/*
336 	 * Lockless access: we're in a wait_event so it's ok if it
337 	 * changes under us.  PTE markers should be handled the same as none
338 	 * ptes here.
339 	 */
340 	if (pte_none_mostly(*pte))
341 		ret = true;
342 	if (!pte_write(*pte) && (reason & VM_UFFD_WP))
343 		ret = true;
344 	pte_unmap(pte);
345 
346 out:
347 	return ret;
348 }
349 
350 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
351 {
352 	if (flags & FAULT_FLAG_INTERRUPTIBLE)
353 		return TASK_INTERRUPTIBLE;
354 
355 	if (flags & FAULT_FLAG_KILLABLE)
356 		return TASK_KILLABLE;
357 
358 	return TASK_UNINTERRUPTIBLE;
359 }
360 
361 /*
362  * The locking rules involved in returning VM_FAULT_RETRY depending on
363  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
364  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
365  * recommendation in __lock_page_or_retry is not an understatement.
366  *
367  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
368  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
369  * not set.
370  *
371  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
372  * set, VM_FAULT_RETRY can still be returned if and only if there are
373  * fatal_signal_pending()s, and the mmap_lock must be released before
374  * returning it.
375  */
376 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
377 {
378 	struct mm_struct *mm = vmf->vma->vm_mm;
379 	struct userfaultfd_ctx *ctx;
380 	struct userfaultfd_wait_queue uwq;
381 	vm_fault_t ret = VM_FAULT_SIGBUS;
382 	bool must_wait;
383 	unsigned int blocking_state;
384 
385 	/*
386 	 * We don't do userfault handling for the final child pid update.
387 	 *
388 	 * We also don't do userfault handling during
389 	 * coredumping. hugetlbfs has the special
390 	 * follow_hugetlb_page() to skip missing pages in the
391 	 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
392 	 * the no_page_table() helper in follow_page_mask(), but the
393 	 * shmem_vm_ops->fault method is invoked even during
394 	 * coredumping without mmap_lock and it ends up here.
395 	 */
396 	if (current->flags & (PF_EXITING|PF_DUMPCORE))
397 		goto out;
398 
399 	/*
400 	 * Coredumping runs without mmap_lock so we can only check that
401 	 * the mmap_lock is held, if PF_DUMPCORE was not set.
402 	 */
403 	mmap_assert_locked(mm);
404 
405 	ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
406 	if (!ctx)
407 		goto out;
408 
409 	BUG_ON(ctx->mm != mm);
410 
411 	/* Any unrecognized flag is a bug. */
412 	VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
413 	/* 0 or > 1 flags set is a bug; we expect exactly 1. */
414 	VM_BUG_ON(!reason || (reason & (reason - 1)));
415 
416 	if (ctx->features & UFFD_FEATURE_SIGBUS)
417 		goto out;
418 	if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
419 	    ctx->flags & UFFD_USER_MODE_ONLY) {
420 		printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
421 			"sysctl knob to 1 if kernel faults must be handled "
422 			"without obtaining CAP_SYS_PTRACE capability\n");
423 		goto out;
424 	}
425 
426 	/*
427 	 * If it's already released don't get it. This avoids to loop
428 	 * in __get_user_pages if userfaultfd_release waits on the
429 	 * caller of handle_userfault to release the mmap_lock.
430 	 */
431 	if (unlikely(READ_ONCE(ctx->released))) {
432 		/*
433 		 * Don't return VM_FAULT_SIGBUS in this case, so a non
434 		 * cooperative manager can close the uffd after the
435 		 * last UFFDIO_COPY, without risking to trigger an
436 		 * involuntary SIGBUS if the process was starting the
437 		 * userfaultfd while the userfaultfd was still armed
438 		 * (but after the last UFFDIO_COPY). If the uffd
439 		 * wasn't already closed when the userfault reached
440 		 * this point, that would normally be solved by
441 		 * userfaultfd_must_wait returning 'false'.
442 		 *
443 		 * If we were to return VM_FAULT_SIGBUS here, the non
444 		 * cooperative manager would be instead forced to
445 		 * always call UFFDIO_UNREGISTER before it can safely
446 		 * close the uffd.
447 		 */
448 		ret = VM_FAULT_NOPAGE;
449 		goto out;
450 	}
451 
452 	/*
453 	 * Check that we can return VM_FAULT_RETRY.
454 	 *
455 	 * NOTE: it should become possible to return VM_FAULT_RETRY
456 	 * even if FAULT_FLAG_TRIED is set without leading to gup()
457 	 * -EBUSY failures, if the userfaultfd is to be extended for
458 	 * VM_UFFD_WP tracking and we intend to arm the userfault
459 	 * without first stopping userland access to the memory. For
460 	 * VM_UFFD_MISSING userfaults this is enough for now.
461 	 */
462 	if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
463 		/*
464 		 * Validate the invariant that nowait must allow retry
465 		 * to be sure not to return SIGBUS erroneously on
466 		 * nowait invocations.
467 		 */
468 		BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
469 #ifdef CONFIG_DEBUG_VM
470 		if (printk_ratelimit()) {
471 			printk(KERN_WARNING
472 			       "FAULT_FLAG_ALLOW_RETRY missing %x\n",
473 			       vmf->flags);
474 			dump_stack();
475 		}
476 #endif
477 		goto out;
478 	}
479 
480 	/*
481 	 * Handle nowait, not much to do other than tell it to retry
482 	 * and wait.
483 	 */
484 	ret = VM_FAULT_RETRY;
485 	if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
486 		goto out;
487 
488 	/* take the reference before dropping the mmap_lock */
489 	userfaultfd_ctx_get(ctx);
490 
491 	init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
492 	uwq.wq.private = current;
493 	uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
494 				reason, ctx->features);
495 	uwq.ctx = ctx;
496 	uwq.waken = false;
497 
498 	blocking_state = userfaultfd_get_blocking_state(vmf->flags);
499 
500 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
501 	/*
502 	 * After the __add_wait_queue the uwq is visible to userland
503 	 * through poll/read().
504 	 */
505 	__add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
506 	/*
507 	 * The smp_mb() after __set_current_state prevents the reads
508 	 * following the spin_unlock to happen before the list_add in
509 	 * __add_wait_queue.
510 	 */
511 	set_current_state(blocking_state);
512 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
513 
514 	if (!is_vm_hugetlb_page(vmf->vma))
515 		must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
516 						  reason);
517 	else
518 		must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
519 						       vmf->address,
520 						       vmf->flags, reason);
521 	mmap_read_unlock(mm);
522 
523 	if (likely(must_wait && !READ_ONCE(ctx->released))) {
524 		wake_up_poll(&ctx->fd_wqh, EPOLLIN);
525 		schedule();
526 	}
527 
528 	__set_current_state(TASK_RUNNING);
529 
530 	/*
531 	 * Here we race with the list_del; list_add in
532 	 * userfaultfd_ctx_read(), however because we don't ever run
533 	 * list_del_init() to refile across the two lists, the prev
534 	 * and next pointers will never point to self. list_add also
535 	 * would never let any of the two pointers to point to
536 	 * self. So list_empty_careful won't risk to see both pointers
537 	 * pointing to self at any time during the list refile. The
538 	 * only case where list_del_init() is called is the full
539 	 * removal in the wake function and there we don't re-list_add
540 	 * and it's fine not to block on the spinlock. The uwq on this
541 	 * kernel stack can be released after the list_del_init.
