userfaultfd: clear flag if remap event not enabled
[sfrench/cifs-2.6.git] / fs / userfaultfd.c
1 /*
2  *  fs/userfaultfd.c
3  *
4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
5  *  Copyright (C) 2008-2009 Red Hat, Inc.
6  *  Copyright (C) 2015  Red Hat, Inc.
7  *
8  *  This work is licensed under the terms of the GNU GPL, version 2. See
9  *  the COPYING file in the top-level directory.
10  *
11  *  Some part derived from fs/eventfd.c (anon inode setup) and
12  *  mm/ksm.c (mm hashing).
13  */
14
15 #include <linux/list.h>
16 #include <linux/hashtable.h>
17 #include <linux/sched/signal.h>
18 #include <linux/sched/mm.h>
19 #include <linux/mm.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
33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
34
35 enum userfaultfd_state {
36         UFFD_STATE_WAIT_API,
37         UFFD_STATE_RUNNING,
38 };
39
40 /*
41  * Start with fault_pending_wqh and fault_wqh so they're more likely
42  * to be in the same cacheline.
43  */
44 struct userfaultfd_ctx {
45         /* waitqueue head for the pending (i.e. not read) userfaults */
46         wait_queue_head_t fault_pending_wqh;
47         /* waitqueue head for the userfaults */
48         wait_queue_head_t fault_wqh;
49         /* waitqueue head for the pseudo fd to wakeup poll/read */
50         wait_queue_head_t fd_wqh;
51         /* waitqueue head for events */
52         wait_queue_head_t event_wqh;
53         /* a refile sequence protected by fault_pending_wqh lock */
54         struct seqcount refile_seq;
55         /* pseudo fd refcounting */
56         refcount_t refcount;
57         /* userfaultfd syscall flags */
58         unsigned int flags;
59         /* features requested from the userspace */
60         unsigned int features;
61         /* state machine */
62         enum userfaultfd_state state;
63         /* released */
64         bool released;
65         /* memory mappings are changing because of non-cooperative event */
66         bool mmap_changing;
67         /* mm with one ore more vmas attached to this userfaultfd_ctx */
68         struct mm_struct *mm;
69 };
70
71 struct userfaultfd_fork_ctx {
72         struct userfaultfd_ctx *orig;
73         struct userfaultfd_ctx *new;
74         struct list_head list;
75 };
76
77 struct userfaultfd_unmap_ctx {
78         struct userfaultfd_ctx *ctx;
79         unsigned long start;
80         unsigned long end;
81         struct list_head list;
82 };
83
84 struct userfaultfd_wait_queue {
85         struct uffd_msg msg;
86         wait_queue_entry_t wq;
87         struct userfaultfd_ctx *ctx;
88         bool waken;
89 };
90
91 struct userfaultfd_wake_range {
92         unsigned long start;
93         unsigned long len;
94 };
95
96 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
97                                      int wake_flags, void *key)
98 {
99         struct userfaultfd_wake_range *range = key;
100         int ret;
101         struct userfaultfd_wait_queue *uwq;
102         unsigned long start, len;
103
104         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
105         ret = 0;
106         /* len == 0 means wake all */
107         start = range->start;
108         len = range->len;
109         if (len && (start > uwq->msg.arg.pagefault.address ||
110                     start + len <= uwq->msg.arg.pagefault.address))
111                 goto out;
112         WRITE_ONCE(uwq->waken, true);
113         /*
114          * The Program-Order guarantees provided by the scheduler
115          * ensure uwq->waken is visible before the task is woken.
116          */
117         ret = wake_up_state(wq->private, mode);
118         if (ret) {
119                 /*
120                  * Wake only once, autoremove behavior.
121                  *
122                  * After the effect of list_del_init is visible to the other
123                  * CPUs, the waitqueue may disappear from under us, see the
124                  * !list_empty_careful() in handle_userfault().
125                  *
126                  * try_to_wake_up() has an implicit smp_mb(), and the
127                  * wq->private is read before calling the extern function
128                  * "wake_up_state" (which in turns calls try_to_wake_up).
129                  */
130                 list_del_init(&wq->entry);
131         }
132 out:
133         return ret;
134 }
135
136 /**
137  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
138  * context.
139  * @ctx: [in] Pointer to the userfaultfd context.
140  */
141 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
142 {
143         refcount_inc(&ctx->refcount);
144 }
145
146 /**
147  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
148  * context.
149  * @ctx: [in] Pointer to userfaultfd context.
150  *
151  * The userfaultfd context reference must have been previously acquired either
152  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
153  */
154 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
155 {
156         if (refcount_dec_and_test(&ctx->refcount)) {
157                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
158                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
159                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
160                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
161                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
162                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
163                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
164                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
165                 mmdrop(ctx->mm);
166                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
167         }
168 }
169
170 static inline void msg_init(struct uffd_msg *msg)
171 {
172         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
173         /*
174          * Must use memset to zero out the paddings or kernel data is
175          * leaked to userland.
176          */
177         memset(msg, 0, sizeof(struct uffd_msg));
178 }
179
180 static inline struct uffd_msg userfault_msg(unsigned long address,
181                                             unsigned int flags,
182                                             unsigned long reason,
183                                             unsigned int features)
184 {
185         struct uffd_msg msg;
186         msg_init(&msg);
187         msg.event = UFFD_EVENT_PAGEFAULT;
188         msg.arg.pagefault.address = address;
189         if (flags & FAULT_FLAG_WRITE)
190                 /*
191                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
192                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
193                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
194                  * was a read fault, otherwise if set it means it's
195                  * a write fault.
196                  */
197                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
198         if (reason & VM_UFFD_WP)
199                 /*
200                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
201                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
202                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
203                  * a missing fault, otherwise if set it means it's a
204                  * write protect fault.
205                  */
206                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
207         if (features & UFFD_FEATURE_THREAD_ID)
208                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
209         return msg;
210 }
211
212 #ifdef CONFIG_HUGETLB_PAGE
213 /*
214  * Same functionality as userfaultfd_must_wait below with modifications for
215  * hugepmd ranges.
216  */
217 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
218                                          struct vm_area_struct *vma,
219                                          unsigned long address,
220                                          unsigned long flags,
221                                          unsigned long reason)
222 {
223         struct mm_struct *mm = ctx->mm;
224         pte_t *ptep, pte;
225         bool ret = true;
226
227         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
228
229         ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
230
231         if (!ptep)
232                 goto out;
233
234         ret = false;
235         pte = huge_ptep_get(ptep);
236
237         /*
238          * Lockless access: we're in a wait_event so it's ok if it
239          * changes under us.
240          */
241         if (huge_pte_none(pte))
242                 ret = true;
243         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
244                 ret = true;
245 out:
246         return ret;
247 }
248 #else
249 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
250                                          struct vm_area_struct *vma,
251                                          unsigned long address,
252                                          unsigned long flags,
253                                          unsigned long reason)
254 {
255         return false;   /* should never get here */
256 }
257 #endif /* CONFIG_HUGETLB_PAGE */
258
259 /*
260  * Verify the pagetables are still not ok after having reigstered into
261  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
262  * userfault that has already been resolved, if userfaultfd_read and
263  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
264  * threads.
