xref: /linux/fs/userfaultfd.c (revision c4101e55974cc7d835fbd2d8e01553a3f61e9e75)
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 #include <linux/miscdevice.h>
34 
35 static int sysctl_unprivileged_userfaultfd __read_mostly;
36 
37 #ifdef CONFIG_SYSCTL
38 static struct ctl_table vm_userfaultfd_table[] = {
39 	{
40 		.procname	= "unprivileged_userfaultfd",
41 		.data		= &sysctl_unprivileged_userfaultfd,
42 		.maxlen		= sizeof(sysctl_unprivileged_userfaultfd),
43 		.mode		= 0644,
44 		.proc_handler	= proc_dointvec_minmax,
45 		.extra1		= SYSCTL_ZERO,
46 		.extra2		= SYSCTL_ONE,
47 	},
48 };
49 #endif
50 
51 static struct kmem_cache *userfaultfd_ctx_cachep __ro_after_init;
52 
53 /*
54  * Start with fault_pending_wqh and fault_wqh so they're more likely
55  * to be in the same cacheline.
56  *
57  * Locking order:
58  *	fd_wqh.lock
59  *		fault_pending_wqh.lock
60  *			fault_wqh.lock
61  *		event_wqh.lock
62  *
63  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
64  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
65  * also taken in IRQ context.
66  */
67 struct userfaultfd_ctx {
68 	/* waitqueue head for the pending (i.e. not read) userfaults */
69 	wait_queue_head_t fault_pending_wqh;
70 	/* waitqueue head for the userfaults */
71 	wait_queue_head_t fault_wqh;
72 	/* waitqueue head for the pseudo fd to wakeup poll/read */
73 	wait_queue_head_t fd_wqh;
74 	/* waitqueue head for events */
75 	wait_queue_head_t event_wqh;
76 	/* a refile sequence protected by fault_pending_wqh lock */
77 	seqcount_spinlock_t refile_seq;
78 	/* pseudo fd refcounting */
79 	refcount_t refcount;
80 	/* userfaultfd syscall flags */
81 	unsigned int flags;
82 	/* features requested from the userspace */
83 	unsigned int features;
84 	/* released */
85 	bool released;
86 	/* memory mappings are changing because of non-cooperative event */
87 	atomic_t mmap_changing;
88 	/* mm with one ore more vmas attached to this userfaultfd_ctx */
89 	struct mm_struct *mm;
90 };
91 
92 struct userfaultfd_fork_ctx {
93 	struct userfaultfd_ctx *orig;
94 	struct userfaultfd_ctx *new;
95 	struct list_head list;
96 };
97 
98 struct userfaultfd_unmap_ctx {
99 	struct userfaultfd_ctx *ctx;
100 	unsigned long start;
101 	unsigned long end;
102 	struct list_head list;
103 };
104 
105 struct userfaultfd_wait_queue {
106 	struct uffd_msg msg;
107 	wait_queue_entry_t wq;
108 	struct userfaultfd_ctx *ctx;
109 	bool waken;
110 };
111 
112 struct userfaultfd_wake_range {
113 	unsigned long start;
114 	unsigned long len;
115 };
116 
117 /* internal indication that UFFD_API ioctl was successfully executed */
118 #define UFFD_FEATURE_INITIALIZED		(1u << 31)
119 
120 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
121 {
122 	return ctx->features & UFFD_FEATURE_INITIALIZED;
123 }
124 
125 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx *ctx)
126 {
127 	return ctx && (ctx->features & UFFD_FEATURE_WP_ASYNC);
128 }
129 
130 /*
131  * Whether WP_UNPOPULATED is enabled on the uffd context.  It is only
132  * meaningful when userfaultfd_wp()==true on the vma and when it's
133  * anonymous.
134  */
135 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
136 {
137 	struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
138 
139 	if (!ctx)
140 		return false;
141 
142 	return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
143 }
144 
145 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
146 				     vm_flags_t flags)
147 {
148 	const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP;
149 
150 	vm_flags_reset(vma, flags);
151 	/*
152 	 * For shared mappings, we want to enable writenotify while
153 	 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
154 	 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
155 	 */
156 	if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
157 		vma_set_page_prot(vma);
158 }
159 
160 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
161 				     int wake_flags, void *key)
162 {
163 	struct userfaultfd_wake_range *range = key;
164 	int ret;
165 	struct userfaultfd_wait_queue *uwq;
166 	unsigned long start, len;
167 
168 	uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
169 	ret = 0;
170 	/* len == 0 means wake all */
171 	start = range->start;
172 	len = range->len;
173 	if (len && (start > uwq->msg.arg.pagefault.address ||
174 		    start + len <= uwq->msg.arg.pagefault.address))
175 		goto out;
176 	WRITE_ONCE(uwq->waken, true);
177 	/*
178 	 * The Program-Order guarantees provided by the scheduler
179 	 * ensure uwq->waken is visible before the task is woken.
180 	 */
181 	ret = wake_up_state(wq->private, mode);
182 	if (ret) {
183 		/*
184 		 * Wake only once, autoremove behavior.
185 		 *
186 		 * After the effect of list_del_init is visible to the other
187 		 * CPUs, the waitqueue may disappear from under us, see the
188 		 * !list_empty_careful() in handle_userfault().
189 		 *
190 		 * try_to_wake_up() has an implicit smp_mb(), and the
191 		 * wq->private is read before calling the extern function
192 		 * "wake_up_state" (which in turns calls try_to_wake_up).
193 		 */
194 		list_del_init(&wq->entry);
195 	}
196 out:
197 	return ret;
198 }
199 
200 /**
201  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
202  * context.
203  * @ctx: [in] Pointer to the userfaultfd context.
204  */
205 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
206 {
207 	refcount_inc(&ctx->refcount);
208 }
209 
210 /**
211  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
212  * context.
213  * @ctx: [in] Pointer to userfaultfd context.
214  *
215  * The userfaultfd context reference must have been previously acquired either
216  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
217  */
218 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
219 {
220 	if (refcount_dec_and_test(&ctx->refcount)) {
221 		VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
222 		VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
223 		VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
224 		VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
225 		VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
226 		VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
227 		VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
228 		VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
229 		mmdrop(ctx->mm);
230 		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
231 	}
232 }
233 
234 static inline void msg_init(struct uffd_msg *msg)
235 {
236 	BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
237 	/*
238 	 * Must use memset to zero out the paddings or kernel data is
239 	 * leaked to userland.
240 	 */
241 	memset(msg, 0, sizeof(struct uffd_msg));
242 }
243 
244 static inline struct uffd_msg userfault_msg(unsigned long address,
245 					    unsigned long real_address,
246 					    unsigned int flags,
247 					    unsigned long reason,
248 					    unsigned int features)
249 {
250 	struct uffd_msg msg;
251 
252 	msg_init(&msg);
253 	msg.event = UFFD_EVENT_PAGEFAULT;
254 
255 	msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
256 				    real_address : address;
257 
258 	/*
259 	 * These flags indicate why the userfault occurred:
260 	 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
261 	 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
262 	 * - Neither of these flags being set indicates a MISSING fault.
263 	 *
264 	 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
265 	 * fault. Otherwise, it was a read fault.
266 	 */
267 	if (flags & FAULT_FLAG_WRITE)
268 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
269 	if (reason & VM_UFFD_WP)
270 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
271 	if (reason & VM_UFFD_MINOR)
272 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
273 	if (features & UFFD_FEATURE_THREAD_ID)
274 		msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
275 	return msg;
276 }
277 
278 #ifdef CONFIG_HUGETLB_PAGE
279 /*
280  * Same functionality as userfaultfd_must_wait below with modifications for
281  * hugepmd ranges.
282  */
283 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
284 					      struct vm_fault *vmf,
285 					      unsigned long reason)
286 {
287 	struct vm_area_struct *vma = vmf->vma;
288 	pte_t *ptep, pte;
289 	bool ret = true;
290 
291 	assert_fault_locked(vmf);
292 
293 	ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
294 	if (!ptep)
295 		goto out;
296 
297 	ret = false;
298 	pte = huge_ptep_get(ptep);
299 
300 	/*
301 	 * Lockless access: we're in a wait_event so it's ok if it
302 	 * changes under us.  PTE markers should be handled the same as none
303 	 * ptes here.
304 	 */
305 	if (huge_pte_none_mostly(pte))
306 		ret = true;
307 	if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
308 		ret = true;
309 out:
310 	return ret;
311 }
312 #else
313 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
314 					      struct vm_fault *vmf,
315 					      unsigned long reason)
316 {
317 	return false;	/* should never get here */
318 }
319 #endif /* CONFIG_HUGETLB_PAGE */
320 
321 /*
322  * Verify the pagetables are still not ok after having reigstered into
323  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
324  * userfault that has already been resolved, if userfaultfd_read and
325  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
326  * threads.