542 	 */
543 	if (!list_empty_careful(&uwq.wq.entry)) {
544 		spin_lock_irq(&ctx->fault_pending_wqh.lock);
545 		/*
546 		 * No need of list_del_init(), the uwq on the stack
547 		 * will be freed shortly anyway.
548 		 */
549 		list_del(&uwq.wq.entry);
550 		spin_unlock_irq(&ctx->fault_pending_wqh.lock);
551 	}
552 
553 	/*
554 	 * ctx may go away after this if the userfault pseudo fd is
555 	 * already released.
556 	 */
557 	userfaultfd_ctx_put(ctx);
558 
559 out:
560 	return ret;
561 }
562 
563 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
564 					      struct userfaultfd_wait_queue *ewq)
565 {
566 	struct userfaultfd_ctx *release_new_ctx;
567 
568 	if (WARN_ON_ONCE(current->flags & PF_EXITING))
569 		goto out;
570 
571 	ewq->ctx = ctx;
572 	init_waitqueue_entry(&ewq->wq, current);
573 	release_new_ctx = NULL;
574 
575 	spin_lock_irq(&ctx->event_wqh.lock);
576 	/*
577 	 * After the __add_wait_queue the uwq is visible to userland
578 	 * through poll/read().
579 	 */
580 	__add_wait_queue(&ctx->event_wqh, &ewq->wq);
581 	for (;;) {
582 		set_current_state(TASK_KILLABLE);
583 		if (ewq->msg.event == 0)
584 			break;
585 		if (READ_ONCE(ctx->released) ||
586 		    fatal_signal_pending(current)) {
587 			/*
588 			 * &ewq->wq may be queued in fork_event, but
589 			 * __remove_wait_queue ignores the head
590 			 * parameter. It would be a problem if it
591 			 * didn't.
592 			 */
593 			__remove_wait_queue(&ctx->event_wqh, &ewq->wq);
594 			if (ewq->msg.event == UFFD_EVENT_FORK) {
595 				struct userfaultfd_ctx *new;
596 
597 				new = (struct userfaultfd_ctx *)
598 					(unsigned long)
599 					ewq->msg.arg.reserved.reserved1;
600 				release_new_ctx = new;
601 			}
602 			break;
603 		}
604 
605 		spin_unlock_irq(&ctx->event_wqh.lock);
606 
607 		wake_up_poll(&ctx->fd_wqh, EPOLLIN);
608 		schedule();
609 
610 		spin_lock_irq(&ctx->event_wqh.lock);
611 	}
612 	__set_current_state(TASK_RUNNING);
613 	spin_unlock_irq(&ctx->event_wqh.lock);
614 
615 	if (release_new_ctx) {
616 		struct vm_area_struct *vma;
617 		struct mm_struct *mm = release_new_ctx->mm;
618 
619 		/* the various vma->vm_userfaultfd_ctx still points to it */
620 		mmap_write_lock(mm);
621 		for (vma = mm->mmap; vma; vma = vma->vm_next)
622 			if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
623 				vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
624 				vma->vm_flags &= ~__VM_UFFD_FLAGS;
625 			}
626 		mmap_write_unlock(mm);
627 
628 		userfaultfd_ctx_put(release_new_ctx);
629 	}
630 
631 	/*
632 	 * ctx may go away after this if the userfault pseudo fd is
633 	 * already released.
634 	 */
635 out:
636 	atomic_dec(&ctx->mmap_changing);
637 	VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
638 	userfaultfd_ctx_put(ctx);
639 }
640 
641 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
642 				       struct userfaultfd_wait_queue *ewq)
643 {
644 	ewq->msg.event = 0;
645 	wake_up_locked(&ctx->event_wqh);
646 	__remove_wait_queue(&ctx->event_wqh, &ewq->wq);
647 }
648 
649 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
650 {
651 	struct userfaultfd_ctx *ctx = NULL, *octx;
652 	struct userfaultfd_fork_ctx *fctx;
653 
654 	octx = vma->vm_userfaultfd_ctx.ctx;
655 	if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
656 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
657 		vma->vm_flags &= ~__VM_UFFD_FLAGS;
658 		return 0;
659 	}
660 
661 	list_for_each_entry(fctx, fcs, list)
662 		if (fctx->orig == octx) {
663 			ctx = fctx->new;
664 			break;
665 		}
666 
667 	if (!ctx) {
668 		fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
669 		if (!fctx)
670 			return -ENOMEM;
671 
672 		ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
673 		if (!ctx) {
674 			kfree(fctx);
675 			return -ENOMEM;
676 		}
677 
678 		refcount_set(&ctx->refcount, 1);
679 		ctx->flags = octx->flags;
680 		ctx->features = octx->features;
681 		ctx->released = false;
682 		atomic_set(&ctx->mmap_changing, 0);
683 		ctx->mm = vma->vm_mm;
684 		mmgrab(ctx->mm);
685 
686 		userfaultfd_ctx_get(octx);
687 		atomic_inc(&octx->mmap_changing);
688 		fctx->orig = octx;
689 		fctx->new = ctx;
690 		list_add_tail(&fctx->list, fcs);
691 	}
692 
693 	vma->vm_userfaultfd_ctx.ctx = ctx;
694 	return 0;
695 }
696 
697 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
698 {
699 	struct userfaultfd_ctx *ctx = fctx->orig;
700 	struct userfaultfd_wait_queue ewq;
701 
702 	msg_init(&ewq.msg);
703 
704 	ewq.msg.event = UFFD_EVENT_FORK;
705 	ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
706 
707 	userfaultfd_event_wait_completion(ctx, &ewq);
708 }
709 
710 void dup_userfaultfd_complete(struct list_head *fcs)
711 {
712 	struct userfaultfd_fork_ctx *fctx, *n;
713 
714 	list_for_each_entry_safe(fctx, n, fcs, list) {
715 		dup_fctx(fctx);
716 		list_del(&fctx->list);
717 		kfree(fctx);
718 	}
719 }
720 
721 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
722 			     struct vm_userfaultfd_ctx *vm_ctx)
723 {
724 	struct userfaultfd_ctx *ctx;
725 
726 	ctx = vma->vm_userfaultfd_ctx.ctx;
727 
728 	if (!ctx)
729 		return;
730 
731 	if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
732 		vm_ctx->ctx = ctx;
733 		userfaultfd_ctx_get(ctx);
734 		atomic_inc(&ctx->mmap_changing);
735 	} else {
736 		/* Drop uffd context if remap feature not enabled */
737 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
738 		vma->vm_flags &= ~__VM_UFFD_FLAGS;
739 	}
740 }
741 
742 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
743 				 unsigned long from, unsigned long to,
744 				 unsigned long len)
745 {
746 	struct userfaultfd_ctx *ctx = vm_ctx->ctx;
747 	struct userfaultfd_wait_queue ewq;
748 
749 	if (!ctx)
750 		return;
751 
752 	if (to & ~PAGE_MASK) {
753 		userfaultfd_ctx_put(ctx);
754 		return;
755 	}
756 
757 	msg_init(&ewq.msg);
758 
759 	ewq.msg.event = UFFD_EVENT_REMAP;
760 	ewq.msg.arg.remap.from = from;
761 	ewq.msg.arg.remap.to = to;
762 	ewq.msg.arg.remap.len = len;
763 
764 	userfaultfd_event_wait_completion(ctx, &ewq);
765 }
766 
767 bool userfaultfd_remove(struct vm_area_struct *vma,
768 			unsigned long start, unsigned long end)
769 {
770 	struct mm_struct *mm = vma->vm_mm;
771 	struct userfaultfd_ctx *ctx;