265  */
266 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
267                                          unsigned long address,
268                                          unsigned long flags,
269                                          unsigned long reason)
270 {
271         struct mm_struct *mm = ctx->mm;
272         pgd_t *pgd;
273         p4d_t *p4d;
274         pud_t *pud;
275         pmd_t *pmd, _pmd;
276         pte_t *pte;
277         bool ret = true;
278
279         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
280
281         pgd = pgd_offset(mm, address);
282         if (!pgd_present(*pgd))
283                 goto out;
284         p4d = p4d_offset(pgd, address);
285         if (!p4d_present(*p4d))
286                 goto out;
287         pud = pud_offset(p4d, address);
288         if (!pud_present(*pud))
289                 goto out;
290         pmd = pmd_offset(pud, address);
291         /*
292          * READ_ONCE must function as a barrier with narrower scope
293          * and it must be equivalent to:
294          *      _pmd = *pmd; barrier();
295          *
296          * This is to deal with the instability (as in
297          * pmd_trans_unstable) of the pmd.
298          */
299         _pmd = READ_ONCE(*pmd);
300         if (pmd_none(_pmd))
301                 goto out;
302
303         ret = false;
304         if (!pmd_present(_pmd))
305                 goto out;
306
307         if (pmd_trans_huge(_pmd))
308                 goto out;
309
310         /*
311          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
312          * and use the standard pte_offset_map() instead of parsing _pmd.
313          */
314         pte = pte_offset_map(pmd, address);
315         /*
316          * Lockless access: we're in a wait_event so it's ok if it
317          * changes under us.
318          */
319         if (pte_none(*pte))
320                 ret = true;
321         pte_unmap(pte);
322
323 out:
324         return ret;
325 }
326
327 /*
328  * The locking rules involved in returning VM_FAULT_RETRY depending on
329  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
330  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
331  * recommendation in __lock_page_or_retry is not an understatement.
332  *
333  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
334  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
335  * not set.
336  *
337  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
338  * set, VM_FAULT_RETRY can still be returned if and only if there are
339  * fatal_signal_pending()s, and the mmap_sem must be released before
340  * returning it.
341  */
342 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
343 {
344         struct mm_struct *mm = vmf->vma->vm_mm;
345         struct userfaultfd_ctx *ctx;
346         struct userfaultfd_wait_queue uwq;
347         vm_fault_t ret = VM_FAULT_SIGBUS;
348         bool must_wait, return_to_userland;
349         long blocking_state;
350
351         /*
352          * We don't do userfault handling for the final child pid update.
353          *
354          * We also don't do userfault handling during
355          * coredumping. hugetlbfs has the special
356          * follow_hugetlb_page() to skip missing pages in the
357          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
358          * the no_page_table() helper in follow_page_mask(), but the
359          * shmem_vm_ops->fault method is invoked even during
360          * coredumping without mmap_sem and it ends up here.
361          */
362         if (current->flags & (PF_EXITING|PF_DUMPCORE))
363                 goto out;
364
365         /*
366          * Coredumping runs without mmap_sem so we can only check that
367          * the mmap_sem is held, if PF_DUMPCORE was not set.
368          */
369         WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
370
371         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
372         if (!ctx)
373                 goto out;
374
375         BUG_ON(ctx->mm != mm);
376
377         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
378         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
379
380         if (ctx->features & UFFD_FEATURE_SIGBUS)
381                 goto out;
382
383         /*
384          * If it's already released don't get it. This avoids to loop
385          * in __get_user_pages if userfaultfd_release waits on the
386          * caller of handle_userfault to release the mmap_sem.
387          */
388         if (unlikely(READ_ONCE(ctx->released))) {
389                 /*
390                  * Don't return VM_FAULT_SIGBUS in this case, so a non
391                  * cooperative manager can close the uffd after the
392                  * last UFFDIO_COPY, without risking to trigger an
393                  * involuntary SIGBUS if the process was starting the
394                  * userfaultfd while the userfaultfd was still armed
395                  * (but after the last UFFDIO_COPY). If the uffd
396                  * wasn't already closed when the userfault reached
397                  * this point, that would normally be solved by
398                  * userfaultfd_must_wait returning 'false'.
399                  *
400                  * If we were to return VM_FAULT_SIGBUS here, the non
401                  * cooperative manager would be instead forced to
402                  * always call UFFDIO_UNREGISTER before it can safely
403                  * close the uffd.
404                  */
405                 ret = VM_FAULT_NOPAGE;
406                 goto out;
407         }
408
409         /*
410          * Check that we can return VM_FAULT_RETRY.
411          *
412          * NOTE: it should become possible to return VM_FAULT_RETRY
413          * even if FAULT_FLAG_TRIED is set without leading to gup()
414          * -EBUSY failures, if the userfaultfd is to be extended for
415          * VM_UFFD_WP tracking and we intend to arm the userfault
416          * without first stopping userland access to the memory. For
417          * VM_UFFD_MISSING userfaults this is enough for now.
418          */
419         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
420                 /*
421                  * Validate the invariant that nowait must allow retry
422                  * to be sure not to return SIGBUS erroneously on
423                  * nowait invocations.
424                  */
425                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
426 #ifdef CONFIG_DEBUG_VM
427                 if (printk_ratelimit()) {
428                         printk(KERN_WARNING
429                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
430                                vmf->flags);
431                         dump_stack();
432                 }
433 #endif
434                 goto out;
435         }
436
437         /*
438          * Handle nowait, not much to do other than tell it to retry
439          * and wait.
440          */
441         ret = VM_FAULT_RETRY;
442         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
443                 goto out;
444
445         /* take the reference before dropping the mmap_sem */
446         userfaultfd_ctx_get(ctx);
447
448         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
449         uwq.wq.private = current;
450         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
451                         ctx->features);
452         uwq.ctx = ctx;
453         uwq.waken = false;
454
455         return_to_userland =
456                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
457                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
458         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
459                          TASK_KILLABLE;
460
461         spin_lock(&ctx->fault_pending_wqh.lock);
462         /*
463          * After the __add_wait_queue the uwq is visible to userland
464          * through poll/read().
465          */
466         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
467         /*
468          * The smp_mb() after __set_current_state prevents the reads
469          * following the spin_unlock to happen before the list_add in
470          * __add_wait_queue.
471          */
472         set_current_state(blocking_state);
473         spin_unlock(&ctx->fault_pending_wqh.lock);
474
475         if (!is_vm_hugetlb_page(vmf->vma))
476                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
477                                                   reason);
478         else
479                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
480                                                        vmf->address,
481                                                        vmf->flags, reason);
482         up_read(&mm->mmap_sem);
483
484         if (likely(must_wait && !READ_ONCE(ctx->released) &&
485                    (return_to_userland ? !signal_pending(current) :
486                     !fatal_signal_pending(current)))) {
487                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
488                 schedule();
489                 ret |= VM_FAULT_MAJOR;
490
491                 /*
492                  * False wakeups can orginate even from rwsem before
493                  * up_read() however userfaults will wait either for a
494                  * targeted wakeup on the specific uwq waitqueue from
495                  * wake_userfault() or for signals or for uffd
496                  * release.
497                  */
498                 while (!READ_ONCE(uwq.waken)) {
499                         /*
500                          * This needs the full smp_store_mb()
501                          * guarantee as the state write must be
502                          * visible to other CPUs before reading
503                          * uwq.waken from other CPUs.
