f46d133c094998c48eada7a4e2bd8f5a69239ffa
[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         atomic_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         /* mm with one ore more vmas attached to this userfaultfd_ctx */
66         struct mm_struct *mm;
67 };
68
69 struct userfaultfd_fork_ctx {
70         struct userfaultfd_ctx *orig;
71         struct userfaultfd_ctx *new;
72         struct list_head list;
73 };
74
75 struct userfaultfd_unmap_ctx {
76         struct userfaultfd_ctx *ctx;
77         unsigned long start;
78         unsigned long end;
79         struct list_head list;
80 };
81
82 struct userfaultfd_wait_queue {
83         struct uffd_msg msg;
84         wait_queue_entry_t wq;
85         struct userfaultfd_ctx *ctx;
86         bool waken;
87 };
88
89 struct userfaultfd_wake_range {
90         unsigned long start;
91         unsigned long len;
92 };
93
94 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
95                                      int wake_flags, void *key)
96 {
97         struct userfaultfd_wake_range *range = key;
98         int ret;
99         struct userfaultfd_wait_queue *uwq;
100         unsigned long start, len;
101
102         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
103         ret = 0;
104         /* len == 0 means wake all */
105         start = range->start;
106         len = range->len;
107         if (len && (start > uwq->msg.arg.pagefault.address ||
108                     start + len <= uwq->msg.arg.pagefault.address))
109                 goto out;
110         WRITE_ONCE(uwq->waken, true);
111         /*
112          * The Program-Order guarantees provided by the scheduler
113          * ensure uwq->waken is visible before the task is woken.
114          */
115         ret = wake_up_state(wq->private, mode);
116         if (ret) {
117                 /*
118                  * Wake only once, autoremove behavior.
119                  *
120                  * After the effect of list_del_init is visible to the other
121                  * CPUs, the waitqueue may disappear from under us, see the
122                  * !list_empty_careful() in handle_userfault().
123                  *
124                  * try_to_wake_up() has an implicit smp_mb(), and the
125                  * wq->private is read before calling the extern function
126                  * "wake_up_state" (which in turns calls try_to_wake_up).
127                  */
128                 list_del_init(&wq->entry);
129         }
130 out:
131         return ret;
132 }
133
134 /**
135  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
136  * context.
137  * @ctx: [in] Pointer to the userfaultfd context.
138  */
139 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
140 {
141         if (!atomic_inc_not_zero(&ctx->refcount))
142                 BUG();
143 }
144
145 /**
146  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
147  * context.
148  * @ctx: [in] Pointer to userfaultfd context.
149  *
150  * The userfaultfd context reference must have been previously acquired either
151  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
152  */
153 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
154 {
155         if (atomic_dec_and_test(&ctx->refcount)) {
156                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
157                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
158                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
159                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
160                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
161                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
162                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
163                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
164                 mmdrop(ctx->mm);
165                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
166         }
167 }
168
169 static inline void msg_init(struct uffd_msg *msg)
170 {
171         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
172         /*
173          * Must use memset to zero out the paddings or kernel data is
174          * leaked to userland.
175          */
176         memset(msg, 0, sizeof(struct uffd_msg));
177 }
178
179 static inline struct uffd_msg userfault_msg(unsigned long address,
180                                             unsigned int flags,
181                                             unsigned long reason,
182                                             unsigned int features)
183 {
184         struct uffd_msg msg;
185         msg_init(&msg);
186         msg.event = UFFD_EVENT_PAGEFAULT;
187         msg.arg.pagefault.address = address;
188         if (flags & FAULT_FLAG_WRITE)
189                 /*
190                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
191                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
192                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
193                  * was a read fault, otherwise if set it means it's
194                  * a write fault.
195                  */
196                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
197         if (reason & VM_UFFD_WP)
198                 /*
199                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
200                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
201                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
202                  * a missing fault, otherwise if set it means it's a
203                  * write protect fault.
204                  */
205                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
206         if (features & UFFD_FEATURE_THREAD_ID)
207                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
208         return msg;
209 }
210
211 #ifdef CONFIG_HUGETLB_PAGE
212 /*
213  * Same functionality as userfaultfd_must_wait below with modifications for
214  * hugepmd ranges.
215  */
216 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
217                                          struct vm_area_struct *vma,
218                                          unsigned long address,
219                                          unsigned long flags,
220                                          unsigned long reason)
221 {
222         struct mm_struct *mm = ctx->mm;
223         pte_t *pte;
224         bool ret = true;
225
226         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
227
228         pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
229         if (!pte)
230                 goto out;
231
232         ret = false;
233
234         /*
235          * Lockless access: we're in a wait_event so it's ok if it
236          * changes under us.
237          */
238         if (huge_pte_none(*pte))
239                 ret = true;
240         if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
241                 ret = true;
242 out:
243         return ret;
244 }
245 #else
246 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
247                                          struct vm_area_struct *vma,
248                                          unsigned long address,
249                                          unsigned long flags,
250                                          unsigned long reason)
251 {
252         return false;   /* should never get here */
253 }
254 #endif /* CONFIG_HUGETLB_PAGE */
255
256 /*
257  * Verify the pagetables are still not ok after having reigstered into
258  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
259  * userfault that has already been resolved, if userfaultfd_read and
260  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
261  * threads.
262  */
263 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
264                                          unsigned long address,
265                                          unsigned long flags,
266                                          unsigned long reason)
267 {
268         struct mm_struct *mm = ctx->mm;
269         pgd_t *pgd;
270         p4d_t *p4d;
271         pud_t *pud;
272         pmd_t *pmd, _pmd;
273         pte_t *pte;
274         bool ret = true;
275
276         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
277
278         pgd = pgd_offset(mm, address);
279         if (!pgd_present(*pgd))
280                 goto out;
281         p4d = p4d_offset(pgd, address);
282         if (!p4d_present(*p4d))
283                 goto out;
284         pud = pud_offset(p4d, address);
285         if (!pud_present(*pud))
286                 goto out;
287         pmd = pmd_offset(pud, address);
288         /*
289          * READ_ONCE must function as a barrier with narrower scope
290          * and it must be equivalent to:
291          *      _pmd = *pmd; barrier();
292          *
293          * This is to deal with the instability (as in
294          * pmd_trans_unstable) of the pmd.
295          */
296         _pmd = READ_ONCE(*pmd);
297         if (!pmd_present(_pmd))
298                 goto out;
299
300         ret = false;
301         if (pmd_trans_huge(_pmd))
302                 goto out;
303
304         /*
305          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
306          * and use the standard pte_offset_map() instead of parsing _pmd.
307          */
308         pte = pte_offset_map(pmd, address);
309         /*
310          * Lockless access: we're in a wait_event so it's ok if it
311          * changes under us.
312          */
313         if (pte_none(*pte))
314                 ret = true;
315         pte_unmap(pte);
316
317 out:
318         return ret;
319 }
320
321 /*
322  * The locking rules involved in returning VM_FAULT_RETRY depending on
323  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
324  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
325  * recommendation in __lock_page_or_retry is not an understatement.
326  *
327  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
328  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
329  * not set.
330  *
331  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
332  * set, VM_FAULT_RETRY can still be returned if and only if there are
333  * fatal_signal_pending()s, and the mmap_sem must be released before
334  * returning it.