327  */
328 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
329 					 struct vm_fault *vmf,
330 					 unsigned long reason)
331 {
332 	struct mm_struct *mm = ctx->mm;
333 	unsigned long address = vmf->address;
334 	pgd_t *pgd;
335 	p4d_t *p4d;
336 	pud_t *pud;
337 	pmd_t *pmd, _pmd;
338 	pte_t *pte;
339 	pte_t ptent;
340 	bool ret = true;
341 
342 	assert_fault_locked(vmf);
343 
344 	pgd = pgd_offset(mm, address);
345 	if (!pgd_present(*pgd))
346 		goto out;
347 	p4d = p4d_offset(pgd, address);
348 	if (!p4d_present(*p4d))
349 		goto out;
350 	pud = pud_offset(p4d, address);
351 	if (!pud_present(*pud))
352 		goto out;
353 	pmd = pmd_offset(pud, address);
354 again:
355 	_pmd = pmdp_get_lockless(pmd);
356 	if (pmd_none(_pmd))
357 		goto out;
358 
359 	ret = false;
360 	if (!pmd_present(_pmd) || pmd_devmap(_pmd))
361 		goto out;
362 
363 	if (pmd_trans_huge(_pmd)) {
364 		if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
365 			ret = true;
366 		goto out;
367 	}
368 
369 	pte = pte_offset_map(pmd, address);
370 	if (!pte) {
371 		ret = true;
372 		goto again;
373 	}
374 	/*
375 	 * Lockless access: we're in a wait_event so it's ok if it
376 	 * changes under us.  PTE markers should be handled the same as none
377 	 * ptes here.
378 	 */
379 	ptent = ptep_get(pte);
380 	if (pte_none_mostly(ptent))
381 		ret = true;
382 	if (!pte_write(ptent) && (reason & VM_UFFD_WP))
383 		ret = true;
384 	pte_unmap(pte);
385 
386 out:
387 	return ret;
388 }
389 
390 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
391 {
392 	if (flags & FAULT_FLAG_INTERRUPTIBLE)
393 		return TASK_INTERRUPTIBLE;
394 
395 	if (flags & FAULT_FLAG_KILLABLE)
396 		return TASK_KILLABLE;
397 
398 	return TASK_UNINTERRUPTIBLE;
399 }
400 
401 /*
402  * The locking rules involved in returning VM_FAULT_RETRY depending on
403  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
404  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
405  * recommendation in __lock_page_or_retry is not an understatement.
406  *
407  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
408  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
409  * not set.
410  *
411  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
412  * set, VM_FAULT_RETRY can still be returned if and only if there are
413  * fatal_signal_pending()s, and the mmap_lock must be released before
414  * returning it.
415  */
416 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
417 {
418 	struct vm_area_struct *vma = vmf->vma;
419 	struct mm_struct *mm = vma->vm_mm;
420 	struct userfaultfd_ctx *ctx;
421 	struct userfaultfd_wait_queue uwq;
422 	vm_fault_t ret = VM_FAULT_SIGBUS;
423 	bool must_wait;
424 	unsigned int blocking_state;
425 
426 	/*
427 	 * We don't do userfault handling for the final child pid update.
428 	 *
429 	 * We also don't do userfault handling during
430 	 * coredumping. hugetlbfs has the special
431 	 * hugetlb_follow_page_mask() to skip missing pages in the
432 	 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
433 	 * the no_page_table() helper in follow_page_mask(), but the
434 	 * shmem_vm_ops->fault method is invoked even during
435 	 * coredumping and it ends up here.
436 	 */
437 	if (current->flags & (PF_EXITING|PF_DUMPCORE))
438 		goto out;
439 
440 	assert_fault_locked(vmf);
441 
442 	ctx = vma->vm_userfaultfd_ctx.ctx;
443 	if (!ctx)
444 		goto out;
445 
446 	BUG_ON(ctx->mm != mm);
447 
448 	/* Any unrecognized flag is a bug. */
449 	VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
450 	/* 0 or > 1 flags set is a bug; we expect exactly 1. */
451 	VM_BUG_ON(!reason || (reason & (reason - 1)));
452 
453 	if (ctx->features & UFFD_FEATURE_SIGBUS)
454 		goto out;
455 	if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY))
456 		goto out;
457 
458 	/*
459 	 * If it's already released don't get it. This avoids to loop
460 	 * in __get_user_pages if userfaultfd_release waits on the
461 	 * caller of handle_userfault to release the mmap_lock.
462 	 */
463 	if (unlikely(READ_ONCE(ctx->released))) {
464 		/*
465 		 * Don't return VM_FAULT_SIGBUS in this case, so a non
466 		 * cooperative manager can close the uffd after the
467 		 * last UFFDIO_COPY, without risking to trigger an
468 		 * involuntary SIGBUS if the process was starting the
469 		 * userfaultfd while the userfaultfd was still armed
470 		 * (but after the last UFFDIO_COPY). If the uffd
471 		 * wasn't already closed when the userfault reached
472 		 * this point, that would normally be solved by
473 		 * userfaultfd_must_wait returning 'false'.
474 		 *
475 		 * If we were to return VM_FAULT_SIGBUS here, the non
476 		 * cooperative manager would be instead forced to
477 		 * always call UFFDIO_UNREGISTER before it can safely
478 		 * close the uffd.
479 		 */
480 		ret = VM_FAULT_NOPAGE;
481 		goto out;
482 	}
483 
484 	/*
485 	 * Check that we can return VM_FAULT_RETRY.
486 	 *
487 	 * NOTE: it should become possible to return VM_FAULT_RETRY
488 	 * even if FAULT_FLAG_TRIED is set without leading to gup()
489 	 * -EBUSY failures, if the userfaultfd is to be extended for
490 	 * VM_UFFD_WP tracking and we intend to arm the userfault
491 	 * without first stopping userland access to the memory. For
492 	 * VM_UFFD_MISSING userfaults this is enough for now.
493 	 */
494 	if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
495 		/*
496 		 * Validate the invariant that nowait must allow retry
497 		 * to be sure not to return SIGBUS erroneously on
498 		 * nowait invocations.
499 		 */
500 		BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
501 #ifdef CONFIG_DEBUG_VM
502 		if (printk_ratelimit()) {
503 			printk(KERN_WARNING
504 			       "FAULT_FLAG_ALLOW_RETRY missing %x\n",
505 			       vmf->flags);
506 			dump_stack();
507 		}
508 #endif
509 		goto out;
510 	}
511 
512 	/*
513 	 * Handle nowait, not much to do other than tell it to retry
514 	 * and wait.
515 	 */
516 	ret = VM_FAULT_RETRY;
517 	if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
518 		goto out;
519 
520 	/* take the reference before dropping the mmap_lock */
521 	userfaultfd_ctx_get(ctx);
522 
523 	init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
524 	uwq.wq.private = current;
525 	uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
526 				reason, ctx->features);
527 	uwq.ctx = ctx;
528 	uwq.waken = false;
529 
530 	blocking_state = userfaultfd_get_blocking_state(vmf->flags);
531 
532         /*
533          * Take the vma lock now, in order to safely call
534          * userfaultfd_huge_must_wait() later. Since acquiring the
535          * (sleepable) vma lock can modify the current task state, that
536          * must be before explicitly calling set_current_state().
537          */
538 	if (is_vm_hugetlb_page(vma))
539 		hugetlb_vma_lock_read(vma);
540 
541 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
542 	/*
543 	 * After the __add_wait_queue the uwq is visible to userland
544 	 * through poll/read().
545 	 */
546 	__add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
547 	/*
548 	 * The smp_mb() after __set_current_state prevents the reads
549 	 * following the spin_unlock to happen before the list_add in
550 	 * __add_wait_queue.
551 	 */
552 	set_current_state(blocking_state);
553 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
554 
555 	if (!is_vm_hugetlb_page(vma))
556 		must_wait = userfaultfd_must_wait(ctx, vmf, reason);
557 	else
558 		must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
559 	if (is_vm_hugetlb_page(vma))
560 		hugetlb_vma_unlock_read(vma);
561 	release_fault_lock(vmf);
562 
563 	if (likely(must_wait && !READ_ONCE(ctx->released))) {
564 		wake_up_poll(&ctx->fd_wqh, EPOLLIN);
565 		schedule();
566 	}
567 
568 	__set_current_state(TASK_RUNNING);
569 
570 	/*
571 	 * Here we race with the list_del; list_add in
572 	 * userfaultfd_ctx_read(), however because we don't ever run
573 	 * list_del_init() to refile across the two lists, the prev
574 	 * and next pointers will never point to self. list_add also
575 	 * would never let any of the two pointers to point to
576 	 * self. So list_empty_careful won't risk to see both pointers
577 	 * pointing to self at any time during the list refile. The
578 	 * only case where list_del_init() is called is the full
579 	 * removal in the wake function and there we don't re-list_add
580 	 * and it's fine not to block on the spinlock. The uwq on this
581 	 * kernel stack can be released after the list_del_init.
582 	 */
583 	if (!list_empty_careful(&uwq.wq.entry)) {
584 		spin_lock_irq(&ctx->fault_pending_wqh.lock);
585 		/*
586 		 * No need of list_del_init(), the uwq on the stack
587 		 * will be freed shortly anyway.
588 		 */
589 		list_del(&uwq.wq.entry);
590 		spin_unlock_irq(&ctx->fault_pending_wqh.lock);
591 	}
592 
593 	/*
594 	 * ctx may go away after this if the userfault pseudo fd is
595 	 * already released.
596 	 */
597 	userfaultfd_ctx_put(ctx);
598 
599 out:
600 	return ret;
601 }
602 
603 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
604 					      struct userfaultfd_wait_queue *ewq)
605 {
606 	struct userfaultfd_ctx *release_new_ctx;
607 
608 	if (WARN_ON_ONCE(current->flags & PF_EXITING))
609 		goto out;
610 
611 	ewq->ctx = ctx;
612 	init_waitqueue_entry(&ewq->wq, current);
613 	release_new_ctx = NULL;
614 
615 	spin_lock_irq(&ctx->event_wqh.lock);
616 	/*
617 	 * After the __add_wait_queue the uwq is visible to userland
618 	 * through poll/read().