504                          */
505                         set_current_state(blocking_state);
506                         if (READ_ONCE(uwq.waken) ||
507                             READ_ONCE(ctx->released) ||
508                             (return_to_userland ? signal_pending(current) :
509                              fatal_signal_pending(current)))
510                                 break;
511                         schedule();
512                 }
513         }
514
515         __set_current_state(TASK_RUNNING);
516
517         if (return_to_userland) {
518                 if (signal_pending(current) &&
519                     !fatal_signal_pending(current)) {
520                         /*
521                          * If we got a SIGSTOP or SIGCONT and this is
522                          * a normal userland page fault, just let
523                          * userland return so the signal will be
524                          * handled and gdb debugging works.  The page
525                          * fault code immediately after we return from
526                          * this function is going to release the
527                          * mmap_sem and it's not depending on it
528                          * (unlike gup would if we were not to return
529                          * VM_FAULT_RETRY).
530                          *
531                          * If a fatal signal is pending we still take
532                          * the streamlined VM_FAULT_RETRY failure path
533                          * and there's no need to retake the mmap_sem
534                          * in such case.
535                          */
536                         down_read(&mm->mmap_sem);
537                         ret = VM_FAULT_NOPAGE;
538                 }
539         }
540
541         /*
542          * Here we race with the list_del; list_add in
543          * userfaultfd_ctx_read(), however because we don't ever run
544          * list_del_init() to refile across the two lists, the prev
545          * and next pointers will never point to self. list_add also
546          * would never let any of the two pointers to point to
547          * self. So list_empty_careful won't risk to see both pointers
548          * pointing to self at any time during the list refile. The
549          * only case where list_del_init() is called is the full
550          * removal in the wake function and there we don't re-list_add
551          * and it's fine not to block on the spinlock. The uwq on this
552          * kernel stack can be released after the list_del_init.
553          */
554         if (!list_empty_careful(&uwq.wq.entry)) {
555                 spin_lock(&ctx->fault_pending_wqh.lock);
556                 /*
557                  * No need of list_del_init(), the uwq on the stack
558                  * will be freed shortly anyway.
559                  */
560                 list_del(&uwq.wq.entry);
561                 spin_unlock(&ctx->fault_pending_wqh.lock);
562         }
563
564         /*
565          * ctx may go away after this if the userfault pseudo fd is
566          * already released.
567          */
568         userfaultfd_ctx_put(ctx);
569
570 out:
571         return ret;
572 }
573
574 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
575                                               struct userfaultfd_wait_queue *ewq)
576 {
577         struct userfaultfd_ctx *release_new_ctx;
578
579         if (WARN_ON_ONCE(current->flags & PF_EXITING))
580                 goto out;
581
582         ewq->ctx = ctx;
583         init_waitqueue_entry(&ewq->wq, current);
584         release_new_ctx = NULL;
585
586         spin_lock(&ctx->event_wqh.lock);
587         /*
588          * After the __add_wait_queue the uwq is visible to userland
589          * through poll/read().
590          */
591         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
592         for (;;) {
593                 set_current_state(TASK_KILLABLE);
594                 if (ewq->msg.event == 0)
595                         break;
596                 if (READ_ONCE(ctx->released) ||
597                     fatal_signal_pending(current)) {
598                         /*
599                          * &ewq->wq may be queued in fork_event, but
600                          * __remove_wait_queue ignores the head
601                          * parameter. It would be a problem if it
602                          * didn't.
603                          */
604                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
605                         if (ewq->msg.event == UFFD_EVENT_FORK) {
606                                 struct userfaultfd_ctx *new;
607
608                                 new = (struct userfaultfd_ctx *)
609                                         (unsigned long)
610                                         ewq->msg.arg.reserved.reserved1;
611                                 release_new_ctx = new;
612                         }
613                         break;
614                 }
615
616                 spin_unlock(&ctx->event_wqh.lock);
617
618                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
619                 schedule();
620
621                 spin_lock(&ctx->event_wqh.lock);
622         }
623         __set_current_state(TASK_RUNNING);
624         spin_unlock(&ctx->event_wqh.lock);
625
626         if (release_new_ctx) {
627                 struct vm_area_struct *vma;
628                 struct mm_struct *mm = release_new_ctx->mm;
629
630                 /* the various vma->vm_userfaultfd_ctx still points to it */
631                 down_write(&mm->mmap_sem);
632                 for (vma = mm->mmap; vma; vma = vma->vm_next)
633                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
634                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
635                                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
636                         }
637                 up_write(&mm->mmap_sem);
638
639                 userfaultfd_ctx_put(release_new_ctx);
640         }
641
642         /*
643          * ctx may go away after this if the userfault pseudo fd is
644          * already released.
645          */
646 out:
647         WRITE_ONCE(ctx->mmap_changing, false);
648         userfaultfd_ctx_put(ctx);
649 }
650
651 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
652                                        struct userfaultfd_wait_queue *ewq)
653 {
654         ewq->msg.event = 0;
655         wake_up_locked(&ctx->event_wqh);
656         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
657 }
658
659 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
660 {
661         struct userfaultfd_ctx *ctx = NULL, *octx;
662         struct userfaultfd_fork_ctx *fctx;
663
664         octx = vma->vm_userfaultfd_ctx.ctx;
665         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
666                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
667                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
668                 return 0;
669         }
670
671         list_for_each_entry(fctx, fcs, list)
672                 if (fctx->orig == octx) {
673                         ctx = fctx->new;
674                         break;
675                 }
676
677         if (!ctx) {
678                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
679                 if (!fctx)
680                         return -ENOMEM;
681
682                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
683                 if (!ctx) {
684                         kfree(fctx);
685                         return -ENOMEM;
686                 }
687
688                 refcount_set(&ctx->refcount, 1);
689                 ctx->flags = octx->flags;
690                 ctx->state = UFFD_STATE_RUNNING;
691                 ctx->features = octx->features;
692                 ctx->released = false;
693                 ctx->mmap_changing = false;
694                 ctx->mm = vma->vm_mm;
695                 mmgrab(ctx->mm);
696
697                 userfaultfd_ctx_get(octx);
698                 WRITE_ONCE(octx->mmap_changing, true);
699                 fctx->orig = octx;
700                 fctx->new = ctx;
701                 list_add_tail(&fctx->list, fcs);
702         }
703
704         vma->vm_userfaultfd_ctx.ctx = ctx;
705         return 0;
706 }
707
708 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
709 {
710         struct userfaultfd_ctx *ctx = fctx->orig;
711         struct userfaultfd_wait_queue ewq;
712
713         msg_init(&ewq.msg);
714
715         ewq.msg.event = UFFD_EVENT_FORK;
716         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
717
718         userfaultfd_event_wait_completion(ctx, &ewq);
719 }
720
721 void dup_userfaultfd_complete(struct list_head *fcs)
722 {
723         struct userfaultfd_fork_ctx *fctx, *n;
724
725         list_for_each_entry_safe(fctx, n, fcs, list) {
726                 dup_fctx(fctx);
727                 list_del(&fctx->list);
728                 kfree(fctx);
729         }
730 }
731
732 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
733                              struct vm_userfaultfd_ctx *vm_ctx)
734 {
735         struct userfaultfd_ctx *ctx;
736
737         ctx = vma->vm_userfaultfd_ctx.ctx;