335  */
336 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
337 {
338         struct mm_struct *mm = vmf->vma->vm_mm;
339         struct userfaultfd_ctx *ctx;
340         struct userfaultfd_wait_queue uwq;
341         int ret;
342         bool must_wait, return_to_userland;
343         long blocking_state;
344
345         ret = VM_FAULT_SIGBUS;
346
347         /*
348          * We don't do userfault handling for the final child pid update.
349          *
350          * We also don't do userfault handling during
351          * coredumping. hugetlbfs has the special
352          * follow_hugetlb_page() to skip missing pages in the
353          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
354          * the no_page_table() helper in follow_page_mask(), but the
355          * shmem_vm_ops->fault method is invoked even during
356          * coredumping without mmap_sem and it ends up here.
357          */
358         if (current->flags & (PF_EXITING|PF_DUMPCORE))
359                 goto out;
360
361         /*
362          * Coredumping runs without mmap_sem so we can only check that
363          * the mmap_sem is held, if PF_DUMPCORE was not set.
364          */
365         WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
366
367         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
368         if (!ctx)
369                 goto out;
370
371         BUG_ON(ctx->mm != mm);
372
373         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
374         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
375
376         if (ctx->features & UFFD_FEATURE_SIGBUS)
377                 goto out;
378
379         /*
380          * If it's already released don't get it. This avoids to loop
381          * in __get_user_pages if userfaultfd_release waits on the
382          * caller of handle_userfault to release the mmap_sem.
383          */
384         if (unlikely(READ_ONCE(ctx->released))) {
385                 /*
386                  * Don't return VM_FAULT_SIGBUS in this case, so a non
387                  * cooperative manager can close the uffd after the
388                  * last UFFDIO_COPY, without risking to trigger an
389                  * involuntary SIGBUS if the process was starting the
390                  * userfaultfd while the userfaultfd was still armed
391                  * (but after the last UFFDIO_COPY). If the uffd
392                  * wasn't already closed when the userfault reached
393                  * this point, that would normally be solved by
394                  * userfaultfd_must_wait returning 'false'.
395                  *
396                  * If we were to return VM_FAULT_SIGBUS here, the non
397                  * cooperative manager would be instead forced to
398                  * always call UFFDIO_UNREGISTER before it can safely
399                  * close the uffd.
400                  */
401                 ret = VM_FAULT_NOPAGE;
402                 goto out;
403         }
404
405         /*
406          * Check that we can return VM_FAULT_RETRY.
407          *
408          * NOTE: it should become possible to return VM_FAULT_RETRY
409          * even if FAULT_FLAG_TRIED is set without leading to gup()
410          * -EBUSY failures, if the userfaultfd is to be extended for
411          * VM_UFFD_WP tracking and we intend to arm the userfault
412          * without first stopping userland access to the memory. For
413          * VM_UFFD_MISSING userfaults this is enough for now.
414          */
415         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
416                 /*
417                  * Validate the invariant that nowait must allow retry
418                  * to be sure not to return SIGBUS erroneously on
419                  * nowait invocations.
420                  */
421                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
422 #ifdef CONFIG_DEBUG_VM
423                 if (printk_ratelimit()) {
424                         printk(KERN_WARNING
425                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
426                                vmf->flags);
427                         dump_stack();
428                 }
429 #endif
430                 goto out;
431         }
432
433         /*
434          * Handle nowait, not much to do other than tell it to retry
435          * and wait.
436          */
437         ret = VM_FAULT_RETRY;
438         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
439                 goto out;
440
441         /* take the reference before dropping the mmap_sem */
442         userfaultfd_ctx_get(ctx);
443
444         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
445         uwq.wq.private = current;
446         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
447                         ctx->features);
448         uwq.ctx = ctx;
449         uwq.waken = false;
450
451         return_to_userland =
452                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
453                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
454         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
455                          TASK_KILLABLE;
456
457         spin_lock(&ctx->fault_pending_wqh.lock);
458         /*
459          * After the __add_wait_queue the uwq is visible to userland
460          * through poll/read().
461          */
462         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
463         /*
464          * The smp_mb() after __set_current_state prevents the reads
465          * following the spin_unlock to happen before the list_add in
466          * __add_wait_queue.
467          */
468         set_current_state(blocking_state);
469         spin_unlock(&ctx->fault_pending_wqh.lock);
470
471         if (!is_vm_hugetlb_page(vmf->vma))
472                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
473                                                   reason);
474         else
475                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
476                                                        vmf->address,
477                                                        vmf->flags, reason);
478         up_read(&mm->mmap_sem);
479
480         if (likely(must_wait && !READ_ONCE(ctx->released) &&
481                    (return_to_userland ? !signal_pending(current) :
482                     !fatal_signal_pending(current)))) {
483                 wake_up_poll(&ctx->fd_wqh, POLLIN);
484                 schedule();
485                 ret |= VM_FAULT_MAJOR;
486
487                 /*
488                  * False wakeups can orginate even from rwsem before
489                  * up_read() however userfaults will wait either for a
490                  * targeted wakeup on the specific uwq waitqueue from
491                  * wake_userfault() or for signals or for uffd
492                  * release.
493                  */
494                 while (!READ_ONCE(uwq.waken)) {
495                         /*
496                          * This needs the full smp_store_mb()
497                          * guarantee as the state write must be
498                          * visible to other CPUs before reading
499                          * uwq.waken from other CPUs.
500                          */
501                         set_current_state(blocking_state);
502                         if (READ_ONCE(uwq.waken) ||
503                             READ_ONCE(ctx->released) ||
504                             (return_to_userland ? signal_pending(current) :
505                              fatal_signal_pending(current)))
506                                 break;
507                         schedule();
508                 }
509         }
510
511         __set_current_state(TASK_RUNNING);
512
513         if (return_to_userland) {
514                 if (signal_pending(current) &&
515                     !fatal_signal_pending(current)) {
516                         /*
517                          * If we got a SIGSTOP or SIGCONT and this is
518                          * a normal userland page fault, just let
519                          * userland return so the signal will be
520                          * handled and gdb debugging works.  The page
521                          * fault code immediately after we return from
522                          * this function is going to release the
523                          * mmap_sem and it's not depending on it
524                          * (unlike gup would if we were not to return
525                          * VM_FAULT_RETRY).
526                          *
527                          * If a fatal signal is pending we still take
528                          * the streamlined VM_FAULT_RETRY failure path
529                          * and there's no need to retake the mmap_sem
530                          * in such case.
531                          */
532                         down_read(&mm->mmap_sem);
533                         ret = VM_FAULT_NOPAGE;
534                 }
535         }
536
537         /*
538          * Here we race with the list_del; list_add in
539          * userfaultfd_ctx_read(), however because we don't ever run
540          * list_del_init() to refile across the two lists, the prev
541          * and next pointers will never point to self. list_add also
542          * would never let any of the two pointers to point to
543          * self. So list_empty_careful won't risk to see both pointers
544          * pointing to self at any time during the list refile. The
545          * only case where list_del_init() is called is the full
546          * removal in the wake function and there we don't re-list_add
547          * and it's fine not to block on the spinlock. The uwq on this
548          * kernel stack can be released after the list_del_init.