619 	 */
620 	__add_wait_queue(&ctx->event_wqh, &ewq->wq);
621 	for (;;) {
622 		set_current_state(TASK_KILLABLE);
623 		if (ewq->msg.event == 0)
624 			break;
625 		if (READ_ONCE(ctx->released) ||
626 		    fatal_signal_pending(current)) {
627 			/*
628 			 * &ewq->wq may be queued in fork_event, but
629 			 * __remove_wait_queue ignores the head
630 			 * parameter. It would be a problem if it
631 			 * didn't.
632 			 */
633 			__remove_wait_queue(&ctx->event_wqh, &ewq->wq);
634 			if (ewq->msg.event == UFFD_EVENT_FORK) {
635 				struct userfaultfd_ctx *new;
636 
637 				new = (struct userfaultfd_ctx *)
638 					(unsigned long)
639 					ewq->msg.arg.reserved.reserved1;
640 				release_new_ctx = new;
641 			}
642 			break;
643 		}
644 
645 		spin_unlock_irq(&ctx->event_wqh.lock);
646 
647 		wake_up_poll(&ctx->fd_wqh, EPOLLIN);
648 		schedule();
649 
650 		spin_lock_irq(&ctx->event_wqh.lock);
651 	}
652 	__set_current_state(TASK_RUNNING);
653 	spin_unlock_irq(&ctx->event_wqh.lock);
654 
655 	if (release_new_ctx) {
656 		struct vm_area_struct *vma;
657 		struct mm_struct *mm = release_new_ctx->mm;
658 		VMA_ITERATOR(vmi, mm, 0);
659 
660 		/* the various vma->vm_userfaultfd_ctx still points to it */
661 		mmap_write_lock(mm);
662 		for_each_vma(vmi, vma) {
663 			if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
664 				vma_start_write(vma);
665 				vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
666 				userfaultfd_set_vm_flags(vma,
667 							 vma->vm_flags & ~__VM_UFFD_FLAGS);
668 			}
669 		}
670 		mmap_write_unlock(mm);
671 
672 		userfaultfd_ctx_put(release_new_ctx);
673 	}
674 
675 	/*
676 	 * ctx may go away after this if the userfault pseudo fd is
677 	 * already released.
678 	 */
679 out:
680 	atomic_dec(&ctx->mmap_changing);
681 	VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
682 	userfaultfd_ctx_put(ctx);
683 }
684 
685 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
686 				       struct userfaultfd_wait_queue *ewq)
687 {
688 	ewq->msg.event = 0;
689 	wake_up_locked(&ctx->event_wqh);
690 	__remove_wait_queue(&ctx->event_wqh, &ewq->wq);
691 }
692 
693 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
694 {
695 	struct userfaultfd_ctx *ctx = NULL, *octx;
696 	struct userfaultfd_fork_ctx *fctx;
697 
698 	octx = vma->vm_userfaultfd_ctx.ctx;
699 	if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
700 		vma_start_write(vma);
701 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
702 		userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
703 		return 0;
704 	}
705 
706 	list_for_each_entry(fctx, fcs, list)
707 		if (fctx->orig == octx) {
708 			ctx = fctx->new;
709 			break;
710 		}
711 
712 	if (!ctx) {
713 		fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
714 		if (!fctx)
715 			return -ENOMEM;
716 
717 		ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
718 		if (!ctx) {
719 			kfree(fctx);
720 			return -ENOMEM;
721 		}
722 
723 		refcount_set(&ctx->refcount, 1);
724 		ctx->flags = octx->flags;
725 		ctx->features = octx->features;
726 		ctx->released = false;
727 		atomic_set(&ctx->mmap_changing, 0);
728 		ctx->mm = vma->vm_mm;
729 		mmgrab(ctx->mm);
730 
731 		userfaultfd_ctx_get(octx);
732 		atomic_inc(&octx->mmap_changing);
733 		fctx->orig = octx;
734 		fctx->new = ctx;
735 		list_add_tail(&fctx->list, fcs);
736 	}
737 
738 	vma->vm_userfaultfd_ctx.ctx = ctx;
739 	return 0;
740 }
741 
742 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
743 {
744 	struct userfaultfd_ctx *ctx = fctx->orig;
745 	struct userfaultfd_wait_queue ewq;
746 
747 	msg_init(&ewq.msg);
748 
749 	ewq.msg.event = UFFD_EVENT_FORK;
750 	ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
751 
752 	userfaultfd_event_wait_completion(ctx, &ewq);
753 }
754 
755 void dup_userfaultfd_complete(struct list_head *fcs)
756 {
757 	struct userfaultfd_fork_ctx *fctx, *n;
758 
759 	list_for_each_entry_safe(fctx, n, fcs, list) {
760 		dup_fctx(fctx);
761 		list_del(&fctx->list);
762 		kfree(fctx);
763 	}
764 }
765 
766 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
767 			     struct vm_userfaultfd_ctx *vm_ctx)
768 {
769 	struct userfaultfd_ctx *ctx;
770 
771 	ctx = vma->vm_userfaultfd_ctx.ctx;
772 
773 	if (!ctx)
774 		return;
775 
776 	if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
777 		vm_ctx->ctx = ctx;
778 		userfaultfd_ctx_get(ctx);
779 		atomic_inc(&ctx->mmap_changing);
780 	} else {
781 		/* Drop uffd context if remap feature not enabled */
782 		vma_start_write(vma);
783 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
784 		userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
785 	}
786 }
787 
788 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
789 				 unsigned long from, unsigned long to,
790 				 unsigned long len)
791 {
792 	struct userfaultfd_ctx *ctx = vm_ctx->ctx;
793 	struct userfaultfd_wait_queue ewq;
794 
795 	if (!ctx)
796 		return;
797 
798 	if (to & ~PAGE_MASK) {
799 		userfaultfd_ctx_put(ctx);
800 		return;
801 	}
802 
803 	msg_init(&ewq.msg);
804 
805 	ewq.msg.event = UFFD_EVENT_REMAP;
806 	ewq.msg.arg.remap.from = from;
807 	ewq.msg.arg.remap.to = to;
808 	ewq.msg.arg.remap.len = len;
809 
810 	userfaultfd_event_wait_completion(ctx, &ewq);
811 }
812 
813 bool userfaultfd_remove(struct vm_area_struct *vma,
814 			unsigned long start, unsigned long end)
815 {
816 	struct mm_struct *mm = vma->vm_mm;
817 	struct userfaultfd_ctx *ctx;
818 	struct userfaultfd_wait_queue ewq;
819 
820 	ctx = vma->vm_userfaultfd_ctx.ctx;
821 	if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
822 		return true;
823 
824 	userfaultfd_ctx_get(ctx);
825 	atomic_inc(&ctx->mmap_changing);
826 	mmap_read_unlock(mm);
827 
828 	msg_init(&ewq.msg);
829 
830 	ewq.msg.event = UFFD_EVENT_REMOVE;
831 	ewq.msg.arg.remove.start = start;
832 	ewq.msg.arg.remove.end = end;
833 
834 	userfaultfd_event_wait_completion(ctx, &ewq);
835 
836 	return false;
837 }
838 
839 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
840 			  unsigned long start, unsigned long end)
841 {
842 	struct userfaultfd_unmap_ctx *unmap_ctx;
843 
844 	list_for_each_entry(unmap_ctx, unmaps, list)
845 		if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
846 		    unmap_ctx->end == end)
847 			return true;
848 
849 	return false;
850 }
851 
852 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
853 			   unsigned long end, struct list_head *unmaps)
854 {
855 	struct userfaultfd_unmap_ctx *unmap_ctx;
856 	struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
857 
858 	if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
859 	    has_unmap_ctx(ctx, unmaps, start, end))
860 		return 0;
861 
862 	unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
863 	if (!unmap_ctx)
864 		return -ENOMEM;
865 
866 	userfaultfd_ctx_get(ctx);
867 	atomic_inc(&ctx->mmap_changing);
868 	unmap_ctx->ctx = ctx;
869 	unmap_ctx->start = start;
870 	unmap_ctx->end = end;
871 	list_add_tail(&unmap_ctx->list, unmaps);
872 
873 	return 0;
874 }
875 
876 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
877 {
878 	struct userfaultfd_unmap_ctx *ctx, *n;
879 	struct userfaultfd_wait_queue ewq;
880 
881 	list_for_each_entry_safe(ctx, n, uf, list) {
882 		msg_init(&ewq.msg);
883 
884 		ewq.msg.event = UFFD_EVENT_UNMAP;
885 		ewq.msg.arg.remove.start = ctx->start;
886 		ewq.msg.arg.remove.end = ctx->end;
887 
888 		userfaultfd_event_wait_completion(ctx->ctx, &ewq);
889 
890 		list_del(&ctx->list);
891 		kfree(ctx);
892 	}
893 }
894 
895 static int userfaultfd_release(struct inode *inode, struct file *file)
896 {
897 	struct userfaultfd_ctx *ctx = file->private_data;
898 	struct mm_struct *mm = ctx->mm;
899 	struct vm_area_struct *vma, *prev;
900 	/* len == 0 means wake all */
901 	struct userfaultfd_wake_range range = { .len = 0, };
902 	unsigned long new_flags;
903 	VMA_ITERATOR(vmi, mm, 0);
904 
905 	WRITE_ONCE(ctx->released, true);
906 
907 	if (!mmget_not_zero(mm))
908 		goto wakeup;
909 
910 	/*
911 	 * Flush page faults out of all CPUs. NOTE: all page faults
912 	 * must be retried without returning VM_FAULT_SIGBUS if
913 	 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
914 	 * changes while handle_userfault released the mmap_lock. So
915 	 * it's critical that released is set to true (above), before
916 	 * taking the mmap_lock for writing.