772 	struct userfaultfd_wait_queue ewq;
773 
774 	ctx = vma->vm_userfaultfd_ctx.ctx;
775 	if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
776 		return true;
777 
778 	userfaultfd_ctx_get(ctx);
779 	atomic_inc(&ctx->mmap_changing);
780 	mmap_read_unlock(mm);
781 
782 	msg_init(&ewq.msg);
783 
784 	ewq.msg.event = UFFD_EVENT_REMOVE;
785 	ewq.msg.arg.remove.start = start;
786 	ewq.msg.arg.remove.end = end;
787 
788 	userfaultfd_event_wait_completion(ctx, &ewq);
789 
790 	return false;
791 }
792 
793 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
794 			  unsigned long start, unsigned long end)
795 {
796 	struct userfaultfd_unmap_ctx *unmap_ctx;
797 
798 	list_for_each_entry(unmap_ctx, unmaps, list)
799 		if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
800 		    unmap_ctx->end == end)
801 			return true;
802 
803 	return false;
804 }
805 
806 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
807 			   unsigned long start, unsigned long end,
808 			   struct list_head *unmaps)
809 {
810 	for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
811 		struct userfaultfd_unmap_ctx *unmap_ctx;
812 		struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
813 
814 		if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
815 		    has_unmap_ctx(ctx, unmaps, start, end))
816 			continue;
817 
818 		unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
819 		if (!unmap_ctx)
820 			return -ENOMEM;
821 
822 		userfaultfd_ctx_get(ctx);
823 		atomic_inc(&ctx->mmap_changing);
824 		unmap_ctx->ctx = ctx;
825 		unmap_ctx->start = start;
826 		unmap_ctx->end = end;
827 		list_add_tail(&unmap_ctx->list, unmaps);
828 	}
829 
830 	return 0;
831 }
832 
833 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
834 {
835 	struct userfaultfd_unmap_ctx *ctx, *n;
836 	struct userfaultfd_wait_queue ewq;
837 
838 	list_for_each_entry_safe(ctx, n, uf, list) {
839 		msg_init(&ewq.msg);
840 
841 		ewq.msg.event = UFFD_EVENT_UNMAP;
842 		ewq.msg.arg.remove.start = ctx->start;
843 		ewq.msg.arg.remove.end = ctx->end;
844 
845 		userfaultfd_event_wait_completion(ctx->ctx, &ewq);
846 
847 		list_del(&ctx->list);
848 		kfree(ctx);
849 	}
850 }
851 
852 static int userfaultfd_release(struct inode *inode, struct file *file)
853 {
854 	struct userfaultfd_ctx *ctx = file->private_data;
855 	struct mm_struct *mm = ctx->mm;
856 	struct vm_area_struct *vma, *prev;
857 	/* len == 0 means wake all */
858 	struct userfaultfd_wake_range range = { .len = 0, };
859 	unsigned long new_flags;
860 
861 	WRITE_ONCE(ctx->released, true);
862 
863 	if (!mmget_not_zero(mm))
864 		goto wakeup;
865 
866 	/*
867 	 * Flush page faults out of all CPUs. NOTE: all page faults
868 	 * must be retried without returning VM_FAULT_SIGBUS if
869 	 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
870 	 * changes while handle_userfault released the mmap_lock. So
871 	 * it's critical that released is set to true (above), before
872 	 * taking the mmap_lock for writing.
873 	 */
874 	mmap_write_lock(mm);
875 	prev = NULL;
876 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
877 		cond_resched();
878 		BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
879 		       !!(vma->vm_flags & __VM_UFFD_FLAGS));
880 		if (vma->vm_userfaultfd_ctx.ctx != ctx) {
881 			prev = vma;
882 			continue;
883 		}
884 		new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
885 		prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
886 				 new_flags, vma->anon_vma,
887 				 vma->vm_file, vma->vm_pgoff,
888 				 vma_policy(vma),
889 				 NULL_VM_UFFD_CTX, anon_vma_name(vma));
890 		if (prev)
891 			vma = prev;
892 		else
893 			prev = vma;
894 		vma->vm_flags = new_flags;
895 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
896 	}
897 	mmap_write_unlock(mm);
898 	mmput(mm);
899 wakeup:
900 	/*
901 	 * After no new page faults can wait on this fault_*wqh, flush
902 	 * the last page faults that may have been already waiting on
903 	 * the fault_*wqh.
904 	 */
905 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
906 	__wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
907 	__wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
908 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
909 
910 	/* Flush pending events that may still wait on event_wqh */
911 	wake_up_all(&ctx->event_wqh);
912 
913 	wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
914 	userfaultfd_ctx_put(ctx);
915 	return 0;
916 }
917 
918 /* fault_pending_wqh.lock must be hold by the caller */
919 static inline struct userfaultfd_wait_queue *find_userfault_in(
920 		wait_queue_head_t *wqh)
921 {
922 	wait_queue_entry_t *wq;
923 	struct userfaultfd_wait_queue *uwq;
924 
925 	lockdep_assert_held(&wqh->lock);
926 
927 	uwq = NULL;
928 	if (!waitqueue_active(wqh))
929 		goto out;
930 	/* walk in reverse to provide FIFO behavior to read userfaults */
931 	wq = list_last_entry(&wqh->head, typeof(*wq), entry);
932 	uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
933 out:
934 	return uwq;
935 }
936 
937 static inline struct userfaultfd_wait_queue *find_userfault(
938 		struct userfaultfd_ctx *ctx)
939 {
940 	return find_userfault_in(&ctx->fault_pending_wqh);
941 }
942 
943 static inline struct userfaultfd_wait_queue *find_userfault_evt(
944 		struct userfaultfd_ctx *ctx)
945 {
946 	return find_userfault_in(&ctx->event_wqh);
947 }
948 
949 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
950 {
951 	struct userfaultfd_ctx *ctx = file->private_data;
952 	__poll_t ret;
953 
954 	poll_wait(file, &ctx->fd_wqh, wait);
955 
956 	if (!userfaultfd_is_initialized(ctx))
957 		return EPOLLERR;
958 
959 	/*
960 	 * poll() never guarantees that read won't block.
961 	 * userfaults can be waken before they're read().
962 	 */
963 	if (unlikely(!(file->f_flags & O_NONBLOCK)))
964 		return EPOLLERR;
965 	/*
966 	 * lockless access to see if there are pending faults
967 	 * __pollwait last action is the add_wait_queue but
968 	 * the spin_unlock would allow the waitqueue_active to
969 	 * pass above the actual list_add inside
970 	 * add_wait_queue critical section. So use a full
971 	 * memory barrier to serialize the list_add write of
972 	 * add_wait_queue() with the waitqueue_active read
973 	 * below.