549          */
550         if (!list_empty_careful(&uwq.wq.entry)) {
551                 spin_lock(&ctx->fault_pending_wqh.lock);
552                 /*
553                  * No need of list_del_init(), the uwq on the stack
554                  * will be freed shortly anyway.
555                  */
556                 list_del(&uwq.wq.entry);
557                 spin_unlock(&ctx->fault_pending_wqh.lock);
558         }
559
560         /*
561          * ctx may go away after this if the userfault pseudo fd is
562          * already released.
563          */
564         userfaultfd_ctx_put(ctx);
565
566 out:
567         return ret;
568 }
569
570 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
571                                               struct userfaultfd_wait_queue *ewq)
572 {
573         if (WARN_ON_ONCE(current->flags & PF_EXITING))
574                 goto out;
575
576         ewq->ctx = ctx;
577         init_waitqueue_entry(&ewq->wq, current);
578
579         spin_lock(&ctx->event_wqh.lock);
580         /*
581          * After the __add_wait_queue the uwq is visible to userland
582          * through poll/read().
583          */
584         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
585         for (;;) {
586                 set_current_state(TASK_KILLABLE);
587                 if (ewq->msg.event == 0)
588                         break;
589                 if (READ_ONCE(ctx->released) ||
590                     fatal_signal_pending(current)) {
591                         /*
592                          * &ewq->wq may be queued in fork_event, but
593                          * __remove_wait_queue ignores the head
594                          * parameter. It would be a problem if it
595                          * didn't.
596                          */
597                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
598                         if (ewq->msg.event == UFFD_EVENT_FORK) {
599                                 struct userfaultfd_ctx *new;
600
601                                 new = (struct userfaultfd_ctx *)
602                                         (unsigned long)
603                                         ewq->msg.arg.reserved.reserved1;
604
605                                 userfaultfd_ctx_put(new);
606                         }
607                         break;
608                 }
609
610                 spin_unlock(&ctx->event_wqh.lock);
611
612                 wake_up_poll(&ctx->fd_wqh, POLLIN);
613                 schedule();
614
615                 spin_lock(&ctx->event_wqh.lock);
616         }
617         __set_current_state(TASK_RUNNING);
618         spin_unlock(&ctx->event_wqh.lock);
619
620         /*
621          * ctx may go away after this if the userfault pseudo fd is
622          * already released.
623          */
624 out:
625         userfaultfd_ctx_put(ctx);
626 }
627
628 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
629                                        struct userfaultfd_wait_queue *ewq)
630 {
631         ewq->msg.event = 0;
632         wake_up_locked(&ctx->event_wqh);
633         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
634 }
635
636 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
637 {
638         struct userfaultfd_ctx *ctx = NULL, *octx;
639         struct userfaultfd_fork_ctx *fctx;
640
641         octx = vma->vm_userfaultfd_ctx.ctx;
642         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
643                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
644                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
645                 return 0;
646         }
647
648         list_for_each_entry(fctx, fcs, list)
649                 if (fctx->orig == octx) {
650                         ctx = fctx->new;
651                         break;
652                 }
653
654         if (!ctx) {
655                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
656                 if (!fctx)
657                         return -ENOMEM;
658
659                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
660                 if (!ctx) {
661                         kfree(fctx);
662                         return -ENOMEM;
663                 }
664
665                 atomic_set(&ctx->refcount, 1);
666                 ctx->flags = octx->flags;
667                 ctx->state = UFFD_STATE_RUNNING;
668                 ctx->features = octx->features;
669                 ctx->released = false;
670                 ctx->mm = vma->vm_mm;
671                 atomic_inc(&ctx->mm->mm_count);
672
673                 userfaultfd_ctx_get(octx);
674                 fctx->orig = octx;
675                 fctx->new = ctx;
676                 list_add_tail(&fctx->list, fcs);
677         }
678
679         vma->vm_userfaultfd_ctx.ctx = ctx;
680         return 0;
681 }
682
683 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
684 {
685         struct userfaultfd_ctx *ctx = fctx->orig;
686         struct userfaultfd_wait_queue ewq;
687
688         msg_init(&ewq.msg);
689
690         ewq.msg.event = UFFD_EVENT_FORK;
691         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
692
693         userfaultfd_event_wait_completion(ctx, &ewq);
694 }
695
696 void dup_userfaultfd_complete(struct list_head *fcs)
697 {
698         struct userfaultfd_fork_ctx *fctx, *n;
699
700         list_for_each_entry_safe(fctx, n, fcs, list) {
701                 dup_fctx(fctx);
702                 list_del(&fctx->list);
703                 kfree(fctx);
704         }
705 }
706
707 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
708                              struct vm_userfaultfd_ctx *vm_ctx)
709 {
710         struct userfaultfd_ctx *ctx;
711
712         ctx = vma->vm_userfaultfd_ctx.ctx;
713         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
714                 vm_ctx->ctx = ctx;
715                 userfaultfd_ctx_get(ctx);
716         }
717 }
718
719 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
720                                  unsigned long from, unsigned long to,
721                                  unsigned long len)
722 {
723         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
724         struct userfaultfd_wait_queue ewq;
725
726         if (!ctx)
727                 return;
728
729         if (to & ~PAGE_MASK) {
730                 userfaultfd_ctx_put(ctx);
731                 return;
732         }
733
734         msg_init(&ewq.msg);
735
736         ewq.msg.event = UFFD_EVENT_REMAP;
737         ewq.msg.arg.remap.from = from;
738         ewq.msg.arg.remap.to = to;
739         ewq.msg.arg.remap.len = len;
740
741         userfaultfd_event_wait_completion(ctx, &ewq);
742 }
743
744 bool userfaultfd_remove(struct vm_area_struct *vma,
745                         unsigned long start, unsigned long end)
746 {
747         struct mm_struct *mm = vma->vm_mm;
748         struct userfaultfd_ctx *ctx;
749         struct userfaultfd_wait_queue ewq;
750
751         ctx = vma->vm_userfaultfd_ctx.ctx;
752         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
753                 return true;
754
755         userfaultfd_ctx_get(ctx);
756         up_read(&mm->mmap_sem);
757
758         msg_init(&ewq.msg);
759
760         ewq.msg.event = UFFD_EVENT_REMOVE;
761         ewq.msg.arg.remove.start = start;
762         ewq.msg.arg.remove.end = end;
763
764         userfaultfd_event_wait_completion(ctx, &ewq);
765
766         return false;
767 }
768
769 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
770                           unsigned long start, unsigned long end)
771 {
772         struct userfaultfd_unmap_ctx *unmap_ctx;
773
774         list_for_each_entry(unmap_ctx, unmaps, list)
775                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
776                     unmap_ctx->end == end)
777                         return true;
778
779         return false;
780 }
781
782 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
783                            unsigned long start, unsigned long end,
784                            struct list_head *unmaps)
785 {
786         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
787                 struct userfaultfd_unmap_ctx *unmap_ctx;
788                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
789
790                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
791                     has_unmap_ctx(ctx, unmaps, start, end))
792                         continue;
793
794                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
795                 if (!unmap_ctx)
796                         return -ENOMEM;
797
798                 userfaultfd_ctx_get(ctx);
799                 unmap_ctx->ctx = ctx;
800                 unmap_ctx->start = start;
801                 unmap_ctx->end = end;
802                 list_add_tail(&unmap_ctx->list, unmaps);
803         }
804
805         return 0;
806 }
807
808 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
809 {
810         struct userfaultfd_unmap_ctx *ctx, *n;
811         struct userfaultfd_wait_queue ewq;
812
813         list_for_each_entry_safe(ctx, n, uf, list) {
814                 msg_init(&ewq.msg);
815
816                 ewq.msg.event = UFFD_EVENT_UNMAP;
817                 ewq.msg.arg.remove.start = ctx->start;
818                 ewq.msg.arg.remove.end = ctx->end;
819
820                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
821
822                 list_del(&ctx->list);
823                 kfree(ctx);
824         }
825 }
826
827 static int userfaultfd_release(struct inode *inode, struct file *file)
828 {
829         struct userfaultfd_ctx *ctx = file->private_data;
830         struct mm_struct *mm = ctx->mm;
831         struct vm_area_struct *vma, *prev;
832         /* len == 0 means wake all */
833         struct userfaultfd_wake_range range = { .len = 0, };
834         unsigned long new_flags;
835
836         WRITE_ONCE(ctx->released, true);
837
838         if (!mmget_not_zero(mm))
839                 goto wakeup;
840
841         /*
842          * Flush page faults out of all CPUs. NOTE: all page faults
843          * must be retried without returning VM_FAULT_SIGBUS if
844          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
845          * changes while handle_userfault released the mmap_sem. So
846          * it's critical that released is set to true (above), before
847          * taking the mmap_sem for writing.