738
739         if (!ctx)
740                 return;
741
742         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
743                 vm_ctx->ctx = ctx;
744                 userfaultfd_ctx_get(ctx);
745                 WRITE_ONCE(ctx->mmap_changing, true);
746         } else {
747                 /* Drop uffd context if remap feature not enabled */
748                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
749                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
750         }
751 }
752
753 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
754                                  unsigned long from, unsigned long to,
755                                  unsigned long len)
756 {
757         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
758         struct userfaultfd_wait_queue ewq;
759
760         if (!ctx)
761                 return;
762
763         if (to & ~PAGE_MASK) {
764                 userfaultfd_ctx_put(ctx);
765                 return;
766         }
767
768         msg_init(&ewq.msg);
769
770         ewq.msg.event = UFFD_EVENT_REMAP;
771         ewq.msg.arg.remap.from = from;
772         ewq.msg.arg.remap.to = to;
773         ewq.msg.arg.remap.len = len;
774
775         userfaultfd_event_wait_completion(ctx, &ewq);
776 }
777
778 bool userfaultfd_remove(struct vm_area_struct *vma,
779                         unsigned long start, unsigned long end)
780 {
781         struct mm_struct *mm = vma->vm_mm;
782         struct userfaultfd_ctx *ctx;
783         struct userfaultfd_wait_queue ewq;
784
785         ctx = vma->vm_userfaultfd_ctx.ctx;
786         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
787                 return true;
788
789         userfaultfd_ctx_get(ctx);
790         WRITE_ONCE(ctx->mmap_changing, true);
791         up_read(&mm->mmap_sem);
792
793         msg_init(&ewq.msg);
794
795         ewq.msg.event = UFFD_EVENT_REMOVE;
796         ewq.msg.arg.remove.start = start;
797         ewq.msg.arg.remove.end = end;
798
799         userfaultfd_event_wait_completion(ctx, &ewq);
800
801         return false;
802 }
803
804 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
805                           unsigned long start, unsigned long end)
806 {
807         struct userfaultfd_unmap_ctx *unmap_ctx;
808
809         list_for_each_entry(unmap_ctx, unmaps, list)
810                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
811                     unmap_ctx->end == end)
812                         return true;
813
814         return false;
815 }
816
817 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
818                            unsigned long start, unsigned long end,
819                            struct list_head *unmaps)
820 {
821         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
822                 struct userfaultfd_unmap_ctx *unmap_ctx;
823                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
824
825                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
826                     has_unmap_ctx(ctx, unmaps, start, end))
827                         continue;
828
829                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
830                 if (!unmap_ctx)
831                         return -ENOMEM;
832
833                 userfaultfd_ctx_get(ctx);
834                 WRITE_ONCE(ctx->mmap_changing, true);
835                 unmap_ctx->ctx = ctx;
836                 unmap_ctx->start = start;
837                 unmap_ctx->end = end;
838                 list_add_tail(&unmap_ctx->list, unmaps);
839         }
840
841         return 0;
842 }
843
844 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
845 {
846         struct userfaultfd_unmap_ctx *ctx, *n;
847         struct userfaultfd_wait_queue ewq;
848
849         list_for_each_entry_safe(ctx, n, uf, list) {
850                 msg_init(&ewq.msg);
851
852                 ewq.msg.event = UFFD_EVENT_UNMAP;
853                 ewq.msg.arg.remove.start = ctx->start;
854                 ewq.msg.arg.remove.end = ctx->end;
855
856                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
857
858                 list_del(&ctx->list);
859                 kfree(ctx);
860         }
861 }
862
863 static int userfaultfd_release(struct inode *inode, struct file *file)
864 {
865         struct userfaultfd_ctx *ctx = file->private_data;
866         struct mm_struct *mm = ctx->mm;
867         struct vm_area_struct *vma, *prev;
868         /* len == 0 means wake all */
869         struct userfaultfd_wake_range range = { .len = 0, };
870         unsigned long new_flags;
871
872         WRITE_ONCE(ctx->released, true);
873
874         if (!mmget_not_zero(mm))
875                 goto wakeup;
876
877         /*
878          * Flush page faults out of all CPUs. NOTE: all page faults
879          * must be retried without returning VM_FAULT_SIGBUS if
880          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
881          * changes while handle_userfault released the mmap_sem. So
882          * it's critical that released is set to true (above), before
883          * taking the mmap_sem for writing.
884          */
885         down_write(&mm->mmap_sem);
886         prev = NULL;
887         for (vma = mm->mmap; vma; vma = vma->vm_next) {
888                 cond_resched();
889                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
890                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
891                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
892                         prev = vma;
893                         continue;
894                 }
895                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
896                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
897                                  new_flags, vma->anon_vma,
898                                  vma->vm_file, vma->vm_pgoff,
899                                  vma_policy(vma),
900                                  NULL_VM_UFFD_CTX);
901                 if (prev)
902                         vma = prev;
903                 else
904                         prev = vma;
905                 vma->vm_flags = new_flags;
906                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
907         }
908         up_write(&mm->mmap_sem);
909         mmput(mm);
910 wakeup:
911         /*
912          * After no new page faults can wait on this fault_*wqh, flush
913          * the last page faults that may have been already waiting on
914          * the fault_*wqh.
915          */
916         spin_lock(&ctx->fault_pending_wqh.lock);
917         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
918         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
919         spin_unlock(&ctx->fault_pending_wqh.lock);
920
921         /* Flush pending events that may still wait on event_wqh */
922         wake_up_all(&ctx->event_wqh);
923
924         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
925         userfaultfd_ctx_put(ctx);
926         return 0;
927 }
928
929 /* fault_pending_wqh.lock must be hold by the caller */
930 static inline struct userfaultfd_wait_queue *find_userfault_in(
931                 wait_queue_head_t *wqh)
932 {
933         wait_queue_entry_t *wq;
934         struct userfaultfd_wait_queue *uwq;
935
936         lockdep_assert_held(&wqh->lock);
937
938         uwq = NULL;
939         if (!waitqueue_active(wqh))
940                 goto out;
941         /* walk in reverse to provide FIFO behavior to read userfaults */
942         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
943         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
944 out:
945         return uwq;
946 }
947
948 static inline struct userfaultfd_wait_queue *find_userfault(
949                 struct userfaultfd_ctx *ctx)
950 {
951         return find_userfault_in(&ctx->fault_pending_wqh);
952 }
953
954 static inline struct userfaultfd_wait_queue *find_userfault_evt(
955                 struct userfaultfd_ctx *ctx)
956 {
957         return find_userfault_in(&ctx->event_wqh);
958 }
959
960 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
961 {
962         struct userfaultfd_ctx *ctx = file->private_data;
963         __poll_t ret;
964
965         poll_wait(file, &ctx->fd_wqh, wait);
966
967         switch (ctx->state) {
968         case UFFD_STATE_WAIT_API:
969                 return EPOLLERR;
970         case UFFD_STATE_RUNNING:
971                 /*
972                  * poll() never guarantees that read won't block.
973                  * userfaults can be waken before they're read().
974                  */
975                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
976                         return EPOLLERR;
977                 /*
978                  * lockless access to see if there are pending faults
979                  * __pollwait last action is the add_wait_queue but
980                  * the spin_unlock would allow the waitqueue_active to
981                  * pass above the actual list_add inside
982                  * add_wait_queue critical section. So use a full
983                  * memory barrier to serialize the list_add write of
984                  * add_wait_queue() with the waitqueue_active read
985                  * below.