974 	 */
975 	ret = 0;
976 	smp_mb();
977 	if (waitqueue_active(&ctx->fault_pending_wqh))
978 		ret = EPOLLIN;
979 	else if (waitqueue_active(&ctx->event_wqh))
980 		ret = EPOLLIN;
981 
982 	return ret;
983 }
984 
985 static const struct file_operations userfaultfd_fops;
986 
987 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
988 				  struct inode *inode,
989 				  struct uffd_msg *msg)
990 {
991 	int fd;
992 
993 	fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
994 			O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
995 	if (fd < 0)
996 		return fd;
997 
998 	msg->arg.reserved.reserved1 = 0;
999 	msg->arg.fork.ufd = fd;
1000 	return 0;
1001 }
1002 
1003 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1004 				    struct uffd_msg *msg, struct inode *inode)
1005 {
1006 	ssize_t ret;
1007 	DECLARE_WAITQUEUE(wait, current);
1008 	struct userfaultfd_wait_queue *uwq;
1009 	/*
1010 	 * Handling fork event requires sleeping operations, so
1011 	 * we drop the event_wqh lock, then do these ops, then
1012 	 * lock it back and wake up the waiter. While the lock is
1013 	 * dropped the ewq may go away so we keep track of it
1014 	 * carefully.
1015 	 */
1016 	LIST_HEAD(fork_event);
1017 	struct userfaultfd_ctx *fork_nctx = NULL;
1018 
1019 	/* always take the fd_wqh lock before the fault_pending_wqh lock */
1020 	spin_lock_irq(&ctx->fd_wqh.lock);
1021 	__add_wait_queue(&ctx->fd_wqh, &wait);
1022 	for (;;) {
1023 		set_current_state(TASK_INTERRUPTIBLE);
1024 		spin_lock(&ctx->fault_pending_wqh.lock);
1025 		uwq = find_userfault(ctx);
1026 		if (uwq) {
1027 			/*
1028 			 * Use a seqcount to repeat the lockless check
1029 			 * in wake_userfault() to avoid missing
1030 			 * wakeups because during the refile both
1031 			 * waitqueue could become empty if this is the
1032 			 * only userfault.
1033 			 */
1034 			write_seqcount_begin(&ctx->refile_seq);
1035 
1036 			/*
1037 			 * The fault_pending_wqh.lock prevents the uwq
1038 			 * to disappear from under us.
1039 			 *
1040 			 * Refile this userfault from
1041 			 * fault_pending_wqh to fault_wqh, it's not
1042 			 * pending anymore after we read it.
1043 			 *
1044 			 * Use list_del() by hand (as
1045 			 * userfaultfd_wake_function also uses
1046 			 * list_del_init() by hand) to be sure nobody
1047 			 * changes __remove_wait_queue() to use
1048 			 * list_del_init() in turn breaking the
1049 			 * !list_empty_careful() check in
1050 			 * handle_userfault(). The uwq->wq.head list
1051 			 * must never be empty at any time during the
1052 			 * refile, or the waitqueue could disappear
1053 			 * from under us. The "wait_queue_head_t"
1054 			 * parameter of __remove_wait_queue() is unused
1055 			 * anyway.
1056 			 */
1057 			list_del(&uwq->wq.entry);
1058 			add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1059 
1060 			write_seqcount_end(&ctx->refile_seq);
1061 
1062 			/* careful to always initialize msg if ret == 0 */
1063 			*msg = uwq->msg;
1064 			spin_unlock(&ctx->fault_pending_wqh.lock);
1065 			ret = 0;
1066 			break;
1067 		}
1068 		spin_unlock(&ctx->fault_pending_wqh.lock);
1069 
1070 		spin_lock(&ctx->event_wqh.lock);
1071 		uwq = find_userfault_evt(ctx);
1072 		if (uwq) {
1073 			*msg = uwq->msg;
1074 
1075 			if (uwq->msg.event == UFFD_EVENT_FORK) {
1076 				fork_nctx = (struct userfaultfd_ctx *)
1077 					(unsigned long)
1078 					uwq->msg.arg.reserved.reserved1;
1079 				list_move(&uwq->wq.entry, &fork_event);
1080 				/*
1081 				 * fork_nctx can be freed as soon as
1082 				 * we drop the lock, unless we take a
1083 				 * reference on it.
1084 				 */
1085 				userfaultfd_ctx_get(fork_nctx);
1086 				spin_unlock(&ctx->event_wqh.lock);
1087 				ret = 0;
1088 				break;
1089 			}
1090 
1091 			userfaultfd_event_complete(ctx, uwq);
1092 			spin_unlock(&ctx->event_wqh.lock);
1093 			ret = 0;
1094 			break;
1095 		}
1096 		spin_unlock(&ctx->event_wqh.lock);
1097 
1098 		if (signal_pending(current)) {
1099 			ret = -ERESTARTSYS;
1100 			break;
1101 		}
1102 		if (no_wait) {
1103 			ret = -EAGAIN;
1104 			break;
1105 		}
1106 		spin_unlock_irq(&ctx->fd_wqh.lock);
1107 		schedule();
1108 		spin_lock_irq(&ctx->fd_wqh.lock);
1109 	}
1110 	__remove_wait_queue(&ctx->fd_wqh, &wait);
1111 	__set_current_state(TASK_RUNNING);
1112 	spin_unlock_irq(&ctx->fd_wqh.lock);
1113 
1114 	if (!ret && msg->event == UFFD_EVENT_FORK) {
1115 		ret = resolve_userfault_fork(fork_nctx, inode, msg);
1116 		spin_lock_irq(&ctx->event_wqh.lock);
1117 		if (!list_empty(&fork_event)) {
1118 			/*
1119 			 * The fork thread didn't abort, so we can
1120 			 * drop the temporary refcount.
1121 			 */
1122 			userfaultfd_ctx_put(fork_nctx);
1123 
1124 			uwq = list_first_entry(&fork_event,
1125 					       typeof(*uwq),
1126 					       wq.entry);
1127 			/*
1128 			 * If fork_event list wasn't empty and in turn
1129 			 * the event wasn't already released by fork
1130 			 * (the event is allocated on fork kernel
1131 			 * stack), put the event back to its place in
1132 			 * the event_wq. fork_event head will be freed
1133 			 * as soon as we return so the event cannot
1134 			 * stay queued there no matter the current
1135 			 * "ret" value.
1136 			 */
1137 			list_del(&uwq->wq.entry);
1138 			__add_wait_queue(&ctx->event_wqh, &uwq->wq);
1139 
1140 			/*
1141 			 * Leave the event in the waitqueue and report
1142 			 * error to userland if we failed to resolve
1143 			 * the userfault fork.
1144 			 */
1145 			if (likely(!ret))
1146 				userfaultfd_event_complete(ctx, uwq);
1147 		} else {
1148 			/*
1149 			 * Here the fork thread aborted and the
1150 			 * refcount from the fork thread on fork_nctx
1151 			 * has already been released. We still hold
1152 			 * the reference we took before releasing the
1153 			 * lock above. If resolve_userfault_fork
1154 			 * failed we've to drop it because the
1155 			 * fork_nctx has to be freed in such case. If
1156 			 * it succeeded we'll hold it because the new
1157 			 * uffd references it.
1158 			 */
1159 			if (ret)
1160 				userfaultfd_ctx_put(fork_nctx);
1161 		}
1162 		spin_unlock_irq(&ctx->event_wqh.lock);
1163 	}
1164 
1165 	return ret;
1166 }
1167 
1168 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1169 				size_t count, loff_t *ppos)
1170 {
1171 	struct userfaultfd_ctx *ctx = file->private_data;
1172 	ssize_t _ret, ret = 0;
1173 	struct uffd_msg msg;
1174 	int no_wait = file->f_flags & O_NONBLOCK;
1175 	struct inode *inode = file_inode(file);
1176 
1177 	if (!userfaultfd_is_initialized(ctx))
1178 		return -EINVAL;
1179 
1180 	for (;;) {
1181 		if (count < sizeof(msg))
1182 			return ret ? ret : -EINVAL;
1183 		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1184 		if (_ret < 0)
1185 			return ret ? ret : _ret;
1186 		if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1187 			return ret ? ret : -EFAULT;
1188 		ret += sizeof(msg);
1189 		buf += sizeof(msg);
1190 		count -= sizeof(msg);
1191 		/*
1192 		 * Allow to read more than one fault at time but only
1193 		 * block if waiting for the very first one.