848          */
849         down_write(&mm->mmap_sem);
850         prev = NULL;
851         for (vma = mm->mmap; vma; vma = vma->vm_next) {
852                 cond_resched();
853                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
854                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
855                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
856                         prev = vma;
857                         continue;
858                 }
859                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
860                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
861                                  new_flags, vma->anon_vma,
862                                  vma->vm_file, vma->vm_pgoff,
863                                  vma_policy(vma),
864                                  NULL_VM_UFFD_CTX);
865                 if (prev)
866                         vma = prev;
867                 else
868                         prev = vma;
869                 vma->vm_flags = new_flags;
870                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
871         }
872         up_write(&mm->mmap_sem);
873         mmput(mm);
874 wakeup:
875         /*
876          * After no new page faults can wait on this fault_*wqh, flush
877          * the last page faults that may have been already waiting on
878          * the fault_*wqh.
879          */
880         spin_lock(&ctx->fault_pending_wqh.lock);
881         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
882         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
883         spin_unlock(&ctx->fault_pending_wqh.lock);
884
885         /* Flush pending events that may still wait on event_wqh */
886         wake_up_all(&ctx->event_wqh);
887
888         wake_up_poll(&ctx->fd_wqh, POLLHUP);
889         userfaultfd_ctx_put(ctx);
890         return 0;
891 }
892
893 /* fault_pending_wqh.lock must be hold by the caller */
894 static inline struct userfaultfd_wait_queue *find_userfault_in(
895                 wait_queue_head_t *wqh)
896 {
897         wait_queue_entry_t *wq;
898         struct userfaultfd_wait_queue *uwq;
899
900         VM_BUG_ON(!spin_is_locked(&wqh->lock));
901
902         uwq = NULL;
903         if (!waitqueue_active(wqh))
904                 goto out;
905         /* walk in reverse to provide FIFO behavior to read userfaults */
906         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
907         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
908 out:
909         return uwq;
910 }
911
912 static inline struct userfaultfd_wait_queue *find_userfault(
913                 struct userfaultfd_ctx *ctx)
914 {
915         return find_userfault_in(&ctx->fault_pending_wqh);
916 }
917
918 static inline struct userfaultfd_wait_queue *find_userfault_evt(
919                 struct userfaultfd_ctx *ctx)
920 {
921         return find_userfault_in(&ctx->event_wqh);
922 }
923
924 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
925 {
926         struct userfaultfd_ctx *ctx = file->private_data;
927         unsigned int ret;
928
929         poll_wait(file, &ctx->fd_wqh, wait);
930
931         switch (ctx->state) {
932         case UFFD_STATE_WAIT_API:
933                 return POLLERR;
934         case UFFD_STATE_RUNNING:
935                 /*
936                  * poll() never guarantees that read won't block.
937                  * userfaults can be waken before they're read().
938                  */
939                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
940                         return POLLERR;
941                 /*
942                  * lockless access to see if there are pending faults
943                  * __pollwait last action is the add_wait_queue but
944                  * the spin_unlock would allow the waitqueue_active to
945                  * pass above the actual list_add inside
946                  * add_wait_queue critical section. So use a full
947                  * memory barrier to serialize the list_add write of
948                  * add_wait_queue() with the waitqueue_active read
949                  * below.
950                  */
951                 ret = 0;
952                 smp_mb();
953                 if (waitqueue_active(&ctx->fault_pending_wqh))
954                         ret = POLLIN;
955                 else if (waitqueue_active(&ctx->event_wqh))
956                         ret = POLLIN;
957
958                 return ret;
959         default:
960                 WARN_ON_ONCE(1);
961                 return POLLERR;
962         }
963 }
964
965 static const struct file_operations userfaultfd_fops;
966
967 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
968                                   struct userfaultfd_ctx *new,
969                                   struct uffd_msg *msg)
970 {
971         int fd;
972         struct file *file;
973         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
974
975         fd = get_unused_fd_flags(flags);
976         if (fd < 0)
977                 return fd;
978
979         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
980                                   O_RDWR | flags);
981         if (IS_ERR(file)) {
982                 put_unused_fd(fd);
983                 return PTR_ERR(file);
984         }
985
986         fd_install(fd, file);
987         msg->arg.reserved.reserved1 = 0;
988         msg->arg.fork.ufd = fd;
989
990         return 0;
991 }
992
993 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
994                                     struct uffd_msg *msg)
995 {
996         ssize_t ret;
997         DECLARE_WAITQUEUE(wait, current);
998         struct userfaultfd_wait_queue *uwq;
999         /*
1000          * Handling fork event requires sleeping operations, so
1001          * we drop the event_wqh lock, then do these ops, then
1002          * lock it back and wake up the waiter. While the lock is
1003          * dropped the ewq may go away so we keep track of it
1004          * carefully.
1005          */
1006         LIST_HEAD(fork_event);
1007         struct userfaultfd_ctx *fork_nctx = NULL;
1008
1009         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1010         spin_lock(&ctx->fd_wqh.lock);
1011         __add_wait_queue(&ctx->fd_wqh, &wait);
1012         for (;;) {
1013                 set_current_state(TASK_INTERRUPTIBLE);
1014                 spin_lock(&ctx->fault_pending_wqh.lock);
1015                 uwq = find_userfault(ctx);
1016                 if (uwq) {
1017                         /*
1018                          * Use a seqcount to repeat the lockless check
1019                          * in wake_userfault() to avoid missing
1020                          * wakeups because during the refile both
1021                          * waitqueue could become empty if this is the
1022                          * only userfault.