986                  */
987                 ret = 0;
988                 smp_mb();
989                 if (waitqueue_active(&ctx->fault_pending_wqh))
990                         ret = EPOLLIN;
991                 else if (waitqueue_active(&ctx->event_wqh))
992                         ret = EPOLLIN;
993
994                 return ret;
995         default:
996                 WARN_ON_ONCE(1);
997                 return EPOLLERR;
998         }
999 }
1000
1001 static const struct file_operations userfaultfd_fops;
1002
1003 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
1004                                   struct userfaultfd_ctx *new,
1005                                   struct uffd_msg *msg)
1006 {
1007         int fd;
1008
1009         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
1010                               O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
1011         if (fd < 0)
1012                 return fd;
1013
1014         msg->arg.reserved.reserved1 = 0;
1015         msg->arg.fork.ufd = fd;
1016         return 0;
1017 }
1018
1019 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1020                                     struct uffd_msg *msg)
1021 {
1022         ssize_t ret;
1023         DECLARE_WAITQUEUE(wait, current);
1024         struct userfaultfd_wait_queue *uwq;
1025         /*
1026          * Handling fork event requires sleeping operations, so
1027          * we drop the event_wqh lock, then do these ops, then
1028          * lock it back and wake up the waiter. While the lock is
1029          * dropped the ewq may go away so we keep track of it
1030          * carefully.
1031          */
1032         LIST_HEAD(fork_event);
1033         struct userfaultfd_ctx *fork_nctx = NULL;
1034
1035         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1036         spin_lock_irq(&ctx->fd_wqh.lock);
1037         __add_wait_queue(&ctx->fd_wqh, &wait);
1038         for (;;) {
1039                 set_current_state(TASK_INTERRUPTIBLE);
1040                 spin_lock(&ctx->fault_pending_wqh.lock);
1041                 uwq = find_userfault(ctx);
1042                 if (uwq) {
1043                         /*
1044                          * Use a seqcount to repeat the lockless check
1045                          * in wake_userfault() to avoid missing
1046                          * wakeups because during the refile both
1047                          * waitqueue could become empty if this is the
1048                          * only userfault.
1049                          */
1050                         write_seqcount_begin(&ctx->refile_seq);
1051
1052                         /*
1053                          * The fault_pending_wqh.lock prevents the uwq
1054                          * to disappear from under us.
1055                          *
1056                          * Refile this userfault from
1057                          * fault_pending_wqh to fault_wqh, it's not
1058                          * pending anymore after we read it.
1059                          *
1060                          * Use list_del() by hand (as
1061                          * userfaultfd_wake_function also uses
1062                          * list_del_init() by hand) to be sure nobody
1063                          * changes __remove_wait_queue() to use
1064                          * list_del_init() in turn breaking the
1065                          * !list_empty_careful() check in
1066                          * handle_userfault(). The uwq->wq.head list
1067                          * must never be empty at any time during the
1068                          * refile, or the waitqueue could disappear
1069                          * from under us. The "wait_queue_head_t"
1070                          * parameter of __remove_wait_queue() is unused
1071                          * anyway.
1072                          */
1073                         list_del(&uwq->wq.entry);
1074                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1075
1076                         write_seqcount_end(&ctx->refile_seq);
1077
1078                         /* careful to always initialize msg if ret == 0 */
1079                         *msg = uwq->msg;
1080                         spin_unlock(&ctx->fault_pending_wqh.lock);
1081                         ret = 0;
1082                         break;
1083                 }
1084                 spin_unlock(&ctx->fault_pending_wqh.lock);
1085
1086                 spin_lock(&ctx->event_wqh.lock);
1087                 uwq = find_userfault_evt(ctx);
1088                 if (uwq) {
1089                         *msg = uwq->msg;
1090
1091                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1092                                 fork_nctx = (struct userfaultfd_ctx *)
1093                                         (unsigned long)
1094                                         uwq->msg.arg.reserved.reserved1;
1095                                 list_move(&uwq->wq.entry, &fork_event);
1096                                 /*
1097                                  * fork_nctx can be freed as soon as
1098                                  * we drop the lock, unless we take a
1099                                  * reference on it.
1100                                  */
1101                                 userfaultfd_ctx_get(fork_nctx);
1102                                 spin_unlock(&ctx->event_wqh.lock);
1103                                 ret = 0;
1104                                 break;
1105                         }
1106
1107                         userfaultfd_event_complete(ctx, uwq);
1108                         spin_unlock(&ctx->event_wqh.lock);
1109                         ret = 0;
1110                         break;
1111                 }
1112                 spin_unlock(&ctx->event_wqh.lock);
1113
1114                 if (signal_pending(current)) {
1115                         ret = -ERESTARTSYS;
1116                         break;
1117                 }
1118                 if (no_wait) {
1119                         ret = -EAGAIN;
1120                         break;
1121                 }
1122                 spin_unlock_irq(&ctx->fd_wqh.lock);
1123                 schedule();
1124                 spin_lock_irq(&ctx->fd_wqh.lock);
1125         }
1126         __remove_wait_queue(&ctx->fd_wqh, &wait);
1127         __set_current_state(TASK_RUNNING);
1128         spin_unlock_irq(&ctx->fd_wqh.lock);
1129
1130         if (!ret && msg->event == UFFD_EVENT_FORK) {
1131                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1132                 spin_lock(&ctx->event_wqh.lock);
1133                 if (!list_empty(&fork_event)) {
1134                         /*
1135                          * The fork thread didn't abort, so we can
1136                          * drop the temporary refcount.
1137                          */
1138                         userfaultfd_ctx_put(fork_nctx);
1139
1140                         uwq = list_first_entry(&fork_event,
1141                                                typeof(*uwq),
1142                                                wq.entry);
1143                         /*
1144                          * If fork_event list wasn't empty and in turn
1145                          * the event wasn't already released by fork
1146                          * (the event is allocated on fork kernel
1147                          * stack), put the event back to its place in
1148                          * the event_wq. fork_event head will be freed
1149                          * as soon as we return so the event cannot
1150                          * stay queued there no matter the current
1151                          * "ret" value.
1152                          */
1153                         list_del(&uwq->wq.entry);
1154                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1155
1156                         /*
1157                          * Leave the event in the waitqueue and report
1158                          * error to userland if we failed to resolve
1159                          * the userfault fork.
1160                          */
1161                         if (likely(!ret))
1162                                 userfaultfd_event_complete(ctx, uwq);
1163                 } else {
1164                         /*
1165                          * Here the fork thread aborted and the
1166                          * refcount from the fork thread on fork_nctx
1167                          * has already been released. We still hold
1168                          * the reference we took before releasing the
1169                          * lock above. If resolve_userfault_fork
1170                          * failed we've to drop it because the
1171                          * fork_nctx has to be freed in such case. If
1172                          * it succeeded we'll hold it because the new
1173                          * uffd references it.
1174                          */
1175                         if (ret)
1176                                 userfaultfd_ctx_put(fork_nctx);
1177                 }
1178                 spin_unlock(&ctx->event_wqh.lock);
1179         }
1180
1181         return ret;
1182 }
1183
1184 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1185                                 size_t count, loff_t *ppos)
1186 {
1187         struct userfaultfd_ctx *ctx = file->private_data;
1188         ssize_t _ret, ret = 0;
1189         struct uffd_msg msg;
1190         int no_wait = file->f_flags & O_NONBLOCK;
1191
1192         if (ctx->state == UFFD_STATE_WAIT_API)
1193                 return -EINVAL;
1194
1195         for (;;) {
1196                 if (count < sizeof(msg))
1197                         return ret ? ret : -EINVAL;
1198                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1199                 if (_ret < 0)
1200                         return ret ? ret : _ret;
1201                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1202                         return ret ? ret : -EFAULT;
1203                 ret += sizeof(msg);
1204                 buf += sizeof(msg);
1205                 count -= sizeof(msg);
1206                 /*
1207                  * Allow to read more than one fault at time but only
1208                  * block if waiting for the very first one.