917 	 */
918 	mmap_write_lock(mm);
919 	prev = NULL;
920 	for_each_vma(vmi, vma) {
921 		cond_resched();
922 		BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
923 		       !!(vma->vm_flags & __VM_UFFD_FLAGS));
924 		if (vma->vm_userfaultfd_ctx.ctx != ctx) {
925 			prev = vma;
926 			continue;
927 		}
928 		new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
929 		vma = vma_modify_flags_uffd(&vmi, prev, vma, vma->vm_start,
930 					    vma->vm_end, new_flags,
931 					    NULL_VM_UFFD_CTX);
932 
933 		vma_start_write(vma);
934 		userfaultfd_set_vm_flags(vma, new_flags);
935 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
936 
937 		prev = vma;
938 	}
939 	mmap_write_unlock(mm);
940 	mmput(mm);
941 wakeup:
942 	/*
943 	 * After no new page faults can wait on this fault_*wqh, flush
944 	 * the last page faults that may have been already waiting on
945 	 * the fault_*wqh.
946 	 */
947 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
948 	__wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
949 	__wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
950 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
951 
952 	/* Flush pending events that may still wait on event_wqh */
953 	wake_up_all(&ctx->event_wqh);
954 
955 	wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
956 	userfaultfd_ctx_put(ctx);
957 	return 0;
958 }
959 
960 /* fault_pending_wqh.lock must be hold by the caller */
961 static inline struct userfaultfd_wait_queue *find_userfault_in(
962 		wait_queue_head_t *wqh)
963 {
964 	wait_queue_entry_t *wq;
965 	struct userfaultfd_wait_queue *uwq;
966 
967 	lockdep_assert_held(&wqh->lock);
968 
969 	uwq = NULL;
970 	if (!waitqueue_active(wqh))
971 		goto out;
972 	/* walk in reverse to provide FIFO behavior to read userfaults */
973 	wq = list_last_entry(&wqh->head, typeof(*wq), entry);
974 	uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
975 out:
976 	return uwq;
977 }
978 
979 static inline struct userfaultfd_wait_queue *find_userfault(
980 		struct userfaultfd_ctx *ctx)
981 {
982 	return find_userfault_in(&ctx->fault_pending_wqh);
983 }
984 
985 static inline struct userfaultfd_wait_queue *find_userfault_evt(
986 		struct userfaultfd_ctx *ctx)
987 {
988 	return find_userfault_in(&ctx->event_wqh);
989 }
990 
991 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
992 {
993 	struct userfaultfd_ctx *ctx = file->private_data;
994 	__poll_t ret;
995 
996 	poll_wait(file, &ctx->fd_wqh, wait);
997 
998 	if (!userfaultfd_is_initialized(ctx))
999 		return EPOLLERR;
1000 
1001 	/*
1002 	 * poll() never guarantees that read won't block.
1003 	 * userfaults can be waken before they're read().
1004 	 */
1005 	if (unlikely(!(file->f_flags & O_NONBLOCK)))
1006 		return EPOLLERR;
1007 	/*
1008 	 * lockless access to see if there are pending faults
1009 	 * __pollwait last action is the add_wait_queue but
1010 	 * the spin_unlock would allow the waitqueue_active to
1011 	 * pass above the actual list_add inside
1012 	 * add_wait_queue critical section. So use a full
1013 	 * memory barrier to serialize the list_add write of
1014 	 * add_wait_queue() with the waitqueue_active read
1015 	 * below.
1016 	 */
1017 	ret = 0;
1018 	smp_mb();
1019 	if (waitqueue_active(&ctx->fault_pending_wqh))
1020 		ret = EPOLLIN;
1021 	else if (waitqueue_active(&ctx->event_wqh))
1022 		ret = EPOLLIN;
1023 
1024 	return ret;
1025 }
1026 
1027 static const struct file_operations userfaultfd_fops;
1028 
1029 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
1030 				  struct inode *inode,
1031 				  struct uffd_msg *msg)
1032 {
1033 	int fd;
1034 
1035 	fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
1036 			O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1037 	if (fd < 0)
1038 		return fd;
1039 
1040 	msg->arg.reserved.reserved1 = 0;
1041 	msg->arg.fork.ufd = fd;
1042 	return 0;
1043 }
1044 
1045 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1046 				    struct uffd_msg *msg, struct inode *inode)
1047 {
1048 	ssize_t ret;
1049 	DECLARE_WAITQUEUE(wait, current);
1050 	struct userfaultfd_wait_queue *uwq;
1051 	/*
1052 	 * Handling fork event requires sleeping operations, so
1053 	 * we drop the event_wqh lock, then do these ops, then
1054 	 * lock it back and wake up the waiter. While the lock is
1055 	 * dropped the ewq may go away so we keep track of it
1056 	 * carefully.
1057 	 */
1058 	LIST_HEAD(fork_event);
1059 	struct userfaultfd_ctx *fork_nctx = NULL;
1060 
1061 	/* always take the fd_wqh lock before the fault_pending_wqh lock */
1062 	spin_lock_irq(&ctx->fd_wqh.lock);
1063 	__add_wait_queue(&ctx->fd_wqh, &wait);
1064 	for (;;) {
1065 		set_current_state(TASK_INTERRUPTIBLE);
1066 		spin_lock(&ctx->fault_pending_wqh.lock);
1067 		uwq = find_userfault(ctx);
1068 		if (uwq) {
1069 			/*
1070 			 * Use a seqcount to repeat the lockless check
1071 			 * in wake_userfault() to avoid missing
1072 			 * wakeups because during the refile both
1073 			 * waitqueue could become empty if this is the
1074 			 * only userfault.
1075 			 */
1076 			write_seqcount_begin(&ctx->refile_seq);
1077 
1078 			/*
1079 			 * The fault_pending_wqh.lock prevents the uwq
1080 			 * to disappear from under us.
1081 			 *
1082 			 * Refile this userfault from
1083 			 * fault_pending_wqh to fault_wqh, it's not
1084 			 * pending anymore after we read it.
1085 			 *
1086 			 * Use list_del() by hand (as
1087 			 * userfaultfd_wake_function also uses
1088 			 * list_del_init() by hand) to be sure nobody
1089 			 * changes __remove_wait_queue() to use
1090 			 * list_del_init() in turn breaking the
1091 			 * !list_empty_careful() check in
1092 			 * handle_userfault(). The uwq->wq.head list
1093 			 * must never be empty at any time during the
1094 			 * refile, or the waitqueue could disappear
1095 			 * from under us. The "wait_queue_head_t"
1096 			 * parameter of __remove_wait_queue() is unused
1097 			 * anyway.
1098 			 */
1099 			list_del(&uwq->wq.entry);
1100 			add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1101 
1102 			write_seqcount_end(&ctx->refile_seq);
1103 
1104 			/* careful to always initialize msg if ret == 0 */
1105 			*msg = uwq->msg;
1106 			spin_unlock(&ctx->fault_pending_wqh.lock);
1107 			ret = 0;
1108 			break;
1109 		}
1110 		spin_unlock(&ctx->fault_pending_wqh.lock);
1111 
1112 		spin_lock(&ctx->event_wqh.lock);
1113 		uwq = find_userfault_evt(ctx);
1114 		if (uwq) {
1115 			*msg = uwq->msg;
1116 
1117 			if (uwq->msg.event == UFFD_EVENT_FORK) {
1118 				fork_nctx = (struct userfaultfd_ctx *)
1119 					(unsigned long)
1120 					uwq->msg.arg.reserved.reserved1;
1121 				list_move(&uwq->wq.entry, &fork_event);
1122 				/*
1123 				 * fork_nctx can be freed as soon as
1124 				 * we drop the lock, unless we take a
1125 				 * reference on it.
1126 				 */
1127 				userfaultfd_ctx_get(fork_nctx);
1128 				spin_unlock(&ctx->event_wqh.lock);
1129 				ret = 0;
1130 				break;
1131 			}
1132 
1133 			userfaultfd_event_complete(ctx, uwq);
1134 			spin_unlock(&ctx->event_wqh.lock);
1135 			ret = 0;
1136 			break;
1137 		}
1138 		spin_unlock(&ctx->event_wqh.lock);
1139 
1140 		if (signal_pending(current)) {
1141 			ret = -ERESTARTSYS;
1142 			break;
1143 		}
1144 		if (no_wait) {
1145 			ret = -EAGAIN;
1146 			break;
1147 		}
1148 		spin_unlock_irq(&ctx->fd_wqh.lock);
1149 		schedule();
1150 		spin_lock_irq(&ctx->fd_wqh.lock);
1151 	}
1152 	__remove_wait_queue(&ctx->fd_wqh, &wait);
1153 	__set_current_state(TASK_RUNNING);
1154 	spin_unlock_irq(&ctx->fd_wqh.lock);
1155 
1156 	if (!ret && msg->event == UFFD_EVENT_FORK) {
1157 		ret = resolve_userfault_fork(fork_nctx, inode, msg);
1158 		spin_lock_irq(&ctx->event_wqh.lock);
1159 		if (!list_empty(&fork_event)) {
1160 			/*
1161 			 * The fork thread didn't abort, so we can
1162 			 * drop the temporary refcount.
1163 			 */
1164 			userfaultfd_ctx_put(fork_nctx);
1165 
1166 			uwq = list_first_entry(&fork_event,
1167 					       typeof(*uwq),
1168 					       wq.entry);
1169 			/*
1170 			 * If fork_event list wasn't empty and in turn
1171 			 * the event wasn't already released by fork
1172 			 * (the event is allocated on fork kernel
1173 			 * stack), put the event back to its place in
1174 			 * the event_wq. fork_event head will be freed
1175 			 * as soon as we return so the event cannot
1176 			 * stay queued there no matter the current
1177 			 * "ret" value.