1023                          */
1024                         write_seqcount_begin(&ctx->refile_seq);
1025
1026                         /*
1027                          * The fault_pending_wqh.lock prevents the uwq
1028                          * to disappear from under us.
1029                          *
1030                          * Refile this userfault from
1031                          * fault_pending_wqh to fault_wqh, it's not
1032                          * pending anymore after we read it.
1033                          *
1034                          * Use list_del() by hand (as
1035                          * userfaultfd_wake_function also uses
1036                          * list_del_init() by hand) to be sure nobody
1037                          * changes __remove_wait_queue() to use
1038                          * list_del_init() in turn breaking the
1039                          * !list_empty_careful() check in
1040                          * handle_userfault(). The uwq->wq.head list
1041                          * must never be empty at any time during the
1042                          * refile, or the waitqueue could disappear
1043                          * from under us. The "wait_queue_head_t"
1044                          * parameter of __remove_wait_queue() is unused
1045                          * anyway.
1046                          */
1047                         list_del(&uwq->wq.entry);
1048                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1049
1050                         write_seqcount_end(&ctx->refile_seq);
1051
1052                         /* careful to always initialize msg if ret == 0 */
1053                         *msg = uwq->msg;
1054                         spin_unlock(&ctx->fault_pending_wqh.lock);
1055                         ret = 0;
1056                         break;
1057                 }
1058                 spin_unlock(&ctx->fault_pending_wqh.lock);
1059
1060                 spin_lock(&ctx->event_wqh.lock);
1061                 uwq = find_userfault_evt(ctx);
1062                 if (uwq) {
1063                         *msg = uwq->msg;
1064
1065                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1066                                 fork_nctx = (struct userfaultfd_ctx *)
1067                                         (unsigned long)
1068                                         uwq->msg.arg.reserved.reserved1;
1069                                 list_move(&uwq->wq.entry, &fork_event);
1070                                 /*
1071                                  * fork_nctx can be freed as soon as
1072                                  * we drop the lock, unless we take a
1073                                  * reference on it.
1074                                  */
1075                                 userfaultfd_ctx_get(fork_nctx);
1076                                 spin_unlock(&ctx->event_wqh.lock);
1077                                 ret = 0;
1078                                 break;
1079                         }
1080
1081                         userfaultfd_event_complete(ctx, uwq);
1082                         spin_unlock(&ctx->event_wqh.lock);
1083                         ret = 0;
1084                         break;
1085                 }
1086                 spin_unlock(&ctx->event_wqh.lock);
1087
1088                 if (signal_pending(current)) {
1089                         ret = -ERESTARTSYS;
1090                         break;
1091                 }
1092                 if (no_wait) {
1093                         ret = -EAGAIN;
1094                         break;
1095                 }
1096                 spin_unlock(&ctx->fd_wqh.lock);
1097                 schedule();
1098                 spin_lock(&ctx->fd_wqh.lock);
1099         }
1100         __remove_wait_queue(&ctx->fd_wqh, &wait);
1101         __set_current_state(TASK_RUNNING);
1102         spin_unlock(&ctx->fd_wqh.lock);
1103
1104         if (!ret && msg->event == UFFD_EVENT_FORK) {
1105                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1106                 spin_lock(&ctx->event_wqh.lock);
1107                 if (!list_empty(&fork_event)) {
1108                         /*
1109                          * The fork thread didn't abort, so we can
1110                          * drop the temporary refcount.
1111                          */
1112                         userfaultfd_ctx_put(fork_nctx);
1113
1114                         uwq = list_first_entry(&fork_event,
1115                                                typeof(*uwq),
1116                                                wq.entry);
1117                         /*
1118                          * If fork_event list wasn't empty and in turn
1119                          * the event wasn't already released by fork
1120                          * (the event is allocated on fork kernel
1121                          * stack), put the event back to its place in
1122                          * the event_wq. fork_event head will be freed
1123                          * as soon as we return so the event cannot
1124                          * stay queued there no matter the current
1125                          * "ret" value.
1126                          */
1127                         list_del(&uwq->wq.entry);
1128                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1129
1130                         /*
1131                          * Leave the event in the waitqueue and report
1132                          * error to userland if we failed to resolve
1133                          * the userfault fork.
1134                          */
1135                         if (likely(!ret))
1136                                 userfaultfd_event_complete(ctx, uwq);
1137                 } else {
1138                         /*
1139                          * Here the fork thread aborted and the
1140                          * refcount from the fork thread on fork_nctx
1141                          * has already been released. We still hold
1142                          * the reference we took before releasing the
1143                          * lock above. If resolve_userfault_fork
1144                          * failed we've to drop it because the
1145                          * fork_nctx has to be freed in such case. If
1146                          * it succeeded we'll hold it because the new
1147                          * uffd references it.
1148                          */
1149                         if (ret)
1150                                 userfaultfd_ctx_put(fork_nctx);
1151                 }
1152                 spin_unlock(&ctx->event_wqh.lock);
1153         }
1154
1155         return ret;
1156 }
1157
1158 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1159                                 size_t count, loff_t *ppos)
1160 {
1161         struct userfaultfd_ctx *ctx = file->private_data;
1162         ssize_t _ret, ret = 0;
1163         struct uffd_msg msg;
1164         int no_wait = file->f_flags & O_NONBLOCK;
1165
1166         if (ctx->state == UFFD_STATE_WAIT_API)
1167                 return -EINVAL;
1168
1169         for (;;) {
1170                 if (count < sizeof(msg))
1171                         return ret ? ret : -EINVAL;
1172                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1173                 if (_ret < 0)
1174                         return ret ? ret : _ret;
1175                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1176                         return ret ? ret : -EFAULT;
1177                 ret += sizeof(msg);
1178                 buf += sizeof(msg);
1179                 count -= sizeof(msg);
1180                 /*
1181                  * Allow to read more than one fault at time but only
1182                  * block if waiting for the very first one.
1183                  */
1184                 no_wait = O_NONBLOCK;
1185         }
1186 }
1187
1188 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1189                              struct userfaultfd_wake_range *range)
1190 {
1191         spin_lock(&ctx->fault_pending_wqh.lock);
1192         /* wake all in the range and autoremove */
1193         if (waitqueue_active(&ctx->fault_pending_wqh))
1194                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1195                                      range);
1196         if (waitqueue_active(&ctx->fault_wqh))
1197                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1198         spin_unlock(&ctx->fault_pending_wqh.lock);
1199 }
1200
1201 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1202                                            struct userfaultfd_wake_range *range)
1203 {
1204         unsigned seq;
1205         bool need_wakeup;
1206
1207         /*
1208          * To be sure waitqueue_active() is not reordered by the CPU
1209          * before the pagetable update, use an explicit SMP memory
1210          * barrier here. PT lock release or up_read(mmap_sem) still
1211          * have release semantics that can allow the
1212          * waitqueue_active() to be reordered before the pte update.
1213          */
1214         smp_mb();
1215
1216         /*
1217          * Use waitqueue_active because it's very frequent to
1218          * change the address space atomically even if there are no
1219          * userfaults yet. So we take the spinlock only when we're
1220          * sure we've userfaults to wake.