1194 		 */
1195 		no_wait = O_NONBLOCK;
1196 	}
1197 }
1198 
1199 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1200 			     struct userfaultfd_wake_range *range)
1201 {
1202 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
1203 	/* wake all in the range and autoremove */
1204 	if (waitqueue_active(&ctx->fault_pending_wqh))
1205 		__wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1206 				     range);
1207 	if (waitqueue_active(&ctx->fault_wqh))
1208 		__wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1209 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1210 }
1211 
1212 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1213 					   struct userfaultfd_wake_range *range)
1214 {
1215 	unsigned seq;
1216 	bool need_wakeup;
1217 
1218 	/*
1219 	 * To be sure waitqueue_active() is not reordered by the CPU
1220 	 * before the pagetable update, use an explicit SMP memory
1221 	 * barrier here. PT lock release or mmap_read_unlock(mm) still
1222 	 * have release semantics that can allow the
1223 	 * waitqueue_active() to be reordered before the pte update.
1224 	 */
1225 	smp_mb();
1226 
1227 	/*
1228 	 * Use waitqueue_active because it's very frequent to
1229 	 * change the address space atomically even if there are no
1230 	 * userfaults yet. So we take the spinlock only when we're
1231 	 * sure we've userfaults to wake.
1232 	 */
1233 	do {
1234 		seq = read_seqcount_begin(&ctx->refile_seq);
1235 		need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1236 			waitqueue_active(&ctx->fault_wqh);
1237 		cond_resched();
1238 	} while (read_seqcount_retry(&ctx->refile_seq, seq));
1239 	if (need_wakeup)
1240 		__wake_userfault(ctx, range);
1241 }
1242 
1243 static __always_inline int validate_range(struct mm_struct *mm,
1244 					  __u64 start, __u64 len)
1245 {
1246 	__u64 task_size = mm->task_size;
1247 
1248 	if (start & ~PAGE_MASK)
1249 		return -EINVAL;
1250 	if (len & ~PAGE_MASK)
1251 		return -EINVAL;
1252 	if (!len)
1253 		return -EINVAL;
1254 	if (start < mmap_min_addr)
1255 		return -EINVAL;
1256 	if (start >= task_size)
1257 		return -EINVAL;
1258 	if (len > task_size - start)
1259 		return -EINVAL;
1260 	return 0;
1261 }
1262 
1263 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1264 				unsigned long arg)
1265 {
1266 	struct mm_struct *mm = ctx->mm;
1267 	struct vm_area_struct *vma, *prev, *cur;
1268 	int ret;
1269 	struct uffdio_register uffdio_register;
1270 	struct uffdio_register __user *user_uffdio_register;
1271 	unsigned long vm_flags, new_flags;
1272 	bool found;
1273 	bool basic_ioctls;
1274 	unsigned long start, end, vma_end;
1275 
1276 	user_uffdio_register = (struct uffdio_register __user *) arg;
1277 
1278 	ret = -EFAULT;
1279 	if (copy_from_user(&uffdio_register, user_uffdio_register,
1280 			   sizeof(uffdio_register)-sizeof(__u64)))
1281 		goto out;
1282 
1283 	ret = -EINVAL;
1284 	if (!uffdio_register.mode)
1285 		goto out;
1286 	if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1287 		goto out;
1288 	vm_flags = 0;
1289 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1290 		vm_flags |= VM_UFFD_MISSING;
1291 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1292 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1293 		goto out;
1294 #endif
1295 		vm_flags |= VM_UFFD_WP;
1296 	}
1297 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1298 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1299 		goto out;
1300 #endif
1301 		vm_flags |= VM_UFFD_MINOR;
1302 	}
1303 
1304 	ret = validate_range(mm, uffdio_register.range.start,
1305 			     uffdio_register.range.len);
1306 	if (ret)
1307 		goto out;
1308 
1309 	start = uffdio_register.range.start;
1310 	end = start + uffdio_register.range.len;
1311 
1312 	ret = -ENOMEM;
1313 	if (!mmget_not_zero(mm))
1314 		goto out;
1315 
1316 	mmap_write_lock(mm);
1317 	vma = find_vma_prev(mm, start, &prev);
1318 	if (!vma)
1319 		goto out_unlock;
1320 
1321 	/* check that there's at least one vma in the range */
1322 	ret = -EINVAL;
1323 	if (vma->vm_start >= end)
1324 		goto out_unlock;
1325 
1326 	/*
1327 	 * If the first vma contains huge pages, make sure start address
1328 	 * is aligned to huge page size.
1329 	 */
1330 	if (is_vm_hugetlb_page(vma)) {
1331 		unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1332 
1333 		if (start & (vma_hpagesize - 1))
1334 			goto out_unlock;
1335 	}
1336 
1337 	/*
1338 	 * Search for not compatible vmas.
1339 	 */
1340 	found = false;
1341 	basic_ioctls = false;
1342 	for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1343 		cond_resched();
1344 
1345 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1346 		       !!(cur->vm_flags & __VM_UFFD_FLAGS));
1347 
1348 		/* check not compatible vmas */
1349 		ret = -EINVAL;
1350 		if (!vma_can_userfault(cur, vm_flags))
1351 			goto out_unlock;
1352 
1353 		/*
1354 		 * UFFDIO_COPY will fill file holes even without
1355 		 * PROT_WRITE. This check enforces that if this is a
1356 		 * MAP_SHARED, the process has write permission to the backing
1357 		 * file. If VM_MAYWRITE is set it also enforces that on a
1358 		 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1359 		 * F_WRITE_SEAL can be taken until the vma is destroyed.
1360 		 */
1361 		ret = -EPERM;
1362 		if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1363 			goto out_unlock;
1364 
1365 		/*
1366 		 * If this vma contains ending address, and huge pages
1367 		 * check alignment.
1368 		 */
1369 		if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1370 		    end > cur->vm_start) {
1371 			unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1372 
1373 			ret = -EINVAL;
1374 
1375 			if (end & (vma_hpagesize - 1))
1376 				goto out_unlock;
1377 		}
1378 		if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1379 			goto out_unlock;
1380 
1381 		/*
1382 		 * Check that this vma isn't already owned by a
1383 		 * different userfaultfd. We can't allow more than one
1384 		 * userfaultfd to own a single vma simultaneously or we
1385 		 * wouldn't know which one to deliver the userfaults to.
1386 		 */
1387 		ret = -EBUSY;
1388 		if (cur->vm_userfaultfd_ctx.ctx &&
1389 		    cur->vm_userfaultfd_ctx.ctx != ctx)
1390 			goto out_unlock;
1391 
1392 		/*
1393 		 * Note vmas containing huge pages
1394 		 */
1395 		if (is_vm_hugetlb_page(cur))
1396 			basic_ioctls = true;
1397 
1398 		found = true;
1399 	}
1400 	BUG_ON(!found);
1401 
1402 	if (vma->vm_start < start)
1403 		prev = vma;
1404 
1405 	ret = 0;
1406 	do {
1407 		cond_resched();
1408 
1409 		BUG_ON(!vma_can_userfault(vma, vm_flags));
1410 		BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1411 		       vma->vm_userfaultfd_ctx.ctx != ctx);
1412 		WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1413 
1414 		/*
1415 		 * Nothing to do: this vma is already registered into this
1416 		 * userfaultfd and with the right tracking mode too.