1209                  */
1210                 no_wait = O_NONBLOCK;
1211         }
1212 }
1213
1214 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1215                              struct userfaultfd_wake_range *range)
1216 {
1217         spin_lock(&ctx->fault_pending_wqh.lock);
1218         /* wake all in the range and autoremove */
1219         if (waitqueue_active(&ctx->fault_pending_wqh))
1220                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1221                                      range);
1222         if (waitqueue_active(&ctx->fault_wqh))
1223                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1224         spin_unlock(&ctx->fault_pending_wqh.lock);
1225 }
1226
1227 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1228                                            struct userfaultfd_wake_range *range)
1229 {
1230         unsigned seq;
1231         bool need_wakeup;
1232
1233         /*
1234          * To be sure waitqueue_active() is not reordered by the CPU
1235          * before the pagetable update, use an explicit SMP memory
1236          * barrier here. PT lock release or up_read(mmap_sem) still
1237          * have release semantics that can allow the
1238          * waitqueue_active() to be reordered before the pte update.
1239          */
1240         smp_mb();
1241
1242         /*
1243          * Use waitqueue_active because it's very frequent to
1244          * change the address space atomically even if there are no
1245          * userfaults yet. So we take the spinlock only when we're
1246          * sure we've userfaults to wake.
1247          */
1248         do {
1249                 seq = read_seqcount_begin(&ctx->refile_seq);
1250                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1251                         waitqueue_active(&ctx->fault_wqh);
1252                 cond_resched();
1253         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1254         if (need_wakeup)
1255                 __wake_userfault(ctx, range);
1256 }
1257
1258 static __always_inline int validate_range(struct mm_struct *mm,
1259                                           __u64 start, __u64 len)
1260 {
1261         __u64 task_size = mm->task_size;
1262
1263         if (start & ~PAGE_MASK)
1264                 return -EINVAL;
1265         if (len & ~PAGE_MASK)
1266                 return -EINVAL;
1267         if (!len)
1268                 return -EINVAL;
1269         if (start < mmap_min_addr)
1270                 return -EINVAL;
1271         if (start >= task_size)
1272                 return -EINVAL;
1273         if (len > task_size - start)
1274                 return -EINVAL;
1275         return 0;
1276 }
1277
1278 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1279 {
1280         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1281                 vma_is_shmem(vma);
1282 }
1283
1284 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1285                                 unsigned long arg)
1286 {
1287         struct mm_struct *mm = ctx->mm;
1288         struct vm_area_struct *vma, *prev, *cur;
1289         int ret;
1290         struct uffdio_register uffdio_register;
1291         struct uffdio_register __user *user_uffdio_register;
1292         unsigned long vm_flags, new_flags;
1293         bool found;
1294         bool basic_ioctls;
1295         unsigned long start, end, vma_end;
1296
1297         user_uffdio_register = (struct uffdio_register __user *) arg;
1298
1299         ret = -EFAULT;
1300         if (copy_from_user(&uffdio_register, user_uffdio_register,
1301                            sizeof(uffdio_register)-sizeof(__u64)))
1302                 goto out;
1303
1304         ret = -EINVAL;
1305         if (!uffdio_register.mode)
1306                 goto out;
1307         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1308                                      UFFDIO_REGISTER_MODE_WP))
1309                 goto out;
1310         vm_flags = 0;
1311         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1312                 vm_flags |= VM_UFFD_MISSING;
1313         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1314                 vm_flags |= VM_UFFD_WP;
1315                 /*
1316                  * FIXME: remove the below error constraint by
1317                  * implementing the wprotect tracking mode.
1318                  */
1319                 ret = -EINVAL;
1320                 goto out;
1321         }
1322
1323         ret = validate_range(mm, uffdio_register.range.start,
1324                              uffdio_register.range.len);
1325         if (ret)
1326                 goto out;
1327
1328         start = uffdio_register.range.start;
1329         end = start + uffdio_register.range.len;
1330
1331         ret = -ENOMEM;
1332         if (!mmget_not_zero(mm))
1333                 goto out;
1334
1335         down_write(&mm->mmap_sem);
1336         vma = find_vma_prev(mm, start, &prev);
1337         if (!vma)
1338                 goto out_unlock;
1339
1340         /* check that there's at least one vma in the range */
1341         ret = -EINVAL;
1342         if (vma->vm_start >= end)
1343                 goto out_unlock;
1344
1345         /*
1346          * If the first vma contains huge pages, make sure start address
1347          * is aligned to huge page size.
1348          */
1349         if (is_vm_hugetlb_page(vma)) {
1350                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1351
1352                 if (start & (vma_hpagesize - 1))
1353                         goto out_unlock;
1354         }
1355
1356         /*
1357          * Search for not compatible vmas.
1358          */
1359         found = false;
1360         basic_ioctls = false;
1361         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1362                 cond_resched();
1363
1364                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1365                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1366
1367                 /* check not compatible vmas */
1368                 ret = -EINVAL;
1369                 if (!vma_can_userfault(cur))
1370                         goto out_unlock;
1371
1372                 /*
1373                  * UFFDIO_COPY will fill file holes even without
1374                  * PROT_WRITE. This check enforces that if this is a
1375                  * MAP_SHARED, the process has write permission to the backing
1376                  * file. If VM_MAYWRITE is set it also enforces that on a
1377                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1378                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1379                  */
1380                 ret = -EPERM;
1381                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1382                         goto out_unlock;
1383
1384                 /*
1385                  * If this vma contains ending address, and huge pages
1386                  * check alignment.
1387                  */
1388                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1389                     end > cur->vm_start) {
1390                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1391
1392                         ret = -EINVAL;
1393
1394                         if (end & (vma_hpagesize - 1))
1395                                 goto out_unlock;
1396                 }
1397
1398                 /*
1399                  * Check that this vma isn't already owned by a
1400                  * different userfaultfd. We can't allow more than one
1401                  * userfaultfd to own a single vma simultaneously or we
1402                  * wouldn't know which one to deliver the userfaults to.
1403                  */
1404                 ret = -EBUSY;
1405                 if (cur->vm_userfaultfd_ctx.ctx &&
1406                     cur->vm_userfaultfd_ctx.ctx != ctx)
1407                         goto out_unlock;
1408
1409                 /*
1410                  * Note vmas containing huge pages
1411                  */
1412                 if (is_vm_hugetlb_page(cur))
1413                         basic_ioctls = true;
1414
1415                 found = true;
1416         }
1417         BUG_ON(!found);
1418
1419         if (vma->vm_start < start)
1420                 prev = vma;
1421
1422         ret = 0;
1423         do {
1424                 cond_resched();
1425
1426                 BUG_ON(!vma_can_userfault(vma));
1427                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1428                        vma->vm_userfaultfd_ctx.ctx != ctx);
1429                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1430
1431                 /*
1432                  * Nothing to do: this vma is already registered into this
1433                  * userfaultfd and with the right tracking mode too.
1434                  */
1435                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1436                     (vma->vm_flags & vm_flags) == vm_flags)
1437                         goto skip;
1438
1439                 if (vma->vm_start > start)
1440                         start = vma->vm_start;
1441                 vma_end = min(end, vma->vm_end);
1442
1443                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1444                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1445                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1446                                  vma_policy(vma),
1447                                  ((struct vm_userfaultfd_ctx){ ctx }));
1448                 if (prev) {
1449                         vma = prev;
1450                         goto next;
1451                 }
1452                 if (vma->vm_start < start) {
1453                         ret = split_vma(mm, vma, start, 1);
1454                         if (ret)
1455                                 break;
1456                 }
1457                 if (vma->vm_end > end) {
1458                         ret = split_vma(mm, vma, end, 0);
1459                         if (ret)
1460                                 break;
1461                 }
1462         next:
1463                 /*
1464                  * In the vma_merge() successful mprotect-like case 8:
1465                  * the next vma was merged into the current one and
1466                  * the current one has not been updated yet.