1221          */
1222         do {
1223                 seq = read_seqcount_begin(&ctx->refile_seq);
1224                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1225                         waitqueue_active(&ctx->fault_wqh);
1226                 cond_resched();
1227         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1228         if (need_wakeup)
1229                 __wake_userfault(ctx, range);
1230 }
1231
1232 static __always_inline int validate_range(struct mm_struct *mm,
1233                                           __u64 start, __u64 len)
1234 {
1235         __u64 task_size = mm->task_size;
1236
1237         if (start & ~PAGE_MASK)
1238                 return -EINVAL;
1239         if (len & ~PAGE_MASK)
1240                 return -EINVAL;
1241         if (!len)
1242                 return -EINVAL;
1243         if (start < mmap_min_addr)
1244                 return -EINVAL;
1245         if (start >= task_size)
1246                 return -EINVAL;
1247         if (len > task_size - start)
1248                 return -EINVAL;
1249         return 0;
1250 }
1251
1252 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1253 {
1254         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1255                 vma_is_shmem(vma);
1256 }
1257
1258 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1259                                 unsigned long arg)
1260 {
1261         struct mm_struct *mm = ctx->mm;
1262         struct vm_area_struct *vma, *prev, *cur;
1263         int ret;
1264         struct uffdio_register uffdio_register;
1265         struct uffdio_register __user *user_uffdio_register;
1266         unsigned long vm_flags, new_flags;
1267         bool found;
1268         bool basic_ioctls;
1269         unsigned long start, end, vma_end;
1270
1271         user_uffdio_register = (struct uffdio_register __user *) arg;
1272
1273         ret = -EFAULT;
1274         if (copy_from_user(&uffdio_register, user_uffdio_register,
1275                            sizeof(uffdio_register)-sizeof(__u64)))
1276                 goto out;
1277
1278         ret = -EINVAL;
1279         if (!uffdio_register.mode)
1280                 goto out;
1281         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1282                                      UFFDIO_REGISTER_MODE_WP))
1283                 goto out;
1284         vm_flags = 0;
1285         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1286                 vm_flags |= VM_UFFD_MISSING;
1287         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1288                 vm_flags |= VM_UFFD_WP;
1289                 /*
1290                  * FIXME: remove the below error constraint by
1291                  * implementing the wprotect tracking mode.
1292                  */
1293                 ret = -EINVAL;
1294                 goto out;
1295         }
1296
1297         ret = validate_range(mm, uffdio_register.range.start,
1298                              uffdio_register.range.len);
1299         if (ret)
1300                 goto out;
1301
1302         start = uffdio_register.range.start;
1303         end = start + uffdio_register.range.len;
1304
1305         ret = -ENOMEM;
1306         if (!mmget_not_zero(mm))
1307                 goto out;
1308
1309         down_write(&mm->mmap_sem);
1310         vma = find_vma_prev(mm, start, &prev);
1311         if (!vma)
1312                 goto out_unlock;
1313
1314         /* check that there's at least one vma in the range */
1315         ret = -EINVAL;
1316         if (vma->vm_start >= end)
1317                 goto out_unlock;
1318
1319         /*
1320          * If the first vma contains huge pages, make sure start address
1321          * is aligned to huge page size.
1322          */
1323         if (is_vm_hugetlb_page(vma)) {
1324                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1325
1326                 if (start & (vma_hpagesize - 1))
1327                         goto out_unlock;
1328         }
1329
1330         /*
1331          * Search for not compatible vmas.
1332          */
1333         found = false;
1334         basic_ioctls = false;
1335         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1336                 cond_resched();
1337
1338                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1339                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1340
1341                 /* check not compatible vmas */
1342                 ret = -EINVAL;
1343                 if (!vma_can_userfault(cur))
1344                         goto out_unlock;
1345                 /*
1346                  * If this vma contains ending address, and huge pages
1347                  * check alignment.
1348                  */
1349                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1350                     end > cur->vm_start) {
1351                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1352
1353                         ret = -EINVAL;
1354
1355                         if (end & (vma_hpagesize - 1))
1356                                 goto out_unlock;
1357                 }
1358
1359                 /*
1360                  * Check that this vma isn't already owned by a
1361                  * different userfaultfd. We can't allow more than one
1362                  * userfaultfd to own a single vma simultaneously or we
1363                  * wouldn't know which one to deliver the userfaults to.
1364                  */
1365                 ret = -EBUSY;
1366                 if (cur->vm_userfaultfd_ctx.ctx &&
1367                     cur->vm_userfaultfd_ctx.ctx != ctx)
1368                         goto out_unlock;
1369
1370                 /*
1371                  * Note vmas containing huge pages
1372                  */
1373                 if (is_vm_hugetlb_page(cur))
1374                         basic_ioctls = true;
1375
1376                 found = true;
1377         }
1378         BUG_ON(!found);
1379
1380         if (vma->vm_start < start)
1381                 prev = vma;
1382
1383         ret = 0;
1384         do {
1385                 cond_resched();
1386
1387                 BUG_ON(!vma_can_userfault(vma));
1388                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1389                        vma->vm_userfaultfd_ctx.ctx != ctx);
1390
1391                 /*
1392                  * Nothing to do: this vma is already registered into this
1393                  * userfaultfd and with the right tracking mode too.
1394                  */
1395                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1396                     (vma->vm_flags & vm_flags) == vm_flags)
1397                         goto skip;
1398
1399                 if (vma->vm_start > start)
1400                         start = vma->vm_start;
1401                 vma_end = min(end, vma->vm_end);
1402
1403                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1404                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1405                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1406                                  vma_policy(vma),
1407                                  ((struct vm_userfaultfd_ctx){ ctx }));
1408                 if (prev) {
1409                         vma = prev;
1410                         goto next;
1411                 }
1412                 if (vma->vm_start < start) {
1413                         ret = split_vma(mm, vma, start, 1);
1414                         if (ret)
1415                                 break;
1416                 }
1417                 if (vma->vm_end > end) {
1418                         ret = split_vma(mm, vma, end, 0);
1419                         if (ret)
1420                                 break;
1421                 }
1422         next:
1423                 /*
1424                  * In the vma_merge() successful mprotect-like case 8:
1425                  * the next vma was merged into the current one and
1426                  * the current one has not been updated yet.
1427                  */
1428                 vma->vm_flags = new_flags;
1429                 vma->vm_userfaultfd_ctx.ctx = ctx;
1430
1431         skip:
1432                 prev = vma;
1433                 start = vma->vm_end;
1434                 vma = vma->vm_next;
1435         } while (vma && vma->vm_start < end);
1436 out_unlock:
1437         up_write(&mm->mmap_sem);
1438         mmput(mm);
1439         if (!ret) {
1440                 /*
1441                  * Now that we scanned all vmas we can already tell
1442                  * userland which ioctls methods are guaranteed to
1443                  * succeed on this range.