1178 			 */
1179 			list_del(&uwq->wq.entry);
1180 			__add_wait_queue(&ctx->event_wqh, &uwq->wq);
1181 
1182 			/*
1183 			 * Leave the event in the waitqueue and report
1184 			 * error to userland if we failed to resolve
1185 			 * the userfault fork.
1186 			 */
1187 			if (likely(!ret))
1188 				userfaultfd_event_complete(ctx, uwq);
1189 		} else {
1190 			/*
1191 			 * Here the fork thread aborted and the
1192 			 * refcount from the fork thread on fork_nctx
1193 			 * has already been released. We still hold
1194 			 * the reference we took before releasing the
1195 			 * lock above. If resolve_userfault_fork
1196 			 * failed we've to drop it because the
1197 			 * fork_nctx has to be freed in such case. If
1198 			 * it succeeded we'll hold it because the new
1199 			 * uffd references it.
1200 			 */
1201 			if (ret)
1202 				userfaultfd_ctx_put(fork_nctx);
1203 		}
1204 		spin_unlock_irq(&ctx->event_wqh.lock);
1205 	}
1206 
1207 	return ret;
1208 }
1209 
1210 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1211 				size_t count, loff_t *ppos)
1212 {
1213 	struct userfaultfd_ctx *ctx = file->private_data;
1214 	ssize_t _ret, ret = 0;
1215 	struct uffd_msg msg;
1216 	int no_wait = file->f_flags & O_NONBLOCK;
1217 	struct inode *inode = file_inode(file);
1218 
1219 	if (!userfaultfd_is_initialized(ctx))
1220 		return -EINVAL;
1221 
1222 	for (;;) {
1223 		if (count < sizeof(msg))
1224 			return ret ? ret : -EINVAL;
1225 		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1226 		if (_ret < 0)
1227 			return ret ? ret : _ret;
1228 		if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1229 			return ret ? ret : -EFAULT;
1230 		ret += sizeof(msg);
1231 		buf += sizeof(msg);
1232 		count -= sizeof(msg);
1233 		/*
1234 		 * Allow to read more than one fault at time but only
1235 		 * block if waiting for the very first one.
1236 		 */
1237 		no_wait = O_NONBLOCK;
1238 	}
1239 }
1240 
1241 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1242 			     struct userfaultfd_wake_range *range)
1243 {
1244 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
1245 	/* wake all in the range and autoremove */
1246 	if (waitqueue_active(&ctx->fault_pending_wqh))
1247 		__wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1248 				     range);
1249 	if (waitqueue_active(&ctx->fault_wqh))
1250 		__wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1251 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1252 }
1253 
1254 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1255 					   struct userfaultfd_wake_range *range)
1256 {
1257 	unsigned seq;
1258 	bool need_wakeup;
1259 
1260 	/*
1261 	 * To be sure waitqueue_active() is not reordered by the CPU
1262 	 * before the pagetable update, use an explicit SMP memory
1263 	 * barrier here. PT lock release or mmap_read_unlock(mm) still
1264 	 * have release semantics that can allow the
1265 	 * waitqueue_active() to be reordered before the pte update.
1266 	 */
1267 	smp_mb();
1268 
1269 	/*
1270 	 * Use waitqueue_active because it's very frequent to
1271 	 * change the address space atomically even if there are no
1272 	 * userfaults yet. So we take the spinlock only when we're
1273 	 * sure we've userfaults to wake.
1274 	 */
1275 	do {
1276 		seq = read_seqcount_begin(&ctx->refile_seq);
1277 		need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1278 			waitqueue_active(&ctx->fault_wqh);
1279 		cond_resched();
1280 	} while (read_seqcount_retry(&ctx->refile_seq, seq));
1281 	if (need_wakeup)
1282 		__wake_userfault(ctx, range);
1283 }
1284 
1285 static __always_inline int validate_unaligned_range(
1286 	struct mm_struct *mm, __u64 start, __u64 len)
1287 {
1288 	__u64 task_size = mm->task_size;
1289 
1290 	if (len & ~PAGE_MASK)
1291 		return -EINVAL;
1292 	if (!len)
1293 		return -EINVAL;
1294 	if (start < mmap_min_addr)
1295 		return -EINVAL;
1296 	if (start >= task_size)
1297 		return -EINVAL;
1298 	if (len > task_size - start)
1299 		return -EINVAL;
1300 	if (start + len <= start)
1301 		return -EINVAL;
1302 	return 0;
1303 }
1304 
1305 static __always_inline int validate_range(struct mm_struct *mm,
1306 					  __u64 start, __u64 len)
1307 {
1308 	if (start & ~PAGE_MASK)
1309 		return -EINVAL;
1310 
1311 	return validate_unaligned_range(mm, start, len);
1312 }
1313 
1314 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1315 				unsigned long arg)
1316 {
1317 	struct mm_struct *mm = ctx->mm;
1318 	struct vm_area_struct *vma, *prev, *cur;
1319 	int ret;
1320 	struct uffdio_register uffdio_register;
1321 	struct uffdio_register __user *user_uffdio_register;
1322 	unsigned long vm_flags, new_flags;
1323 	bool found;
1324 	bool basic_ioctls;
1325 	unsigned long start, end, vma_end;
1326 	struct vma_iterator vmi;
1327 	bool wp_async = userfaultfd_wp_async_ctx(ctx);
1328 
1329 	user_uffdio_register = (struct uffdio_register __user *) arg;
1330 
1331 	ret = -EFAULT;
1332 	if (copy_from_user(&uffdio_register, user_uffdio_register,
1333 			   sizeof(uffdio_register)-sizeof(__u64)))
1334 		goto out;
1335 
1336 	ret = -EINVAL;
1337 	if (!uffdio_register.mode)
1338 		goto out;
1339 	if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1340 		goto out;
1341 	vm_flags = 0;
1342 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1343 		vm_flags |= VM_UFFD_MISSING;
1344 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1345 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1346 		goto out;
1347 #endif
1348 		vm_flags |= VM_UFFD_WP;
1349 	}
1350 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1351 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1352 		goto out;
1353 #endif
1354 		vm_flags |= VM_UFFD_MINOR;
1355 	}
1356 
1357 	ret = validate_range(mm, uffdio_register.range.start,
1358 			     uffdio_register.range.len);
1359 	if (ret)
1360 		goto out;
1361 
1362 	start = uffdio_register.range.start;
1363 	end = start + uffdio_register.range.len;
1364 
1365 	ret = -ENOMEM;
1366 	if (!mmget_not_zero(mm))
1367 		goto out;
1368 
1369 	ret = -EINVAL;
1370 	mmap_write_lock(mm);
1371 	vma_iter_init(&vmi, mm, start);
1372 	vma = vma_find(&vmi, end);
1373 	if (!vma)
1374 		goto out_unlock;
1375 
1376 	/*
1377 	 * If the first vma contains huge pages, make sure start address
1378 	 * is aligned to huge page size.
1379 	 */
1380 	if (is_vm_hugetlb_page(vma)) {
1381 		unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1382 
1383 		if (start & (vma_hpagesize - 1))
1384 			goto out_unlock;
1385 	}
1386 
1387 	/*
1388 	 * Search for not compatible vmas.
1389 	 */
1390 	found = false;
1391 	basic_ioctls = false;
1392 	cur = vma;
1393 	do {
1394 		cond_resched();
1395 
1396 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1397 		       !!(cur->vm_flags & __VM_UFFD_FLAGS));
1398 
1399 		/* check not compatible vmas */
1400 		ret = -EINVAL;
1401 		if (!vma_can_userfault(cur, vm_flags, wp_async))
1402 			goto out_unlock;
1403 
1404 		/*
1405 		 * UFFDIO_COPY will fill file holes even without
1406 		 * PROT_WRITE. This check enforces that if this is a
1407 		 * MAP_SHARED, the process has write permission to the backing
1408 		 * file. If VM_MAYWRITE is set it also enforces that on a
1409 		 * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1410 		 * F_WRITE_SEAL can be taken until the vma is destroyed.
1411 		 */
1412 		ret = -EPERM;
1413 		if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1414 			goto out_unlock;
1415 
1416 		/*
1417 		 * If this vma contains ending address, and huge pages
1418 		 * check alignment.
1419 		 */
1420 		if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1421 		    end > cur->vm_start) {
1422 			unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1423 
1424 			ret = -EINVAL;
1425 
1426 			if (end & (vma_hpagesize - 1))
1427 				goto out_unlock;
1428 		}
1429 		if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1430 			goto out_unlock;
1431 
1432 		/*
1433 		 * Check that this vma isn't already owned by a
1434 		 * different userfaultfd. We can't allow more than one
1435 		 * userfaultfd to own a single vma simultaneously or we
1436 		 * wouldn't know which one to deliver the userfaults to.
1437 		 */
1438 		ret = -EBUSY;
1439 		if (cur->vm_userfaultfd_ctx.ctx &&
1440 		    cur->vm_userfaultfd_ctx.ctx != ctx)
1441 			goto out_unlock;
1442 
1443 		/*
1444 		 * Note vmas containing huge pages
1445 		 */
1446 		if (is_vm_hugetlb_page(cur))
1447 			basic_ioctls = true;
1448 
1449 		found = true;
1450 	} for_each_vma_range(vmi, cur, end);
1451 	BUG_ON(!found);
1452 
1453 	vma_iter_set(&vmi, start);
1454 	prev = vma_prev(&vmi);
1455 	if (vma->vm_start < start)
1456 		prev = vma;
1457 
1458 	ret = 0;
1459 	for_each_vma_range(vmi, vma, end) {
1460 		cond_resched();
1461 
1462 		BUG_ON(!vma_can_userfault(vma, vm_flags, wp_async));
1463 		BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1464 		       vma->vm_userfaultfd_ctx.ctx != ctx);
1465 		WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1466 
1467 		/*
1468 		 * Nothing to do: this vma is already registered into this
1469 		 * userfaultfd and with the right tracking mode too.