1467                  */
1468                 vma->vm_flags = new_flags;
1469                 vma->vm_userfaultfd_ctx.ctx = ctx;
1470
1471         skip:
1472                 prev = vma;
1473                 start = vma->vm_end;
1474                 vma = vma->vm_next;
1475         } while (vma && vma->vm_start < end);
1476 out_unlock:
1477         up_write(&mm->mmap_sem);
1478         mmput(mm);
1479         if (!ret) {
1480                 /*
1481                  * Now that we scanned all vmas we can already tell
1482                  * userland which ioctls methods are guaranteed to
1483                  * succeed on this range.
1484                  */
1485                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1486                              UFFD_API_RANGE_IOCTLS,
1487                              &user_uffdio_register->ioctls))
1488                         ret = -EFAULT;
1489         }
1490 out:
1491         return ret;
1492 }
1493
1494 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1495                                   unsigned long arg)
1496 {
1497         struct mm_struct *mm = ctx->mm;
1498         struct vm_area_struct *vma, *prev, *cur;
1499         int ret;
1500         struct uffdio_range uffdio_unregister;
1501         unsigned long new_flags;
1502         bool found;
1503         unsigned long start, end, vma_end;
1504         const void __user *buf = (void __user *)arg;
1505
1506         ret = -EFAULT;
1507         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1508                 goto out;
1509
1510         ret = validate_range(mm, uffdio_unregister.start,
1511                              uffdio_unregister.len);
1512         if (ret)
1513                 goto out;
1514
1515         start = uffdio_unregister.start;
1516         end = start + uffdio_unregister.len;
1517
1518         ret = -ENOMEM;
1519         if (!mmget_not_zero(mm))
1520                 goto out;
1521
1522         down_write(&mm->mmap_sem);
1523         vma = find_vma_prev(mm, start, &prev);
1524         if (!vma)
1525                 goto out_unlock;
1526
1527         /* check that there's at least one vma in the range */
1528         ret = -EINVAL;
1529         if (vma->vm_start >= end)
1530                 goto out_unlock;
1531
1532         /*
1533          * If the first vma contains huge pages, make sure start address
1534          * is aligned to huge page size.
1535          */
1536         if (is_vm_hugetlb_page(vma)) {
1537                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1538
1539                 if (start & (vma_hpagesize - 1))
1540                         goto out_unlock;
1541         }
1542
1543         /*
1544          * Search for not compatible vmas.
1545          */
1546         found = false;
1547         ret = -EINVAL;
1548         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1549                 cond_resched();
1550
1551                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1552                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1553
1554                 /*
1555                  * Check not compatible vmas, not strictly required
1556                  * here as not compatible vmas cannot have an
1557                  * userfaultfd_ctx registered on them, but this
1558                  * provides for more strict behavior to notice
1559                  * unregistration errors.
1560                  */
1561                 if (!vma_can_userfault(cur))
1562                         goto out_unlock;
1563
1564                 found = true;
1565         }
1566         BUG_ON(!found);
1567
1568         if (vma->vm_start < start)
1569                 prev = vma;
1570
1571         ret = 0;
1572         do {
1573                 cond_resched();
1574
1575                 BUG_ON(!vma_can_userfault(vma));
1576
1577                 /*
1578                  * Nothing to do: this vma is already registered into this
1579                  * userfaultfd and with the right tracking mode too.
1580                  */
1581                 if (!vma->vm_userfaultfd_ctx.ctx)
1582                         goto skip;
1583
1584                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1585
1586                 if (vma->vm_start > start)
1587                         start = vma->vm_start;
1588                 vma_end = min(end, vma->vm_end);
1589
1590                 if (userfaultfd_missing(vma)) {
1591                         /*
1592                          * Wake any concurrent pending userfault while
1593                          * we unregister, so they will not hang
1594                          * permanently and it avoids userland to call
1595                          * UFFDIO_WAKE explicitly.
1596                          */
1597                         struct userfaultfd_wake_range range;
1598                         range.start = start;
1599                         range.len = vma_end - start;
1600                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1601                 }
1602
1603                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1604                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1605                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1606                                  vma_policy(vma),
1607                                  NULL_VM_UFFD_CTX);
1608                 if (prev) {
1609                         vma = prev;
1610                         goto next;
1611                 }
1612                 if (vma->vm_start < start) {
1613                         ret = split_vma(mm, vma, start, 1);
1614                         if (ret)
1615                                 break;
1616                 }
1617                 if (vma->vm_end > end) {
1618                         ret = split_vma(mm, vma, end, 0);
1619                         if (ret)
1620                                 break;
1621                 }
1622         next:
1623                 /*
1624                  * In the vma_merge() successful mprotect-like case 8:
1625                  * the next vma was merged into the current one and
1626                  * the current one has not been updated yet.
1627                  */
1628                 vma->vm_flags = new_flags;
1629                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1630
1631         skip:
1632                 prev = vma;
1633                 start = vma->vm_end;
1634                 vma = vma->vm_next;
1635         } while (vma && vma->vm_start < end);
1636 out_unlock:
1637         up_write(&mm->mmap_sem);
1638         mmput(mm);
1639 out:
1640         return ret;
1641 }
1642
1643 /*
1644  * userfaultfd_wake may be used in combination with the
1645  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1646  */
1647 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1648                             unsigned long arg)
1649 {
1650         int ret;
1651         struct uffdio_range uffdio_wake;
1652         struct userfaultfd_wake_range range;
1653         const void __user *buf = (void __user *)arg;
1654
1655         ret = -EFAULT;
1656         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1657                 goto out;
1658
1659         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1660         if (ret)
1661                 goto out;
1662
1663         range.start = uffdio_wake.start;
1664         range.len = uffdio_wake.len;
1665
1666         /*
1667          * len == 0 means wake all and we don't want to wake all here,
1668          * so check it again to be sure.
1669          */
1670         VM_BUG_ON(!range.len);
1671
1672         wake_userfault(ctx, &range);
1673         ret = 0;
1674
1675 out:
1676         return ret;
1677 }
1678
1679 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1680                             unsigned long arg)
1681 {
1682         __s64 ret;
1683         struct uffdio_copy uffdio_copy;
1684         struct uffdio_copy __user *user_uffdio_copy;
1685         struct userfaultfd_wake_range range;
1686
1687         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1688
1689         ret = -EAGAIN;
1690         if (READ_ONCE(ctx->mmap_changing))
1691                 goto out;
1692
1693         ret = -EFAULT;
1694         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1695                            /* don't copy "copy" last field */
1696                            sizeof(uffdio_copy)-sizeof(__s64)))
1697                 goto out;
1698
1699         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1700         if (ret)
1701                 goto out;
1702         /*
1703          * double check for wraparound just in case. copy_from_user()
1704          * will later check uffdio_copy.src + uffdio_copy.len to fit
1705          * in the userland range.