1444                  */
1445                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1446                              UFFD_API_RANGE_IOCTLS,
1447                              &user_uffdio_register->ioctls))
1448                         ret = -EFAULT;
1449         }
1450 out:
1451         return ret;
1452 }
1453
1454 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1455                                   unsigned long arg)
1456 {
1457         struct mm_struct *mm = ctx->mm;
1458         struct vm_area_struct *vma, *prev, *cur;
1459         int ret;
1460         struct uffdio_range uffdio_unregister;
1461         unsigned long new_flags;
1462         bool found;
1463         unsigned long start, end, vma_end;
1464         const void __user *buf = (void __user *)arg;
1465
1466         ret = -EFAULT;
1467         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1468                 goto out;
1469
1470         ret = validate_range(mm, uffdio_unregister.start,
1471                              uffdio_unregister.len);
1472         if (ret)
1473                 goto out;
1474
1475         start = uffdio_unregister.start;
1476         end = start + uffdio_unregister.len;
1477
1478         ret = -ENOMEM;
1479         if (!mmget_not_zero(mm))
1480                 goto out;
1481
1482         down_write(&mm->mmap_sem);
1483         vma = find_vma_prev(mm, start, &prev);
1484         if (!vma)
1485                 goto out_unlock;
1486
1487         /* check that there's at least one vma in the range */
1488         ret = -EINVAL;
1489         if (vma->vm_start >= end)
1490                 goto out_unlock;
1491
1492         /*
1493          * If the first vma contains huge pages, make sure start address
1494          * is aligned to huge page size.
1495          */
1496         if (is_vm_hugetlb_page(vma)) {
1497                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1498
1499                 if (start & (vma_hpagesize - 1))
1500                         goto out_unlock;
1501         }
1502
1503         /*
1504          * Search for not compatible vmas.
1505          */
1506         found = false;
1507         ret = -EINVAL;
1508         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1509                 cond_resched();
1510
1511                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1512                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1513
1514                 /*
1515                  * Check not compatible vmas, not strictly required
1516                  * here as not compatible vmas cannot have an
1517                  * userfaultfd_ctx registered on them, but this
1518                  * provides for more strict behavior to notice
1519                  * unregistration errors.
1520                  */
1521                 if (!vma_can_userfault(cur))
1522                         goto out_unlock;
1523
1524                 found = true;
1525         }
1526         BUG_ON(!found);
1527
1528         if (vma->vm_start < start)
1529                 prev = vma;
1530
1531         ret = 0;
1532         do {
1533                 cond_resched();
1534
1535                 BUG_ON(!vma_can_userfault(vma));
1536
1537                 /*
1538                  * Nothing to do: this vma is already registered into this
1539                  * userfaultfd and with the right tracking mode too.
1540                  */
1541                 if (!vma->vm_userfaultfd_ctx.ctx)
1542                         goto skip;
1543
1544                 if (vma->vm_start > start)
1545                         start = vma->vm_start;
1546                 vma_end = min(end, vma->vm_end);
1547
1548                 if (userfaultfd_missing(vma)) {
1549                         /*
1550                          * Wake any concurrent pending userfault while
1551                          * we unregister, so they will not hang
1552                          * permanently and it avoids userland to call
1553                          * UFFDIO_WAKE explicitly.
1554                          */
1555                         struct userfaultfd_wake_range range;
1556                         range.start = start;
1557                         range.len = vma_end - start;
1558                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1559                 }
1560
1561                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1562                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1563                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1564                                  vma_policy(vma),
1565                                  NULL_VM_UFFD_CTX);
1566                 if (prev) {
1567                         vma = prev;
1568                         goto next;
1569                 }
1570                 if (vma->vm_start < start) {
1571                         ret = split_vma(mm, vma, start, 1);
1572                         if (ret)
1573                                 break;
1574                 }
1575                 if (vma->vm_end > end) {
1576                         ret = split_vma(mm, vma, end, 0);
1577                         if (ret)
1578                                 break;
1579                 }
1580         next:
1581                 /*
1582                  * In the vma_merge() successful mprotect-like case 8:
1583                  * the next vma was merged into the current one and
1584                  * the current one has not been updated yet.
1585                  */
1586                 vma->vm_flags = new_flags;
1587                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1588
1589         skip:
1590                 prev = vma;
1591                 start = vma->vm_end;
1592                 vma = vma->vm_next;
1593         } while (vma && vma->vm_start < end);
1594 out_unlock:
1595         up_write(&mm->mmap_sem);
1596         mmput(mm);
1597 out:
1598         return ret;
1599 }
1600
1601 /*
1602  * userfaultfd_wake may be used in combination with the
1603  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1604  */
1605 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1606                             unsigned long arg)
1607 {
1608         int ret;
1609         struct uffdio_range uffdio_wake;
1610         struct userfaultfd_wake_range range;
1611         const void __user *buf = (void __user *)arg;
1612
1613         ret = -EFAULT;
1614         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1615                 goto out;
1616
1617         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1618         if (ret)
1619                 goto out;
1620
1621         range.start = uffdio_wake.start;
1622         range.len = uffdio_wake.len;
1623
1624         /*
1625          * len == 0 means wake all and we don't want to wake all here,
1626          * so check it again to be sure.
1627          */
1628         VM_BUG_ON(!range.len);
1629
1630         wake_userfault(ctx, &range);
1631         ret = 0;
1632
1633 out:
1634         return ret;
1635 }
1636
1637 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1638                             unsigned long arg)
1639 {
1640         __s64 ret;
1641         struct uffdio_copy uffdio_copy;
1642         struct uffdio_copy __user *user_uffdio_copy;
1643         struct userfaultfd_wake_range range;
1644
1645         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1646
1647         ret = -EFAULT;
1648         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1649                            /* don't copy "copy" last field */
1650                            sizeof(uffdio_copy)-sizeof(__s64)))
1651                 goto out;
1652
1653         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1654         if (ret)
1655                 goto out;
1656         /*
1657          * double check for wraparound just in case. copy_from_user()
1658          * will later check uffdio_copy.src + uffdio_copy.len to fit
1659          * in the userland range.
1660          */
1661         ret = -EINVAL;
1662         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1663                 goto out;
1664         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1665                 goto out;
1666         if (mmget_not_zero(ctx->mm)) {
1667                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1668                                    uffdio_copy.len);
1669                 mmput(ctx->mm);
1670         } else {
1671                 return -ESRCH;
1672         }
1673         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1674                 return -EFAULT;
1675         if (ret < 0)
1676                 goto out;
1677         BUG_ON(!ret);
1678         /* len == 0 would wake all */
1679         range.len = ret;
1680         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1681                 range.start = uffdio_copy.dst;
1682                 wake_userfault(ctx, &range);
1683         }
1684         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1685 out:
1686         return ret;
1687 }
1688
1689 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1690                                 unsigned long arg)
1691 {
1692         __s64 ret;
1693         struct uffdio_zeropage uffdio_zeropage;
1694         struct uffdio_zeropage __user *user_uffdio_zeropage;
1695         struct userfaultfd_wake_range range;
1696
1697         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1698
1699         ret = -EFAULT;
1700         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1701                            /* don't copy "zeropage" last field */
1702                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1703                 goto out;
1704
1705         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1706                              uffdio_zeropage.range.len);
1707         if (ret)
1708                 goto out;
1709         ret = -EINVAL;
1710         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1711                 goto out;
1712
1713         if (mmget_not_zero(ctx->mm)) {
1714                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1715                                      uffdio_zeropage.range.len);
1716                 mmput(ctx->mm);
1717         } else {
1718                 return -ESRCH;
1719         }
1720         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1721                 return -EFAULT;
1722         if (ret < 0)
1723                 goto out;
1724         /* len == 0 would wake all */
1725         BUG_ON(!ret);
1726         range.len = ret;
1727         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1728                 range.start = uffdio_zeropage.range.start;
1729                 wake_userfault(ctx, &range);
1730         }
1731         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1732 out:
1733         return ret;
1734 }
1735
1736 static inline unsigned int uffd_ctx_features(__u64 user_features)
1737 {
1738         /*
1739          * For the current set of features the bits just coincide
1740          */
1741         return (unsigned int)user_features;
1742 }
1743
1744 /*
1745  * userland asks for a certain API version and we return which bits
1746  * and ioctl commands are implemented in this kernel for such API
1747  * version or -EINVAL if unknown.