1470 		 */
1471 		if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1472 		    (vma->vm_flags & vm_flags) == vm_flags)
1473 			goto skip;
1474 
1475 		if (vma->vm_start > start)
1476 			start = vma->vm_start;
1477 		vma_end = min(end, vma->vm_end);
1478 
1479 		new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1480 		vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1481 					    new_flags,
1482 					    (struct vm_userfaultfd_ctx){ctx});
1483 		if (IS_ERR(vma)) {
1484 			ret = PTR_ERR(vma);
1485 			break;
1486 		}
1487 
1488 		/*
1489 		 * In the vma_merge() successful mprotect-like case 8:
1490 		 * the next vma was merged into the current one and
1491 		 * the current one has not been updated yet.
1492 		 */
1493 		vma_start_write(vma);
1494 		userfaultfd_set_vm_flags(vma, new_flags);
1495 		vma->vm_userfaultfd_ctx.ctx = ctx;
1496 
1497 		if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1498 			hugetlb_unshare_all_pmds(vma);
1499 
1500 	skip:
1501 		prev = vma;
1502 		start = vma->vm_end;
1503 	}
1504 
1505 out_unlock:
1506 	mmap_write_unlock(mm);
1507 	mmput(mm);
1508 	if (!ret) {
1509 		__u64 ioctls_out;
1510 
1511 		ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1512 		    UFFD_API_RANGE_IOCTLS;
1513 
1514 		/*
1515 		 * Declare the WP ioctl only if the WP mode is
1516 		 * specified and all checks passed with the range
1517 		 */
1518 		if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1519 			ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1520 
1521 		/* CONTINUE ioctl is only supported for MINOR ranges. */
1522 		if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1523 			ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1524 
1525 		/*
1526 		 * Now that we scanned all vmas we can already tell
1527 		 * userland which ioctls methods are guaranteed to
1528 		 * succeed on this range.
1529 		 */
1530 		if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1531 			ret = -EFAULT;
1532 	}
1533 out:
1534 	return ret;
1535 }
1536 
1537 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1538 				  unsigned long arg)
1539 {
1540 	struct mm_struct *mm = ctx->mm;
1541 	struct vm_area_struct *vma, *prev, *cur;
1542 	int ret;
1543 	struct uffdio_range uffdio_unregister;
1544 	unsigned long new_flags;
1545 	bool found;
1546 	unsigned long start, end, vma_end;
1547 	const void __user *buf = (void __user *)arg;
1548 	struct vma_iterator vmi;
1549 	bool wp_async = userfaultfd_wp_async_ctx(ctx);
1550 
1551 	ret = -EFAULT;
1552 	if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1553 		goto out;
1554 
1555 	ret = validate_range(mm, uffdio_unregister.start,
1556 			     uffdio_unregister.len);
1557 	if (ret)
1558 		goto out;
1559 
1560 	start = uffdio_unregister.start;
1561 	end = start + uffdio_unregister.len;
1562 
1563 	ret = -ENOMEM;
1564 	if (!mmget_not_zero(mm))
1565 		goto out;
1566 
1567 	mmap_write_lock(mm);
1568 	ret = -EINVAL;
1569 	vma_iter_init(&vmi, mm, start);
1570 	vma = vma_find(&vmi, end);
1571 	if (!vma)
1572 		goto out_unlock;
1573 
1574 	/*
1575 	 * If the first vma contains huge pages, make sure start address
1576 	 * is aligned to huge page size.
1577 	 */
1578 	if (is_vm_hugetlb_page(vma)) {
1579 		unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1580 
1581 		if (start & (vma_hpagesize - 1))
1582 			goto out_unlock;
1583 	}
1584 
1585 	/*
1586 	 * Search for not compatible vmas.
1587 	 */
1588 	found = false;
1589 	cur = vma;
1590 	do {
1591 		cond_resched();
1592 
1593 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1594 		       !!(cur->vm_flags & __VM_UFFD_FLAGS));
1595 
1596 		/*
1597 		 * Check not compatible vmas, not strictly required
1598 		 * here as not compatible vmas cannot have an
1599 		 * userfaultfd_ctx registered on them, but this
1600 		 * provides for more strict behavior to notice
1601 		 * unregistration errors.
1602 		 */
1603 		if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
1604 			goto out_unlock;
1605 
1606 		found = true;
1607 	} for_each_vma_range(vmi, cur, end);
1608 	BUG_ON(!found);
1609 
1610 	vma_iter_set(&vmi, start);
1611 	prev = vma_prev(&vmi);
1612 	if (vma->vm_start < start)
1613 		prev = vma;
1614 
1615 	ret = 0;
1616 	for_each_vma_range(vmi, vma, end) {
1617 		cond_resched();
1618 
1619 		BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
1620 
1621 		/*
1622 		 * Nothing to do: this vma is already registered into this
1623 		 * userfaultfd and with the right tracking mode too.
1624 		 */
1625 		if (!vma->vm_userfaultfd_ctx.ctx)
1626 			goto skip;
1627 
1628 		WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1629 
1630 		if (vma->vm_start > start)
1631 			start = vma->vm_start;
1632 		vma_end = min(end, vma->vm_end);
1633 
1634 		if (userfaultfd_missing(vma)) {
1635 			/*
1636 			 * Wake any concurrent pending userfault while
1637 			 * we unregister, so they will not hang
1638 			 * permanently and it avoids userland to call
1639 			 * UFFDIO_WAKE explicitly.
1640 			 */
1641 			struct userfaultfd_wake_range range;
1642 			range.start = start;
1643 			range.len = vma_end - start;
1644 			wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1645 		}
1646 
1647 		/* Reset ptes for the whole vma range if wr-protected */
1648 		if (userfaultfd_wp(vma))
1649 			uffd_wp_range(vma, start, vma_end - start, false);
1650 
1651 		new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1652 		vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1653 					    new_flags, NULL_VM_UFFD_CTX);
1654 		if (IS_ERR(vma)) {
1655 			ret = PTR_ERR(vma);
1656 			break;
1657 		}
1658 
1659 		/*
1660 		 * In the vma_merge() successful mprotect-like case 8:
1661 		 * the next vma was merged into the current one and
1662 		 * the current one has not been updated yet.
1663 		 */
1664 		vma_start_write(vma);
1665 		userfaultfd_set_vm_flags(vma, new_flags);
1666 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1667 
1668 	skip:
1669 		prev = vma;
1670 		start = vma->vm_end;
1671 	}
1672 
1673 out_unlock:
1674 	mmap_write_unlock(mm);
1675 	mmput(mm);
1676 out:
1677 	return ret;
1678 }
1679 
1680 /*
1681  * userfaultfd_wake may be used in combination with the
1682  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1683  */
1684 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1685 			    unsigned long arg)
1686 {
1687 	int ret;
1688 	struct uffdio_range uffdio_wake;
1689 	struct userfaultfd_wake_range range;
1690 	const void __user *buf = (void __user *)arg;
1691 
1692 	ret = -EFAULT;
1693 	if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1694 		goto out;
1695 
1696 	ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1697 	if (ret)
1698 		goto out;
1699 
1700 	range.start = uffdio_wake.start;
1701 	range.len = uffdio_wake.len;
1702 
1703 	/*
1704 	 * len == 0 means wake all and we don't want to wake all here,
1705 	 * so check it again to be sure.