1417 		 */
1418 		if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1419 		    (vma->vm_flags & vm_flags) == vm_flags)
1420 			goto skip;
1421 
1422 		if (vma->vm_start > start)
1423 			start = vma->vm_start;
1424 		vma_end = min(end, vma->vm_end);
1425 
1426 		new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1427 		prev = vma_merge(mm, prev, start, vma_end, new_flags,
1428 				 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1429 				 vma_policy(vma),
1430 				 ((struct vm_userfaultfd_ctx){ ctx }),
1431 				 anon_vma_name(vma));
1432 		if (prev) {
1433 			vma = prev;
1434 			goto next;
1435 		}
1436 		if (vma->vm_start < start) {
1437 			ret = split_vma(mm, vma, start, 1);
1438 			if (ret)
1439 				break;
1440 		}
1441 		if (vma->vm_end > end) {
1442 			ret = split_vma(mm, vma, end, 0);
1443 			if (ret)
1444 				break;
1445 		}
1446 	next:
1447 		/*
1448 		 * In the vma_merge() successful mprotect-like case 8:
1449 		 * the next vma was merged into the current one and
1450 		 * the current one has not been updated yet.
1451 		 */
1452 		vma->vm_flags = new_flags;
1453 		vma->vm_userfaultfd_ctx.ctx = ctx;
1454 
1455 		if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1456 			hugetlb_unshare_all_pmds(vma);
1457 
1458 	skip:
1459 		prev = vma;
1460 		start = vma->vm_end;
1461 		vma = vma->vm_next;
1462 	} while (vma && vma->vm_start < end);
1463 out_unlock:
1464 	mmap_write_unlock(mm);
1465 	mmput(mm);
1466 	if (!ret) {
1467 		__u64 ioctls_out;
1468 
1469 		ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1470 		    UFFD_API_RANGE_IOCTLS;
1471 
1472 		/*
1473 		 * Declare the WP ioctl only if the WP mode is
1474 		 * specified and all checks passed with the range
1475 		 */
1476 		if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1477 			ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1478 
1479 		/* CONTINUE ioctl is only supported for MINOR ranges. */
1480 		if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1481 			ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1482 
1483 		/*
1484 		 * Now that we scanned all vmas we can already tell
1485 		 * userland which ioctls methods are guaranteed to
1486 		 * succeed on this range.
1487 		 */
1488 		if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1489 			ret = -EFAULT;
1490 	}
1491 out:
1492 	return ret;
1493 }
1494 
1495 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1496 				  unsigned long arg)
1497 {
1498 	struct mm_struct *mm = ctx->mm;
1499 	struct vm_area_struct *vma, *prev, *cur;
1500 	int ret;
1501 	struct uffdio_range uffdio_unregister;
1502 	unsigned long new_flags;
1503 	bool found;
1504 	unsigned long start, end, vma_end;
1505 	const void __user *buf = (void __user *)arg;
1506 
1507 	ret = -EFAULT;
1508 	if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1509 		goto out;
1510 
1511 	ret = validate_range(mm, uffdio_unregister.start,
1512 			     uffdio_unregister.len);
1513 	if (ret)
1514 		goto out;
1515 
1516 	start = uffdio_unregister.start;
1517 	end = start + uffdio_unregister.len;
1518 
1519 	ret = -ENOMEM;
1520 	if (!mmget_not_zero(mm))
1521 		goto out;
1522 
1523 	mmap_write_lock(mm);
1524 	vma = find_vma_prev(mm, start, &prev);
1525 	if (!vma)
1526 		goto out_unlock;
1527 
1528 	/* check that there's at least one vma in the range */
1529 	ret = -EINVAL;
1530 	if (vma->vm_start >= end)
1531 		goto out_unlock;
1532 
1533 	/*
1534 	 * If the first vma contains huge pages, make sure start address
1535 	 * is aligned to huge page size.
1536 	 */
1537 	if (is_vm_hugetlb_page(vma)) {
1538 		unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1539 
1540 		if (start & (vma_hpagesize - 1))
1541 			goto out_unlock;
1542 	}
1543 
1544 	/*
1545 	 * Search for not compatible vmas.
1546 	 */
1547 	found = false;
1548 	ret = -EINVAL;
1549 	for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1550 		cond_resched();
1551 
1552 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1553 		       !!(cur->vm_flags & __VM_UFFD_FLAGS));
1554 
1555 		/*
1556 		 * Check not compatible vmas, not strictly required
1557 		 * here as not compatible vmas cannot have an
1558 		 * userfaultfd_ctx registered on them, but this
1559 		 * provides for more strict behavior to notice
1560 		 * unregistration errors.
1561 		 */
1562 		if (!vma_can_userfault(cur, cur->vm_flags))
1563 			goto out_unlock;
1564 
1565 		found = true;
1566 	}
1567 	BUG_ON(!found);
1568 
1569 	if (vma->vm_start < start)
1570 		prev = vma;
1571 
1572 	ret = 0;
1573 	do {
1574 		cond_resched();
1575 
1576 		BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1577 
1578 		/*
1579 		 * Nothing to do: this vma is already registered into this
1580 		 * userfaultfd and with the right tracking mode too.
1581 		 */
1582 		if (!vma->vm_userfaultfd_ctx.ctx)
1583 			goto skip;
1584 
1585 		WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1586 
1587 		if (vma->vm_start > start)
1588 			start = vma->vm_start;
1589 		vma_end = min(end, vma->vm_end);
1590 
1591 		if (userfaultfd_missing(vma)) {
1592 			/*
1593 			 * Wake any concurrent pending userfault while
1594 			 * we unregister, so they will not hang
1595 			 * permanently and it avoids userland to call
1596 			 * UFFDIO_WAKE explicitly.
1597 			 */
1598 			struct userfaultfd_wake_range range;
1599 			range.start = start;
1600 			range.len = vma_end - start;
1601 			wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1602 		}
1603 
1604 		new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1605 		prev = vma_merge(mm, prev, start, vma_end, new_flags,
1606 				 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1607 				 vma_policy(vma),
1608 				 NULL_VM_UFFD_CTX, anon_vma_name(vma));
1609 		if (prev) {
1610 			vma = prev;
1611 			goto next;
1612 		}
1613 		if (vma->vm_start < start) {
1614 			ret = split_vma(mm, vma, start, 1);
1615 			if (ret)
1616 				break;
1617 		}
1618 		if (vma->vm_end > end) {
1619 			ret = split_vma(mm, vma, end, 0);
1620 			if (ret)
1621 				break;
1622 		}
1623 	next:
1624 		/*
1625 		 * In the vma_merge() successful mprotect-like case 8:
1626 		 * the next vma was merged into the current one and
1627 		 * the current one has not been updated yet.