1706          */
1707         ret = -EINVAL;
1708         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1709                 goto out;
1710         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1711                 goto out;
1712         if (mmget_not_zero(ctx->mm)) {
1713                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1714                                    uffdio_copy.len, &ctx->mmap_changing);
1715                 mmput(ctx->mm);
1716         } else {
1717                 return -ESRCH;
1718         }
1719         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1720                 return -EFAULT;
1721         if (ret < 0)
1722                 goto out;
1723         BUG_ON(!ret);
1724         /* len == 0 would wake all */
1725         range.len = ret;
1726         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1727                 range.start = uffdio_copy.dst;
1728                 wake_userfault(ctx, &range);
1729         }
1730         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1731 out:
1732         return ret;
1733 }
1734
1735 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1736                                 unsigned long arg)
1737 {
1738         __s64 ret;
1739         struct uffdio_zeropage uffdio_zeropage;
1740         struct uffdio_zeropage __user *user_uffdio_zeropage;
1741         struct userfaultfd_wake_range range;
1742
1743         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1744
1745         ret = -EAGAIN;
1746         if (READ_ONCE(ctx->mmap_changing))
1747                 goto out;
1748
1749         ret = -EFAULT;
1750         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1751                            /* don't copy "zeropage" last field */
1752                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1753                 goto out;
1754
1755         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1756                              uffdio_zeropage.range.len);
1757         if (ret)
1758                 goto out;
1759         ret = -EINVAL;
1760         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1761                 goto out;
1762
1763         if (mmget_not_zero(ctx->mm)) {
1764                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1765                                      uffdio_zeropage.range.len,
1766                                      &ctx->mmap_changing);
1767                 mmput(ctx->mm);
1768         } else {
1769                 return -ESRCH;
1770         }
1771         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1772                 return -EFAULT;
1773         if (ret < 0)
1774                 goto out;
1775         /* len == 0 would wake all */
1776         BUG_ON(!ret);
1777         range.len = ret;
1778         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1779                 range.start = uffdio_zeropage.range.start;
1780                 wake_userfault(ctx, &range);
1781         }
1782         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1783 out:
1784         return ret;
1785 }
1786
1787 static inline unsigned int uffd_ctx_features(__u64 user_features)
1788 {
1789         /*
1790          * For the current set of features the bits just coincide
1791          */
1792         return (unsigned int)user_features;
1793 }
1794
1795 /*
1796  * userland asks for a certain API version and we return which bits
1797  * and ioctl commands are implemented in this kernel for such API
1798  * version or -EINVAL if unknown.
1799  */
1800 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1801                            unsigned long arg)
1802 {
1803         struct uffdio_api uffdio_api;
1804         void __user *buf = (void __user *)arg;
1805         int ret;
1806         __u64 features;
1807
1808         ret = -EINVAL;
1809         if (ctx->state != UFFD_STATE_WAIT_API)
1810                 goto out;
1811         ret = -EFAULT;
1812         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1813                 goto out;
1814         features = uffdio_api.features;
1815         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1816                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1817                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1818                         goto out;
1819                 ret = -EINVAL;
1820                 goto out;
1821         }
1822         /* report all available features and ioctls to userland */
1823         uffdio_api.features = UFFD_API_FEATURES;
1824         uffdio_api.ioctls = UFFD_API_IOCTLS;
1825         ret = -EFAULT;
1826         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1827                 goto out;
1828         ctx->state = UFFD_STATE_RUNNING;
1829         /* only enable the requested features for this uffd context */
1830         ctx->features = uffd_ctx_features(features);
1831         ret = 0;
1832 out:
1833         return ret;
1834 }
1835
1836 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1837                               unsigned long arg)
1838 {
1839         int ret = -EINVAL;
1840         struct userfaultfd_ctx *ctx = file->private_data;
1841
1842         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1843                 return -EINVAL;
1844
1845         switch(cmd) {
1846         case UFFDIO_API:
1847                 ret = userfaultfd_api(ctx, arg);
1848                 break;
1849         case UFFDIO_REGISTER:
1850                 ret = userfaultfd_register(ctx, arg);
1851                 break;
1852         case UFFDIO_UNREGISTER:
1853                 ret = userfaultfd_unregister(ctx, arg);
1854                 break;
1855         case UFFDIO_WAKE:
1856                 ret = userfaultfd_wake(ctx, arg);
1857                 break;
1858         case UFFDIO_COPY:
1859                 ret = userfaultfd_copy(ctx, arg);
1860                 break;
1861         case UFFDIO_ZEROPAGE:
1862                 ret = userfaultfd_zeropage(ctx, arg);
1863                 break;
1864         }
1865         return ret;
1866 }
1867
1868 #ifdef CONFIG_PROC_FS
1869 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1870 {
1871         struct userfaultfd_ctx *ctx = f->private_data;
1872         wait_queue_entry_t *wq;
1873         unsigned long pending = 0, total = 0;
1874
1875         spin_lock(&ctx->fault_pending_wqh.lock);
1876         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1877                 pending++;
1878                 total++;
1879         }
1880         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1881                 total++;
1882         }
1883         spin_unlock(&ctx->fault_pending_wqh.lock);
1884
1885         /*
1886          * If more protocols will be added, there will be all shown
1887          * separated by a space. Like this:
1888          *      protocols: aa:... bb:...
1889          */
1890         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1891                    pending, total, UFFD_API, ctx->features,
1892                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1893 }
1894 #endif
1895
1896 static const struct file_operations userfaultfd_fops = {
1897 #ifdef CONFIG_PROC_FS
1898         .show_fdinfo    = userfaultfd_show_fdinfo,
1899 #endif
1900         .release        = userfaultfd_release,
1901         .poll           = userfaultfd_poll,
1902         .read           = userfaultfd_read,
1903         .unlocked_ioctl = userfaultfd_ioctl,
1904         .compat_ioctl   = userfaultfd_ioctl,
1905         .llseek         = noop_llseek,
1906 };
1907
1908 static void init_once_userfaultfd_ctx(void *mem)
1909 {
1910         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1911
1912         init_waitqueue_head(&ctx->fault_pending_wqh);
1913         init_waitqueue_head(&ctx->fault_wqh);
1914         init_waitqueue_head(&ctx->event_wqh);
1915         init_waitqueue_head(&ctx->fd_wqh);
1916         seqcount_init(&ctx->refile_seq);
1917 }
1918
1919 SYSCALL_DEFINE1(userfaultfd, int, flags)
1920 {
1921         struct userfaultfd_ctx *ctx;
1922         int fd;
1923
1924         BUG_ON(!current->mm);
1925
1926         /* Check the UFFD_* constants for consistency.  */
1927         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1928         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1929
1930         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1931                 return -EINVAL;
1932
1933         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1934         if (!ctx)
1935                 return -ENOMEM;
1936
1937         refcount_set(&ctx->refcount, 1);
1938         ctx->flags = flags;
1939         ctx->features = 0;
1940         ctx->state = UFFD_STATE_WAIT_API;
1941         ctx->released = false;
1942         ctx->mmap_changing = false;
1943         ctx->mm = current->mm;
1944         /* prevent the mm struct to be freed */
1945         mmgrab(ctx->mm);
1946
1947         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
1948                               O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1949         if (fd < 0) {
1950                 mmdrop(ctx->mm);
1951                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1952         }
1953         return fd;
1954 }
1955
1956 static int __init userfaultfd_init(void)
1957 {
1958         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1959                                                 sizeof(struct userfaultfd_ctx),
1960                                                 0,
1961                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1962                                                 init_once_userfaultfd_ctx);
1963         return 0;
1964 }
1965 __initcall(userfaultfd_init);