1706 	 */
1707 	VM_BUG_ON(!range.len);
1708 
1709 	wake_userfault(ctx, &range);
1710 	ret = 0;
1711 
1712 out:
1713 	return ret;
1714 }
1715 
1716 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1717 			    unsigned long arg)
1718 {
1719 	__s64 ret;
1720 	struct uffdio_copy uffdio_copy;
1721 	struct uffdio_copy __user *user_uffdio_copy;
1722 	struct userfaultfd_wake_range range;
1723 	uffd_flags_t flags = 0;
1724 
1725 	user_uffdio_copy = (struct uffdio_copy __user *) arg;
1726 
1727 	ret = -EAGAIN;
1728 	if (atomic_read(&ctx->mmap_changing))
1729 		goto out;
1730 
1731 	ret = -EFAULT;
1732 	if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1733 			   /* don't copy "copy" last field */
1734 			   sizeof(uffdio_copy)-sizeof(__s64)))
1735 		goto out;
1736 
1737 	ret = validate_unaligned_range(ctx->mm, uffdio_copy.src,
1738 				       uffdio_copy.len);
1739 	if (ret)
1740 		goto out;
1741 	ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1742 	if (ret)
1743 		goto out;
1744 
1745 	ret = -EINVAL;
1746 	if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1747 		goto out;
1748 	if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1749 		flags |= MFILL_ATOMIC_WP;
1750 	if (mmget_not_zero(ctx->mm)) {
1751 		ret = mfill_atomic_copy(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1752 					uffdio_copy.len, &ctx->mmap_changing,
1753 					flags);
1754 		mmput(ctx->mm);
1755 	} else {
1756 		return -ESRCH;
1757 	}
1758 	if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1759 		return -EFAULT;
1760 	if (ret < 0)
1761 		goto out;
1762 	BUG_ON(!ret);
1763 	/* len == 0 would wake all */
1764 	range.len = ret;
1765 	if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1766 		range.start = uffdio_copy.dst;
1767 		wake_userfault(ctx, &range);
1768 	}
1769 	ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1770 out:
1771 	return ret;
1772 }
1773 
1774 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1775 				unsigned long arg)
1776 {
1777 	__s64 ret;
1778 	struct uffdio_zeropage uffdio_zeropage;
1779 	struct uffdio_zeropage __user *user_uffdio_zeropage;
1780 	struct userfaultfd_wake_range range;
1781 
1782 	user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1783 
1784 	ret = -EAGAIN;
1785 	if (atomic_read(&ctx->mmap_changing))
1786 		goto out;
1787 
1788 	ret = -EFAULT;
1789 	if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1790 			   /* don't copy "zeropage" last field */
1791 			   sizeof(uffdio_zeropage)-sizeof(__s64)))
1792 		goto out;
1793 
1794 	ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1795 			     uffdio_zeropage.range.len);
1796 	if (ret)
1797 		goto out;
1798 	ret = -EINVAL;
1799 	if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1800 		goto out;
1801 
1802 	if (mmget_not_zero(ctx->mm)) {
1803 		ret = mfill_atomic_zeropage(ctx->mm, uffdio_zeropage.range.start,
1804 					   uffdio_zeropage.range.len,
1805 					   &ctx->mmap_changing);
1806 		mmput(ctx->mm);
1807 	} else {
1808 		return -ESRCH;
1809 	}
1810 	if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1811 		return -EFAULT;
1812 	if (ret < 0)
1813 		goto out;
1814 	/* len == 0 would wake all */
1815 	BUG_ON(!ret);
1816 	range.len = ret;
1817 	if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1818 		range.start = uffdio_zeropage.range.start;
1819 		wake_userfault(ctx, &range);
1820 	}
1821 	ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1822 out:
1823 	return ret;
1824 }
1825 
1826 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1827 				    unsigned long arg)
1828 {
1829 	int ret;
1830 	struct uffdio_writeprotect uffdio_wp;
1831 	struct uffdio_writeprotect __user *user_uffdio_wp;
1832 	struct userfaultfd_wake_range range;
1833 	bool mode_wp, mode_dontwake;
1834 
1835 	if (atomic_read(&ctx->mmap_changing))
1836 		return -EAGAIN;
1837 
1838 	user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1839 
1840 	if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1841 			   sizeof(struct uffdio_writeprotect)))
1842 		return -EFAULT;
1843 
1844 	ret = validate_range(ctx->mm, uffdio_wp.range.start,
1845 			     uffdio_wp.range.len);
1846 	if (ret)
1847 		return ret;
1848 
1849 	if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1850 			       UFFDIO_WRITEPROTECT_MODE_WP))
1851 		return -EINVAL;
1852 
1853 	mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1854 	mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1855 
1856 	if (mode_wp && mode_dontwake)
1857 		return -EINVAL;
1858 
1859 	if (mmget_not_zero(ctx->mm)) {
1860 		ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1861 					  uffdio_wp.range.len, mode_wp,
1862 					  &ctx->mmap_changing);
1863 		mmput(ctx->mm);
1864 	} else {
1865 		return -ESRCH;
1866 	}
1867 
1868 	if (ret)
1869 		return ret;
1870 
1871 	if (!mode_wp && !mode_dontwake) {
1872 		range.start = uffdio_wp.range.start;
1873 		range.len = uffdio_wp.range.len;
1874 		wake_userfault(ctx, &range);
1875 	}
1876 	return ret;
1877 }
1878 
1879 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1880 {
1881 	__s64 ret;
1882 	struct uffdio_continue uffdio_continue;
1883 	struct uffdio_continue __user *user_uffdio_continue;
1884 	struct userfaultfd_wake_range range;
1885 	uffd_flags_t flags = 0;
1886 
1887 	user_uffdio_continue = (struct uffdio_continue __user *)arg;
1888 
1889 	ret = -EAGAIN;
1890 	if (atomic_read(&ctx->mmap_changing))
1891 		goto out;
1892 
1893 	ret = -EFAULT;
1894 	if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1895 			   /* don't copy the output fields */
1896 			   sizeof(uffdio_continue) - (sizeof(__s64))))
1897 		goto out;
1898 
1899 	ret = validate_range(ctx->mm, uffdio_continue.range.start,
1900 			     uffdio_continue.range.len);
1901 	if (ret)
1902 		goto out;
1903 
1904 	ret = -EINVAL;
1905 	if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1906 				     UFFDIO_CONTINUE_MODE_WP))
1907 		goto out;
1908 	if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1909 		flags |= MFILL_ATOMIC_WP;
1910 
1911 	if (mmget_not_zero(ctx->mm)) {
1912 		ret = mfill_atomic_continue(ctx->mm, uffdio_continue.range.start,
1913 					    uffdio_continue.range.len,
1914 					    &ctx->mmap_changing, flags);
1915 		mmput(ctx->mm);
1916 	} else {
1917 		return -ESRCH;
1918 	}
1919 
1920 	if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1921 		return -EFAULT;
1922 	if (ret < 0)
1923 		goto out;
1924 
1925 	/* len == 0 would wake all */
1926 	BUG_ON(!ret);
1927 	range.len = ret;
1928 	if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1929 		range.start = uffdio_continue.range.start;
1930 		wake_userfault(ctx, &range);
1931 	}
1932 	ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1933 
1934 out:
1935 	return ret;
1936 }
1937 
1938 static inline int userfaultfd_poison(struct userfaultfd_ctx *ctx, unsigned long arg)
1939 {
1940 	__s64 ret;
1941 	struct uffdio_poison uffdio_poison;
1942 	struct uffdio_poison __user *user_uffdio_poison;
1943 	struct userfaultfd_wake_range range;
1944 
1945 	user_uffdio_poison = (struct uffdio_poison __user *)arg;
1946 
1947 	ret = -EAGAIN;
1948 	if (atomic_read(&ctx->mmap_changing))
1949 		goto out;
1950 
1951 	ret = -EFAULT;
1952 	if (copy_from_user(&uffdio_poison, user_uffdio_poison,
1953 			   /* don't copy the output fields */
1954 			   sizeof(uffdio_poison) - (sizeof(__s64))))
1955 		goto out;
1956 
1957 	ret = validate_range(ctx->mm, uffdio_poison.range.start,
1958 			     uffdio_poison.range.len);
1959 	if (ret)
1960 		goto out;
1961 
1962 	ret = -EINVAL;
1963 	if (uffdio_poison.mode & ~UFFDIO_POISON_MODE_DONTWAKE)
1964 		goto out;
1965 
1966 	if (mmget_not_zero(ctx->mm)) {
1967 		ret = mfill_atomic_poison(ctx->mm, uffdio_poison.range.start,
1968 					  uffdio_poison.range.len,
1969 					  &ctx->mmap_changing, 0);
1970 		mmput(ctx->mm);
1971 	} else {
1972 		return -ESRCH;
1973 	}
1974 
1975 	if (unlikely(put_user(ret, &user_uffdio_poison->updated)))
1976 		return -EFAULT;
1977 	if (ret < 0)
1978 		goto out;
1979 
1980 	/* len == 0 would wake all */
1981 	BUG_ON(!ret);
1982 	range.len = ret;
1983 	if (!(uffdio_poison.mode & UFFDIO_POISON_MODE_DONTWAKE)) {
1984 		range.start = uffdio_poison.range.start;
1985 		wake_userfault(ctx, &range);
1986 	}
1987 	ret = range.len == uffdio_poison.range.len ? 0 : -EAGAIN;
1988 
1989 out:
1990 	return ret;
1991 }
1992 
1993 bool userfaultfd_wp_async(struct vm_area_struct *vma)
1994 {
1995 	return userfaultfd_wp_async_ctx(vma->vm_userfaultfd_ctx.ctx);
1996 }
1997 
1998 static inline unsigned int uffd_ctx_features(__u64 user_features)
1999 {
2000 	/*
2001 	 * For the current set of features the bits just coincide. Set
2002 	 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
2003 	 */
2004 	return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
2005 }
2006 
2007 static int userfaultfd_move(struct userfaultfd_ctx *ctx,
2008 			    unsigned long arg)
2009 {
2010 	__s64 ret;
2011 	struct uffdio_move uffdio_move;
2012 	struct uffdio_move __user *user_uffdio_move;
2013 	struct userfaultfd_wake_range range;
2014 	struct mm_struct *mm = ctx->mm;
2015 
2016 	user_uffdio_move = (struct uffdio_move __user *) arg;
2017 
2018 	if (atomic_read(&ctx->mmap_changing))
2019 		return -EAGAIN;
2020 
2021 	if (copy_from_user(&uffdio_move, user_uffdio_move,
2022 			   /* don't copy "move" last field */
2023 			   sizeof(uffdio_move)-sizeof(__s64)))
2024 		return -EFAULT;
2025 
2026 	/* Do not allow cross-mm moves. */
2027 	if (mm != current->mm)
2028 		return -EINVAL;
2029 
2030 	ret = validate_range(mm, uffdio_move.dst, uffdio_move.len);
2031 	if (ret)
2032 		return ret;
2033 
2034 	ret = validate_range(mm, uffdio_move.src, uffdio_move.len);
2035 	if (ret)
2036 		return ret;
2037 
2038 	if (uffdio_move.mode & ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES|
2039 				  UFFDIO_MOVE_MODE_DONTWAKE))
2040 		return -EINVAL;
2041 
2042 	if (mmget_not_zero(mm)) {
2043 		mmap_read_lock(mm);
2044 
2045 		/* Re-check after taking mmap_lock */
2046 		if (likely(!atomic_read(&ctx->mmap_changing)))
2047 			ret = move_pages(ctx, mm, uffdio_move.dst, uffdio_move.src,
2048 					 uffdio_move.len, uffdio_move.mode);
2049 		else
2050 			ret = -EINVAL;
2051 
2052 		mmap_read_unlock(mm);
2053 		mmput(mm);
2054 	} else {
2055 		return -ESRCH;
2056 	}
2057 
2058 	if (unlikely(put_user(ret, &user_uffdio_move->move)))
2059 		return -EFAULT;
2060 	if (ret < 0)
2061 		goto out;
2062 
2063 	/* len == 0 would wake all */
2064 	VM_WARN_ON(!ret);
2065 	range.len = ret;
2066 	if (!(uffdio_move.mode & UFFDIO_MOVE_MODE_DONTWAKE)) {
2067 		range.start = uffdio_move.dst;
2068 		wake_userfault(ctx, &range);
2069 	}
2070 	ret = range.len == uffdio_move.len ? 0 : -EAGAIN;
2071 
2072 out:
2073 	return ret;
2074 }
2075 
2076 /*
2077  * userland asks for a certain API version and we return which bits
2078  * and ioctl commands are implemented in this kernel for such API
2079  * version or -EINVAL if unknown.