1628 		 */
1629 		vma->vm_flags = new_flags;
1630 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1631 
1632 	skip:
1633 		prev = vma;
1634 		start = vma->vm_end;
1635 		vma = vma->vm_next;
1636 	} while (vma && vma->vm_start < end);
1637 out_unlock:
1638 	mmap_write_unlock(mm);
1639 	mmput(mm);
1640 out:
1641 	return ret;
1642 }
1643 
1644 /*
1645  * userfaultfd_wake may be used in combination with the
1646  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1647  */
1648 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1649 			    unsigned long arg)
1650 {
1651 	int ret;
1652 	struct uffdio_range uffdio_wake;
1653 	struct userfaultfd_wake_range range;
1654 	const void __user *buf = (void __user *)arg;
1655 
1656 	ret = -EFAULT;
1657 	if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1658 		goto out;
1659 
1660 	ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1661 	if (ret)
1662 		goto out;
1663 
1664 	range.start = uffdio_wake.start;
1665 	range.len = uffdio_wake.len;
1666 
1667 	/*
1668 	 * len == 0 means wake all and we don't want to wake all here,
1669 	 * so check it again to be sure.
1670 	 */
1671 	VM_BUG_ON(!range.len);
1672 
1673 	wake_userfault(ctx, &range);
1674 	ret = 0;
1675 
1676 out:
1677 	return ret;
1678 }
1679 
1680 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1681 			    unsigned long arg)
1682 {
1683 	__s64 ret;
1684 	struct uffdio_copy uffdio_copy;
1685 	struct uffdio_copy __user *user_uffdio_copy;
1686 	struct userfaultfd_wake_range range;
1687 
1688 	user_uffdio_copy = (struct uffdio_copy __user *) arg;
1689 
1690 	ret = -EAGAIN;
1691 	if (atomic_read(&ctx->mmap_changing))
1692 		goto out;
1693 
1694 	ret = -EFAULT;
1695 	if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1696 			   /* don't copy "copy" last field */
1697 			   sizeof(uffdio_copy)-sizeof(__s64)))
1698 		goto out;
1699 
1700 	ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1701 	if (ret)
1702 		goto out;
1703 	/*
1704 	 * double check for wraparound just in case. copy_from_user()
1705 	 * will later check uffdio_copy.src + uffdio_copy.len to fit
1706 	 * in the userland range.
1707 	 */
1708 	ret = -EINVAL;
1709 	if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1710 		goto out;
1711 	if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1712 		goto out;
1713 	if (mmget_not_zero(ctx->mm)) {
1714 		ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1715 				   uffdio_copy.len, &ctx->mmap_changing,
1716 				   uffdio_copy.mode);
1717 		mmput(ctx->mm);
1718 	} else {
1719 		return -ESRCH;
1720 	}
1721 	if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1722 		return -EFAULT;
1723 	if (ret < 0)
1724 		goto out;
1725 	BUG_ON(!ret);
1726 	/* len == 0 would wake all */
1727 	range.len = ret;
1728 	if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1729 		range.start = uffdio_copy.dst;
1730 		wake_userfault(ctx, &range);
1731 	}
1732 	ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1733 out:
1734 	return ret;
1735 }
1736 
1737 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1738 				unsigned long arg)
1739 {
1740 	__s64 ret;
1741 	struct uffdio_zeropage uffdio_zeropage;
1742 	struct uffdio_zeropage __user *user_uffdio_zeropage;
1743 	struct userfaultfd_wake_range range;
1744 
1745 	user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1746 
1747 	ret = -EAGAIN;
1748 	if (atomic_read(&ctx->mmap_changing))
1749 		goto out;
1750 
1751 	ret = -EFAULT;
1752 	if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1753 			   /* don't copy "zeropage" last field */
1754 			   sizeof(uffdio_zeropage)-sizeof(__s64)))
1755 		goto out;
1756 
1757 	ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1758 			     uffdio_zeropage.range.len);
1759 	if (ret)
1760 		goto out;
1761 	ret = -EINVAL;
1762 	if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1763 		goto out;
1764 
1765 	if (mmget_not_zero(ctx->mm)) {
1766 		ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1767 				     uffdio_zeropage.range.len,
1768 				     &ctx->mmap_changing);
1769 		mmput(ctx->mm);
1770 	} else {
1771 		return -ESRCH;
1772 	}
1773 	if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1774 		return -EFAULT;
1775 	if (ret < 0)
1776 		goto out;
1777 	/* len == 0 would wake all */
1778 	BUG_ON(!ret);
1779 	range.len = ret;
1780 	if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1781 		range.start = uffdio_zeropage.range.start;
1782 		wake_userfault(ctx, &range);
1783 	}
1784 	ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1785 out:
1786 	return ret;
1787 }
1788 
1789 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1790 				    unsigned long arg)
1791 {
1792 	int ret;
1793 	struct uffdio_writeprotect uffdio_wp;
1794 	struct uffdio_writeprotect __user *user_uffdio_wp;
1795 	struct userfaultfd_wake_range range;
1796 	bool mode_wp, mode_dontwake;
1797 
1798 	if (atomic_read(&ctx->mmap_changing))
1799 		return -EAGAIN;
1800 
1801 	user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1802 
1803 	if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1804 			   sizeof(struct uffdio_writeprotect)))
1805 		return -EFAULT;
1806 
1807 	ret = validate_range(ctx->mm, uffdio_wp.range.start,
1808 			     uffdio_wp.range.len);
1809 	if (ret)
1810 		return ret;
1811 
1812 	if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1813 			       UFFDIO_WRITEPROTECT_MODE_WP))
1814 		return -EINVAL;
1815 
1816 	mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1817 	mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1818 
1819 	if (mode_wp && mode_dontwake)
1820 		return -EINVAL;
1821 
1822 	if (mmget_not_zero(ctx->mm)) {
1823 		ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1824 					  uffdio_wp.range.len, mode_wp,
1825 					  &ctx->mmap_changing);
1826 		mmput(ctx->mm);
1827 	} else {
1828 		return -ESRCH;
1829 	}
1830 
1831 	if (ret)
1832 		return ret;
1833 
1834 	if (!mode_wp && !mode_dontwake) {
1835 		range.start = uffdio_wp.range.start;
1836 		range.len = uffdio_wp.range.len;
1837 		wake_userfault(ctx, &range);
1838 	}
1839 	return ret;
1840 }
1841 
1842 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1843 {
1844 	__s64 ret;
1845 	struct uffdio_continue uffdio_continue;
1846 	struct uffdio_continue __user *user_uffdio_continue;
1847 	struct userfaultfd_wake_range range;
1848 
1849 	user_uffdio_continue = (struct uffdio_continue __user *)arg;
1850 
1851 	ret = -EAGAIN;
1852 	if (atomic_read(&ctx->mmap_changing))
1853 		goto out;
1854 
1855 	ret = -EFAULT;
1856 	if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1857 			   /* don't copy the output fields */
1858 			   sizeof(uffdio_continue) - (sizeof(__s64))))
1859 		goto out;
1860 
1861 	ret = validate_range(ctx->mm, uffdio_continue.range.start,
1862 			     uffdio_continue.range.len);
1863 	if (ret)
1864 		goto out;
1865 
1866 	ret = -EINVAL;
1867 	/* double check for wraparound just in case. */
1868 	if (uffdio_continue.range.start + uffdio_continue.range.len <=
1869 	    uffdio_continue.range.start) {
1870 		goto out;
1871 	}
1872 	if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE)
1873 		goto out;
1874 
1875 	if (mmget_not_zero(ctx->mm)) {
1876 		ret = mcopy_continue(ctx->mm, uffdio_continue.range.start,
1877 				     uffdio_continue.range.len,
1878 				     &ctx->mmap_changing);
1879 		mmput(ctx->mm);
1880 	} else {
1881 		return -ESRCH;
1882 	}
1883 
1884 	if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1885 		return -EFAULT;
1886 	if (ret < 0)
1887 		goto out;
1888 
1889 	/* len == 0 would wake all */
1890 	BUG_ON(!ret);
1891 	range.len = ret;
1892 	if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1893 		range.start = uffdio_continue.range.start;
1894 		wake_userfault(ctx, &range);
1895 	}
1896 	ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1897 
1898 out:
1899 	return ret;
1900 }
1901 
1902 static inline unsigned int uffd_ctx_features(__u64 user_features)
1903 {
1904 	/*
1905 	 * For the current set of features the bits just coincide. Set
1906 	 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1907 	 */
1908 	return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1909 }
1910 
1911 /*
1912  * userland asks for a certain API version and we return which bits
1913  * and ioctl commands are implemented in this kernel for such API
1914  * version or -EINVAL if unknown.