1748  */
1749 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1750                            unsigned long arg)
1751 {
1752         struct uffdio_api uffdio_api;
1753         void __user *buf = (void __user *)arg;
1754         int ret;
1755         __u64 features;
1756
1757         ret = -EINVAL;
1758         if (ctx->state != UFFD_STATE_WAIT_API)
1759                 goto out;
1760         ret = -EFAULT;
1761         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1762                 goto out;
1763         features = uffdio_api.features;
1764         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1765                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1766                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1767                         goto out;
1768                 ret = -EINVAL;
1769                 goto out;
1770         }
1771         /* report all available features and ioctls to userland */
1772         uffdio_api.features = UFFD_API_FEATURES;
1773         uffdio_api.ioctls = UFFD_API_IOCTLS;
1774         ret = -EFAULT;
1775         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1776                 goto out;
1777         ctx->state = UFFD_STATE_RUNNING;
1778         /* only enable the requested features for this uffd context */
1779         ctx->features = uffd_ctx_features(features);
1780         ret = 0;
1781 out:
1782         return ret;
1783 }
1784
1785 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1786                               unsigned long arg)
1787 {
1788         int ret = -EINVAL;
1789         struct userfaultfd_ctx *ctx = file->private_data;
1790
1791         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1792                 return -EINVAL;
1793
1794         switch(cmd) {
1795         case UFFDIO_API:
1796                 ret = userfaultfd_api(ctx, arg);
1797                 break;
1798         case UFFDIO_REGISTER:
1799                 ret = userfaultfd_register(ctx, arg);
1800                 break;
1801         case UFFDIO_UNREGISTER:
1802                 ret = userfaultfd_unregister(ctx, arg);
1803                 break;
1804         case UFFDIO_WAKE:
1805                 ret = userfaultfd_wake(ctx, arg);
1806                 break;
1807         case UFFDIO_COPY:
1808                 ret = userfaultfd_copy(ctx, arg);
1809                 break;
1810         case UFFDIO_ZEROPAGE:
1811                 ret = userfaultfd_zeropage(ctx, arg);
1812                 break;
1813         }
1814         return ret;
1815 }
1816
1817 #ifdef CONFIG_PROC_FS
1818 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1819 {
1820         struct userfaultfd_ctx *ctx = f->private_data;
1821         wait_queue_entry_t *wq;
1822         struct userfaultfd_wait_queue *uwq;
1823         unsigned long pending = 0, total = 0;
1824
1825         spin_lock(&ctx->fault_pending_wqh.lock);
1826         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1827                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1828                 pending++;
1829                 total++;
1830         }
1831         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1832                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1833                 total++;
1834         }
1835         spin_unlock(&ctx->fault_pending_wqh.lock);
1836
1837         /*
1838          * If more protocols will be added, there will be all shown
1839          * separated by a space. Like this:
1840          *      protocols: aa:... bb:...
1841          */
1842         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1843                    pending, total, UFFD_API, ctx->features,
1844                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1845 }
1846 #endif
1847
1848 static const struct file_operations userfaultfd_fops = {
1849 #ifdef CONFIG_PROC_FS
1850         .show_fdinfo    = userfaultfd_show_fdinfo,
1851 #endif
1852         .release        = userfaultfd_release,
1853         .poll           = userfaultfd_poll,
1854         .read           = userfaultfd_read,
1855         .unlocked_ioctl = userfaultfd_ioctl,
1856         .compat_ioctl   = userfaultfd_ioctl,
1857         .llseek         = noop_llseek,
1858 };
1859
1860 static void init_once_userfaultfd_ctx(void *mem)
1861 {
1862         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1863
1864         init_waitqueue_head(&ctx->fault_pending_wqh);
1865         init_waitqueue_head(&ctx->fault_wqh);
1866         init_waitqueue_head(&ctx->event_wqh);
1867         init_waitqueue_head(&ctx->fd_wqh);
1868         seqcount_init(&ctx->refile_seq);
1869 }
1870
1871 /**
1872  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1873  * @flags: Flags for the userfaultfd file.
1874  *
1875  * This function creates a userfaultfd file pointer, w/out installing
1876  * it into the fd table. This is useful when the userfaultfd file is
1877  * used during the initialization of data structures that require
1878  * extra setup after the userfaultfd creation. So the userfaultfd
1879  * creation is split into the file pointer creation phase, and the
1880  * file descriptor installation phase.  In this way races with
1881  * userspace closing the newly installed file descriptor can be
1882  * avoided.  Returns a userfaultfd file pointer, or a proper error
1883  * pointer.
1884  */
1885 static struct file *userfaultfd_file_create(int flags)
1886 {
1887         struct file *file;
1888         struct userfaultfd_ctx *ctx;
1889
1890         BUG_ON(!current->mm);
1891
1892         /* Check the UFFD_* constants for consistency.  */
1893         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1894         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1895
1896         file = ERR_PTR(-EINVAL);
1897         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1898                 goto out;
1899
1900         file = ERR_PTR(-ENOMEM);
1901         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1902         if (!ctx)
1903                 goto out;
1904
1905         atomic_set(&ctx->refcount, 1);
1906         ctx->flags = flags;
1907         ctx->features = 0;
1908         ctx->state = UFFD_STATE_WAIT_API;
1909         ctx->released = false;
1910         ctx->mm = current->mm;
1911         /* prevent the mm struct to be freed */
1912         mmgrab(ctx->mm);
1913
1914         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1915                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1916         if (IS_ERR(file)) {
1917                 mmdrop(ctx->mm);
1918                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1919         }
1920 out:
1921         return file;
1922 }
1923
1924 SYSCALL_DEFINE1(userfaultfd, int, flags)
1925 {
1926         int fd, error;
1927         struct file *file;
1928
1929         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1930         if (error < 0)
1931                 return error;
1932         fd = error;
1933
1934         file = userfaultfd_file_create(flags);
1935         if (IS_ERR(file)) {
1936                 error = PTR_ERR(file);
1937                 goto err_put_unused_fd;
1938         }
1939         fd_install(fd, file);
1940
1941         return fd;
1942
1943 err_put_unused_fd:
1944         put_unused_fd(fd);
1945
1946         return error;
1947 }
1948
1949 static int __init userfaultfd_init(void)
1950 {
1951         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1952                                                 sizeof(struct userfaultfd_ctx),
1953                                                 0,
1954                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1955                                                 init_once_userfaultfd_ctx);
1956         return 0;
1957 }
1958 __initcall(userfaultfd_init);