2080  */
2081 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
2082 			   unsigned long arg)
2083 {
2084 	struct uffdio_api uffdio_api;
2085 	void __user *buf = (void __user *)arg;
2086 	unsigned int ctx_features;
2087 	int ret;
2088 	__u64 features;
2089 
2090 	ret = -EFAULT;
2091 	if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
2092 		goto out;
2093 	features = uffdio_api.features;
2094 	ret = -EINVAL;
2095 	if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
2096 		goto err_out;
2097 	ret = -EPERM;
2098 	if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
2099 		goto err_out;
2100 
2101 	/* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2102 	if (features & UFFD_FEATURE_WP_ASYNC)
2103 		features |= UFFD_FEATURE_WP_UNPOPULATED;
2104 
2105 	/* report all available features and ioctls to userland */
2106 	uffdio_api.features = UFFD_API_FEATURES;
2107 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2108 	uffdio_api.features &=
2109 		~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
2110 #endif
2111 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2112 	uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
2113 #endif
2114 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2115 	uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
2116 	uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
2117 	uffdio_api.features &= ~UFFD_FEATURE_WP_ASYNC;
2118 #endif
2119 	uffdio_api.ioctls = UFFD_API_IOCTLS;
2120 	ret = -EFAULT;
2121 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2122 		goto out;
2123 
2124 	/* only enable the requested features for this uffd context */
2125 	ctx_features = uffd_ctx_features(features);
2126 	ret = -EINVAL;
2127 	if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
2128 		goto err_out;
2129 
2130 	ret = 0;
2131 out:
2132 	return ret;
2133 err_out:
2134 	memset(&uffdio_api, 0, sizeof(uffdio_api));
2135 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2136 		ret = -EFAULT;
2137 	goto out;
2138 }
2139 
2140 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2141 			      unsigned long arg)
2142 {
2143 	int ret = -EINVAL;
2144 	struct userfaultfd_ctx *ctx = file->private_data;
2145 
2146 	if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2147 		return -EINVAL;
2148 
2149 	switch(cmd) {
2150 	case UFFDIO_API:
2151 		ret = userfaultfd_api(ctx, arg);
2152 		break;
2153 	case UFFDIO_REGISTER:
2154 		ret = userfaultfd_register(ctx, arg);
2155 		break;
2156 	case UFFDIO_UNREGISTER:
2157 		ret = userfaultfd_unregister(ctx, arg);
2158 		break;
2159 	case UFFDIO_WAKE:
2160 		ret = userfaultfd_wake(ctx, arg);
2161 		break;
2162 	case UFFDIO_COPY:
2163 		ret = userfaultfd_copy(ctx, arg);
2164 		break;
2165 	case UFFDIO_ZEROPAGE:
2166 		ret = userfaultfd_zeropage(ctx, arg);
2167 		break;
2168 	case UFFDIO_MOVE:
2169 		ret = userfaultfd_move(ctx, arg);
2170 		break;
2171 	case UFFDIO_WRITEPROTECT:
2172 		ret = userfaultfd_writeprotect(ctx, arg);
2173 		break;
2174 	case UFFDIO_CONTINUE:
2175 		ret = userfaultfd_continue(ctx, arg);
2176 		break;
2177 	case UFFDIO_POISON:
2178 		ret = userfaultfd_poison(ctx, arg);
2179 		break;
2180 	}
2181 	return ret;
2182 }
2183 
2184 #ifdef CONFIG_PROC_FS
2185 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2186 {
2187 	struct userfaultfd_ctx *ctx = f->private_data;
2188 	wait_queue_entry_t *wq;
2189 	unsigned long pending = 0, total = 0;
2190 
2191 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
2192 	list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2193 		pending++;
2194 		total++;
2195 	}
2196 	list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2197 		total++;
2198 	}
2199 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2200 
2201 	/*
2202 	 * If more protocols will be added, there will be all shown
2203 	 * separated by a space. Like this:
2204 	 *	protocols: aa:... bb:...
2205 	 */
2206 	seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2207 		   pending, total, UFFD_API, ctx->features,
2208 		   UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2209 }
2210 #endif
2211 
2212 static const struct file_operations userfaultfd_fops = {
2213 #ifdef CONFIG_PROC_FS
2214 	.show_fdinfo	= userfaultfd_show_fdinfo,
2215 #endif
2216 	.release	= userfaultfd_release,
2217 	.poll		= userfaultfd_poll,
2218 	.read		= userfaultfd_read,
2219 	.unlocked_ioctl = userfaultfd_ioctl,
2220 	.compat_ioctl	= compat_ptr_ioctl,
2221 	.llseek		= noop_llseek,
2222 };
2223 
2224 static void init_once_userfaultfd_ctx(void *mem)
2225 {
2226 	struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2227 
2228 	init_waitqueue_head(&ctx->fault_pending_wqh);
2229 	init_waitqueue_head(&ctx->fault_wqh);
2230 	init_waitqueue_head(&ctx->event_wqh);
2231 	init_waitqueue_head(&ctx->fd_wqh);
2232 	seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2233 }
2234 
2235 static int new_userfaultfd(int flags)
2236 {
2237 	struct userfaultfd_ctx *ctx;
2238 	int fd;
2239 
2240 	BUG_ON(!current->mm);
2241 
2242 	/* Check the UFFD_* constants for consistency.  */
2243 	BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2244 	BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2245 	BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2246 
2247 	if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2248 		return -EINVAL;
2249 
2250 	ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2251 	if (!ctx)
2252 		return -ENOMEM;
2253 
2254 	refcount_set(&ctx->refcount, 1);
2255 	ctx->flags = flags;
2256 	ctx->features = 0;
2257 	ctx->released = false;
2258 	atomic_set(&ctx->mmap_changing, 0);
2259 	ctx->mm = current->mm;
2260 	/* prevent the mm struct to be freed */
2261 	mmgrab(ctx->mm);
2262 
2263 	fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2264 			O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2265 	if (fd < 0) {
2266 		mmdrop(ctx->mm);
2267 		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2268 	}
2269 	return fd;
2270 }
2271 
2272 static inline bool userfaultfd_syscall_allowed(int flags)
2273 {
2274 	/* Userspace-only page faults are always allowed */
2275 	if (flags & UFFD_USER_MODE_ONLY)
2276 		return true;
2277 
2278 	/*
2279 	 * The user is requesting a userfaultfd which can handle kernel faults.
2280 	 * Privileged users are always allowed to do this.
2281 	 */
2282 	if (capable(CAP_SYS_PTRACE))
2283 		return true;
2284 
2285 	/* Otherwise, access to kernel fault handling is sysctl controlled. */
2286 	return sysctl_unprivileged_userfaultfd;
2287 }
2288 
2289 SYSCALL_DEFINE1(userfaultfd, int, flags)
2290 {
2291 	if (!userfaultfd_syscall_allowed(flags))
2292 		return -EPERM;
2293 
2294 	return new_userfaultfd(flags);
2295 }
2296 
2297 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2298 {
2299 	if (cmd != USERFAULTFD_IOC_NEW)
2300 		return -EINVAL;
2301 
2302 	return new_userfaultfd(flags);
2303 }
2304 
2305 static const struct file_operations userfaultfd_dev_fops = {
2306 	.unlocked_ioctl = userfaultfd_dev_ioctl,
2307 	.compat_ioctl = userfaultfd_dev_ioctl,
2308 	.owner = THIS_MODULE,
2309 	.llseek = noop_llseek,
2310 };
2311 
2312 static struct miscdevice userfaultfd_misc = {
2313 	.minor = MISC_DYNAMIC_MINOR,
2314 	.name = "userfaultfd",
2315 	.fops = &userfaultfd_dev_fops
2316 };
2317 
2318 static int __init userfaultfd_init(void)
2319 {
2320 	int ret;
2321 
2322 	ret = misc_register(&userfaultfd_misc);
2323 	if (ret)
2324 		return ret;
2325 
2326 	userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2327 						sizeof(struct userfaultfd_ctx),
2328 						0,
2329 						SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2330 						init_once_userfaultfd_ctx);
2331 #ifdef CONFIG_SYSCTL
2332 	register_sysctl_init("vm", vm_userfaultfd_table);
2333 #endif
2334 	return 0;
2335 }
2336 __initcall(userfaultfd_init);
2337