1915  */
1916 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1917 			   unsigned long arg)
1918 {
1919 	struct uffdio_api uffdio_api;
1920 	void __user *buf = (void __user *)arg;
1921 	unsigned int ctx_features;
1922 	int ret;
1923 	__u64 features;
1924 
1925 	ret = -EFAULT;
1926 	if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1927 		goto out;
1928 	/* Ignore unsupported features (userspace built against newer kernel) */
1929 	features = uffdio_api.features & UFFD_API_FEATURES;
1930 	ret = -EPERM;
1931 	if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1932 		goto err_out;
1933 	/* report all available features and ioctls to userland */
1934 	uffdio_api.features = UFFD_API_FEATURES;
1935 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1936 	uffdio_api.features &=
1937 		~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1938 #endif
1939 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1940 	uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
1941 #endif
1942 #ifndef CONFIG_PTE_MARKER_UFFD_WP
1943 	uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
1944 #endif
1945 	uffdio_api.ioctls = UFFD_API_IOCTLS;
1946 	ret = -EFAULT;
1947 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1948 		goto out;
1949 
1950 	/* only enable the requested features for this uffd context */
1951 	ctx_features = uffd_ctx_features(features);
1952 	ret = -EINVAL;
1953 	if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
1954 		goto err_out;
1955 
1956 	ret = 0;
1957 out:
1958 	return ret;
1959 err_out:
1960 	memset(&uffdio_api, 0, sizeof(uffdio_api));
1961 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1962 		ret = -EFAULT;
1963 	goto out;
1964 }
1965 
1966 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1967 			      unsigned long arg)
1968 {
1969 	int ret = -EINVAL;
1970 	struct userfaultfd_ctx *ctx = file->private_data;
1971 
1972 	if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
1973 		return -EINVAL;
1974 
1975 	switch(cmd) {
1976 	case UFFDIO_API:
1977 		ret = userfaultfd_api(ctx, arg);
1978 		break;
1979 	case UFFDIO_REGISTER:
1980 		ret = userfaultfd_register(ctx, arg);
1981 		break;
1982 	case UFFDIO_UNREGISTER:
1983 		ret = userfaultfd_unregister(ctx, arg);
1984 		break;
1985 	case UFFDIO_WAKE:
1986 		ret = userfaultfd_wake(ctx, arg);
1987 		break;
1988 	case UFFDIO_COPY:
1989 		ret = userfaultfd_copy(ctx, arg);
1990 		break;
1991 	case UFFDIO_ZEROPAGE:
1992 		ret = userfaultfd_zeropage(ctx, arg);
1993 		break;
1994 	case UFFDIO_WRITEPROTECT:
1995 		ret = userfaultfd_writeprotect(ctx, arg);
1996 		break;
1997 	case UFFDIO_CONTINUE:
1998 		ret = userfaultfd_continue(ctx, arg);
1999 		break;
2000 	}
2001 	return ret;
2002 }
2003 
2004 #ifdef CONFIG_PROC_FS
2005 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2006 {
2007 	struct userfaultfd_ctx *ctx = f->private_data;
2008 	wait_queue_entry_t *wq;
2009 	unsigned long pending = 0, total = 0;
2010 
2011 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
2012 	list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2013 		pending++;
2014 		total++;
2015 	}
2016 	list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2017 		total++;
2018 	}
2019 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2020 
2021 	/*
2022 	 * If more protocols will be added, there will be all shown
2023 	 * separated by a space. Like this:
2024 	 *	protocols: aa:... bb:...
2025 	 */
2026 	seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2027 		   pending, total, UFFD_API, ctx->features,
2028 		   UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2029 }
2030 #endif
2031 
2032 static const struct file_operations userfaultfd_fops = {
2033 #ifdef CONFIG_PROC_FS
2034 	.show_fdinfo	= userfaultfd_show_fdinfo,
2035 #endif
2036 	.release	= userfaultfd_release,
2037 	.poll		= userfaultfd_poll,
2038 	.read		= userfaultfd_read,
2039 	.unlocked_ioctl = userfaultfd_ioctl,
2040 	.compat_ioctl	= compat_ptr_ioctl,
2041 	.llseek		= noop_llseek,
2042 };
2043 
2044 static void init_once_userfaultfd_ctx(void *mem)
2045 {
2046 	struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2047 
2048 	init_waitqueue_head(&ctx->fault_pending_wqh);
2049 	init_waitqueue_head(&ctx->fault_wqh);
2050 	init_waitqueue_head(&ctx->event_wqh);
2051 	init_waitqueue_head(&ctx->fd_wqh);
2052 	seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2053 }
2054 
2055 SYSCALL_DEFINE1(userfaultfd, int, flags)
2056 {
2057 	struct userfaultfd_ctx *ctx;
2058 	int fd;
2059 
2060 	if (!sysctl_unprivileged_userfaultfd &&
2061 	    (flags & UFFD_USER_MODE_ONLY) == 0 &&
2062 	    !capable(CAP_SYS_PTRACE)) {
2063 		printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
2064 			"sysctl knob to 1 if kernel faults must be handled "
2065 			"without obtaining CAP_SYS_PTRACE capability\n");
2066 		return -EPERM;
2067 	}
2068 
2069 	BUG_ON(!current->mm);
2070 
2071 	/* Check the UFFD_* constants for consistency.  */
2072 	BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2073 	BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2074 	BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2075 
2076 	if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2077 		return -EINVAL;
2078 
2079 	ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2080 	if (!ctx)
2081 		return -ENOMEM;
2082 
2083 	refcount_set(&ctx->refcount, 1);
2084 	ctx->flags = flags;
2085 	ctx->features = 0;
2086 	ctx->released = false;
2087 	atomic_set(&ctx->mmap_changing, 0);
2088 	ctx->mm = current->mm;
2089 	/* prevent the mm struct to be freed */
2090 	mmgrab(ctx->mm);
2091 
2092 	fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2093 			O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2094 	if (fd < 0) {
2095 		mmdrop(ctx->mm);
2096 		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2097 	}
2098 	return fd;
2099 }
2100 
2101 static int __init userfaultfd_init(void)
2102 {
2103 	userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2104 						sizeof(struct userfaultfd_ctx),
2105 						0,
2106 						SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2107 						init_once_userfaultfd_ctx);
2108 	return 0;
2109 }
2110 __initcall(userfaultfd_init);
2111